content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Working Group 6 (WG6), theory and simulation, of the third edition of the European Advanced Accelerator Workshop has been characterized by a strong numerical connotation: massively parallel particle in cell numerical simulations. It became evident that plasma and laser wakefield acceleration studies are getting constantly more complex and demanding so that numerical tools are inevitable. While theory and analytical models remain necessary to clarify the relevant underlying physical mechanisms, numerics is the never ending help to support, drive and validate these theories.
WG6 presentations could be subdivided into three main streams approaches: numerical simulations in support of physical problems or experimental campaigns, numerical simulations used to uncover new underlying mechanisms and the development of new numerical schemes in support of the previous goals. WG6 focused its attention on three main scientific topics: the mitigation of the numerical Cherenkov radiation, the mitigation of the hose-instability and on start-to-end simulations.
The theoretical approaches has focused on the understanding of the beam evolution and matching for a plasma wakefield acceleration case, with specific attention to the driver bunch for short and long running distances, and to exact symplectic approaches.
In summary the four sessions, and so the discussed topics, could be itemized as follow:
\begin{enumerate}
\item Analytical models
\begin{itemize}
\item Transverse bunch matching
\end{itemize}
\item Numerical Results
\begin{itemize}
\item Code development
\item Code update
\end{itemize}
\item Numerical Methods
\begin{itemize}
\item Numerical Cherenkov mitigation
\item Spectral methods PSATD
\item Adaptive Mesh Refinement (AMR)
\item Envelope
\end{itemize}
\item Scientific computing
\begin{itemize}
\item Dynamic load balancing
\item Data Reduction
\item In situ visalization
\item Use of new libraries
\end{itemize}
\end{enumerate}
\section{Analytical models and numerical results}
The presented analytical models focused on the development of a closed form model to study the channel centroid in hosing instability for PWFA \cite{robson}, the approach considered leverage on a moments approach (very similar to the Boltzmann one, but in the form and fashion originally developed by Maxwell himself). A similar approach has also been used to study and predict the emittance evolution for an external injetion case \cite{aschikhin}. Along the same problem a model to predict the equilibrium condition for a driver bunch has been discusses in \cite{lotov}. The model consider that the head erosion plays a key role, but after several betatron oscillations, the driver bunch will eventually reach a stable shape and conformation, the theory is used to find this final configuration. The same author also highlights how numerical simulations generally are in need of a relatively short window simulation box, but one has to consider that the AWAKE experiment would need boxes that are three orders of magnitude longer \cite{lotov}, requiring new strategies and new approaches. Some theory has also been proposed for TNSA (target normal sheet acceleration), where with a simple electrostatic approach mixed with some initial conditions obtained by particle in cell code simulation, it is possible to calculate the maximum energy that can be reached by particles \cite{sinigardi}.
At the transition between theory and simulation, we find the evolution of the symplectic approach that allows for exact solution but (at present) still with computational constraints. Moving to more computationally oriented presentations, a possible strategy to mitigate the beam hosing instability with a betatron chirp in the quasi linear regime has been presented \cite{lehe}. External injection, with some start to end approach, loading particles from external tracking codes and remapping the background density where the bunches evolve from a MHD code, has also been presented \cite{marocchino}. Ionization has also played a key role in the presentations, and it has been considered as a valuable approach to produce in a controlled self-injected way a high brightness trailing bunch; this technique has been proposed both for the case of PWFA \cite{mira} as well as for LWFA \cite{lee}. LWFA new approaches, such as a bi-color stack of sub-joule pulses has been considered for Thomson scattering based gamma ray source \cite{Kalmykov}. The bi-color is resilient to degradation of the dense plasma and electron beam quality is preserved.
\section{Numerical methods and implementations}
Numerical methods and their implementations are quickly improving in quality but also evolving in their implementations in order to follow the fast evolution of super-computers.
Avoiding numerical Cherenkov effects, resulting from the unphysical interaction of relativistic macro-particles with their own field, is a priority
in accelerator simulations.
It appears that Pseudo-Spectral Analytical Time Domain (PSATD) solvers, virtually free of numerical dispersion, are not subject to numerical Cherenkov radiation while retaining satisying parallelization capabilities\cite{Jalas}.
Moreover, they can be used in a Galilean frame, co-moving with the plasma, in order to suppress numercial Cherenkov instability as well.
This allows simulations in a Lorentz boosted frame and significant speed-ups without compromising the accuracy of the results\cite{Kirchen}.
Another suggested way to improve accuracy is the numerical noise control technique. Controlled additional noise is introduced in a series of simulations in order to extrapolate the results to the very low noise level domain. This ultra low noise level is otherwise unreachable because it would require unreasonable numerical resolution\cite{Spitsyn}.
The WarpX project, targeting PIC simulations on exascale systems was presented\cite{Vay}.
It relies on the already existing Warp code and the PICSAR library for the high performance computing (HPC) kernels thus combining very advanced algorithms.
It is also going to provide adaptative mesh refinement (AMR) by having overlapping grids of different resolutions.
Refined grids are framed by perfectly matched layers (PML) boundary conditions and driven by the surrounding coarser grid.
The latest advances in the PIConGPU code demonstrated that a very efficient cross-platform portability is achieved with a reasonable code size and complexity by using smart abstraction techniques and external libraries \cite{Huebl}.
Advanced physics and diagnostics were also reported in the PIConGPU code \cite{Garten, Pausch}.
The ability to describe and take into account photon generation and their complex interaction with plasmas, including atomic physics, leads to a
brand new kind of simulations.
This puts additional experiments within the reach of PIC simulations even if, in some cases, the PIC cycle represents only a fraction of the compute time.
\section{Code Comparison and Benchmarking}
Code comparison and code benchmarking is a common problem in the whole field and is still an \textit{open question}. Numerical codes are generally tested against theoretical solutions, however full theoretical solutions only exist for very simplified cases where numerical solutions themselves would not offer much more than pure mathematics. The interest is to use numerical codes where theory can, at present, hardly get. In \textit{classical numerics}, by solving simplified problems we can control the error that depends on the chosen numerical scheme and the resolution. In a \textit{multi-physics} scenario, as today codes are sometimes referred to, due to the variety of techniques used to integrate, interpolate and weighting, it is rather difficult to understand how far is the numerical solution from an ideal closed-form analytical solution. A possible way to tackle this issue is by comparing different codes with each other. Code comparison is a strategy that allows to verify that implementations are consistent and that base models are well aligned and coherent (theory is correct). Furthermore, understanding if all the relevant physical mechanisms have been taken into account requires comparisons with experiments. These might be rather difficult due to the large number of experimentally uncontrollable variables and physical mechanisms that concur during a single experiment. Designing a simple and clear experiment against which codes could be benchmarked might not be an easy task. Presentations during the workshop showed a clear motivation to work toward these goals. A clear example is the effort towards the \textit{synthetic diagnostics}: numerical codes are trying to reproduce laboratory diagnostics to shorten the distance between theory and laboratory experiments.
\section{Conclusions}
Theories and simulations of advanced accelerators is progressing quickly.
Software developpers are doing their best to keep up with most experimental ambitions but also with the constantly evolving super-computers.
Very few setups are now out of reach of these tools but challenges of a different kind remains.
The first challenge is the increasing variety of techniques eventually leading to a lack of confidence in the different methods.
As explained in the previous section, thorough cross checking and benchmarking is now absolutely mandatory to guarantee the accuracy of the results.
Nevertheless, with the noticeable exception of openPMD, the lack of accepted standards makes this work terribly difficult.
The second challenge is the unprecedented numerical complexity which creates very high level entry barriers to the most efficient tools.
In this regard, collaboration between physicists and scientific computing experts is absolutely recommended.
\renewcommand{\baselinestretch}{1.0}
|
\section{Introduction}
Ever since the original proposal of Kaluza-Klein (KK) regarding the existence of extra spatial
dimension(s) it is often believed that our universe is a 3-brane embedded in a higher dimensional
spacetime and is described through a low energy
effective theory on the brane carrying the signatures of extra dimensions \cite{kanno,shiromizu}. Depending on
different possible compactification schemes for the extra dimensions, a large number of models
\cite{arkani,horava,RS,kaloper,cohen,burgess,chodos} have been constructed, and
their predictions are yet to be observed in the current experiments.\\
Among various extra dimensional models proposed over the last several decades, Randall-Sundrum (RS) warped extra dimensional model \cite{RS} earned a
special attention since it can resolve the gauge hierarchy problem without introducing any
intermediate scale (between Planck and TeV) in the theory. In RS model
the interbrane separation (known as modulus or radion) is assumed to be $\sim$ Planck length and
generates the required hierarchy between the branes.\\
A suitable potential with a stable minimum is therefore needed for modulus stabilization. Goldberger
and Wise (GW) proposed a useful stabilization mechanism \cite{GW} by introducing
a massive scalar field in the bulk with appropriate boundary values. Though the backreaction
of the stabilizing scalar field was originally neglected
in GW proposal, its implications are subsequently studied in \cite{csaki, tp2}. It has been demonstrated in \cite{tp2}
that the modulus of RS scenario can be stabilized using GW prescription even by incorporating the
backreaction of the stabilizing field. Not only that, even the
stable value of the modulus appears as a parameter in the low
energy effective theory on the brane, but it's fluctuation about that stable value leads
to dynamical modulus (or radion) field which couples to the fields on the
observable brane. This attracted
a large volume of work on phenomenological and cosmological implications \cite{csaki,GW_radion,julien,wolfe,sumanta,sorbo}
of modulus field in RS warped geometry model. This radion phenomenology
along with the study of RS graviton \cite{dhr,rizzo,yong,dhr1,thomas}
are considered to be the testing ground of warped extra dimensional models in collider experiments \cite{atlas1,atlas2}.\\
Apart from phenomenological setup, here we are interested to provide a possible testing ground for the existence of warped extra dimension
from the angle of stable stellar structure namely the Buchdahl's limit.\\
There have been considerable interest in the compactness limit of any stellar structure,
which, originally initiated by Buchdahl, indicates that under reasonable assumptions the minimum radius of a star
has to be greater than $(9/8)$ of its Schwarzschild radius \cite{1,2,3}. These assumptions involve nature of the density
of the star, which has to be decreasing outwards and also the interior solution has to be matched with an exterior one. This raises an intriguing
question, how is the above limit modified if one considers a theory of gravity different from general relativity. This resulted
into a large number of work quite extensively in recent times \cite{4,5,6,7,8,10,11,12,13,14} (see also \cite{15,16,17,18,19}).\\
The important questions that remain are:
\begin{enumerate}
\item How does the compactness limit (known as Buchdahl's limit) of a stellar structure modify due to the presence
of radion field which carries the footprint of compactified warped extra dimension on our visible universe?
\item Can the modified Buchdahl's limit be a possible testing ground for such compactified extra dimension?
\end{enumerate}
We address these questions in the present paper from the perspective of four dimensional effective theory.\\
Our paper is organized as follows. In section [2], we describe the model. In section [3] we find the possible modifications of
Buchdahl's limit and its implications are discussed. In section [4], the interior spacetime of the stellar object is
matched with a suitable exterior one. Finally we end the paper with some conclusive remarks.\\
\section{The model}
We consider a five dimensional AdS spacetime involving one warped and compact
extra spacelike dimension. The spacetime is $S^1/Z_2$ orbifolded along
the extra dimensional angular coordinate $\phi$, where the fixed points $\phi=(0,\pi)$ are
identified with two 3-branes ($3+1$ dimensional), known as Planck (or hidden), TeV (or visible) brane
respectively. Our usual four dimensional universe is the TeV brane and emerges as 4D effective theory.
The opposite brane tensions along with the finely
tuned five dimensional cosmological constant serve as energy-momentum tensor of the aforementioned configuration.\\
In higher dimensional braneworld scenario, the stabilization of extra dimensional
modulus is a crucial aspect and needs to be addressed carefully.
It has been demonstrated by Goldberger and Wise that the modulus corresponding to the
radius of the extra dimension in warped geometry models can be stabilized \cite{GW} by invoking a massive
scalar field in the bulk with non zero value on the branes.\\
Keeping the stabilization mechanism in mind, the braneworld setup considered in the
present context is represented by the following action:
\begin{eqnarray}
S&=&\int d^5x \sqrt{-G} \bigg[M^3R - \Lambda - (1/2)G^{MN}\partial_M\Phi\partial_N\Phi-V(\Phi)\bigg]\nonumber\\
&-&\int d^5x \sqrt{-G}\bigg[\lambda_{hid}\delta(\phi) + \lambda_{vis}\delta(\phi-\pi)\bigg]
\label{action}
\end{eqnarray}
where $M$ is the five dimensional Planck scale, $G_{MN}$ is the
five dimensional metric. $\Lambda$
symbolizes the bulk cosmological constant, $\Phi$ is the
scalar field endowed with a potential $V(\Phi)$,
$\lambda_{hid}$, $\lambda_{vis}$ are the self interactions of scalar field (including brane
tensions) on Planck, TeV branes.\\
We consider the metric ansatz as follows,
\begin{equation}
ds^2 = \exp{[-2A(\phi)]}\eta_{\mu\nu}dx^{\mu}dx^{\nu} + r_c^2d\phi^2
\label{metric ansatz}
\end{equation}
where $r_c$ is the compactification radius and $A(\phi)$ is termed as warp factor. For simplicity we assume that
the bulk scalar field depends only on the extra dimensional coordinate ($\phi$).
Thus the 5-dimensional Einstein's and scalar field equations
for this metric can be written as,
\begin{eqnarray}
\frac{4}{r_c^2} A'^2(\phi) - \frac{1}{r_c^2}A''(\phi)&=&-(2\kappa^2/3)V(\Phi), \nonumber\\
&-&(\kappa^2/3)\sum \lambda_{i}(\Phi)\delta(\phi-\phi_{i}),
\label{equation1}
\end{eqnarray}
\begin{equation}
\frac{1}{r_c^2}A'^2(\phi)= \frac{\kappa^2}{12r_c^2}\Phi'^2 -(\kappa^2/6)V(\Phi)
\label{equation2}
\end{equation}
\begin{eqnarray}
\frac{1}{r_c^2}\Phi''(\phi)=\frac{4}{r_c^2}A'\Phi' + \frac{\partial V}{\partial \Phi}
+ \sum \frac{\partial \lambda_{i}}{\partial \Phi}\delta(\phi-\phi_{i}),
\label{equation3}
\end{eqnarray}
where $M^3=(1/2\kappa^2)$. Here index $i$ is used to designate the two branes and prime denotes the derivative with
respect to $\phi$. From the above equations,
the boundary conditions of $A(\phi)$ and $\Phi(\phi)$ are obtained as,
\begin{equation}
\frac{1}{r_c}[A'(\phi)]_{i} = (\kappa^2/3)\lambda_{i}(\Phi_{i})
\label{bc1}
\end{equation}
and
\begin{equation}
\frac{1}{r_c}[\Phi'(\phi)]_{i}=\partial_{\Phi}\lambda_{i}(\Phi_{i}).
\label{bc2}
\end{equation}
Square brackets in the above two equations represent the jump of the
corresponding variables on the branes. In order to get an analytic
solution, let us consider the form of the scalar field potential
as \cite{csaki},
\begin{equation}
V(\Phi) = (1/2)\Phi^2(u^2+4uk) - (\kappa^2/6)u^2\Phi^4
\label{scalar potential}
\end{equation}
where $k=\sqrt{-\kappa^2\Lambda/6}$. The potential contains quadratic
as well as quartic self interaction of the scalar field. Moreover it may be
noticed that the mass and quartic
coupling of the field $\Phi(\phi)$ are connected by a common free parameter $u$.
Using this form of the potential, one obtains a
solution of $A(\phi)$ and $\Phi(\phi)$ as follows,
\begin{eqnarray}
A(\phi) = kr_c|\phi| + (\kappa^2/12)\Phi_{P}^2\exp{(-2ur_c|\phi|)}
\label{solution of warp factor}
\end{eqnarray}
and
\begin{eqnarray}
\Phi(\phi) = \Phi_{P}\exp{(-ur_c|\phi|)},
\label{solution of scalar field}
\end{eqnarray}
where $\Phi_{P}$ is taken as the value of the scalar field on the Planck brane.
Using these solutions, $\lambda_{hid}$ and $\lambda_{vis}$ can be obtained from the
boundary conditions (eqn.(\ref{bc1}) and eqn.(\ref{bc2})) as,
\begin{eqnarray}
\lambda_{hid} = 6k/\kappa^2 - u\Phi_{P}^2
\label{planck brane tension}
\end{eqnarray}
and
\begin{eqnarray}
\lambda_{vis} = -6k/\kappa^2 + u\Phi_{P}^2\exp{(-2u\pi r_c)}.
\label{planck brane tension}
\end{eqnarray}
In order to introduce the radion field, we consider a fluctuation of the inter-brane separation
around the stable configuration $r_c$ \cite{GW_radion}. This fluctuation can be treated as a field $T(x)$ (known as radion field)
and for simplicity this new field is assumed to be the function of brane coordinates only. Then the metric takes
the following form \cite{GW_radion}:
\begin{eqnarray}
ds^2 = \exp{[-2A(x,\phi)]}g_{\mu\nu}(x)dx^{\mu}dx^{\nu} + T^2(x)d\phi^2,
\label{metric ansatz 2}
\end{eqnarray}
where $g_{\mu\nu}(x)$ is the induced on-brane metric and $A(x,\phi)$ has the following form,
\begin{eqnarray}
A(x,\phi) = k\phi T(x) + \frac{\kappa^2\Phi_{P}^2}{12}\exp{[-2u\phi T(x)]}.
\label{warp factor 2}
\end{eqnarray}
Consequently $\Phi(x,\phi)$ can be obtained from eqn.(\ref{solution of scalar field}) by replacing
$r_c$ by $T(x)$ i.e.
\begin{eqnarray}
\Phi(x,\phi) = \Phi_{P}\exp{(-uT(x)\phi)}
\label{solution of scalar field 2}
\end{eqnarray}
Plugging back the solutions presented in eqns. (\ref{warp factor 2}), (\ref{solution of scalar field 2})
into original five dimensional action (in eqn. (\ref{action})) and integrating over $\phi$ yields
the four dimensional effective action as follows
\begin{eqnarray}
S_{eff} = \int d^4x \sqrt{-g} \bigg[\frac{M^3}{k}R_{(4)} - \frac{1}{2}g^{\mu\nu}\partial_{\mu}\Psi\partial_{\nu}\Psi
- U_{rad}(\Psi)\bigg]
\label{effective action}
\end{eqnarray}
where $R_{(4)}$ is the Ricci scalar formed by $g_{\mu\nu}(x)$. Moreover,
$\Psi(x) = \sqrt{\frac{24M^3}{k}} e^{-A(x,\pi)}$
(with $A(x,\pi)$ given in eqn.(\ref{warp factor 2})) is the
canonical radion field and $U_{rad}(\Psi)$ is the radion potential with the following form \cite{tp2}
\begin{eqnarray}
U_{rad}(\Psi) = u^2\Phi_P^2 \bigg[\frac{1}{u}\bigg(1 - e^{(2u+4k)\pi T(\Psi)}\bigg)
- \frac{\kappa^2\Phi_P^2}{8(u+k)}\bigg(1 - e^{(4u+4k)\pi T(\Psi)}\bigg)\bigg],
\label{radion potential}
\end{eqnarray}
where $\kappa\Phi_P$ ($=\frac{\Phi_P}{M^{3/2}}$) is taken to be less than unity in order to ensure
to validity of the classical solution. Moreover, $T(\Psi)$ is given by the expression:
$\Psi(x) = \sqrt{\frac{24M^3}{k}} \exp{\big[-k\pi T(x) - \frac{\kappa^2\Phi_P^2}{12}e^{-2u\pi T(x)}\big]}$, as defined
earlier. Using this relation between $T(x)$ and $\Psi(x)$, we obtain the minimum of the radion potential $U_{rad}(\Psi)$ at,
\begin{eqnarray}
<\Psi> = \sqrt{\frac{24M^3}{k}} \bigg(\frac{2\sqrt{1+\frac{2k}{u}}}{\kappa\Phi_P}\bigg)^{k/u}
\exp{\bigg[-\frac{1}{3}\bigg(1+\frac{2k}{u}}\bigg)\bigg],
\nonumber
\end{eqnarray}
which immediately leads to the stabilized value of the modulus as \cite{tp2},
\begin{eqnarray}
k\pi <T(x)>&=&k\pi r_c\nonumber\\
&=&\frac{k}{u}\ln{\{\frac{\kappa\Phi_{P}}{2\sqrt{1+\frac{2k}{u}}}\}}.
\label{stabilized modulus}
\end{eqnarray}
However our entire analysis of finding the stabilization condition in eqn.(\ref{stabilized modulus})
is valid only for $u>0$. In this context one can easily
check that the radion potential produces no minima for $u<0$. Hence the parameter
$u$ is confined in positive regime in order to make a stable configuration for this braneworld scenario.\\
On projecting the bulk gravity on the brane, the extra degrees of freedom of $R^{(5)}$ (with respect
to $R_4$) appears as a scalar field (the radion field), symbolized by $\Psi(x)$ in the four dimensional
effective action (see eqn.(\ref{effective action})). For such on-brane theory, we are interested to explore
the effect of radion field on stellar structure. Thus we further consider an extra matter density
($L_{mat}$, confined on the brane) which acts as the ingredients of the star. Taking $L_{mat}$ into
account, the final form of 4D effective action is as follows,
\begin{eqnarray}
A_{eff}&=&S_{eff} + \int d^4x \sqrt{-g} L_{mat}\nonumber\\
&=&\int d^4x \sqrt{-g} \bigg[\frac{M^3}{k}R_{(4)} - \frac{1}{2}g^{\mu\nu}\partial_{\mu}\Psi\partial_{\nu}\Psi
- U_{rad}(\Psi) + L_{mat}\bigg].
\label{final effective action}
\end{eqnarray}
Therefore the radion field $\Psi$ (originated from extra dimension) and
$L_{mat}$ serve as energy-momentum tensor in the
four dimensional effective action. As the stellar interior is concerned, $L_{mat}$
is taken to be a perfect fluid with energy-momentum tensor given by
$T^{\mu}_{\nu}$(matter) $= diag\big(-\rho, p, p, p\big)$. Moreover the radion field
contributes as,
\begin{eqnarray}
T_{\mu\nu}(\Psi) = \partial_{\mu}\Psi\partial_{\nu}\Psi
- g_{\mu\nu}\bigg[\frac{1}{2}g^{\mu\nu}\partial_{\mu}\Psi\partial_{\nu}\Psi + U_{rad}(\Psi)\bigg].
\label{em for radion}
\end{eqnarray}
This completes our preliminary discussion and provides the necessary steps
that we will require in the next section while discussing the effect of radion field on a stellar structure
from the perspective of effective four dimensional theory (described by eqn. (\ref{final effective action})).\\
\section{Buchdahl's limit on stellar structure in presence of radion field}
As mentioned earlier, we want to investigate the possible modifications of Buchdahl's limit
due to the presence of radion field, thus the spacetime that fit our purpose is static and a spherically symmetric.
Therefore the metric ansatz for the interior star is taken as,
\begin{eqnarray}
ds_{-}^2&=&g_{\mu\nu} dx^{\mu}dx^{\nu}\nonumber\\
&=&-e^{\nu(r)}dt^2 + e^{\lambda(r)}dr^2 + r^2\big(d\theta^2 + \sin^2\theta d\phi^2\big),
\label{4D metric ansatz}
\end{eqnarray}
where $\nu(r)$ and $\lambda(r)$ are arbitrary functions of radial coordinate $r$ that we need
to determine from gravitational equations. Such spherically symmetric spacetime
ensures that $\Psi$ as well as $\rho$ and $p$ are the functions of $r$ only. Hence
eqn.(\ref{em for radion}) can be simplified and as a consequence the
energy-momentum tensors of matter field, radion field are given by
\begin{eqnarray}
T^{\mu}_{\nu}(matter) = diag\bigg(-\rho(r), p(r), p(r), p(r)\bigg),
\label{em for matter 2}
\end{eqnarray}
\begin{eqnarray}
T^{\mu}_{\nu}(\Psi) = diag\bigg(&-&f^2(r)-U_{rad}(r), f^2(r)-U_{rad}(r),\nonumber\\
&-&f^2(r)-U_{rad}(r), -f^2(r)-U_{rad}(r)\bigg)
\label{em for radion 2}
\end{eqnarray}
respectively, where $f^2(r)$ is defined as $f^2(r) = \frac{1}{2}e^{-\lambda(r)}\Psi'(r)^2$
and $U_{rad}(r) = U_{rad}(\Psi(r))$. Considering the interior of the stellar object
to be filled with perfect fluid having energy-momentum tensor presented in eqn.(\ref{em for matter 2}), the gravitational
equations (for the metric ansatz mentioned in eqn.(\ref{4D metric ansatz})) become,
\begin{eqnarray}
e^{-\lambda}\bigg[\frac{1}{r^2} - \frac{\lambda'}{r}\bigg] - \frac{1}{r^2} = 8\pi G_{4}\bigg[-\rho - f^2 - U_{rad}(r)\bigg],
\label{eqn1}
\end{eqnarray}
\begin{eqnarray}
e^{-\lambda}\bigg[\frac{\nu'}{r} + \frac{1}{r^2}\bigg] - \frac{1}{r^2} = 8\pi G_{4}\bigg[p + f^2 - U_{rad}(r)\bigg],
\label{eqn2}
\end{eqnarray}
where $'$ denotes the derivative with respect to $r$ and $\frac{1}{8\pi G_4} = \frac{M^3}{k} \sim 10^{38}$(GeV)$^2$.
There exists another Einstein's equation corresponding to angular coordinate, but that can be derived from the above two
and hence is not independent.\\
On the other hand, the conservation equation for the fluid and the field equation for radion field takes the
following simple form in the context of spherically symmetric spacetime,
\begin{eqnarray}
p' + \frac{\nu'}{2}\big(\rho + p\big) = 0
\label{eqn3}
\end{eqnarray}
and
\begin{eqnarray}
f' + \frac{\nu'}{2}f + \frac{2}{r}f = \frac{U_{rad}'}{2f}
\label{eqn4}
\end{eqnarray}
respectively. To derive the radion field equation, we use the definition of $f^2(r)$ as mentioned earlier.
At this stage, it deserves mentioning that there are four independent differential equations governing the
behaviour of the system considered in the present case, while there are five unknowns, $\lambda(r)$,
$\nu(r)$, $\rho(r)$, $p(r)$, $f(r)$. This problem is generally resolved by assuming an equation of state
for the perfect fluid. However this equation of state is not needed in the present context,
because here we are interested on the upper bound
of the mass-radius ratio of the star (in presence of modulus field), for which the complete interior solutions
are not necessary\\
Next we try to get some information about the functions $\lambda(r)$, $\nu(r)$ from the above equations of motion.
It is easy to show that one can integrate eqn. (\ref{eqn1}), resulting into the following form of $\lambda(r)$,
\begin{eqnarray}
e^{-\lambda(r)}&=&1 - \frac{2G_4}{r} \int_{0}^{r} 4\pi r^2 \bigg[\rho + f^2 + U_{rad}(r)\bigg]\nonumber\\
&=&1 - \frac{2G_4 m(r)}{r},
\label{information of lambda}
\end{eqnarray}
where $m(r) = \int_{0}^{r} 4\pi r^2 \bigg[\rho + f^2 + U_{rad}(r)\bigg]$, the mass of the star up to radius $r$.
It is clear from the expression of $m(r)$ that due to the presence of radion field, the total gravitational
mass is different from that the actual matter density present inside. The extra gravitating mass comes from
the radion field strength and its potential.\\
However, eqn. (\ref{eqn3}) can be rewritten as,
\begin{eqnarray}
\nu'(r) = \frac{-2p'}{\big(\rho + p\big)}
\label{information of nu}
\end{eqnarray}
which is the famous Tolman-Oppenheimer-Volkoff (TOV) equation.
The matter density ($\rho$) as well as the pressure ($p$) inside the star generally decreases with an increase of $r$.
This behaviour of $p$ along with eqn. (\ref{information of nu})
indicate that the function $\nu(r)$ increases with $r$ (i.e $\nu'(r)>0$). Moreover eqn.(\ref{eqn4}) can be
integrated and has a solution of $f(r)$ given by: $f^2(r) = \frac{1}{r^4}e^{-\nu} + U_{rad}(r)$. Considering $U_{rad}' < 0$ and
due to the fact $\nu'(r) > 0$, the above expression of $f(r)$ clearly implies that $f(r)$ decreases with the radial coordinate $r$.
Hence the effective density (inside the star) $\rho + f^2 + U_{rad}$ also decreases as the surface of the star is approached. We
will use this result later on.\\
With these ingredients, let us now derive Buchdahl's limit explicitly and for that let us start by
differentiating both sides of eqn.(\ref{eqn2}) (with respect to $r$) and get,
\begin{eqnarray}
e^{-\lambda}\bigg[\frac{\nu''}{r} - \frac{\nu'}{r^2} - \frac{2}{r^3} - \frac{\lambda'\nu'}{r}&-&\frac{\lambda'}{r^2}\bigg]
+ \frac{2}{r^3} = 8\pi G_4\nonumber\\
&\bigg[&p' + 2ff' - U_{rad}'(r)\bigg].
\label{diff 1}
\end{eqnarray}
Using the conservation equations for the fluid and the radion field from eqn.(\ref{eqn3}) and eqn.(\ref{eqn4}), one can
evaluate the right hand side of eqn.(\ref{diff 1}), leading to
\begin{eqnarray}
8\pi G_4 \bigg[p' + 2ff' - U_{rad}'(r)\bigg]&=&8\pi G_4 \bigg[-\frac{\nu'}{2}\big(\rho + p\big)
+ 2f\big(-\frac{\nu'}{2}f - \frac{2}{r}f\big)\bigg]\nonumber\\
&=&e^{-\lambda}\bigg[-\frac{\nu'^2}{2r} - \frac{\nu'\lambda'}{2r}\bigg] - \frac{32\pi G_4}{r}f^2
\label{rhs}
\end{eqnarray}
Substituting the above expression back to eqn.(\ref{diff 1}) and a little simplification leads to the following equation,
\begin{eqnarray}
2r\nu'' + r\nu'^2 - r\nu'\lambda' - 2\nu' = \frac{4}{r}\big(1 - e^{\lambda}\big) + 2\lambda' - 64\pi G_4 rf^2e^{\lambda}.
\label{simplification 1}
\end{eqnarray}
By using the following two identities namely
\begin{eqnarray}
\frac{d}{dr}\bigg[\frac{1}{r}e^{-\lambda/2}\frac{de^{\nu/2}}{dr}\bigg]&=&\frac{e^{(\nu-\lambda)/2}}{4r^2}
\big[2r\nu'' + r\nu'^2 - r\nu'\lambda' - 2\nu'\big]
\label{identity 1}\\
\frac{d}{dr}\bigg[\frac{1 - e^{-\lambda}}{2r^2}\bigg]&=&\frac{e^{-\lambda}}{2r^3} \big[r\lambda' - 2\big(e^{\lambda} - 1\big)\big]
\label{identity 2}
\end{eqnarray}
Eqn.(\ref{simplification 1}) can be rewritten as follows,
\begin{eqnarray}
e^{-(\nu+\lambda)/2} \frac{d}{dr}\bigg[\frac{1}{r}e^{-\lambda/2}\frac{de^{\nu/2}}{dr}\bigg] =
\frac{d}{dr}\bigg[\frac{1 - e^{-\lambda}}{2r^2}\bigg] - \frac{16\pi G_4}{r}f^2
\label{simplification 2}
\end{eqnarray}
At this point, we put forward some sensible requirements: the average energy density inside the star $\rho_{av} = m(r)/r^3$ should decrease
with the radial coordinate. Even though the average density involves contribution from the radion field, since
the radion field strength itself decreases outwards, the above requirement will trivially hold. Further, the form of
$e^{-\lambda}$ given in eqn.(\ref{information of lambda}) indicates that the first term of the right hand side of
eqn.(\ref{simplification 2}) is essentially $d\rho_{av}/dr$ i.e. $\frac{d}{dr}\bigg[\frac{1 - e^{-\lambda}}{2r^2}\bigg] = \frac{d\rho_{av}}{dr}$.
This expression along with the decreasing character of $\rho_{av}$ (with $r$) makes the right hand side of eqn.(\ref{simplification 2})
negative. As a consequence, we get the following inequality:
\begin{eqnarray}
\frac{d}{dr}\bigg[\frac{1}{r}e^{-\lambda/2}\frac{de^{\nu/2}}{dr}\bigg] < 0.
\label{inequality 1}
\end{eqnarray}
Integrating the above relation from some radius $r$ within the star to the surface of the star,
given by the radius $r_0$, we obtain,
\begin{eqnarray}
\frac{de^{\nu/2}}{dr} > \frac{\nu_0'}{2r_0} re^{\lambda/2} e^{(\nu_0 - \lambda_0)/2},
\label{inequality 2}
\end{eqnarray}
where the quantities with the subscript $'0'$ denotes that they are to be evaluated at the surface of the star
i.e. at $r = r_0$. Furthermore, to derive the above inequality, we consider that both the metric and its first derivative
are continuous at $r = r_0$. However, later, in section [4],
we explicitly find the continuity conditions of metric and its first derivative at the boundary of
the star by considering a generalized Vaidya metric for the exterior spacetime of the stellar object.\\
Integrating again the both sides of eqn.(\ref{inequality 2}) from the origin to the surface of the star, we obtain
\begin{eqnarray}
e^{\nu_0/2} - e^{\nu/2}(r=0)&>&\frac{\nu_0'}{2r_0} e^{(\nu_0 - \lambda_0)/2}\int_0^{r_0} re^{\lambda/2} dr\nonumber\\
&=&\frac{\nu_0'}{2r_0} e^{(\nu_0 - \lambda_0)/2}\int_0^{r_0} dr \frac{r}{\sqrt{1 - \frac{2G_4 m(r)}{r}}}.
\label{inequality 3}
\end{eqnarray}
In the last line, we use the solution of $e^{\lambda}$ (see eqn.(\ref{information of lambda})).
As the average energy density decreases towards the boundary of the star, it immediately follows that
$m(r)/r^3 > M/r^3$, where $M = m(r_0)$, the total effective mass of the star. Thus the inequality
in eqn.(\ref{inequality 3}) holds more strongly if $\frac{m(r)}{r}$ is replaced by $\frac{M}{r_0^3}r^2$.
With this modification, we arrive at,
\begin{eqnarray}
e^{\nu_0/2} - e^{\nu/2}(r=0)&>&\frac{\nu_0'}{2r_0} e^{(\nu_0 - \lambda_0)/2}\int_0^{r_0} dr \frac{r}{\sqrt{1 - \frac{2G_4M}{r_0^3}r^2}}\nonumber\\
&=&\frac{\nu_0'}{2r_0} e^{(\nu_0 - \lambda_0)/2} \bigg(\frac{r_0^3}{2G_4M}\bigg) \bigg[1 - \sqrt{1- \frac{2G_4M}{r_0}}\bigg].
\label{inequality 4}
\end{eqnarray}
Both the pressure and the contribution of the radion field are positive and finite at the origin, it
follows that $e^{\nu/2}(r=0) > 0$. Applying this result into eqn.(\ref{inequality 4}), we immediately obtain the
following inequality,
\begin{eqnarray}
1 - \frac{\nu_0'}{2r_0} e^{-\lambda_0/2} \bigg(\frac{r_0^3}{2G_4M}\bigg) \bigg[1 - \sqrt{1- \frac{2G_4M}{r_0}}\bigg] > 0.
\label{inequality 5}
\end{eqnarray}
The factor $\nu_0'$ can be obtained in terms of $\lambda_0$ by considering eqn.(\ref{eqn2}) at the surface
of the star ($r=r_0$, where the pressure is zero i.e. $p(r_0)=0$) as,
\begin{eqnarray}
\frac{\nu_0'}{2r_0} = -\frac{1}{2r_0^2} + \frac{1}{2}e^{\lambda_0} \bigg[\frac{1}{r_0^2} + 8\pi G_4\big(f_0^2 - U_{rad}^0\big)\bigg].
\label{intermediate}
\end{eqnarray}
Recall $f_0^2 = \frac{1}{2}e^{-\lambda_0}\Psi'(r_0)^2$ and $U_{rad}^0$ is the radion potential at $r_0$. Plugging the
above expression of $\nu_0'$ into eqn.(\ref{inequality 5}) and further using the solution of $e^{\lambda}$, we finally lands up with the
inequality as follows:
\begin{eqnarray}
\sqrt{1 - \frac{2G_4M}{r_0}}&-&\frac{r_0^3}{2G_4M}\bigg(1 - \sqrt{1 - \frac{2G_4M}{r_0}}\bigg)\nonumber\\
&\bigg[&\frac{1}{2r_0^2}\bigg(\frac{2G_4M}{r_0}\bigg) + 4\pi G_4\big(f_0^2 - U_{rad}^0\big)\bigg] > 0.
\label{inequality 6}
\end{eqnarray}
Simplification of eqn.(\ref{inequality 6}) yields a quadratic expression for $\frac{2G_4M}{r_0}$, one root of which
corresponds to a negative value and hence can be safely ignored, while the other root provides the necessary limit
on mass-radius ratio (i.e. $\frac{M}{r_0}$) of the star, as follows,
\begin{eqnarray}
\frac{2G_4M}{r_0} < \frac{b}{2a} \bigg[1 + \sqrt{1 + \frac{4ac}{b^2}}\bigg]
\label{final inequality}
\end{eqnarray}
where $a$, $b$, $c$ have the following expressions:
\begin{eqnarray}
a = \frac{9}{4} - 6\pi G_4r_0^2\Psi_0'^2 + 4\pi^2G_4^2r_0^4\Psi_0'^4,
\label{a}
\end{eqnarray}
\begin{eqnarray}
b = 2 - 10\pi G_4r_0^2\Psi_0'^2 + 8\pi^2G_4^2r_0^4\Psi_0'^4 + 12\pi G_4r_0^2U_{rad}^0 - 16\pi^2G_4^2r_0^4\Psi_0'^2U_{rad}^0,
\label{b}
\end{eqnarray}
\begin{eqnarray}
c&=&4\pi G_4r_0^2\Psi_0'^2 - 4\pi^2G_4^2r_0^4\Psi_0'^4 - 8\pi G_4r_0^2U_{rad}^0 - 16\pi^2G_4^2r_0^4(U_{rad}^0)^2\nonumber\\
&+&16\pi^2G_4^2r_0^4\Psi_0'^2U_{rad}^0.
\label{c}
\end{eqnarray}
Eqns.(\ref{a}) to (\ref{c}) indicate that in the absence of the radion field, $a$, $b$, $c$ take the value as
$\frac{9}{4}$, $2$, $0$ respectively, for which one immediately recovers the usual Buchdahl's limit in General Relativity,
i.e $\frac{2G_4M}{r_0} < \frac{8}{9}$.\\
However, in the presence of the radion field ($\Psi$), the upper limit on mass-radius ratio of a stable stellar object
gets modified compared to general relativity and obviously depends on the strength of the radion field
(and its potential) on the surface of the star. From eqn.(\ref{final inequality}) (along with the expressions of $a$, $b$, $c$), we obtain
figure (\ref{plot stellar limit}) demonstrating the variation of $\frac{2G_4M}{r_0}$ with $f_0$ and $U_{rad}^0$.\\
\begin{figure}[!h]
\begin{center}
\centering
\includegraphics[width=4.2in,height=4.2in]{buchdahl_limit1.eps}
\includegraphics[width=0.3in,height=2.0in]{buchdahl_limit2.eps}
\caption{$\frac{2G_4M}{r_0}$ vs $G_4r_0^2\Psi_0'^2 (=x)$ and $G_4r_0^2U_{rad}^0 (=y)$ (left figure), corresponding contour plot (right figure)}
\label{plot stellar limit}
\end{center}
\end{figure}
Figure (\ref{plot stellar limit}) (with the contour plot) clearly reveals that in the presence of higher
dimensional modulus field, the upper bound on $\frac{2G_4M}{r_0}$ can be larger than $8/9$
and reach up to unity. Therefore extra mass
can be packed into the stellar structure in comparison to the Einstein
gravity. This provides an interesting testbed for existence of the modulus field (or radion field)
which carries the footprint of compactified extra dimension on our visible universe.
Thus, if a compact stellar object (say a neutron star)
is observed whose $2G_4M/r_0$ is larger than $8/9$ $\big($i.e $\frac{2G_4M}{r_0}$ lies between [8/9,1]$\big)$,
then it can possibly signal towards the existence of such higher dimension.\\
\section{Matching of the interior spacetime with an exterior geometry}
To complete the model, the interior spacetime geometry of the spherical star needs to be matched to
an exterior geometry. For the required matching, the Israel conditions are used,
where the metric coefficients and extrinsic curvatures (first and second fundamental forms
respectively) are matched at the boundary of the sphere.\\
However in the presence of a scalar field (which is the radion field in the present context), the interior
spacetime can not be smoothly matched to a vacuum exterior (i.e. the Schwarzschild one). If the exterior is a vacuum, the
scalar field has to behave as a delta function at the boundary resulting in a square of delta function for the energy density.\\
To avoid this problem, in the present case, we match
the interior spacetime with a generalized Vaidya exterior spacetime at the boundary hypersurface $\Sigma$
given by $r = r_0$. The metric inside and outside of $\Sigma$ are given by,
\begin{eqnarray}
ds_{-}^2 = -e^{\nu(r)}dt^2 + e^{\lambda(r)}dr^2 + r^2\big(d\theta^2 + \sin^2\theta d\phi^2\big)
\label{inside metric}
\end{eqnarray}
and
\begin{eqnarray}
ds_{+}^2 = -\bigg(1 - \frac{2G_4M_+(r_v)}{r_v}\bigg)dv^2 + 2dvdr_v + r_v^2\big(d\theta^2 + \sin^2{\theta}d\varphi^2\big)
\label{outside metric}
\end{eqnarray}
respectively, where $r_v,v,\theta$ and $\varphi$ are the exterior coordinates and $M_+(r_v)$ (the suffix $'+'$ stands for exterior)
is exterior mass function, which is independent of $v$ due to the reason that the spacetime is static.
The same hypersurface $\Sigma$ can alternatively be defined by the exterior
coordinates as $r_v = R(t)$ and $v = T(t)$. Then the metrics on $\Sigma$ from inside and
outside coordinates turn out to be,
\begin{eqnarray}
ds_{-,\Sigma}^2 = -e^{\nu_0}dt^2 + r_0^2 d\Omega^2
\nonumber
\end{eqnarray}
and
\begin{eqnarray}
ds_{+,\Sigma}^2 = -\bigg[\bigg(1 - \frac{2G_4M_{+,\Sigma}}{R(t)}\bigg)\dot{T}^2 - 2\dot{T}\dot{R}\bigg]dt^2 + R(t)^2 d\Omega^2,
\nonumber
\end{eqnarray}
where $M_{+,\Sigma}$ is the exterior mass function on $\Sigma$, $d\Omega^2$ denotes the line
element on a unit two sphere and dot represents $\frac{d}{dt}$. Matching the first fundamental
form on $\Sigma$ (i.e. $ds_{-,\Sigma}^2 = ds_{+,\Sigma}^2$) yields the following two conditions :
\begin{eqnarray}
\frac{dT(t)}{dt} = \frac{e^{\nu_0/2}}{\sqrt{1 - \frac{2G_4M_{+,\Sigma}}{r_0}}}
\label{con 1}
\end{eqnarray}
and
\begin{eqnarray}
R(t) = r_0.
\label{con 2}
\end{eqnarray}
In order to match the second fundamental form, we calculate the normal of the hypersurface $\Sigma$
from inside ($\vec{n}_{-} = n_{-}^t$, $n_{-}^r$, $n_{-}^{\theta}$, $n_{-}^{\varphi}$) and outside
($\vec{n}_{+} = n_{+}^v$, $n_{+}^{r_v}$, $n_{+}^{\theta}$, $n_{+}^{\varphi}$) coordinates as follows,
\begin{eqnarray}
n_{-}^t = 0~~,~~~~~~~~n_{-}^r = e^{-\lambda_0/2}~~,~~~~~~~n_{-}^{\theta} = n_{-}^{\varphi} = 0\label{inside normal}
\end{eqnarray}
and
\begin{eqnarray}
n_{+}^v = \frac{1}{\sqrt{1 - \frac{2G_4M_{+,\Sigma}}{r_0}}}~~,\nonumber
\end{eqnarray}
\begin{eqnarray}
n_{+}^{r_v} = \sqrt{1 - \frac{2G_4M_{+,\Sigma}}{r_0}}~~,\nonumber
\end{eqnarray}
\begin{eqnarray}
n_{+}^{\theta} = n_{+}^{\varphi} = 0.
\label{outside normal}
\end{eqnarray}
To derive the normal vectors, we use eqn.(\ref{con 2}).
The above expressions of $\vec{n}_{-}$ and $\vec{n}_{+}$ leads to the extrinsic
curvature of $\Sigma$ from interior and exterior coordinates respectively, and are given by,
\begin{eqnarray}
K_{tt}^- = -\frac{\nu_0'}{2}e^{\nu_0}e^{-\lambda_0/2}~~,\nonumber
\end{eqnarray}
\begin{eqnarray}
K_{\theta\theta}^- = r_0 e^{-\lambda_0/2}~~,~~~~~~~K_{\varphi\varphi}^- = r_0 e^{-\lambda_0/2}\sin^2{\theta},
\label{inside extrinsic}
\end{eqnarray}
from interior metric, and
\begin{eqnarray}
K_{tt}^+ = -G_4 \frac{e^{\nu_0}}{\sqrt{1 - 2\frac{G_4M_{+,\Sigma}}{r_0}}} \bigg[\frac{M_{+,\Sigma}}{r_0^2}
- \frac{1}{r_0}\frac{\partial M_{+}(r_v)}{\partial r_v}\bigg{|}_{\Sigma}\bigg]~~,
\nonumber
\end{eqnarray}
\begin{eqnarray}
K_{\theta\theta}^+ = r_0 \sqrt{1 - \frac{2G_4M_{+,\Sigma}}{r_0}}~~,
\nonumber
\end{eqnarray}
\begin{eqnarray}
K_{\varphi\varphi}^+ = r_0 \sqrt{1 - \frac{2G_4M_{+,\Sigma}}{r_0}} \sin{\theta}^2
\label{outside extrinsic}
\end{eqnarray}
from exterior metric.\\
The equality of the extrinsic curvatures of $\Sigma$ from both sides is therefore equivalent to the following
two conditions :
\begin{eqnarray}
e^{-\lambda_0)/2} = \sqrt{1 - \frac{2G_4M_{+,\Sigma}}{r_0}}
\label{con 3}
\end{eqnarray}
and
\begin{eqnarray}
\frac{\partial M_{+}(r_v)}{\partial r_v}\bigg{|}_{\Sigma} = \frac{M_{+,\Sigma}}{r_0}
- \frac{r_0}{2G_4}\nu_0' \bigg[1 - \frac{2G_4M_{+,\Sigma}}{r_0}\bigg].
\label{con 4}
\end{eqnarray}
The total mass of the spherical star is given by: $M = \int_0^{r_0} 4\pi r^2 \big(\rho + f^2 + U_{rad}(r)\big)$.
Eqn.(\ref{con 3}) along with the solution of $\lambda$ (see eqn.(\ref{information of lambda})) relates the exterior
mass function on $\Sigma$ (i.e. $M_{+,\Sigma}$) with the total mass of the stellar object as,
\begin{eqnarray}
M_{+,\Sigma}&=&M\nonumber\\
&=&\int_0^{r_0} 4\pi r^2 \big(\rho + f^2 + U_{rad}(r)\big).
\label{con 5}
\end{eqnarray}
Eqns. (\ref{con 1}), (\ref{con 2}), (\ref{con 4}), (\ref{con 5}) completely
specify the matching at the boundary of the star with an exterior generalized Vaidya geometry.\\
\section{Conclusion}
We consider a five dimensional AdS compactified warped geometry model with two 3-branes embedded within the spacetime.
For the purpose of modulus stabilization, a massive scalar field is invoked in the bulk and its backreaction
on spacetime geometry is taken into account. In such a scenario, our universe is identified with a 3-brane and emerges
as a four dimensional effective theory. On projecting the bulk gravity on the brane, the extra degrees of freedom of
$R^{(5)}$ appears as a scalar field (known as radion field) in the 4D effective on-brane theory.
From the perspective of such on-brane theory, we explore the effect of radion field on the limit
of mass-radius ratio ($M/r_0$) for a stable stellar structure. \\
We match the interior spacetime of the star with a suitable exterior geometry on the boundary ($\Sigma$). For this
matching, the Israel junction conditions are used where the metric coefficients and extrinsic curvatures are matched on $\Sigma$.
At this stage, it deserves mention that in presence of a scalar field (which is the radion field in the present context),
the matching of interior spacetime with exterior Schwarzschild geometry leads to some inconsistency. For instance, since
Schwarzschild has zero scalar field, such a matching would lead to a discontinuity
in the scalar field, which means a delta function in the gradient of the scalar field.
As a consequence, there will appear square of a delta function in the stress-energy, which is definitely an inconsistency.
To avoid such problems, here we consider the exterior geometry as a generalized Vaidya spacetime. With this consideration,
we determine the matching conditions given in eqns. (\ref{con 1}), (\ref{con 2}), (\ref{con 4}), (\ref{con 5}). \\
The main conclusion of the present investigation is the following. Due to the presence of radion field, the upper limit on mass-radius ratio (generally known as Buchdahl's limit) of
a compact stellar object gets modified in comparison to general relativity and obviously depends on the
strength of the radion field (and its potential) on the surface of the star. The variation of Buchdahl's limit
with the radion field strength is shown in figure (\ref{plot stellar limit}), which
clearly demonstrates that in the presence of the higher dimensional modulus field, the upper bound of $\frac{2G_4M}{r_0}$ can go
beyond the value $8/9$ and reach up to unity; while the general relativity prediction is given by: $\frac{2G_4M}{r_0} < \frac{8}{9}$.
Therefore extra mass can be packed into the stellar structure in comparison to the Einstein gravity.
This provides an interesting testbed for existence of modulus field (or radion field)
which carries the footprint of compactified extra dimension on our visible universe. Hence if it is possible to detect
a compact object with mass-radius ratio larger than the general relativity prediction, then one can infer about
the possible presence of such higher dimension.\\
\section*{Acknowledgements}
The author would like to thank Narayan Banerjee and Soumitra SenGupta for illuminating discussions.
|
\section{Introduction}
Generally, in coastal and ocean waters, current velocity profiles are established by bottom friction and wind stress at the sea surface, and consequently are vertically varying. Ebb and flood currents due to the tide may have an important effect on water wave properties. In any region where the wind blows, the generated current affects the behavior of the waves. The present work focuses on the nonlinear evolution of two-dimensional gravity waves propagating in shallow water on a shear current which varies linearly with depth. Consequently, the waves are travelling on a flow of constant vorticity. Considering constant vorticity is an approximation which allows the analytical derivation of a unique partial differential equation governing the nonlinear evolution of hyperbolic or dispersive waves in shallow water.
\newline
Within the framework of long waves, there are very few papers devoted to unsteady nonlinear evolution of water waves propagating on an underlying vertically sheared current. \citet{Freeman} derived a KdV equation governing the time evolution of gravity waves on shear flows of arbitrary vorticity distribution. Using an asymptotic expansion method \citet{Choi} derived a Green-Naghdi system for long gravity waves in uniform shear flows (constant vorticity) and for weakly nonlinear waves he deduced from this system a Boussinesq-type equation and a KdV equation. The derivation of a Boussinesq-type equation and a Camassa-Holm equation with constant vorticity was carried out by \citet{Johnson}. Very recently, \citet{Richard} derived a dispersive shallow water model which is a generalisation of the classical Green-Naghdi model to the case of shear flows. \citet{Castro} investigated rigorously a Green-Naghdi type system including a general vorticity. At the same time and separately \citet{Kharif} and \citet{Hur} have derived shallow water wave equations with constant vorticity. \citet{Kharif} have considered the effect of a vertically sheared current on rogue wave properties whereas \citet{Hur} investigated the effect of vorticity on modulational instability and onset of breaking.
\newline
In the coastal zone where the vorticity is an important ingredient, water wave dynamics is governed by the Euler equations with boundary conditions at the free surface which is nonlinear and unknown {\em a priori}. Numerical integration of this system of equations is not a trivial task and so more simple models have been derived in the past to describe and investigate the dynamics of water waves phenomena in shallow water such as undular bores and nonlinear long wave propagation. For a review one can refer to the paper by \citet{Lannes}.
\newline
Following \citet{Whitham}, we propose a new model derived from the Euler equation for water waves propagating on a vertically sheared current of constant vorticity in shallow water. This new approach is easier to handle numerically. The heuristic introduction of dispersion allows the study of strongly nonlinear two-dimensional long gravity waves in the presence of vorticity and as well that of undular bores. From this model we derive, within the framework of weakly nonlinear waves satisfying the exact linear dispersion a Whitham equation and for weakly nonlinear and weakly dispersive waves the KdV equation previously obtained by \citet{Freeman} and \citet{Choi}. These different equations are then used to investigate the effect of constant vorticity on breaking time of dispersive waves and hyperbolic waves as well.
\section{Derivation of the new approach: The generalised Whitham equation}
We consider two-dimensional gravity water waves propagating at the free surface of a vertically sheared current of uniform intensity $\Omega$ which is the opposite of the vorticity. The wave train moves along the $x-\mathrm{axis}$ and the $z-\mathrm{axis}$ is oriented upward opposite to the gravity. The origin $z=0$ is the undisturbed free surface and $z=-h(x)$ is the rigid bottom.
\vspace{0.2cm}
\newline
The continuity equation is
\begin{equation}
u_x+w_z=0
\label{continuity}
\end{equation}
where $u$ and $w$ are the longitudinal and vertical components of the wave induced velocity,
respectively. The underlying current is $U=U_0+\Omega z$ where $U_0$ is the constant surface velocity.
\vspace{0.2cm}
\newline
Integration of equation (\ref{continuity}) gives
\begin{equation}
w(z=\eta)-w(z=-h)=-\int_{-h(x)}^{\eta(x,t)} u_x dz
\label{interal-continuite}
\end{equation}
where $\eta$ is the surface elevation.
\vspace{0.2cm}
\newline
Note that
\begin{equation}
\int_{-h(x)}^{\eta(x,t)} u_x dz=\frac{\partial}{\partial x}\int_{-h(x)}^{\eta(x,t)} u dz-u(z=\eta)\eta_x-u(z=-h)h_x
\label{interal-continuite-2}
\end{equation}
\vspace{0.2cm}
\newline
The kinematic boundary condition at the free surface is
\begin{equation}
\eta_t+(u+U_0+\Omega \eta)\eta_x-w=0 \qquad \mathrm{on} \qquad z=\eta(x,t)
\label{kinematic}
\end{equation}
\vspace{0.2cm}
\newline
The bottom boundary condition writes
\begin{equation}
(u+U_0-\Omega h)h_x+w=0 \qquad \mathrm{on} \qquad z=-h(x)
\end{equation}
\vspace{0.2cm}
\newline
We assume $h$ constant, then
\begin{equation}
w=0 \qquad \qquad \mathrm{on} \qquad \qquad z=-h
\end{equation}
\vspace{0.3cm}
\newline
From equation (\ref{kinematic}) it follows that
\begin{eqnarray}
w(z=\eta)=\eta_t+(u+U_0+\Omega \eta)_{z=\eta}\eta_x \\
w(z=\eta)=\eta_t+[u(z=\eta)+U_0+\Omega \eta]\eta_x
\end{eqnarray}
Equation (\ref{interal-continuite}) becomes
\begin{equation}
-\int_{-h}^{\eta(x,t)} u_x dz = \eta_t+[u(z=\eta)+U_0+\Omega \eta]\eta_x
\end{equation}
Using equation (\ref{interal-continuite-2}) with $h_x=0$ we obtain
\begin{equation}
\frac{\partial}{\partial x}\int_{-h}^{\eta(x,t)}udz + \eta_t + (U_0+\Omega \eta) \eta_x=0
\end{equation}
We assume $u$ does not depend on $z$, then
\begin{equation}
\eta_t + \frac{\partial}{\partial x}[u(\eta+h)+\frac{\Omega}{2}\eta^2 + U_0 \eta]=0
\label{mass_conservation}
\end{equation}
Equation (\ref{mass_conservation}) corresponds to mass conservation in shallow water in the presence of constant vorticity.
\vspace{0.2cm}
\newline
Under the assumption of hydrostatic pressure, the Euler equation in $x$-direction is
\begin{equation}
u_t + (u+U_0+\Omega z)u_x + \Omega w +g \eta_x=0
\end{equation}
where $g$ is the gravity.
\vspace{0.2cm}
\newline
Using the continuity equation and boundary conditions that $w$ satisfies on the bottom and at the free surface, we obtain
\begin{equation}
w=-(z+h)u_x
\end{equation}
It follows that the Euler equation becomes
\begin{equation}
u_t + (u+U_0-\Omega h)u_x + g \eta_x=0
\label{x_Euler}
\end{equation}
The dynamics of non dispersive shallow water waves on a vertically sheared current of constant vorticity is governed by equations (\ref{mass_conservation}) and (\ref{x_Euler}).
\vspace{0.2cm}
\newline
The pair of equations (\ref{mass_conservation}) and (\ref{x_Euler}) admits the following Riemann invariants
\begin{equation}
u+\frac{\Omega H}{2} \pm \bigg\lbrace\sqrt{gH+\Omega^2H^2/4}+\frac{g}{\Omega}\ln\left[1+\frac{\Omega}{2g}(\Omega H +2 \sqrt{gH+\Omega^2H^2/4})\right]\bigg\rbrace=\rm{constant} \nonumber
\end{equation}
on characteristic lines
\begin{equation}
\frac{dx}{dt}=u+U_0+\frac{1}{2}\Omega (\eta-h)\pm \sqrt{gH+\frac{\Omega^2 H^2}{4}}
\label{characteristic}
\end{equation}
where $H=\eta + h$.
\vspace{0.2cm}
\newline
The constant is determined for $u=0$ and $\eta = 0$ or $H=h$.
\vspace{0.2cm}
\newline
Finally
\begin{eqnarray}
u+\frac{\Omega \eta}{2} &+& \sqrt{gH+\Omega^2H^2/4}-\sqrt{gh+\Omega^2h^2/4} \nonumber \\ &+& \frac{g}{\Omega}
\ln\left[\frac{1+\frac{\Omega}{2g}(\Omega H
+2 \sqrt{gH+\Omega^2H^2/4})}{1+\frac{\Omega}{2g}(\Omega h
+2 \sqrt{gh+\Omega^2h^2/4})}\right]=0
\end{eqnarray}
\begin{eqnarray}
u+\frac{\Omega \eta}{2} &-& \sqrt{gH+\Omega^2H^2/4}+\sqrt{gh+\Omega^2h^2/4} \nonumber \\ &-& \frac{g}{\Omega}
\ln\left[\frac{1+\frac{\Omega}{2g}(\Omega H
+2 \sqrt{gH+\Omega^2H^2/4})}{1+\frac{\Omega}{2g}(\Omega h
+2 \sqrt{gh+\Omega^2h^2/4})}\right]=0
\end{eqnarray}
Let us consider a wave moving rightwards
\begin{eqnarray}
u= -\frac{\Omega \eta}{2} &+& \sqrt{gH+\Omega^2H^2/4}-\sqrt{gh+\Omega^2h^2/4} \nonumber \\ &+& \frac{g}{\Omega}
\ln\left[\frac{1+\frac{\Omega}{2g}(\Omega H
+2 \sqrt{gH+\Omega^2H^2/4})}{1+\frac{\Omega}{2g}(\Omega h
+2 \sqrt{gh+\Omega^2h^2/4})}\right]
\end{eqnarray}
Substituting this expression into equation (\ref{mass_conservation}) gives
\begin{equation}
\eta_t + \bigg\lbrace U_0-\frac{\Omega h}{2} + 2\sqrt{g(\eta+h)+\Omega^2(\eta+h)^2/4}-\sqrt{gh+\Omega^2h^2/4}
\nonumber
\end{equation}
\begin{equation}
+ \frac{g}{\Omega} \ln\left[1+ \frac{\Omega}{2g}\frac{\Omega \eta +2(\sqrt{g(\eta+h)+\Omega^2(\eta+h)^2/4}-\sqrt{gh+\Omega^2h^2/4})}{1+\frac{\Omega}{g}(\frac{\Omega h}{2}+\sqrt{gh+\Omega^2h^2/4})} \right] \bigg\rbrace\eta_x =0
\label{nonlin-nondispersive}
\end{equation}
Equation (\ref{nonlin-nondispersive}) is fully nonlinear and describes the spatio-temporal evolution of hyperbolic water waves in shallow water in the presence of constant vorticity. This equation is equivalent to the system of equations (\ref{mass_conservation}) and (\ref{x_Euler}) for waves moving rightwards.
\vspace{0.3cm}
\newline
Following \citet{Whitham}, full linear dispersion is introduced heuristically
\begin{equation}
\eta_t + \bigg\lbrace U_0-\frac{\Omega h}{2} + 2\sqrt{g(\eta+h)+\Omega^2(\eta+h)^2/4}-\sqrt{gh+\Omega^2h^2/4}
\nonumber
\end{equation}
\begin{equation}
+ \frac{g}{\Omega} \ln\left[1+ \frac{\Omega}{2g}\frac{\Omega \eta +2(\sqrt{g(\eta+h)+\Omega^2(\eta+h)^2/4}-\sqrt{gh+\Omega^2h^2/4})}{1+\frac{\Omega}{g}(\frac{\Omega h}{2}+\sqrt{gh+\Omega^2h^2/4})} \right] \bigg\rbrace\eta_x + K*\eta_x=0
\label{nonlin-dispersive}
\end{equation}
where $K*\eta_x$ is a convolution product. The kernel $K$ is given as the inverse Fourier transform of the fully linear dispersion relation of gravity waves in finite depth in the presence of constant vorticity $\Omega$: $K=F^{-1}(c)$ with
\[c=U_0-\frac{\Omega \tanh (kh)}{2k} + \sqrt{\frac{g \tanh(kh)}{k} + \frac{\Omega^2 \tanh^2(kh)}{4 k^2}}\]
Equation (\ref{nonlin-dispersive}) governs the propagation of nonlinear long gravity waves in a fully linear dispersive medium. For $\Omega=0$ and $U_0=0$ (\ref{nonlin-nondispersive}) reduces to equation (13.97) of \citet{Whitham}.
\vspace{0.2cm}
\newline
For weakly nonlinear waves ($\eta/h\ll1$) equation (\ref{nonlin-dispersive}) becomes the Whitham equation with constant vorticity given by
\begin{equation}
\eta_t + \frac{3gh+h^2 \Omega^2}{h\sqrt{gh(4gh+h^2 \Omega^2)}} \eta \eta_x + K*\eta_x=0
\label{Whitham}
\end{equation}
Note that in the Whitham equation the exact linear dispersion is considered unlike the KdV equation. In the absence of vorticity, \citet{Ehrnstrom} proved rigorously that equation (\ref{Whitham}) admits small-amplitude periodic travelling wave solutions and computed numerically approximations of small- and finite-amplitude waves as well as those close to the highest. The existence of solitary waves has been proven by \citet{Ehrnstrom1} without considering vorticity effect.
\newline
The numerical method that allows the computation of periodic travelling wave solutions of equation (\ref{Whitham}) is included in Appendix A. In figure \ref{fig:bifurcation_vor} is plotted the maximum elevation of the solutions of equation (\ref{Whitham}) for branches corresponding to several values of the vorticity. For a constant maximum elevation the phase velocity decreases as $\Omega$ increases. In figure \ref{fig:profiles_Whitham_vor} are shown the profiles of the elevation $\eta$ for several values of the vorticity for a constant phase velocity and a constant maximum elevation, respectively. The profiles become larger as $\Omega$ is decreasing (or as the vorticity is increasing).
\begin{figure}
\center
\includegraphics[width=0.6\linewidth]{a_V_1_.eps}
\includegraphics[width=0.6\linewidth]{a_V_amp_015.eps}
\caption{Bifurcation diagrams for several values of the vorticity in the (maximum elevation, phase velocity) plane. Top: solid line ($\Omega=0$), dashed line ($\Omega=-0.08$) and -$\circ$- ($\Omega=0.08$). Bottom: solid line ($\Omega=0$), dashed line ($\Omega=-0.5$) and -$\circ$- ($\Omega=0.5$)}
\label{fig:bifurcation_vor}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.6\linewidth]{phi_x_Om_1_.eps}
\includegraphics[width=0.6\linewidth]{phi_x_Om_amp_015.eps}
\caption{Elevation profiles. Top: solid line ($\Omega=0$), dashed line ($\Omega=-0.08$) and -$\circ$- ($\Omega=0.08$) for $V=0.8$. Bottom: solid line ($\Omega=0$), dashed line ($\Omega=-0.5$) and -$\circ$- ($\Omega=0.5$) for the maximum of elevation equal to $0.15$}.
\label{fig:profiles_Whitham_vor}
\end{figure}
\vspace{0.2cm}
\newline
We call equation (\ref{nonlin-dispersive}) which generalises the Whitham equation (\ref{Whitham}) to fully nonlinear waves the generalised Whitham equation in the presence of vorticity.
\vspace{0.1cm}
\newline
For weakly nonlinear ($\eta/h\ll1$) and weakly dispersive ($kh\ll1$) water waves, equation (\ref{nonlin-dispersive}) reduces to the KdV equation with vorticity derived by \citet{Freeman} and \citet{Choi} who used multiple scale methods, different to the approach used herein. To set the KdV equation in dimensionless form, $h$ and $\sqrt{h/g}$ are chosen as reference length and reference time which corresponds to $h=1$ and $g=1$. The equation reads
\begin{equation}
\eta_t+c_0(\Omega) \eta_x+c_1(\Omega)\eta \eta_x+c_2(\Omega)\eta_{xxx}=0
\label{vor-KdV}
\end{equation}
with
\[c_0=U_0-\frac{\Omega}{2}+\sqrt{1+\Omega^2/4}, \quad\, \quad c_1=\frac{3+\Omega^2}{\sqrt{4+\Omega^2}}, \quad \quad c_2=\frac{2+\Omega^2-\Omega\sqrt{4+\Omega^2}}{6\sqrt{4+\Omega^2}}\]
Equation (\ref{vor-KdV}) is known to admit as solutions the solitary wave and cnoidal wave whose expressions are
\[\eta=a\, \mathrm{sech}^2 (\frac{x-ct}{\Delta_s})\]
with $c=c_0 + \frac{c_1}{3}a$ and $\Delta_s=\sqrt{\frac{12 c_2}{c_1 a}}$
\newline
and
\[\eta=\frac{a}{m}(1-m-\frac{E(m)}{K(m)}) + a\, \mathrm{cn}^2(\frac{(x-ct)}{\Delta_{cn}}|m) \]
with $\Delta_{cn} = \sqrt{12 m c_2/(c_1 a)}$ and $c=c_0 + c_1(2-m-E(m)/K(m))a/6$
\newline
Note that $\Delta_s=\Delta_{cn}(m=1)$. An increase of $\Delta_s$ and $\Delta_{cn}$ gives a wider wave profile. The profiles of the solitary wave and cnoidal wave in the presence of vorticity ($-\Omega$) for various values of $\Omega$ are plotted in figure \ref{fig:profiles:waves}. In both cases, for fixed height $a$ the width of the profiles is increasing with the vorticity. In figure \ref{fig:width_vor} is shown $\sqrt{a/12}\Delta_{cn}$ as a function of $\Omega$ for two values of the elliptic parameter $m=1$ and $m=0.5$. Solitary waves and cnoidal waves of height $a$ propagating on currents of positive vorticity ($\Omega<0$) are wider than solitary waves and cnoidal waves of height $a$ propagating on currents of negative vorticity ($\Omega>0$).
\begin{figure}
\includegraphics[width=0.5\linewidth]{plot_soliton_omega_new_1_.eps}
\includegraphics[width=0.5\linewidth]{vor_cnoidal_1_.eps}
\caption{Profiles of solitary waves and cnoidal waves for various values of the vorticity. Solid line ($\Omega=0$), dashed line ($\Omega=-1$) and dotted line ($\Omega=1$)}
\label{fig:profiles:waves}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.7\linewidth]{soliton_cnoidal_width.eps}
\caption{Width of the profiles of solitary waves (solid line for $m=1$) and cnoidal waves (dashed line for $m=0.5$) as a function of the vorticity}
\label{fig:width_vor}
\end{figure}
\vspace{0.2cm}
\newline
The dimensionless form of the Whitham equation (\ref{Whitham}) is
\begin{equation}
\eta_t+c_1(\Omega)\eta \eta_x+K*\eta_x=0
\end{equation}
with
\[c^*=U_0-\frac{\Omega \tanh k}{2k} + \sqrt{\tanh k(1+\Omega^2 \tanh k/(4k))/k}, \quad c_1=\frac{3+\Omega^2}{\sqrt{4+\Omega^2}}\]
Herein $K=F^{-1}(c^*)$
\section{Numerical integration}
The equations (\ref{nonlin-nondispersive}), (\ref{nonlin-dispersive}) and (\ref{vor-KdV}) are solved numerically in a periodic domain of length $2L$. The length $L$ is chosen $O(400\delta)$ where $\delta$ is a characteristic length scale of the initial condition. The number of grid points is $N_x=2^{14}$. Spatial derivatives are computed in the Fourier space and nonlinear terms in the physical space. The link between the two spaces is made by the Fast Fourier Transform.
For the time integration, a splitting technique is used. The equations (\ref{nonlin-nondispersive}), (\ref{nonlin-dispersive}) and (\ref{vor-KdV}) could be written as
\begin{equation}
\eta_t+L+N=0,
\end{equation}
where $L$ and $N$ are linear and nonlinear differential operators in $\eta$, respectively. Note that in general the operators $L$ and $N$ do not commute. If the initial condition is $\eta_0$, the exact solution of the previous equation is
\begin{equation}
\eta(t)=e^{-(L+N)t}\eta_0.
\end{equation}
This equation is discretized as follows. Let $t_n=n\Delta t$. We have
\begin{equation}
\eta(t_n)=e^{-(L+N)n\Delta t}\eta_0=(e^{-L\Delta t/2}e^{-N\Delta t}e^{-L\Delta t/2})^n\eta_0+O(\Delta t^2),
\end{equation}
and the scheme is globally second order in time. The operator $e^{-L\Delta t/2}$ is computed exactly in the Fourier space. However, the operator $e^{-N\Delta t}$ is approximated using a Runge-Kutta scheme of order 4. The time step is chosen as $\Delta t=0.005$. With this time step, and when the KdV equation is integrated, the mass $I_1$, momentum $I_2$ and energy $I_3$ invariants are conserved to the machine precision for $I_1$ and with a relative error $\epsilon(I_i)=(\tilde I_i(t)-I_i)/I_i=O(10^{-9})$ for $I_2$ and $I_3$ ($\tilde I_i$ being the computed value of $I_i$) as shown in figure \ref{fig:rel:error}.
\section{Validation of the numerical method}
The efficiency and accuracy of the numerical method has been checked against the nonlinear analytical solution of the St-Venant equations for the dam-break problem and the experiments of \citet{Favre}, in the absence of current and vorticity ($\Omega=0$ and $U_0=0$).
\vspace{0.2cm}
\newline
For $U_0=0$ and $\Omega=0$ equation (\ref{nonlin-nondispersive}) reduces to
\begin{equation}
H_t +(3\sqrt{gH}-2\sqrt{gh})H_x=0, \quad \mathrm{with} \quad H=\eta+h.
\label{dam-break}
\end{equation}
For $t>0$, the nonlinear analytical solution of equation (\ref{dam-break}) is
\begin{eqnarray}
H(x,t)=h, & u(x,t)=0; & \frac{x}{t}\ge\sqrt{gh} \nonumber \\
H(x,t)=\frac{h}{9}\left(2+\frac{x}{\sqrt{gh}\ t}\right)^2, & u(x,t)=-\frac{2}{3}\left(\sqrt{gh}-\frac{x}{t}\right); & -2\sqrt{gh}\le\frac{x}{t}\le\sqrt{gh} \nonumber \\
H(x,t)=0,& u(x,t)=0; & \frac{x}{t} \le -2\sqrt{gh}
\label{exact-solution}
\end{eqnarray}
At time $t=0$ the initial condition is $H(x,0)=h(1+\tanh(2x))/2$ and $u(x,0)=0$ everywhere. A numerical simulation of equation (\ref{dam-break}) has been carried out with $g=1$ and $h=1$. The numerical and analytical surface profiles at $t=0$ and after the dam has broken are plotted in figure \ref{fig:dam_break}.
\newline
Within the framework of the KdV equation in the presence of vorticity, we have also checked that solitary waves are propagated with the right velocity that depends on $\Omega$.
\begin{figure}
\includegraphics[width=0.3\linewidth]{eps_I1.eps}
\includegraphics[width=0.3\linewidth]{eps_I2.eps}
\includegraphics[width=0.3\linewidth]{eps_I3.eps}
\caption{Relative errors of the computed invariants}
\label{fig:rel:error}
\end{figure}
\begin{figure}
\center
\includegraphics[width=0.7\linewidth]{validation_barrage_Vor_Riemann_2_.eps}
\caption{Dam-break: comparison between analytical (solid line) and numerical solutions ($\circ$) after the dam has broken. The dashed line represents the initial condition at $t=0$ }
\label{fig:dam_break}
\end{figure}
\vspace{0.2cm}
\newline
An undular bore is formed when a sudden discharge of water at rest of depth $h(1+\Delta)$ is initiated into a still water of depth $h$ (see figure \ref{fig:undular_sketch}). Herein, the bore is the region of transition between two uniform depths. To consider non-breaking undular bore, the initial relative difference in water level, $\Delta$, is chosen less than $0.28$. \citet{Favre} showed experimentally that beyond this value undular bores evolve to breaking. He conducted experiments on undular bores in a facility $73.58 \, m$ long, $0.42 \, m$ wide and $0.40 \, m$ high. He carried out a series of experiments on non breaking and breaking undular bores. Non breaking undular bores correspond to his experiments $N^{\circ}$ $2,\, 4, \, 6, \, 8, \, 10, \, 12$ for a water depth $h=0.205 \, m$ and $21, \, 22, \, 23$ for a water depth $h=0.1075 \, m$. The height of the leading wave, $h_{\mathrm{max}}$, which corresponds to a maximum waveheight was recorded at the end of the tank, after travelling on distances close to $300$ and $600$ depths. During these distances of propagation we can not ignore the shear stress acting at the bottom. Following \citet{Holloway} and \citet{Caputo} we introduce the following approximate dimensionless damping term based on the drag law for modelling the stress at the bottom
\begin{equation}
D(\eta)=k \eta |\eta|
\label{damping}
\end{equation}
where $k \sim 0.001 - 0.0026$ is an empirical dimensionless coefficient describing frictional effects at the bottom.
\newline
The time of diffusion, $t_d$, of negative vorticity generated at the bottom is $\mathcal{O}(h^2/\nu)$ where $\nu$ is the kinematic viscosity of water whereas the time of propagation, $t_p$, of the bore is $\mathcal{O}(L/\sqrt{gh})$ where $L$ is the length of the tank. For water, $t_p \ll t_d$ so that the vorticity due to molecular viscosity is assumed to have a negligible effect on the waves during their propagation.
\newline
Let $\zeta(x,t)=\eta(x,t)-\eta_0(x)$ where $\eta_0(x)=\eta(x,0)$ is the initial condition. We substitute $\eta=\zeta+\eta_0$, $h=1$, $g=1$ and $U_0=0$ into equations (\ref{nonlin-dispersive}) and (\ref{vor-KdV}), so these equations are dimensionless. The initial condition is
$\eta_0(x)=\Delta(1-\tanh(\alpha x))/2$ with $\alpha=1.25$,
so that $\lim \zeta = 0, \, x \rightarrow \pm \infty$.
\newline
To compare our numerical results with those experimental of \citet{Favre} we introduce in equations (\ref{nonlin-dispersive}) and (\ref{vor-KdV}) in dimensionless form the damping term given by equation (\ref{damping}) and set $U_0=0$ and $\Omega=0$. We have neglected vortical effects due to an underlying current because experimental bores considered herein were generated in water at rest. Experimental and numerical results are shown in figure \ref{fig:undular_comparison_Favre}. The dashed line corresponds to the experimental fit obtained by Favre in his figure 49. For strongly nonlinear waves ($0.30 \lesssim h_{\mathrm{max}} \lesssim 0.60$) results given by equation (\ref{nonlin-dispersive}) are close to those of experiments whereas KdV's results are in quite good agreement with experiments for weakly nonlinear waves ($h_{\mathrm{max}} \lesssim 0.20$). Beyond $h_{\mathrm{max}}=0.20$, the Boussinesq regime is no longer valid because the Ursell number becomes large and equation (\ref{nonlin-dispersive}) with its full nonlinearity is more appropriate. Within the framework of the KdV equation, we found that the profiles of the leading wave at $300$ depths are very close to those of solitary waves.
\section{Vorticity effect on undular bore properties}
We have ignored vorticity effects due to an underlying current in our numerical simulations of Favre's experiments. Generally, in natural conditions waves travel in the presence of currents and we cannot ignore the influence of vorticity. As emphasized by \citet{Teles} undular bores which travel upstream estuaries feel positive vorticity due to the boundary layer of the downstream current. To determine the kinematical and dynamical effects of the vorticity on the maximal height of the head wave and wavelength of the following wave train of the undular bore, numerical simulations haved been run for several values of $\Omega$. Waveheights are recorded at $100$ depths. In figure \ref{fig:undular_vor_Riemann_hmax} is shown the dimensionless height of the leading wave, $h_{\mathrm{max}}/\Delta$, as a function of the initial relative difference in water level, $\Delta$, for several values of the vorticity. In comparison to the case without vorticity and for a fixed value of $\Delta$, negative vorticity or positive vorticity increases or decreases the height of the leading wave, respectively. In figure \ref{fig:undular_vor_Riemann_lambda} is plotted the wavelength of the following waves as a function of $\Delta$ for several values of the vorticity. For a fixed value of $\Delta$, positive vorticity shortens the wavelength whereas negative vorticity lengthens the wavelength.
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.6\linewidth]{scheme_undular_bore.eps}
\caption{Sketch of the evolution of an undular bore from its initial position}
\label{fig:undular_sketch}
\end{figure}
\newline
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.4\linewidth]{comparison_favre_kdv_VorRiemann_hmax_delta_Frottement_bw.eps}
\includegraphics[width=0.4\linewidth]{comparison_favre_KdV_VorRiemann_lambda_1_2_delta_Forttement_bw.eps}
\caption{Dimensionless height of the leading wave (left) and dimensionless wavelength of the trailing waves (right) as a function of the initial relative difference in water level. Favre's experiments ($\circ$), equation (\ref{nonlin-dispersive}) with damping and $(U_0,\Omega)=(0,0)$ ($\ast$), KdV equation with damping and $(U_0,\Omega)=(0,0)$ ($\square$) }
\label{fig:undular_comparison_Favre}
\end{figure}
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.7\linewidth]{Vor_Riemann_hmax_Omega_delta.eps}
\caption{Dimensionless height of the leading wave as a function of the initial relative difference in water level for several values of the vorticity. $\Omega=0.20$ ($\square$), $\Omega=0$ ($\circ$) and $\Omega=-0.20$ ($\ast$).}
\label{fig:undular_vor_Riemann_hmax}
\end{figure}
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.7\linewidth]{Vor_Riemann_lambda_Omega_delta.eps}
\caption{Dimensionless wavelength of the following waves as a function of the initial relative difference in water level for several values of the vorticity. $\Omega=0.20$ ($\square$), $\Omega=0$ ($\circ$) and $\Omega=-0.20$ ($\ast$).}
\label{fig:undular_vor_Riemann_lambda}
\end{figure}
\section{Breaking time for hyperbolic waves and dispersive waves in the presence of vorticity}
\subsection{Within the framework of the hyperbolic equation (\ref{nonlin-nondispersive})}
Herein, we consider equation (\ref{nonlin-nondispersive}) in its dimensionless form by setting $g=1$ and $h=1$. Let us assume that $U_0=0$.
\newline
It is well known that nonlinear water waves in shallow water may evolve to breaking. The problem of breaking waves within the framework of nonlinear hyperbolic system has been tackled by many authors. The corresponding theory for the computation of the breaking time can be found in the book of \citet{Whitham}. For most water wave models in shallow water, wave breaking corresponds to the occurrence of an infinite slope of the wave profile.
\newline
The dimensionless form of equation (\ref{nonlin-nondispersive}) can be rewritten as follows
\begin{equation}
\eta_t+ \mathcal{C}(\eta) \eta_x=0
\label{hyperbolic-eq}
\end{equation}
where $\eta$ is now dimensionless and
\[\mathcal{C}(\eta)=-\frac{\Omega}{2}+ 2\sqrt{(\eta+1)+\Omega^2(\eta+1)^2/4}\]
\[ -\sqrt{1+\Omega^2/4} + \frac{1}{\Omega} \ln\left[1+ \frac{\Omega}{2}\frac{\Omega \eta +2(\sqrt{(\eta+1)+\Omega^2(\eta+1)^2/4}-\sqrt{1+\Omega^2/4})}{1+\Omega(\frac{\Omega}{2}+\sqrt{1+\Omega^2/4})} \right]\]
Equation (\ref{hyperbolic-eq}) is equivalent to the following system
\[\frac{d\eta}{dt}=0, \quad \mathrm{along \, the \, characteristic \, curve} \quad
\frac{dx}{dt}=\mathcal{C}(\eta)\]
The characteristic curves are straight lines in the $(x,t)$-plane.
\vspace{0.2cm}
\newline
Let $\eta_0(x)=\eta(x,0)$ be the initial condition and $x_0$ the point where the characteristic curve intersect the $x$-axis ($t=0$). The equation of this characteristic curve is
\[x=x_0+\mathcal{C}(\eta_0(x_0))t \]
\[x=x_0+\mathcal{V}(x_0)t \]
One can easily demonstrate that the slope of the profile at $t$ is
\[\frac{\partial{\eta}}{\partial x}=\frac{d\eta_0/dx_0}{1+\frac{d\mathcal{V}}{dx_0}t}\]
On any charateristic for which $\frac{d\mathcal{V}}{dx_0} < 0$ the slope of the profile becomes infinite when $t=-(d\mathcal{V}/dx_0)^{-1}$.
Consequently, breaking wave first occurs on the characteristic curve intersecting the $x$-axis at $x_0=x_{0_B}$ for which $\frac{d\mathcal{V}}{dx_0}(x_{0_B}) < 0$ with $|\frac{d\mathcal{V}}{dx_0}(x_{0_B})|$ is a maximum. The breaking time is
\begin{equation}
t_B=-(\frac{d\mathcal{V}}{dx_0}(x_{0_B}))^{-1}
\label{breaking_time}
\end{equation}
Herein, the breaking wave phenomenon can be understood as the blow-up of the slope in finite time $t_B$.
\vspace{0.2cm}
\newline
For dispersive waves in shallow water, there is no analytical expression of the breaking time similar to equation (\ref{breaking_time}). Nevertheless, the determination of the breaking time can be carried out numerically. To that purpose we have tested the efficiency of a numerical method for the detection of blow-up in finite time against the expression given by (\ref{breaking_time}). The chosen numerical method which allows the detection of possible occurrence of blow-up in finite time for a large class of nonlinear evolution equations was proposed by \citet{Sulem83}. The method is briefly presented in Appendix B. In figure \ref{fig:breaking_time_hyperbolic} is plotted the breaking time as a function of the vorticity. The initial condition is
\begin{equation}
\eta(x,0)=a \cos(kx)+ \frac{3-\sigma^2}{4\sigma^3}a^2 k \cos(2kx+\varphi)
\label{wind_profile}
\end{equation}
where $\sigma=\tanh(kh)$, $h=1$, $k=1$, $a=0.10$ and $\varphi = 0$.
\newline
Note that for $\varphi=0$ the initial profile is symmetric whereas for $\varphi \neq 0$ it is asymmetric.
In the latter case the initial asymmetric profile is typical of wind wave profiles with the leeward side steeper than the backward side.
\newline
The agreement between the analytical and numerical results is excellent. We can observe that the breaking time decreases as the current intensity increases in agreement with the results of Hur (2017). The presence of the shear current reduces the breaking time and speeds up the breaking phenomenon.
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.7\linewidth]{t_breaking_Om_th_num_New_sym_Hyperbolic_Whitham_2_.eps}
\caption{Theoretical and numerical breaking times as a function of the vorticity within the framework of equation (\ref{nonlin-nondispersive}). The solid line corresponds to the theoritical solution whereas the circles correspond to numerical values.}
\label{fig:breaking_time_hyperbolic}
\end{figure}
\subsection{Within the framework of the Whitham equation}
Since the KdV equation is not the appropriate model for describing wave evolution to breaking, \citet{Whitham} suggested as model the equation (\ref{Whitham}) (with $\Omega=0$). The Whitham equation and KdV equation have the same nonlinear term and different dispersive terms. The dispersive term of the Whitham equation corresponds to exact linear dispersion and consequently allows the introduction of small scales which are important in the breaking phenomenon. Constantin \& Escher(1998) have considered from a mathematical point of view breaking waves for Whitham-type equations without vorticity effect. They proved rigorously that a sufficiently asymmetric initial profile yields wave breaking. As emphasized by the latter authors, within the framework of weakly nonlinear waves we do not find a blow-up of the slope for symmetric initial profiles without vorticity as shown in figure \ref{fig:breaking_time_vor_Whitham}.
Due to imbalance in nonlinearity and dispersion for large values of $\Omega$, the breaking time is equal to that given by the hyperbolic equation (\ref{nonlin-nondispersive}). Note that for $\Omega$ larger than approximately $0.50$ breaking occurs. On the opposite, we found that a sufficiently asymmetric initial profile evolves to breaking without vorticity as shown in figure \ref{fig:breaking_time_vor_Whitham}. For that purpose we have considered the initial profile given by equation (\ref{wind_profile}) with $a=0.16$ and $\varphi \neq 0$. In that case the initial profile is asymmetric, typical of wind wave profiles with the leeward slope steeper than the backward slope, and breaking occurs for $\Omega$ larger than a negative threshold value close to zero ($ \approx -0.10$). Note that for amplitude $a<0.16$ and phase $\varphi=0$ we did not observe a blow-up of the slope of the leeward side of the crest. In addition to the asymmetry of the wave profile a condition on the nonlinearity of the initial profile is required to obtain wave breaking. Consequently, for an insufficient asymmetry of the initial condition an increase of its amplitude is required to observe breaking of the wave. Negative values of the vorticity stimulate the breaking phenomenon (opposing current) even though breaking occurred for weak positive values of the vorticity (advancing current). Generally, we can observe that there is no blow-up of the leeward slope when $\Omega < 0$. In that case, because the wave propagates on advancing current, we can expect that both the waveheight and wave slope will not increase. We have run numerical simulations for several negative values of $\Omega$ with symmetric and asymmetric initial wave profiles. The long-time evolution of the wave profiles has shown that the waveheight and leeward slope do not increase. The effect of the advancing current is to prevent the blow-up occurrence. Note that in the case of weak negative values of $\Omega$ the blow-up occurs. In that case, the intensity of advancing current is not sufficiently strong to prevent the onset of breaking wave.
\newline
In order to satisfy the criterion of weakly nonlinear waves we have reduced the value of the amplitude of the initial condition and increase the asymmetry of its profile. A forcing term $F$ is applied during a short period of time which consists of a sine progressive wave with the phase velocity $c$, in quadrature with the surface elevation $\eta(x,t)$.
\begin{eqnarray}
F=\epsilon \sin t \sin (x-ct), \qquad 0 \leq t \leq \delta t \nonumber \\
F=0, \qquad \qquad \qquad \hspace{1.9cm} t > \delta t \nonumber
\label{forcing}
\end{eqnarray}
with $\epsilon=0.50$, $c=1$ and $\delta t =0.8$.
\newline
The initial condition is given by equation (\ref{wind_profile}) with $a=0.10$ and $\varphi=0$, and the forcing $F$ is applied during $\delta t$. In that case a blow-up of the slope is obtained. We can conclude that within the framework of weakly nonlinear waves a sufficient asymmetry of the initial profile yields to wave breaking as demonstrated by Constantin \& Escher (1998).
\newline
Figure \ref{fig:breaking_time_vor_Whitham} shows that the breaking time decreases as $\Omega$ increases.
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.6\linewidth]{t_breaking_init_cond_Sym_Vor_Whitham_a016.eps}
\includegraphics[width=0.6\linewidth]{t_breaking_init_cond_asym_Vor_Whitham_a016.eps}
\caption{Breaking time as a function of the vorticity for the symmetric initial condition ($a=0.16, \,\varphi=0$) (top) and for the asymmetric initial condition ($a=0.16, \,\varphi=3\pi/2$) (bottom). The solid line corresponds to the hyperbolic equation (\ref{nonlin-nondispersive}) and the circle to the Whitham equation (\ref{Whitham}).}
\label{fig:breaking_time_vor_Whitham}
\end{figure}
\subsection{Within the framework of the generalised Whitham equation}
The Whitham equation governs the evolution of weakly nonlinear waves whereas the generalised Whitham equation describes the evolution of fully nonlinear waves. In figures \ref{fig:breaking_time_Gen_Whitham} is plotted the breaking time as a function of the vorticity for symmetric and asymmetric initial profiles given by equation (\ref{wind_profile}), respectively. For an initial amplitude $a=0.20$ we observe the blow-up of the slope for both symmetric and asymmetric initial profiles as soon as $\Omega$ becomes larger than a given threshold ($0.20$ and $-0.71$ for the symmetric and asymmetric initial conditions, respectively). Note that for this value of the amplitude the symmetric initial profile does not evolve to breaking in the absence of vorticity dispersive effects prevent the blow-up occurrence of the slope. The trend observed within the framework of the Whitham equation is amplified due to stronger nonlinearity. Like in the case of the Whitham equation, sufficiently powerful advancing current can eliminate the onset of breaking wave.
\newline
Figure \ref{fig:breaking_time_Gen_Whitham} shows that the breaking time decreases as the current intensity $\Omega$ increases.
\newline
For $\Omega=1$, figures \ref{fig:evolution_profile_Om1_Gen_Whitham} show the profiles of the wave at different times before breaking for symmetric and asymmetric initial conditions, respectively. The amplitude and asymmetry of the profile increase as the wave evolves to breaking. We can observe the incipient formation of a jet at the crest of the wave.
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.6\linewidth]{t_breaking_Gen_Whitham_Om_CI_sym.eps}
\includegraphics[width=0.6\linewidth]{t_breaking_Gen_Whitham_Om_CI_asym.eps}
\caption{Breaking time as a function of the vorticity within the framework of the generalised Whitham equation with a symmetric initial condition ($a=0.20, \, \varphi=0$) (top) and an asymmetric initial condition ($a=0.20, \, \varphi=3 \pi/2$) (bottom). The solid line corresponds to the hyperbolic case.}
\label{fig:breaking_time_Gen_Whitham}
\end{figure}
\begin{figure}
\vspace{0.2cm}
\center
\includegraphics[width=0.6\linewidth]{formes_Om_1_sym_Gen_Whitham.eps}
\includegraphics[width=0.6\linewidth]{formes_Om_1_asym_Gen_Whitham.eps}
\caption{Time evolution of the initial symmetric profile ($a=0.20, \, \varphi=0$) (top) and asymmetric initial profile ($a=0.20, \, \varphi=3 \pi/2$) (bottom) to breaking for $\Omega=1$.}
\label{fig:evolution_profile_Om1_Gen_Whitham}
\end{figure}
\section{Conclusion}
Following \citet{Whitham} approach, we have derived from Riemann invariants, in the presence of constant vorticity, a new single partial differential equation governing the spatio-temporal evolution of the free surface elevation, then the velocity can be determined through an algebraic equation by using Riemann invariants. This system of partial differential equation and algebraic equation is fully equivalent to the St Venant equations for $1D$-propagation and allows the resolution of the evolution of hyperbolic waves of arbitrary height in the presence of vorticity. To take account of dispersion, we have introduced in the previous partial differential equation the exact linear dispersion of gravity waves on finite depth and in the presence of vorticity. From the latter equation we have derived the Whitham equation for weakly nonlinear water waves propagating in the presence of constant vorticity. Under the assumption of weakly nonlinear and weakly dispersive waves we have rediscovered in more straightforward manner the KdV equation with vorticity derived previously by \citet{Freeman} and \citet{Choi}. We found that solitary wave and cnoidal wave profiles are broader for positive vorticity whereas they are narrower for negative vorticity. Furthermore, positive vorticity increases their phase velocity whereas it is the opposite in the presence of negative vorticity. The Whitham equation with non zero vorticity has been tackled, too.
\newline
To follow the evolution of non breaking undular bores, we have implemented a pseudo-spectral numerical method. The validity of the numerical method has been checked against the exact analytical solution of the St-Venant equations for the dam-break problem in the absence of current and vorticity. A second validation has been carried out by using Favre's experiments on undular bores with frictional effect at the bottom and without vorticity.
\newline
A particular attention has been paid to non breaking undular bores in the presence of vorticity, too. We have shown that negative vorticity increases the height of the leading wave whereas it decreases in the presence of positive vorticity. The wavelength of the following waves is also modified by vortical effects, it is shortened in the presence of negative vorticity and stretched for positive vorticity.
Within the framework of deep water, \citet{Touboul} using a Boundary Integral Element Method found similar results.
\newline
An investigation on the breaking time of dispersive waves in the presence of constant vorticity has been carried out within the framework of the Whitham equation and generalised Whitham equation, respectively. The numerical method to capture the blow-up of the slope corresponding to the onset of the breaking has been checked against the analytical solution of the St-Venant equation in 1D. Numerical and exact results are in excellent agreement. We found that the breaking time of hyperbolic waves decreases as the vorticity magnitude increases whatever its sign. The breaking time of dispersive waves decreases as the shear intensity, $\Omega$, increases. However, the cases we have considered have shown that dispersive waves propagating in the presence of a strong positive vorticity ($\Omega<0$) do not evolve to breaking.
\vspace{1cm}
\newline
{\bf Appendix A. Periodic travelling-wave solutions of the Whitham equation}
\vspace{0.3cm}
\newline
The Whitham equation (\ref{Whitham}) is rewritten as
\[
\eta_{t}+c_{1}(\Omega)\eta\eta_{x}+K*\eta_{x}=0,\quad c_{1}(\Omega)=\frac{3gh+h^{2}\Omega^{2}}{h\sqrt{gh(4gh+h^{2}\Omega^{2})}}.
\]
Travelling-wave solutions of the form $\eta(x,t)=\phi(x-ct)$, for
a given phase velocity $c$, are sought. They are solutions of the stationary
equation:
\[
-c\phi+c_{1}(\Omega)\frac{\phi^{2}}{2}+K*\phi=0.
\]
For $\Omega=0$, \citet{Ehrnstrom} proved the
existence of a branch of $2\pi$-periodic traveling-wave solutions
for the Whitham equation. We will show, numerically, the existence
of $2\pi$-periodic branches of solutions for $\Omega\ne0.$ Let the
approximate solution be:
\[
\phi_{N}(X)=\sum_{n=0}^{N}a_{n}\cos(nX),\quad X\in[0,2\pi].
\]
The residual is then:
\[
R_{N}(a_{0},a_{1},...,a_{N})=-V\phi_{N}+c_{1}(\Omega)\frac{\phi_{N}^{2}}{2}+K*\phi_{N}.
\]
We use a pseudo-spectral method. The nonlinear term is computed in
the physical space ($[0,2\pi]$ is discretized using a regular grid).
The convolution is evaluated in the spectral space using Fast Fourier
Transforms. To make the residual minimal, we use a Galerkin method:
\[
\langle R_{N},\cos(nX)\rangle=\int_{0}^{2\pi}R_{N}\cos(nX)dX=0,\quad n=0,N,
\]
and this gives $N+1$ nonlinear equations for the unknowns $a_{n}$.
These equations are solved using the fsolve routine from matlab. The
initial guess is a cosine wave with a small amplitude and a phase velocity
$c$ estimated using the linear dispersion relation. We used $N=20$,
and the computations are stopped when the residual norm is $O(10^{-14})$.
Then, we employ a continuation method to obtain solutions for different
values of $c$. For $\Omega=0$ we checked that our results are in
agreement with those of Sanford et al \citet{Sanford}. The results
of computations for different values of $\Omega$ are given in section 2.
\vspace{0.3cm}
\newline
{\bf Appendix B. Detection of blow-up occurrence in finite-time}
\newline
The blow-up after a finite time is detected using the method of the analyticity strip presented in \citet{Sulem83}. The essence of the method is the following.
When $\eta(x,t)$ is an analytic function, its Fourier coefficients (with respect to $x$) decay faster than any power of $1/k$ ($k$ is the wavenumber) in the limit $k\rightarrow\infty$. When $\eta$ is singular, its Fourier coefficients decay algebraically with $1/k$. Hence, to detect the time of the appearence of the singularity we assume that the Fourier coefficients of the solution $\eta(x,t)$ behave as:
$$\hat\eta_k(t)=C(t)k^{-\alpha(t)}e^{-\delta(t)k}.$$
The adjustable coefficients $C$, $\alpha$, $\delta$ are calculated using a least sqaure method. The coeffecient $\delta$ is known as the analyticity strip width. Loss of regularity correponds to the vanishing of $\delta(t)$ and the corresponding time gives the time of breaking $t_B$. This method is validated against the time of breaking when all the equations studied here are hyperbolic (vanishing dispesion). In those cases, the expression of the time of breaking could be obtained analytically as a function of $\Omega$. An example of this validation is given in figure \ref{fig:breaking_time_hyperbolic}.
|
\section{Introduction}
\vspace{-2mm}
\input{intro}
\vspace{-2mm}
\section{Related Work}
\vspace{-2mm}
\input{related}
\vspace{-2mm}
\section{Approach}
\vspace{-2mm}
\seclabel{learning}
\input{learning}
\vspace{-2mm}
\section{Experiments}
\vspace{-2mm}
\seclabel{experiments}
\input{experiments}
\vspace{-2mm}
\section{Discussion}
\vspace{-1mm}
We presented a framework that allows learning single-view prediction of 3D structure without direct supervision for shape or pose. While this is an encouraging result that indicates the feasibility of using natural forms of supervision for this task, a number of challenges remain to be addressed. As our supervisory signal, we rely on consistency with validation images of unoccluded objects and it would be useful to deal with unknown occlusions. It would also be interesting to apply similar ideas for learning the 3D structure of general scenes though this might additionally require leveraging alternate 3D representations and allowing for object motion to handle dynamic scenes.
\vspace{2mm}
\noindent \textbf{Acknowledgements.}
{We thank David Fouhey for insightful discussions, and Saurabh Gupta and Tinghui Zhou for helpful comments. This work was supported in part by Intel/NSF VEC award IIS-1539099 and NSF Award IIS-1212798. We gratefully acknowledge NVIDIA corporation for the donation of GPUs used for this research.}
{\small
\bibliographystyle{ieee}
\subsection{Empirical Analysis using ShapeNet}
\subsubsection{Experimental Setup}
\vspace{-1mm}
\renewcommand{\arraystretch}{1.2}
\setlength{\tabcolsep}{4pt}
\begin{table*}[t]
\centering
\footnotesize
\hfill %
\begin{subtable}{}
\begin{tabular}{l c cc cc cc}
\toprule
Training & \multicolumn{1}{c}{3D} & \multicolumn{2}{c}{{Multi-view}} & \multicolumn{2}{c}{{Multi-view}} & \multicolumn{2}{c}{{Multi-view w/o}} \\
Data & & \multicolumn{2}{c}{{\& GT Pose}} & \multicolumn{2}{c}{{ w/o Rot}} & \multicolumn{2}{c}{{Rot \& Trans}}
\\
\cmidrule(lr){3-4}
\cmidrule(lr){5-6}
\cmidrule(lr){7-8}
{class} & & {Mask} & {Depth} & {Mask} & {Depth} & {Mask} & {Depth} \tabularnewline
\midrule
{aero} & 0.57 & 0.55 & 0.43 & 0.52 & 0.44 & 0.38 & 0.37 \tabularnewline
{car} & 0.79 & 0.75 & 0.69 & 0.74 & 0.71 & 0.48 & 0.68 \tabularnewline
{chair} & 0.49 & 0.42 & 0.45 & 0.40 & 0.43 & 0.35 & 0.37 \tabularnewline
\midrule {mean} & 0.62 & 0.57 & 0.52 & 0.55 & 0.53 & 0.40 & 0.47 \tabularnewline
\bottomrule
\end{tabular}
\end{subtable}%
\hfill %
\begin{subtable}{}
\begin{tabular}{l cc cc cc cc cc}
\toprule
Training & \multicolumn{2}{c}{GT} & \multicolumn{4}{c}{MV w/o Rot} & \multicolumn{4}{c}{MV w/o Rot \& Trans}
\\
Data & \multicolumn{2}{c}{Pose} & \multicolumn{2}{c}{Mask} & \multicolumn{2}{c}{Depth} & \multicolumn{2}{c}{Mask} & \multicolumn{2}{c}{Depth}
\\
\cmidrule(lr){2-3}
\cmidrule(lr){4-5}
\cmidrule(lr){6-7}
\cmidrule(lr){8-9}
\cmidrule(lr){10-11}
{class} & Acc & Err & Acc & Err & Acc & Err & Acc & Err & Acc & Err \tabularnewline
\midrule
{aero} & 0.79 & 10.7 & 0.69 & 14.3 & 0.60 & 21.7 & 0.53 & 26.9 & 0.63 & 12.3 \tabularnewline
{car} & 0.90 & 7.4 & 0.87 & 5.2 & 0.85 & 4.9 & 0.53 & 24.8 & 0.56 & 20.6 \tabularnewline
{chair} & 0.85 & 11.2 & 0.81 & 7.8 & 0.83 & 8.6 & 0.55 & 24.0 & 0.62 & 19.1 \tabularnewline
\midrule {mean} & 0.85 & 10.0 & 0.79 & 9.0 & 0.76 & 11.7 & 0.54 & 25.1 & 0.61 & 17.4 \tabularnewline
\bottomrule
\end{tabular}
\end{subtable}%
\hfill %
\vspace{2mm}
\caption{Analysis of the performance for single-view shape (Left) and pose (Right) prediction. a) Shape Accuracy: Mean IoU on the test set using various supervision settings. b) Pose Accuracy/Error: Acc$_\frac{\pi}{6}$ and Med-Err across different supervision settings.}
\tablelabel{snetEval}
\vspace{-2mm}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{figures/resultsPoseMesh.png}
\caption{\small Rotation predictions on a random subset of the validation images. For visualization, we render the ground-truth voxel occupancies using the corresponding rotation. Left to Right: a) Input Image b) Ground-truth Rotation c) GT Supervised Prediction d,e) Multi-view w/o Rot Supervision (Mask, Depth), and f,g) Multi-view w/o Rot and Trans Supervision (Mask, Depth)}
\vspace{-2mm}
\figlabel{posePred}
\end{figure*}
\noindent \textbf{Dataset.}
We use the ShapeNet dataset~\cite{shapenet2015} to empirically validate our approach. We evaluate on three representative object categories with a large number of models : airplanes, cars, and chairs. We create random train/val/test splits with $(0.7,0.1,0.2)$ fraction of the models respectively. For each training model, we use $N_i = 5$ images available from different (unknown) views with corresponding depth/mask observations. The images are rendered using blender and correspond to a viewpoint from a randomly chosen azimuth $\in [0,360)$ degrees and elevation $\in [-20,40]$ degrees. We additionally use random lighting variations during rendering.
We also render the training objects under two settings - a) origin centred, or b) randomly translated around the origin. As the camera is always at a fixed distance away from the origin, the first setting corresponds to training with a known camera translation, but unknown rotation. The second corresponds to training with both translation and rotation unknown. To have a common test set across various control setting (and compare to~\cite{drcTulsiani17}), we use the origin centered renderings for our validation and test sets. We note that these rendering settings are rather challenging and correspond to significantly more variation than commonly examined by previous multi-view supervised methods which examine settings with fixed translation~\cite{drcTulsiani17}, and sometimes only consider 24~\cite{yan2016perspective} or even 8~\cite{gadelha20163d} possible discrete views.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{figures/alignment.png}
\vspace{-4mm}
\caption{\small Visualization of 8 random predicted shapes from azimuth $=60^{\circ}$, elevation $=30^{\circ}$. Left: Original predictions from the shape CNN. Right: Shape predictions transformed according to the optimal rotation.}
\vspace{-2mm}
\figlabel{alignment}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{figures/posePrior.png}
\vspace{-4mm}
\caption{\small Visualization of the predicted pose distribution under various training settings. Each small image is placed at the (predicted/known) location of the corresponding camera. The reference grid in the centre depicts the space in which shape is predicted. Left to Right : a) Ground-truth poses b) No pose prior c) True pose prior d) Incorrect pose prior, discarded midway through training. See text for details.}
\vspace{-3mm}
\figlabel{posePrior}
\end{figure*}
\vspace{1mm}
\noindent \textbf{Control Settings.} In addition to reporting the performance in the scenario where pose and shape supervision is unavailable, we also examine the settings where stronger supervision \eg shape or pose can be used. These experiments serve to highlight the upper bound performance. In all the experiments, we train a separate model per object category. The various settings studied are :
\emph{3D Supervision.} To mimic the setup used by 3D supervised approaches~\cite{choy20163d,Girdhar16b}, we assume known ground-truth 3D models for each training image and train the shape CNN using a cross-entropy loss.
\emph{Multi-view with Ground-truth Pose.} In this supervision setting used by previous multi-view supervised approaches, pose (but not shape) supervision is available for the multiple observations. We use our loss function but train the shape prediction CNN $f_s$ using the ground-truth pose instead of predicted poses. We separately train the pose prediction CNN $f_p$ using squared L2 loss in quaternion space (after accounting for antipodal symmetry of quaternions).
\emph{Multi-view without Pose Supervision.} This represents our target setting with the weakest form of supervision available. We train the shape and pose prediction CNNs jointly using our proposed loss. Further, we consider two variants of this setting - one where camera translation is known, one where both camera translation and rotation are unknown.
\vspace{1mm}
\noindent \textbf{Evaluation Metrics.} We report the results using predictions for 2 images per test model. For evaluating the shape prediction CNN, we report the mean intersection over union (IoU) between the ground-truth and predicted shapes. Since different CNNs can be calibrated differently, we search for the optimal threshold (per CNN on the validation set) to binarize the predictions.
To evaluate the rotation prediction, we measure the angular distance between the predicted and ground-truth rotation (in degrees) and report two metrics : a) Fraction of instances with error less than 30 degrees (Acc$_\frac{\pi}{6}$), and b) Median Angular Error (Med-Err).
\vspace{-1mm}
\subsubsection{Results}
\vspace{-1mm}
\seclabel{mainExp}
\noindent \textbf{Prediction Frame Alignment.} The ShapeNet models are all aligned in a canonical frame where X and Y axes represent lateral and upward directions. The shape and pose prediction CNNs learned using our approach are not constrained to adhere to this frame and in practice, learn to predict shape and pose w.r.t some arbitrary frame.
However, to evaluate these predictions, we compute an optimal rotation to best align the predictions to the canonical ShapeNet frame. We use 8 \emph{random} images per category (the first validation mini-batch) alongwith the ground-truth 3D voxelizations and search for a rotation that maximizes the voxel overlap between the ground-truth and the rotated predicted shapes. We visualize the prediction frame alignment for car and chair CNNs trained using multi-view observations w/o pose via depth verification images in \figref{alignment}. Note that the prediction frames across classes vary arbitrarily. After the alignment process, the predictions for both categories are in the canonical ShapeNet frame.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{figures/ebay_vis.png}
\vspace{-4mm}
\caption{\small Visualization of predictions using the Stanford Online Product Dataset. (Top) Input image. (Middle) Predicted shape in the emergent canonical pose. (Bottom) Predicted shape rotated according to the predicted pose.}
\vspace{-2mm}
\figlabel{ebay}
\end{figure*}
\vspace{1mm}
\noindent \textbf{Role of a Pose prior.} While the empirical results reported below correspond to using the correct pose prior, we first show that the primary benefit of this prior is that it encourages the CNN to predict diverse poses and avoid local minima, and that even an approximate prior is sufficient. To further support this point, we conducted an experiment where we used an incorrect pose prior (elevation uniform $\in [-40,80]$ instead of $\in [-20,40]$) and removed the prior loss midway through training. We observed that this network also trained successfully, indicating that we do not require the true pose prior, rather only an approximate one. \figref{posePrior} visualizes the pose distributions inferred under various settings. While using no prior results in a local optima, using the approximate prior (or the correct prior) does not.
\vspace{1mm}
\noindent \textbf{Single-view Shape Prediction.} Our results and the performance under various control settings with stronger supervision is reported in \tableref{snetEval} and visualized in \figref{shapePred}. In general, we observe that the results using our approach are encouragingly close to those obtained using much stronger forms of supervision. This clearly indicates that our approach is able to learn single-view shape prediction despite the lack of either shape or pose information during training. As expected, we also observe that we cannot learn about concavities in chairs via consistency against mask validation images, though we can do so using depth images. e observe a noticeable performance drop in case of mask supervision with unknown translation, as this settings results in scale ambiguities which our evaluation does not account for \eg we learn to predict larger cars, but further away, and this results in a low empirical score.
\vspace{1mm}
\noindent \textbf{Single-view Pose Estimation.} The results of our approach are reported in \tableref{snetEval} and visualized in \figref{posePred}. We observe a similar trend for the task of pose prediction -- that our approach performs comparably to directly supervised learning using ground-truth pose supervision. Interestingly, we often get lower median errors than the supervised setting. We attribute this to the different topologies of the loss functions. The squared L2 loss used in the supervised setting yields small gradients if the pose is almost correct. Our consistency loss however, would want the observation image to perfectly align with the shape via the predicted pose.
\vspace{1mm}
\noindent \textbf{Interpretation.}
The main takeaway from these results is that it is indeed possible to learn shape and pose prediction without direct supervision for either. We empirically and qualitatively observe competitive performances for both these tasks when compared to approaches that leverage stronger forms of supervision. We see that we always learn meaningful shape and pose prediction systems across observation types (mask/depth) and that performance degrades gracefully when using less supervision (known/unknown translation).
\vspace{-1mm}
\subsection{Learning from Online Product Images}
\vspace{-1mm}
\noindent \textbf{Dataset.}
We examined the `chair' object category from the Stanford Online Products Dataset~\cite{song16product} which comprises of automatically downloaded images from eBay.com~\cite{eBay}. Since multiple images (views) of the same product are available, we can leverage our approach to learn from this data. As we also require associated foreground masks for these images, we use an out-of-the-box semantic segmentation system~\cite{chen2017deeplab} to obtain these. However, the obtained segmentation masks are often incorrect. Additionally, many of the product images were not suited for our setting as they only comprised of a zoom-in of a small portion of the instance (\eg chair wheel). We therefore manually selected images of unoccluded/untruncated instances with a reasonably accurate (though still noisy) predicted segmentation. We then used the object instances with atleast 2 valid views for training. This results in a filtered dataset of $N = $ 282 instances with $N_i = $ 3.65 views on average per instance.
\vspace{1mm}
\noindent \textbf{Results.}
We can apply our approach to learn from this dataset comprising of multiple views with associated (approximate) foreground masks. Since the camera intrinsics are unknown, we assume a default intrinsic matrix (see appendix). We then learn to predict the (unknown) translation and rotation via $f_p$ and the (unknown) shape via $f_s$ using the available multi-view supervision. Note that the learned CNNs are trained from scratch, and that we use the same architecture/hyperparameters as in the ShapeNet experiments.
Some results (on images of novel instances) using our learned CNN are visualized in \figref{ebay}. We see that we can learn to predict meaningful 3D structure and infer the appropriate shape and pose corresponding to the input image. Since only foreground mask supervision is leveraged, we cannot learn to infer the concavities in shapes. We also observe confusion across poses which result in similar foreground masks. However, we feel that this result using training data derived from a challenging real world setting, concretely demonstrates our method's ability to learn despite the lack of direct shape or pose supervision. To the best of our knowledge, this is the first such result and it represents an encouraging step forward.
\subsection{Pose-Differentiable Ray Consistency}
\vspace{1mm}
\noindent \textbf{Sampling Occupancies along a Ray.}
To define the consistency cost function $L_p(\bar{x}, C; v_p)$, we need to consider the ray as it is passing through the probabilistically occupied voxel grid $\bar{x}$. We do so by looking at discrete points sampled along the ray. Concretely, we sample points at a pre-defined set of $N = 80$ depth values $\{d_i | 1 \leq i \leq N\}$ along each ray. We denote by $x^p_i$ the occupancy value at the $i^{th}$ sample along this ray. To determine $x^p_i$, we look at the 3D coordinate of the corresponding point. Note that this can be determined using the camera parameters. Given the camera intrinsic parameters ($f_u,f_v,u_0,v_0$), the ray corresponding to the image pixel $(u,v)$ travels along the direction $(\frac{u-u_0}{f_u}, \frac{v-v_0}{f_v},1)$ in the camera frame. Therefore, the $i^{th}$ point along the ray, in the camera coordinate frame, is located at $l_i \equiv (\frac{u-u_0}{f_u}d_i, \frac{v-v_0}{f_v}d_i,d_i)$. Then, given the camera extrinsics $(R, t)$, we can compute the location of his point in the coordinate frame of the predicted shape $\bar{x}$. Finally, we can use trilinear sampling to determine the occupancy at this point by sampling the value at this using the occupancies $\bar{x}$. Denoting by $T(G,pt)$ a function that samples a volumetric grid $G$ at a location $pt$, we can compute the occupancy sampled at the $i^{th}$ as below.
\begin{gather}
x^p_i = \mathcal{T}( \bar{x} , R \times (l_i + t)~); \\
l_i \equiv (\frac{u-u_0}{f_u}d_i, \frac{v-v_0}{f_v}d_i,d_i)
\end{gather}
Note that since the trilinear sampling function $T$ is differentiable w.r.t its arguments, the sampled occupancy $x^p_i$ is differentiable w.r.t the shape $\bar{x}$ and the camera $C$.
\vspace{1mm}
\noindent \textbf{Probabilistic Ray Tracing.}
We have so far considered the ray associated with a pixel $p$ and computed samples with corresponding occupancy probabilities along it. We now trace this ray as it travels forward and use the samples along the ray as checkpoints. In particular, we assume that when the ray reaches the point corresponding to the $i^{th}$ sample, it either travels forward or terminates at that point. Conditioned on the ray reaching this sample, it travels forward with probability $x^p_i$ and terminates with likelihood $(1-x^p_i)$. We denote by $\mathbf{z}^p \in \{1, \cdots , N+1\}$ a random variable corresponding to the sample index where the ray (probabilistically) terminates, where $z^p = N+1$ implies that the ray escapes. We call these probabilistic ray terminations as \emph{ray termination events} and can compute the probability distribution $q(z_p)$ for these.
\begin{gather}
\eqlabel{pzuv}
q(z^p = i) = (1 - x^p_i) \prod_{j=1}^{i-1} x^p_j~~~~\forall (i \leq N) ; \\ q(z^p = N+1) = \prod_{j=1}^{N} x^p_j;
\end{gather}
\noindent \textbf{Event Costs.} Each event corresponds to the ray terminating at a particular point. It is possible to assign a cost to each event based on how inconsistent it is to w.r.t the pixel value $v_p$. If we have a depth observation $v_p \equiv d_p$, we can penalize the event $z^p = i$ by measuring the difference between $d_p$ and $d_i$. Alternatively, if we have a foreground image observation \ie $v_p \equiv s_p \in \{0,1\}$ where $s_p = 1$ implies a foreground pixel, we can penalize all events which correspond to a different observation. We can therefore define a cost function $\psi_p(i)$ which computes the cost associated with event $z_p=i$.
\begin{gather}
\psi_p^{depth}(i) = |d_p - d_i|; \\
\psi_p^{mask}(i) = |s_p - \mathds{1}(i \leq N)|;
\end{gather}
\vspace{1mm}
\noindent \textbf{Ray Consistency Cost.}
We formulated the concept of ray termination events, and associated a probability and a cost to these. The ray consistency cost is then defined as the expected event cost.
\begin{gather}
L_p(\bar{x},C; v_p)~~=~~\underset{z_p}{\mathbb{E}}~\psi_p(z_p)~~=~~\sum_{i=1}^{N} q(z_p=i) \psi_p(i)
\end{gather}
Note that the probabilities $q(z_p=i)$ are a differentiable function of $x_p$ which, in turn, is a differentiable function of shape $\bar{x}$ and camera $C$. The view consistency loss, which is simply a sum of multiple ray consistency terms, is therefore also differentiable w.r.t the shape and pose.
\vspace{1mm}
\textbf{Relation to Previous Work.} The formulation presented draws upon previous work on differentiable ray consistency~\cite{drcTulsiani17} and leverages the notions of probabilistic ray termination events and event costs to define the ray consistency loss. A crucial difference however, is that we, using trilinear sampling, compute occupancies for point samples along the ray instead of directly using the occupancies of the voxels in the ray's path. Unlike their formulation, this allows our loss to also be differentiable w.r.t pose which is a crucial requirement for our scenario. Yan \etal~\cite{yan2016perspective} also use a similar sampling trick but their formulation is restricted to specifically using mask verification images and is additionally not leveraged for learning about pose. Tulsiani \etal~\cite{drcTulsiani17} also discuss how their formulation can be adapted to use more general verification images \eg color, semantics \etc using additional per-voxel predictions. While our experiments presented in the main text focus on leveraging mask or depth verification images, a similar generalization is possible for our formulation.
\subsection{Geometric Consistency as Supervision}
\vspace{-1mm}
Multiple images of the same instance are simply renderings of a common geometry from diverse viewpoints. Therefore, to correctly `explain' multiple observations of an instance, we need the correct geometry (shape) of the instance and the corresponding viewpoints (pose) for each image. Our approach, which is depicted in \figref{overview}, builds on this insight and proposes to predict \emph{both}, shape and pose s.t. the available multi-view observations can be explained.
Concretely, during training, we use one image of an instance to predict the instance shape. In parallel, we use a \emph{different} image of the same instance to predict pose. Then, we enforce that the predicted shape, when viewed according to the predicted pose, should be consistent with a depth/mask image from the latter view. We therefore use the notion of \emph{consistency} as a form of meta-supervision \ie while the ground-truth shape and pose are unknown, we know that they should be consistent with the available verification image. After the training stage, our learned models can infer shape and pose from a single view of a novel instance.
A crucial aspect of the designed training setup is that the shape and pose estimates are \emph{independently} obtained from \emph{different} images of the same instance. This enforces that the optimal solution corresponds to predicting the correct shape and pose. Another interesting property is that the shape is predicted in an emergent canonical, view-independent frame, and the predicted pose is with respect to this frame.
\vspace{1mm}
\noindent \textbf{Correctness of Optimal Shape and Pose.}
We consider \figref{overview} and first examine the shape prediction CNN $f_s$. It predicts a shape $f_s(I_1)$ given some input image. This shape is verified against $V$ from a different view which is unknown to $f_s$. The optimal predicted shape should therefore be consistent with \emph{all} possible novel views of this instance, and therefore correspond to the true shape (upto some inherent ambiguities \eg concavities in case of mask supervision). Similarly, the pose prediction CNN $f_p$ is required to infer a viewpoint under which the predicted geometry can explain the verification image $V$. As $V$ is chosen to be from the same viewpoint as the image $I_2$, the pose CNN should predict the correct viewpoint corresponding to its input image ($I_2$).
\vspace{1mm}
\noindent \textbf{Emergent Canonical Frame.}
Under our proposed setup, the predicted pose $f_p(I_2)$ is agnostic to the image $I_1$. However, to explain the verification image $V$, the pose CNN is required to predict a pose w.r.t the inferred shape $f_s(I_1)$. So how can $f_p$ infer pose w.r.t $f_s(I_1)$ when it does not even have access to $I_1$? The resolution to this is that the shape prediction CNN $f_s$ automatically learns to predict shape in some (arbitrary) view-agnostic canonical frame (e.g. `front' of chairs may always face towards the X axis), and the pose CNN $f_p$ learns to predict pose w.r.t this frame. Therefore, even though it is not explicitly enforced, our approach of independently inferring shape and pose makes the learnt CNNs automatically adhere to some emergent canonical frame.
Towards implementing our framework, we require a consistency loss $L(\bar{x}, C; V)$ which measures whether the (predicted) shape $\bar{x}$ and camera pose $C$ can geometrically explain a depth/mask image $V$. We present a formulation for this loss in \secref{pdrc} and then describe the training process in \secref{training}. We finally describe some modifications required to make the training more robust.
\vspace{-1mm}
\subsection{Pose-differentiable Consistency Loss}
\vspace{-1mm}
\seclabel{pdrc}
We formulate a view consistency loss $L(\bar{x},C; V)$ that measures the inconsistency between a shape $\bar{x}$ viewed according to camera $C$ and a depth/mask image $V$. Our formulation builds upon previously proposed differentiable ray consistency formulation~\cite{drcTulsiani17}. However, unlike the previous formulation, our proposed view consistency loss is differentiable w.r.t pose (a crucial requirement for usage in our learning framework). Here, we very briefly recall the previous formulation and mainly highlight our proposed extension. A more detailed and complete formulation of the view consistency loss can be found in the appendix.
\vspace{1mm}
\noindent \textbf{Differentiable Ray Consistency~\cite{drcTulsiani17}.} The view consistency loss formulated by Tulsiani \etal~\cite{drcTulsiani17} could be decomposed into per-pixel (or ray) based loss terms where $L_p(\bar{x},C; v_p)$ denotes the consistency of the shape and camera with the observation $v_p$ at pixel $p$. The per-pixel loss is defined as the \emph{expected event cost}:
\vspace{-2mm}
\begin{gather}
L_p(\bar{x},C; v_p)~~=~~\sum_{i=1}^{N} q_p(i) \psi_p(i)
\vspace{-2mm}
\end{gather}
Here, $\psi_p(i)$ denotes the cost for each event, determined by $v_p$, and $q_p(i)$ indicates the \emph{event probability} \ie the likelihood of the ray stopping at the $i^{th}$ voxel in its path. The event probability, $q_p(i)$ is in turn instantiated using the probabilities $\{x_p^i\}$ - where $x_p^i$ denotes the occupancy probability of the $i^{th}$ voxel in the ray's path. See appendix for details.
\vspace{1mm}
\noindent \textbf{Sampling Occupancies along a Ray.}
The loss function as defined above is differentiable w.r.t shape $\bar{x}$, but not the camera parameters. This is because the quantity $\{x_p^i\}$ is not a differentiable function of the camera (since the ordering of voxels on a ray's path is a discrete function). Our insight is that instead of looking up \emph{voxels} on the ray's path, we can consider \emph{samples} along its path. Thus, our formulation is similar to that proposed by Tulsiani \etal~\cite{drcTulsiani17}, with the difference that the variable $\{x_p^i\}$ is redefined to correspond to the occupancy at the $i^{th}$ point sample along the ray.
Concretely, we sample points at a fixed set of $N = 80$ depth values $\{d_i | 1 \leq i \leq N\}$ along each ray. To determine $x^p_i$, we look at the 3D coordinate of the corresponding point (determined using camera parameters), and trilinearly sample the shape $\bar{x}$ to determine the occupancy at this point.
\vspace{-2mm}
\begin{gather}
l_i \equiv (\frac{u-u_0}{f_u}d_i, \frac{v-v_0}{f_v}d_i,d_i) \\
x^p_i = \mathcal{T}( \bar{x} , R \times (l_i + t)~)
\end{gather}
As the trilinear sampling function $\mathcal{T}$ is differentiable w.r.t its arguments, the sampled occupancy $x^p_i$ is differentiable w.r.t the shape $\bar{x}$ and the camera $C$. We note that Yan \etal~\cite{yan2016perspective} also used a similar sampling trick but their formulation is restricted to specifically using mask verification images and is additionally not leveraged for learning about pose.
\vspace{-1mm}
\subsection{Learning}
\seclabel{training}
\noindent \textbf{Training Objective.}
To train the shape and pose predictors, we leverage the view consistency loss previously defined (\secref{pdrc}) and train $f_s, f_p$ jointly to minimize $L_{data} = \sum\limits_{i=1}^{N}~\sum\limits_{u=1}^{N_i}~\sum\limits_{v=1}^{N_i}~ L(f_s(I^i_u), f_p(I^i_v); V^i_v)$. Therefore, the shape predicted using every image $f_s(I^i_u)$ should be consistent with \emph{all} available verification images of the same instance ($\{V^i_v\}$) when viewed from the corresponding (predicted) poses ($\{f_p(I^i_v)\}$). As detailed earlier, the independent prediction of shape and pose from different images ensures that the CNNs learn to infer the correct shape and pose under some emergent canonical frame.
\vspace{1mm}
\noindent \textbf{Architecture and Optimization Details.} We use a mini-batch size of 8 images $I_u^i$ for which shape is predicted. For each of these images, we randomly sample at least 2, and upto 3 if available, out of $N_i$, views $I_v^i$ of the same instance \ie the mini-batch size for the pose prediction CNN is between 16 and 24. We use extremely simple CNN architectures (depicted in \figref{overview}) corresponding to $f_s$ and $f_p$. Note that both these CNNs are initialized randomly (without any pre-training) and trained using ADAM~\cite{adam}.
\emph{Shape Prediction.} Our shape prediction CNN has an encoder-decoder structure similar to the one used by Tulsiani \etal~\cite{drcTulsiani17}. The input to the CNN is an RGB image of size $64 \times 64$ and the outputs are corresponding voxel occupancy probabilities for a $32 \times 32 \times 32$ grid.
\emph{Pose Prediction.} Our pose prediction CNN $f_p$ has a similar encoder to $f_s$, but outputs the predicted pose via fully connected layers. The rotation aspect of the pose is parametrized using two euler angles (azimuth, elevation) and the predicted translation $\in \mathbb{R}^3$. However, for some analysis experiments, we also assume that the object is at a known location w.r.t the camera and only predict the camera rotation. While in this work we assume known intrinsic parameters, the pose prediction CNN could in principle be extended to infer these.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{figures/resultsShapeMesh.png}
\vspace{-2mm}
\caption{\small Shape predictions on the validation set using a single RGB input image. We visualize the voxel occupancies by rendering the corresponding mesh (obtained via marching cubes) from a canonical pose. Left to Right: a) Input Image b) Ground-truth c) 3D Supervised Prediction d,e) Multi-view \& Pose Supervision (Mask, Depth) f,g) Mult-view w/o Rotation Supervision (Mask, Depth), and h,i) Mult-view w/o Rotation and Translation Supervision (Mask, Depth)}
\vspace{-2mm}
\figlabel{shapePred}
\end{figure*}
\vspace{-1mm}
\subsection{Overcoming Local Minima}
\vspace{-1mm}
We observed that our training is susceptible to local minima, in particular for the pose prediction CNN $f_p$. This is not too surprising since we have to learn both shape and pose from scratch, and erroneous estimates for one could confound the learning for the other, particularly in the in the initial stages We observe that the $f_p$ learns to predict only a small range of poses and \eg instead of predicting back-facing chairs, it confuses them with front-facing chairs. To avoid such local minima, we introduce two changes to the setup previously described.
\vspace{1mm}
\noindent \textbf{Incorporating a Pose Prior.}
We encourage the distribution of the predicted poses to be similar to a prior distribution (uniform azimuth $\in [0,360)$, elevation $\in [-20,40)$ degrees). We do so by adding an adversarial loss for the predictions of $f_p$ where the `real' samples are drawn from the prior distribution and `generated' samples are those predicted by $f_p$. We empirically show that our training is robust to the exact prior and that it can be different from the true distribution.
\vspace{1mm}
\noindent \textbf{Allowing Diverse Predictions.}
While the adversarial loss encourages diverse predictions, we also need some architectural changes to easily capture these. Instead of directly regressing to a single pose estimate in the last layer, we predict $N_p = 8$ estimates and additionally predict a probability distribution over these. We then sample a pose according to the predicted distribution. We use Reinforce~\cite{williams1992simple} to obtain gradients for the probability predictions.
|
\section{Introduction}
Performance models of high performance computer (HPC) systems enable
architects and administrators to gain insight into optimal system
configuration for their applications. Models also provide valuable
comparison against measured values that can be used to evaluate the
behavior of system components (for example network, storage nodes). A
common approach is to directly observe a mean value for the property of
interest (bandwidth or latency for example) and with domain expertise,
historical knowledge, and instantaneous system insight, judge if the
measured value is tolerable.
In the paper ``A parallel workload has extreme variability''
\cite{henwood16} it is claimed that the duration of time taken for a
sufficiently parallel workload to complete should follow the Generalized
Extreme Value (GEV) distribution shown in Eqn \ref{eqn:gevpdf}. A parallel
write makes a good choice for study as it is common, well understood, and a
good proxy for other common parallel operations, including map-reduce, and
sharded databases.
\begin{equation}
P_{\mu,\sigma,\xi}(x) =
\begin{cases}
\exp \left( - \left( 1+\xi \left(\frac{x-\mu}{\sigma}\right)\right)^{-1/\xi} \right) & \text{if} 1+\xi\frac{(x-\mu)}{\sigma} > 0 \\
\exp \left( - e^(\frac{x-\mu}{\sigma})\right) & \text{if} \xi = 0
\end{cases}
\label{eqn:gevpdf}
\end{equation}
Extreme value theory and significant observations of the GEV across many
scientific disciplines. The modern theory can be traced back to Fisher and
Tippett in the 1920s \cite{fishertippett1928}. A brief description of their
discovery and recent important observations of extreme phenomena are recorded
in \cite{henwood16} and references therein.
The argument presented for GEV in parallel environments is, in essence,
that the parallel task is executed by workers that are completely
independent with identical distribution (IID). The duration of the parallel
task ($T_g$) is measured when the worker ($S_n$) is complete:
\begin{equation}
T_g = \max\{S_1, S_2 ... S_m\}.
\label{eqn:max}
\end{equation}
These criteria broadly meet the preconditions for extreme value theory to be
applicable. And hence $T_g$ is approximated by $P(x)$ in Eqn
\ref{eqn:gevpdf}. The authors then present results for a carefully
constructed setup that attempts to minimize external perturbations and
focuses on solely on the period when the workload is running under IID
conditions.
This paper builds on work by studying a parallel workload in a production
HPC environment. The chosen workload is more typical to an actual workload
and the duration studied includes some serialized setup time. The
environment is a multi-user environment and no precautions have been taken
to limit any other simultaneous activity on the system. This paper argues
that the extreme distribution should follow a Fisher-Tippet Type I
(``Gumbel'') distribution. A framework for data analysis of extreme
statistics \cite{chapman01} is applied to distinguish between the different
Fisher-Tippett asymptotes.
\section{Experiment}
In changing from the controlled environment constructed in \cite{henwood16}
to a multi-user production HPC environment in this study, two notable
differences must be understood. Namely: traffic from external activates and
network topology.
\subsection{Traffic from external activities}
$T_g$ in Eqn \ref{eqn:max} is defined in \cite{henwood16} as the quiescent
performance of the file system under study. Let us simply assume that the
observed duration of a task on a system with congestion (from other jobs
accessing the shared resource) is the quiescent duration $T_q$ with
a delay of $T_s$ due to external congestion. i.e.
\begin{equation} T_g = T_s + T_q.
\label{basicvar}
\end{equation}
\begin{itemize}
\item $T_g$ is the observed write time of a fixed size transfer.
\item $T_s$ is the quiescent write time of a fixed size transfer.
\item $T_q$ is the time required to resolve resource contention on a multi-user system.
\end{itemize}
By formalizing the affect of external congestion we now make the following
claim: provided the congestion factor ($T_s$) is constant throughout
the period of experimental observation, it can be ignored for the purposes of identifying the distribution present.
\subsection{Network topology}
From the perspective of the task in question, the worker nodes (storage
nodes in this case) are all assumed to be the same distance away. Depending
on the configuration of the HPC custer, this may, or may not be the case. In
this study, we chosen a HPC cluster (Ranger) is chosen where this
requirement typically holds.
The experiment has two parts. The first part is designed to measure the
distribution of a typical write. The second part of the experiment applies
the framework described by Chapman, Rowlands, and Watkins \cite{chapman01}
to verify the presence of a member of the GEV family with $\xi = 0$ in
Eqn. \ref{eqn:gevpdf}. With $\xi=0$, the GEV distribution would
specialize to the Gumbel distribution \cite{gumbel1958,coles2001}.
\subsection{Environment}
Experimental data were collected during the summer of 2012 on the Ranger HPC
environment \cite{minyard08experiences} operated by the Texas Advanced
Computer Center at The University of Texas at Austin. Ranger is a Sun
Constellation Linux cluster with system components connected via a full-CLOS
InfiniBand interconnect. Eighty-two compute racks house the quad-socket
compute infrastructure. A high-speed parallel Lustre file systems is
provided running across 72 I/O servers.
The experimental runs were executed as normal user jobs during a normal
production operation. The parallel file system is a Lustre file system (v
1.8.5) on Red Hat$^*$ Enterprise Linux$^*$ 5, the client nodes are Red
Hat$^*$ Enterprise Linux$^*$ 5.
\subsection{Workload definition}
For the purposes of both identifying the underlying variability and testing
for particular member of the GEV family a typical write is defined as:
\begin{itemize}
\item Total file size is 16GB (half the available client memory.)
\item 16 storage nodes are used, a single 1GB stripe is written to each storage node.
\end{itemize}
A 1GB stripe size is chosen to ensure the effect of a serialized metadata
request (a restriction imposed by the Lustre 1.8 file system) is small and
the write time from the perspective of the client is dominated by variability
in the storage nodes. 16 nodes are chosen as a level of parallelism
sufficient to exhibit GEV behavior.
The tool {\tt{dd}} (v 5.2.1) is used to complete the write. A block size of
16GB is specified. A single directory is used for all writes.
{\tt{/dev/zero}} is used as the data for the write and is chosen to ensure
the client node does not stall waiting for data to write. During this phase,
the write is repeated 100 times to ensure sufficient fidelity of the
resulting distribution.
\begin{verbatim}
$LUSTRE = '/scratch/writetest' # directory
# on the parallel file system
$SSZ = 1024 # each stripe is 1GB stripes
$SCT = 16 # number of stripes to write is 16
$FILESIZE = $SSZ * $SCT
lfs setstripe -c $SCT -s ${SSZ}M $LUSTRE
for i in {0..100} do; \
sleep 30
time dd if=/dev/zero of=$LUSTRE/${i}.dat \
bs=${FILESIZE}M count=1 \
>> ~/results.txt;
done
\end{verbatim}
To ensure 16 storage nodes are written in parallel, a Lustre control code
({\tt{lfs setstripe}}) is used to define the striping on the destination
folder. The client node enters sleep for 30 seconds between writes to mimic a
computation load that may be present during a typical application. Once the
write is complete, {\tt{time}} diagnostics are written including the absolute
elapsed real time, which is the measure of interest $T_g$. Diagnostics are
written from the client to a separate file system. Note: the {\tt{dd}} tool on the Ranger environment is unable to write a block size of larger than 2GB.
If a blocksize is specified larger than this limit, 2GB is written and the task ends.
\subsection{Testing for a specific member of the GEV family}
Chapman, Rowlands, and Watkins \cite{chapman01} provide a method for
identifying the presence of values distributed with the Gumbel distribution
(where $\xi=0$ in Eqn. \ref{eqn:gevpdf}) by exploiting the fact the value of
the third central moment (skewness) of the Gumbel distribution is constant
($\simeq 1.14$). The tools and setup are identical to the previous experimental
phase. Instead of fixing on 16 storage nodes, the method requires a sweep of
increasing storage node counts.
The method is as follows: One hundred 1GB writes are completed against a single
storage node and the skewness calculated from the distribution of the elapsed
time measurements. The experiment is repeated, this time writing one GB each to
two storage nodes. Storage nodes and GBs are added until a limit is reached and
for each addition 100 writes are performed and the skewness is calculated for
the 100 experimental runs. In practice, {\tt{\$SCT}} is incremented in the code
example above.
As the number of storage nodes increases, the measured skewness of the
data should converge on the Gumbel skewness value. Chapman, Rowlands,
and Watkins \cite{chapman01} point out that the speed of the convergence
is dependent on the underlying distribution from which the maximum (or
minimum) is selected.
\section{Results}
Fig.\ref{fig:1nresults} shows the distribution of 100 typical writes
designed to deliver a load characteristic of large file count per directory
large I/O's checkpoint. The plot {\bf{(a)}} focuses attention to the state
of the fit around the distribution's head. The quantile plot {\bf{(b)}}
emphasises the tail of the fit. Points lying on the line indicate perfect
agreement between observation and model. Both the probability plot
{\bf{(a)}} and the quantile plot {\bf{(b)}} show good agreement with the
model. The return level plot {\bf{(c)}} also indicates good agreement with
the model as the points appear within the 95\% confidence level. {\bf{(d)}}
provides further evidence showing the fitted GEV density function is in good
agreement with the data histogram. All figures was generated using the R
language \cite{rmanual} using the ismev library \cite{ismev}.
\begin{figure}
\centering
\def\svgwidth{\columnwidth}
\input{16_gevfit.pdf_tex}
\caption{A model using the GEV distribution of a typical write against 16
storage nodes repeated 100 times.
A GEV distribution (Eqn. \ref{eqn:gevpdf}) is fitted to the data with a
location $\mu = 19.36$ ($\pm 0.20$), scale $ \sigma = 1.75$ ($\pm 0.14$),
and shape $\xi = -0.01$ ($\pm 0.07$). {\bf{(a)}} shows the cumulative
value of the observation against the model value. {\bf{(b)}} is the
observed quantity plotted against the modeled quantity. {\bf{(c)}} shows
the 95\% confidence interval of the value of $\xi$. {\bf{(d)}} presents
the observation histogram with the probability density plotted over the
data.}
\label{fig:1nresults}
\end{figure}
Fig.\ref{fig:addostsresults} skewness and mean of write times of the
individual distributions obtained as the number of storage nodes is
increased. A large spread of distributions is present as storage nodes
increase from one to 12. Between 12 and 27 the data skewness is near the
value for the Gumbel distribution (1.14). Beyond 27 storage nodes, the
skewness of the data appears to spread away from the Gumbel skewness.
\begin{figure}
\centering
\def\svgwidth{\columnwidth}
\input{100_runs_addosts.pdf_tex}
\caption{The skewness of the individual distributions obtained by increasing
the number of storage nodes in the system. For each run with a fixed number
of storage nodes, 100 typical writes were completed. The skewness of the
Gumbel distribution is shown as a horizontal line at 1.14.}
\label{fig:addostsresults}
\end{figure}
The maximum number of storage nodes included within this experiment was
limited to 29. Beyond this count, the client node exhausted local memory and
did not successfully complete the write.
\section{Discussion}
The results obtained from the Ranger system are encouraging. They were
obtained with a simple experiment, on a production system running many
simultaneous jobs.
Fig.\ref{fig:1nresults} shows general agreement with the GEV model. The
value of the shape parameter $\xi = -0.01 \pm 0.07$ is close to the shape
value for the Gumbel distribution ($\xi = 0$.) Observations within the tail
may appear to diverge slightly from the model, but remain well within the
margin of error (95\%).
If we consider Amhdal's Law:
\begin{equation}
S_{latency}(s) = {{1}\over{(1-p) + {{p}\over{s}}}}
\label{eqn:amdahl}
\end{equation}
This equation give the theoretical speedup in latency of a parallel task
given the proportion of the task that executes in parallel. In the previous
study of GEV in parallel environments, the proportion of the application that
executed in parallel was made as close to 100\% ($p = 1$) as possible. With
the use of {\tt{dd}} to dispatch the parallel workload in study, there is a
non-zero period of the task spent performing the serial metadata operations
before the parallel work begins. This suggest that GEV varibility is also
present for workloads where $p \approx 1$.
Fig.\ref{fig:addostsresults} indicates a rapid convergence towards the
Gumbel distribution from one to 12 storage nodes. Between 12 and 27 storage
nodes the skewness of the distribution is close to the Gumbel skewness.
Beyond that value, the skewness is less like Gumbel.
One possible explanation for the trend away from Gumbel skewness with higher
storage node counts is congestion from external traffic. As the storage node
count increases on a multi-user system, it is much less likely the
assumption to reduce $T_q \rightarrow 0$ in Eqn. \ref{basicvar} will hold.
i.e., while there are a large number of storage nodes available, as the
client writes to an increasing number, the probability another job will also
be requesting the same storage node resources increases. If a storage node
is under higher load it may service requests with a distribution that is not
identical to an unloaded storage node.
Fig.\ref{fig:gevmodel} presents the cumulative distribution of the GEV model.
The model uses the parameters extracted from the fit in Fig.
\ref{fig:1nresults}. A graphical technique for classifying the likelihood that
a particular write will take a given duration is shown with dotted lines. In
this case, the example write is observed to take 26 seconds. A vertical line is
drawn from the measure of 26 on the abscissa to intersect the cumulative
distribution function. From this point of intersection, horizontal line is
drawn to the ordinate intersecting at a probability of about 98\%.
\begin{figure}
\centering
\def\svgwidth{\columnwidth}
\input{gevModel.pdf_tex}
\caption{A model of a write on a parallel file system to 16 storage nodes.
The model is a cumulative generalised extreme distribution function with
values of location $\mu = 19.36$ ($\pm 0.20$), scale $ \sigma = 1.75$
($\pm 0.14$), and shape $\xi = -0.01$ ($\pm 0.07$). The probability a
write will be completed in 26 seconds or less is illustrated graphically.
The probable value is about 98\%.}
\label{fig:gevmodel}
\end{figure}
The result in the example presented in Fig.\ref{fig:gevmodel} is that the
probability a write will take 26 seconds or less is about 98\%. In
contrast, writes that take more than 26 seconds have a probability of about
2\% - or occur approximately once for every 50 writes. Given this, if a
client observes more than one write in 50 taking longer than 26 seconds an
alert could be issued.
\section{Conclusions}
Understanding and predicting the behavior of high performance computers is a
challenging task. Even after they are constructed and operating the
measurement, benchmarking, and monitoring of large computers is still an
art, practiced by a small number of experts. This paper presents extreme
value theory as a new tool to enhance diagnostics. Taking the required
measurements to reveal the I/O distribution for a given machine is shown
here to be simple and quick to complete. Performing such measurements on a
quiet system (where $T_q = 0$) enables the baseline variation of the
environment to be captured. The model presented in this paper can be
calibrated for a given environment and periodic observations can reassure a
systems operator that the machine is performing optimally. Furthermore,
depending on the operations policy of the environment an accurate model for
variability may be valuable in pricing service level agreements.
It is interesting to consider a practical outcome of the GEV result for
increasing parallel-ism in data transfers: the case where the throughput of the
client interface is the narrowest (or equally narrow) bottle neck in the
system. In this case, a parallel transfer is constricted at the point of the
client. Adding parallelism will have the effect of reducing performance, and
increasing variability. Hence, one might conclude that, if possible, a single
storage target should be specified for all data transfers.
As high performance computing continues to develop, new libraries become
available to simplify interfacing with data objects. For example, the
{\tt{t3pio}} library provides automatic configuration for MPI applications
that use HDF5. Such a library could introduce auto-tuning behavior to avoid
congestion from competing file system tasks if the assumption that a write
time will have a GEV distribution. A single client node could be calibrated
for ideal parallel behavior and measure deviations from this behavior as
values that are unlikely according to the GEV model. A more sophisticated
client could choose to modify file layout or cross-reference with other
anomalous results to identify congestion areas within the system.
Predicting the behavior of a large machine while it is still in design is
valuable if accurate. The GEV model for I/O enables a designer to sweep away
the complexity of modeling the file system as a series of components. By
simulating the parallel file system as a simple GEV (or Gumbel) source it may
be possible to arrive at an accurate file system performance metric. It is
conceivable that by starting with the specified performance of a single disk
and successively applying the GEV model one can now arrive at the probable
behavior of the whole storage system.
\section{Acknowledgements}
RH is grateful to John Hammond, Doug James and Andreas Dilger for sharing
their expertise on the subtleties of a Lustre file system on Ranger. Linda
Berbernes and Victor Eijkhout provided valuable input to enhance the text and
content. SCC acknowledges a Fulbright Lloyd's of London Scholarship and
support from the AFOSR on grant FA9550-17-1-0054.
\bibliographystyle{abbrv}
|
\section{Introduction}
Blazars are a subclass of Active Galactic Nuclei (AGN) in which the relativistic jet presents a small viewing angle towards the observer and thus where relativistic effects on the observed emission are more extreme. Conventionally, blazars are subdivided in BL Lac objects and FSRQs depending on the characteristic of their optical spectrum: while BL Lac objects are dominated by the featureless continuum emission from the jet, FSRQs typically show wide optical emission lines.
The blazar S4 0954+65 hosts a black hole of mass $M_{\mathrm{BH}}\sim3.3\times 10^8 M_{\odot}$, estimated from the width of the H$_\alpha$ line \citep{blackholemasses}. The detection of the H$_\alpha$ line is not confirmed by \cite{landoni2015} (see the discussion on the redshift determination) so that the mass estimation cannot be confirmed either.
This blazar presents strong variability in the optical band, already well studied by \citet[][]{wagner1990} and by \citet{morozova2014}. Intra night variability has been found both in optical and radio wavelengths \citep{wagner1993}.
The optical high brightness state of February 2015, presented here, is however exceptional for the object, with a brightening of more than 3 magnitudes in the R-band with respect to the average monitored state\footnote{\url{http://users.utu.fi/kani/1m/S4\_0954+65.html}}. This not only spurred many alerts in the community \citep[see ATel \#6996, \#7001, \#7057,~\#7083, \#7093;][]{atel6996,atel7001,atel7057,atel7083,atel7093}, but also the first and only detection of the object at very high energies (VHE, E$\gtrsim$100 GeV), thanks to observations by the MAGIC Telescopes. This detection by MAGIC and the MWL data collected alongside it are the focus of the present work.
The source GRO J0957+65, detected with the EGRET telescope on board the \textit{Compton Gamma-Ray Observatory}, has been associated through optical and radio observations with S4 0954+65 by \citet{egret95}. S4 0954+65 has been afterwards always included in the released catalogs of sources detected by the Large Area Telescope (LAT) instrument on board the \textit{Fermi} satellite \citep{1FGL,2FGL,1FHL,3FGL,2FHL,3FHL}, with the exclusion of the bright source list released after the first 3 months of \textit{Fermi}-LAT~ data integration.
The classification of the object, based on the available literature, is still unclear. In most of the ATels mentioned above S4 0954+65 is referenced as a FSRQ, but in most of the literature this is classified as a BL Lac object due to the small equivalent width of the emission lines in its spectrum \citep[see, e.g.][]{stickel91}. \citet{sambruna1996} classified the SED of S4 0954+65 as ``FSRQ-like'', in a sample limited to the sources with a detection from EGRET data. It indeed presents a flatter spectral index than most BL Lac objects, in both X-ray and $\gamma$-ray\ bands \citep[see][ and references therein]{raiteri}.
Among BL Lac objects, a further phenomenological subdivision can be made based on the frequency of the synchrotron peak, ranging from optical to X-ray frequency and identifying the classes of low-, intermediate- or high-peaked BL Lac object (LBL, IBL, HBL respectively). \citet{ghisellini2011} classified this object as a LBL based on the SED.
When including the kinematic features from the radio jet in the classification templates, \citet{hervetetal2016} classify this as their kinematic class II, mostly composed of FSRQ.
S4 0954+65 can thus be interpreted as a transitional object between FSRQ and classical BL Lac objects.
The most numerous extragalactic sources detected at VHE from Imaging Air Cherenkov Telescopes (IACTs), presently, belong to the HBL class. Therefore the VHE detection of an object such as S4 0954+65 provides a rare opportunity to study VHE emission conceivably produced in a different kind of environment. Indeed, while emission in HBL can mostly be satisfactorily modeled taking into account only processes in a compact feature in the jet, for FSRQs the inclusion of the interactions of such a feature with the surrounding ambient becomes of greater importance \citep[see e.g. ][]{tavecchiogamma}. The structure of the broadband SED collected here will also be put in context with other common characteristics of a FSRQ classification, such as intrinsic brightness, peak of the synchrotron component and Compton dominance.
Also the question of S4 0954+65 redshift is still not settled, as claims of line detection in the optical spectrum are not always confirmed. The redshift of the source was first determined at $z$=0.368 by the identification of
lines by \citet{law86,law96}. \citet{stickel93} obtained, from different measurements, the same redshift estimate based on line identification. None of these lines were confirmed by the observations reported in \citet{landoni2015}, who instead pose a lower limit of z$\geq0.45$. The latter results were obtained with a superior resolution spectra. At the time of the observation the magnitude in R-band of the object was 15.5, while it is known from variability studies that it could be even 2 magnitudes lower.
In the following we will adopt the redshift $z$=0.368.
The outline of this paper is as follows. In Section \ref{sec:magicdata}, we will present the MAGIC telescopes and the relative data set on S4 0954+65. Section \ref{sec:mwlcoverage} reviews all the MWL data that were collected during this exceptional burst, whereas Section \ref{sec:lightcurveandsed} discusses the implication of this burst for the source state and inner jet structure. Additional information on the MAGIC data analysis, the parameters derived from the VLBA data and the full dataset for Swift-XRT X-ray data will be found in Appendix \ref{app:magicdata} and \ref{app:vlbadata}, respectively.
\section{MAGIC Observations}\label{sec:magicdata}
The MAGIC telescopes are an array of two IACTs located in the Island of La Palma (Spain) at an altitude of $\sim$ 2200 m asl. The system is sensitive down to an energy threshold of $E\sim50$ GeV \citep{magicperformancepaper} for low zenith angle observations. This is of particular relevance for the monitoring of variable sources and of those that tend to exhibit a steep spectrum at VHE.
The full data have been analyzed using the standard MAGIC analysis chain and the MAGIC Standard Analysis Software \citep[MARS,][]{standardMARS, magicperformancepaper}.
The MAGIC collaboration supports a program of Targets of Opportunity (ToO), triggered by MWL monitoring. The ToO program was activated for observations of S4 0954+65 at the end of January 2015 after the first hints of enhanced optical state (triggered by the Tuorla monitoring in R-band, see Section \ref{subsec:optical}).
We observed the source with the MAGIC telescopes for 2 nights (MJD 57049-57050, 2015 January 27 and 28), for a total of 1 hour high-quality dark time data, but obtained no detection.
We resumed the ToO observations in February after the Tuorla monitoring revealed a very exceptional flux state, later confirmed by other monitoring programs (see Section \ref{subsec:optical}). We obtained a detection at a significance of $\sim7.4\sigma$ from observations during 2015 February 14 \citep[MJD 57067, ATel \#7080][]{atelmagic}.
We continued observing S4 0954+65, barring adverse atmospheric conditions, until full moon days when standard MAGIC observations are not possible due to the elevated level of background light (last day of observation, with already large moonlight contamination, on 2015 March 1, MJD 57082). A detailed breakdown of the observation conditions and relative results can be found in Appendix \ref{app:magicdata}.
The total excess from the dark-time data is consistent with a point source emission (see Fig. \ref{fig:odie_le_fullset}). No other significant emission is found in the field of view apart from the one coincident with S4 0954+65 at the center.
\begin{figure}
\centering
\includegraphics[width=\hsize]{figure1.eps}
\caption{Distribution of the squared angular distance ($\theta^2$) between the reconstructed event direction and the nominal source direction. The filled histogram is the background estimation, obtained from sky regions within the field of view with similar detector acceptance. We show only data taken in dark condition (condition 1, see Appendix \ref{app:magicdata}). The standard MAGIC low energy (LE) cuts are applied to the data (see Appendix \ref{app:magicdata} and Table \ref{table:magicdata}). The vertical line corresponds to the optimal cut ($\theta^2=0.02\mathrm{\ deg}^2$) for point source analysis in LE cuts, used to derive significance values.}
\label{fig:odie_le_fullset}
\end{figure}
The SED points presented in Section \ref{sec:lightcurveandsed} below are derived for the day of the flare (MJD 57067, 2015 February 14), using only data taken in dark conditions (that allow for the lowest threshold and lowest systematic uncertainty, Appendix \ref{app:magicdata}). We follow the standard MAGIC unfolding procedure \citep{unfoldingmagic} to obtain the intrinsic spectrum.
The $\gamma$-ray\ emission from sources at high redshift is absorbed via photon-photon pair production on photons from the extragalactic background light \citep[EBL, see e.g.][]{dominguez11,finke2010}. S4 0954+65 redshift is assumed to be $z=0.368$.
The spectral shape of the intrinsic emission, i.e. after the correction for the EBL absorption, can be fitted with a simple power law:
\begin{equation}
\frac{dN}{dE} = N_0 \left(\frac{E}{E_0}\right)^{-\Gamma}
\end{equation}
with normalization $N_0 = \left(13.8\pm2.1^{\rm stat}\pm 1.5^{\rm sys}\right)\times 10^{-10} \rm{\ TeV}^{-1} {\rm cm}^{-2} {\rm s}^{-1}$ at $E_0=0.15$ TeV and spectral index $\Gamma=3.98\pm 0.67^{\rm stat}\pm 0.15^{\rm sys}$. The quoted systematic uncertainties are derived from the standard evaluation in MAGIC data presented by \citet{magicperformancepaper}. Note that the calculated systematic uncertainty on $N_0$ does not contain the uncertainty on the energy scale, that is about 15\%. The unfolded MAGIC spectrum is shown in Fig.~\ref{fig:magic_spec}.
The unfolded observed spectrum, i.e. without correcting for the EBL absorption, can be described also by a simple power law with $N_0 = \left(9.9\pm1.5^{\rm stat}\pm 1.1^{\rm sys}\right)\times 10^{-10} \rm{\ TeV}^{-1} {\rm cm}^{-2} {\rm s}^{-1}$ at $E_0=0.15$ TeV and spectral index $\Gamma=4.58\pm 0.66^{\rm stat}\pm 0.15^{\rm sys}$.
\begin{figure}[!htb]
{\includegraphics[width=\linewidth]{figure2.eps}}
\caption{Spectrum for the VHE MAGIC detection. MAGIC data are for flare night only (2015 February 14, MJD 57067.14). Violet filled circles are for the unfolded observed points, while open circles are de-absorbed for EBL absorption \citep[EBL model by][]{dominguez11}. The solid line is the fit for the observed points and the dashed line is the fit for the de-absorbed ones, with details in the text.}
\label{fig:magic_spec}
\end{figure}
\section{The Multiwavelength coverage}\label{sec:mwlcoverage}
All the data presented in this section are collected to produce the light curves and SED, whose interpretation is later presented in Section \ref{sec:lightcurveandsed}.
\begin{figure*}[p]
\resizebox{\hsize}{!}
{\includegraphics[]{figure3.eps}}
\caption{MWL light curves and polarization evolution of S4 0954+65 ranging from MJD 56970 (2014 November 9) to MJD 57200 (2015 June 27). The energy range of each panel and the corresponding instrument can be found in the legend. Please refer to the text for details on the data taking and reduction for each instrument. }
\label{fig:mwl_lc}
\end{figure*}
\subsection{Fermi-LAT}
The LAT on board the \textit{Fermi} satellite scans the entire sky every 3 hours. From the data of the first 4 years of operation, S4 0954+65 was detected with an average significance of $27.2\sigma$ in the energy range from 100 MeV to 300 GeV as reported in the 3FGL catalog \citep{3FGL}. A dedicated analysis from MJD 56952 (2014 October 22) to MJD 57208 (2015 July 05) is presented in this work. We selected Pass 8 source class events within a 10$^\circ$ circular region centered on the position of S4 0954+65, in the energy range 0.1-500 GeV. The spectral analysis was performed through an unbinned likelihood fit, using the ScienceTools software package version v11-05-00 along with the instrument response functions P8R2\_SOURCE\_V6. The model of the likelihood fit includes a Galactic diffuse emission model and an isotropic component\footnote{Model available at \url{https://fermi.gsfc.nasa.gov/ssc/data/access/lat/BackgroundModels.html}.}. In addition, we included the sources in the 3FGL catalog within a 20$^\circ$ circular region centered on S4 0954+65. The spectral indexes and fluxes of the 3FGL sources located within a region of 10$^\circ$ from S4 0954+65 were left free to vary, while the sources in the region from 10$^\circ$ to 20$^\circ$ were fixed to their catalog values. The results were obtained from two iterations of maximum-likelihood analysis, after the sources with a test statistics \citep[][]{egretlikelihood} TS$<$10 were removed. The strongest source located beyond 10$^\circ$ from S4 0954+65 is at an angular distance of 10.8$^\circ$. This source has a variability index of 42.4 in the 3FGL catalog, that allows us to treat it as a non-variable source and thus to fix its spectral index and flux to the values reported in the 3FGL catalog.
The light curve was calculated in day timescale bins, modeling the source with a single power-law spectrum (as it is also described in the 3FGL). Both the flux and spectral index of S4 0954+65 were left free during the likelihood fits, while the rest of the point sources were fixed and only the diffuse Galactic and isotropic models were allowed to vary. In case of TS<4, an upper limit on the flux was calculated fixing the spectral index to 2.38 as given in the 3FGL catalog. The results are shown in Fig. \ref{fig:mwl_lc}. The figure shows also the light curve calculated in 15-days bin as comparison. The light curve was obtained with the same procedure described above for the 1-day binning. During the HE flare in November 2014 \citep[MJD 56976, ATel \#6709;][]{atel6709} the LAT spectral index is compatible with its 3FGL value of $2.38 \pm 0.04$, averaged from 4 years of data. Moreover, the visibility of the source by MAGIC was at an unfavorable zenith angle of 60$^\circ$ (implying a high energy threshold). Therefore, no ToO observation was activated with MAGIC for this flare. MAGIC observations were activated later on during the strong flare on February 2015 when the LAT detected a hardening of the spectrum as shown previously by Tanaka et al. (2016) where the LAT analysis using Pass 7 reprocessed data is presented.
The spectral analysis for the MWL SED corresponds to 1-day integration centered in the MAGIC observation (MJD 57067.14, 2015 February 14). From a first likelihood fit we found the best spectral fit was a power-law spectral index of $1.87 \pm 0.09$ (significantly harder than its average 3FGL value) and was fixed in the model for the spectral points calculation. Moreover, all the sources included in the model except the diffuse Galactic and isotropic models were also fixed. The source was detected during this period with a TS of 379.7. A curved spectral model is not significantly favored in this day (TS for a log parabola fit is TS$_\textrm{LP}$=380.10 to be compared with a simple power law fit with TS$_\textrm{PWL}$=379.74).
\subsection{\textit{Swift} dataset}
The 22 multi epochs event-list obtained by the X-ray Telescope \citep[XRT,][]{swiftXRT} on board the \textit{Swift} satellite in the period of 2014 November 17 (MJD 56978.96395) to 2015 March 11 (MJD 57092.26632) with a total exposure time of $\sim$11.12 hours were processed using the procedure described by \citet{vandad2017}. All these observations had been performed in photon counting (PC) mode, with an average integration time of 1.8 ks each. The equivalent Galactic hydrogen column density is fixed to the value of $n_H = 5.17 \times 10^{20} \textrm{[cm}^{-2} \textrm{]}$ \citep{LABHI}.
The average integral photon X-ray flux (0.3-10 keV) in this period is $1.64 \times 10^{-11}$ $\rm erg/cm^2/s$. The X-ray flux is peaking at MJD 57070.76523 with $F_{(0.3-10 \mathrm{keV})} = 3.18 \times 10^{-11}$ $\rm erg cm^{-2} s^{-1}$ which is a factor of about 2 higher than the average flux of the analyzed period. The average flux outside the flare period (2006-2015) is $F_{(0.3-10 keV)} = 4.3 \times 10^{-12}$ $\rm erg cm^{-2} s^{-1}$, that we derived from a sample of XRT data comprising 25 X-ray exposures in the XRT database, not including the 22 multi epochs event-list described above. This indicates that the source was clearly in its X-ray high state during the VHE $\gamma$-ray detection. The X-ray spectral index during the analyzed period varies between $1.15\pm0.06 \le \Gamma_X \le 1.82\pm0.1$. It is notable that the softest spectral index was obtained a night prior to the VHE $\gamma$-ray flare while the spectra starts to harden after 2015 February 14 and reach its historical hardest spectra 10 days after the VHE $\gamma$-ray flare. The X-ray spectra on the night before and after the VHE $\gamma$-ray flare can be well described with a power-law with spectral index of $\Gamma_{\rm X, Feb. 13} = 1.82 \pm 0.05$ ($ \chi^2/$d.o.f.=1.024/41) and $\Gamma_{\rm X, Feb. 15} = 1.49 \pm 0.07$ (1.025/24 $ \chi^2/$d.o.f.) respectively.
The full dataset analysis is given in Appendix \ref{app:swiftxrt}.
The \textit{Swift} satellite hosts an additional instrument, the Ultraviolet/Optical Telescope \citep[UVOT, ][]{uvotmainref}. The data taken during the period of interest for this work have already been presented by \citet{tanaka2016}. They follow the behavior of the optical light curve that we will present next. Therefore they are not reproduced again nor shown in Fig. \ref{fig:mwl_lc}. The UVOT bands are however important for the SED modeling presented in Section \ref{sec:lightcurveandsed} and will therefore be included there for MJD 57067 (2015 February 14, day of the VHE detection). The dataset presented by \citet{tanaka2016} suffers from an incorrect exposure calculation by a factor of 2, related to the deadtime correction, and thus a lower reconstructed flux. We therefore have performed a re-analysis here for the two exposures taken with UVOT on MJD 57066.76. Data reduction has been done on all the available filters ($v,\ b,\ u,\ w1,\ m2,\ w2$), following the standard UVOT data analysis prescriptions\footnote{\url{https://swift.gsfc.nasa.gov/analysis/}}. We present both exposures separately, due to the high variability in this night (e.g for the V-band there is a variation of $\sim$0.3 magnitudes in $\sim$1.5 hours).
\subsection{The optical domain}\label{subsec:optical}
Optical data were collected with: 35cm KVA telescope (La Palma Island, Spain) used in the Tuorla monitoring program; 1.8~m Perkins telescope of Lowell Observatory (Flagstaff, Arizona); 70 cm telescope AZT-8 at the Crimean Astrophysical Observatory (Nauchny, Russia); 40 cm telescope LX-200 of St. Petersburg State University (St. Petersburg, Russia); IAC80/Camelot at the Teide Observatory (Tenerife, Spain). The data analysis from KVA was
performed with the semi-automatic pipeline using the standard analysis procedures (Nilsson et al. in prep). The differential photometry was performed using the comparison star magnitudes from \citet{villata}. For the Perkins telescope see \citet{jorstad2010} and references therein. The details of observations and data reductions with AZT-8 and LX-200 are given by \citet{larionov2008}. IAC80/Camelot data were automatically processed by the pipeline Redcam and calibrated astrometrically using XParallax, both available at the telescope. Instrumental magnitudes for IAC80/Camelot data were extracted using Sextractor \citep{bertin96} and calibration of the source magnitude was obtained with respect to the reference stars provided by \citet{raiteri}.
All the above mentioned telescopes provide R-band photometry. We have applied the calibration of \citet{mead1990} for all optical measurements to transform magnitudes into flux densities, and dereddened the flux according to the absorption by \citet{schlafly}. The host galaxy is not detected for this object.
From the Perkins, AZT-8+ST7 and LX-200 telescopes we collect also polarization information. In Fig.~\ref{fig:mwl_lc} we show the optical photometry data and time evolution of the fractional linear polarization and the electrical vector position angle (EVPA) in R-band. The EVPA measurements have been arranged such to minimize the impact of the $\pm180^\circ$ ambiguity, i.e. adding or subtracting $180^\circ$ whenever two subsequent measurements differ by more than $90^\circ$.
In the same timeframe of the VHE detection and the optical flare, a substantial change in the optical EVPA can be identified (see Fig. \ref{fig:mwl_lc}). The EVPA rotation starts just before the optical and VHE flare and reaches a total change of roughly 100$^\circ$.
The optical flare in February 2015 is a factor of about 3 larger in flux than the 2011 flare \citep[see][]{morozova2014}, that was already exceptional and concurrent with a series of $\gamma$-ray\ flares evident in \textit{Fermi}-LAT~ data. During the most extreme flare in 2011, the EVPA rotated by about 300$^\circ$.
\begin{figure*}[!htb]
\resizebox{\hsize}{!}
{\includegraphics[]{figure4.eps}}
\caption{ A sequence of total (contours) and polarized (segments) intensity images of S4 0954+658 at 43 GHz, convolved with a beam of 0.24$\times$0.15 mas$^2$ at PA=-10$^\circ$. The global total intensity peak is 1606 mJy/ beam and the global polarized intensity peak is 104 mJy/ beam; black line segments within each image show the direction of polarization and their length is proportional to the polarized intensity. The black horizontal line indicates the position of the core, A0, and grey, blue, and red circles show the locations of knots K14a, K14b, and K15, respectively. The size of the circles is proportional to the estimated average size in each epoch. The detailed characteristics of the knots can be found in Table \ref{table:knots} and in Table \ref{table:knots_LC}. }
\label{fig:43ghzimages}
\end{figure*}
\begin{figure}[!htb]
\resizebox{\hsize}{!}
{\includegraphics[]{figure5.eps}}
\caption{Apparent distance from the radio core A0 of the new emerging knots, K14a,b and K15, as a function of time. The images from which the apparent distances are calculated can be found in Fig. \ref{fig:43ghzimages}. It can be noted that the K15 knot presents the highest apparent speed.}
\label{fig:43ghzknots_characteristics}
\end{figure}
\begin{table*}[!htb]
\caption{Characteristics of the new radio knots observed from the jet of S4 0954+65. The evolution of parameters with the monitoring snapshots can be found in Appendix \ref{app:vlbadata}. }
\label{table:knots}
\centering
\begin{tabular}{c c c c c c c c}
\hline\hline
Knot & Average Flux & Maximum Flux & Average PA & Average Size & Proper motion & Apparent Speed & Time of Ejection \\
& mJy & mJy & deg ($^\circ$) & (FWHM) mas & mas/yr & c & MJD \\
\hline
K14a & $120\pm7$ & $286\pm10$ & $-17.6\pm2.4$ & $0.15\pm0.07$ & $0.55\pm0.04$ & $12.49\pm0.91$ & $56708\pm26$ \\
K14b & $76\pm25$ & $118\pm6$ & $-16.2\pm2.6$ & $0.07\pm0.06$ & $0.59\pm0.04$ & $13.47\pm0.86$ & $56891\pm15$ \\
K15 &$109\pm14$ & $121\pm5$ & $-5.9\pm1.9$ & $0.05\pm0.01$ & $1.11\pm0.08$ & $25.27\pm1.20$ &$57081\pm18$ \\
\hline
\end{tabular}
\end{table*}
\subsection{The radio and millimeter ranges}
S4 0954+65 was monitored at 3.5 mm (86 GHz) and 1.3 mm (229 GHz) wavelengths from the IRAM 30 m Millimeter Radiotelescope under the POLAMI (Polarimetric Monitoring of AGN at Millimeter Wavelengths)\footnote{\url{https://polami.iaa.es}} program. The program monitors the four Stokes parameters of a sample of the brightest ~40 northern blazars with a cadence better than a month
\citep[see][]{agudo2018a,agudo2018b,thum2018}. Results from the observations are presented in Fig. \ref{fig:mwl_lc}. The data reduction, calibration, and flagging procedures were described in detail by Agudo et al. 2017a, submitted \citep[see also][]{agudo2010,agudo2014}. Fig. \ref{fig:mwl_lc} includes also the 1.3mm flux density data that were obtained at the Submillimeter Array (SMA) located in Hawaii. S4 0954+65 is included in an ongoing monitoring program at the SMA to determine the fluxes of compact extragalactic radio sources that can be used as calibrators at mm wavelengths \citep{gurwell2007}. Observations of available potential calibrators are from time to time observed for 3 to 5 minutes, and the measured source signal strength calibrated against known standards, typically solar system objects (Titan, Uranus, Neptune, or Callisto). Data from this program are updated regularly and are available at the SMA website\footnote{\url{http://sma1.sma.hawaii.edu/callist/callist.html}}. The largest flux in the considered period is at MJD 57072-57076, showing an increase of the flux at 1 mm and 3 mm wavelengths.
It is to be noted however the lack of exactly simultaneous data to the MAGIC peak detection (MJD 57067).
S4 0954+65 is monitored monthly by the Boston University (BU) group with the Very Long Baseline Array (VLBA) at 43 GHz within a sample of bright $\gamma$-ray\ blazars through the VLBA-BU-BLAZAR program\footnote{\url{http://www.bu.edu/blazars/VLBAproject.html}}. The VLBA data are calibrated and imaged in the same manner as discussed by \citet{jorstad2005,jorstad2017}.
The VLBA imaging monitoring program allows us to study the kinematics of the inner jet at pc scale. The inner jet has been monitored also for months after the VHE flare (see Fig. \ref{fig:43ghzimages}). In addition to the stable core at mm wavelengths (dubbed A0, see Fig. \ref{fig:43ghzimages}) it was possible to identify the emergence of three new knots whose characteristics are tabulated in Table \ref{table:knots}. The nomenclature of the knots follows in sequential order from the beginning of the VLBA monitoring program. Previous knots characteristics can be found in \citet{morozova2014}.
Of particular interest is knot K15, which is very compact, with a FWHM average size of $0.05\pm0.01$ mas and presents the largest apparent speed of ($25.27\pm1.20$)c, cf Fig. \ref{fig:43ghzknots_characteristics}. The zero-epoch separation of this knot is consistent with the VHE flare considering its 18-day uncertainty. The intensity of the core is increasing in the epoch of MJD 57067 observation, but no significant change in the core polarization can be appreciated. The detailed information on the time evolution of the radio knot can be found in Table \ref{table:knots_LC}, while the polatization evolution details are shown in Table \ref{table:CorePol_LC}. During November 2014, while the source was high in the HE band as observed by \textit{Fermi}-LAT~ but without optical enhancement, no new knot appears. The zero epoch-separation from the core of knots K14a and K14b are not coincident with the high state in \textit{Fermi}-LAT~ data of November 2014, but happen months before. We analyzed \textit{Fermi}-LAT~ data in the period included within the error band for K14a and K14b zero epoch-separation and found no particular enhancements.
To be noted is also the position angle of K15 with respect to the core, (PA=$-5.9^\circ\pm1.9^\circ$). This is different than the values reconstructed from previous knots, ranging from roughly PA=$-15^\circ$ to PA=$-25^\circ$ in \citet{morozova2014}, that are in turn consistent with the values for K14a/b. The mean jet direction is at PA$\simeq -20^\circ$. A difference in PA and in apparent speed could be simply related to a small difference in the angle to the observer. However, the highest apparent speed can be used to estimate the Doppler factor, considering the upper limit to largest possible viewing angle $\theta_\textrm{obs}<\textrm{arcsin}(1/\beta_\textrm{app})$ and ultimately leading to $\delta_\textrm{app}\sim\beta_\textrm{app}$. Applying this to the above mentioned knots (averaging the apparent speed to $\beta_\textrm{app}\sim13$c for K14a/b): $\theta_\textrm{obs,K15}<2.3^\circ$ and $\delta_\textrm{app,K15}\sim25$; $\theta_\textrm{obs,K14}<4.4^\circ$ and $\delta_\textrm{app,K14}\sim13$.
The 37 GHz observations were made with the 13.7 m diameter telescope at Aalto University Mets\"ahovi Radio Observatory. A detailed description of the data reduction and analysis is given by \citet{metsa}. The error estimate in the flux density includes contributions from the measurement RMS and the uncertainty of the absolute calibration. The S4 0954+65 observations were done as part of the regular monitoring program and the GASP-WEBT campaign. There are no strictly simultaneous 37GHz data to the MAGIC detection, however an increase in flux can be seen when comparing observation taken one day before (2015 February 13, MJD 57066.15, $F_\nu=1.27\pm0.07$ Jy) and one day after the MAGIC detection (2015 February 15, MJD 57068.15, $F_\nu=1.65\pm0.09$ Jy).
The OVRO 40 m uses off-axis dual-beam optics and a cryogenic pseudo-correlation receiver with a
15.0 GHz center frequency and 3 GHz bandwidth.
Calibration is achieved using a temperature-stable diode noise source to remove receiver gain
drifts and the flux density scale is derived from observations of
3C~286 assuming the \citet{1977A&A....61...99B} value of 3.44 Jy at
15.0 GHz. The systematic uncertainty of about 5\% in the flux density
scale is not included in the error bars. Complete details of the
reduction and calibration procedure are found in \citet{ovro}.
The long-term monitoring program at OVRO (Owens Valley Radio Observatory) monitors the variability of this source at 15GHz over a longer time than what shown here. While it is obvious
that the source was variable also during February 2015, it is not an exceptionally bright flux state of the source in the radio band. From a decade long monitoring, the source shows brighter levels (highest at $F_\textrm{15 GHz} =2.53$ Jy) and fainter levels (lowest at $F_\textrm{15 GHz} =0.85$ Jy).
Both 15 GHz and 37 GHz data seem to be in agreement with the behavior seen from mm wavelength data. Again note the lack of strictly simultaneous data to the MAGIC peak detection (MJD 57067). emission.
\section{Discussion}\label{sec:lightcurveandsed}
The coverage of flaring states at VHE is helpful to understand the jet dynamics. We present a discussion of the SED for the day of the flare (2015 February 14). We do not attempt a SED modeling for other days, for which the MAGIC data would provide only non-constraining upper limits to emission at VHE.
The day of the VHE detection is instead put in context with a longer time span behavior in the MWL dataset. However the VHE sampling of the state is too scarce to attempt a numerical correlation study of the light curves.
\subsection{Light curve phenomenology}
The MWL light curves of the source for all the instruments involved in the present work are reported in Fig. \ref{fig:mwl_lc}, and cover a time range of 7 months, from MJD 56970 (2014 November 19) to MJD 57200 (2015 June 27). The panels of Fig. \ref{fig:mwl_lc}, in order of decreasing energy, show in the top panel the MAGIC detection at VHE, while the radio data collected by OVRO, POLAMI and the other instruments in the radio band are shown in the bottom one. The red region indicates the time window where the knot K15 was ejected in the VLBA analysis, as reported in Table \ref{table:knots}: a time range of 36 days centered in MJD 57081 (2015 February 28). The VHE detection and the enhanced activity in the other bands are found inside the K15 ejection time window, making this event important for the understanding of the whole scenario. The spectral index at HE as inferred from the \textit{Fermi}-LAT data is harder than the average spectral index of $\Gamma=2.38\pm0.04$ from the 3FGL catalog dataspan. \
In the presented timeframe, the X-ray emission peaks around the observation on MJD 57070.76434 (2015 February 17), with a delay with respect to the detection in VHE. The $\sim$ 3 hours of observations in VHE during the same night did not lead to a detection (see Table \ref{table:magicdata}). However during the period of enhanced MWL activity, there is a clear hardening of the X-ray spectrum. Hardening at both X-ray and $\gamma$-ray\ energies points toward the emergence of a new component in the non-thermal spectrum. \par
The optical band is very bright during the VHE detection, reaching peaks of more than 20 mJy of flux density when the average behavior of the source is found around a few mJy (see the optical monitoring from Tuorla observatory).
The optical emission is polarized by a fraction of $\gtrsim10\%$ and the polarization angle rotates by $\sim100^\circ$ during the flare: \citet{blinov2015} have shown that from a systematic monitoring (Robopol monitoring) of both $\gamma$-ray\ loud and $\gamma$-ray\ quiet sources, only the former class of object displays polarization angle rotation similar to the one seen here for S4 0954+65. \citet{blinov2015} studied the change of EVPA as a function of time for smooth changes of $>90^\circ$. Requesting the same smoothness requirements, no smooth rotation of $>90^\circ$ can be identified in the dataset presented here, see Fig.~\ref{fig:optical_lc}. A non smooth variation of $\Delta_\textrm{EVPA}\simeq105^\circ$ can however be identified between MJD 57060 and MJD 57075. This variation would imply a change of the EVPA curve slope of $\Delta_\textrm{EVPA}/\Delta_t=7 \textrm{deg/day}$, compatible with the bulk of the variations studied by \citet{blinov2015}.
The rotations of the polarization angle are often physically linked to high flaring states of the objects in the $\gamma$-ray\ band. While individual occurrences of $\gamma$-ray\ flares and rotations cannot be firmly linked to each other, there is a low probability that all the occurrences are due to chance coincidence \citep[from MonteCarlo simulations in][]{blinov2015}. This hypothesis is still confirmed from 3 years of Robopol monitoring data in \citet{blinov2018}. \citet{kiehlmann2017} also study whether a simple stochastic variation can account for the observed rotations in the Robopol monitoring. While their model is failing to recover all the observational characteristics in the monitoring, it also highlights a larger discrepancy from the expectations of stochastic model with respect to the occurrence of large variations of EVPA ($>90^\circ$), however not significant.
Smooth variations seem also to be more firmly linked to deterministic processes and not to a random walk effect \citep{kiehlmann2016}.
Robopol monitoring data are also used in \citet{angelakis2016}, to study the difference in the amount of polarization seen on average in $\gamma$-ray\ loud and $\gamma$-ray\ quiet sources. The median fraction variability of the S4 0954+65 dataset presented here is $16.4\%$. This value can be compared with the average $10\%$ for the $\gamma$-ray\ loud subset of the Robopol monitoring and a value of $17.1\%$ for S4 0954+65 computed for the observations on year 2013 and 2014. According to the interpretation by \citet{angelakis2016}, a higher fractional polarization is also expected in LSP/ISP blazars, due to the fact that in such sources the optical synchrotron emission relates to the peak synchrotron emission. Therefore, the particles associated with this emission are the most energetic, with faster cooling and thus probing a small volume of the emission region near the acceleration region, where it is expected to have a stronger ordered (helical) magnetic field, leading to higher polarization fraction.
\begin{figure}[!htbp]
\resizebox{\hsize}{!}
{\includegraphics[]{figure6.eps}}
\caption{Light curves for R-band polarimetry of S4 0954+65. Please refer to the text for details on the data taking and reduction for each instrument. }
\label{fig:optical_lc}
\end{figure}
Images at 43 GHz show the emergence of new knots. In \citet{morozova2014}, a series of optical flares of S4 0954+65 in 2011 are studied, and the emission of knots is found correlated to the simultaneous flaring of the optical and HE bands. The maximum flux in the 2011 state is a factor of 3 lower in optical than the state presented here. The polarization fraction in this 2011 flare was similar to that seen in the present work. In \citet{morozova2014} the chance coincidence of high optical state and knot emission has very low probability. \par
The phenomenology of the 2015 flare described here agrees very well with the model put forward by \citet{marschernature} and applied to the S4 0954+65 dataset of \citet{morozova2014}. In that model, the flare is due to a newly appearing knot accelerating at the base of the jet and propagating through an helical flow streamline. The helical streamline can be expected due to the anchoring of the accelerating flow to the rotating base of the accretion disk or black hole magnetosphere, depending on modeling. The magnetic field topology in the jet is also helical and ordered. Geometrical effects and the propagation through the helical magnetic field account for the rotation of the EVPA.
In \citet{zhang2014}, a model is proposed where the EVPA rotation is also related to the propagation through an helical magnetic field, but the streamline of propagation is not necessarily helical itself. In this model the magnitude of the swing can depend on the assumptions on the settings for the flare, specifically the magnetic field strength and orientation, the acceleration efficiency and the continuous injection of freshly accelerated particles.
The model described in \citet{marschernature} allows the emission at radio wavelengths in a flaring state which is not simultaneous with the VHE flare. In this scenario the radio activity could be delayed several days, even months, with respect to the VHE detection. This is expected if synchrotron self absorption is involved, and hence the emission region is located closer to the central engine than the radio core (A0 in Fig. \ref{fig:43ghzimages}). The peak of radio emission is expected to be lagging behind and appear when the disturbance has propagated further down the jet, where the absorption is not an issue. The X-ray emission peak, then, could also be delayed with respect to the optical outburst. As the X-ray emission is probably due to IC of an external soft photon field by electrons in the jet (see above), the X-ray variability traces both the accelerated particle distribution and a change in the soft photon field. This retraces similar interpretation drawn for flares of other sources where the dataset was however richer and more detailed \citep{marschernature,marscherpks1510,magic15102012,magic15102017}.
\subsection{Emission model for the flare SED}
\begin{figure*}[!htb]
{\includegraphics[width=\textwidth]{figure7.eps}}
\caption{Spectral energy distribution for the VHE MAGIC detection. Red symbols are strictly simultaneous to the VHE detection, blue symbols are for data taken during the same day and black symbols are for the closest observations. MAGIC spectral data (red circles) are for flare night only (2015 February 14, MJD 57067.14). Red filled circles are for the unfolded observed data points. The red shaded band shows the region of additional systematic uncertainty. \textit{Fermi}-LAT~ data are the PASS8 data for 2015 February 14 (1-day integration centered on the MAGIC observation, blue squares). \textit{Swift}-XRT data are for 2015 February 13 (MJD 57066.70992, blue squares). \textit{Swift}-UVOT data are given for the two separate exposure taken on 2015 February 13 (MJD 57066.76, blue triangles and dark blue triangles). R-band data are for 14th Feb (Tuorla, MJD 57067.16375 and AZT-8+ST7 MJD 57067.1, red diamonds). POLAMI data are for the 18th February (MJD 57071.5, black stars at 100 GHz and 300 GHz). OVRO data are for 2015 February 10 and 19 (black circles at 15 GHz). Mets\"{a}hovi for 2015 February 13 and 15 (MJD 57066.15, MJD 57068.68, black squares at 37 GHz). The gray data are for NED (light) and SSDC (dark) SED historical data points. The model from \citet[][gray dashed curve]{tanaka2016} as well as the model presented here (black solide curve) include an emission component from synchrotron plus inverse Compton on a dusty torus (see text for details and Table \ref{table:SEDmodel} for the values of the physical parameters). The effect of the EBL attenuation is included in the modeling using the model by \citet{finke2010} and a redshift of z=0.368.}
\label{fig:mwl_sed}
\end{figure*}
The SED of blazars are dominated by their non-thermal emission and can usually be described by two broad components. The low energy non-thermal emission is explained as synchrotron emission, while the high energy emission is most commonly modeled through inverse Compton (IC) emission, where soft photons are upscattered to $\gamma$-ray energies by electrons within the jet emitting region. The origin of the soft photon field itself can vary for different blazar subclasses. In particular, for most of the classical BL Lac objects, the VHE emission can be reasonably modeled through Synchrotron self-Compton emission \citep[SSC, see e.g.][]{rees67,maraschi92}. Instead, for the case of FSRQs, the modeling of the emission usually requires the inclusion of external soft photon fields from e.g. the infrared dusty torus or the optical-ultraviolet emission from the Broad Line Region (BLR) for the IC process \citep[see e.g.][]{tavecchiogamma}.
A broadband SED is compiled for 2015 February 14 (MJD 57067). We collect, from the MWL sample described in Section \ref{sec:mwlcoverage}, the data closest in time to the MAGIC observation. \textit{Fermi}-LAT~ data points are obtained from a 1-day integration centered on the MAGIC observation. The specific dates of other wavelength observations are given in the caption of Fig. \ref{fig:mwl_sed}.
\citet{tanaka2016} model the SED of S4 0954+65 during a similar integration time as the 2015 flare studied in this work. The data shown in Fig. \ref{fig:mwl_sed} include, in addition to what is shown by \citet{tanaka2016}, the VHE data from the MAGIC observation, the AZT-8+ST7 and POLAMI data. Moreover, the \textit{Fermi}-LAT~ data are reanalyzed as described in Section \ref{sec:mwlcoverage} to be centered at the MAGIC observation time and benefit from the latest \textit{Fermi}-LAT~ PASS8. The \textit{Swift}-XRT and \textit{Swift}-UVOT data are also reanalyzed for this work.
\citet{tanaka2016} report that a SSC modeling of the data is challenging, requiring very low magnetic field ($\textrm{B}\sim1\mu$G in contrast to the $\textrm{B}\sim1G$ expected in blazar jet components). Alternatively, an External Compton (EC) modeling was able to reproduce the data. In their model, the soft photon field for the EC {model} was the dusty torus from the source. In Fig. \ref{fig:mwl_sed}, we plot the model from \citet{tanaka2016}. This model reproduces the \textit{Fermi}-LAT~ and MAGIC data, although their paper did not include any MAGIC data. However, the model fails to reproduce properly the optical observations. Such underestimation at optical frequencies in the model of \citet{tanaka2016} is driven by a misreconstruction of the UVOT fluxes, explained in Section \ref{sec:mwlcoverage}. With the reanalyzed UVOT dataset presented here, we use a new model, using the same code and most of the same assumptions as in \citet{tanaka2016}, including a redshift of $z=0.368$. The code is explained in detail in \citet{finke08} and \citet{dermer09}. Note that the presented SED model curves already include the effect of EBL absorption, i.e the intrinsic emission is absorbed according to the EBL model by \citet{finke2010}. The new EC model provides a good description of the MWL data and is shown in Fig. \ref{fig:mwl_sed}. The parameters of both models are reported in Table \ref{table:SEDmodel}. The break in the underlying electron population is similar to what expected by classical cooling, with the slope of the electron distribution before of the break ($s_1$) and after the break ($s_2$) differing by $s_2-s_1=1.2$. The use of VHE spectral information is crucial to model the falling part of the high energy peak of the blazars SEDs, which is crucial to constrain the most energetic electrons within the leptonic framework scenario (SSC and EC models).
\begin{table*}[!htbp]
\caption{SED model parameters}
\label{table:SEDmodel}
\centering
\begin{tabular}{c c c c}
\hline\hline
Parameter & Symbol & Model A & Model B \\
& & \citet{tanaka2016} & this work \\
\hline
Redshift & $z$ & \multicolumn{2}{c}{0.368} \\
Bulk Lorentz Factor & $\Gamma$ & 30 & 35 \\
Doppler factor & $\delta_D$ & 30 &35 \\
Variability Timescale [s]& $t_v$ & $1.0\times10^5$ & $4\times10^4$\\
Comoving radius of blob [cm]& $R^{\prime}_b$ & 6.6$\times$10$^{16}$ & 3.0$\times$10$^{16}$\\
Magnetic Field [G]& $B$ & 0.6 & 0.4 \\
\hline
Low-Energy Electron Spectral Index & $s_1$ & 2.4 & 2.4 \\
High-Energy Electron Spectral Index & $s_2$ & 4.5 & 3.6 \\
Minimum Electron Lorentz Factor & $\gamma^{\prime}_{\rm min}$ & $1.0$ & $1.0$ \\
Break Electron Lorentz Factor & $\gamma^{\prime}_{\rm brk}$ & $8.0\times10^3$ & $4.0\times10^3$ \\
Maximum Electron Lorentz Factor & $\gamma^{\prime}_{\rm max}$ & $2.0\times10^4$ & $4.0\times10^4$ \\
\hline
Black hole Mass [$M_\odot]$ & $M_{\rm BH}$ & \multicolumn{2}{c}{$3.4\times10^8$} \\
Disk luminosity [$\mathrm{erg} \ \mathrm{s}^{-1}$] & $L_{\rm disk}$ & \multicolumn{2}{c}{$3.0\times10^{43}$} \\
Inner disk radius [$R_g$] & $R_{\rm in}$ & \multicolumn{2}{c}{$6.0$}\\
Seed photon source energy density [$\mathrm{erg}\ \mathrm{cm}^{-3}$] & $u_{\rm seed}$ & $2.4\times10^{-4}$ & $4.4\times10^{-5}$\\
Seed photon source photon energy [$m_e c^2$ units] & $\epsilon_{\rm seed}$ & $7.5\times10^{-7}$ & $5\times10^{-7}$\\
Dust Torus luminosity [$\mathrm{erg}\ \mathrm{s}^{-1}$] & $L_{\rm dust}$ & $3.9\times10^{42}$ & $1.5\times10^{42}$ \\
Dust Torus radius [cm] & $R_{\rm dust}$ & $2.1\times10^{17}$ & $6.1\times10^{17}$ \\
Dust temperature [K] & $T_{\rm dust}$ & $1500$ & $1000$\\
\hline
Jet Power in Magnetic Field [$\mathrm{erg}\ \mathrm{s}^{-1}$] & $P_{j,B}$ & $1.0\times10^{46}$ & $1.4\times10^{45}$ \\
Jet Power in Electrons [$\mathrm{erg}\ \mathrm{s}^{-1}$] & $P_{j,e}$ & $1.1\times10^{45}$ & $6.6\times10^{45}$ \\
\hline
\end{tabular}
\end{table*}
As mentioned in the introduction, the classification of a blazar can be aided by the study of its SED characteristics. According to the SED model presented above, the peak of the synchrotron emission is at $\nu_\textrm{syn}\sim8\times10^{14}$ Hz, making it an intermediate synchrotron peaked BL Lac object \citep{3LAC}\footnote{intermediate-synchrotron-peaked blazar (ISP) are defined with rest-frame synchrotron peak frequencies of $10^{14}\textrm{Hz}<\nu_\textrm{syn}<10^{15}\textrm{Hz}$}. The Compton dominance, calculated comparing the luminosity at the peak of the synchrotron emission to that of the IC peak, is $L_\textrm{IC}/L_\textrm{syn}\sim 7$. Such Compton dominance value is at least 3.5 times the values obtained by \cite{finke_cd} for long-term blazar studies.
\section{Conclusions}
The census of extragalactic objects that present VHE emission is still limited. We present here the first detection at VHE of the blazar S4 0954+65 obtained through observations with the MAGIC Telescopes. The observations were conducted during an exceptional flare of the source in February 2015, originally identified in the optical band. We collected MWL simultaneous data to better characterize the state of the source.
The HE emission is also found in elevated state from the analysis of \textit{Fermi}-LAT~ data, which reveal the hardest state of the HE emission to be concurrent with the detection at VHE. The X-ray emission peak is delayed by a few days with respect to the VHE detection and shows a trend of spectral hardening during the period presented here. The radio and mm wavelength emission reveal a moderate elevation of the flux, that is however not exceptional in the long term behavior of the source.
The source is classified in the literature as a BL Lac, but we have shown here that it presents similarities with the FSRQ class. Results from the monitoring of optical polarization and 43 GHz jet component analysis were compared to archival observation of S4 0954+65 and of statistical behaviour of other sources. Three main measurements were considered: the day of the VHE detection of S4 0954+65 is included in the error box for the zero epoch separation of knot K15; the optical polarization fraction is increasing in the same period; a rotation of optical EVPA of $\sim100^\circ$ can be identified, also in the same period, possibly related to the helical structure of the magnetic field in the acceleration region. We have discussed how these measurements point to a common behaviour with ISP/LSP sources. Both the best emission model (EC on dust torus) and the MWL light curve behavior show points of contact with other sources that are either clear FSRQ (like PKS 1510-089) or are transitional objects (like BL Lac itself). This is also supported from the moderate Compton dominance in the SED model and the fact that the synchtrotron peak show that the source can be classified as ISP source.
The work presented here reiterates the importance of VHE $\gamma$ ray\ and detailed MWL studies of blazars during different flux states to test their intrinsic characteristics and shed light on the physical processes taking place within their jets.
\begin{acknowledgements}
We would like to thank the Instituto de Astrof\'{\i}sica de Canarias for the excellent working conditions at the Observatorio del Roque de los Muchachos in La Palma. The financial support of the German BMBF and MPG, the Italian INFN and INAF, the Swiss National Fund SNF, the ERDF under the Spanish MINECO (FPA2015-69818-P, FPA2012-36668, FPA2015-68378-P, FPA2015-69210-C6-2-R, FPA2015-69210-C6-4-R, FPA2015-69210-C6-6-R, AYA2015-71042-P, AYA2016-76012-C3-1-P, ESP2015-71662-C2-2-P, CSD2009-00064), and the Japanese JSPS and MEXT is gratefully acknowledged. This work was also supported by the Spanish Centro de Excelencia ``Severo Ochoa'' SEV-2012-0234 and SEV-2015-0548, and Unidad de Excelencia ``Mar\'{\i}a de Maeztu'' MDM-2014-0369, by the Croatian Science Foundation (HrZZ) Project IP-2016-06-9782 and the University of Rijeka Project 13.12.1.3.02, by the DFG Collaborative Research Centers SFB823/C4 and SFB876/C3, the Polish National Research Centre grant UMO-2016/22/M/ST9/00382 and by the Brazilian MCTIC, CNPq and FAPERJ.
The \textit{Fermi} LAT Collaboration acknowledges generous ongoing support
from a number of agencies and institutes that have supported both the
development and the operation of the LAT as well as scientific data analysis.
These include the National Aeronautics and Space Administration and the
Department of Energy in the United States, the Commissariat \`a l'Energie Atomique
and the Centre National de la Recherche Scientifique / Institut National de Physique
Nucl\'eaire et de Physique des Particules in France, the Agenzia Spaziale Italiana
and the Istituto Nazionale di Fisica Nucleare in Italy, the Ministry of Education,
Culture, Sports, Science and Technology (MEXT), High Energy Accelerator Research
Organization (KEK) and Japan Aerospace Exploration Agency (JAXA) in Japan, and
the K.~A.~Wallenberg Foundation, the Swedish Research Council and the
Swedish National Space Board in Sweden.
This work performed in part under DOE
Contract DE-AC02-76SF00515
This research has made use of the NASA/IPAC Extragalactic Database (NED),
which is operated by the Jet Propulsion Laboratory, California Institute of Technology,
under contract with the National Aeronautics and Space Administration.
Part of this work is based on archival data, software or online services provided by the Space Science Data Center - ASI.
The OVRO 40-m monitoring program is supported in part by NASA grants NNX08AW31G, NNX11A043G and NNX14AQ89G, and NSF grants AST-0808050 and AST-1109911
The research at Boston University was supported by NASA Fermi Guest Investigator program grant 80NSSC17K0694 and US National Science Foundation grant AST-1615796. The VLBA is an instrument of the Long Baseline Observatory.
The Long Baseline Observatory is a facility of the National Science Foundation operated by Associated Universities, Inc.
This paper is partly based on observations carried out with the IRAM 30 m Telescope. IRAM is supported by INSU/CNRS (France), MPG (Germany) and IGN (Spain). IA acknowledges support by a Ram\'on y Cajal grant of the Ministerio de Econom\'ia, Industria y Competitividad (MINECO) of Spain. The research at the IAA--CSIC was supported in part by the MINECO through grants AYA2016--80889--P, AYA2013--40825--P, and AYA2010--14844, and by the regional government of Andaluc\'{i}a through grant P09--FQM--4784.
St. Petersburg University team acknowledges support from Russian Science
Foundation grant 17-12-01029.
The Submillimeter Array is a joint project between the Smithsonian Astrophysical Observatory and the Academia Sinica Institute of Astronomy and Astrophysics and is funded by the Smithsonian Institution and the Academia Sinica.
\end{acknowledgements}
|
\section{Related work}
Recent results in popularity modeling feature a range of theoretical models to explain and predict online content popularity~\cite{bakshy2011everyone, Martin2016, Mishra2016,Zhao2015,Rizoiu2017}.
However, the choice of publicly available platforms that implement these models for the non-scientific user is rather limited.
\citet{khosla2014makes} implement a web-based demonstration\footnote{Demonstration from \citet{khosla2014makes}: \url{http://popularity.csail.mit.edu/}} for predicting image popularity before publishing the images.
Their system computes a popularity confidence by taking image content and a range of social factors into account.
\citet{castillo2014characterizing} create FAST\footnote{Platform by \citet{castillo2014characterizing}: \url{http://fast.qcri.org/}} (Forecast and Analytics of Social Media and Traffic) for predicting views of news article, by leveraging anonymous real-time page view data.
\citet{xie2016topicsketch} provide a tool\footnote{Tool from \citet{xie2016topicsketch}: \url{http://research.pinnacle.smu.edu.sg/clear/}} that uses real-time Twitter data to identify potential viral topics and predict their total popularity.
HIPie\xspace differs from the aforementioned works in several ways.
First, it aims to be more than a demonstration of a scientific algorithm: it is a visualization platform dedicated to users.
It has multiple visualization components and the user-friendly interaction enables this platform to easily convey the modeling outcome.
Second, this platform is built for Youtube video popularity modeling, while other platforms generally deal with Twitter, news article, etc.
Third, it is open-source and it allows developers to integrate it with other popularity models and additional data sources.
\section{Introduction}
The popularity of online videos is typically measured by the number of views they attract from viewers.
Understanding online popularity can help content producers to propose better content, and content consumers to deal with information overload.
Viral videos quickly catch the attention of the viewers and achieve very high popularity in short periods of time.
Explaining what makes videos go viral, and identifying them early would prove useful for advertisers and content providers.
Our tool aims to fill several gaps about the systems that enable users to reason about the popularity of online videos.
The first gap concerns the availability of such systems.
Despite the range of theoretical models that have been recently proposed for modeling online popularity~\cite{bakshy2011everyone,Martin2016,Zhao2015,Rizoiu2017},
\textbf{there is no readily available software that allows regular users to easily
examine the popularity over time for online videos and forecast their future popularity.}
The second gap concerns content producers and advertisers who need to choose which videos to promote and to identify potentially viral videos.
\textbf{How can content producers quantify virality and simulate video reaction to online promotions?}
The third gap sits for content consumers.
Most distribution platforms (e.g. Youtube) feature personalized recommendation systems; these usually act as black boxes and make the decision for the user.
The open question is \textbf{how can the user be empowered by enabling her to compare and select content on the fly?}
In this work, we answer the above three questions, building upon the current state-of-the-art popularity model, the Hawkes Intensity Process (HIP)~\cite{Rizoiu2017}.
We introduce the \emph{HIP Insights Explorer} (HIPie\xspace),
an interactive web-based application designed to assist users to reason about the popularity and the virality of Youtube videos.
It exposes a series of measures derived from HIP -- such as the sensitivity to external promotions and the endogenous amplification.\eat{ -- and it}
It allows \rev{one} to
conduct various tasks, including identifying prospective popular videos, simulating video reaction to promotion schedules, comparing videos from different authors (i.e. Youtube channels) and visualizing the popularity series fitted and predicted by HIP.
The most important visualization of HIPie\xspace is the endo-exo map~\citep{Rizoiu2017}, a projection of videos in the two-dimensional space defined by \rev{the} endogenous response and exogenous sensitivity.
The relative positions of the videos in this space indicate their potential of becoming viral.
HIPie\xspace allows adding any Youtube video on-the-fly as long as its popularity series are available.
The main contributions of this work include:
\squishlist
\item A web-based interactive tool to visualize and predict future video popularity using the HIP~\citep{Rizoiu2017} popularity model;
\item The endo-exo map visualization, on which \rev{the} viral potential of videos is compared;
\item HIPie\xspace enables a series of \rev{applications} concerning online popularity, such as comparing videos and channels, identifying future popular videos and simulating video reaction to promotion.
\squishend
\section{Prerequisites}
In this section we briefly review the HIP model and how it is used to quantify virality and simulate the effect of promotions.
\textbf{Explain and forecast popularity with HIP.}
HIP~\citep{Rizoiu2017} is a novel generative model that explains online popularity series by linking exogenous inputs from public social media platforms, such as Twitter or Youtube, to endogenous responses within the Youtube content platform,
which account for the word-of-mouth process occurring around videos.
HIP models popularity using:
\begin{equation} \label{eq:HIP}
\vspace{-0mm}} %{\vspace{-2.4mm}
\xi(t) = \mu s(t) + C \int^{t}_{0} \xi(t-\tau)(\tau + c)^{-(1+\theta)}d\tau
\end{equation}
where $\xi(t)$ is the number of views that the video receives during day $t$ and $s(t)$ is the volume of exogenous inputs (tweets, shares or promotions).
Eq.~\eqref{eq:HIP} can intuitively be understood as: the number of views a video receives during day $t$ is dependent on its popularity at each previous day $1, .. , t-1$ decayed by how fast people forget ($\theta$ is the exponent controlling the power-law decay of social memory).
The parameters of HIP ($\mu, C, c, \theta$) are fit on an observed prefix of the views and the shares series.
If future exogenous stimuli \rev{are} known, the popularity series can be ``run forward'' by plugging $s(t)$ and the past popularity into Eq.~\eqref{eq:HIP}.
\textbf{The endo-exo map.}
Two metrics derived from HIP describe a video's virality.
The exogenous sensitivity $\mu$ quantifies the video sensitivity to the external stimuli $s(t)$.
The endogenous response $A_{\hat{\xi}}$ is computed as $A_{\hat{\xi}} = \sum^{\infty}_{t=0} \hat{\xi}(t)$, where $\hat{\xi}(t)$ is the popularity series generated by a single initial exogenous impulse.
Intuitively, $A_{\hat{\xi}}$ represents the total amount of endogenous amplification that each view generates.
\citet{Rizoiu2017} introduce the \emph{endo-exo map}, which is a two-dimensional space of the exogenous sensitivity and the endogenous response.
It is used to identify potentially viral videos, videos with high scores on both dimensions, as well as unpromotable videos.
\textbf{Viral potential and the reaction to promotions.}
\citet{Rizoiu2017b} use HIP in an advertisement application, in which the aim is to quantify the effect of promotion on content popularity.
The \emph{viral potential} $\nu = \mu A_{\hat\xi}$ is the \emph{return on investment}, or the total number of views generated by a single promotion.
They also study the effect of promotion schedules on the views series.
They construct the promotion series by allocating \rev{to} each day an amount of promotions, and by \rev{introducing} it into HIP alongside with the organic exogenous stimuli to obtain the promoted view series.
\section{Applications}
\label{sec:applications}
HIPie\xspace has a series of functionalities that enable users to understand, reason and interact with the popularity of Youtube videos.
\textbf{Explain and predict Youtube video popularity (Fig.~\ref{subfig:popularity-series}).}
For any video, HIPie\xspace depicts several popularity series: observed, fitted and forecasted by HIP.
Fig.~\ref{subfig:popularity-series} shows the example of a Music video ``Footprints'' from the Dutch DJ \textit{Tiesto} (Youtube id: \textit{bcJI2DMPk40}).
It shows the popularity series for the first 120 days after video upload.
The dotted blue line represents the observed view counts (i.e., real data) and the red line is the external promotion series.
Fitting HIP and forecasting future popularity are performed in a temporal holdout setup.
The views and shares series in the first 90 days are used to fit the parameters of HIP.
The green line shows the fitted view count series.
The orange line represents the predicted view counts series, using the previously fitted parameters and the external promotion series from day 91 to 120.
\rev{As shown in Fig.~{\ref{subfig:popularity-series}} and in the online public installation (described in Sec.~{\ref{sc:obtaining}}), the popularity series fitted by HIP follows closely the observed popularity series.
Furthermore, \citet{Rizoiu2017} have shown HIP to be able to forecast future popularity with less than $5\%$ mean absolute percentile error when using shares as the exogenous stimuli series, and $5.35\%$ when using tweets.}
\textbf{Compare videos (Fig.~\ref{subfig:endo-exo}).}
HIPie\xspace enables users to comparatively analyze videos using the endo-exo map, by showing the amount of views and shares they receive, alongside with the exogenous sensitivity and the endogenous reaction.
Fig.~\ref{subfig:endo-exo} shows six of the most popular \rev{pop} songs on Youtube on the endo-exo map.
Each video is presented as a bubble, where the x coordinate is the endogenous response and the y coordinate is the exogenous sensitivity.
The color depth of bubbles indicates the amount (in percentage scale) of external promotions that the video receives and the size of bubbles shows the amount of views it receives.
``Gentleman'' (id \textit{ASO\_zypdnsQ}) and ``Gangnam style'' (id \textit{ 9bZkp7q19f0}) from the Korean singer Psy occupy the most privileged position on the map, both having a high exogenous sensitivity and a high endogenous response.
In comparison, ``Hello'' by Adele (id \textit{YQHsXMglC9A}) has lower exogenous sensitivity, while ``PPAP'' by the Japanese singer Pikotaro (id \textit{0E00Zuayv9Q}) has a lower endogenous response.
\textbf{Compare channels (Fig.~\ref{subfig:separating-channels}).}
In HIPie\xspace we use the endo-exo map to visualize groups of videos that belong to the same user-assigned content type, or are from the same author (called channel in YouTube).
Fig.~\ref{subfig:separating-channels} shows in black color a scatter plot of videos in the category News\&Activism, posted by the reporter \emph{Anatolii Sharij} covering the 2014 events in Ukraine, and in red color a user (\emph{VEGETTA777}) focusing on recordings of Game sessions.
The game recording videos are generally more popular (bigger bubbles) than the news
videos, and this is explained by the former group having higher exogenous sensitivity -- higher values of $\mu$.
\textbf{Identify potentially viral videos (Fig.~\ref{subfig:future-popular}).}
HIPie\xspace allows to identity videos that have the potential of going viral, but are yet to.
These are videos with high exogenous sensitivity and high endogenous response (top-right corner of the endo-exo map), but which \rev{have received} very little external stimuli.
Video \emph{H} from in Fig.~\ref{subfig:future-popular} -- a Japaneese Film\&Animation video -- is an example of a potentially viral video: it has both high endogenous response and exogenous sensitivity, it has received few promotions (light color) and has achieved little popularity (small bubble size).
Every external stimuli that this video receives will generates \eat{in time }450 views.
\textbf{Simulate video response to promotions (Fig.~\ref{subfig:promotions-before} \&~\ref{subfig:promotions}).}
HIPie\xspace allows simulating ``what-if'' scenarios: what would be the popularity of a video if it \rev{received} an additional volume of external stimulation.
One can promote (or demote) a video by adding or subtracting a volume of promotion, spread equally among the first 90 days (the \emph{even} promotion schedule studied by \citet{Rizoiu2017b}).
Video \textit{A} in Fig.~\ref{subfig:promotions-before} is a collection of ``ice bucket'' challenges (id: \textit{3hSIh-tbiKE}) which receives little external stimuli (light color) and low popularity (small bubble).
Video \textit{B} (a Gaming video, id \textit{0lTTWeavl1c}) has a similar position on the endo-exo map, but it has a higher popularity due to having received more external stimuli.
After promoting video \textit{A} with an amount of promotions similar to \textit{B} (i.e. similar color in Fig.~\ref{subfig:promotions}), \textit{A} achieves a similar popularity \rev{level} as \textit{B} (similar bubble size).
\section{Description of the demo}
In this section, we introduce the main interfaces of HIPie\xspace (Sec.~\ref{sc:interfaces}),
present some implementation choices (Sec.~\ref{sc:implementation})
and how to obtain, test and use HIPie\xspace (Sec.~\ref{sc:obtaining}).
\subsection{Main Interfaces}
\label{sc:interfaces}
HIPie\xspace is designed as an interactive web application.
Fig.~\ref{fig:overview} gives an overview of the main interface of HIPie\xspace, containing four panels on the right and a navigation menu on the left.
In this section, we describe the main interfaces and functionalities.
\textbf{The endo-exo map (top-left)}.
This is the most important panel in HIPie\xspace.
In addition to the characteristics described in Sec.~\ref{sec:applications}, the plot is interactive:
hovering \rev{over} bubbles shows a pop up with additional information about \rev{the} video (YoutubeID, Author, Title, endogenous response and exogenous sensitivity values, acquired percentiles of views and shares).
The user can zoom into certain areas of the map and drag the map.
Clicking on bubbles in endo-exo map causes \rev{other three panels} to switch to the current active video.
The rest of the panels in Fig.~\ref{fig:overview} are updated to show detailed information of the current active video.
\textbf{The popularity panel (top-right)}
shows the popularity series of the current active video: observed, fitted and forecasted view counts and observed share counts.
Hovering \rev{over} the plot gives the values of each series at the hovering time point.
\textbf{The preview panel (bottom-left).}
This panel allows to play the selected video, while interacting with the other panels.
\textbf{The video metadata panel (bottom-right).}
This panel shows detailed metadata information about the current active video: Youtube ID, Title, Author, Category, Upload date, number of views, number of shares and endo-exo values.
\textbf{Collection management menu.}
HIPie\xspace allows \rev{one} to create, manage and visualize multiple video collections, using the navigation menu and controls located on the left side of the interface.
The default collections cannot be changed.
\textbf{Adding New Videos.}
HIPie\xspace allows \rev{one} to add or remove videos from the current collection, by using the buttons in the top-right corner of the interface.
A new video is added by inputing its
Youtube video ID or video link.
A backend process will crawl the video metadata and popularity series using the \textit{youtube-insight}~\cite{wu2017beyond} package.
Once crawling completes, the first 90 days of the popularity series are used to fit the parameters of HIP.
Any Youtube videos can be added to the system as long as its popularity series are publicly accessible and at least 120 days data.
After the fitting is completed, the video appears in the corresponding collection.
\subsection{Implementation}
\label{sc:implementation}
HIPie\xspace is built in R~\citep{R} using the open-source package Shiny~\citep{shiny}, which is dedicated to creating interactive web applications in R.
Shiny enables developers to focus on visualization by simplifying front-end design and backend configuration and it provides web-based input and output tools.
\rev{The other employed packages are chosen for their compatibility with Shiny and for their efficiency}.
\rev{For example, we} employ Plotly to construct the interactive visualization experience.
\rev{It allows \rev{one} to easily convert static R plots into interactive visualization.}
We also use ShinyJs to build customized JavaScript interactions between R and the frontend, \rev{as it is compatible with default Shiny elements and has many useful APIs}.
Table~\ref{t:packages} lists all the packages employed in HIPie\xspace, together with their project links and brief descriptions.
\subsection{Obtaining and running the demo}
\label{sc:obtaining}
HIPie\xspace is open-source and publicly accessible.
To download the source code, access a live demo or a quick tour, we provide the following options:
\begin{itemize}
\item A \textbf{public installation} of HIPie\xspace\footnote{HIPie\xspace public installation: \url{http://www.hipie.ml/}} is live for testing, requiring only a web browser.
\item The source code of HIPie\xspace can be accessible from a \textbf{Github repository}\footnote{Github repository of HIPie\xspace: \url{https://github.com/computationalmedia/hipie/}}.
Simply clone the repository and follow the instructions in the README file to run a local installation.
\item For a quick tour of the various usages and capabilities of HIPie\xspace, we provide a \textbf{short Youtube video}\footnote{Screencast for HIPie\xspace: \url{https://youtu.be/x5xIf4vUScI/}}.
\end{itemize}
\input{tab1-summary-packages}
\section*{Authors}
\newcommand\mytextwidth{19.5em}
\newcommand\myheight{0.09}
\begin{tabular} {ll}
\parbox[c]{\myheight\textheight}{
\includegraphics[height=\myheight\textheight]{me}
} & \parbox[c]{\mytextwidth}{\textbf{Quyu Kong} received his Bachelor of Agronomy at the Zhejiang University in 2015 and Master of Computing (Advanced) degree in Artificial Intelligence at the Australian National University in 2017. His research interests include Information Diffusion Modeling and Machine Learning.} \\
& \\
\parbox[c]{\myheight\textheight}{
\includegraphics[height=\myheight\textheight]{andrei}
} & \parbox[c]{\mytextwidth}{\textbf{Marian-Andrei Rizoiu} is a Research Fellow with the Australian National University.
He employs big social data to develop theoretical models for novel societal phenomena, such as information diffusion in online networks and online popularity modeling and prediction.
} \\
& \\
\parbox[c]{\myheight\textheight}{
\includegraphics[height=\myheight\textheight]{siqi}
}
& \parbox[c]{\mytextwidth}{\textbf{Siqi Wu} is a PhD student at the Australian National University. His research focuses on the analysis of user engagement in online media and influential network in social graph.} \\
& \\
\parbox[c]{\myheight\textheight}{
\includegraphics[height=\myheight\textheight]{lexing}
}
& \parbox[c]{\mytextwidth}{\textbf{Lexing Xie} is an Associate Professor in computer science with the College of Engineering and Computer Science in the Australian National University, leading the ANU Computational Media Lab. Her research focuses broadly on innovative use and design of machine learning algorithms and systems, especially on structured and multi-relational graph data.} \\
\end{tabular}
\section{Inter-event time probabilities in non-homogeneous Poisson processes}
\label{ap-subsec:probs-NHPP}
In this section, we revisit the Non-Homogeneous Poisson Process (NHPP) and we compute the formula for the probabilities of observing inter-arrival times.
We also show that NHPP is a non-Markovian process and we derive a simple proof for the formula for the log-likelihood of a NHPP, which is widely used in CS literature, but an accessible proof of which is currently missing.
\subsection{Inter-arrival times probabilities}
Here we compute the probability of observing $t_i$ -- the arrival of an event.
We denote by $\tau_i$ the inter-arrival time between event $i-1$ and event $i$.
It follows that $\tau_i = t_i - t_{i-1}$ and $t_i = \sum_1^j \tau_j$.
We study in parallel the Homogeneous Poisson Process (HPP) and NHPP.
For ease of understanding, we further consider the two cases when $i = 1$ and $i > 1$.
\textbf{The arrival of the first event $t_1$}.
In a HPP of intensity $\lambda$, the probability of having no events in the time interval $[0, t)$ is:
\begin{equation} \label{eq:CCDF-waiting-time}
\mathbbm{P}[t_1 \ge t] = e^{-\lambda t} \enspace .
\end{equation}
This can be interpreted as the probability of waiting at least $t$ units of time until the first event.
Consequently, Eq.~\eqref{eq:CCDF-waiting-time} is the CCDF (Complementary Cumulative Distribution Function) of the waiting time until the first event.
The PDF is $ PDF = \fp{}{t} (1 - CCDF) = - \fp{}{t} CCDF$.
Consequently \emph{the waiting time to the first event in a HPP is distributed exponentially}, with parameter $\lambda$:
\begin{equation}
\mathbbm{P}[t_1 = t] = - e^{-\lambda t} \fp{-\lambda t}{t} = \lambda e^{-\lambda t} \enspace.
\end{equation}
For a NHPP with the event rate $\lambda(t)$, we first define the function $\Lambda(t) = \int_0^t \lambda(\tau) d\tau$.
The inverse relation between $\lambda(t)$ and $\Lambda(t)$ is $\lambda(t) = \fp{}{t} \Lambda(t)$.
We have:
\begin{equation} \label{eq:NHPP-wait-0-t}
\mathbbm{P}[t_1 \ge t] = e^{-\Lambda(t)} \enspace ,
\end{equation}
and we compute
\begin{equation} \label{eq:NHPP-waiting-t1}
\mathbbm{P}[t_1 = t] = \fp{}{t} e^{-\Lambda(t)} = - e^{-\Lambda(t)} \fp{}{t} \Lambda(t) = \lambda(t) e^{-\Lambda(t)} \enspace.
\end{equation}
Note that \emph{the waiting time to the first event is not exponentially distributed in the case of NHPP}.
An intuitive interpretation of Eq.~\eqref{eq:NHPP-waiting-t1} is that the probability of observing an event at time $t$ is the product of
the probability of observing an event in the infinitesimal time interval $[t, t + \partial t]$ -- equal to the event rate $\lambda(t)$ --
and the probability having observed no event in $[0, t]$ -- as defined in Eq.\eqref{eq:NHPP-wait-0-t}.
\textbf{The arrival of $t_2, t_3, \ldots, t_n$}.
For a HPP of rate $\lambda$, the probability of not observing an event in the interval $[t, t + s]$ -- after having observed a first event at time $t_1 = t$ -- is:
\begin{equation*}
\mathbbm{P}[t_2 - t_1 \ge s | t_1 = t] = e^{- \lambda (t+s - t)} = e^{- \lambda s} \enspace .
\end{equation*}
does not depend of $t$. By denoting $\tau_2 = t_2 - t_1$ and $\tau_1 = t_1$, we obtain
\begin{equation} \label{eq:HPP-inter-arrival-times-prob-distribution}
\mathbbm{P}[\tau_2 = s | \tau_1 = t] = \lambda e^{- \lambda s} \Longrightarrow \mathbbm{P}[\tau_i = s ] = \lambda e^{- \lambda s} \enspace.
\end{equation}
Inter-arrival times in a HPP are exponentially distributed with parameters $\lambda$, and the probability of observing a $\tau_i$ does not depend on the previous inter-arrival times $\tau_1, \tau_2, \ldots, \tau _{i-1}$.
This property is called \emph{memorylessness} -- and it is equivalent to the Markovian property~\cite{Allen2008} -- as the next state of the process depends only on the current state and not on the past.
For the NHPP of rate $\lambda(t)$, we have
\begin{align}
\mathbbm{P}[t_2 - t_1 \ge s | t_1 = t] &= e^{ \Lambda (t) - \Lambda(t+s)} \nonumber \\
\Rightarrow \; \mathbbm{P}[t_2 - t_1 = s | t_1 = t] &= \fp{}{s} \mathbbm{P}[t_2 - t_1 \ge s | t_1 = t] \nonumber \\
&= \lambda(t+s) e^{ \Lambda (t) - \Lambda(t+s)} \nonumber
\end{align}
$\Lambda (t) - \Lambda(t+s)$ can be interpreted as the minus area under the curve of $\lambda(t)$.
We can further show that
\begin{equation} \label{eq:NHPP-inter-event-times-prob}
\mathbbm{P}[\tau_{i+1} = s | \mathcal{H}_i] = \lambda(t_i+s) e^{ \Lambda (t_i) - \Lambda(t_i+s)}
\end{equation}
where $\mathcal{H}_i = \{ t_1, t_2, \ldots, t_i\}$ is the history of the process up to event $t_i$.
Note that when $\lambda(t) = \lambda$ -- i.e. a HPP -- we have $\Lambda(t) = \lambda t$ and Eq.~\ref{eq:HPP-inter-arrival-times-prob-distribution} and~\ref{eq:NHPP-inter-event-times-prob} are identical.
We can express Eq.~\eqref{eq:NHPP-inter-event-times-prob} in terms of event times (rather than inter-event times):
\begin{equation} \label{eq:NHPP-event-times-prob}
\mathbbm{P}[t_{i+1} | \mathcal{H}_i] = \lambda(t_{i+1}) e^{ \Lambda (t_i) - \Lambda(t_{i+1})}
\end{equation}
\subsection{Two follow-up conclusions}
We study the Markovian property of NHPP and we derive its likelihood function.
\textbf{NHPP is not Markovian}.
One direct consequence of Eq~\eqref{eq:NHPP-inter-event-times-prob} is that inter-arrival times in a NHPP are not exponentially distributed.
We further study if the process is memoryless -- i.e. if it has the Markovian property.
For this, we compute the join probability of having an event in the interval $[0, t]$ and a second event in $[t, s]$.
\begin{align}
\mathbbm{P}[t_1 = t, t_2 = t + s] &= \mathbbm{P}[t_1 = t] \mathbbm{P}[t_2 = t + s | t_1 = t] \nonumber \\
&= \lambda(t) \lambda(t + s) e^{- \Lambda(t + s)} \label{eq:non-markov}
\end{align}
which shows that $t_2$ is not independent of $t_1$.
The implication is that the next state of a NHPP -- i.e. $t_{i+1}$ -- is dependent on all previous states -- $t_j, j \in [1 \dots i]$.
\emph{This shows that NHPP is not Markovian.}
Note that this is a general results, for non-specific functions $\lambda(t)$
Specific functions $\lambda(t)$ can be constructed so that the NHPP becomes Markovian.
As a sanity check, we write Eq.~\eqref{eq:non-markov} for a HPP.
We obtain
\begin{align}
\mathbbm{P}[t_1 = t, t_2 = t + s] &= \lambda ^2 e^{-\lambda (t+s)} \nonumber \\
&= \lambda e^{-\lambda t} \lambda e^{- \lambda s} = \mathbbm{P}[t_1 = t] \mathbbm{P}[t_2 = t + s]
\end{align}
therefore the inter-arrival times $\tau_1$ and $\tau_2$ are independent and exponentially distributed -- as expected.
\textbf{The likelihood function for NHPP}.
Given $\mathcal{H}_i$, which includes the parameter of the process $\theta$ and the history of the process up to event $t_i$, the probability of an event at time $t_{i+1}$ is defined (according to Eq.~\eqref{eq:NHPP-event-times-prob} as the probability of observing an event at time $t_{i+1}$ -- $\Lambda(t_{i+1})$ -- and the probability of not having observed any event in the interval $[t_i, t_{i+1}]$.
We construct the likelihood function as
\begin{align}
Likelihood(\theta) &= \mathbbm{P}[t_1, t_2, \dots, t_n | \theta] \nonumber \\
&= \mathbbm{P}[t_1 | \theta] \mathbbm{P}[t_2 | t_1, \theta] \mathbbm{P}[t_3 | t_2, t_1, \theta] \ldots \mathbbm{P}[t_n | t_{n-1}, \ldots t_1, \theta] \nonumber \\
&= \prod_{i = 1}^n \mathbbm{P}[t_i | \mathcal{H}_{i-1}] = e^{-\Lambda(t_1)+\Lambda(t_1)-\Lambda(t_2) + \ldots - \Lambda(t_n)} \prod_{i = 1}^n \lambda(t_i) \nonumber \\
&= \prod_{i = 1}^n \lambda(t_i) e^{-\Lambda(t_n)} \nonumber
\end{align}
Finally, we derive the expression of the log-likelihood widely used in literature:
\begin{align}
log(Likelihood(\theta)) &= \sum_{i = 1}^n log \left( \lambda(t_i) \right) - \Lambda(t_n) \nonumber \\
&= \sum_{i = 1}^n log \left( \lambda(t_i) \right) - \int_0^{t_n} \lambda(\tau)d\tau . \label{eq:log-likelihood}
\end{align}
\section{Fitting HawkesN with AMPL -- implementation}
We fit the parameters of the HawkesN model to observed data by maximizing the log-likelihood function Eq.~\eqref{eq:log-likelihood}.
We use AMPL, an industry standard for modeling optimization problems and with a transparent interfaces to powerful solvers.
We start with an introduction of AMPL (Sec.~\ref{subsec:ampl-intro}), we describe our optimization setup and the employed solvers (Sec.~\ref{subsec:optimization-setup}) and we finish with the R interface that we constructed for AMPL (Sec.~\ref{subsec:ampl-r-interface}).
\subsection{AMPL introduction}
\label{subsec:ampl-intro}
Since the first commercial release in 1993, AMPL -- which stands for A Mathematical Programming Language -- has provided a convenient interface between mathematic modelers and implemented solvers~\cite{fourer1987ampl}.
It now also offers a complete tool set including many solvers for modeling different optimization problems.
Our optimization problem used to involve much more than just deducing log-likelihood functions before utilizing APML.
Special effort had to be expanded to derive some components because of specific requirements from solving algorithms.
For example, to apply IPOPT solver to our model estimation, we were required to sketch out all parameter derivatives of log-likelihood functions and Jacobian matrix.
AMPL, however, allows us to solve the problem by only defining the problem and formulating the constraints.
To run AMPL on models, it needs two parts as input including model files and data files. Model files define the problem, while data files specify constants and initial values for variables. AMPL translator will read in those files and translate them into languages that solvers can understand.
AMPL is particularly notable for its general syntax, including variable definitions and data structures.
\subsection{Used solvers and optimization setup}
\label{subsec:optimization-setup}
AMPL supports a comprehensive set of solvers including solvers for linear programming, quadratic programming and non-linear programming~\cite{fourer1993ampl}.
This link\footnote{\url{http://www.ampl.com/solvers.html}} gives a full list of solvers for AMPL.
\textbf{Solvers Applied in Implementation}.
We used two solvers in our fitting procedure:
\begin{itemize}
\item \textbf{LGO}: a \emph{global optimizer} for non-linear problems, which is capable of finding approximate solutions when the problems have multiple local optimal solutions (\cite{pinter1997lgo}). This is also one of the default solvers provided by AMPL.
\item \textbf{IPOPT}: an open-source large-scale \emph{local optimizer} for non-linear programming, which is released in 2006~\cite{Wachter2006}.
\end{itemize}
Local solvers rely on improving an existing solution, employing complex techniques to avoid getting stuck in local minima.
They require an initial point from which to start exploring the space of solutions.
Global solvers attempt to search for the optimal solution in the entire space of solutions (one solution would be, for example, to divide the solution space into hyper-squares and apply local optimization in each one of them).
Global solvers tend to find solutions which are not too far from the optimal, but they lack the precision of specialized local solvers
In summary: local solvers achieve solutions very close to the optimal, but run the risk of getting stuck in horrible local optima;
global solvers achieve imprecise solutions close to the optimal.
\textbf{Optimization implementation setup}.
Our optimization setup is constructed to account for the weaknesses of each class of solvers.
A classical solution to the problem of local optima with local solvers is to repeat the function optimization multiple times, from different starting points.
We generate 8 random sets of initial parameters, within the definition range of parameters, and we use the IPOPT solver using each of these as initial point.
We also combine the global and the local solver: we use LGO to search in the space of solutions for an approximate solution, which we feed into IPOPT as initial point for further optimization.
Lastly, we run IPOPT without any initial parameters, leveraging IPOPT's internal strategy for choosing the starting point based on the parameters' range of definition.
After completing these 10 rounds of optimization, we select the solution with the maximum training log likelihood values.
This tends to be the combination of global and local optimizer (LGO + IPOPT).
\subsection{Interfacing AMPL with R}
\label{subsec:ampl-r-interface}
Our entire code base is using the R language, but AMPL has its own modeling language.
Therefore, we need to interface between R and AMPL.
Inspired by a blog post\footnote{\url{https://www.rmetrics.org/Rmetrics2AMPL}}, we implemented our own interface between AMPL and R language. The core ideas are described as follow:
\begin{itemize}
\item \textbf{Generating model files and data files}: one of the major components of this interface is to generate temporary model and data files, which express the problem to be solved and the used data into AMPL language.
As our experiments involve a large amount of cascades, we prefixed all temporal files with process ids so that running AMPL in parallel becomes possible.
\item \textbf{Interacting with AMPL}: this is also implemented by file I/O in the disk. After model files and data files are generated, we call AMPL via system commands through R and AMPL will then start optimization. Selections of solvers can be specified in the system commands used for starting AMPL. Optimization results will be saved in result files and our interface will extract results and return them.
\item \textbf{Exception handling}: this is another important component of this interface as solvers are easily encountering errors during optimization process (such as computing $log(0)$) when float number running out of precision.
\end{itemize}
\section{Relation between deterministic SIR and stochastic SIR}
\citet{Allen2008} analyzes in details the relation between the deterministic SIR and the stochastic SIR and shows that the mean behavior of the stochastic version converges asymptotically to the deterministic version.
She shows that the mean of the random function $I(t)$ in the stochastic SIR epidemic process is less than the solution $I(t)$ to the deterministic differential equation in~Eq.\eqref{eq:det-sir-size}.
We study the equivalence of the two flavors of SIR through simulation.
We simulate 100 realizations of the stochastic SIR and the deterministic SIR from the same set of parameters.
Fig.~\ref{fig:stochastic-vs-deterministic-SIR} shows the sizes of the population of Susceptible $S(t)$, Infected $I(t)$, Recovered $R(t)$ and the cumulated infected $C_t$.
For the stochastic version, we show the median and the 2.5\% / 97.5\% percentiles.
This result complements the analysis in Sec.~\ref{subsec:sir-model}.
\input{fig3-stochastic-vs-deterministic-sir}
\input{5-simulation-fitting}
\section{Robustness of fit -- additional graphics}
Fig.~\ref{fig:robustness-fit-extra-graphs} shows the robustness of fit for parameters $\kappa$, $\beta$ and $\theta$ for Hawkes \emph{(a)-(c)} and HawkesN \emph{(d)-(f)}.
This result complements Sec.~\ref{subsec:robustness-of-fit}.
\begin{figure}[tbp]
\centering
\newcommand\myheight{0.2}
\subfloat[] {
\includegraphics[page=1,height=\myheight\textheight]{robustness-HAWKES}
}
\subfloat[] {
\includegraphics[page=2,height=\myheight\textheight]{robustness-HAWKES}
}\\
\subfloat[] {
\includegraphics[page=3,height=\myheight\textheight]{robustness-HAWKES}
}
\subfloat[] {
\includegraphics[page=1,height=\myheight\textheight]{robustness-HAWKESN}
}\\
\subfloat[] {
\includegraphics[page=2,height=\myheight\textheight]{robustness-HAWKESN}
}
\subfloat[] {
\includegraphics[page=3,height=\myheight\textheight]{robustness-HAWKESN}
}
\caption{
Robustness of estimating parameters $\kappa$, $\beta$ and $\theta$ for Hawkes \emph{(a)-(c)} and HawkesN \emph{(d)-(f)}.
One set of parameters for each model was simulated 100 times and fitted on increasingly longer prefixes of each simulation.
One value for parameter is obtained for each fit and the median and the 15\%/85\% percentile values are shown.
}
\label{fig:robustness-fit-extra-graphs}
\eqmoveup\vspace{-0mm}} %{\vspace{-2.4mm}
\end{figure}
\section{Generalization performance -- Hawkes}
Fig.~\ref{fig:news1k-increasing-perc} shows the generalization performance of Hawkes, for increasing amounts of data.
Each cascade in a random sample of 1000 cascades in {\sc News}\xspace is observed for increasing periods of time.
This result complements the analysis in Sec.~\ref{subsec:explain-holdout}.
\begin{figure}[tbp]
\centering
\newcommand\myheight{0.18}
\includegraphics[page=1,width=0.45\textwidth]{News1k-holdout-ll-increasing-perc}
\caption{
Performances of Hawkes explaining unobserved data, using holdout negative log likelihood.
The performance over 1000 randomly sampled cascades in {\sc News}\xspace are summarized using boxplots, lower is better.
The percentage of observed events in each cascade used to train Hawkes is varied between 10\% and 95\%.
}
\label{fig:news1k-increasing-perc}
\eqmoveup\vspace{-0mm}} %{\vspace{-2.4mm}
\end{figure}
|
\section{Introduction}
A feature of interaction-based Internet communities is that direct connections and relationships between users do not have a significant influence on assessing their reputations. Rather, the most influential aspect for such an assessment is the behavior and the activities of the users within a digital community. The computation of user reputation and the assessment of user rating are directly connected because reputation is used for comparing users and at the same time rating is based upon that comparison. Rating systems are used in Internet communities where people communicate with each other, share opinions, information as well as find new contacts. One type of Internet community include web-sites where questioning and answering (Q\&A) is possible. Examples include \emph{Ask.fm} and \emph{Yahoo! answers} that allow users to ask questions on a wide range of topics. Other examples include platforms such as \emph{StackOverflow (SO)} that focuses on more narrow topics as computer science. Q\&A sites are built upon the notion of community contributions. Here, users generate content by asking specific questions to the community. In turn, other users of the same community can answers them, thus generating peer-reviewed content. The quality of this content depends mainly on the human expertise and knowledge. Hence an open problem is how to assess the level of expertise of users. StackOverflow has its own model for the assessment of the reputation of its users. This is mainly based upon a voting mechanism that allows users to recommend (like) or disapprove (dislike) the quality of questions or answers. This mechanism helps to determine the expertise and reputation of each user within the community. Here, reputation is an integer value from zero to infinity. As a consequence, users can be ordered and compared by this reputation value.
This study is focused on the investigation of how dynamic factors - factors that add dynamism to reputation, can be successfully used for rating users. The hypothesis is that dynamic aspects such as past activity, cumulative past knowledge and forgetting (inactivity) can be meaningfully used in computing the reputation of users as well as their trustworthiness in interaction-based Internet communities. This hypothesis is exploited with the data generated by the StackOverflow platform.\\
The paper is organized as follows: section 2 describes related works on reputation and trust. Section 3 focuses on the design of a novel model of reputation, called DIB-RM, employing dynamic factors. Section 4 evaluates this model highlighting the impacts of the dynamic factors on the assessment of reputation. Eventually, section 5 concludes the study and presents future work.
\section{Related Work}
\subsection{Reputation and trust}
Trust can be defined by person's positive or negative expectations of another person's actions. Reputation is a collective measure of trustworthiness based on the referrals or ratings from members of a community.
In \cite{rakoczy2016users}, authors systematize knowledge about trust and reputation. They highlight the problem that many researchers use these terms as equal and therefore they explain and separate them. Authors propose the schema depicted in Fig. \ref{fig:trust} which shows the hierarchy of trust types. Reputation is a type of trust called "Global trust". The first level classification is based on the number of people who participate in trust evaluation:
\begin{enumerate}
\item Local trust - trust which exists between two people.
\item Global trust - trust is the resultant of deposing of the many users' opinions towards one particular user.
\end{enumerate}
Another separation is performed by a method of collecting information:
\begin{enumerate}
\item Explicit - the value is directly given by users.
\item Implicit - the value is based on users activity and interaction, according to available data and made assumptions.
\end{enumerate}
The concept of trust has been investigate thoroughly, and several properties were defined: \textit{context-specific, dynamic, transitive, asymmetric, direction}. As mentioned in \cite{rakoczy2016users}, reputation has only three of them:
\begin{itemize}
\item \textit{Context-specific}. Reputation can be different between the same units of a system in a different scope. Rousseau discussed this specific nature of trust in social and psychological sciences \cite{rousseau1998not}.
\item \textit{Dynamic}. Chang E. \cite{chang2005fuzzy} describes this property in a way that reputation changes on time perspective continuously. Also, new interactions have more influence on reputation value because they are more relevant and important than old ones. A lot of techniques have been invented and they implement this concept \cite{kamvar2003eigentrust}, \cite{zhang2007fine}, \cite{zhou2007powertrust}.
\item \textit{Transitive}. This is the most common property which is widely used in several models. The reputation of a person depends on indirect connections of other people. There are several examples \cite{sabater2002reputation}, \cite{mui2002computational}.
\end{itemize}
Non-commercial trust-based platform have been proposed in the past \cite{MBGDMQN2013}. However, temporal factors have been rarely used as an exclusive factor in the computation of trust.
\subsection{Reputation models}
Nowadays, size of Internet communities increases, more and more people around the world connect to different platforms, such as Facebook, MySpace or Twitter. However, the users meet many problems related to trust. For example, a user needs to know a level of trustworthiness of a service provider or a product supplier before making a choice, or evaluate in new person before accepting his/her request. \cite{hamdi2016computational} Due to the incredible growth of social networks, researchers give their attention to trust and reputation management problems. Measurement of trust in social networks is based on several principles.
Wanita Sherchan separates reputation models into three groups \cite{Sherchan:2013:STS:2501654.2501661}:
\begin{enumerate}
\item Network Structure/Graph-Based models.
\item Interaction-Based models.
\item Hybrid models.
\end{enumerate}
This separation is based on the type of technique which is used in the model.
Models which have network structure use the concept of "Web of trust" or FOAF (Friend-Of-A-Friend). This concept uses "Transitivity" property and direct connections among people to evaluate the trust value between two people. Kutter et Golbeck \cite{kuter2007sunny} invented their model for calculating inference trust in social networks which are called SUNNY. Jiang and Wang \cite{jiang2011swtrust} proposed SWTrust algorithm, it generates a small graph from a big online social network (OSN). Authors in \cite{golbeck2006generating} presented a model which provides a movie recommendation and it is based on an average score of users ratings of films. However, this type of models does not take into account interactions between members. The activity of users and the nature of their communications particularly affect the trust or reputation value.\\
\begin{figure}[!t]
\centering
\includegraphics[width=0.47\textwidth]{Trust_hierarchy}
\caption{Taxonomy of trust}
\label{fig:trust}
\end{figure}
In the previous paragraphs graph based algorithms were mentioned. In contrast to these models, some trust models consider only interactions between system nodes. The name of the group is interaction-based models. Liu et al. actively use in \cite{liu2008predicting} interactions between users in online platforms for predicting trust value. They take into account two groups of parameters: metrics of user's activity with data such as frequency of reviews and ratings and taxonomy of different connections between two users. Kamvar et al. \cite{kamvar2003eigentrust} propose EigenTrust algorithm which performs reputation evaluation on history and state of interactions with the system. It uses aging to differentiate importance of new interactions and old ones.
Hybrid models combine graph structure of system and interactions between units of that system. Anupam et. al provides "SecuredTrust" model \cite{das2012securedtrust} which evaluates trust between multi-agent system units for load balancing and finding malicious agents. This model accounts for a historical information that does not allow malicious units to change their trust value in short period of time. They also implement decreasing of trust value of previous interactions that increases the influence of current activity of the unit. \\
Longo et al. in \cite{longo2007temporal} check hypothesis that temporal based factors, such as activity, frequency, regularity and presence, can be used as an evidence of an entity's trustworthiness. They introduce new algorithm and provide tests on Wikipedia database, there are 12000 users and 94000 articles. They compared prediction metrics with Wikipedia ratings and had satisfactory results. Good prediction rate was 60\%, bad prediction rate was less than 20\%, so this approach can be useful in trust measurement and can be aggregated with more traditional methods. The main drawback of using temporal factors is the amount of information required. A lot of data is needed to evaluate the trustworthiness of article and compare them to each other, because interactions are distributed on time interval where the article exists.
The same author proposed a methodology to continuously align a trust model in force with the changing context within dynamic applications such as forums, blogs, p2p systems. The self-adaptation is reflected in the auto-organisation of the trust function aimed at assessing an agents' trustworthiness \cite{longo2009enabling}.
Adali et al. evaluated trust in a social network, which is based on interaction behavior between two users and propagation of messages of each other \cite{adali2010measuring}. The first feature is called conversation trust, the second - propagative trust. These trust metrics depends only on communication traffic stream, so models are interaction-based ones. Only information about sender, receiver and time parameters of messages were used. Authors investigated the relevance of using this features on Twitter social network database. They divide messages into several sets by proximity of time. These sets are called conversations. Long conversations are also more confidently balanced conversations. Propagative trust is higher if users share messages to third parties.
Several models were designed for trust and reputation evaluation. They solve different problems from implementing recommendation system to reaching the high quality of service and system load balancing.
\subsection{Research question and hypothesis}
Some researchers improve models by making them more complex and harder in computation to achieve better results. On another hand, some of them try to create more simple models without significant decreasing of results but with better performance. We select the second approach. So, if reputation model based on interactions will give suitable results, implementing a reputation system which needs to store additional data and requires to create and manage new logic is redundant. The research question of this paper is:
\emph{To what extent a model, built upon dynamic interaction factors, can approximate subjective voting of users within the StackOverflow community?
}
\section{Dynamic Interaction Based Reputation Model}
\textit{Trust} can be seen as the amount of interaction among people: the more interaction occurs between two individuals the more one of them trust the other. This makes \textit{trust} very unstable, it actually changes continuously over time \cite{multidisciplinary}. We introduce Dynamic Interaction Based Reputation Model (DIB-RM), model that captures this dynamic property of trust.
DIB-RM is an interaction-based model among users of a community over time. The model computes a reputation value for each user on the system combining different factors: forgetting factor, the continuous decrease of reputation of an individual; cumulative factor, the importance of users' activities; and activity period factor, the period of time in which the change in the reputation value happened.
DIB-RM updates the reputation value of each interaction using a fixed number of parameters. This removes the need for storing information about previous interactions. Also, it works in dynamic environments that means a model can update the reputation value of users while they provide some action.
The following sections explain the assumptions made by the model, the mathematical background for DIB-RM and the metrics used to test the hypothesis.
\subsection{Trust Properties}
DIB-RM is built upon the following two properties of trust behavior:
\begin{enumerate}
\item if two individuals have not interactions for a long period of time, the trust level between them starts to decrease;
\item if two individuals interact very frequently and regularly, the trust level between them should increase faster than when they communicate rarely.
\end{enumerate}
The first property is based on the dynamic property of trust. It requires the continuous change of trust levels over time. The second property comes from \cite{multidisciplinary}. Authors use ``fragile trust" concept to represent that trust levels can change rapidly during short period of time depending on the activity of the user.
\subsection{Model description}
In Internet Communities, interactions occur when there is an activity between two individuals. As an example, in \textit{StackOverflow} there is an interaction between a user and the system when a user posts a question, or between users when a user answers an already posted question.
Interactions in DIB-RM are modeled by $I_n$
$$I_n=I_{b_n}+I_{c_n}$$
where $n \in 0\ldots N$ is the index of the interaction and $N$ is the total number of interactions of a user. $I_n$ contains a time stamp, when the interaction takes place, and
a value that describes the contribution to the reputation. They can be enumerated by time stamp to form historical chain of user's activity.
Interactions have different effects to the trust value. Each interaction has a basic value $I_{b_n}$.
Depending on the state of communication between a user and the system characterized by activity and frequency, an interaction can be perceived differently. $I_{c_n}$ capture the cumulative part of the interaction, the second property of trust held by DIB-RM.
It is defined as:
$$I_{c_n}=I_{b_n}*\alpha*(1-\frac{1}{A_{n}+1})$$
where $\alpha$ is the weight of the cumulative part. It shows how big $I_{c_n}$ can grow (if $\alpha = 1$ then $I_{c_n} \in 0\ldots I_{b_n}$). $A_{n}$ is the number of sequential activity periods.
Figure \ref{fig:interaction} depicts the dependency of interaction values from a number of activity periods for different weights of $\alpha$ and $I_{b} = 2$.
So for $\alpha=1$ $I_{c}$ can be maximum 2, for $\alpha=2$ - maximum 4, for $\alpha=3$ - maximum 6.
\begin{figure}[!t]
\centering
\includegraphics[width=0.3\textwidth]{Interaction_graph}
\caption{Interaction value graph for different $\alpha$, weight of ``cumulative" effect.}
\label{fig:interaction}
\end{figure}
Social communities have different contexts and features that affect the properties of the system. One of these properties is the frequency of user communication which is defined as the period of time between the last two activities. DIB-RM models this property as $t_a$. As an example, $t_a$ for Wikipedia can be one week, when a user creates or edits some article whereas for StackOverflow it can be one day, when user answers to a question.
$$\Delta_n=[\frac{t_n - t_{n-1}}{t_a}]$$
is the number of periods between the 2 last interactions. If the difference between $t_n$ and $t_{n-1}$ is less than $t_a$ the number of activity periods will increase by one. It means, the user continues to communicate frequently.
The final formula for trust is
$$T_n = T_{n-1}*\beta^{\Delta_n} + I_n, \beta \in [0,1]$$
where $\beta$ is the forgetting factor that is chosen by each system individually. If $\beta$ is close to 1, the trust value decreases.
Also, if save DIB-RM reputation values of a user for each day and represent results as a graph, it will look like the line which is depicted in Fig. $\ref{fig:dynamic_reputation}$. Another parameter which can be calculated is the sum of previous reputation values. This parameter is close to a value of an area which is under the graph line. We also use it to compare DIB-RM and StackOverflow model because it accumulates historical information about user's reputation. Even if a user currently has low reputation value but was very active before and done a lot of operations, the sum can be high in comparison with other users. We call this parameter $historical$ reputation.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{dynamic_reputation}
\caption{DIB-RM user reputation graph}
\label{fig:dynamic_reputation}
\end{figure}
In order to achieve objective results several components should be presented. On one hand it is a reputation model, on another hand, it is data which will be used for evaluations.
\subsection{Metric of approximation}
Reputation values mean nothing in isolation. It is a relative value used for comparison of users. In general, if the reputation value of user $A$ is higher than the reputation of user $B$, the trustworthiness of user $A$ is also higher.
To measure the efficiency of DIB-RM, we applied the model StackOverflow and compared the results to StackOverflow's own rating system. The results of this comparison will give information about how DIB-RM approximates StackOverflow voting system. The metric is defined as
$$\mu_D= 1 - \frac{1}{N^2}*\sum^{N}_{i=1}{(\frac{1}{D}*\sum^{D}_{j=1}{|R_{S_{ij}} - R_{D_{ij}}|})}$$
where $N$ is the number of users, $D$ the number of days between first and last dates, $R_{S_{ij}}$ the StackOverflow reputation value of user $i$ on day $j$ and $R_{D_{ij}}$ is the DIB-RM reputation value of user $i$ on day $j$. $|R_{S_{ij}} - R_{D_{ij}}|$ is the absolute difference between rating places of individual $i$ on particular day $j$. This value shows how close DIB-RM rating is to StackOverflow. Then we calculate average difference of ratings for user $i$ $\frac{1}{D}*\sum^{D}_{j=1}{|R_{S_{ij}} - R_{D_{ij}}|}$ in all-days period. It allows to avoid focusing on one estimation and analyze general behavior of the model. After that average difference of all users is estimated. The last step is subtracting from 1 the average difference, which divided to the number of rating places $N$, gives information about how DIB-RM rating system close to StackOverflow's one.
Another approach is measuring rating of users by historical reputation value. The formula of metric remains the same but instead of $R_{D_{ij}}$ (reputation rating place of user $i$ on day $j$) $R_{H_{ij}}$ (historical reputation rating) is used.
$$\mu_H= 1 - \frac{1}{N^2}*\sum^{N}_{i=1}{(\frac{1}{D}*\sum^{D}_{j=1}{|R_{S_{ij}} - R_{H_{ij}}|})}$$
Moreover, error of metric should be estimated to have clear picture of DIB-RM working. If the model has a small error, it gives expected results. Error estimation is performed by calculating standard deviation of metric, $\mu$. For reputation it is $\sigma_D$, for historical reputation it is $\sigma_H$.
\section{Evaluation and Discussion}
We used StackOverflow to evaluate DIB-RM. StackOverflow defines its own users reputation system. We used DIB-RM to evaluate the reputation of users based on interaction. Then, we calculated their difference using the metrics defined in the previous section. The StackOverflow database is available online and can be downloaded from an open access repository. This resource provides $xml$ dumps for all needed files about posts, posts' history, posts' links, comments, users, votes, badges among others. For the computation of DIB-RM and StackOverflow reputations we need posts, comments, users and votes. This is because other entities contain only details about interactions, for example, a post's history stores texts of questions and answers. This dataset includes history of user's activity from September $15^{th}$ 2008, launch day of StackOverflow, to September $14^{th}$ 2012.
\subsection{Analysis of used data}
We build a program which internal structure is shown in Fig. \ref{fig:DIB-RM}. Information which is contained in $xml$ files can be represented in a form of a table. Therefore we performed conversion from $ xml$ to $csv$ format because it can be manged by programing tools that we used for creating DIB-RM model. We wrote a parser which optimized to generate output results. It operates only with required fields without converting all file to about$csv$.
The next step is creating internal structure of data from $csv$ files that provides the DIB-RM fast access to information of interactions. By interactions we consider both posts and comments because they show activity of a user and his/her contribution to the system. Post is a general concept of content which users produce. It can be of two types: question or answer. In this paper we do not distinguish types of interactions and assign the same reputation value to them. A typical post tuple is $<PostId, CreationDate, PostTypeId, ParentId, UserId>$, a typical tuple of comment is $<CommentId, CreationDate, UserId>$.
\begin{enumerate}
\item PostId, CommentId - positive integer which represent unique identifier of entity.
\item CreationDate - date and time when post or comment was created.
\item UserId - positive integer which represents unique identifier of the user who is author.
\end{enumerate}
Those two tuples have similar domains so we can store them together. Sorting of interactions dataset by $(UserId, CreationDate)$ key pair will give historical sequence for each user. We do not add votes as interactions to DIB-RM because the purpose is to compare it with StackOverflow model which is based on voting system.
Vote entities are required to make simulation of StackOverflow model. We created a program which is fully based on rules of calculating users' reputation in StackOverflow. Votes' tuple has the structure $<VoteId, CreatinDate, VoteTypeId, PostId, UserId>$.
\begin{enumerate}
\item VoteId - positive integer which represent unique identifier of vote.
\item PostId - positive integer which represent unique identifier of post. Vote is related to this post.
\item VoteTypeId - positive integer which represent type of vote. It can have a value in the range from 0 to 9.
\item CreationDate - date and time when vote was created.
\end{enumerate}
Each post has a $UserId$ attribute and we can connect a vote with its' recipient and change his reputation.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{DIB-RM}
\caption{DIB-RM structure.}
\label{fig:DIB-RM}
\end{figure}
Total amount of users that we used for computations is 15.000. Between the minimum and maximum StackOverflow's reputations we identified 10 equal intervals and extracted 1.500 users from each group. This method allows to have a representative set of users. During 4 years those users generated 8.630.000 posts, 16.067.000 comments and received 33.269.000 votes.
\subsection{Activity period factor}
The first step of our experimentation is to understand the
importance of the $t_a$ constant (activity period) essential for computing value of "cumulative" part of interaction. We performed a set of computations changing the $t_a$ constant (in days) obtaining ratings difference averages and standard deviation values shown in table \ref{table:1} and in table \ref{table:2}. DIB-RM model has three factors which can be changed: $t_a$, $\alpha$ (cumulative factor), $\beta$ (forgetting factor). Computations are provided with fixed $\alpha=1$ and $\beta=0.99$ for both parameters reputation and historical reputation.
The results which are provided in tables \ref{table:1} and \ref{table:2} show that if $t_a$ (activity period) increases, metric value also increases for both parameters. It comes from the nature of the StackOverflow model which calculates reputation by adding value of a new vote to the sum of previous ones and does not decrease over time. So when $t_a$ increases, reputation value starts to decrease after a longer period of time, users have wider window to interact and increase cumulative part of interactions' reputation value. That means reputation keeps almost the same or increases to the high value of interaction because cumulative part also decreases less often. It makes reputation lose a dynamic property and become static as the StackOverflow model.
Historical reputation does not decrease as StackOverflow reputation also and contains values of all interactions. That is why historical reputation approximates better than the StackOverflow's model reputation value. Values of metric in table \ref{table:2} are between 0.88 and 0.882. That means increasing of $t_a$ does not have significant influence on metric results.
\begin{table}[!t]
\begin{center}
\begin{tabular}{ | l | l | l | l |}
\hline
$\#$ & $t_a$ & $\mu_D$ & $\sigma_D$ \\ \hline
1 & 1 & 0,8122 & 0,1100 \\ \hline
2 & 2 & 0,8313 & 0,0936 \\ \hline
3 & 4 & 0,8510 & 0,0744 \\ \hline
4 & 8 & 0,8605 & 0,0604 \\ \hline
\end{tabular}
\end{center}
\caption{Table of reputation metric results for different $t_a$ values}
\label{table:1}
\end{table}
\begin{table}[!t]
\begin{center}
\begin{tabular}{ | l | l | l | l |}
\hline
$\#$ & $t_a$ & $\mu_H$ & $\sigma_H$ \\ \hline
1 & 1 & 0,8816 & 0,0128 \\ \hline
2 & 2 & 0,8805 & 0,0026 \\ \hline
3 & 4 & 0,8813 & 0,0086 \\ \hline
4 & 8 & 0,8808 & 0,0021 \\ \hline
\end{tabular}
\end{center}
\caption{Table of historical reputation metric results for different $t_a$ values}
\label{table:2}
\end{table}
Four graphs are depicted in Fig. \ref{fig:reputation_graph_set} for different $t_a$ (activity period) values. They show reputation changes for two users over time. The red line belongs to user with $id = 300$, the blue line to user with $id =235$. Comparison of these lines shows that high value of $t_a$ increases the distance between reputation values. At first sub-graph user $235$ has four times has greater reputation than user $300$. In one period of time in range from 800'th day to 1100'th day blue line is higher than red line. However, on the fourth sub-graph, where $t_a$ parameter equals to 8, the blue line has higher value just at the beginning.
\subsection{Forgetting factor}
In this section we analyze the influence of forgetting factor to metric results and to reputation value. Forgetting factor is used to decrease importance of previous interactions, so new ones have more influence to a reputation. We use two forgetting factor values $\beta=0.99$ and $\beta=0.9$ that means a reputation reduces to 1\% or to 10\% for each activity period. Hence a combination of forgetting factor and activity period factor is also important. The results of computations are presented in tables \ref{table:3} and \ref{table:4}.
We provide metric values for four cases where $\alpha$ is fixed and equals to 1, $t_a$ has two variants, 2 and 8, and $\beta$ equals to 0.99 and 0.90. Increasing the forgetting factor leads to raising of the metric value that means previous interactions' values are also important at reputation evaluation.
If rating of users is built on DIB-RM reputation, changing a $\beta$ value has significant influence to metric. In case of $t_a=2$ $\mu_D$ grows from 0.79 to 0.83 when $t_a=8$ $\mu_D$ grows from 0.81 to 0.86.
\begin{figure}[!t]
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=1\linewidth]{D_rep12_8}
\caption{$t_a=1$}
\label{fig:d_rep1}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=1\linewidth]{D_rep12_2}
\caption{$t_a=2$}
\label{fig:d_rep2}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=1\linewidth]{D_rep12_4}
\caption{$t_a=4$}
\label{fig:d_rep4}
\end{subfigure}
\begin{subfigure}{0.24\textwidth}
\includegraphics[width=1\linewidth]{D_rep12_81}
\caption{$t_a=8$}
\label{fig:d_rep8}
\end{subfigure}
\caption{Dynamic reputation graph for different $t_a$}
\label{fig:reputation_graph_set}
\end{figure}
\begin{table}[h]
\begin{center}
\begin{tabular}{ | l | l | l | l | l |}
\hline
$\#$ & $t_a$ & $\beta$ & $\mu_D$ & $\sigma_D$ \\ \hline
1 & 2 & 0.90 & 0,7900 & 0,1285 \\ \hline
2 & 2 & 0.99 & 0,8313 & 0,0936 \\ \hline
3 & 8 & 0.90 & 0,8193 & 0,0992 \\ \hline
4 & 8 & 0.99 & 0,8605 & 0,0604 \\ \hline
\end{tabular}
\end{center}
\caption{Table of reputation metric results for different $\beta$ values}
\label{table:3}
\end{table}
\begin{table}[h]
\begin{center}
\begin{tabular}{ | l | l | l | l | l |}
\hline
$\#$ & $t_a$ & $\beta$ & $\mu_H$ & $\sigma_H$ \\ \hline
1 & 2 & 0.90 & 0,8803 & 0,0024 \\ \hline
2 & 2 & 0.99 & 0,8805 & 0,0026 \\ \hline
3 & 8 & 0.90 & 0,8808 & 0,0023 \\ \hline
4 & 8 & 0.99 & 0,8808 & 0,0021 \\ \hline
\end{tabular}
\end{center}
\caption{Table of historical reputation metric results for different $\beta$ values}
\label{table:4}
\end{table}
\subsection{Cumulative factor}
Cumulative factor $\alpha$ represents the proportion of basic part and cumulative part of interaction. Cumulative part directly depends on the activity of a user. If a user sequentially performs interactions that have an interval between each other less than the activity period, the value of cumulative part increases. We provide evaluation for four cases when $t_a=2$, $\beta=0.99$, $\alpha=\{1, 2, 4, 8\}$. The result values are shown in tables \ref{table:5} and \ref{table:6}.
\begin{table}[!t]
\begin{center}
\begin{tabular}{ | l | l | l | l | l |}
\hline
$\#$ & $\alpha$ & $\mu_D$ & $\sigma_D$ \\ \hline
1 & 1 & 0,8313 & 0,0936 \\ \hline
2 & 2 & 0,8441 & 0,0774 \\ \hline
3 & 4 & 0,8426 & 0,0814 \\ \hline
4 & 8 & 0,8515 & 0,0723 \\ \hline
\end{tabular}
\end{center}
\caption{Table of reputation metric results for different $\alpha$ values}
\label{table:5}
\end{table}
\begin{table}[!t]
\begin{center}
\begin{tabular}{ | l | l | l | l | l |}
\hline
$\#$ & $\alpha$ & $\mu_H$ & $\sigma_H$ \\ \hline
1 & 1 & 0,8805 & 0,0026 \\ \hline
2 & 2 & 0,8806 & 0,0025 \\ \hline
3 & 4 & 0,8808 & 0,0030 \\ \hline
4 & 8 & 0,8808 & 0,0023 \\ \hline
\end{tabular}
\end{center}
\caption{Table of historical reputation metric results for different $\alpha$ values}
\label{table:6}
\end{table}
\section{Conclusion and Future Work}
In this paper, we investigated the usage of dynamic factors for reputation evaluation. We formally defined reputation model which combines all factors: forgetting factor, cumulative factor and active period factor. Our evaluation was performed in the context of StackOverflow web site. Results based on 4 year history, covering 15.000 users, more than 8.000.000 posts and 33.000.000 votes. We tested our factors and hypothesis by comparing ratings of users that are created by DIB-RM and StackOverflow model. We used two values for creating ratings: reputation and historical reputation. Historical reputation value gave better results, around 88\% similarity between DIB-RM and StackOverflow ratings. Results of evaluation show that this value is resistant to factors' changes, so it allow to adopt model to some environments by selecting different values of factors without decreasing of metric value. We believe that this factors can be used as an evidence of users' trustworthiness in combination with more traditional ones. Our further works will be addressed to determining environments in which context dynamic factors can be used as a strong evidence of trustworthiness.
\bibliographystyle{plain}
|
\section{INTRODUCTION}
\label{sec:introduction}
Human detection and tracking is an important task in service robotics, where knowledge of human motion properties such as position, velocity and direction can be used to improve the behavior of the robot, for example to improve its collision avoidance and adapt its velocity to that of the surrounding people.
Using multiple sensors to track people has advantages over a single one.
The most obvious one is that multiple sensors can often do the task with a wider field of view and thus track more people within a larger range \cite{schulz03ijrr,linder16icra}.
Another advantage is that multiple sensors providing redundant information can increase tracking accuracy and reliability \cite{kobilarov06icra,held13icra,misu14ias,koide16iros}.
Different sensors have different properties.
The 3D LiDAR in our robot platform (Fig.~\ref{fig:flobot}) has 16 scan channels, 360$^{\circ}$ horizontal and 30$^{\circ}$ vertical fields of view, and up to 100~m range.
However, this sensor provides only sparse point clouds, from which human detection can be very difficult because some useful features, such as color and texture, are missing.
2D LiDARs have obviously similar problems, with further limitations due the availability of a single scan channel and reduced field of view.
However, these sensors are also cheaper than the previous, and have been used in mobile robotics long enough to stimulate the creation of many human detection algorithms~\cite{arras07icra,bellotto2009}.
RGB-D cameras, instead, can detect humans more reliably but only within short range and limited field of view~\cite{munaro14auro,jafari14icra}.
\begin{figure}[t]
\centering
\includegraphics[height=4.8cm]{figures/flobot.jpg}
\hfil
\includegraphics[height=4.8cm]{figures/flobot2.png}
\caption{Left: Pioneer 3-AT mobile robot with Velodyne VLP-16 3D LiDAR (1), ASUS Xtion PRO LIVE RGB-D camera (2), and HOKUYO UTM-30LX 2D LiDAR (3).
Right: visualized sensor data (only depth information is shown for RGB-D camera) annotated with corresponding sensor numbers.
}
\label{fig:flobot}
\end{figure}
In the literature, several algorithms already exist which can reliably detect human under particular conditions and with specific sensors (e.g. close range RGB-D detection).
Other sensors, however, are not yet so popular to benefit from good human detection software (e.g. 3D~LiDAR).
In some cases, there are simply not enough datasets with such sensors to learn robust human classifiers for many real-world applications.
In this paper, therefore, we wish to train a 3D~LiDAR-based human classifier in a semi-supervised way by learning from existing RGB-D and 2D~LiDAR-based detectors.
Although better human detectors ultimately lead to better people tracking systems, here we focus on the first part only and leave the second for future work.
Typically, data collection and training of the classifier are done offline, with obvious labor cost and potential human errors.
In the proposed transfer learning framework, instead, a 3D~LiDAR-based human classifier is trained online while exploiting spatio-temporal information from the tracking sub-system, which uses static (i.e. pre-trained) human detectors for the 2D~LiDAR and RGB-D sensors.
This framework enables a new sensor to learn from trajectories of the tracked people, each one with an associated confidence, or probability, of being human-generated, and fusing different model-based (labeled) and model-free (unlabeled) detections according to a semi-supervised learning scheme~\cite{zhu2009introduction}.
In contrast to previous approaches~\cite{Yan2017,Teichman2012}, our solution does not need any hand-labeled data, performing online learning completely from scratch.
Besides reducing the burden of data annotation, this feature makes our system easily adaptable to the environment where the robot is deployed.
The contributions of this paper can be summarized as follows:
$1)$ we propose an online transfer learning framework for multisensor people tracking based on a new \emph{trajectory probability}, which takes into account both sensor independence (in the detection) and multisensor interaction (in the trajectory estimation);
$2)$ we present an experimental evaluation of our system for 3D~LiDAR-based human classification with a mobile robot on a real-world dataset using different sensor combinations.
The remainder of the paper is organized as follows:
Section~\ref{sec:related_work} provides an overview of relevant literature in human detection and tracking.
Section~\ref{sec:solution_framework} presents our solution framework for online transfer learning.
Section~\ref{sec:case_study} describes the application of the proposed framework to the problem of 3D~LiDAR-based human classification.
Section~\ref{sec:evaluation} illustrates the experimental results for different sensor configurations.
Finally, Section~\ref{sec:conclusions} concludes the paper summarizing the contributions and suggesting future research work.
\section{RELATED WORK}
\label{sec:related_work}
The problem of multitarget and multisensor tracking has been extensively studied during the past few decades.
Most of present systems are based on Bayesian methods~\cite{shalom95multitarget}, which compute an estimate of the correspondence between features detected in the sensor data and the different humans to be tracked.
Regarding robotic applications, multiple sensors can be deployed in single- or multi-robot systems~\cite{yz13ijars}, while the former is the concern of this paper.
RGB/RGB-D camera plus 2D LiDAR is the most frequently used combination in the literature.
\cite{kobilarov06icra} presented two different methods for mobile robot tracking and following of a fast-moving person in an outdoor environment.
The robot was equipped with an omnidirectional camera and a 2D LiDAR.
\cite{spinello09ijrr} presented an integrated system to detect and track people and cars in outdoor scenarios, based on the information retrieved from a camera and a 2D LiDAR on an autonomous car.
\cite{linder16icra} introduced a people tracking system for mobile robots in very crowded and dynamic environments.
Their system was evaluated with a robot equipped with two RGB-D cameras, a stereo camera and two 2D LiDARs.
Our previous work with the aforementioned sensor combination includes \cite{bellotto2009} and \cite{Dondrup2015a}.
The former presented a human tracking system for mobile service robots in populated environments, while the latter extended this system to a fully integrated perception pipeline for people detection and tracking in close vicinity to the robot.
The proposed tracking system tracks people by detecting legs extracted from a 2D LiDAR and fusing this with the faces or the upper-bodies detected with a camera using a sequential implementation of the Unscented Kalman Filter (UKF).
The combination with 3D LiDAR is increasing with the development of the 3D LiDAR technology.
Taking advantage of its high accuracy, \cite{held13icra} developed an algorithm to align 3D LiDAR data with high-resolution camera images obtained from five cameras, in order to accurately track moving vehicles.
Other reported results include \cite{premebida14iros} and \cite{gonzalez15iv}, which mainly focused on pedestrian detection rather than tracking.
In addition, earlier work presented multitarget tracking with a mobile robot equipped with two 2D LiDARs, respectively located at the front and back~\cite{schulz03ijrr}.
Thus the robot can have a 360$^{\circ}$ horizontal field of view, where each scan of these two sensors covers the whole surrounding of the robot at an angular resolution of 1$^{\circ}$.
The use of machine learning algorithms for tracking has particular advantages.
The closest work to ours is \cite{Teichman2012}, where the authors proposed a semi-supervised learning approach to the problem of track classification in 3D LiDAR data, based on Expectation-Maximization (EM) algorithm.
In contrast to our approach, their learning procedure needs a small set of seed tracks and a large set of background tracks, that need to be manually or semi-manually labeled at first, whereas we do not need any hand-labeled data.
To our knowledge, no existing work in the robotics field explicitly exploits information from multisensor-based tracking to implement transfer learning between different sensors as in this paper.
Our work combines the advantages of multiple sensors with the efficiency of semi-supervised learning, and integrates them into an single online framework applied to 3D~LiDAR-based human detection.
\section{ONLINE TRANSFER LEARNING}
\label{sec:solution_framework}
An overview of our solution framework for online transfer learning can be seen in Fig.~\ref{fig:framework}.
It contains four main components: static detectors denoted by $\mathcal{D}_s$, dynamic detectors $\mathcal{D}_d$, a target tracker $\mathcal{T}$ and a label generator $\mathcal{G}$.
In order to facilitate the explanation, we present each component following the sequence of an entire iteration, starting with human detection.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/framework.pdf}
\caption{Block diagram of the online transfer learning framework.}
\label{fig:framework}
\end{figure}
\subsection{Human Detection and Tracking}
\label{sec:detection_tracking}
$\mathcal{D}_s$ can detect humans with offline-trained or heuristic detectors, typically with high confidence, while $\mathcal{D}_d$ acquires this ability through the online framework.
Both detectors provide new observations for $\mathcal{T}$ and their corresponding probabilities for $\mathcal{G}$.
$\mathcal{D}_s$ provides labeled detections, while $\mathcal{D}_d$ provides both labeled and unlabeled ones.
Here we assume the initial training set is substituted, instead, by a transfer learning process between the initial $\mathcal{D}_s$ and the final $\mathcal{D}_d$.
The tracking process $\mathcal{T}$ gathers the observations, fuses them and generates human motion estimates.
Both moving and stationary targets are tracked.
For the latter, the trajectory length is supposed to be null or at least very small.
$\mathcal{T}$ associates human detections from different sensors to the same corresponding estimates, linking $\mathcal{D}_s$ and $\mathcal{D}_d$ detections and therefore making the transfer learning possible.
In order to enable this on a mobile robot with multiple sensors, $\mathcal{T}$ should:
$a)$ be robust to sensor noise and partial occlusions;
$b)$ fuse multisensor data;
$c)$ be able to deal with multiple targets simultaneously;
$d)$ cope with noise introduced by robot motion.
\subsection{Transfer Learning}
The label generator $\mathcal{G}$ fuses the information coming from $\mathcal{D}_s$, $\mathcal{D}_d$ and $\mathcal{T}$, then generates training labels for $\mathcal{D}_d$.
A \emph{trajectory probability} is measured by $\mathcal{G}$, based on Bayes' theorem.
The idea is to measure the likelihood that a trajectory belongs to a human, which is defined as follow.
Given an objectness proposal $x_i$ and its category label $y_i$, $P(y_i|x_i,d_j)$ denote the predictive probability that sample $x_i$ is a human observed by detector $d_j \in \mathcal{D}$ ($\mathcal{D} = \mathcal{D}_s \cup \mathcal{D}_d$) at time $t$.
For a whole trajectory of detections $X_T\{x_i\}$ and its category label $Y_T$, the predictive probability of the whole trajectory $P(Y_T|X_T,\mathcal{D})$ is computed by integrating the observations of the different detectors according to the following formula:
\begin{equation}
\label{eq:px}
\small
P(Y_T|X_T,\mathcal{D}) = \frac{odds_{X_{T}}}{1+odds_{X_{T}}}
\end{equation}
where
\begin{equation}
\label{eq:odds_prod}
\small
odds_{X_{T}} = \prod_{i=1}^{t} \prod_{j=1}^{K} odds_{x_{i}}^j
\end{equation}
and
\begin{equation}
\label{eq:odds}
\small
odds_{x_{i}}^j = \frac{P(y_i|x_i,d_j)}{1-P(y_i|x_i,d_j)}
\end{equation}
Here, $P(y_i|x_i,d_j)$ may be necessary to get through some transformations based on specific circumstances, that are illustrated in our application case (Section \ref{sec:static_detecors} and \ref{sec:dynamic_detecor}).
The above formula has been proved effective by theoretical and experimental analysis in generating grid occupancy map from different sensors \cite{burgard00icra,moravec88sensor},
since the odds method takes sensor interactions into account.
To the best of our knowledge, applying it to human trajectory analysis is new.
An intuitive approach to generate labels is to threshold the trajectory probability in Eq.~\ref{eq:px}.
Let $\sigma_t$ be a predefined threshold value, then:
\begin{equation}
\label{eq:track_label}
\small
\begin{split}
\forall x_i \in X_T & \text{~assign~} y_i = z (z \in \mathbb{Z}^+) \text{,} ~\text{get}~ X_T, Y_T = \{x_i,y_i\}_T\\
& \text{~if~} P(Y_T|X_T,\mathcal{D}) \geq \sigma_t ~~~~
\end{split}
\end{equation}
where $\mathbb{Z}^+$ refers to positive integer values.
Our framework uses a Batch Incremental Training (BIT) policy to learn a classifier in $\mathcal{D}_d$, because:
$a)$~the number of new generated training examples is always more than one;
$b)$~the batch of examples can be used to approximate the distribution of the whole examples in online learning~\cite{read12ida}.
For a certain interval, a new batch of training samples $[X_{new}, Y_{new}]$ can be generated:
\begin{equation}
\label{eq:batch}
\small
\begin{split}
[X_{new}, Y_{new}] = \{(x_{i}, y_{i}) ~|~ & x_{i} \in X_T \subseteq \mathbb{R}^n,\\
& y_{i} \in Y_T \subseteq \mathbb{Z}^+, 1 \leq i \leq n_i\}_{\forall T}
\end{split}
\end{equation}
where $X_T$ is obtained by the tracker and $Y_T$ is obtained from Eq.~\ref{eq:track_label}.
The classifier of $\mathcal{D}_d$ will be updated in the BIT procedure:
\begin{equation}
\small
\mathcal{D}_d^i = \text{BIT}([X_{new}, Y_{new}], \mathcal{D}_d^{i-1}))
\end{equation}
where subscript $i$ refers to the iteration rather than time.
\subsection{Convergence of the Learning Process}
The learning procedure converges when the number of correct detections output by $\mathcal{D}_d$ reaches a steady state:
\begin{equation}
\label{eq:halting}
\small
\lim_{I \to \infty}\frac{stability_I}{I} = 0
\end{equation}
where $I$ is the total number of re-training iterations.
In our system, the stability at iteration $i$ is defined as follows:
\begin{equation}
\label{eq:stability}
\small
stability(I) = \sum_{i=1}^{I} \|u_i-u_{i+1} \|
\end{equation}
\begin{equation}
\small
\begin{aligned}
u_i~=~&\sum_{N_i} y_i \cdot \psi (P(y|X_{valid}^i, \mathcal{D}_d^{i}))\\
& + (1-y_{i}) \cdot \psi (1-P(y|X_{valid}^{i}, \mathcal{D}_d^{i}))
\end{aligned}
\end{equation}
where $u_i$ is the number of correct classified examples, $X_{valid}$ is the validation set, $\psi$ is binary function via thresholding 0.5.
One can halt the learning process when the stability stops increasing, or other stopping conditions (e.g. number of iterations) are triggered.
\section{APPLICATION TO MULTISENSOR HUMAN DETECTION}
\label{sec:case_study}
In this section, we present an online transfer learning for human classification in 3D~LiDAR scans using the robot shown in Fig.~\ref{fig:flobot}.
The sensor configuration resembles the one adopted for an industrial floor washing robot developed by the EU project FLOBOT\footnote{\url{http://www.flobot.eu}} and, besides a 3D~LiDAR on the top, includes an RGB-D camera and a 2D~LiDAR mounted on the front.
We describe in the following paragraphs how to use the state-of-the-art detectors to train a 3D LiDAR-based human detector online, instead of training it offline using manually-labeled samples.
The detailed block diagram of our implementation can be seen in Fig.~\ref{fig:case}.
At each iteration, 3D LiDAR scans are first segmented into point clusters.
The 2D position and velocity of these clusters are estimated in real-time by a multitarget tracking system, which outputs the trajectories of all the clusters.
At the same time, a classifier is trained to classify the clusters as human or not, assigning a normalized confidence value to each of them.
This confidence is the predictive probability $P(y_i|x_i, d_j)$ for the 3D~LiDAR-based detector, which is needed for the calculation of the trajectory probability in Eq.\ref{eq:px}-\ref{eq:odds}.
The classifier is initialized and retrained online.
The trajectories and the probabilities are sent to a label generator, which generates the training labels for the next iteration.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/case.pdf}
\caption{Process details of the online transfer learning for human detection.}
\label{fig:case}
\end{figure}
The upper-body detector~\cite{jafari14icra} and the leg detector~\cite{arras07icra}, respectively based on the RGB-D camera and the 2D~LiDAR, are the static detectors $\mathcal{D}_s$.
Both enable human tracking by sending the position of the detections.
In addition, they provide the corresponding probabilities $P(y_i|x_i, d_j)$ (i.e. normalized detection confidence) to the label generator.
The combination of 3D-LiDAR-based cluster detector and the human classifier, instead, constitutes the dynamic detector $\mathcal{D}_d$ that we want retrain online.
For an intuitive understanding of the various detectors and their outputs, please refer to the example in Fig.~\ref{fig:screenshot}.
The following paragraphs describe each module in detail.
\begin{figure}[t]
\centering
\includegraphics[width=\columnwidth]{figures/screenshot.png}
\caption{A screenshot of our multisensor-based detection and tracking system in action.
The sparse colored dots represent the laser beams with reflected intensity from the 3D~LiDAR.
The white dots indicate the laser beams from the 2D~LiDAR.
The colored point clouds are RGB images projected on depth data of the RGB-D camera.
The robot is at the center of the 3D LiDAR beam rings.
The numbers are the tracking IDs and the colored lines represent the people trajectories generated by the tracker.
For example, the person with tracking ID~49 has been detected by the RGB-D upper-body detector (green cube), the 2D~LiDAR leg detector (green circle), and the 3D~LiDAR cluster detector (blue bounding box).}
\label{fig:screenshot}
\end{figure}
\subsection{Upper Body Detector and Leg Detector}
\label{sec:static_detecors}
The upper-body detector identifies upper-bodies (shoulders and head) in 2D range (depth) images, taking advantage of a pre-defined template.
The confidence of the detection is inversely proportional to the observation range.
The leg detector detects legs in 2D LiDAR scans based on 14 features, including the number of beams, circularity, radius, mean curvature, mean speed, and more.
Its detection performance is limited by moving and crowd people.
As the upper-body detector and leg detector are not probabilistic methods and for the sake of simplifying mathematical conversion, a probability of 0.5 is assigned if $x_i$ is detected as a human.
\subsection{Cluster Detector and Human Classifier}
\label{sec:dynamic_detecor}
The 3D~LiDAR-based cluster detector and the human classifier are originally from our recent work~\cite{Yan2017}, while the former has been incorporated in different problems \cite{ls18icra, ls18ral}.
As input of this module, a 3D LiDAR scan is first properly segmented into different clusters using an adaptive clustering approach.
The latter enables to use different optimal thresholds for point cloud clustering according to the scan ranges.
Then, a Support Vector Machine (SVM)-based classifier \cite{svm} with six features (a total of 61 dimensions) is trained online.
These features are selected to meet robots' requirements for real-time and online computing performance.
For more implementation details, please refer to~\cite{Yan2017}.
In our approach (based on LIBSVM~\cite{libsvm}), the uncalibrated error function of SVM is squashed into a logistic function (here is the sigmoid function) to get the predictive probability $P(y_i|x_i,d_j)$ (used in Eq.~\ref{eq:odds}).
To be more specific, a binary classifier (i.e. human or non-human) is trained at each iteration.
The ratio of positive to negative training samples is set to $1:1$, and all data are scaled to $[-1, 1]$, generating probability outputs and using a Gaussian Radial Basis Function kernel~\cite{rbf}.
Since LIBSVM does not currently support incremental learning, the system stores all the training samples accumulated from the beginning and retrains the entire classifier at each new iteration.
The solution framework, however, also allows for other classifiers and learning algorithms.
\subsection{Bayesian Tracker}
People tracking is performed by a robust multisensor-multitarget Bayesian tracker, which has been widely used and described in previous works~\cite{linder16icra,bellotto2009,Yan2017,Dondrup2015a,BayesianTracking}.
The estimation consists of two steps.
In the first step, a constant velocity model is used to predict the target state at time $t$ given the previous state at ${t-1}$.
In the second step, if one or more new observations are available from the detectors, the predicted states are updated using a Cartesian or a Polar observation model, depending on the type of the sensor.
An efficient implementation of Nearest Neighbor data association is finally included to resolve ambiguities and assign each person the correct detections, in case more than one are simultaneously generated by the same or different sensors. As reported in~\cite{linder16icra,Dondrup2015a,BayesianTracking}, our tracker fulfills the requirements listed in Sec.~\ref{sec:detection_tracking}.
We refer the reader to these publications for further details.
\subsection{Label Generator}
The positive training labels $X^+$ are generated according to Eq.~\ref{eq:track_label}, while the negatives $X^-$ are generated based on a volume filter:
\begin{equation}
\label{eq:volume_filter}
\small
\begin{aligned}
X^-~=~&\{x_i~|~\textsc{w}_i < 0.2,~\textsc{w}_i > 1.0,\\
& \textsc{d}_i < 0.2,~\textsc{d}_i > 1.0,~\textsc{h}_i < 0.2,~\textsc{h}_i > 2.0\}
\end{aligned}
\end{equation}
where $\textsc{w}_i$, $\textsc{d}_i$, $\textsc{h}_i$ are the width, depth and height of a 3D cluster $x_i$.
The idea is that clusters without a pre-defined human-like volumetric model will be considered as negative samples for the next training iteration.
In our application, the dynamic classifier was trained from scratch without any manually-labeled initial sets.
As the validation set is not available, the maximum number of iterations was used as halting criteria.
\section{EVALUATION}
\label{sec:evaluation}
\subsection{Dataset}
We evaluated our framework on a real-world dataset\footnote{\url{https://lcas.lincoln.ac.uk/wp/research/data-sets-software/l-cas-multisensor-people-dataset/}} collected in an indoor public area by the robot shown in Fig.~\ref{fig:flobot}.
The robot was running the Robot Operating System (ROS)~\cite{ros} and it was manually driven with a gamepad.
Several \emph{rosbag} files were recorded, and the total length of which is about 49 minutes, including two continuous recordings of 19 and 30 minutes.
Sensor data were recorded in their original frame of reference and the coordinate transformations were handled by the ROS \emph{tf} package.
\subsection{Experimental Setup}
The experiments were conducted on the 19 minutes segment of continuous data, in which our binary SVM human classifier was learned online.
The classifier was retrained once every 300 new positive (human) and 300 new negative (non-human) samples, \revised{labeled by the label generator}, corresponding to one iteration.
We report the results for the first seven iterations, collecting a total of 2,100 positive and 2,100 negative.
In addition, a classifier was trained offline, using 2,100 manually labeled positive samples with an equal amount of randomly selected negative samples, to serve as a baseline for comparison.
Furthermore, we arbitrarily selected 100 scan frames from the dataset and fully annotated these (including standing and sitting people) as a test set.
This contains 1,197 human labels with varying distances from the robot between few centimeters and twenty meters.
A detection was considered a true positive if the overlap between it and the ground truth was more than~50\%.
Our framework has been fully implemented within a modular ROS architecture.
All components are ready for download\footnote{\url{https://github.com/LCAS/online_learning/tree/multisensor}} and use by other researchers.
The dataset collection and all the experiments reported in this paper were carried out on the robot embedded PC, with an Intel i7-4785T processor and 8~GB memory, using Linux Ubuntu 14.04~LTS (64-bit) and ROS Indigo.
It is worth noting that our system is fast and cost effective, since it can learn a human detector within minutes and using only inexpensive CPUs, rather than training for hours or days with expensive GPUs.
\subsection{Human Classification}
We first evaluate the performance of the 3D~LiDAR-based human classification after every online training iteration.
We compare the results for all the possible sensor combinations: 3D~LiDAR only, i.e. without any knowledge transfer \revised{but learned from trajectories only}; 3D~LiDAR with RGB-D camera; 3D~LiDAR with 2D~LiDAR; 3D~LiDAR with both RGB-D camera and 2D~LiDAR.
We measure the average precision (AP)~\cite{voc}, rather than the classification accuracy (ACC) used in other methods~\cite{Teichman2012}, because more informative.
Indeed the number of true negatives in our binary classification was far larger than the number of true positives, leading to an ACC always higher than 80\% (with probability threshold $\sigma_t = 0.5$), for each training iteration and each sensor combination.
In all the experiments, the trajectory probability threshold $\sigma_t$ of Eq.~\ref{eq:track_label} was set to $0.7$.
The experimental results in Fig.~\ref{fig:detection_results}, show that the AP for ``3D~LiDAR only'' and (3D~LiDAR) ``with 2D~LiDAR'' increases with the iterations, while there are not significant changes ``with RGB-D camera''.
However, the results ``with RGB-D camera and 2D~LiDAR'' show an interesting trend, first decreasing until the 5th iteration, and then increasing well above any other combination.
This final outcome shows the advantage of our multisensor system, which can eventually improve the online transfer learning process.
\begin{figure}[t]
\centering
\includegraphics[width=0.92\columnwidth]{figures/detection.pdf}
\caption{Comparison of different 3D~LiDAR-based human classifiers trained offline (red line) and online using different sensor combinations. Best viewed in color.}
\label{fig:detection_results}
\end{figure}
We additionally evaluate the Precision-Recall of the offline and final (i.e. after the 7th iteration) online trained classifiers.
The experimental results are shown in Fig.~\ref{fig:classification_results}.
Once again, the combination of RGB-D camera and 2D~LiDAR achieves the best performance.
Differently from the ``3D LiDAR only'' case, which learned only from moving people, the solution ``with RGB-D camera and 2D LiDAR'' was able to learn from moving, standing and sitting people, which greatly improved the human classification performance.
It is also worth pointing out that, despite showing a relatively high precision, the true positive rate (recall) of the offline trained classifier is generally lower.
This is due to a lack of long-distance samples in the offline training set, which are difficult to label by a human annotator.
\begin{figure}[t]
\centering
\includegraphics[width=0.92\columnwidth]{figures/pr.pdf}
\caption{Performance evaluation of human detection. Best viewed in color.
\label{fig:classification_results}
\end{figure}
\section{CONCLUSION}\label{sec:conclusions}
In this paper, we presented a framework for online transfer learning, applied to 3D~LiDAR-based human classification, taking advantage of multisensor-based tracking.
The framework, which relies on the computation of human trajectory probabilities, enables a robot to learn a new human classifier over time with the help of existing human detectors.
To this end, we proposed a semi-supervised learning method, which fuses both model-based (labeled) and model-free (unlabeled) detections from different sensors.
A very promising feature of the proposed solution is that the new human classifier can be learned directly from the deployment environment, thus removing the dependence on pre-annotated data.
The experimental results, based on a real-world dataset, demonstrated the efficiency of our system.
The proposed framework has been fully implemented into ROS with a high level of modularity.
The software and the dataset are publicly available to the research community, with the intention to perform objective and systematic comparisons between the recognition capabilities of different robots.
Moreover, our framework is easy to extend to other sensors and moving objects, such as cars, bicycles and animals.
Despite these encouraging results, there are several aspects which could be improved.
For example, the AP of the online learned classifier is still relatively low, due to the complexity of the environment recorded in our dataset.
This can be further improved by using a more advanced model for negative sample generation.
In addition, it remains to be verified how a new human detector, based on the online trained classifier, will affect the stability of the system and its tracking performance.
\section*{ACKNOWLEDGMENT}
This work has received funding from the European Union's Horizon 2020 research and innovation programme under grant agreement No. 645376 (FLOBOT) and No. 732737 (ILIAD).
\bibliographystyle{IEEEtran}
|
\section{INTRODUCTION}\label{sec:intro}
Galactic scale gas outflows (or winds) are ubiquitous in star forming
galaxies \citep[see][for
reviews]{2005ARA&A..43..769V,2017arXiv170109062H} and believed to be
essential to distribution of the gas and metals in
galaxies and the circumgalactic/intergalactic medium (CGM/IGM) and hence
to regulating cosmic star formation history
\citep[see][for reviews]{2015ARA&A..53...51S,2017ARA&A..55...59N}.
Theoretical models of the stellar mass-halo mass relation
constructed by abundance matching of
observational stellar mass functions to
simulated halo mass functions
\citep[e.g.,][]{2013MNRAS.428.3121M,2013ApJ...770...57B,2017MNRAS.470..651R}
indicate
that galaxies (or dark matter halos) are very inefficient
at converting gas into stars. At low redshift, at most
$\sim10-20\%$ of the available
baryonic mass has been converted into stars at halo mass of
$\sim10^{12}\Msun$, while the ratio of stellar mass to halo mass
declines steeply toward both lower and higher masses. Recent
cosmological hydrodynamic simulations of large volumes of the Universe
require strong outflows driven by both star formation
and active galactic nuclei feedback to explain low baryonic abundance
in galaxies compared to the cosmic fraction
\citep[e.g.,][]{2003MNRAS.339..289S,2014Natur.509..177V,2015MNRAS.446..521S}.
Direct inclusion of feedback processes in large-volume cosmological
galaxy formation simulations is still not feasible in practice.
For star formation feedback by supernovae (SNe),
implementation via simple thermal energy dumps suffers
``overcooling,'' with energy radiated away without
preventing in situ star formation or driving winds
\citep[e.g.,][]{1992ApJ...391..502K}; this is because resolving
the cooling radii of SN remnants requires much higher resolution
\citep{2015ApJ...802...99K} than is practicable in large-volume
simulations. In cosmological zoom-in simulations,
more careful treatments of SNe allowing for a
``momentum prescription'' at low resolution
can solve at least some aspects of the overcooling problem
\citep[e.g.,][]{2014ApJ...788..121K,2015MNRAS.451.2900K,
2014MNRAS.445..581H,2017arXiv170206148H}, especially for dwarfs.
However, given the constraints of computational expense,
treating unresolved physics with parameterized models
is unavoidable in many situations,
including in simulations of galaxy groups/clusters, and
in the large boxes needed for
fully-sampled statistics of cosmic galaxy populations
\citep[e.g.,][]{2015MNRAS.446..521S,2017arXiv170703406P}.
When star formation feedback physics is not directly simulated,
galactic winds are not an outcome but an input that is
part of the ``subgrid'' treatment \citep{2015ARA&A..53...51S}.
Currently, however,
subgrid models of wind driving by stellar feedback
often either adopt highly simplified scaling
prescriptions for wind mass loss rates (relative to the star formation rate)
and velocities (relative to the halo potential depth), or else are
calibrated using empirical results from a limited set of galaxies
(and hence are not fully predictive).
Better theoretical models are clearly needed.
Towards this end, the first step is to provide a physical
understanding and detailed characterization of outflowing gas (including
both winds and fountains) in
galaxies, informed and calibrated based on high-resolution
three-dimensional numerical magnetohydrodynamic (MHD) simulations.
To fully capture the
interaction between stellar feedback and a realistic multiphase interstellar
medium (ISM), it is crucial to self-consistently include the gravitational
collapse that produces star clusters and to resolve the local injection of
energy from individual massive stars.
In classical theoretical models of galactic winds motivated by observed
starburst galaxies \citep[e.g.][]{1985Natur.317...44C}, a
steady, adiabatic flow is produced
by a central energy source. In this approach,
hot, overpressured gas flows are characterized by ``mass loading''
and ``energy loading''
factors, respectively defined by the ratios of mass and energy outflow rates
to star formation rates and energy injection rates
at the wind base. Simple spherical wind models can also be constructed
that allow for radiative cooling, such that the temperature precipitously
drops at some radius in certain parameter regimes
\citep[e.g.][]{1995ApJ...444..590W,2016ApJ...819...29B,2016MNRAS.455.1830T}.
Observed galactic outflows are multiphase in nature.
Systematic observations reveal
prevalent multiphase structure of galactic winds with cold molecular
\citep[e.g.,][]{1999A&A...345L..23W,2015ApJ...814...83L}, neutral \citep[e.g.,][]{2000ApJS..129..493H,2005ApJ...621..227M,2005ApJS..160..115R,2010AJ....140..445C,2013A&A...549A.118C},
ionized \citep[e.g.,][]{2001ApJ...554..981P,2003ApJ...588...65S,2010ApJ...717..289S,2012ApJ...759...26E,2015ApJ...809..147H,2017MNRAS.469.4831C}, and hot gas phases \citep[e.g.,][]{2000MNRAS.314..511S,2007ApJ...658..258S,2017MNRAS.467.4951L}.
For the best studied example, local starburst M82, \citet[][see also \citealt{2016ApJ...830...72C}]{2015ApJ...814...83L}
have shown a clear signature of decreasing {outward}
mass fluxes in molecular and neutral gas
as a function of the distance from the disk midplane, implying a fountain
\citep{1976ApJ...205..762S,1980ApJ...236..577B} rather
than a wind
for the cooler gas.
In one conceptual framework, warm and cold gas in outflows
results when a hot medium accelerated by its
own pressure gradient cools radiatively; an alternative concept is
that overdense warm and cold
ISM clouds are ``entrained'' by a high-velocity, low-density
hot wind. More realistically, both effects can in principle
occur, and in general there is
a complex interaction between the multiple phases that are present.
The mass and energy loading factors are key quantities that
characterize winds and quantify their significance in controlling
baryonic mass cycles of galaxies. Measuring these loading factors has
been of intense observational interest, but uncertainties are still
large. In particular, the reported mass loading factor ranges widely
from 0.01 to 10 \citep{2005ARA&A..43..769V}. Depending on the
assumed geometry, metallicity, and ionization state, the mass outflow rate
can easily be reduced by a factor of 10
\citep[e.g.,][]{2016MNRAS.463..541C,2017MNRAS.469.4831C}. In
addition, uncertain deprojection may result in an
overestimate the velocity, incorrectly leading to
interpretation of a low-temperature outflow as a wind rather than a fountain.
If gas is not really escaping, the {outward} mass flux will be a decreasing
function of distance from the wind launching region, and mass fluxes
estimated at small radii would exceed the true losses from a galaxy.
Predicting wind loading factors theoretically requires
modeling the interaction between SN remnants (including
from clustered SNe) and the ISM.
Expansion of superbubbles driven by multiple SNe
has been studied by idealized analytic models and simulations
\citep[e.g.,][]{
1987ApJ...317..190M,1988ApJ...324..776M,
1989ApJ...337..141M,1999ApJ...513..142M,2017ApJ...834...25K}.
While these idealized models provide essential physical insight
and quantitative estimates,
firm theoretical measurements of mass loading in multiphase winds
from galactic disks require ISM models with realistic spatio-temporal
distribution of SNe and vertical stratification.
A number of local stratified-disk
simulations, with increasingly high resolution, have modeled
the multiphase ISM with SN feedback
\citep[e.g.,][]{1999ApJ...514L..99K,2000MNRAS.315..479D,
2004A&A...425..899D,2006ApJ...653.1266J,2012ApJ...750..104H,
2013MNRAS.432.1396G,2015MNRAS.454..238W,2016MNRAS.456.3432G,
2017ApJ...841..101L}, albeit with SN rates and locations imposed
rather responding self-consistently to star formation.
Some recent numerical work has focused specifically on the outflow properties
driven by SN feedback
\citep[e.g.,][]{2013MNRAS.429.1922C,2016MNRAS.459.2311M,2017MNRAS.470L..39F},
although with a cooling cutoff at $10^4\Kel$ that does not allow for
the full range of ISM phases.
Very recently, it has become possible to evolve the turbulent, magnetized,
multiphase ISM in local galactic disks with cooling and heating,
galactic differential rotation, and self-gravity, including
fully self-consistent resolved star formation
and SN feedback over durations of several 100 Myr
\citep[][Paper~I hereafter]{2017ApJ...846..133K}.
A few other recent simulations have also included self-gravity to model
SN rates and positions self-consistently with star formation
\citep[][]{2014A&A...570A..81H,2017MNRAS.466.1903G,2017MNRAS.466.3293P,2017A&A...604A..70I},
but given their relatively short simulation duration ($\le 100\Myr$),
they have not reached a statistically quasi-steady state and wind properties
may be subject to transient effects from the simulation start-up.
In this paper, we analyze our fiducial model from
the TIGRESS (Three-phase ISM in Galaxies Resolving Evolution with Star formation
and SN feedback) simulation suite introduced in Paper~I, in order to provide
more comprehensive understanding of multiphase gas outflows
in the realistic ISM.
Our analysis here mainly
focuses on characterizing differences among outflows of
different thermal phases.
In a subsequent paper, we will analyze
models with different galactic conditions
to investigate systematic scaling relations of wind properties
\citep[e.g.,][]{2015MNRAS.454.2691M,2015ApJ...809..147H,
2016ApJ...822....9H,2017MNRAS.469.4831C}.
In Section~\ref{sec:overview},
we review equations for steady, adiabatic flows
and summarize key physical quantities to be measured from the simulation.
We then present an overview of gas flows in the simulation, including overall
mass fluxes and vertical profiles of each gas phase; this demonstrates
the necessity of a phase-by-phase analysis.
In Section~\ref{sec:wind}, we analyze the hot gas component, showing
that it is consistent with a wind having
well-defined mass flux and specific energy
(or Bernoulli parameter) that are {approximately} constant
as a function of distance from the midplane.
Section~\ref{sec:fountain} presents an analysis of the warm gas component,
showing characteristics of a fountain flow that has
decreasing mass and energy fluxes
as a function of distance from the midplane.
Section~\ref{sec:loading} provides
mass and energy loading factors of each phase, comparing these to previous work
and to observations.
Section~\ref{sec:summary} summarizes our main conclusions.
\section{Overview of Gas Flows in TIGRESS}\label{sec:overview}
{\subsection{Outflow terminology, vertical profiles, and classical adiabatic winds}}
In galactic disks, star formation takes place in dense gas
near the midplane, within the scale height of the ISM.
Prodigious energy is injected by SNe within this thin layer,
and the high-entropy, overpressured gas expands outward.
Strong shocks heat and accelerate both dense, cold cloudlets and the warm,
diffuse intercloud medium surrounding individual SNRs and superbubbles,
with most of this gas cooling back to its original temperature relatively
quickly \citep[e.g.][]{2015ApJ...802...99K,2017ApJ...834...25K}.
Depending on the level of remaining specific energy with respect to the
gravitational potential, outflows {of cooled, SN-accelerated warm
(or cold) gas may either keep moving out of the disk,
or may turn around at some height.
While most of the energy deposited by SNe is radiated away,
some of the hot gas created in strong SN shocks is at low enough
density that it has very limited cooling. This accelerates away from
the midplane towards higher-latitude, lower-pressure regions, achieving
high enough velocity that it can escape from the galactic potential well.}
In this paper, the term ``galactic winds'' refers to outflows
launched with high enough energy to escape the
galactic gravitational potential, while the term ``galactic fountains''
refers to outflows launched with insufficient energy that eventually fall back.
{ In our simulations (and in real galaxies), the motions of gas
are three-dimensional. Within any given patch of the ISM in the disk, followed
on its orbit about the galactic center, horizontal averaging can be
used to define a mean density, velocity, and other flow properties
as a function of height $z$. In general, the instantaneous mean velocity
will have both horizontal (radial-azimuthal) and vertical
components. The horizontally-averaged flow quantities may be further
averaged over time (with window comparable to an orbit time, so that
epicyclic motions are averaged away). If the accretion rate is low,
the resulting temporally-averaged velocity at any height
will be dominated by vertical motion.
Thus, horizontal- and temporal-averaging
of the gas yields an effectively one-dimensional profile
as a function of height, consisting of average values
$\langle \rho(z)\rangle$, $\langle v_z(z)\rangle$, $\langle P(z) \rangle$,
etc. If there is a net outflow $sign(v_z)=sign(z)$, and if
there is a net inflow the sign is reversed.}
{Similar to horizontal- and time-averaged flows, classical SN-driven
wind solutions are one-dimensional, and the simplest solutions are also adiabatic.}
For steady-state one-dimensional adiabatic gas flows,
the equations of mass and total energy conservation including source terms
can be written as
\begin{equation}\label{eq:mass}
\divergence{\rho \vel}=\dot{\rho}_{\rm inj}(z),
\end{equation}
\begin{equation}\label{eq:energy}
\divergence{\rho \vel\mathcal{B}}=\dot{e}_{\rm inj}(z),
\end{equation}
where $\dot{\rho}_{\rm inj}$ and $\dot{e}_{\rm inj}$
are the volumetric mass and energy injection rates, respectively,
arising from SN feedback.
{In Equations~(\ref{eq:mass}) and (\ref{eq:energy}), $z$ represents the
vertical direction for a flow perpendicular to the plane of a galactic disk;
an approximately spherical galactic center
flow \citep[e.g.][]{1985Natur.317...44C} would
instead have $\dot{\rho}_{\rm inj}(r)$ and $\dot{e}_{\rm inj}(r)$.
Note that $\dot{\rho}_{\rm inj}$ does {\it not} represent SN ejecta itself.
Rather, shock heating of ambient ISM gas near SNe increases the
mass injection rate above that of
the initial SN ejecta, while cooling tends to reduce the rate.
Allowing for this shock heating and cooling,
$\dot{\rho}_{\rm inj}$ is simply the mean local rate
at which hot material is added the steady flow.
Similarly, $\dot{e}_{\rm inj}$ represents the
mean local energy input rate to the flow,
which is bounded above by the initial
energy carried by SN ejecta.}
In Equation (\ref{eq:energy}), the
total specific energy (the Bernoulli parameter)
is defined by
\begin{equation}\label{eq:ber}
\mathcal{B}\equiv\frac{v^2}{2}+\frac{\gamma}{\gamma-1}\frac{P}{\rho}+\Phi,
\end{equation}
consisting of specific kinetic energy, specific enthalpy
$h=[\gamma/(\gamma -1)]P/\rho$,
and gravitational potential terms.
Note that here, for simplicity, we consider adiabatic, unmagnetized gas, but
Appendix~\ref{sec:appendix} presents {
the full equations for general
MHD flows in a local shearing box, and shows that
horizontal- and time-averaging yields a set of
simple steady-state 1D flow equations, which can be
applied} to our simulations.
In this paper, the $\Phi=0$ reference point is at the midplane.
For flows in a local Cartesian box, like ours,
with energy sources near the midplane,
{any time-averaged steady winds that may exist }
are launched vertically along the $\hat z$ axis.
Through Gauss's Law, volume integration of Equations~(\ref{eq:mass})
and (\ref{eq:energy}) gives
mass and total energy fluxes
through surface area $A=L_xL_y$ at
$z=\pm d$
as
\begin{equation}\label{eq:Mflux}
\Mflux{}(d)\equiv F_M(z=d)-F_M(z=-d)=\dot{M}_{\rm inj}(d)/A
\end{equation}
and
\begin{equation}\label{eq:Eflux}
\Eflux{}(d)\equiv F_E(z=d)-F_E(z=-d)=\dot{E}_{\rm inj}(d)/A.
\end{equation}
Here, $d$ is the distance from the disk midplane,
and total mass and energy injection rates within $|z|<d$ are
$\dot{M}_{\rm inj}(d)=\int_{-d}^{d} \dot{\rho}_{\rm inj} dV$
and $\dot{E}_{\rm inj}(d)=\int_{-d}^{d} \dot{e}_{\rm inj} dV$.
The quantities $F_M(z)=\langle \rho v_z \rangle$
and $F_E(z)=\langle \rho v_z\mathcal{B}\rangle$
stand for mass and total energy fluxes
averaged over horizontal area at height $z$.\footnote{
For the energy flux, we hereafter use subscripts KE, TE, GE, and ME
to denote different energy components. These are respectively
$F_{\rm KE}=\rho v_z v^2/2$ (kinetic), $F_{\rm TE}=\rho v_z h$ (thermal,
with $h$ enthalpy),
$F_{\rm GE}=\rho v_z \Phi $ (gravitational),
and $F_{\rm ME}=S_z$ (magnetic;
see Section~\ref{sec:appendix} and
Equation (\ref{eq:a_Poynting_z})).
The subscript E will denote the total energy term, the sum of all four
components.}
The above relations assume periodic boundary conditions in both of the
horizontal directions; energy terms associated with the background shear from
integrals over faces perpendicular to $\hat x$ are discussed in
Appendix~\ref{sec:appendix}.
Since SN explosions are usually concentrated within a thin layer near the midplane,
$\dot{M}_{\rm inj}(d)$ and $\dot{E}_{\rm inj}(d)$ are expected to approach
constant values for $d\gg H$, where H is the disk scale height.
Hence,
if $A$ is constant as in the local Cartesian coordinates,
for a steady wind the (areal) mass and energy fluxes $\Mflux{}$ and $\Eflux{}$
would also approach constants at $d\gg H$ above the source region.
For general geometries (e.g., \citealt{1985Natur.317...44C}
for spherical coordinates), however,
where the area that encloses a given set of
streamlines varies as a function of distance (e.g., $A\propto r^2$ for the
spherical case),
the fluxes would also vary (e.g., $\propto r^{-2}$).
To the extent that horizontal correlations (or anticorrelations)
among variables at a given $z$ may be neglected,
$F_E=F_M \mathcal{B}$, and if we may assume symmetry across the midplane
this yields $\mathcal{B} = \mathcal{F}_E/\dot{\Sigma}_{\rm wind}=\dot{E}_{\rm inj}/\dot{M}_{\rm inj}$.
This implies that beyond the
source region where $\dot{E}_{\rm inj}$ and $\dot{M}_{\rm inj}$ have reached
their final values, the Bernoulli parameter becomes constant
{for a steady (or time-averaged) flow.}
In the more general case with $A$ an increasing function of distance,
because the area perpendicular to streamlines is the same for both mass and
energy flux, $F_E$ and $F_M$ would vary {$\propto A^{-1}$} but
$\mathcal{B}$ would still be conserved along
streamlines (beyond the source region) irrespective of geometry. Therefore,
for steady {adiabatic winds (or equivalently
for any time-averaged, adiabatic portion in
a more general outflow),}
the Bernoulli parameter is a key quantity
that enables robust extrapolation
of flow evolution to large distances.\footnote{This holds true for
MHD flows if the Poynting flux is negligible,
as in our simulations; see Figure~\ref{fig:hot}.}
As applied to the present problem,
this suggests that evaluation of $\mathcal{B}$ with our local disk simulations
should provide predictions for properties {of the hot portion of the wind}
at large distance
(outside our simulation domain) where
wind streamlines open up (becoming more radial than vertical).
This motivates the need to quantify mass and energy fluxes
in the launching region -- just above the source region -- in our simulations.
The simple analysis above provides helpful intuition for gas flows driven
by localized energy sources, but the real ISM -- and our simulations -- is not
a single phase, adiabatic gas. In fact, material in the ISM spans a
wide range of density and temperature.
SN shocks are responsible for
generating the hot gas phase ($T\sim 10^6-10^7\Kel$),
which interacts with surrounding warm ($T\sim 10^4\Kel$)
and cold ($T\sim 10^2\Kel$) phases.
Considering each thermal component
\textit{individually}, radiative heating and cooling as
well as mass and energy transfers between components
would act as source and sink terms in the conservation equations of each phase.
Because SN ejecta strongly interact with the surrounding
ISM in extremely
complex ways to heat and accelerate gas, some of which may be able to
escape from a galaxy, it is not at all obvious how one would estimate
$\dot{M}_{\rm inj}$ and $\dot{E}_{\rm inj}$
{for individual thermal components of a multiphase outflow.}
Moreover, star formation and hence SN events
are very bursty, and this burstiness may affect yields.
Clearly, {the total and individual-phase}
$\dot{M}_{\rm inj}$ and $\dot{E}_{\rm inj}$ are
only quantifiable with self-consistent numerical simulations that capture
the full physics of the ISM.
{ Nevertheless, while simulations are essential for obtaining the detailed properties of
realistic multiphase outflows, we can
still expect certain aspects of classical wind solutions to hold when
suitably applied. In particular, as we shall
show, the space-time-averaged hot portion of the wind, when considered
separately from other phases, shares many
similarities with adiabatic one-dimensional winds. }
{\subsection{TIGRESS simulation model and analysis}}
In Paper~I, we presented a novel framework for
multi-physics numerical simulations of the star-forming
ISM implemented in the {\it Athena} MHD code
\citep{2008ApJS..178..137S,2009NewA...14..139S}.
We solve the ideal MHD equations in a local, shearing
box, representing a small patch of a differentially rotating
galactic disk. This treatment allows us to
achieve uniformly
high spatial resolution compared to what is possible in
a global simulation of an entire
galaxy (or galaxies) \citep[e.g.,][]{2012MNRAS.421.3522H,2015MNRAS.454.2691M},
which is crucial to resolve both star formation and SN feedback
as well as all thermal phases of the ISM both near and far from the midplane.
We include gaseous and (young) stellar self-gravity, a fixed external
gravitational potential to represent the old stellar disk and dark matter halo,
galactic differential rotation, and optically thin cooling and grain
photoelectric heating. We utilize sink particles to follow formation of and
accretion onto star clusters in dense, cold gas. Massive young stars in these
star clusters feed energy back to the ISM, by emitting far-ultraviolet
radiation (FUV) and exploding as SNe. The former heats the diffuse
warm and cold ISM, while the latter creates hot ISM gas, drives turbulence, and
induces outflows.
Our simulations yield
realistic, fully self-consistent three-phase ISM models with self-regulated
star formation.\footnote{{Of course, the absence of global geometry
means that we are unable to follow effects of strong noncircular flows in
the disk, transport of gas from one radius to another in a fountain
(which would also require significant angular momentum exchange), or
the transition of hot winds through a sonic point. Nevertheless, the
high resolution afforded by our local scope is extremely valuable for
limiting artificial mixing, which is essential for understanding
key characteristics of multiphase flows}.}
Paper~I presented results from a fiducial model with
parameters similar to those of the Solar neighborhood. There, we showed that
after $t\sim100\Myr$ a quasi-steady state is reached.
When stars form, massive stars
enhance heating in the warm and cold ISM,
and the SN rate increases, driving turbulence throughout the ISM.
Both feedback processes disperse dense clouds and puff the gas disk up,
{temporarily}
shutting star formation off. With the corresponding reduction in
star formation feedback, the gas can settle back to the midplane and
collect into dense clouds which then form a new generation of stars.
In Paper~I, we evaluated several basic ISM and wind properties, and
demonstrated their convergence as a function of the numerical spatial resolution.
In this paper, we analyze detailed properties of galactic winds and fountains
for a vertically extended version of the fiducial model of Paper~I.
The simulation domain covers
$1\kpc\times1\kpc$ horizontally and $-4.5\kpc<z<4.5\kpc$ vertically, at a
uniform resolution $\Delta x=4\pc$.
Representative snapshots displaying a
volume rendering of density and velocity vectors
during outflow- and inflow-dominated periods are shown in
Figure~\ref{fig:snap}.
Figure~\ref{fig:slice} displays slices of temperature and vertical velocity
through the $y=0$ plane for the same snapshots shown in Figure~\ref{fig:snap}.
The outflows and inflows seen in Figures~\ref{fig:snap} and \ref{fig:slice}
are part of an overall
cycle that repeats, representing the response to large amplitude
temporal fluctuations in star formation rates
(see grey line in Figure~\ref{fig:flux_tevol}(a)).
\begin{figure*}
\plottwo{fig1a.pdf}{fig1b.pdf}
\caption{Sample snapshots illustrating
(left) outflow-dominated, and (right) inflow-dominated periods,
at $t=300$ and $360\Myr$, respectively. Gas density is shown in
color scale volume rendering and the velocity field is shown with
vectors; vector colors (rather than length)
indicate velocity magnitudes.
A fast-moving, low-density (dark blue) outflow is evident in the left snapshot,
while moderate-density (green) gas that was previously
blown out to large distances is falling back toward the midplane in the right.
Velocity fields are turbulent near the midplane ($d<H$), but ordered
in either outflowing or inflowing directions at large distances.
For visual clarity, only the upper half of the simulation box
($z=0$ to $4\kpc$) is shown, with a full horizontal crossection
$-512\pc \le x,y \le 512\pc$.
\label{fig:snap}}
\end{figure*}
\begin{figure}
\plottwo{fig2a.pdf}{fig2b.pdf}
\caption{Sample slices through $y=0$ showing temperature and vertical velocity
at $t=300$ (left) and $360\Myr$ (right),
representing outflow- and inflow-dominated periods.
During outflow-dominated periods (left), the hot component
fills most of the volume at high $|z|$ and flows outward at high velocity,
while the warm component is confined in small cloudlets.
During inflow-dominated periods (right),
the warm component occupies most of the volume and falls toward the midplane.
As in the volume renderings of Fig. \ref{fig:snap},
only the upper half of the simulation box ($z=0$ to $4\kpc$) is shown.
\label{fig:slice}}
\end{figure}
\begin{figure}
\plotone{fig3.pdf}
\caption{
Time evolution of (a) mass and (b) energy fluxes driven by
bursts in the star formation rate.
Mass and energy fluxes per unit area at distances
$d=$1, 2, and 3 kpc from the midplane are measured
(Equations~\ref{eq:Mflux} and \ref{eq:Eflux}).
Each burst in $\Sigma_{\rm SFR}$ (shown in grey in upper panel)
leads to outflows in both mass and energy near the midplane.
The net mass flux decreases at larger $d$ because most of the outflow near
the midplane is warm ``fountain'' gas with low velocity that turns around
{(resulting in periods of negative mass flux at $d=1\kpc$).}
Energy fluxes are substantial at all heights, due to the dominance of
hot gas that escapes as a wind.
\label{fig:flux_tevol}}
\end{figure}
To characterize vertical gas flows, we first construct
horizontally-averaged quantities. We then calculate the {net}
mass and energy fluxes
that pass through horizontal planes (both upper and lower sides)
at $d=1$, $2$ and $3\kpc$
(cf. Equations~\ref{eq:Mflux} and \ref{eq:Eflux}).
Figure~\ref{fig:flux_tevol} shows time evolution of
(a) mass flux,
along with the areal star formation rate ($\Sigma_{\rm SFR}$),
and (b) energy flux.
Every star formation burst is followed by a burst of energy
injection, and {this burstiness is reflected in large temporal
variations in the mass and energy fluxes. The fluxes can become
negative, meaning that the mass and/or energy of gas flowing inward
exceeds that of the gas flowing outward, at a given height.}
As distance $d$ from the midplane increases, the
net mass flux significantly decreases,
while the net energy flux even through $d=2$ and $3\kpc$ remains large.
While net negative mass fluxes (implying fallback) occur at $d=1\kpc$
after each burst of mass outflow,
net energy fluxes almost always remain positive. The differing
behavior of mass and energy fluxes is a signature of multiphase flows.
{Although star formation and hence SN feedback are impulsive
rather than continuous, the system
approaches a quasi-equilibrium state.
This state is a limit cycle mediated by the feedback loop, in which
epochs of cooling and collapse alternate with epochs of heating and
expansion.\footnote{{We note, however, that a quasi-steady
state is not guaranteed for all galactic conditions
\citep[e.g.,][]{2017MNRAS.467.2301T}, and may
only hold in parts of the parameter space in which the vertical
oscillation time (which controls collapse and star formation) is
sufficiently long compared to the stellar evolution timescale
(which controls feedback and expansion).}}
Given that a
quasi-steady state exists in the present simulation,
horizontal- and temporal averages can be
constructed to characterize this mean state. Since different
thermal phases coexist at all heights, to understand the outflows
and inflows of mass and energy it is further necessary to separately
construct horizontal averages of each thermal phase.}
\begin{figure}
\plotone{fig4.pdf}
\caption{Vertical distributions of
mass, vertical momentum, and energy densities, averaged horizontally
and over upper ($z>0$) and
lower ($z<0$) sides of the disk
and over time $t=250-500\Myr$.
Colored lines for separate thermal components
show profiles of (a) hydrogen number density $n_H$,
(b) vertical momentum density
$\rho v_{\rm out}$ (which is the same as the mass flux), and
(c) total energy density (excluding gravity) $\cal E$, each as a function of
distance $d$ from the midplane. Color-coded shaded regions represent
one-sigma temporal fluctuations. In order to properly
visualize both the magnitude and sign of the momentum density,
we use a linear scale
for $|\rho \vout|<10^{-4}$ and a log scale for $|\rho\vout|>10^{-4}$.
The warm and hot phases respectively dominate mass and energy
densities above the disk scale height (indicated by vertical dotted line),
and the hot component also has the largest vertical momentum density
(net outward mass flux).
\label{fig:zprof}}
\end{figure}
In Figure~\ref{fig:zprof}
we plot
horizontally- and temporally-averaged
profiles of
mass, momentum, and energy distributions for thermally separated gas phases;
{these profiles average over $t=250-500\Myr$ } and also average
over upper ($z>0$) and lower ($z<0$) sides. Profiles show
hydrogen number density, $n_H\equiv\rho/(\mu_Hm_H)$,
outward vertical momentum density, $\rho \vout\equiv \rho v_z{\rm sign}(z)$,
and total energy density (excluding gravity)
\begin{equation}\label{eq:Etot}
\mathcal{E}\equiv \frac{1}{2}\rho v^2 +\frac{P}{\gamma-1} +
\frac{B^2}{8\pi}.
\end{equation}
{Note that this energy density differs from the gas density multiplied by
the Bernoulli parameter of Equation (\ref{eq:ber}),
as it includes
just thermal energy (rather than enthalpy) and also includes
magnetic energy density.}
The four thermal phases plotted in Figure~\ref{fig:zprof} are
cold ($T<5050\Kel$), warm ($5050\Kel<T<2\times10^4\Kel$), ionized
($2\times10^4\Kel<T<5\times10^5\Kel$), and hot ($T>5\times10^5\Kel$).\footnote{
Note that we omit the ``unstable'' phase ($184\Kel < T < 5050\Kel$)
defined in Paper~I and merge it into the ``cold'' phase since (1) these phases
are not of primary interest in this paper since we are focusing on gas
above the disk scale height, and
(2) the sum of these two phases numerically converges better than the individual
phases.}
Above the warm/cold layer ($d>H$, where $H\approx 400\pc$),
the mass density is dominated by the warm component and the
energy density is dominated by the hot component.
As the individual terms in the energy density
are proportional to corresponding terms in the momentum flux ($\rho v^2$, $P$),
the hot medium also
dominates the momentum flux
away from the midplane.
The hot medium is the largest contributor to the
time-averaged vertical momentum
density, which is the same as the time-averaged net mass flux.
Although the mean value of vertical momentum density (or net mass flux)
of the warm medium is effectively
zero, there are
large temporal fluctuations (indicated by the green shaded region)
at small and intermediate $d$
because the warm gas contributes significantly to
both outgoing and incoming mass fluxes at different times
(as in Figure~\ref{fig:flux_tevol}).
The phase-separated momentum and energy density profiles in
Figure~\ref{fig:zprof} (and
corresponding profiles of mass and momentum flux) reflect essential
differences of gas flow dynamics between the warm and hot phases,
which we separately analyze in the following sections.
\section{Hot Winds}\label{sec:wind}
In this section, we focus on the hot component, defined by $T>5\times10^5\Kel$,
representing gas that has been shock heated by SN blastwaves.
The hot medium fills most of the volume above the disk scale height,
and cooling in tenuous hot gas is inefficient. With pressure
gradients that accelerate it outward (cf. Figure~\ref{fig:zprof}(c)) and
a source from SN shocks propagating into the surrounding warm and cold medium
near the midplane,
the {horizontally and temporally averaged} hot medium naturally fits the criteria for a (quasi-)steady, adiabatic wind.
We therefore might expect that the mass flux and energy flux
(and thus the Bernoulli parameter)
of the time-averaged hot component would be (approximately)
conserved as the gas flows outward.
Figure~\ref{fig:hot} explicitly shows vertical profiles
of (a) the mass flux, $\Mflux{,h}$
and (b) the specific energy, $\Eflux{,h}/\Mflux{,h}$,
based on time averages of each horizontally averaged flux.
{To distinguish between outflows and inflows, for the mass flux
in addition to net flux we separately show the flux of outflowing
($sign(v_z)=sign(z)$) and inflowing ($sign(v_z)=-sign(z)$) gas.}
\begin{figure*}
\plotone{fig5.pdf}
\caption{Time averaged ($t=250-500\Myr$) vertical profiles
of hot gas (a) mass fluxes and (b) specific energies.
(a) The net mass flux is shown as a blue line with
one-sigma temporal fluctuations as the blue
shaded region. The hot gas does not show any
significant inflows (yellow line),
implying that any decrement of the mass flux is due to
a phase transition to cooler phases and not direct inflows of the hot gas.
(b) Total Bernoulli parameter $\cal B$
as well as individual components
(i.e. kinetic, thermal, gravitational, and magnetic
terms; see Equations (\ref{eq:ber}) and (\ref{eq:spec_mag})).
Mean values are shown as colored lines,
and one-sigma temporal fluctuations are shown as shaded regions.
The specific enthalpy
and kinetic energy are much larger than the gravitational potential.
The magnetic term plays a minor role.
In both (a) and (b), the vertical dotted line indicates the disk scale
height ($H=400\pc$).
\label{fig:hot}}
\end{figure*}
\begin{figure*}
\plotone{fig6.pdf}
\caption{
{
Mass flux and specific energy profiles for the hot wind as in
Figure~\ref{fig:hot}, except at a single instant $t=300 \Myr$.}
\label{fig:hot300}}
\end{figure*}
The mass flux of the hot gas shows net outflow
(blue line in Figure~\ref{fig:hot}(a)),
with negligible inflow flux at all heights
(yellow line in Figure~\ref{fig:hot}(a)).
As the SNe that create the hot medium are concentrated near the midplane,
the hot gas mass flux within the warm/cold layer
($d<H$; below the vertical dotted line in Figure~\ref{fig:hot})
first increases with $d$
and then decreases as
hot SNRs have a maximum size before the onset of cooling.
Above the point where the hot gas in the interior of SN remnants
(or superbubbles) breaks out of the warm/cold layer the hot gas
mass flux is nearly constant with $d$.
The slight decrease of the mass flux at large $d$ is caused by non-zero cooling
(radiative and mixing of hot gas with cooler phases).
Figure~\ref{fig:hot}(b) shows near constancy of the time-averaged
total specific energy $\mathcal{B}$ of the hot gas
above $d=1\kpc$, as expected for a steady, adiabatic wind
(given the negligible Poynting flux).
We also calculate individual components of the (averaged) energy flux,
and divide them by the (averaged) mass flux.
This provides the components of the mass-flux-weighted specific energy
$\Eflux{,h}/\Mflux{,h}$, consisting
of kinetic, thermal (enthalpy), gravitational,
and magnetic terms
(see Equations \ref{eq:ber} and \ref{eq:a_Poynting_z}).
Once the hot gas breaks out of the warm/cold layer, the
gas flow approximately preserves mass and energy fluxes
because except for limited cooling
there are no sources (or sinks) for the hot gas mass and energy.
We have checked that the individual cooling and heating terms,
including Reynolds and Maxwell stresses arise from the shearing box
(see Equation
(\ref{eq:a_energyz})), are indeed small compared to the SN energy injection.
From $d=2-4\kpc$, the enthalpy of the hot gas implies a temperature in the
range $1.2-1.5\times10^6\Kel$
(or sound speed in the range $130-150\kms$), and the kinetic energy
of the hot gas translates to a velocity of $170-200\kms$. The Poynting flux
contribution to $\mathcal{B}$ is negligible, corresponding to an Alfv\'en
speed of $30-35\kms$.
In SN-driven hot winds, the enthalpy (specific heat) term
dominates over other components
of the Bernoulli parameter, including the gravitational potential, at
heights less than the disk radius.
At large distance where streamlines open up in angle,
one could expect hot galactic winds
with constant $\cal B$ to accelerate past a sonic point
to an asymptotic velocity of
$v_{\rm wind}\equiv\sqrt{2}(\mathcal{B}-\Phi)^{1/2}$ as they adiabatically cool,
similar to classical Parker stellar wind solutions.\footnote{Note that for galactic potentials, $\Phi$ is generally
computed relative to the midplane, whereas for point mass potentials
the $\Phi=0$ reference point is at infinite distance; the zero point in
$\Phi$ is irrelevant as long as it is consistent in $v_{\rm wind}$ and
$\mathcal{B}$.}
{Global simulations indeed show the expected behavior at large distance
\citep[e.g.][]{2017MNRAS.470L..39F}.}
In general, hot
winds are accelerated by pressure gradients at the same time as
enthalpy is converted to kinetic energy, and a sonic transition in
a steady wind is only possible if the crossectional area increases as
the flow moves outward \citep[e.g.][]{1992pavi.book.....S}.
In the Cartesian geometry of the present simulations, streamlines cannot
open up and there is no associated adiabatic cooling, limiting the pressure
and density gradients and therefore the acceleration of the flow.
However, the constancy of both the Bernoulli parameter and mass flux with
$d$ for the hot medium in our simulation suggests
that it properly represents the near-disk regions for a generalized galactic
disk wind,
in which streamlines emerge from the disk vertically (with $A=const$
when $z\ll R_{\rm disk}$) and would open up ($A$ increasing with distance) when
$z\simgt R_{\rm disk}$.
{The mass and energy fluxes carried by the
hot wind are controlled} by the
interaction between SN shocks and the warm-cold medium
that creates the hot ISM well within the disk, in processes
that are unlikely to be affected by large-scale global galactic and
CGM properties and geometry. Therefore, the
Bernoulli parameter we calculate
should be a robust estimator of asymptotic {hot} wind speed
irrespective of the constraints of our Cartesian box.
As shown in Figure~\ref{fig:flux_tevol}
(and demonstrated by the shaded region in Figure~\ref{fig:hot}(b)),
star formation and hence outflows are bursty,
resulting in large temporal fluctuations.
Based on analysis of
{${\cal B}$ for the hot component within}
short time ranges,
the maximum asymptotic velocity\footnote{Here, this maximum wind velocity
is calculated using $\Phi$ at a height $d=4\kpc$, although in reality the wind
velocity would continue to decrease slowly with distance due to
the logarithmic increase of $\Phi(r)$
with $r$ at large distances in dark matter halo potentials.}
of the hot wind could
reach up to $v_{\rm wind}\sim 500\kms$, while the
mean asymptotic wind speed would be $v_{\rm wind}\sim 350\kms$.
Defining the speed required to escape to $\bf r$ as
$v_{\rm esc}({\bf r}) \equiv [2(\Phi({\bf r})-\Phi(0))]^{1/2}$,
in our simulation domain
$v_{\rm esc}(z=4\kpc)\sim 130\kms$, so the
hot wind easily escapes,
{i.e. ${\cal B}\gg \Phi$ even at large $d$} for the hot component.
A hot wind launched with the local conditions of our simulation would
also be able to propagate far into the halo for the Milky Way,
where the escape velocities are $v_{\rm esc}(50\kpc)\sim 350\kms$ and
$v_{\rm esc}(150\kpc)\sim450\kms$
using {\tt MWPotential2014} in {\tt galpy} \citep{2015ApJS..216...29B}.
More generally, the far-field velocity for a hot wind with given local
launching conditions can be estimated based on $\mathcal{B}$
and the large-scale galactic potential.
{Finally, we note that while we have mostly based the discussion above
on temporal averages of horizontally-averaged profiles, the properties
of instantaneous profiles are quite similar. To illustrate this,
Figure \ref{fig:hot300} shows the mass flux and specific energy of the
hot component at $t=300$ Myr, the ``outflow'' snapshot shown in Fig.
(\ref{fig:slice}). Except for local fluctuations, the instantaneous
mass flux and specific energy profiles are overall very similar to the
corresponding time-averaged profiles. In particular, the specific energy
profiles in Fig. \ref{fig:hot} and \ref{fig:hot300} are quantitatively
almost the same, while the mass flux is higher in Fig. \ref{fig:hot300} than in
Fig. \ref{fig:hot} because the latter includes non-outburst epochs as well as
outburst epochs in the temporal averaging.}
\section{Warm Fountains}\label{sec:fountain}
Over the duration of the simulation,
the mean net mass flux in warm gas out of the simulation domain is
$\sim 1.1\times10^{-4} \sfrunit$, about 28 percent of the net mass flux in hot gas.
As we shall show, the outward
mass flux of the warm medium secularly decreases
with $d$, such that if our box were taller we would expect the mean net
mass flux out of the simulation domain to be even smaller.
{With negligible time-averaged net outward mass flux at large distance,
the warm medium at $d\simgt 1\kpc$ in our simulation is not a true galactic wind.}
Even though the inflows and outflows of warm gas over long timescales
are essentially balanced, time variations in the warm gas flux
$\rho v_{\rm out}$ are
quite large and include both positive and negative values
(green shaded region of Figure~\ref{fig:zprof}(b)).
{The fluctuating behavior of the warm medium
can be contrasted with the much smaller temporal fluctuations of the
mass flux in the hot gas (red shaded region of Figure~\ref{fig:zprof}(b)).}
Evidently, the warm gas does not produce an escaping
wind like that in the hot gas
but a fluctuating fountain that at any time consists of both outflow and inflow.
{Figure~\ref{fig:flux_tevol} shows that alternating inflow
and outflow dominance in the warm gas is reflected in the alternating signs
of the mass flux for the whole medium at $d=1\kpc$. In addition,
the decrease in the magnitude of the total (phase-integrated)
mass flux with increasing
$d$ reflects the secular decrease in the net mass flux of the warm fountain
(difference between inward and outward fluxes) with distance.}
\begin{figure*}
\plotone{fig7.pdf}
\caption{Vertical profiles of warm gas mass fluxes (top)
and specific energies (bottom)
averaged
over outflow (left; $\Mflux{}(1\kpc)>0.001\sfrunit$) and inflow
(right; $\Mflux{}(1\kpc)<-0.001\sfrunit$) periods.
Inflows and outflows are comparable with each other,
implying little net mass and energy outflows (or inflows)
from the simulation domain associated with the warm medium.
Note that, for visualization purpose, we take absolute values
of the inflow mass flux and net flux during the inflow period;
these have negative signs by definition.
For the warm medium, the kinetic term exceeds both the thermal and
magnetic terms in the specific energy, but is always lower than the
gravitational term above $\sim 1\kpc$. This explains why the
warm medium creates a fountain rather than a wind.
The vertical dotted lines in all panels indicates the gas scale
height ($H=400\pc$).
\label{fig:warm_flux}}
\end{figure*}
To quantify the characteristics of warm fountain flows,
we take time averages selectively for
outflow ($\Mflux{}(1\kpc)>0.001\sfrunit$)
and inflow ($\Mflux{}(1\kpc)<-0.001\sfrunit$) periods.
Outflows of warm gas occur when many correlated SNe from a
star formation burst lead to a
superbubble expanding into the warm and cold layer, while
inflows occur when the disk is in a quiescent state with reduced
star formation after the cold medium has been dispersed by a
previous burst.
Figure~\ref{fig:warm_flux} plots time averaged mass flux (top) and
specific energy (bottom) for outflow (left) and inflow (right) periods.
Although one flow dominates the other during each period,
the opposite flows always exist at all heights
and are more significant compared to the case for
hot gas (Figure~\ref{fig:hot}).
For the warm medium, the kinetic term in the specific energy
exceeds the enthalpy,
but remains below the gravitational potential term at $d\simgt 1\kpc$.
This explains why most of
the warm gas outflow turns around at $d\sim1-2\kpc$ and
falls back toward the midplane.
During the outflow period, the mean velocity of the warm gas is in the
range $\sim60-80\kms$ for $d=1-4\kpc$.
Overall, the warm medium occupies more volume above $d=1\kpc$
during the inflow period than the outflow period.
This is because the hot gas is mainly generated
during the outflow period by shocking the warm gas, and
the high-pressure hot gas confines the warm gas into small cloudlets.
When the disk becomes quiescent, the warm gas expands into previous
hot wind channels
(see Figure~\ref{fig:slice}).
About a factor of 5 to 10 more volume
is occupied
by the warm gas in the inflow period than the outflow period for
$d=2$ and $3\kpc$ slabs.
\begin{figure*}
\plotone{fig8.pdf}
\caption{Outward vertical velocity ($\vout\equiv v_z{\rm sign}(z)$)
probability distributions of the warm medium.
Distributions of volume (top),
mass (middle), and mass flux (bottom)
as a function of velocity
are calculated for $20\pc$ thickness slabs above and below the
midplane ($z=\pm d$) {centered at $d=0,0.5,1,2,3\kpc$.}
In the bottom row, the mass flux within each velocity bin is in units of $\sfrunit/\kms$, and
negative mass fluxes are shown with dotted lines.
Warm gas with both signs of $\vout$ is present
during both ``outflow'' and ``inflow'' periods, but the corresponding
mean velocity {(for $d\ne 0\kpc$)} changes sign.
Acceleration of the warm gas to high velocities,
especially during outflow periods, is evident in the difference of
profile shapes between the midplane ($d=0\kpc$) and higher
latitude, with an exponential tail at high velocity developing by $d=1\kpc$.
However,
an overall deceleration with height in the increasing gravitational potential
is also evident in comparison of profiles at increasing $d\ge 1\kpc$.
The magenta dashed lines in the left column
are fits to the fast-moving gas $\vout>50\kms$ at $d=1\kpc$
given by {Equations (\ref{eq:fit}) and (\ref{eq:fit_mf})}.
\label{fig:warm_vpdf}}
\end{figure*}
Although in the current simulations the warm gas is
almost entirely
confined by the
galactic gravitational potential,
this would not necessarily be true if the potential were shallower,
as in dwarf galaxies.
Direct simulations for different galactic conditions, including
shallower potentials, are underway using the same TIGRESS framework.
However, we can also use our
current simulation to provide information on what might be expected
by quantifying the fraction of fast-moving warm gas.
Figure~\ref{fig:warm_vpdf}
shows outward velocity probability distribution functions (PDFs)
weighted by volume (top),
mass (middle), and mass flux (bottom)
within slabs of thickness $\Delta_{\rm slab}=20\pc$
during outflow (left) and inflow (right) periods.
We show results at several distances $d$ averaged over
both sides of the disk ($z=\pm d$).
Note that the volume-weighted PDF is normalized by slab volume
($V_{\rm slab}=L_xL_y2\Delta_{\rm slab}$),
and the mass-weighted PDF is normalized by the mean mass within the same
volume
($\overline{M}_{\rm slab}\equiv M_{\rm tot} 2\Delta_{\rm slab}/L_z$)
($M_{\rm tot}$ is the mass in the whole domain at a given time),
while the mass flux PDF is in physical units of $\sfrunit/\kms$.
During both outflow and inflow periods,
Figure~\ref{fig:warm_vpdf} shows that the warm gas velocity
has a broad distribution with both outward and inward velocities.
During the outflow period, Figures~\ref{fig:warm_vpdf}(a) and (c) show that
the volume and mass of high velocity ($>50\kms$) warm gas increases
from the midplane to $1\kpc$, where the specific kinetic
energy is larger than the gravitational potential
(see Figure~\ref{fig:warm_flux}(c)).
The increase in mass of high-velocity, high-altitude warm gas
between the midplane and $d=1\kpc$
is due to acceleration of the warm medium pushed by expanding
superbubbles; this includes warm gas that was shock-heated to the hot phase
(and accelerated to high velocity) and subsequently cooled back down.
Figures~\ref{fig:warm_vpdf}(c) and (e) show that as $d$ increases
the peaks of the distributions of mass and mass flux move to higher velocity
and the overall outflowing gas fraction decreases.
This general trend represents
dropout/turnaround of warm gas fluid elements with low (and
decelerating) velocities that are unable to climb to large $d$ in the
gravitational potential.
In principle, acceleration of warm clouds driven by hot-gas ram pressure,
cooling of fast hot gas, and dropout of low-velocity warm fluid elements
could all contribute to the gradual increase of warm-medium
specific kinetic energy at $d>H$ shown in Figure~\ref{fig:warm_flux}(c).
Figure~\ref{fig:warm_vpdf}(c) shows, however, that overall the mass of
high-velocity warm gas is {\it decreasing} with increasing $d$. This has
the important implications that in our simulation (1)
warm clouds at large $d$ are {\it not} significantly
accelerated by ram pressure of the hot, high-velocity gas that
is flowing out around them, and
(2) relatively little hot gas is converted to the warm phase through
cooling at large $d$. Rather, the warm medium is primarily
accelerated via direct energy input from SNe at $d\simlt 1\kpc$, and at higher
altitudes warm fluid elements slow and turn around according to the
competition between the gravitational potential and the kinetic energy they
initially acquired at small $d$.
Figure~\ref{fig:zprof}(c) and Figure~\ref{fig:hot}(b) are also telling in this
regard: the total energy density (and individual components)
of the hot medium declines very slowly for $d\simgt 1\kpc$,
while the energy density
of the warm medium declines steeply; since momentum flux {terms} are
proportional to energy density {terms},
this indicates that there is no significant transfer of
momentum from the hot to the warm gas.
During the inflow period, the majority of the warm gas is falling.
Since SN feedback is never completely turned off, however,
some warm gas is still accelerated outward.
The fraction of (outgoing) fast-moving warm gas ($v_{\rm out} > 50\kms$)
is reduced by a factor of 5 to 10
in inflow compared to outflow periods.
Combined with the total outflowing mass fluxes of
$\Mflux{,w}\sim 10^{-2}$ and $10^{-3}\sfrunit$
at $d=1\kpc$ during outflow and inflow periods
(see Figure~\ref{fig:warm_flux}(a,b)), respectively,
the mass fluxes of the fast-moving warm gas are about $6\times10^{-4}$
and $10^{-5}\sfrunit$
(see Figure~\ref{fig:warm_vpdf}(e) and (f)).
This can be compared to a mean mass flux of
hot gas at the same height of
$\sim 1.3\times10^{-3}$ and $2\times10^{-4}\sfrunit$ during outflow
and inflow periods, respectively.
Although the ``fast'' warm outflow has comparable mass flux
at $d=1\kpc$ to that of the hot
medium during outflow periods,
even at $v_{\rm out} > 50\kms$ the warm medium yields
very little mass escaping from the large-scale potential in the present
simulation, while the hot medium mostly escapes
($0.12$ and $0.034\Surf$ of hot and warm gas have respectively escaped
during the time interval of $250\Myr$). This emphasizes
the importance of measuring not just mass fluxes, but mass fluxes and
specific energies in comparison to galactic escape speeds.
At large distances the large-scale gravitational potential
strongly affects the warm-gas velocity distribution by enforcing
dropout of lower-velocity material, but closer to the midplane
this is less of an issue. At around the disk scale height, the
gravitational potential is small compared to the specific kinetic energy
of the gas, and global geometric effects are not important yet.
We therefore consider the velocity distribution at $d=1\kpc$ as
representative of the launching conditions for a warm outflow,
which would apply relatively independently of the global galaxy (e.g.
in a dwarf as well as a large galaxy for given local conditions).
Here, we find the PDFs during outflow periods of the fast-moving warm gas
($\vout > 50\kms$)
at $d=1\kpc$ are well fitted by a single exponential function,
\begin{equation}\label{eq:fit}
\frac{df}{dv_{\rm out}} = A_f\exp\rbrackets{-\frac{v_{\rm out}}{v_f}},
\end{equation}
{for $f=V/V_{\rm slab}$ or $M/\overline{M}_{\rm slab}$,
where the normalization factors for volume and mass weighted PDFs are
$A_V= 0.63/v_V$ and $A_M=0.79/v_M$, respectively, and the characteristic velocities are
$v_V=18\kms$ and $v_M=15\kms$.
Using the mass PDFs, the mass-flux PDF is given by
\begin{equation}\label{eq:fit_mf}
\frac{d(\rho v_{\rm out})}{dv_{\rm out}}=
\frac{\overline{M}_{\rm slab}}{V_{\rm slab}} A_M v_{\rm out} \exp\rbrackets{-\frac{v_{\rm out}}{v_M}},
\end{equation}
where the mean density, $\overline{M}_{\rm slab}/V_{\rm slab}$,
is given by $\rho = 1.2 \times10^{-3} \Msun \pc^{-3}$,
corresponding to hydrogen number density $n_H=0.033\pcc$.}
By quantifying the mass flux PDF in
warm gas for models with different local conditions, it should be
possible to develop a comprehensive quantitative characterization of
warm wind launching by star formation feedback.
{These local results
could then be used to make global predictions.
For example, integration of Equation (\ref{eq:fit_mf}) for velocities
$\vout > v_H$ for would yield a mass flux
$(v_H/v_M + 1) \exp(-v_H/v_M) \dot\Sigma_{\rm wind,w}(d=1\kpc)$.
If this holds in general, it means that
measurement of the launching-region warm mass flux
$\dot\Sigma_{\rm wind,w}(d=1\kpc)$ for given local disk conditions could be used to
predict the flux that actually escapes into the halo for a galaxy with
arbitrary halo velocity $v_H$.
}
This statistical
characterization can be combined with measurements of mass flux and
Bernoulli parameter for the hot medium to develop subgrid models
of multiphase wind driving for implementation in galaxy formation simulations.
\section{Mass and Energy Loading}\label{sec:loading}
\subsection{Simulation Results}\label{sec:loading_result}
In this section, we discuss key quantities of
gas outflows driven by SNe, mass and energy loading factors.
The mass loading factor, $\beta$, through the surfaces at $d$
(including both sides of the disk plane)
is conventionally defined by the ratio of ``outgoing''
mass flux to the star formation rate as
\begin{equation}\label{eq:mass_loading}
\beta\equiv \frac{\Mflux{}^+}{\Sigma_{\rm SFR}},
\end{equation}
while the energy loading factor, $\alpha$, is defined by the ratio of ``outgoing''
kinetic + thermal (enthalpy) energy flux to the energy production rate of SNe,
\begin{equation}\label{eq:energy_loading}
\alpha\equiv \frac{\Eflux[KE]{}^++\Eflux[TE]{}^+}{E_{\rm SN}\Sigma_{\rm SFR}/m_*},
\end{equation}
where $E_{\rm SN}=10^{51}\erg$ is the total energy per SN, and
$m_*=95.5\Msun$ is the total mass of new stars per SN (see Paper~I).
To compute $\Mflux{}^+$, $\Eflux[KE]{}^+$, and $\Eflux[TE]{}^+$ we
select only zones with outflowing gas,
i.e. with ${\rm sign}(v_z)={\rm sign}(z)$.
The areal star formation rate averaged over $t=250-500\Myr$ is $\Sigma_{\rm SFR}=0.006\sfrunit$;
we use this average $\Sigma_{\rm SFR}$ in computing all loading factors.
Note that a rolling mean of $\Sigma_{\rm SFR}$ with
100Myr time window gives 50\% variation with respect to the mean.
Because there is generally a temporal offset between star formation bursts
and winds (see Figure~\ref{fig:flux_tevol}),
the instantaneous ratio of mass or energy fluxes to the star
formation rate can significantly over- or under-estimate the true physical
loading. The ratio of time-averaged wind fluxes to the time-averaged star
formation rate is more meaningful.
\begin{figure*}
\plotone{fig9.pdf}
\caption{Mass and energy loading factors in each thermal phase.
For both cold and warm gas,
both mass and energy loading factors stiffly decrease as a function of $d$.
Mass loading also drops off with $d$ for the ionized phase.
Only the hot gas is a true wind, with well-defined mass and
energy loading factors of $\beta_h\sim0.1$ and $\alpha_h\sim0.02$, respectively.
\label{fig:loading}}
\end{figure*}
We decompose the gas into four thermal components and present the loading factors
of each thermal phase as a function of $d$ (Figure~\ref{fig:loading}).
The energy flux is always dominated by the hot gas,
with the energy loading factor of $\alpha_h\sim0.01-0.05$ above $d>1\kpc$.
SN feedback causes large outgoing mass fluxes of warm and cold gas
within the disk scale height $d<H$,
but the majority of the warm and cold medium has low velocity and
cannot travel far from the midplane; this mass flux at small $d$
is best thought of
as the ``upwelling'' of turbulent motions within the disk
driven by expanding SNRs and
superbubbles.\footnote{Allowing for the work done by shock-heated hot gas,
both isolated and spatially correlated SNe
inject a mean spherical momentum/SN to the ISM of
$\sim 10^5 \Msun\kms$ \citep[see][and references therein]{2015ApJ...802...99K,2017ApJ...834...25K},
an order of magnitude greater than the momenta
of the initial SN ejecta. Most of this momentum goes into maintaining
quasi-equilibrium force balance with gravity
in the bulk of the ISM, rather than driving a wind.}
The mass loading factor of the ionized gas also decreases significantly
with $d$.
By the time the flow reaches the edge of the simulation domain,
$\beta_h$ is at least a factor $\sim3$
larger than all the other components. Therefore,
as we have concluded in Sections~\ref{sec:wind} and \ref{sec:fountain},
only the hot gas forms a genuine``galactic wind,''
with a mass loading factor of
$\beta_h\sim 0.1$ that is nearly constant above $d>1\kpc$.
{We note that with the above definitions of $\alpha$ and $\beta$,
the Bernoulli parameter for the hot wind is given by
${\cal B} = \Phi + (\alpha_h/\beta_h)E_{\rm SN}/m_*$}.
The vertical dependence of the mass loading factors
of non-hot phases shown in Figure~\ref{fig:loading}(a)
implies that it is important to provide careful distinctions when
reporting on the outflows measured in numerical simulations, as
``wind'' mass loading can be greatly overestimated if this is not done.
In particular, the steep decrease of $\beta_w$ with $d$ implies that
if this were measured at $d\sim 1\kpc$ it would overestimate the value at
the edge of our box ($d \sim 4\kpc$) by a factor $\sim 30$, and the
true value at larger distance would be even smaller
\citep[see also][who reached similar conclusions about total $\beta$]{2016MNRAS.459.2311M}.
Thus, a measurement
of $\beta_w$ from a simulation with small vertical domain cannot by
itself provide a prediction for warm gas mass loss in a galactic wind.
However, even with a limited vertical domain, it is possible to discriminate
between fountain flow and wind by combining measurements of the warm
gas mass flux and its vertical velocity (as in
Section~\ref{sec:fountain}). Similar considerations apply to observations of
gas at $T\sim 10^4\Kel$ at high latitudes in edge-on galaxies, but in that
case uncertain projection effects make this even more problematic:
without an unambiguous measurement of velocity (which is subject to assumed
wind geometry), it is impossible to distinguish between fountain
flow and wind from observations of the emission measure.
More generally, it is essential to decompose
outflows in simulations into separate thermal phases to distinguish
winds from fountain flows. The integrated $\beta$ cannot be taken as
a true wind mass loading unless the measurement is made at very large
distance in a global simulation. Nevertheless, individual measurements of
mass fluxes in different phases, when combined with analysis of the
components of their individual Bernoulli parameters, can be used to
distinguish fountains from winds even within a limited vertical domain.
The mass loading factor of the hot gas obtained here
is consistent with simple estimates based on idealized experiments
of superbubbles driven by multiple SNe in the warm-cold ISM
\citep{2017ApJ...834...25K}.
The shock from an individual SN sweeps up $\sim 10^3\Msun$
before it cools and forms a shell \citep{2015ApJ...802...99K}. For
superbubbles created by multiple SNe, the maximum mass in hot gas per SN
is $\hat M_h\sim 400-2000\Msun$ prior to shell formation, but subsequently
this drops to $\hat M_h\sim 10-100\Msun$ (lower at higher ambient
density). When star formation rates are self-regulated
\citep[e.g.,][]{2013ApJ...776....1K,2015ApJ...815...67K},
the mean interval between SNe within projected area $\pi H^2$
in the disk of a galaxy is always $\sim 0.3\Myr$,
and breakout of superbubbles is expected to occur after shell formation
(see Section~5 of \citealt{2017ApJ...834...25K}).
For a SN interval $\ge 0.1\Myr$, Figure 11 of \citet{2017ApJ...834...25K}
shows that by the time superbubbles reach a radius of $2H$,
$\hat M_h \sim 10- 30 \Msun$, depending only weakly on ambient density.
This corresponds to
$\beta_h=\hat{M}_h/m_*\sim 0.1-0.3$.
The same idealized superbubble simulations show
a hot-gas energy loading factor of a few percent when superbubbles
expand beyond $\sim H$, since most of energy has already been transferred
to acceleration and heating of ambient gas and lost via radiative cooling
at the time of shell formation.
\citet{2017ApJ...834...25K} argued that
$\beta_h< 1$ is expected quite generally
unless the temporal and spatial correlations of SNe
are extremely enhanced compared to their mean values,
requiring more than a factor of 40 elevation
compared to the average conditions in self-regulated galactic disks
where $\Sigma_{\rm SFR} \ge 10^{-3}\sfrunit$.
Although our simulation does exhibit large temporal fluctuations
(see Figure~\ref{fig:flux_tevol}),
the peak upward fluctuation compared to the mean star formation rate
is only a factor of 5. Further systematic investigations for different
galactic conditions will be needed to confirm whether the
predictions for low $\beta_h$ apply universally in star-forming galaxies.
If local or global conditions make star formation inherently extremely
bursty, then $\beta_h$ may be higher.
\subsection{Comparison with Observations}\label{sec:loading_obs}
The low mass loading factor of the hot gas in our simulations
is comparable to
({or slightly smaller than})
the observed mass loading factor estimated
in the best studied local starburst M82. Using Chandra X-ray observations,
\citet{2009ApJ...697.2030S} constrained
the ``central'' mass loading factor of the hot gas to
about $\beta_h \sim 0.3-1$,
with the ``central'' energy loading factor about
$\alpha_h\sim0.3-1$.
Here, they constrained quantities using
a large number of hydrodynamical models to explain
the observed diffuse and hard X-rays, which come from the central 500pc region.
\footnote{However, we note that in their modeling the hot wind
freely expands into a very tenuous medium rather than expanding
into a dense ISM. By comparison, it is evident
e.g. in Fig. 8 of \citet{2017ApJ...834...25K} that before superbubble
breakout from the warm/cold ISM, the hot gas has very high velocity but
mostly remains subsonic. This suggests that lack of a dense surrounding
medium in a hydrodynamic wind comparison model might lead to a
Mach number higher than would be realistic, and hence
for density and temperature constrained by emission properties
the mass loss rate could be overestimated by a factor of a few.}
It is difficult to make a direct comparison,
but if we consider the state of the hot medium within the
energy-injecting layer ($d<H$) in our simulations, we have $\beta_h\sim 0.5$
and $\alpha_h\sim 0.1-0.2$. Although there are no systematic
observational studies of mass loading factors of the hot gas,
$\beta_h<1$ is suggested for a wide range of star formation rates
from dwarf starbursts to ultraluminous infrared galaxies,
utilizing the \citet{1985Natur.317...44C} wind model
with observational constraints of
the X-ray luminosity and star formation rates
\citep{2014ApJ...784...93Z}.
For the solar neighborhood conditions investigated in the present simulation,
the warm gas accelerated by the SNe cannot reach velocities
fast enough to escape the gravitational potential.
Typically, star formation bursts launch
warm outflows with velocity up to $\sim 100\kms$,
but with most gas at lower velocity (see Figure~\ref{fig:warm_vpdf})
which can climb to $\sim 1\kpc$ but
no further (compare the kinetic and gravitational curves in
Figure~\ref{fig:warm_flux}(c)).
Without additional energy and momentum input at high-$|z|$, SN feedback alone
cannot drive warm winds in the Milky Way-type galaxies.
Above $d=1\kpc$, Figure~\ref{fig:loading}(a) shows that $\beta_w<1$.
Even though relatively fast warm gas could escape from dwarf galaxies
with a shallower potential well than the Milky Way,
Equation (\ref{eq:fit}) implies that $\beta_w\sim 1/3$
for $\vout>50\kms$ at most.
This suggests that achieving $\beta>1$ in dwarfs, as appears necessary
to explain current-day galaxy-halo relationships and cosmic history,
would require an additional acceleration mechanism such as interaction
with an outflowing cosmic ray fluid
\citep[e.g.,][]{2013ApJ...777L..38H,2016ApJ...816L..19G,2017ApJ...834..208R}.
In cosmic-ray driven winds,
SN feedback may still be crucial for pushing warm gas out to
large $d$ where cosmic ray pressure gradients are sufficient to
produce efficient acceleration (Mao \& Ostriker 2017, submitted).
Observational studies of warm phase outflows
(possibly including the ionized phase according to our definition)
indicate a wide range of the mass loading factor, $\beta_{w+i}\sim 0.1-10$
\citep[e.g.,][]{2015ApJ...809..147H,2017MNRAS.469.4831C},
with systematically decreasing
trends for increasing galaxy mass, circular velocity, and total SFR.
As the level of wind mass loading may vary with both local conditions
(including $\Sigma_{\rm gas}$ and $\Sigma_{\rm *}$) as well as the
global gravitational potential, it will be quite interesting to measure outflow
properties in simulations under widely varying galactic conditions,
for comparison with current empirical scaling trends and future observations.
\subsection{Wind Driving Simulation Context}\label{sec:loading_sim}
Recently, a number of other research
groups have performed simulations with similar
local Cartesian box setups to study galactic outflows driven by SNe
\citep[e.g.,][]{2013MNRAS.429.1922C,2015MNRAS.446.2125C,2016ApJ...816L..19G,2016MNRAS.456.3432G,2016MNRAS.459.2311M,2017ApJ...841..101L,2017MNRAS.466.1903G}.
Most of these simulations have adopted fixed SN rates,
while varying the SN placement (e.g. random vs. in high density regions).
In contrast, in our simulation, SN rates and locations are self-consistent with
star formation, which we believe is crucial in creating realistic
multiphase outflows.
\citet{2016MNRAS.456.3432G} ran a set of simulations with solar neighborhood
conditions, focusing on the effect of the SN placement and of SN clustering.
The authors
emphasize that due to the short duration of their simulations ($100\Myr$),
definitive
conclusions cannot be drawn regarding wind driving. However, their
measurements of outflow in the $d=1\kpc$ plane can be compared to
our fountain flow measurements.
For a range of different SN feedback treatments,
they report mass loading factors $\sim 5-10$.
Although this is larger than what we find
{($\beta_w\sim 1$ at $d\sim 1\kpc$)},
their simulations do show a decline in outflow over time,
{and our measurements are at $t>250\Myr$.}
As pointed out by the authors, the majority of the outflowing mass is
relatively dense
($n_H\sim 0.1$)
and moves slowly ($\vout\sim20-40\kms$),
similar to the properties
of our warm fountain at low $d$. Their fraction of high-velocity
gas ($\vout>500\kms$) is only $2-8\times10^{-5}$, which correspondingly
reduces the mass loading factor of gas that is certain to escape to
large distance.
In simulations including cosmic rays,
\citet{2016ApJ...816L..19G} found slowly moving
($\vout\sim10-50\kms$) warm outflows with a mass loading factor near unity.
The SN-only comparison model of \citet{2016ApJ...816L..19G}
had an order of magnitude lower mass loading, with
just a fast, low density component
passing through the $d=1\kpc$ surface where the mass flux is measured.
While this comparison suggests that cosmic rays may be crucial to
accelerating warm outflows, a concern is that the role of SNe may not
have been properly captured in \citet{2016ApJ...816L..19G}.
Due to the high computational cost of including cosmic rays, relatively
low spatial resolution $\Delta x\sim16\pc$ was adopted for this pair of
comparison simulations.
The reported mass outflow rate for SN-only models appears to
differ significantly between the high resolution
($\Delta x\sim4\pc$) simulation of \citet{2016MNRAS.456.3432G},
with mass flux at $d=1\kpc$ of $4\times10^{-2}\sfrunit$, and the
low-resolution ($\Delta x \sim 16\pc$) simulation of
\citealt{2016ApJ...816L..19G}, with mass flux of $10^{-3}\sfrunit$.
In our own resolution study for a similar parameter regime (see Paper~I),
we found that numerical convergence of SN driven ISM properties
is not guaranteed at $\Delta x\sim16\pc$, and the low-resolution
outcomes are quite
sensitive to the exact prescription for SN feedback.
This suggests that higher resolution simulations will be needed
to assess the role of interactions between the cosmic ray fluid and
gas in driving galactic winds.
The \citet{2017MNRAS.466.1903G} models are most similar to our
simulations in that they self-consistently model star formation and SN
feedback, rather than adopting a fixed SN rate. The main conclusion
they draw is that the outflow rate strongly depends on the volume
filling factor of hot gas. Above a 50\% volume filling factor, their
measured mass loading factors at $d=1\kpc$ are $\sim 1-100$
(but note that this may be exaggerated since loading is evaluated
instantaneously rather than based on temporal averages; see discussion after
Equation~\ref{eq:energy_loading} in Section~\ref{sec:loading_result}).
While the simulation durations of \citet{2017MNRAS.466.1903G} are
$<100\Myr$, such that star formation and wind mass-loss rates are
likely both subject to ``startup'' transient effects, we agree with
the conclusion that significant driving of fountain flows
(which dominate at $d=1\kpc$ based on our work) is associated with prominent
superbubbles near the disk midplane. In Paper~I, we found the hot
gas fills $\sim 20-60\%$ of the volume at $|z|<H$.
\citet{2016MNRAS.459.2311M} performed simulations for solar neighborhood
conditions, as well as environments with higher gas surface density and
SN rates. For high-velocity gas ($\vout>300\kms$) that could potentially
escape, they reported mass loading factors of
0.02-0.005 (lower for higher gas surface density models).
They also noted the absence of wind acceleration from subsonic
to supersonic velocities as a limitation of local Cartesian box simulations.
We have argued in Section~\ref{sec:wind}
that provided the hot-gas Bernoulli
parameter and mass flux have both approached constant values at large $d$
in a Cartesian simulation, they can be combined to make predictions for
wind properties at large distance. In this case, the asymptotic
mass flux at high velocity would be larger than the near-disk value
as high-enthalpy gas is further accelerated when streamlines open, but
the total mass flux of hot gas would change little. However, it is not
possible to compare hot-gas mass fluxes
as \citet{2016MNRAS.459.2311M} did not separate by phase.
\citet{2017ApJ...841..101L} conducted simulations
using fixed SN rates, but decomposed by thermal phase in
reporting mass and energy (as well as metal) loading factors.
Their measurements are at $d=1-2.5\kpc$, with their ``warm'' component
including most of what we consider ``ionized,'' and their ``hot'' extending
down to $T=3\times10^5\Kel$, slightly lower than our ``hot'' definition.
For their solar neighborhood model, the hot and total
mass loading factors are about $0.8$ and $2-3$, which are larger than ours by
a factor of 3 to 8
(see also their discussion in comparison to work of
\citealt{2013MNRAS.429.1922C,2015MNRAS.446.2125C,2016MNRAS.456.3432G}).
The energy loading factor is also about an order of
magnitude larger in their model.
{The reason for this large difference in mass and energy loading
with respect to our findings is mainly due to the difference in the
vertical scale height of SNe (relative to the gas scale height), as
also pointed out in \citet{2017ApJ...841..101L}. By placing SNe
randomly with a fixed SN scale height of 250pc, we find that we are
able to reproduce their results (see
Appendix~\ref{sec:appendix_sn}). When we adopt a fixed SN
scale height of 250pc, the majority of SN
explosions occur outside of the main gas layer. Each SN remnant can then
expand into the hot, rarefied disk atmosphere with little
interaction with warm gas. Most of the injected energy is carried
outward before cooling. The
resulting ISM properties are also quite different: the gas scale height
is smaller ($H\sim150\pc$), star formation rates are higher, and the
hot and warm/cold phases are almost completely segregated (single
phase outflow). In Appendix~\ref{sec:appendix_sn}, we also test
models with no runaway O stars, and with
random SN placement with a smaller scale height of 50pc. Both alternatives
result in loading factors and warm gas velocity distribution very similar to
the fiducial model.
Within the standard TIGRESS framework, the spatial distribution of SNe is
subject to the adopted prescription for runaways, but is otherwise
self-consistently determined with respect to the gas
by the distribution of star formation sites.
Better theoretical and observational constraints for runaway OB
stars would lead to more accurate modeling of the SN distribution,
which in principle could change the wind mass loading. However,
our tests suggest that a very large proportion of high-velocity
runaways would be needed to significantly increase the wind mass-loading,
while the corresponding star formation rate and ISM phase segregation
in that case might not be consistent with observations.}
Finally, we remark that a fine enough grid to spatially resolve both
low filling-factor warm gas and high filling-factor hot gas in the
wind launching region is crucial for proper physical characterization
of galactic winds. If warm and hot gas are artificially mixed,
e.g. if the flow in AMR
and semi-Lagrangian simulations
moves from a higher resolution
region near the midplane to a lower resolution region at high
latitude, the result can be unphysical in ways that would compromise
the implications for real galactic systems. For example, {consider
mixing of warm and hot flows that have
total horizontally-averaged mass, momentum, and energy vertical fluxes of
$F_\rho$, $F_{\rho v}$, and $F_E$.
If we further assume a steady state and neglect
gravity and
magnetic fields, the
outgoing fluxes of the mixed gas must be the same as the sum of the
horizontally-averaged incoming fluxes of the warm and hot gas. That is,
$F_\rho = \rho_{\rm mix} v_{\rm mix} = \langle \rho_w v_w\rangle + \langle\rho_h v_h\rangle$, and similarly
$F_{\rho v} = \rho_{\rm mix} v_{\rm mix}^2 + P_{\rm mix}$
and
$F_{E} = \rho_{\rm mix}v_{\rm mix}
[v_{\rm mix}^2/2 + \gamma P_{\rm mix}/((\gamma-1)\rho_{\rm mix})]$
are the sum of ``warm'' and ``hot'' terms based on horizontal averages
over multiple zones.}
{
The post-mixing mean velocity will depend on the total fluxes as
\begin{equation}\label{eq:vmix}
v_{\rm mix} = \frac{\gamma}{\gamma+1}\frac{F_{ \rho v}}{F_\rho}
\sbrackets{1-\rbrackets{1-\frac{2(\gamma^2-1)}{\gamma^2}\frac{F_E F_{\rho}}{F_{\rho v}^2}}^{1/2}}.
\end{equation}
Typically, $F_\rho$ (near the disk)
will be dominated by the warm medium contribution (see Fig.
\ref{fig:loading}a), $F_{\rho v}$ will be dominated by the hot medium
contribution
(see Fig. \ref{fig:zprof}c), and
$F_E$ will be dominated by the hot medium contribution (see Fig.
\ref{fig:loading}b).
}
{Figure~\ref{fig:vmix} shows this hypothetical
post-mixing velocity as calculated from
our simulation. This is based on horizontal averages of the
outgoing (i.e. zones with $sign(v_z)=sign(z)$) fluxes, temporally
averaged over the outflow period and over both sides of the disk.
Also shown, for comparison, are the mean outflow velocities of the warm and
the hot gas.
The implication of the comparison shown in Figure~\ref{fig:vmix}
is that if numerical mixing were to occur
at $d>1\kpc$, the result would be
$v_{\rm mix}\sim100\kms$, which is intermediate between the
mean hot and mean warm outflow velocities.
}
Depending on the galaxy/halo global
properties, this could have two
(opposite) unphysical consequences. On the one hand, if $v_{\rm mix}$
is greater than the galaxy escape speed but the original $v_w$ is less
than the galaxy escape speed, then the artificially-mixed flow would be
able to escape with a much larger mass-loading factor than would be
realistic (i.e. $\sim \beta_w$ rather than $\beta_h$). On the other
hand, if $v_{\rm mix}$ is smaller than the galaxy escape speed, then the
artificially-mixed flow would not be able to escape at all, whereas in
reality the hot wind should escape with mass-loading factor $\beta_h$,
and potentially loaded with more than its share of metals.
To avoid these unphysical consequences, it is necessary to separately resolve
the multiphase gas even above the dense midplane region.
\begin{figure}
\plotone{fig10.pdf}
\caption{The hypothetical post-mixing velocity $v_{\rm mix}$ (black)
if artificial
numerical mixing were to happen at a distance $d$ from the midplane.
We use horizontally-averaged outgoing fluxes to calculate
$v_{\rm mix}$ using Equation (\ref{eq:vmix}), and take
a time average over the outflow period. Also shown are mean values of
the warm-gas and hot-gas outflowing velocities.
The nominal value of $v_{\rm mix}$ is similar to the warm medium velocity
at small $d$, and gets closer to the hot medium velocity at large $d$.
\label{fig:vmix}}
\end{figure}
\section{SUMMARY}\label{sec:summary}
Gas flows blown out of galactic midplane regions
by energetic stellar feedback, most
notably type II SNe, will either escape as a wind or turn around
as a fountain, depending on the specific kinetic and thermal energy of the gas
compared to the gravitational potential.
Due to the multiphase structure of the ISM,
the simplest steady, adiabatic solutions \citep[e.g.,][]{1985Natur.317...44C}
are not directly applicable, although aspects of these solutions are
informative when considering a phase-decomposed analysis.
More generally, the question of ``how much mass and energy are carried out
by each phase of outflowing gas?'' depends strongly on when and where
star formation and SNe occur and how feedback transfers energy to the ISM.
Given the complexity involved, the properties of galactic
winds and fountain flows created by the star-forming ISM must
be investigated via fully self-consistent numerical simulations.
The TIGRESS implementation described in Paper~I provides
comprehensive, self-consistent simulations of the multiphase
star-forming ISM that can be used for high-resolution investigations
of outflows in a wide variety of environments. In this paper, we
analyze the solar neighborhood model of Paper~I to characterize
vertical
gas flows phase-by-phase.
Our principal finding is that the outflowing gas consists of a superposition
of a hot wind and a warm fountain flow. We summarize key conclusions from
our analysis of each component below.
\subsection{\it Hot wind}
1. The hot gas component, which is created by SN shocks, behaves very similarly
to expectations for a steady, adiabatic flow as it expands away from
the midplane.
{For the horizontally-averaged, time-averaged hot component}, the
mass flux and the Bernoulli parameter are close to constant as a
function of the distance from the midplane $d$ above $d>1\kpc$, indicating
that there is relatively
little mass added or subtracted by heating or cooling from/to
another phase, and little energy added or subtracted either through shocks,
radiation, {pressure work on other phases}, or mixing.
2. The hot gas mass loading factor, defined as the
ratio of outflowing mass flux to the star formation rate per unit area,
is $\beta_h\sim 0.1$, decreasing by
$40\%$ from $d=1$ to $4\kpc$. The measured value of $\beta_h$ is
consistent with estimates from idealized numerical experiments of
superbubble expansion in a warm-cold ISM.
3. Above $d=1\kpc$, the hot outflow contains a tiny fraction of the
energy originally injected in the ISM by SNe. Another tiny fraction
is converted to kinetic energy of the warm and
cold medium accelerated by superbubble expansion, but most of
the original SN energy is
radiated away within the disk scale height.
The ratio of outflowing hot-gas energy flux to the
SN energy injection rate per unit area is $\alpha_h\sim 0.02$,
decreasing by
$65\%$ from $d=1$ to $4\kpc$.
For $d=2-4\kpc$, the mean temperature of the hot gas is $\sim1.5\times10^6\Kel$,
and its mean velocity is
$\sim200\kms$.
Magnetic energy in the outflow is negligible,
with mean Alfv\'en speed of $\sim30\kms$.
4. In the present simulation, the energy
of the hot medium
is carried mainly as heat because the outflow is constrained to remain
vertical in our local Cartesian box. Allowing for the opening of streamlines
at larger distance (beyond our simulation domain), much of the enthalpy
would be converted to kinetic energy. The hot medium is barely affected
by gravity within the simulation domain, and with an asymptotic maximum
wind speed of
{$\sim \sqrt{2 ({\cal B} - \Phi)}\sim 350\kms$}
would easily escape to very large distances.
More generally, local simulations with varying galactic conditions
can provide measures of $\beta_h$ and the Bernoulli parameter $\mathcal{B}$
that can be used to provide predictions for hot wind properties
at large distances in a wide range of galaxies.
\subsection{\it Warm Fountain}
1. The warm (including both neutral and ionized) medium, which dominates
the total gas mass in the simulation, acquires energy from SN feedback,
mostly in kinetic form. However, the typical outgoing velocity of the warm
medium above $d=1\kpc$ is only
$v_{\rm out}\sim 60\kms$, which is not enough to overcome
the large-scale gravitational potential in the present simulation.
As a result, most of the warm gas that is blown out of the midplane
eventually turns around and falls back, forming a fountain.
2. Star formation bursts lead to a succession of
outflow-dominated and inflow-dominated periods of the fountain flow.
During both outflow-dominated and inflow-dominated periods, gas
with both signs of velocity (i.e. outflows and inflows)
is present on both sides of the disk. Considering just the outflowing gas,
owing to deceleration and turnaround the
outgoing warm-medium mass loading factor $\beta_w$ is a
\emph{decreasing} function of height $d$.
The mean warm-gas mass flux in outflowing ``fountain'' gas at $d\sim 1\kpc$ is
$\beta_w\sim1$, but this drops steeply to $\beta_w\sim0.03$ at $d=4\kpc$.
3. The value of
{the warm-medium energy loading factor}
$\alpha_w$ drops from $\sim 0.002$ to $10^{-4}$
over $d=1$ to $4\kpc$. Because $\alpha_w$ drops slightly less than
$\beta_w$, the mean specific energy of the warm medium increases with $d$.
The corresponding
(volume-weighted)
mean velocities of the warm gas at $d=1$ and $4\kpc$ are
$\sim60$ and $80\kms$. However, it is important to note that
the increase of mean velocity in the warm medium with $d$ is primarily
due to dropout of low-velocity fluid elements, and \textit{does not} reflect
acceleration with height.
Detailed distributions show a secular decrease in the mass and mass flux
of high-velocity warm gas with increasing $d$.
4. A promising way to characterize warm outflows is via the velocity
distribution where they are launched at $d\sim H$. Here, we find that
the high-velocity warm gas has an exponential distribution. For a given
PDF in velocity, the portion of the warm gas mass flux that is able to
escape as a wind will depend on the halo potential depth. For the
large scale galactic potential in the present simulation,
very little warm gas escapes, but in
a dwarf galaxy the same distribution could lead to a wind with
$\beta_w\sim 1/3$.
\subsection{Caveats and prospects}
Our simulations have two main caveats for direct comparison with observations.
Firstly, we use a local Cartesian box to achieve high resolution.
The uniformly high resolution is crucial for distinguishing different
phases and limiting numerical mixing.
However, the local Cartesian box prevents us from
following the hot outflow's evolution under global geometry with a
realistic galactic potential \citep[see][]{2016MNRAS.459.2311M}.
Without the opening of streamlines,
the hot medium cannot accelerate through a sonic point to reach its
asymptotic velocity; we
therefore cannot follow this process directly in our simulations.
However, proper decomposition of the gas phases allows us to provide
well-defined mass and energy loading factors for the hot gas,
which can be robustly extrapolated to obtain predictions for
asymptotic wind properties at large scales.
The hot-gas loading factors
{at large $d$ are slightly lower than} the
observational constraints deduced
from M82 \citep{2009ApJ...697.2030S}.
{This is likely because in M82 the strong starburst has successfully
cleared much of the cooler gas away from the midplane; we find that when
SNe explode above the denser phases, the mass and energy loading of hot
winds increases.}
As discussed above, to the extent that the warm phase is ballistic
above a scale height, its measured velocity distributions could be used to
extrapolate its properties to large distance. However, it is clear
that the gravitational potential even within our local box affects the
warm-medium properties at large $d$, so it is important to treat conclusions
regarding warm-medium outflows cautiously.
Secondly, in the present simulations we do not include photoionization.
This affects the properties of the warm and ionized phases,
for which the photoionization heating
can be important. Inclusion of photoionization is necessary for
direct comparison with observations, where the line diagnostics are sometimes
better explained by photoionization rather than a shock model
\citep[e.g.,][]{2016MNRAS.457.3133C}. As a first step towards this,
it may be sufficient
to compute ionization in post-processing, as photoionization
is unlikely to be important
to the dynamics of high-velocity warm gas even though it may dominate heating.
Most observations of highly ionized absorption lines are however based on
large apertures, so direct theoretical comparisons would also require
simulations with global geometry.
We are able to run our simulations over an extended period, long after
initial transients that may affect quantitative results reported
by others for
outflows in simulations with similar physics and resolution to
the TIGRESS implementation, but shorter durations. Other recent
high-resolution simulations do not have self-consistent star formation
and feedback, which may affect outflow loading because this is
sensitive to the spatio-temporal correlation of supernovae with ISM
gas of different phases. In the future, as more groups run
high-resolution simulations with self-consistent star formation and SN
feedback for an extended duration, it will be informative to
compare phase-separated results for mass and energy loading, Bernoulli
parameters, and velocity distributions, for both fountain flows and
winds.
Application of the TIGRESS implementation to other galactic
environments is currently underway, and promises to be quite interesting.
Varying the basic model
input parameters (especially gas and stellar surface density and
metallicity) will enable predictions for multiphase outflow properties
in a wide range of galaxies, and will provide detailed information needed
to build subgrid models for winds in cosmological simulations of galaxy
formation. By extending TIGRESS and other self-consistent multiphase
ISM/star formation numerical implementations to include additional feedback
(especially radiation and cosmic rays), understanding the complex physics
behind galactic winds and fountains is within reach.
\acknowledgements
{We are grateful to the referee for helpful report, and to
Miao Li and Drummond Fielding for fruitful discussions.}
This work was supported by grants AST-1312006 from the National Science
Foundation, and NNX14AB49G and NNX17AG26G from NASA.
Resources supporting this work were provided in part by the NASA
High-End Computing (HEC) Program through the NASA Advanced
Supercomputing (NAS) Division at Ames Research Center and in part by
the Princeton Institute for Computational Science and Engineering
(PICSciE) and the Office of Information Technology's High Performance
Computing Center.
\software{This work made use of the {\tt Athena} MHD code \citep{2008ApJS..178..137S,2009NewA...14..139S}.
This work also made use of analysis and visualization softwares including {\tt yt \citep{2011ApJS..192....9T},
astropy \citep{2013A&A...558A..33A}, matplotlib \citep{Hunter:2007}, numpy \citep{vanderWalt2011},
IPython \citep{Perez2007}, pandas \citep{mckinney-proc-scipy-2010}}, and {\tt VisIt} \citep{HPV:VisIt}.}
\bibliographystyle{aasjournal}
|
\section{Introduction}
The question of interpolating a time-sequence of shapes with a curve and representing a shape evolution with few parameters have been addressed in the literature related to shape analysis and medical imaging since the last ten years. Several methods have been proposed and studied and they essentially rely on extension of standard tools available in Euclidean geometry to shape spaces. In this direction, we mention geodesic regression, cubic regression, kernel methods... The generalization of these tools to infinite dimensional setting are sometimes complicated by the fact that the shape space is not a flat space, nor a finite dimensional space. However, the shape space is usually endowed with a Riemannian structure and most often in infinite dimensions. The generalizations of this Euclidean tools are often introduced by variational formulations, the simplest example being the case of shortest path between two shapes, i.e. geodesics on the space of shapes. Even in that particular example, finding a variational setting in which the object of interest is well defined is of interest, since the existence of an extremum is not guaranteed in general and is complicated by the infinite dimensional setting. For example, in the case of group of diffeomorphisms, this question is addressed in \cite{CompletenessDiffeomorphismGroup}, in which the authors prove that the group of diffeomorphisms of the Euclidean space endowed with a right-invariant Sobolev metric of high enough order is complete in the sense of the Hopf-Rinow theorem. The case of the group of diffeomorphisms with right-invariant metric is relevant for applications in medical imaging and in particular for the problem of diffeomorphic image matching \cite{laurentbook,0855.57035,BegIJCV}. It is also natural to study and develop higher-order interpolations in the space of shapes, which has been actively developed in finite dimensions \cite{birkhoff65,VariationalStudySplines,Noakes1,splinesCk,Crouch,splinesanalyse,Koiso}. It was also extensively used and numerically developed in image processing and computer vision \cite{MumfordElastica,Masnou1,PamiSplines,Samir,CAD,Chan02eulerselastica}. In the past few years, these higher-order models have been introduced in biomedical imaging for interpolation of a time sequence of shapes. They have been proposed in \cite{TrVi2010} for a diffeomorphic group action on a finite dimensional manifold and further developed for general invariant higher-order lagrangians in \cite{HOSplines1,HOSplines2} on a group. A numerical implementation together with a generalized model have been proposed in \cite{SinghVN15} in the context of medical imaging applications. However, in all these articles, the question of existence of an extremum is not treated. An attempt is given in \cite{Vialard2016} where the exact relaxation of the problem is shown on the group of diffeomorphisms of the interval $[0,1]$. The main result of \cite{Vialard2016} consists in providing the existence of a minimizer in a larger space where the relaxation is defined. Although it does not completely solve the problem, it shows that existence of cubic splines for a group of diffeomorphisms with a right-invariant metric is non trivial. Let us discuss where the difficulty comes in a Riemannian setting. Riemannian splines are minimizers of
\begin{equation}\label{SplinesFunctional}
\mathcal{J}(x) = \int_0^1 g\left( \frac{D}{Dt} \dot{x},\frac{D}{Dt}\dot{x} \right) \,\mathrm{d} t\,,
\end{equation}
where $(M,g)$ is a Riemannian manifold, $\frac{D}{Dt}$ is its associated covariant derivative and $x$ is a sufficiently smooth curve from $[0,1]$ in $M$ satisfying first order boundary conditions, i.e. $x(0),\dot{x}(0)$ and $x(1),\dot{x}(1)$ are fixed.
The term $$\frac{D}{Dt} \dot{x} = \ddot{x} + \Gamma(x)(\dot{x},\dot{x})$$ (written in coordinates, with $\Gamma$ the Christoffel symbols) contains nonlinearities which contribute in the variational problem \eqref{SplinesFunctional} by possibly generating high-frequency oscillations in the space variable.
\par
Although this notion of Riemannian cubics could not be well defined in general, it is possible to slightly modify it to make it well-posed. A modification of this type has recently been proposed in \cite{WirthSplines} in their framework.
In this paper, we propose a simple variational setting which makes the second-order variational problem well-posed at the expense of increasing the regularity of the group of diffeomorphisms on which the second-order interpolation is feasible. For practical applications, this gain of smoothness, or loss of controllability of the diffeomorphism does not matter so much since smoothness is preferred in medical image registration. However, the theoretical existence is guaranteed. The main result of the paper is the following
\begin{theorem}[Main result]
Let $\Omega$ be a bounded domain in ${\mathbb R}^d$ and $s'\geq s+1$.
There exists a minimizer to the functional
\begin{equation}\label{SplinesFunctional}
\mathcal{J}(x) = \int_0^1 \left\| \frac{D}{Dt} \dot{x} \o x\i \right\|^2_{H^{s'}} \,\mathrm{d} t\,,
\end{equation}
on the loop space $\Omega_{0,1}(\mc G_{H^{s'}(\Omega,{\mathbb R}^d)})$ where $ \frac{D}{Dt}$ is the covariant derivative associated with the right-invariant $H^s$ metric.
\end{theorem}
In other words, the acceleration is measured in a stronger space than the ambient space so that it will prevent from creating oscillations.
\section{Background on right-invariant metrics on diffeomorphisms group} \label{Background}
\par \textbf{Sobolev right-invariant metrics on the group of diffeomorphisms.} In \cite{CompletenessDiffeomorphismGroup}, the authors proved the following theorem
Let $M$ be either ${\mathbb R}^d$ or a compact manifold without boundary of dimension $d$. We define hereafter a group of diffeomorphisms of $M$ which is a complete metric space.
Consider a space $V$ a Hilbert space of vector fields on $M$ (rapidly decreasing at infinity in the unbounded case), left invariant by their flows, such that the inclusion map $V \hookrightarrow W^{1,\infty}(M,{\mathbb R}^d)$ is continuous. This hypothesis implies that the flow of a time dependent vector field in $L^2([0,1],V)$ is well defined, see \cite[Appendix C]{laurentbook}. Then, the set of flows at time $1$ defines a group of diffeomorphisms denoted by $\mathcal{G}_V$. Denoting
\begin{equation}
\on{Fl}_1(\xi) = \varphi(1)
\end{equation}
where $\varphi$ solves the flow equation
\begin{align} \label{EqFlow}
&\partial_t \varphi(t,x) = \xi(t,\varphi(t,x))\\
&\varphi(0,x) = x\, \text{ } \forall x \in D\,,
\end{align}
we define
\begin{equation}\label{EqDefGV}
\mathcal{G}_V \ensuremath{\stackrel{\mbox{\upshape\tiny def.}}{=}} \{ \varphi(1) \,: \, \exists \, \xi \in L^2([0,1],V) \text{ s.t. } \on{Fl}_1(\xi) \}\,,
\end{equation}
which has been introduced by Trouvé in \cite{0855.57035}. On this group, Trouvé defines a metric
\begin{equation}
\on{dist}(\psi_1,\psi_0)^2 = \inf_{} \left\{\int_0^1 \| \xi \|_V^2 \,\mathrm{d} t \,: \, \xi \in L^2([0,1],V) \text{ s.t. } \psi_1 = \on{Fl}_1(\xi) \circ \psi_0 \,\right\}
\end{equation}
under which he proves that $\mathcal{G}_V$ is complete. In full generality, that is for a general space of vector fields $V$, very few properties are known on this group. For instance, it is a priori not a topological group, or more precisely, there is no known topological structure making it a topological group (the inversion need not be continuous). Moreover, there does not need to be a differentiable structure on this group.
However, for certain choices of spaces $V$, such structures are available and therefore more properties can be derived in this situation. Indeed, consider the group $\mc D^s(M)$, with $s > d/2+1$, which consists of all $C^1$-diffeomorphisms of Sobolev regularity $H^s$. It is known since the work of Ebin and Marsden \cite{Ebin1970} that $\mc D^s(M)$ is a smooth Hilbert manifold and a topological group. It only remains to prove that $\mathcal{G}_{H^s} = \mc D^s(M)_0$ (the connected component of identity) which is done in \cite[Section 8]{CompletenessDiffeomorphismGroup} and its main result is
\begin{theorem}\label{Thmdiff_hopf_rinow}
Let $M$ be ${\mathbb R}^d$ or a closed manifold and $s > d/2+1$. If $G^s$ is a smooth, right-invariant Sobolev-metric of order $s$ on $\mc D^s(M)$, then
\begin{enumerate}
\item
$(\mc D^s(M), G^s)$ is geodesically complete;
\item
$(\mc D^s(M)_0, \on{dist}^s)$ is a complete metric space;
\item
Any two elements of $\mc D^s(M)_0$ can be joined by a minimizing geodesic.
\end{enumerate}
The statements also hold for the subgroups $\mc D^s_\mu(M)$ and $\mc D^s_\om(M)$ of diffeomorphisms preserving a volume form $\mu$ or a symplectic structure $\om$.
\end{theorem}
The crucial ingredient in the proof is showing that the flow map
\begin{equation}
\label{eq:intro_flow}
\on{Fl}_t : L^1(I, \mf X^s(M)) \to \mc D^s(M)
\end{equation}
exists and is continuous.
\par
In \cite{TrVi2010}, we introduced the use of cubic splines \textit{in the space of shapes} to interpolate a sequence of shapes that are time dependent.
Riemannian cubics (also called Riemannian splines) and probably more famous, its constrained alternative called Elastica belong to a class of problems that have been studied since the work of Euler (see the discussion in \cite{MumfordElastica}). Let us present the variational problem in a Riemannian setting. Riemannian splines are minimizers of
\begin{equation}\label{SplinesFunctional}
\mathcal{J}(x) = \int_0^1 g\left( \frac{D}{Dt} \dot{x},\frac{D}{Dt}\dot{x} \right) \,\mathrm{d} t\,,
\end{equation}
where $(M,g)$ is a Riemannian manifold, $\frac{D}{Dt}$ is its associated covariant derivative and $x$ is a sufficiently smooth curve from $[0,1]$ in $M$ satisfying first order boundary conditions, i.e. $x(0),\dot{x}(0)$ and $x(1),\dot{x}(1)$ are fixed.
The case of Elastica consists in restricting the previous optimization problem to the set of curves that are parametrized by unit speed (when the problem is feasible), namely $g(\dot{x},\dot{x})=1$ for all time. To the best of our knowledge, the only paper that deals with analytical questions is \cite{splinesanalyse} where the authors show in particular the existence of minimizers of a second-order functional on the space curves on a complete finite dimensional Riemannian manifold.
In \cite{HOSplines1}, higher-order models are proposed on groups of diffeomorphisms but for the standard Riemannian cubics functional, no analytical study was provided. Indeed, in the case of a Lie group $G$ (and $\mathfrak{g}$ its Lie algebra) with a right-invariant metric ($\| \cdot \|_\mathfrak{g}$ denoting the norm on the Lie algebra), the covariant derivative can be written as follows:
Let $V(t) \in T_{g(t)}G$ be a vector field along a curve $g(t) \in G$
\begin{equation}\label{Covariant_derivative_vf_prop}
\frac{D}{Dt} V =\Big( \dot{\nu} + \frac{1}{2} \operatorname{ad}^\dagger_\xi \nu + \frac{1}{2} \operatorname{ad}^\dagger_\nu \xi - \frac{1}{2} [\xi, \nu]\Big) _G(g).
\end{equation}
where $\operatorname{ad}^\dagger$ is the metric adjoint defined by
\begin{equation}\label{ad-dagger-def}
\operatorname{ad}^\dagger_{\nu}{\kappa} :=
(\operatorname{ad}^*_\nu (\kappa^\flat))^\sharp
\end{equation}
for any $\nu, \kappa\in \mathfrak{g}$ and $\flat$ and $\sharp$ are the musical operator for the
cometric and metric operator. Therefore, the reduced lagrangian for \eqref{SplinesFunctional} is
\begin{equation}\label{ReducedSplinesFunctional}
\mathcal{J}(x) = \int_0^1 \| \dot{\xi} + \ad^\dagger_\xi \xi\|^2_\mathfrak{g} \,\mathrm{d} t\,.
\end{equation}
where $\operatorname{ad}^\dagger$ is the metric adjoint, i.e.,
it is written as
\begin{equation}\label{ad-dagger-def}
\operatorname{ad}^\dagger_{\nu}{\kappa} :=
(\operatorname{ad}^*_\nu (\kappa^\flat))^\sharp
\end{equation}
for any $\nu, \kappa\in \mathfrak{g}$.
We can also formulate the variational problem on the dual of the Lie algebra $\mathfrak{g}^*$ by
\begin{equation}\label{ReducedSplinesFunctionalDualLieAlgebra}
\mathcal{J}(x) = \int_0^1 \| a(t) \|^2_{\mathfrak{g}^*} \,\mathrm{d} t\,,
\end{equation}
under the constraint
\begin{equation}
\dot{m} + \ad^*_{\xi} m = a\,.
\end{equation}
In infinite dimensions, there is a clear obstacle to use reduction since the operator $ \ad^\dagger$ is unbounded on the tangent space at identity due to a loss of derivative.
However, using the smooth Riemannian structure on $\mc D^s$, functional \eqref{SplinesFunctional} is well defined.
The following proposition of \cite{splinesanalyse} is valid in infinite dimensions:
\begin{proposition}
Let $(M,g)$ be an infinite dimensional Riemannian manifold and $$\Omega_{0,1}(M) := \{ x \in H^2([0,1],M) \, | \, x(i) = x_i \,,\, \dot{x}(i) = v_i \text{ for } i=0,1 \}$$ be the space of paths with the first order boundary constraints for given $(x_0,v_0) \in TM$ and $(x_1,v_1) \in TM$.
\noindent
The functional \eqref{SplinesFunctional} is smooth on $\Omega_{0,1}(M)$ and
\begin{equation}
\mathcal{J}'(x)(v) = \int_0^1 g(\frac{D^2}{Dt^2}\dot{v},\frac{D}{Dt}\dot{x}) - g(R(\dot{x},\frac{D}{Dt}\dot{x}),v) \, \,\mathrm{d} t\,.
\end{equation}
A critical point of $\mathcal{J}$ is a smooth curve that satisfies the Riemannian cubic equation
\begin{equation}
\frac{D^3}{Dt^3}\dot{x} - R(\dot{x},\frac{D}{Dt}\dot{x})\dot{x} = 0\,.
\end{equation}
\end{proposition}
The critical points of $\mathcal{J}$ are sometimes called Riemannian cubics or cubic polynomials. The existence of minimizers does not follow from the corresponding proof in \cite{splinesanalyse} since it strongly relies on the finite dimension hypothesis to have compactness properties.
\section{The main result}
We formulate the main result on the flat torus but it can be generalized to bounded domains in ${\mathbb R}^d$ in a straightforward way.
\begin{theorem}\label{ThmSplines}
Let $\ms T^d$ be the $d$ dimensional flat torus and $s'\geq s+1$.
There exists a minimizer to the functional
\begin{equation}\label{SplinesFunctional}
\mathcal{J}(x) = \int_0^1 \left\| \frac{D}{Dt} \dot{\varphi} \o \varphi \i \right\|^2_{H^{s'}} \,\mathrm{d} t\,,
\end{equation}
on the loop space $\Omega_{0,1}(\mc G_{H^{s'}(\ms T^d,{\mathbb R}^d)})$ where $ \frac{D}{Dt}$ is the covariant derivative associated with the right-invariant $H^s$ metric.
\end{theorem}
Before proving the theorem, we prove the following lemmas:
\begin{lemma}\label{ThmExistenceLagrangian}
Let $\alpha \in L^2([0,1],H^s)$, then there exists a unique solution defined on $[0,1]$ to the system
\begin{subequations} \label{ReducedForcedEvolution}
\begin{align}
&\dot{\varphi} = v\\
&\dot{v} = -\Gamma(\varphi)(v,v) + \alpha \circ \varphi \,, \label{Acceleration}
\end{align}
\end{subequations}
for given initial conditions $\varphi(0) = \varphi_0 \in D^{s+1}$ and $v(0) = v_0 \in H^{s+1}$.
\end{lemma}
\begin{proof}
Solutions exist for short time since the system is Lipschitz on $\mc D^s \times H^s$ and the existence theorem for Caratheodory equation gives the result.
\\
Existence for all time is not guaranteed a priori since the second equation of is quadratic in $v$. Let us denote $u = v \circ \varphi^{-1}$ and $f(t):= \frac 12 g(\varphi)(v,v) = \frac 12 \| u \|_{H^s}^2$. Deriving it in time gives $f'(t) = \langle \alpha(t), u(t)\rangle_{H^s}$, so that by the Cauchy-Schwarz inequality, we get
\begin{align}\label{GronwallEnergy}
& f(t) \leq \int_0^t \sqrt{f(z)} \| \alpha \|_{H^s} \,\mathrm{d} z \leq \| \alpha \|_{L^2([0,1],H^s)} \sqrt{\int_0^t f(z) \,\mathrm{d} z}\\
& f(t) \leq \| \alpha \|_{L^2([0,1],H^s)} \left(1 + \int_0^t f(z) \,\mathrm{d} z \right) \,.
\end{align}
Using Gronwall's lemma, it implies that $f(t)$ is bounded for on $[0,T]$ where $T$ is the supremum (possible blow-up) time of definition. Therefore, $r := \int_0^T \| u \|^2_{H^s} \, \,\mathrm{d} t = \int_0^T f(t) \,\mathrm{d} t < \infty$ which means that $u \in L^2([0,T],H^s)$. As a consequence, for all time $t \in [0,T[$, $\varphi(t) \in B(x_0,r)$, which is the ball of radius $r$ for the geodesic distance on $D^s$. In addition, $\lim_{t\to T} \varphi(t)$ is well defined since $D^s$ is metrically complete.
Remark that $\Gamma(\varphi(t))$ is bounded (uniformly in time) as an operator on $H^s \times H^s$ since $\Gamma$ is smooth on $\mc D^s$ and thus continuous on the path $\varphi(t)$. In particular, the right-hand side of the equation \eqref{Acceleration} belongs to $L^1([0,T],H^s)$ so that $v(T):= \int_0^T - \Gamma(\varphi(t))(v(t),v(t)) + \alpha(t) \o \varphi(t) \,\mathrm{d} t$. Using short time existence on $(\varphi(T),v(T))$, the solution can be extended for short time from time $T$ so that in fact $T=1$.
\end{proof}
In the following lemma, we study the solutions of the system \eqref{ReducedForcedEvolution} but written on the dual of the tangent space at identity. As mentioned in Section \ref{Background}, one can rewrite the minimization as in Equation \eqref{ReducedSplinesFunctionalDualLieAlgebra}, however the "dual" acceleration is measured using with the corresponding dual norm. In our case, the dual norm (w.r.t. to $H^s$) associated with $H^{s+1} \subset H^s$ is $(H^{s-1})^* \subset (H^s)^*$ as can be seen by a direct computation in Fourier spaces.
\begin{lemma}\label{ThmExistenceEulerian}
Let $a \in L^2([0,1],(H^{s-2})^*)$ then the following integral equation
\begin{equation}\label{IntegralFormulation}
m(t) = Ad_{g(t)^{-1}}^*(m(0)) + \int_0^t Ad_{g_{t,s}}^*(a(s)) ds \,,
\end{equation}
with initial condition $m(0) \in (H^{s-2})^*$ has a unique solution in $C^0([0,T],(H^{s})^*)$.
If $a \in L^2([0,1],(H^{s-1})^*)$, then there exists a solution to the integral equation with initial condition in $m(0) \in (H^{s-1})^*$.
\end{lemma}
\begin{proof}
The proof of the existence for $a \in L^2([0,1],(H^{s-2})^*)$ follows a standard fixed point method. Let $\Psi : C^0([0,T],(H^{s})^*) \to C^0([0,T],(H^{s})^*)$ defined by Formula \eqref{IntegralFormulation} namely
\begin{equation} \label{IntegralFormulaBis}
\Psi(m)(t) = Ad_{g(t)^{-1}}^*(m(0)) + \int_0^t Ad_{g_{t,s}}^*(a(s)) ds\,,
\end{equation}
where $g_{t,s}$ is the flow of diffeomorphims generated by the vector field associated with $m$.
Remark first that $\Psi(m)$ lies in $C^0([0,T],(H^{s})^*)$ which is well-defined since the integrand is integrable.
Now we claim that the map is a contraction on $(C^0([0,T],(H^{s})^*),\|\cdot\|_{\infty})$ for $T$ small enough,
\begin{multline}
\|\Psi(m_1) - \Psi(m_2)\|_{\infty} \leq \sup_{t \in [0,T]} \| Ad_{g_1(t)^{-1}}^*(m(0)) - Ad_{g_2(t)^{-1}}^*(m(0)) \|_{(H^{s})^*} + \\
\int_0^T \|Ad_{{g_1}_{t,s}}^*a(s)-Ad_{{g_2}_{t,s}}^*a(s)\|_{(H^{s})^*} \, ds \,.
\end{multline}
We need to estimate for $\alpha \in (H^{s-2})^*$ and $w \in H^{s}$. There exists a constant $M_0 > 0$ such that
\begin{align*}
\langle Ad_{g_1}^*(\alpha) - Ad_{g_2}^*(\alpha),w \rangle_{L^2} &\leq \| \alpha \|_{(H^{s-2})^*} \| Ad_{g_1}(w) - Ad_{g_2}(w) \|_{H^{s-2}}\\
&M_0 \| \alpha \|_{(H^{s-2})^*} \| g_1 - g_2 \|_{H^{s-2}} \| w\|_{H^{s}}
\end{align*}
Moreover, there exists a constant $M_1 > 0$ s.t. $$\sup_{s,t \in [0,T]} \| {g_1}_{t,s} - {g_2}_{t,s} \|_{H^{s-2}} \leq M_0 \| m_1 - m_2\|_{L^2([0,T],(H^s)^*)}\,. $$
Therefore there exists $M_1$ s.t.
\begin{equation} \label{LipProp}
\|\Psi(m_1) - \Psi(m_2)(t)\|_\infty \leq M_1 \sqrt{t} \| m_1 - m_2\|_{\infty} ( \| m(0) \|_{(H^s)^*} + \int_0^T \|a(s)\|_{(H^s)^*} \, ds ) \,.
\end{equation}
Hence the map $\Psi$ is a contraction for $t$ small enough and it therefore proves the existence and uniqueness of a solution of formula \eqref{IntegralFormulation} for short times.
Remark that Equation \eqref{LipProp} gives an upper bound $t_{lip} := \left(M_1 (\| m(0) \|_{W_2^*} + \int_0^T \|a(s)\|_{W_2^*} \, ds) \right)^{-2}$ such that for every $t<t_{lip}$, $\Psi$ is a contraction. In addition this upper bound is valid at any time $t \in [0,T]$.
Then the existence and uniqueness until time $T$ follows straightforwardly by an iterative application of the short-time result.
\par
The second part of the proof consists in showing existence of solutions for $a \in L^2([0,1],(H^{s-1})^*)$ which is done using a compactness argument. Let $a_n$ converging to $a$ in $L^2([0,1],(H^{s-1})^*)$, then the solution $m_n \in C^0([0,T],(H^{s})^*)$ actually belongs to $H^1([0,1],(H^{s-1})^*)$ because $\| Ad_g^*(m) \|_{(H^{s-1})^*} \leq M_2 \| g \|_{H^s} \| m \|_{(H^{s-1})^*}$. By the Aubin-Lions-Simon theorem, $H^1([0,1],(H^{s-1})^*)$ is compactly embedded in $L^2([0,1],(H^{s})^*)$, thus one can extract a strongly convergent sequence in $L^2([0,1],(H^{s})^*)$. By theorem \ref{Thmdiff_hopf_rinow}, the flow associated with the momentum $m_n$, denoted by $g_{s,t}^n$ strongly converges in $H^s$ (actually uniformly in $s,t$) to $g_{s,t}$ the flow associated with the limit $m$. Then, it implies that the integrand $Ad_{g^n_{t,s}}^*(a_n(s))$ in Formula \eqref{IntegralFormulaBis} converges to $Ad_{g_{t,s}}^*(a(s))$ in $(H^{s-1})^*$. Since the integrand is bounded uniformly, the Lebesgue convergence theorem applies and the result is obtained.
\end{proof}
\begin{remark}
\begin{enumerate}
\item The reason why we are not able to treat the case of $a \in L^2([0,1],(H^{s-1})^*)$ is because the flow map in Theorem \ref{Thmdiff_hopf_rinow} is only continuous and (possibly) not Lipschitz.
\item Note that Lemma \ref{ThmExistenceEulerian} can be considered as the Eulerian version of Lemma \ref{ThmExistenceLagrangian} and the latter achieves a better result since uniqueness of the solution is proven in $H^{s+1}$. However, our proof of the main theorem will require the use of Lemma \ref{ThmExistenceEulerian} which gives the fact that the Eulerian velocity of the solutions of Lemma \ref{ThmExistenceLagrangian} are bounded in $H^1([0,1],(H^s)^*)$.
\end{enumerate}
\end{remark}
\begin{lemma}\label{ThmWeakComposition}
Let $s > d/2 + 1$,
$\alpha_n \in H^s$ weakly converging to $\alpha$ and $\varphi_n \in \on{Diff}^s$ which strongly converges to $\varphi$, then the composition $\alpha_n \circ \varphi_n$ weakly converges to $\alpha \circ \varphi$.
\end{lemma}
\begin{proof}
We prove the weak convergence by proving that the sequence is bounded and that it weakly converges on a dense set of $H^s$. First, remark that $\| \alpha_n \circ \varphi_n\|_{H^s}$ is bounded in $H^s$ since the composition by a diffeomorphism in $\on{Diff}^s$ is bounded (see \cite[Lemma 2.2]{CompletenessDiffeomorphismGroup}).
\par
Let $m \in (H^{s})^* \cap \mathcal{M}$ where $\mathcal{M}$ denotes the space of Radon measures, consider $\langle \alpha_n \circ \varphi_n , m \rangle_{L^2}$, which can be written by a change of variable as
\begin{equation}
\langle \alpha_n \circ \varphi_n , m \rangle_{L^2} = \langle \alpha_n , (\varphi_n)_*(m) \rangle_{L^2}\,,
\end{equation}
where $(\varphi_n)_*(m)$ is the pushforward of $m$ by $\varphi_n$. Since $\varphi_n$ is strongly convergent in $\on{Diff}^s$, we have that $(\varphi_n)_*(m)$ strongly converge in $(H^{s})^*$ to $\varphi_*(m)$. Therefore, $\langle \alpha_n , (\varphi_n)_*(m) \rangle_{L^2}$ converges to $\langle \alpha , \varphi_*(m) \rangle_{L^2}$, which gives the result.
\end{proof}
\begin{proof}[Proof of the theorem]
First note that the space $\Omega_{0,1}(\mc G_{H^{s'}(\Omega,{\mathbb R}^d)})$ is non-empty: Consider a path connecting $\varphi_0$ and $\varphi_1$ in $\mc G_{H^{s'}(\Omega,{\mathbb R}^d)}$, thus, by concatenation of paths, the problem is reduced to $\varphi_0 = \varphi_1 $ where the path can be easily defined on the tangent space in a neighborhood of $\on{Id}$ or in a local chart.
We will use natural coordinates, i.e. $\on{Id} + H^{s}$ to describe elements of the loop space $\Omega_{0,1}(\mc G_{H^{s'}(\Omega,{\mathbb R}^d)})$.
The term $\left\| \frac{D}{Dt} \dot{\varphi} \o \varphi\i \right\|^2_{H^{s'}}$ can be written in coordinates $$\| \left( \ddot{\varphi} + \Gamma(\varphi)(\dot{\varphi},\dot{\varphi}) \right) \circ \varphi^{-1} \|_{H^{s'}}^2$$ so that, it is natural to introduce the change of variable $\left( \dot{v} + \Gamma(\varphi)(v,v) \right) \circ \varphi^{-1} := \alpha$ where $\dot{\varphi} = v$.
The variational problem \eqref{SplinesFunctional} can be rewritten as the minimization of
the functional defined on the Hilbert space $L^2([0,1],H^{s'})$
\begin{equation}
\ell (\alpha) = \int_0^1 \| \alpha(t) \|^2_{H^{s'}} \,\mathrm{d} t \,,
\end{equation}
under the constraint
\begin{equation} \label{ReducedForcedEvolution}
\begin{cases}
\dot{\varphi} = v \\
\dot{v} = -\Gamma(\varphi)(v,v) + \alpha \o \varphi \,,
\end{cases}
\end{equation}
and the boundary conditions, $\varphi(0) = \varphi_0$, $\varphi(1) = \varphi_1$ and $v(0) = v_0$, $v(1) = v_1$.
\par
This functional is lower semi-continuous on $L^2([0,1],H^{s'})$. Let $\alpha_n$ be a minimizing sequence in $L^2([0,1],H^{s'})$ weakly converging to $\alpha$. The condition to be checked is the constraints that have to be satisfied at the limit. Using Lemma \ref{ThmExistenceEulerian}, the sequence $v_n \circ \varphi_n^{-1} \in H^1([0,1],H^{s-1})$ is bounded and one can extract a strongly converging sequence in $C^0([0,1],H^{s})$ which implies the strong convergence of $\varphi_n(1)$ and $v_n(1)$ in $H^s$. The first consequence is that the boundary constraints $\varphi(1),v(1)$ are satisfied at the limit. It also implies that the term $\Gamma(\varphi_n)(v_n,v_n)$ is strongly convergent in $H^s$ and the term $\alpha_n \circ \varphi_n$ is weakly convergent to $\alpha \circ \varphi$, using Lemma \ref{ThmWeakComposition}. Therefore, we have the equality $v(t) = v(0) + \int_0^t -\Gamma(\varphi(s))(v(s),v(s)) + \alpha \circ \varphi \,\mathrm{d} s$, which implies that the couple $(\varphi,v)$ is the solution of the integral equation associated with System \eqref{ReducedForcedEvolution}.
\end{proof}
\begin{theorem}[Spline interpolation of time sequences]
Let $\varphi_1,\ldots,\varphi_n$ be $n$ diffeomorphisms in $\on{Diff}_0^{s+1}$ and $t_1 <\ldots < t_n$ be a sequence of $n$ positive reals. There exists a path $\varphi(t) \in \on{Diff}_0^{s+1}$, which minimize the acceleration functional
\begin{equation}\label{EqSplinesSequence}
\| \dot{\varphi}(0)\circ \varphi\i \|^2_{H^{s'}} + \int_{t_1}^{t_n} \left\| \frac{D}{Dt} \dot{\varphi} \o \varphi\i \right\|^2_{H^{s'}} \,\mathrm{d} t\,,
\end{equation}
among all curves satisfying $\varphi(t_i) = \varphi_i$ for $i \in 1,\ldots,n$.
\end{theorem}
\begin{proof}
The proof is similar to that of Theorem \ref{ThmSplines} and we do not repeat the arguments here. Note that the penalization on the initial speed seems necessary in order for the curve to stay in a bounded metric ball.\footnote{On the flat 2D torus, straight lines with irrational slopes are dense and they can be parametrized with arbitrarily high velocity so that the infimum of \eqref{EqSplinesSequence} is $0$ without speed penalization.}
\end{proof}
\section{Conclusion}
This theoretical proof of existence was provided to fill in the gap of the variational models proposed in \cite{SinghVN15}. However, we do have treated the case of the induced metric on the space of images, which was also implemented in \cite{SinghVN15}. However, with minor modifications, the approach developed in this article can possibly adapted. On a more theoretical point of view, we leave the open question if the approach can be adapted for $s'>s$.
\bibliographystyle{alpha}
\small{ |
\section{Introduction}
\label{sec:introduction}
In this paper, we analyze the massless Spin-Boson model which is a simple but non-trivial model of quantum field theory. It can be seen as a model of a two-level atom interacting with its second quantized scalar field, and hence, provides a widely employed model for quantum optics.
The unperturbed energies of the two-level atom shall be denoted by real numbers $e_0<e_1$. After switching on the interaction with a massless scalar field that may induce transitions between the atom levels, the free ground-state energy $e_0$ is shifted to the interacting ground-state energy $\lambda_0$ while the free excited state with energy $e_1$ turns into a resonance with complex energy $\lambda_1$.
Unfortunately, in the massless case, neither $e_0$ nor $e_1$ are isolated points in the spectrum of the free Hamiltonian which is why the interacting ground-state and the resonance ($\lambda_0$ and $\lambda_1$) cannot be constructed using standard results from regular perturbation theory.
Several technologies were developed to overcome this difficulty. Two succesful methods that recently received a lot of attention are: The so-called multiscale method (see e.g.\ \cite{pizzo1,pizzo2,bbp,bach}) and the renormalization group method (see e.g.\ \cite{bfs1,bfs2,bfs3,bcfs,bfs100,bbf,fgs100,feshbach,s,f,bffs}). In both cases, a family of spectrally dilated Hamiltonians is analyzed since this allows for complex eigenvalues.
In this paper, we will employ the multiscale technique. This technique invokes an infrared cut-off which is then removed using an inductive scheme. In each step of the induction, lower and lower boson momenta are added to the interaction and regular perturbation theory is used to construct the respective ground-state and resonance. In order to reach the limit of no infrared cut-off good control over the closing gap is essential.
Note that such a procedure has been introduced for the construction of resonances in the Pauli-Fierz model \cite{bach,bbp}. In this work we also construct resonances (and ground-state eigenvalues) but this is not our main purpose. Our main purpose is to prove that resonances are analytic with respect to the dilation parameter and coupling constant, and furthermore, to provide certain spectral and resolvent estimates that allow for the control of the dynamics including the scattering regime. In a forthcoming paper, these estimates are employed to address scattering theory for the model at hand. We want to remark that in \cite{bmw} the time evolution of this model is studied using the spectral renormalization method. Some results derived therein are similar to some of ours, however, utilizing different methods, respectively.
What we call resonances and ground-states multiscale analysis is an inductive construction of a sequence of Hamiltonians that enjoy infrared cutoffs and satisfy certain properties. As the parameter of the sequence tends to infinity these cutoffs are removed.
Our mutliscale analysis is the content of Theorem \ref{thm:ind}. Its basic scheme is presented in Section \ref{sec:indscheme} and the proof of Theorem \ref{thm:ind} is carried out in Section \ref{sec:proof-induction}. Our proofs of analyticity and resolvent and spectral estimates are not part of our multiscale analysis employed in the contruction of the resonance and ground-state energy, they only use it as mathematical input. The latter results are presented in Section \ref{resolvent-estimates} (spectral and resolvent estimates) and Section \ref{analyticity} (analyticity). Theorem \ref{thm:ind} is only an intermediate but necessary step. As we mention above, the method of multiscale analysis for resonances is introduced in \cite{bach,bbp}. However, the results in \cite{bach,bbp} cannot be used directly to prove analyticity because many of the estimations therein consider the dilation parameter, $\theta$, to be purely imaginary whereas analyticity requires estimates that are uniform for $\theta$ in an open set. Thus, although it does not involve major obstacles, for the sake of analyticity, many of the given calculations and some of the proofs need to be redone. For the convenience of the reader and in order to keep this work self-contained we provide them in the proof of Theorem \ref{thm:ind}. Note that Theorem \ref{thm:ind} is applied to the Spin-Boson model while \cite{bach,bbp} address the Pauli-Fierz model. This gives us the opportunity to review the multiscale technique for a non-trivial but more tractable model.
In Section \ref{resolvent-estimates} (in particular, in Subsection \ref{resspec}) we introduce a new inductive scheme that is used to study resolvent and spectral estimates. This scheme is independent and different from the scheme used to construct resonances. It allows to localize the spectrum in two cones with vertices at the location of the resonance and ground-state energy, respectively, and allows for arbitrary small apex angles provided the coupling constant is sufficiently small. We want to emphasize that such a result requires a more subtle analysis than localizing the spectrum in cuspidal domains. Additionally, we provide estimates for the resolvent operator in the vicinity of the cones.
The study of analytic properties of resonances and ground sate eigenvalues in the context of non-relativistic quantum field theory has been the source of several studies. These papers use the method of spectral renormalization.
In \cite{gh} a large class of models of quantum field theory was analyzed and analyticity of the ground-state with respect to the coupling constant was proven under the assumption this ground-state is non-degenerate. The existence of an unique ground-state and its analyticity with respect to the coupling constant was shown in \cite{hasler1} for the Spin-Boson model without an infrared regularization, and in \cite{hh101} for the Pauli-Fierz model. Furthermore, in \cite{bffs}, a model describing the interaction of an atom with its quantized electromagnetic field was studied and it was proven,
that the excited states are analytic functions of the momentum
of the atom and of the coupling constant. Likewise, in \cite{ah}, it is shown that the ground-state energy of the translationally invariant Nelson model is an analytic function of the coupling constant and the total momentum.
Here, to the best of our knowledge, we give the first extension of the multiscale method that provides a ready access to analyticity properties that essentially amounts to proving it for isolated eigenvalues only and exploiting that uniform limits of analytic functions are analytic.
In \cite{bfs1,bfs2,bfs3}, the renormalized group technique was invented and applied in order to construct the ground state and resonances for the confined Pauli-Fierz model. Moreover, resolvent and spectral estimates were obtained therein. Based on this new method, several simplifications and applications were developed in a variety of works \cite{bcfs,bfs100,bffs,feshbach,fgs100,bbf,bffs,s,f,bbf}. The multiscale analysis was first invented in \cite{pizzo1,pizzo2} and then adapted in order to gain access to spectral and resolvent estimates and the construction of ground-states in \cite{bach,bbp}. In \cite{bmw}, resolvent and spectral estimates are derived in order to control the time evolution in the Spin-Boson model and, in \cite{bf}, smoothness of the resolvent and local decay of the photon dynamics for quantum states in a spectral interval just above the ground state energy was proven.
Next, we will introduce the massless Spin-Boson model together with the well-known mathematical facts and tools in order to present our main results in Section \ref{sec:mainresult}.
\subsection{The Spin-Boson model}
\label{sec:defmodel}
In the rest of this Section 1 we will state preliminary definitions and well-known tools and facts from which we start our analysis. \\
The non-interacting Spin-Boson Hamiltonian is defined as
\begin{align}
\label{h0def}
H_0:=K \otimes \mathbbm 1 + \mathbbm 1 \otimes H_f , \qquad K:= \begin{pmatrix}
e_1 & 0 \\
0 & e_0
\end{pmatrix} ,
\qquad
H_f:=\int \mathrm{d^3}k \, \omega(k) a(k)^* a(k) .
\end{align}
We regard $K$ as an idealized free Hamiltonian of a two-level atom. As already stated in the introduction, its two energy levels are denoted by the real numbers $ 0 = e_0 <e_1$. Furthermore, $H_f$ denotes the free Hamiltonian of a massless scalar field having dispersion relation $\omega(k)=|k|$, and $a,a^*$ are the annihilation and creation operators on standard Fock space which will be defined below. In the following we will sometimes call $K$ the atomic part, and $H_f$ the free field part of the Hamiltonian.
The sum of the free two-level atom Hamiltonian $K$ and the free field Hamiltonian $H_f$ will simply be referred to as the ``free Hamiltonian'' $H_0$.
The interaction term reads
\begin{align}
V:= \sigma_1\otimes \left( a(f) + a(f)^*\right) ,
\end{align}
where the boson form factor is given by
\begin{align}
f: \mathbb{R}^3 \setminus \{0\}\to \mathbb{R} , \qquad k\mapsto e^{-\frac{k^2}{\Lambda^2}}|k|^{-\frac{1}{2}+\mu} .
\label{eq:f}
\end{align}
Note that the relativistic form factor of a scalar field should rather be $f(k)=(2\pi)^{-\frac{3}{2}}(2|k|)^{-\frac{1}{2}}$, which however renders the model ill-defined due to the fact that such an $f$ would not be square integrable. This is referred to as ultraviolet divergence. In our case, the Gaussian factor in \eqref{eq:f} acts as an ultraviolet cut-off for $\Lambda>0$ being the ultraviolet cut-off parameter and, in addition, the fixed number
\begin{align}
\label{const:mu}
\mu \in (0, 1/2)
\end{align}
implies a regularization of the infrared singularity at $k=0$ which is a technical assumption chosen to keep the proofs of this work more tractable. With a lot of additional work one can also treat the case $\mu=0$ using methods described in \cite{bbkm}. The missing factor of $2^{-\frac{1}{2}}(2\pi)^{-\frac{3}{2}}$ will be absorbed in the coupling constant in our notation.
The full Spin-Boson Hamiltonian is then defined as
\begin{align}
\label{eq:H}
H:= H_0 + g V
\end{align}
for some coupling constant $g \in \mathbb{C} $. The operator $H$ is densely defined on the Hilbert space
\begin{align}
\mathcal H := \mathcal K \otimes \mathcal F\left[ \mathfrak{h}\right] , \qquad \mathcal K:= \mathbb C^2 ,
\end{align}
where
\begin{align} \label{Fock}
\mathcal F\left[ \mathfrak{h}\right] := \bigoplus^\infty_{n=0} \mathcal F_n\left[ \mathfrak{h}\right] ,
\qquad
F_n\left[ \mathfrak{h}\right] :=
\mathfrak{h}^{\odot n},
\qquad
\mathfrak h:= L^2(\mathbb R^3,\mathbb{C})
\end{align}
denotes the standard bosonic Fock space and superscript $\odot n$ denotes the n-th symmetric tensor product where by convention $\mathfrak{h}^{\odot 0}\equiv\mathbb{C}$. Throughout this work we will use the notation $\mathbb{N}_0:=\mathbb{N}\cup\{0\}$. Thanks to the direct sum, an element $\Psi \in \mathcal F[\mathfrak{h}]$
can be represented as a family $(\psi_n)_{n\in\mathbb{N}_0}$ of wave functions $\psi_n \in \mathfrak{h}^{\odot n}$. The state $\Psi$ with $\psi_0=1$ and $\psi_n=0$ for all $n\geq 1$ is called the vacuum and is denoted by
\begin{align}
\label{Omega}
\Omega:=(1,0,0,\dots)\in \mathcal F\left[ \mathfrak{h}\right] .
\end{align}
For any $h\in \mathfrak{h}$ and $\Psi = (\psi_n)_{n\in\mathbb{N}_0} \in \mathcal F[\mathfrak{h}]$, we define the creation operator
\begin{align}
\left( a(h)^*\Psi \right)_{n\in\mathbb{N}_0}:=\left( 0,h\odot \psi_0 , \sqrt{2} h\odot \psi_1, \dots \right), \quad \left( a(h)^*\Psi \right)_{n} = \sqrt{n} h \odot \psi_{n-1} , \quad \forall n\in\mathbb{N} ,
\end{align}
and the annihilation operator $a(h)$ as the respective adjoint. Occasionally, we shall also use the physics notation
\begin{align}\label{aastar}
a(h)^*=\int \mathrm{d^3}k \, h(k) a(k)^* , \qquad h\in \mathfrak{h},
\end{align}
where formally, the action of these operators in the $n$ boson sector of a vector $\Psi = (\psi_n)_{n\in\mathbb{N}_0} \in\mathcal F[\mathfrak{h}]$ can be seen as:
\begin{align}
\left(a(k) \psi \right)^{(n)}(k_1,...,k_n)&=\sqrt{n+1} \psi^{(n+1)}(k,k_1,...,k_n) \\
\left(a(k)^* \psi \right)^{(n)}(k_1,...,k_n)&=\frac{1}{\sqrt{n}}\sum^n_{i=1} \delta^{(3)}(k-k_i) \psi^{(n-1)}(k_1,...,\tilde k_i,...,k_n).
\end{align}
Here, the notation $\tilde \cdot $ means that the corresponding variable is omitted.
Note that $a$ and $a^*$ fulfill the canonical commutation relations:
\begin{align}
\label{eq:ccr}
\left[a(h),a^*(l) \right]=\left\langle h, l\right\rangle_{\mathfrak{h}} , \qquad \left[a (h),a(l) \right]=0 , \qquad \left[a^*(h),a^*(l) \right]=0 \qquad \forall h,l\in\mathfrak{h} .
\end{align}
Let us recall some well-known facts about the introduced model.
Clearly, $K$ is self-adjoint on $\mathcal K$ and its spectrum
consists of two eigenvalues $e_0$ and $e_1$. The corresponding eigenvectors are
\begin{align}
\label{varphi}
\varphi_0= \begin{pmatrix}
0 \\ 1
\end{pmatrix} \qquad \text{and} \qquad \varphi_1= \begin{pmatrix}
1 \\ 0
\end{pmatrix} \qquad \text{with} \qquad K \varphi_i =e_i \varphi_i , \quad i=0,1.
\end{align}
Moreover, $H_f$ is self-adjoint on its natural domain $\mathcal D(H_f)\subset \mathcal F[\mathfrak{h}]$ and its spectrum $\sigma (H_f)= [0, \infty )$
is absolutely continuous (see \cite{reedsimon2}). Consequently, the spectrum of $H_0$ is given by
$\sigma (H_0)= [e_0, \infty )$, and $e_0,e_1$ are eigenvalues embedded in the absolutely continuous part of the spectrum of $H_0$ (see \cite{reedsimon1}).
Finally, also the closedness of the full Hamiltonian $H$ is well-known (see e.g.\ \cite{spohnspin}), however, for the sake of completeness, we give a proof in the Appendix \ref{app:sa}.
\begin{proposition}
\label{thm:Hsa}
The operator $V$ is relatively bounded by $H_0$ with infinitesimal bound and, consequently, $H$ is a closed operator on the domain $\mathcal D(H) = \mathcal K \otimes \mathcal D(H_f )$.
\end{proposition}
\begin{remark}\label{R} In this work we omit spelling out identities whenever unambiguous. For every vector spaces $V_1$, $V_2$ and operators $ A_1 $ and $A_2$ defined on $V_1$ and $V_2$, respectively, we identify
\begin{equation}\label{iden}
A_1 \equiv A_1 \otimes \mathbbm 1_{V_2}, \hspace{2cm} A_2 \equiv \mathbbm 1_{V_1} \otimes A_2 .
\end{equation}
In order to simplify our notation further, and whenever unambiguous, we do not utilize specific notations for every inner product or norm that we employ.
\end{remark}
\subsubsection{Complex dilation}
In the following we introduce the tools necessary for the complex dilation of Hamiltonians that allows to study resonances as eigenvalues.
\begin{definition}
For every $\theta \in \mathbb R$, we define the unitary transformation
\begin{align}
u_\theta: \mathfrak{h}&\rightarrow \mathfrak{h}
, \qquad \psi(k) \rightarrow e^{-\frac{3\theta}{2}} \psi(e^{-\theta}k) .
\end{align}
Similarly, we define its second quantization
$U_\theta: \mathcal F [\mathfrak{h}]\rightarrow \mathcal F [\mathfrak{h}]$
by its action on $\mathcal F_n [\mathfrak{h}]$:
\begin{align}
U_\theta \Psi^{(n)}(k_1,...,k_n)=e^{-\frac{3\theta}{2}n}\Psi^{(n)}(e^{-\theta}k_1,...,e^{-\theta}k_n) .
\end{align}
\end{definition}
A straight forward calculation yields, for every $\theta \in \mathbb{R}$,
\begin{align}
\label{dila}
a^\theta(h): =U_\theta a(h) U^{-1}_\theta= e^{-\frac{3}{2}\theta}a(h_\theta) \qquad \text{where} \qquad h_\theta(k):= h(e^{-\theta}k)
\end{align}
for $h\in \mathfrak{h}$
and (again, for $\theta \in \mathbb{R}$)
\begin{align} \label{Htheta}
H^\theta := U_\theta H U^{-1}_\theta = H_0^{\theta} + g V^\theta,
\end{align}
where
\begin{align}
H_0^{\theta} : = K \otimes \mathbbm 1 + \mathbbm 1 \otimes H^\theta_f, \hspace{1cm}
H_f^\theta= e^{-\theta} H_f , \qquad V^\theta= \sigma_1 \otimes \left(a(f^{\overline \theta})+ a(f^{ \theta})^* \right)
\end{align}
and
\begin{align}
f^\theta: \mathbb{R}^3\to\mathbb{R} , \quad k\mapsto e^{-\theta (1+\mu)} e^{-e^{2\theta}\frac{k^2}{\Lambda^2}}|k|^{-\frac{1}{2}+\mu}.
\end{align}
The expressions for $ H_f^{\theta} $ and $ f^{\theta} $ above are well-defined for $\theta \in \mathbb{C}$. It follows from Appendix \ref{app:sa} that, as long as
$ f^\theta ,f^\theta/\sqrt{\omega} \in \mathfrak{h}$, $V^{\theta}$ is relatively bounded, with infinitesimal bound, with respect to $H_0^{\theta}$ provided, e.g., $ |\theta| < \pi /16$. In this case we define $ H^{\theta} $ using the right hand side of Eq.\ \eqref{Htheta}. Employing similar bounds as in Appendix \ref{app:sa} it turns out that $ H^{\theta} $ is a closed operator and $ \mathcal D (H^{\theta}) = \mathcal D (H_0 )$. Then, it is easy to see that the family $ \left\lbrace H^\theta \right\rbrace_{ |\theta | < \pi /16 } $ is an analytic family of type $A$. Notice that the expression in the middle of \eqref{Htheta} does not make sense for non-real $\theta$ because we have not defined $ U_{\theta} $ for those $\theta$'s (we can define it, but it turns out to be an unbounded operator and therefore the meaning of the middle of \eqref{Htheta} is still unclear).
From the explicit formula of $ H_{0}^{\theta} $ we deduce that it has only two eigenvalues, namely $e_0$ and $e_1$, and
\begin{equation}
\sigma ( H_{0}^{\theta} ) = \Big \{ e_i + e^{-\theta} r \: : \: r \geq 0, i \in \{0, 1 \} \Big \}.
\end{equation}
In this text we use the notation
\begin{align}
\label{eq:def-disc}
D(x, r):=\left\{ z\in\mathbb{C} : |z-x|<r \right\}, \qquad x\in \mathbb{C} , r>0 ,
\end{align}
and we will assume in the remaining of this work that $\theta \in D(0, \pi /16) $.
\subsubsection{Infrared cut-offs and definition of the ground-state and the resonance}
In the following, we introduce a family of Hamiltonians $H^{(n),\theta}$ which have two isolated (complex) eigenvalues $\lambda^{(n)}_i$ in small neighborhoods of $e_i$, $ i \in \{ 0, 1 \}$. For every $n$, $H^{(n),\theta}$ enjoys an infared cutoff that is removed as $n$ tends to infinity.
\begin{definition}
\label{def:Hntheta} We fix a real number $ \boldsymbol{\nu} \in (0, \pi/16) $ and for every $\theta \in \mathbb{C}$ we set $ \nu :=\Im \theta $.
We define
\begin{align}
\label{def:setS}
\mathcal S:=\left\{\theta\in\mathbb{C}: -10^{-3} < \Re \theta < 10^3 \quad \text{and} \quad \boldsymbol{\nu} < \Im \theta <\pi/16 \right\} .
\end{align}
For $\theta\in \mathcal S$ and $n\in \mathbb{N}$, we define:
\begin{enumerate}
\item[(i)] The sequence of infrared cut-offs $\{\rho_n\}_{n\in\mathbb{N}}$ with $\rho_n:=\rho_0 \rho^n$ for real $ 0 < \rho_0 < \min( 1, e_1/4) $ and $0<\rho<1$. In Definition \ref{sequence} below we specify additional properties of it.
\item[(ii)]
The cutoff-Hilbert space of one particle, $ \mathfrak{h}^{(n)} $:
\begin{align}
\mathfrak{h}^{(n)}:=L^2(\mathbb{R}^3\setminus \mathcal B_{\rho_n}), \quad \mathcal B_{\rho_n}:=\left\{ x\in\mathbb{R}^3 : |x|<\rho_n \right\} .
\end{align}
The Fock space with one particle sector $ \mathfrak{h}^{(n)} $ is defined as in Eq.\ \eqref{Fock}, and we denote it by $ \mathcal F[\mathfrak{h}^{(n)}] $. We denote its vacuum state by $\Omega^{(n)} $. We set
\begin{align}
\mathcal H^{(n)}:=\mathcal K \otimes \mathcal F[\mathfrak{h}^{(n)}] .
\end{align}
The free boson energy operator with an infrared cutoff is is defined on $ F[\mathfrak{h}^{(n)}] $ by Eq.\ \eqref{h0def}, we denote it by $ H_f^{(n), 0} \equiv H_f^{(n)}$. We set
\begin{align}\label{Hntheta}
H_f^{(n), \theta }:= e^{-\theta} H_f^{(n), 0}.
\end{align}
For every function $ h \in \mathfrak{h}^{(n)} $ we define creation an annihilation operators, $a_n(h), \, a_n^*(h) $, on $ F[\mathfrak{h}^{(n)}] $ according to Eq.\ \eqref{aastar}. We use the same formula for functions $ h \in \mathfrak{h} $, then it is understood that we take the restriction of $h$ to
$ \mathbb{R}^3\setminus \mathcal B_{\rho_n} $.
We define the following family of Hamiltonians (densely defined on $ \mathcal H^{(n)} $ - see Remark \ref{R})
\begin{align} \label{operators111}
H_0^{(n), \theta} : = K + H^{(n), \theta}_f, \hspace{1cm} V^{(n), \theta} := \sigma_1 \otimes \left(a_n(f^{\overline \theta})+ a_n(f^{ \theta})^* \right)
\end{align}
and
\begin{align}
H^{(n), \theta} : = H_0^{(n), \theta} + g V^{(n), \theta} .
\end{align}
\end{enumerate}
\end{definition}
The Hamiltonians $H^{(n),\theta}$ turn out to have gaps between the eigenvalues $\lambda^{(n)}_i$ and the rest of the spectrum of $H^{(n),\theta}$. This allows us to define Riesz projections, $ P_i^{(n)} $, corresponding to the eigenvalues $ \lambda_i^{(n)} $ and use regular perturbation theory for each $n\in\mathbb{N}$. In an inductive scheme, one can obtain explicit estimates on the resolvents and the eigenvalues in each step. Below, we prove that the sequences $ ( \lambda^{(n)}_i )_{n \in \mathbb{N}} $ converge and the interacting ground-state energy $\lambda_0$ and resonance energy $\lambda_1$ of $H^\theta$ are the limits
\begin{align}
\lambda_i:= \lim\limits_{n\to\infty}\lambda^{(n)}_i , \qquad i=0,1.
\end{align}
We define
\begin{align} \label{pet1tt}
\mathfrak{h}^{(n,\infty)}:= L^2(\mathcal B_{\rho_n} ) .
\end{align}
We denote the corresponding Fock space by $ \mathcal F[\mathfrak{h}^{(n, \infty)}] $ (it is defined as in \eqref{Fock}), with vacuum state $\Omega^{(n, \infty)}$. It is straightforward to verify that $ \mathcal H $ is isomorphic to $ \mathcal H^{(n)}\otimes \mathcal F[\mathfrak{h}^{(n,\infty)}] $ and, therefore, we identify
\begin{align} \label{pet2tt}
\mathcal H \equiv \mathcal H^{(n)}\otimes \mathcal F[\mathfrak{h}^{(n,\infty )}].
\end{align}
We prove below that the sequence $ ( P^{(n)}_i \otimes P_{ \Omega^{(n, \infty)} } )_{n \in \mathbb{N}} $, where $ P_{ \Omega^{(n, \infty)} } $ is the orthogonal projection in the vector space generated by $ \Omega^{(n, \infty)} $, converges to an eigenprojection corresponding to the eigenvalue $\lambda_i$.
\section{Main results}
\label{sec:mainresult}
Here, we state the main results of our work. All proofs are presented in the next sections. In Proposition \ref{thm:res} below we prove existence of the ground-state eigenvalue and the resonance of $ H^{\theta} $. A similar result, for a more complicated model (Pauli-Fierz), is proved in \cite{bach}. The strategy of proof of Proposition \ref{thm:res} is based on the methods introduced in \cite{bach} but it differs from the proof therein because, here, all our estimates must be independent of $\theta \in \mathcal S$.
As emphasized earlier, the existence of the resonance and the ground-state is not our focus but is only provided in order for this work to be self-contained.
The next proposition is proved in Section \ref{sec:proof-res}.
\begin{proposition}[Construction of the ground-state and the resonance]
\label{thm:res}
For every $\rho, \rho_0 $ sufficiently small (see Definition \ref{sequence}) there is a constant $ g_0 > 0 $ (that depends on $\rho, \rho_0$ and $ \boldsymbol{ \nu } $) such that, for every $\theta\in\mathcal S$ (see \eqref{def:setS}) and every $ g \in D(0, g_0) $, the (complex) number
\begin{align}
\lambda_i :=\lim\limits_{n\to\infty}\lambda^{(n)}_i , \quad i=0,1
\end{align}
is an eigenvalue of $H^\theta$ and the range of
\begin{align}
\label{limP}
P_i :=\lim\limits_{n\to\infty}P^{(n)}_i\otimes P_{ \Omega^{(n, \infty)} } , \quad i=0,1
\end{align}
consists of eigenvectors corresponding to $\lambda_i$. An explicit formula for $ g_0 $ is presented in Definition \ref{gzero} below.
\end{proposition}
The non degeneracy of the eigenvalues in Proposition \ref{thm:res} as well as estimates for the imaginary part of the resonance can be derived from the corresponding results for the Pauli-Fierz model in \cite{bach} and \cite{bbp}. Since their proofs do not need the new features of our multiscale scheme and they are not relevant for our main results, we only state them without proofs and refer to \cite{bach}.
\begin{remark}[Fermi golden rule]
\label{prop:fermi}
The eigenvalues $\lambda_0$ and $\lambda_1$ are non-degenerate, this follows from Section 6.4.3 in \cite{bach} (we do not repeat the proof here).
The leading order of the imaginary part of the resonance $\lambda_1$ can be explicitly calculated. This is presented in Theorem 5.6 in \cite{bach} for the Pauli-Fierz model and, using a different method, in \cite{bfs100}. We do not include a proof here because it follows, for the model at hand, without much change from the proof in \cite{bach}.
We assume that $ |g| > 0$ is small enough and define
\begin{align}
E_I:=-4\pi^2 (e_1-e_0)^2 |f(e_1-e_0)|^2 .
\end{align}
Then, there is a constant $C_\eqref{eq:impartres1}>0$ and a constant $\epsilon>0$ such that for all $n\in \mathbb{N} $ large enough
\begin{align}
\label{eq:impartres1}
\left| \Im \lambda_1^{(n)} -g^2 E_I \right| \leq g^{2+\epsilon} C_\eqref{eq:impartres1}.
\end{align}
\end{remark}
The next theorems are our main results. We prove analyticity of the resonance and the ground-state, and the corresponding eigen-projections, with respect to the dilation parameter and coupling constant.
The next theorem is proved in Section \ref{analyticity} (see Theorem \ref{thm:anaPp}).
\begin{theorem}[Analyticity with respect to the dilation parameter]
\label{thm:anaP}
For $\rho, \rho_0 $ sufficiently small and $g \in D(0, g_0)$ (see Proposition \ref{thm:res}), the functions
\begin{align}
\mathcal S \ni \theta \mapsto P_i, \hspace{1cm} \mathcal S \ni \theta \mapsto \lambda_i
\end{align}
are analytic. Moreover, this implies that $ \lambda_i (\theta) \equiv \lambda_i$ is constant for $\theta\in\mathcal S$ (see \eqref{def:setS}).
\end{theorem}
\begin{remark}
\label{rem:lambda0equal0}
Our bounds in the inductive scheme (see Theorem \ref{thm:ind} below) which are used to prove Theorem \ref{thm:anaP} blow up as we take $\nu\to 0$. We study simultaneously the cases $i = 0$ and $i = 1$ and, therefore, our estimations blow up also for $i=0$. However, it is easy to see from our method that for $ i= 0 $ alone we can take $ \theta $ in a neighborhood of $0$ and prove analyticity in this neighborhood. This implies that
$ \lambda_0 $ is real, because $ H^{\theta} $ is selfadjoint for $\theta = 0$. It is the ground-state energy constructed in \cite{bbkm,spohnspin}.
\end{remark}
Next theorem is proved in Section \ref{analyticity} (Theorem \ref{thm:anaPp}).
\begin{theorem}[Analyticity with respect to the coupling constant]
\label{thm:anaPg}
For every $\rho, \rho_0 $ sufficiently small and $g \in D(0, g_0)$,
the functions
\begin{align}
g \mapsto P_i, \hspace{1cm} g \mapsto \lambda_i
\end{align}
are analytic.
\end{theorem}
Our next two theorems provide an estimate for the spectrum of $H^{\theta}$ in neighborhoods of $\lambda_0$ and $\lambda_1$, and resolvent estimates in these neighborhoods. As discussed in the introduction, similar results on spectral estimates can be found in \cite{bfs1,bfs2,bfs3,bfs100} in which the spectrum is located in cuspidal domains using the spectral renormalization method based on the Feshbach-Schur map method.
Here, we localize the spectrum in cones. For every $ z\in\mathbb{C} $, we define
\begin{align}
\label{eq:defcone}
\mathcal C_m(z) :=\left\{ z+xe^{-i\alpha} : x\geq 0 , |\alpha-\nu |\leq \nu/m \right\},
\end{align}
where we assume that $m \geq 4$, allover this work.
The next theorem is proved in the proofs of Theorems \ref{resuno} and \ref{resdos} below.
\begin{theorem}[Resolvent estimates]\label{resolventestimates}
There is a constant $ \bold C $ (see Definition \ref{C} and \eqref{ggg}) that depends on $ \boldsymbol{\nu} $ but not on $g$ nor in $\rho $ and $ \rho_0$ such that for every $ m \geq 4$ and $\rho, \rho_0 $ sufficiently small, there exists $g^{(m)} > 0$ with the following properties:
for every $ \theta \in \mathcal{S} $ and $g \in D(0, g_0)$ (see Proposition \ref{thm:res}) with $|g | \leq g^{(m)}$ ,
\begin{align}
\norm{\frac{1}{ H^{\theta}-z}} \leq 16
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z, \mathcal{C}_m(\lambda_i ) )} ,
\end{align}
for every $ z \in B_i^{(1)} \setminus \mathcal C_m \left(\lambda_i -
2 \rho_n^{1+ \mu/4} e^{-i\nu}\right) $ and
\begin{align}
\norm{\frac{1}{ H^{\theta}-z}} \leq 8
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z, \mathcal{C}_m(\lambda_i^{(n)} ) )} ,
\end{align}
for every $ z \in B_i^{(1)} \setminus \mathcal C_m \left(\lambda_i^{(n)}-
\rho_n^{1+ \mu/4} e^{-i\nu}\right)$. Here, the symbol ${\rm dist}$ denotes the the Euclidean distance in $\mathbb{C}$.
\end{theorem}
Explicit bounds for $\boldsymbol{C}$, $\rho_0 $ and $\rho$, $g_0$ and $ g^{(m)} $ are given in Definitions \ref{C}, \ref{sequence}, \ref{gzero} and Eq.\ \eqref{ggg}, respectively. We remark that we intentionally do not provide optimal estimates because these would render the proof unnecessary opaque.
The next theorem is proved in the proof of Theorem \ref{resdos} below.
\begin{theorem}[Spectral estimates]\label{spectralestimates} For every $\rho, \rho_0 $ sufficiently small, $ \theta \in \mathcal{S} $ and $g \in D(0, g_0)$ with $ |g| \leq g^{(m)} $, there is a neighborhood $
B_i^{(1)} $ of $ \lambda_i $ (that depends on $\nu $ but not on $g$) such that the spectrum of $H^{\theta}$ in $ B_i^{(1)} $ is contained in $ \mathcal C(\lambda_i) $ (recall that $\nu$ is the imaginary part of $\theta$). An explicit
formula for $ B_i^{(1)} $ is given in \eqref{region:Bi1}.
\end{theorem}
\section{Resolvent estimates far away from the spectrum and detailed analysis of $ H^{(1), \theta} $ } \label{RES}
In this subsection we derive resolvent estimates for $ H^{(n), \theta} $ and $ H^{ \theta} $ for complex numbers $z$ that are far away from their respective spectra. For the first Hamiltonian, $ H^{(1), \theta} $, having an infrared cutoff, we present resolvent estimates for points that are close to its spectrum. Here, we do not need any restrictions on the sequence $\{\rho_n\}_{n \in \mathbb{N}}$ other than $ 0 < \rho_0 < \min(1, e_1/4) $, $ 0 < \rho < 1 $.
In the forthcoming sections (see Definition \ref{sequence}) we need to assume other properties for the sequence $\{\rho_n\}_{n \in \mathbb{N}}$.
We emphasize that the particular choice of numbers $\rho_n$ does not imply any physical constraint, it only specifies the rate at which the infrared cut-off is removed.
In this section and (in the whole paper) we denote by $ c > 0 $ any generic (indeterminate) constant (it can change from line to line) that is independent of
the parameters $n$, $\theta$, $\rho_0$, $\rho$, $\theta$, $\boldsymbol \nu$ and $g$. It might depend on the set $\mathcal S$, as a whole, but not on its elements $ \theta \in \mathcal S$ and nor either on the parameter $\boldsymbol \nu$.
Moreover, by stating that $|g|$ is small enough, we mean that there is a constant such that uniformly for $|g|$ smaller than this constant the referred statement holds true. We employ that such a constant does not depend on $\theta$ and $n$ but it depends on the set $\mathcal S$ and on the remaining parameters.
\subsection{Resolvent estimates far away from the spectrum }
We define regions in the complex plane in which we derive resolvent estimates.
\begin{definition}
\label{def:regionsAB}
We set $ \delta: = e_1 - e_0 = e_1 $ and define the region
\begin{align}
\label{region:A}
A:&=
A_1\cup A_2\cup A_3 ,
\end{align}
where
\begin{align}
A_1:&=\left\{ z\in\mathbb{C} : \Re z <e_0-\frac{1}{2}\delta \right\}
\\
A_2:&= \left\{ z\in\mathbb{C} : \Im z >\frac{1}{8}\delta \sin (\nu) \right\}
\\
A_3:&= \left\{ z\in\mathbb{C} : \Re z >e_1+\frac{1}{2}\delta , \Im z \geq -\sin \Big (\frac{\nu}{2}\Big ) \left(\Re (z) -(e_1+\frac{1}{2}\delta) \right)\right\} ,
\end{align}
and for $i\in \{0,1\}$,
\begin{align}
\label{region:Bi1}
B_i^{(1)}:=\left\{ z\in\mathbb{C} : |\Re z-e_i| \leq \frac{1}{2}\delta, -\frac{1}{2}\rho_1 \sin(\nu)\leq \Im z \leq \frac{1}{8}\delta \sin (\nu) \right\} .
\end{align}
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{RegionsAB.eps}
\caption{Subsets of the complex plane (see Definition \ref{def:regionsAB}) }
\label{fig:regionsAB}
\end{figure}
\end{definition}
In this subsection, we estimate the resolvent of $ H^{(n), \theta} $ and $H^{\theta}$ far away from their spectra, namely in the region $A$ defined in \eqref{region:A}. This estimates are applied for the induction basis in our inductive scheme described in Section \ref{sec:indscheme}.
\begin{lemma}
\label{lemma:resestinA}
Let $\theta\in\mathcal S$ (see \eqref{def:setS}) and $n\in\mathbb{N}$. There is a constant $ C_{\eqref{const:resestinA} } $ (independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $) such that
for small enough $|g|$ (depending on $ \boldsymbol \nu $),
for every $i\in\{0,1\}$:
\begin{align}
\label{const:resestinA}
\norm{\frac{1}{H^{(n),\theta}-z}} \leq \frac{ C_{\eqref{const:resestinA} } }{ \sin(\boldsymbol \nu/2 ) }
\frac{1}{ |e_i-z|}, \hspace{.5cm} \norm{\frac{1}{H^{\theta}-z}} \leq
\frac{ C_{\eqref{const:resestinA} } }{ \sin(\boldsymbol \nu/2 ) } \frac{1}{|e_i-z|}, \qquad \forall z\in A .
\end{align}
\end{lemma}
\begin{proof}
Let $z\in A$ and $n\in \mathbb{N} $. Then, arguing as in Appendix \ref{app:sa} and using functional calculus, we obtain that
\begin{align}
\label{eq:prA0}
\norm{V^{(n),\theta}\frac{1}{H_0^{(n),\theta}-z}}
&\leq \norm{V^{(n),\theta}\frac{1}{(H_0^{(n)}+1)^\frac{1}{2}}} \norm{\frac{H_0^{(n)}+1}{H_0^{(n),\theta}-z}}
\notag \\
&\leq \left( \norm{f^\theta}_2 +2\norm{f^\theta/\sqrt{\omega}}_2 \right) \sup_{y\in [0,\infty),i=0,1}\left| \frac{e_i + y+1}{e_i +e^{-\theta}y-z} \right|
.
\end{align}
Geometrical considerations imply that there is constant $c>0$ such that
\begin{align}
\text{dist}\left(\{e_i+e^{-\theta} y : \, i=0,1 \},A \right)\geq \frac{c}{\sin(\boldsymbol \nu/2)}(1+y) \qquad \forall y\geq 0 ,
\end{align}
and hence, there there is constant $c>0$ such that
\begin{align}
\label{eq:prA000}
\norm{V^{(n),\theta}\frac{1}{H_0^{(n),\theta}-z}}
&\leq \frac{c}{\sin(\boldsymbol \nu/2)}, \qquad \forall z\in A
.
\end{align}
Then, we choose $|g|$ small enough such that
\begin{align}
\norm{ g V^{(n),\theta}\frac{1}{H_0^{(n),\theta}-z}}
\leq \frac{1}{2}
\end{align}
and hence,
\begin{align}\label{nom}
H^{(n),\theta}-z=\left(1+gV^{(n),\theta}\frac{1}{H_0^{(n),\theta}-z} \right) \left( H_0^{(n),\theta}-z \right)
\end{align}
is invertible for all $z\in A$, since $A\cap \sigma (H_0^{(n),\theta})=\emptyset$. Thanks to the particular geometry, there is a constant $ c > 0 $ such that $|e_{j}+ye^{-\theta}-z|\geq c \sin ( \boldsymbol \nu/2 ) |e_i-z|$, for every $z\in A$, every $j \in \{0, 1 \}$, and every positive number $y$. This and \eqref{nom} imply
\begin{align}
\label{eq:resestA1}
\norm{\frac{1}{H^{(n),\theta}-z}} \leq 2\norm{\frac{1}{H_0^{(n),\theta}-z}} = \sup_{i=0,1}\sup_{y\geq \rho_n}\frac{2}{|e_i+ye^{-\theta}-z|} \leq \frac{c}{ \sin(\boldsymbol \nu/2) |e_i-z|}
\end{align}
for all $z\in A$, $i=0,1$ and some constant $c>0$. This completes the proof for the first equation in \eqref{const:resestinA}. Since the second equation can be shown in a very similar fashion we omit the proof here.
\end{proof}
\subsection{Analysis of $ H^{(1), \theta} $ }
\label{sec:ib}
\begin{lemma}
\label{lemma:ib-reswithoutdisc}
Let $\theta\in\mathcal S$ (see \eqref{def:setS}) and $|g|$ small enough (depending on $\boldsymbol{\nu}$ and $\rho_1$). Then,
\begin{align}
\norm{\frac{1}{H^{(1),\theta}-z}} \leq 2\norm{\frac{1}{H_0^{(1),\theta}-z}} \qquad \forall z\in E^{(1)}_i , i=0,1 ,
\end{align}
where
\begin{align}\label{E1i}
E^{(1)}_i:=B^{(1)}_i\setminus D\left(e_i, \frac{\rho_1
\sin (\nu) }{8}\right) .
\end{align}
\end{lemma}
\begin{proof}
Let $z\in E^{(1)}_i$ and $i=0,1$. Then, we have, arguing as in Appendix \ref{app:sa},
\begin{align} \label{pin1}
&\norm{V^{(1),\theta}\frac{1}{H_0^{(1),\theta}-z}} \leq \norm{V^{(1),\theta}\frac{1}{(H_0^{(1)}+1)^\frac{1}{2}}} \norm{\frac{H_0^{(1)}+1}{H_0^{(1),\theta}-z}}
\notag \\
&\leq \left( \norm{f^\theta}_2 +2\norm{f^\theta/\sqrt{\omega}}_2 \right) \sup_{y\in \{0\} \cup [\rho_1,\infty),i=0,1}\left| \frac{e_i+y+1}{e_i+e^{-\theta}y-z} \right| .
\end{align}
Take $y\in \{0\} \cup [\rho_1,\infty)$ and $ i \in \{ 0,1 \}$. It follows that
\begin{align}\label{pin2}
\left| \frac{e_i+y+1}{e_i+e^{-\theta}y-z} \right|
= & \left| \frac{e_i+ e^{\theta}( e_i+e^{-\theta}y-z )+1 - e^{\theta}( e_i -z ) }{e_i+e^{-\theta}y-z} \right|
\\ \notag
\leq & |e^{\theta}| + c \frac{1}{ | e_i+e^{-\theta}y-z |} \leq\frac{c}{\rho_1 \sin( \nu)},
\end{align}
where the last inequality is due to the considered geometry.
From \eqref{pin1} and \eqref{pin2} we obtain that there is a finite constant $c>0$ such that
\begin{align}
\label{eq:lemma-reswithoutdisc-standtardest1}
\norm{V^{(1),\theta}\frac{1}{H_0^{(1),\theta}-z}} \leq \frac{c}{\rho_1 \sin( \nu)}
\leq \frac{c}{\rho_1 \sin(\boldsymbol \nu)} .
\end{align}
For $|g|$ small enough (depending on $ \rho_1 $ and $\boldsymbol{\nu}$), we arrive at
\begin{align}
\label{eq:lemma-reswithoutdisc-standtardest}
\norm{gV^{(1),\theta}\frac{1}{H_0^{(1),\theta}-z}}
&\leq \frac{1}{2} ,
\end{align}
and hence,
\begin{align}
H^{(1),\theta}-z=\left(1+gV^{(1),\theta}\frac{1}{H_0^{(1),\theta}-z} \right) \left( H_0^{(1),\theta}-z \right)
\end{align}
is invertible for all $z\in E^{(1)}_i$, since $E^{(1)}_i\cap \sigma (H_0^{(1),\theta})=\emptyset$. Then, we obtain
\begin{align}
\norm{\frac{1}{H^{(1),\theta}-z}} \leq 2\norm{\frac{1}{H_0^{(1),\theta}-z}},
\end{align}
which completes the proof.
\end{proof}
\begin{definition}
\label{def_ib-proj}
We define the projections
\begin{align}
\label{eq:ib-proj}
P^{(1)}_i: =-\frac{1}{2\pi i}\int_{\hat \gamma^{(1)}_i}\mathrm{d}z \, \frac{1}{H^{(1),\theta}-z}
\end{align}
and
\begin{align}
\label{eq:ib-projat}
P^{(1)}_{\text{at},i}: =-\frac{1}{2\pi i}\int_{\hat \gamma^{(1)}_i}\mathrm{d}z \, \frac{1}{H_0^{(1),\theta}-z}=P_{e_i}\otimes P_{\Omega^{(1)}}
\end{align}
where
\begin{align}
\label{eq:ib-gamma}
\hat \gamma^{(1)}_i: [0,2\pi ] \to \mathbb{C} , \quad
t \mapsto e_i +\frac{1}{4}\rho_1 \sin (\nu) e^{it},
\end{align}
$P_{e_i}$ is the projection onto the eigenspace space corresponding to $e_i$ of the Hamiltonian $K$ and $P_{\Omega^{(1)}}$ is the projection onto the vector space generated by the vacuum, $\Omega^{(1)} $, of $\mathcal F[\mathfrak{h}^{(1)}]$.
\end{definition}
\begin{remark}\label{remp}
The right-hand side of Eq.\ \eqref{eq:ib-projat} follows from the following computation:
\begin{align}
&-\frac{1}{2\pi i}\int_{\hat \gamma^{(1)}_i}\mathrm{d}z \, \frac{1}{H_0^{(1),\theta}-z}
=
-\frac{1}{2\pi i}\int_{\hat \gamma^{(1)}_i}\mathrm{d}z \, \frac{1}{H_0^{(1),\theta}-z} (P_{e_i}\otimes P_{\Omega^{(1)}} +\overline{P_{e_i}\otimes
P_{\Omega^{(1)}}
})
\notag \\
&=P_{e_i}\otimes P_{\Omega^{(1)}} -\frac{1}{2\pi i}\int_{\hat \gamma^{(1)}_i}\mathrm{d}z \, \frac{1}{H_0^{(1),\theta}-z}\overline{P_{e_i}\otimes
P_{\Omega^{(1)}} } ,
\end{align}
where
\begin{align}
\overline{P_{e_i}\otimes P_{\Omega^{(1)}} }=\overline{P_{e_i}}\otimes 1+P_{e_i}\otimes\overline{ P_{\Omega^{(1)}} }
\end{align}
implies that $ -\frac{1}{2\pi i}\int_{\hat \gamma^{(1)}_i}\mathrm{d}z \, \frac{1}{H_0^{(1),\theta}-z}\overline{P_{e_i}\otimes P_{\Omega^{(1)}} }=0$.
\end{remark}
\begin{lemma}
\label{lemma:ib-proj}
Let $\theta\in\mathcal S$ (see \eqref{def:setS}) and let $|g|$ be small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $). Take $i\in\{ 0,1 \}$. Then, there is a constant $C_{\eqref{eq:pconst}}>0$
(independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $) such that
\begin{align}
\label{eq:pconst}
\norm{ P_i^{(1)}-P^{(1)}_{\text{at},i}} \leq | g| \frac{ C_{\eqref{eq:pconst}} }{\rho_1 \sin(\boldsymbol \nu)} <1 \qquad \text{and} \qquad \norm{ P_i^{(1)}}\leq 1+ |g| \frac{ C_{\eqref{eq:pconst}} }{\rho_1 \sin(\boldsymbol \nu)} <2 ,
\end{align}
where $\hat \gamma^{(1)}_i$, $ P_i^{(1)}$ and $P^{(1)}_{\text{at},i}$ are introduced in Definition \ref{def_ib-proj}.
\end{lemma}
\begin{proof}
First, we observe that
\begin{align}
\norm{ P_i^{(1)}-P^{(1)}_{\text{at},i}}
&\leq \frac{1}{2\pi }\norm{ \int_{\hat \gamma^{(1)}_i}\mathrm{d}z \, \left( \frac{1}{H^{(1),\theta}-z} -\frac{1}{H_0^{(1),\theta}-z} \right)} .
\end{align}
Note that $\hat \gamma^{(1)}_i\subset E^{(1)}_i$ (see \eqref{E1i}). Eq.\ \eqref{eq:lemma-reswithoutdisc-standtardest1} implies that there is a finite constant $c>0$ such that for every $ z$ in the (image of the) curve $ \hat \gamma^{(1)}_i $
\begin{align}
\norm{gV^{(1),\theta}\frac{1}{H_0^{(1),\theta}-z}}
&< |g| \frac{c}{\rho_1 \sin(\boldsymbol \nu)} \leq \frac{1}{2} ,
\end{align}
for $|g|$ small enough (depending on $ \boldsymbol \nu $ and $\rho_1$). Next, we obtain
\begin{align}
& \norm{ P_i^{(1)}-P^{(1)}_{\text{at},i}}
\leq \frac{1}{2\pi }\norm{ \int_{\hat \gamma^{(1)}_i}\mathrm{d}z \,\frac{1}{H_0^{(1),\theta}-z} \sum^\infty_{l=1} \left( -gV^{(1),\theta}\frac{1}{H_0^{(1),\theta}-z} \right)^l }
\notag \\
&\leq \frac{\rho_1 \sin (\nu)}{4} \sup_{ |z - e_i | = \frac{\rho_1 \sin (\nu)}{4} }
\Big ( \Big \| \frac{1}{H_0^{(1),\theta}-z} \Big \| \Big )
| g| \frac{c}{\rho_1 \sin(\boldsymbol \nu)}
\sum^\infty_{l=0} \Big(\frac{1}{2}\Big ) ^l
< |g| \frac{c}{\rho_1 \sin(\boldsymbol \nu)} .
\end{align}
This proves the first part of the lemma. Furthermore, it follows from \eqref{eq:ib-projat} that $\norm{P^{(1)}_{\text{at},i}}=1$,
and hence,
\begin{align}
\norm{ P_i^{(1)}} \leq \norm{P^{(1)}_{\text{at},i}} +\norm{\hat P_i^{(1)}-P^{(1)}_{\text{at},i}} \leq 1+ |g| \frac{c}{\rho_1 \sin(\boldsymbol \nu)} <2,
\end{align}
for sufficiently small $|g|$.
This proves proves the second part of the lemma.
\end{proof}
\begin{remark} \label{missing}
Lemmas \ref{lemma:ib-reswithoutdisc} and \ref{lemma:ib-proj}, together with the fact that $ P^{(1)}_{\text{at},i}$ is a rank-one projection imply the following:
Let $\theta\in\mathcal S$ (see \eqref{def:setS}) and $i\in\{ 0,1 \}$.
Suppose that $|g|$ is small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $). Then, there is a unique spectral point
$ \lambda^{(1)}_i $ of $ H^{(1), \theta} $ in $B_i^{(1)}$. Moreover $ \lambda^{(1)}_i $ is a simple eigenvalue and it is contained in $ D\left(e_i, \frac{\rho_1
\sin (\nu) }{8}\right) $.
\end{remark}
Lemma \ref{lemma:ib-reswithoutdisc} together with Lemma \ref{lemma:ib-proj} yield a resolvent estimate in the whole region $B_i^{(1)}\setminus \{\lambda_i^{(1)}\}$ by making use of the maximum modulus principle of complex analysis.
\begin{lemma}
\label{lemma:ib-res}
Let $\theta\in\mathcal S$ (see \eqref{def:setS})
let $|g|$ be small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $). Take $i\in\{ 0,1 \}$. Then, there is a constant $C_{\eqref{eq.ib-res}}>0$
(independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $)
such that
\begin{align}
\label{eq.ib-res}
\norm{\frac{1}{H^{(1),\theta}-z}\overline{ P_i^{(1)}}}
\leq \frac{ C_{\eqref{eq.ib-res}} }{\rho_1 \sin ( \nu) }
\leq \frac{ C_{\eqref{eq.ib-res}} }{\rho_1 \sin (\boldsymbol \nu) },
\qquad \forall z\in B_i^{(1)}.
\end{align}
\end{lemma}
\begin{proof}
Note that the function
\begin{align}
\overline{D\left(e_i, \frac{1}{8}\rho_1 \sin (\nu) \right)} \ni z \mapsto G_{\phi,\psi}(z):=\left\langle \phi, \frac{1}{H^{(1),\theta}-z}\overline{ P_i^{(1)}} \psi \right\rangle
\end{align}
is continuous, and furthermore, analytic on $D\left(e_i, \frac{1}{8}\rho_1 \sin (\nu) \right)$, for all $\phi, \psi \in \mathcal H$. Then, it follows from the maximum modulus principle that this function attains its maximum on the boundary of its domain. This together with the Cauchy-Schwarz inequality and Lemma \ref{lemma:ib-reswithoutdisc} and \ref{lemma:ib-proj} implies that there is a finite constant $c>0$ such that
\begin{align}
\label{eq:lemma-ib-res-maxmod}
\left| G_{\phi,\psi}(z) \right|\leq \frac{c}{\rho_1 \sin ( \nu) } \| \phi \| \psi \| , \qquad \forall z\in\overline{D\left(e_i, \frac{1}{8}\rho_1 \sin (\nu) \right)}.
\end{align}
Consequently, there is a finite constant $c>0$ such that
\begin{align}
\label{eq:lemma-ib-res-est1}
\norm{\frac{1}{H^{(1),\theta}-z}\overline{ P_i^{(1)}}}
&\leq \frac{c}{\rho_1 \sin ( \nu) }, \qquad \forall z\in\overline{D\left(e_i, \frac{1}{8}\rho_1 \sin (\nu) \right)} ,
\end{align}
and moreover, Lemma \ref{lemma:ib-reswithoutdisc} (again) guarantees that there is a finite constant $c>0$ such that
\begin{align}
\label{eq:lemma-ib-res-est2}
\norm{\frac{1}{H^{(1),\theta}-z}} &\leq 2\norm{\frac{1}{H_0^{(1),\theta}-z}}
\leq \frac{c}{\rho_1 \sin ( \nu) },
\end{align}
for all $z\in B^{(1)}_i \setminus D\left(e_i, \frac{1}{8}\rho_1 \sin (\nu) \right) $.
This together with \eqref{eq:lemma-ib-res-est1} completes the proof.
\end{proof}
Applying Lemma \ref{lemma:ib-res} to our particular geometry allows to formulate the following corollary.
\begin{corollary}
We define $ v^{(1)}_i := \lambda^{(1)}_i +
\frac{1}{4} \rho_1 e^{-i \nu} $ and recall Eq.\ \eqref{eq:defcone}.
\label{coro:ib-res}
Let $\theta\in\mathcal S$ ( see \eqref{def:setS})
let $|g|$ be small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $). Take $i\in\{ 0,1 \}$. Then, there is a constant $C_{\eqref{coroeq.ib-res}}>0$
(independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $)
such that
\begin{align}
\label{coroeq.ib-res}
\norm{\frac{1}{H^{(1),\theta}-z}\overline{ P_i^{(1)}}}
\leq \frac{1}{ \sin(\nu/m) }\frac{ C_{\eqref{eq.ib-res}} }{{ \rm dist }( z, \mathcal{C}_m(v_1^{(1)}) ) }, \qquad \forall z \in B^{(1)}_i \setminus \mathcal{C}_m(v_i^{(1)}) .
\end{align}
\end{corollary}
\begin{lemma}
\label{lemma:ib-dist-energy}
Let $\theta\in\mathcal S$ (see \eqref{def:setS}) and let $|g|$ be small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $). Then, there is a constant $C_{\eqref{constgggg}}>0$ (independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $) such that for every $i \in \{0, 1 \}$
\begin{align}
\label{constgggg}
\left| \lambda_i^{(1)} -e_i \right| \leq |g| C_{\eqref{constgggg}} .
\end{align}
\end{lemma}
\begin{proof}
It follows from Lemma \ref{lemma:ib-proj} that $\Big |\left\langle \varphi_i \otimes \Omega, P^{(1)}_i \varphi_i \otimes \Omega \right\rangle \Big |>\frac{1}{2}$ for $|g|$ small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $).
We calculate
\begin{align}
\lambda_i^{(1)}
&= \frac{\left\langle \varphi_i \otimes \Omega^{(1)}, H^{(1),\theta} P^{(1)}_i \varphi_i \otimes \Omega^{(1)} \right\rangle}{\left\langle \varphi_i \otimes \Omega^{(1)} , P^{(1)}_i \varphi_i \otimes \Omega^{(1)} \right\rangle}
=e_i+ g\frac{\left\langle V^{(1),\overline{\theta}} \varphi_i \otimes \Omega^{(1)} , P^{(1)}_i \varphi_i \otimes \Omega^{(1)} \right\rangle}{\left\langle \varphi_i \otimes \Omega^{(1)} , P^{(1)}_i \varphi_i \otimes \Omega^{(1)} \right\rangle}.
\end{align}
Let now $z\in\mathbb{C}$ such that $|e_i-z|=\frac{1}{4}\rho_1 \sin (\nu) $. Eq.\ \eqref{eq:pconst} (which requires $|g|$ to be small enough -depending on $ \boldsymbol{\nu} $ and $ \rho_1 $) then allows to obtain
\begin{align}
&\left| \lambda_i^{(1)} -e_i \right| \leq 4 \norm{gV^{(1),\overline{\theta}} \varphi_i \otimes \Omega^{(1)} }\leq
4 \left| e_i-z \right| \norm {gV^{(1),\overline{\theta}}\frac{1}{H^{(1),\overline \theta}_0-z}}
\leq |g| c,
\end{align}
for some constant $c$ (independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $) .
Here, we have used \eqref{eq:lemma-reswithoutdisc-standtardest1} from Lemma \ref{lemma:ib-reswithoutdisc} in the last step. Notice that in this work we assume that the imaginary part of $\theta$, $ \nu $, is positive. Then, strictly speaking, we do not have the right to use our results for $\overline \theta$. However, the restriction we impose by assuming that $ \nu $ is not negative is irrelevant. This is assumed only for convenience in order to simplify our notation. Of course, the same results hold true if we take $ - \pi/16 < \nu < - \boldsymbol{\nu} $.
\end{proof}
\begin{remark}
\label{lemma:ib-proj2}
Let $\theta\in\mathcal S$ (see \eqref{def:setS}) and let $|g|$ be small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $). Then
\begin{align}
P_i^{(1)}=-\frac{1}{2\pi i}\int_{ \gamma^{(1)}_i}\mathrm{d}z \, \frac{1}{H^{(1),\theta}-z},
\end{align}
where $\gamma^{(1)}_i: [0,2\pi ] \to \mathbb{C} , \quad
t \mapsto \lambda^{(n)}_i +\frac{1}{4}\rho_1 \sin (\nu) e^{it}$. This follows from Remark \ref{missing}, because, for small enough $|g|$, $\gamma^{(1)}_i ,\hat \gamma^{(1)}_i \subset B_i^{(1)}\setminus \{\lambda_i^{(1)}\} $, see Lemma
\ref{lemma:ib-dist-energy}.
\end{remark}
\begin{lemma}
\label{lemma:ib-resprima}
Let $\theta\in\mathcal S$ (see \eqref{def:setS})
and let $|g|$ be small enough (depending on $ \boldsymbol{\nu} $ and $ \rho_1 $). Take $i\in\{ 0,1 \}$. Then, there is a constant $C_{\eqref{eq.ib-resprima}}>0$
(independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $)
such that
\begin{align}
\label{eq.ib-resprima}
\norm{\frac{1}{H^{(1),\theta}-z}\overline{ P_i^{(1)}}}
\leq
\frac{C_{\eqref{eq.ib-resprima}}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_1+\left| \lambda^{(1)}_i-z \right|},
\qquad \forall z\in B_i^{(1)}.
\end{align}
\end{lemma}
\begin{proof}
We use Lemma \ref{lemma:ib-res} and calculate, for $ | z - e_i | \leq \rho_1 $,
\begin{align}
\frac{ 1 }{\rho_1 \sin ( \nu) } = \frac{\rho_1+|\lambda_i^{(1)}-z|}{ \rho_1 \sin ( \nu) } \frac{1}{\rho_1+|\lambda_i^{(1)}-z|} \leq
\frac{ c }{ \sin (\boldsymbol \nu) } \frac{1}{\rho_1+|\lambda_i^{(1)}-z|},
\end{align}
where we use Lemma \ref{lemma:ib-dist-energy} and choose $|g|$ small enough. For $ | z - e_i | > \rho_1 $ we use Lemma
\ref{lemma:ib-reswithoutdisc}. The spectral theorem and the explicit
form of the spectrum of non-interacting Hamiltonian $ H_0^{(1),\theta} $ allow us to estimate the norm of its resolvent. Then, similar estimates as above imply the desired result.
\end{proof}
\begin{lemma}
\label{lemma:resestinAprima}
Let $\theta\in\mathcal S$ (see \eqref{def:setS}) and $n\in\mathbb{N}$. There is a constant $ C_{\eqref{const:resestinAprima} } $ (independent of $\theta $, $n$,
$g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $) such that
for small enough $|g|$ (depending on $ \boldsymbol \nu $),
for every $i\in\{0,1\}$:
\begin{align}
\label{const:resestinAprima}
\norm{\frac{1}{H^{(n),\theta}-z}} \leq \frac{C_{\eqref{const:resestinAprima}}}{\sin (\boldsymbol \nu/2 )} \frac{1}{\rho_l+\left| \widetilde{\lambda}_i-z \right|}, \hspace{.5cm} \norm{\frac{1}{H^{\theta}-z}} \leq\frac{C_{\eqref{const:resestinAprima}}}{\sin (\boldsymbol \nu/2)} \frac{1}{ \rho_l+\left| \widetilde{\lambda}_i-z \right|} ,
\end{align}
for every $ z\in A \cup \Big ( B^{(1)}_1 - [0, \infty) e^{-i \nu} \Big) \setminus B^{(1)}_1 , $ every $l \in \mathbb{N}$ and every $ \widetilde{\lambda}_i \in D( \lambda^{(1)}_i, 3 g ) .$
Moreover,
\begin{align}
\label{coroeq.ib-resT}
\norm{\frac{1}{H^{(n),\theta}-z}\overline{ P_i^{(1)}}}
\leq \frac{1}{ \sin(\nu/m) }\frac{ C_{\eqref{coroeq.ib-resT}} }{{ \rm dist }( z, \mathcal{C}_m(v_1^{(n)}) ) },
\end{align}
for every $ z\in \Big ( B^{(1)}_1 - [0, \infty) e^{-i \nu} \Big) \setminus B^{(1)}_1$.
\end{lemma}
\begin{proof} We take $z \in A$. Lemma \ref{lemma:ib-dist-energy} implies that
$ | \widetilde{\lambda}_i - e_i | \leq |g| ( C_{\eqref{constgggg}} + 3 ) $.
Notice that
\begin{align}
\frac{1}{ | e_i - z | } \leq \frac{ \rho_m+ | \widetilde{\lambda}_i-z | }{ | e_i - z | }
\frac{1}{\rho_m+ | \widetilde{\lambda}_i-z |} \leq &
\frac{ \rho_m+ | \widetilde{\lambda}_i - e_i| + | e_i -z | }{ | e_i - z | }
\frac{1}{\rho_m+\left| \widetilde{\lambda}_i-z \right|} \\ \notag
\leq & c \frac{1}{\rho_m+\left| \widetilde{\lambda}_i-z \right|},
\end{align}
since $ | e_i - z | $ is bounded from below uniformly for $z \in A$. Then, the result follows from Lemma \ref{lemma:resestinA}. The result for
$ z \in \Big ( B^{(1)}_1 - [0, \infty) e^{-i \nu} \Big) $ can be found similarly which is why we omit the proof. The proof of \eqref{coroeq.ib-resT} follows from a similar argument as in Corollary \ref{coro:ib-res},and therefore, it is also omitted.
\end{proof}
\begin{definition} \label{compile}
In this subsection (Section \ref{RES}) we assumed a finite number of times that $ |g| $ is small enough (depending on $\boldsymbol{\nu}$ and $\rho_1$). We set $ \boldsymbol{g} > 0 $ such that for every $ |g| \leq \boldsymbol{g}$ all results of this section hold true. Similarly, in the statements of our results we use a finite number of times constants (that are independent of $\theta $, $n$, $g$, $\rho_0$,
$ \rho $ and $ \boldsymbol \nu $) in order to estimate from above norms of various entities. We denote by $ \boldsymbol{c} \geq 1 $ a, fixed, upper bound of the set of all these constants. We additionally take $\boldsymbol{g}$ small enough such that
$$
\norm{ P_i^{(1)}-P^{(1)}_{\text{at},i}} < 10^{-3},
$$
see Lemma \ref{lemma:ib-proj}.
\end{definition}
\section{Resonance and ground-state multiscale analysis}
\subsection{Notation: the sequence $(\rho_{n})_{n \in \mathbb{N} \cup \{ 0 \}}$ and the coupling constant $g$ }\label{notation}
Next, we introduce a constant, $ \boldsymbol{D} $, that includes all constants involved in estimations for our multiscale construction. This constant does not depend on $\theta \in \mathcal{S} $, $g$, $\boldsymbol{\nu} $, $n$, $\rho$ and $\rho_0$. Our bounds do depend on $\boldsymbol{\nu}$. They blow up as $ \boldsymbol{\nu} $ tends to zero with a rate that is not worse than $\sin( \boldsymbol{\nu}/2 )^{-3} $.
The constant $ \boldsymbol{D} $ does not depend on $ \boldsymbol{\nu} $. The dependence on $ \boldsymbol{\nu} $ in our bounds is reflected in a function
$\boldsymbol{C} \equiv \boldsymbol{C}( \boldsymbol{\nu} )$ that is bounded from below by the constant $\boldsymbol{D}$ multiplied by the factor $\sin( \boldsymbol{\nu}/2 )^{-3} $. As already explained, the constant $ \boldsymbol{D}$ and the blow-up rate
$\sin( \boldsymbol{\nu}/2 )^{-3} $ are intentionally not optimal but often estimated from above to increase the readability of the proofs.
\begin{definition}[The function $\boldsymbol{C} \equiv \boldsymbol{C}( \boldsymbol{\nu} ) $]
\label{C}
We fix a constant, $\boldsymbol{D}$, that does not depend on $g\in D(0,\boldsymbol g)$, $\theta \in \mathcal{S} $, and $ \boldsymbol{\nu} $. The only property that it must satisfy is the following (see Definition \ref{compile})
\begin{align}
\boldsymbol{D} \geq 10^6 + 10 \boldsymbol{c}.
\end{align}
Next, we fix a function $\boldsymbol{C} \equiv \boldsymbol{C}( \boldsymbol{\nu} ) $
satisfying
\begin{align}\label{dorm1}
\boldsymbol{C} \equiv \boldsymbol{C}( \boldsymbol{\nu} ) \geq \boldsymbol{D} \sin( \boldsymbol{\nu}/2 )^{-3} .
\end{align}
\end{definition}
The sequence $(\rho_{n})_{n \in \mathbb{N}_0}$ that we introduce above is defined in the following manner.
\begin{definition}[ Parameters $\rho_0$ and $\rho$] \label{sequence}
The parameters $\rho_0$ and $\rho$ in the definition of the sequence $\rho_n=\rho_0\rho^n , n\in\mathbb{N}_0$, see Definition \ref{def:Hntheta}(i), have to fulfill the following constraints:
\begin{align}\label{dorm2}
\boldsymbol{C}^8 \rho_0^{\mu} \leq 1, \qquad \boldsymbol{C}^4 \rho^{\mu} \leq 1/4.
\end{align}
\end{definition}
We recall that in this work we require $ |g|$ to be small enough. The next definition summarizes all requirements that it must satisfy.
\begin{definition}[The coupling constant $g$] \label{gzero} We set a constant $g_0 \leq \boldsymbol{g}$ satisfying the following (see Definition \ref{compile}):
\begin{align}\label{dorm3}
g_0 \leq \frac{\rho_1\sin( \boldsymbol{\nu} /2)^2}{10^4 \boldsymbol{c}}.
\end{align}
Henceforth, we always require $|g| \leq g_0 $.
\end{definition}
\begin{remark} \label{recuerda}
The selection of $\boldsymbol{C} $, the sequence $(\rho_{n})_{n \in \mathbb{N}_0}$ and the constant $g_0$ will later allow to set up the infrared induction scheme, and is therefore, rather involved. This remark is intended to help the reader to understand this procedure. Below (in this section), we use boldface fonts whenever we use the properties of $\boldsymbol{C}$ and $g$ (or $g_0$) that we specified above.
The requirements for $(\rho_{n})_{n \in \mathbb{N}_0}$ are only present in order to satisfy the last inequalities in Eqs.\ \eqref{eq:P1} and \eqref{eq:P3} below. Then, it will turn out that it is only necessary to assume that $ \boldsymbol{C}^4 \rho_0^{\mu} \leq 1 $ and
$\boldsymbol{C}^2 \rho^{\mu} \leq 1/2$ in order to close our induction. We assume stronger conditions again, for notational convenience, and because it implies a faster convergence rate in \eqref{eq:P3}, which will be used in a forthcoming paper (see Remark \ref{scattering}).
\end{remark}
\subsection{Induction scheme and the strategy of our multiscale construction}
\label{sec:indscheme}
We denote by
\begin{align} \label{pet1p}
\mathfrak{h}^{(n,n+1)}:= L^2(\mathcal B_{\rho_n}\setminus \mathcal B_{\rho_{n+1}})
\end{align}
the Hilbert space of one particle bosons with energies in the interval $ [ \rho_{n+1}, \rho_n ). $ We denote the corresponding Fock space by $ \mathcal F[\mathfrak{h}^{(n,n+1)}] $ (it is defined as in \eqref{Fock}). Note that $ \mathcal H^{(n+1)} $ is isomorphic to $ \mathcal H^{(n)}\otimes \mathcal F[\mathfrak{h}^{(n,n+1)}] $, and, therefore, we identify
\begin{align} \label{pet2p}
\mathcal H^{(n+1)}\equiv \mathcal H^{(n)}\otimes \mathcal F[\mathfrak{h}^{(n,n+1)}].
\end{align}
For $i=0,1$, we inductively (and simultaneously) construct sequences $\{\lambda^{(n)}_i\}_{n\in\mathbb{N}_0}$ of complex numbers, sequences $\{B^{(n)}_i\}_{n\in\mathbb{N}}$ of subsets of the complex plane and sequences $\{P^{(n)}_i\}_{n\in\mathbb{N}_0}$ of operators that satisfy the properties listed below.
\begin{enumerate}
\item[($\mathcal P 1$)] We set $\lambda^{(0)}_i\equiv \lambda^{(1)}_i $. For $n\in\mathbb{N}$, $\lambda^{(n)}_i$ is a simple eigenvalue of $H^{(n),\theta}$ and
\begin{align}
\label{eq:P1}
\left| \lambda^{(n)}_i-\lambda^{(n-1)}_i \right|< |g| \bold C^{n+1}(\rho_{n-1})^{1+\mu}\leq |g| \left(\frac{1}{2}\right)^{n-1}\rho_{n-1} .
\end{align}
The second inequality follows from Definition \ref{sequence}.
\item[($\mathcal P 2$)] For $n\in\mathbb{N}$, we define (recall that $ \nu = \Im \theta $)
\begin{align}
\label{eq:P21}
B_i^{(n)}:= B^{(1)}_i\setminus \left\{ z\in\mathbb{C} : \Im z <\Im \lambda^{(n)}_i - \frac{1}{4} \rho_n \sin (\nu) \right\} .
\end{align}
$\lambda^{(n)}_i$ is the only point in the spectrum of $H^{(n),\theta}$ intersected with $B_i^{(n)}$.
\item[($\mathcal P 3$)] We set $P^{(0)}_i\equiv P^{(1)}_i $. For $n\in\mathbb{N}$, we define
\begin{align}
\label{eq:projin}
P^{(n)}_i=-\frac{1}{2\pi i}\int_{\gamma^{(n)}_i}\mathrm{d}z \, \frac{1}{H^{(n),\theta}-z},
\end{align}
where
\begin{align}
\label{eq:gammain}
\gamma^{(n)}_i: [0,2\pi ] \to \mathbb{C} , \quad
t \mapsto \lambda^{(n)}_i +\frac{1}{4}\rho_n \sin (\nu) e^{it}.
\end{align}
The projections $P^{(n)}_i$ satisfy
\begin{align}
\label{eq:P3}
\norm{P^{(n)}_i - P^{(n-1)}_i\otimes P_{\Omega^{(n-1,n)}}} \leq \frac{|g|}{\rho}\bold C^{2n+2} \rho^\mu_{n-1}\leq \frac{|g|}{\rho} \left( \frac{1}{2}
\right)^{n-1},
\end{align}
where $P_{\Omega^{(n-1,n)}}$ is the projection onto the vacuum vector $\Omega^{(n-1,n)}\in \mathcal F[\mathfrak{h}^{(n-1,n)}]$ (see \eqref{pet1p}-\eqref{pet2p}). In \eqref{eq:P3} we omit the tensor product for $n=1$. The second inequality follows from Definition \ref{sequence}.
\item[($\mathcal P 4$)] Set $n\in\mathbb{N}$. For any $z\in B_i^{(n)}$, we have that
\begin{align}
\label{eq:P5}
\norm{\frac{1}{H^{(n),\theta}-z}\overline{P^{(n)}_i}}\leq \frac{\bold C^{n+1}}{\rho_n+\left| z-\lambda^{(n)}_i \right|} ,
\end{align}
where $\overline{P^{(n)}_i}=\mathbbm 1_{\mathcal H^{(n)}}-P^{(n)}_i$.
\end{enumerate}
\begin{theorem}[Multiscale analysis for resonances and ground state eigenvalues]
\label{thm:ind}
For every $ i \in \{ 0, 1 \} $ and $ \theta \in \mathcal{S} $ (see \eqref{def:setS}), there exist sequences of complex numbers $\{\lambda^{(n)}_i\}_{n\in\mathbb{N}_0}$, subsets of the complex plane $\{B^{(n)}_i\}_{n\in\mathbb{N}}$ and projection operators $\{P^{(n)}_i\}_{n\in\mathbb{N}_0}$ satisfying Properties ($\mathcal P 1$)-($\mathcal P 4$). Recall that we assume that $ |g | \leq g_0 $.
\end{theorem}
The proof for this theorem is given in Section \ref{sec:proof-induction}.
Similar results, for the Pauli-Fierz model, are derived in \cite{bbp}.
In the present paper we need uniform estimates with respect to $\theta \in \mathcal S$ and $ g \in D(0, g_0) $, in order to obtain uniform convergence with respect to these parameters (which is an important ingredient for the proof of analyticity). This is not the case in \cite{bbp} where analyticity is not an issue at stake.
\begin{remark}
\label{rem.ib}
Note that $(\mathcal P1)$ and $(\mathcal P3)$ hold true for $n=1$, by definition. Remark \ref{missing} implies that $(\mathcal P2)$ holds true for $n=1$. Moreover, $(\mathcal P4)$, for $n=1$, follows from Lemma \ref{lemma:ib-resprima} (Recall Definitions \ref{compile}, \ref{C} and \ref{gzero}).
\end{remark}
\subsection{Proof of Theorem~\ref{thm:ind}}
\label{sec:proof-induction}
We recall that in the remainder of this work we always assume that $| g| \leq
g_0 $ (see Definition \ref{gzero}) and $ \theta \in \mathcal{S} $ (see \eqref{def:setS}).
In Section \ref{sec:indkey}, we prove some key ingredients which are then used in Section \ref{sec:is} in order to conclude the induction step.
\subsubsection{Key estimates for the induction step}
\label{sec:indkey}
In this section, we assume that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\leq n \in\mathbb{N}$ and we derive some key estimates which we apply in the next section in order to show the induction step in the proof of Theorem \ref{thm:ind}.
By Eq.\ \eqref{h0def}, we define free boson energy operator restricted to $\mathcal F[\mathfrak{h}^{(n, n+1)}] $ and denote it by $ H_f^{(n, n+1), 0} \equiv H_f^{(n, n+1)}$ (see \eqref{pet1p}-\eqref{pet2p}). We set
\begin{align}\label{Hnthetaprima}
H_f^{(n, n+1), \theta }:= e^{-\theta} H_f^{(n, n+1), 0}.
\end{align}
For every function $ h \in \mathfrak{h}^{(n, n+1)}$, we denote the creation and annihilation operators, $a_{n, n+1}(h), \, a_{n, n+1}^*(h) $, on $ F[\mathfrak{h}^{(n,n+1)}] $ according to Eq.\ \eqref{aastar}. We use the same notation for functions $ h \in \mathfrak{h} $ but then understand the argument as $h$ restricted to $\mathfrak{h}^{(n,n+1)}$.
\noindent Furthermore, we fix the following operator (defined on $\mathcal K \otimes \mathcal F[\mathfrak{h}^{(n,n+1)}]$, and hence, on $ \mathcal H^{(n + 1)} $ - see Remark \ref{R})
\begin{align} \label{operators222}
V^{(n, n+1), \theta} := \sigma_1 \otimes \left(a_{n, n+1}(f^{\overline \theta})+ a_{n, n+1}(f^{\theta})^* \right).
\end{align}
\noindent In this notation we obtain (see Remark \ref{R}):
\begin{align} \label{HnHnm1}
H^{(n+1),\theta}= H^{(n),\theta}+H^{(n,n+1),\theta}_f+gV^{(n,n+1),\theta}.
\end{align}
\begin{lemma}
\label{lemma:is-normproj}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ such that $m\leq n$. Then,
\begin{align}
\norm{P^{(n)}_i} \leq 2+\frac{2|g|}{\rho} \leq 3 , \qquad i=0,1
\end{align}
\textbf{(notice that $|g| \leq \rho / 2$, see Definition \ref{gzero}- recall Remark \ref{recuerda})}
and
\begin{align}\label{obv}
|\lambda_i^{(n)} - \lambda_i^{(1)} | \leq 2 |g|.
\end{align}
\end{lemma}
\begin{proof}
Eq.\ \eqref{obv} is a consequence of Property ($\mathcal P 1$).
We estimate
\begin{align}
\norm{P^{(n)}_i}
&\leq \norm{ P^{(1)}_i}+ \sum^n_{j=2}\norm{P^{(j)}_i-P^{(j-1)}_i\otimes P_{\Omega^{(j-1,j)}}}
\notag \\
& \leq 2+\frac{|g|}{\rho} \sum^{n-1}_{j=0}\left( \frac{1}{2}
\right)^{j}
\leq 2+\frac{2|g|}{\rho},
\end{align}
where apply the induction hypothesis ($\mathcal P 3$) for $j\leq n$ and use Definition \ref{compile} \textbf{and Definition \ref{gzero}}.
\end{proof}
\begin{definition}
\label{def:regionM}
Let $n\in\mathbb{N}$ and $i\in\{0,1\}$. We define the region
\begin{align}
M_i^{(n)}:=B_i^{(n)}\setminus\left\{ z\in\mathbb{C} : \Im (z) \in \left(-\infty, \Im (\lambda^{(n)}_i) - \frac{2}{5}\rho_{n+1} \sin (\nu) \right) \right\}.
\end{align}
\end{definition}
\begin{lemma}
\label{lemma:is-keyest1}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ such that $m\leq n$.
Then, for $i\in\{0,1\}$:
\begin{align}
\norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}\overline{P^{(n,n+1)}_i}}\leq \frac{24+ 4 \bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} ,
\end{align}
for all $z\in M_i^{(n)}$, where
we have used the notation $P^{(n,n+1)}_i:=P^{(n)}_i \otimes P_{\Omega^{(n,n+1)}}$.
\end{lemma}
\begin{proof}
Let $z\in M_i^{(n)}$.
Note that (see Remark \ref{R})
\begin{align}
\label{eq:lemma-is-Pident}
\overline{P^{(n)}_i}+P^{(n)}_i\otimes \overline{P_{\Omega^{(n,n+1)}}}
&= 1-P^{(n)}_i+P^{(n)}_i\otimes \left( 1- P_{\Omega^{(n,n+1)}}\right)= 1-P^{(n)}_i \otimes P_{\Omega^{(n,n+1)}}
\notag \\
&=\overline{P^{(n,n+1)}_i},
\end{align}
and consequently, we obtain from functional calculus (notice that $\left[ H^{(n),\theta} , H_f^{(n,n+1),\theta} \right]=0$) that
\begin{align}
\label{eq:is-keyest10}
&\norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}\overline{P^{(n,n+1)}_i}}
\notag \\
&\leq \norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}\overline{P^{(n)}_i}}
+\norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}P^{(n)}_i\otimes \overline{P_{\Omega^{(n,n+1)}}}}
\notag \\
&=\sup_{s\in \{0\} \cup [\rho_{n+1},\infty)}
\norm{\frac{1}{H^{(n),\theta}-(z-e^{-\theta}s)}\overline{P^{(n)}_i}}
+\sup_{s\in [\rho_{n+1},\infty)}\frac{\norm{P^{(n)}_i}}{|\lambda^{(n)}_i-(z-e^{-\theta}s)|} .
\end{align}
Lemma \ref{lemma:resestinAprima}, Definition \ref{compile} and induction hypothesis ($\mathcal P4$), together with Lemma
\ref{lemma:is-normproj} \textbf{and the Definition of $\boldsymbol{C}$, see Remark \ref{recuerda}, in Definition \ref{C} (notice that
$ \boldsymbol{C} \geq \frac{4\boldsymbol{c}}{ \sin( \boldsymbol{\nu} /2 ) } \geq \frac{4 C_{\eqref{const:resestinAprima} } }{ \sin( \boldsymbol{\nu} /2 ) } $ and $ \| \overline{P^{(n)}_i} \| \leq 4$)}, imply that
\begin{align}\label{malech1}
\norm{\frac{1}{H^{(n),\theta}-(z-e^{-\theta}s)}\overline{P^{(n)}_i}} \leq \frac{\bold C^{n+1}}{\rho_n+\left| \lambda^{(n)}_i - (z-e^{-\theta}s) \right|} ,
\end{align}
for every $s \in \{0\} \cup [\rho_{n+1},\infty)$.
From the definitions of the sets $ M^{(n)}_i $ and
$ \mathcal{S} $, it follows that
\begin{align}
\label{eq:is-firstdist}
|\lambda^{(n)}_i-(z-e^{-\theta}s)| \geq \frac{1}{4} \rho_{n+1} \sin (\nu)
\end{align}
for all $z\in M^{(n)}_i$ and $s\in [\rho_{n+1},\infty)$.
Moreover, we define the sets
\begin{align}
\label{region:Gni}
G^{(n)}_i:= \left\{ z\in M^{(n)}_i :\Re (z) \geq \Re (\lambda^{(n)}_i) \right\}, \quad i=0,1 ,
\end{align}
and for $d \geq 0$
\begin{align}
\label{region:Lndi}
L^{(n),d}_i:= \left\{ \lambda^{(n)}_i+e^{-\theta}(x+id):x\in\mathbb{R} \right\}, \quad i=0,1 .
\end{align}
Furthermore, we define
\begin{align}
\label{region:Lni}
L^{(n)}_i:= \bigcup_{d\geq 0} L^{(n),d}_i \cap G^{(n)}_i, \quad i=0,1 .
\end{align}
Note that, by construction, we have
\begin{align}
\text{dist}\left(L^{(n),d}_i, \lambda^{(n)}_i\right)= e^{- \Re \theta} d , \quad i=0,1,
\end{align}
and, by definition of the sets $ M^{(n)}_i $ and
$ \mathcal{S} $, it follows that
\begin{align}
\label{eq:estwithoutL}
\left| z- \lambda^{(n)}_i \right|\leq \left| z- \lambda^{(n)}_i -e^{-\theta}s\right|, \qquad \forall z\in M^{(n)}_i\setminus L^{(n)}_i, \quad \forall s\in [\rho_{n+1},\infty) ,
\end{align}
where we have used the factor $ \frac{2}{5} $ in the definition of $ M^{(n)}_i $.
Let $Z^{1,d}_i$ and $Z^{2,d}_i$ be the intersections of $L^{(n),d}_i$ with the lines
\begin{align}
\lambda^{(n)}_i-i \frac{2}{5} \rho_{n+1}\sin(\nu)+\mathbb{R} \quad \text{and} \quad \lambda^{(n)}_i+\mathbb{R},
\end{align}
respectively. Furthermore, we define $Z^{3,d}_i:=\lambda^{(n)}_i+de^{i\frac{\pi}{2}-\theta}$ and recall that $\nu <\pi/16$. Then, we obtain
\begin{align}
\label{eq:regionest1}
&\sup_{z\in L^{(n),d}_i\cap G^{(n)}_i}\left| z- \lambda^{(n)}_i \right|^2
=\left| Z^{1,d}_i - \lambda^{(n)}_i \right|^2 =e^{-2\Re \theta}d^2 + \left| Z^{3,d}_i - Z^{1,d}_i \right|^2
\notag \\
&=e^{-2\Re \theta}d^2 + \left(\left| Z^{3,d}_i - Z^{2,d}_i \right|+\left| Z^{2,d}_i - Z^{1,d}_i \right|\right)^2
=e^{-2\Re \theta}d^2 + \left(\frac{e^{-\Re \theta}d}{\tan (\nu)}+
\frac{2}{5}\rho_{n+1}\right)^2 .
\end{align}
This yields the bound
\begin{align}
\label{eq:regionest2}
&\frac{\left| z- \lambda^{(n)}_i \right|}{\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|}
\notag \\
&\leq \left[ \frac{e^{-2\Re \theta}d^2}{\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|^2}+\left( \frac{e^{-\Re \theta}d}{ \tan (\nu) \left| z- \lambda^{(n)}_i-e^{-\theta}s \right|}+ \frac{ 2 \rho_{n+1}}{ 5\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|} \right)^2 \right]^\frac{1}{2}
\end{align}
for $s\in [\rho_{n+1},\infty)$ and $z\in L^{(n),d}_i\cap G^{(n)}_i$.
Note that $\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|\geq e^{-\Re \theta}d$ for all $z\in L^{(n),d}_i$ and together with \eqref{eq:is-firstdist} we obtain
\begin{align}
\label{eq:regionest3}
\frac{\left| z- \lambda^{(n)}_i \right|}{\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|}
\leq \left[ 1+\left( \frac{\cos (\nu)}{\sin (\nu) }+ \frac{8}{ 5\sin (\nu) }\right)^2 \right]^\frac{1}{2}\leq \frac{4}{\sin ( \boldsymbol \nu)},
\end{align}
for all $z\in L^{(n)}_i \cap G^{(n)}_i $. This and \eqref{eq:estwithoutL} guarantees
\begin{align}
\label{eq:regionestwithL}
\frac{1}{\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|}
\leq \frac{4}{\sin (\boldsymbol \nu)}\frac{1}{\left| z- \lambda^{(n)}_i \right|} \qquad \forall z\in M^{(n)}_i, \quad \forall s\in [\rho_{n+1},\infty) ,\quad i=0,1.
\end{align}
Now, if $\left| z- \lambda^{(n)}_i \right|\geq \rho_{n+1}$, we use \eqref{eq:regionestwithL} and compute
\begin{align}
\label{eq:finregest1}
\frac{1}{\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|}
&\leq \frac{4}{\sin (\boldsymbol \nu)}\frac{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|}{\left| z- \lambda^{(n)}_i \right|} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|}
\notag \\
&\leq \frac{8}{\sin (\boldsymbol \nu)}\frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} ,
\end{align}
and if $\left| z- \lambda^{(n)}_i \right|< \rho_{n+1}$, we use \eqref{eq:is-firstdist} and find
\begin{align}
\label{eq:finregest2}
\frac{1}{\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|}
&\leq \frac{4}{\sin (\boldsymbol \nu)}\frac{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|}{\rho_{n+1}} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|}
\notag \\
&\leq \frac{8}{\sin (\boldsymbol \nu)}\frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} .
\end{align}
We conclude from \eqref{eq:finregest1} and \eqref{eq:finregest2} that for $i=0,1$
\begin{align}
\label{eq:finregest0}
\frac{1}{\left| z- \lambda^{(n)}_i-e^{-\theta}s \right|}
\leq \frac{8}{\sin (\boldsymbol \nu)}\frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} \qquad \forall z\in M^{(n)}_i, \quad \forall s\in [\rho_{n+1},\infty)
\end{align}
holds true.
Eqs.\ \eqref{eq:is-keyest10}, \eqref{malech1}, \eqref{eq:regionestwithL}, together with Lemma \ref{lemma:is-normproj} and Eq.\ \eqref{eq:finregest0}
yield
\begin{align}
\norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}\overline{P^{(n,n+1)}_i}}
\leq \frac{24+ 4 \bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} .
\end{align}
This completes the proof.
\end{proof}
\begin{lemma}
\label{lemma:is-keyest2-help}
For all $z\in M_i^{(n)}\setminus \{\lambda_i^{(n)}\}$, all $ 0 \leq r \leq \left|z-\lambda_i^{(n)}\right| $ and every $i\in\{0,1\}$:
\begin{align}
\norm{\frac{H_f^{(n,n+1)}+r}{H_f^{(n,n+1),\theta}-\left(z-\lambda_i^{(n)}\right)}}\leq \frac{ 10 }{\sin (\boldsymbol \nu)} .
\end{align}
\end{lemma}
\begin{proof}
We calculate:
\begin{align}
\label{eq:helplemma1}
&\norm{\frac{H_f^{(n,n+1)}+r}{H_f^{(n,n+1),\theta}-(z-\lambda_i^{(n)})}}
= \sup_{y\in \{0\} \cup [\rho_{n+1},\infty)} \left| \frac{y+r}{e^{-\theta}y+\lambda_i^{(n)}-z} \right|
\notag \\
&\leq |e^{\theta}| + |e^{\theta}| \sup_{y\in \{0\} \cup [\rho_{n+1},\infty)} \left| \frac{e^{-\theta}r-\lambda_i^{(n)}+z}{e^{-\theta}y+\lambda_i^{(n)}-z} \right|
\leq |e^{\theta}| + (1 + |e^{\theta}| ) \sup_{y\in \{0\} \cup [\rho_{n+1},\infty)} \frac{\left| z- \lambda_i^{(n)}\right| }{\left| e^{-\theta}y+\lambda_i^{(n)}-z\right| }
\notag \\
&\leq |e^{\theta}| + \frac{ 4 (1 + |e^{\theta}| ) }{\sin ( \boldsymbol{\nu})}\leq \frac{ 10}{\sin (\boldsymbol \nu)},
\end{align}
where we have used \eqref{eq:regionestwithL} in the second last step.
\end{proof}
\begin{lemma}
\label{lemma:is-keyest2}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ such that $m\leq n$.
Then,
\begin{align}
\norm{V^{(n,n+1),\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}}\leq \frac{ 2500}{\rho \sin (\boldsymbol \nu)^2} \bold C^{n+1}\rho_n^{\mu}
\end{align}
for all $z\in M_i^{(n)}\setminus \{ \lambda_i^{(n)}\}$ such that $\left| z- \lambda_i^{(n)}\right| \geq \frac{1}{10}\rho_{n+1} \sin(\nu)$ and for all $i\in\{0,1\}$.
\end{lemma}
\begin{proof}
Set $r =\left| z- \lambda_i^{(n)}\right| \geq \frac{1}{10}\rho_{n+1}\sin(\nu)$. We observe that
\begin{align}
\label{eq:keyest20}
&\norm{V^{(n,n+1) ,\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}}
\leq \norm{V^{(n,n+1) ,\theta}\frac{1}{H_f^{(n,n+1)}+r}}
\norm{\frac{H_f^{(n,n+1)}+r}{H_f^{(n,n+1),\theta}-(z-\lambda_i^{(n)})}}
\notag \\
& \hspace{4.4cm} \cdot \norm{\left( H_f^{(n,n+1),\theta}-(z-\lambda_i^{(n)})\right)\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}} .
\end{align}
Lemma \ref{lemma:is-keyest2-help} yields
\begin{align}
\label{eq:keyest21}
\norm{\frac{H_f^{(n,n+1)}+r}{H_f^{(n,n+1),\theta}-(z-\lambda_i^{(n)})}}\leq \frac{ 10}{\sin (\boldsymbol \nu)} ,
\end{align}
and furthermore, we obtain from functional calculus that
\begin{align}
\label{eq:similar}
&\norm{\left( H_f^{(n,n+1),\theta}-(z-\lambda_i^{(n)})\right)\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}}
\leq \sup_{y\in\{0\}\cup [\rho_{n+1},\infty)}\norm{\frac{e^{-\theta}y+\lambda_i^{(n)}-z}{H^{(n),\theta}+e^{-\theta}y-z}}
\notag \\
&\leq \norm{ P_i^{(n)}}+\sup_{y\in\{0\}\cup [\rho_{n+1}, \infty)}\norm{\frac{\lambda_i^{(n)}-(z-e^{-\theta}y)}{H^{(n),\theta}-(z-e^{-\theta}y)}\overline{P_i^{(n)}}} \leq3+\bold C^{n+1} \leq 4 \bold C^{n+1}.
\end{align}
In the last step, we use Lemma \ref{lemma:is-normproj} for the first term. For the second term, we utilize Lemma \ref{lemma:resestinAprima}, Definition \ref{compile} and induction hypothesis ($\mathcal P4$), together with Lemma
\ref{lemma:is-normproj} \textbf{and the Definition of $\boldsymbol{C}$ in Definition \ref{C} - see Remark \ref{recuerda} } (notice that
$ \boldsymbol{C} \geq \frac{\boldsymbol{4 c}}{ \sin( \boldsymbol{\nu} /2 ) } \geq \frac{4 C_{\eqref{const:resestinAprima} } }{ \sin( \boldsymbol{\nu} /2 ) } $ and $ \| \overline{P^{(n)}_i} \| \leq 4$).
Using the proofs in Appendix \ref{app:sa}, we obtain
\begin{align}
\label{eq:keyest22}
&\norm{V^{(n,n+1) ,\theta}\frac{1}{H_f^{(n,n+1)}+r}}
\notag \\
& \leq \frac{1}{\sqrt{r}}\left( \norm{a_{n, n+1}(f^{\overline\theta}) \left( H_f^{(n,n+1)}+r\right)^{-\frac{1}{2}}}
+\norm{{a_{n, n+1}(f^\theta)}^*\left( H_f^{(n,n+1)}+r\right)^{-\frac{1}{2}}}\right)
\notag \\
&\leq \frac{1}{r} \norm{f^\theta}_{\mathfrak{h}^{(n,n+1)}} + \frac{2}{\sqrt{r}} \norm{f^\theta/\sqrt{\omega}}_{\mathfrak{h}^{(n,n+1)}}.
\end{align}
We estimate
\begin{align}\label{tuto1}
\norm{f^\theta}_{\mathfrak{h}^{(n,n+1)}}
&= \sqrt{\int_{\mathcal B_{\rho_n} \setminus \mathcal B_{\rho_{n+1}}} \mathrm{d}^3k \, |f^\theta(k)|^2}
= |e^{ - \theta (1+ \mu)} | \sqrt{4\pi\int^{\rho_{n}}_{\rho_{n+1}} \mathrm{d}u\, u^{1+2\mu}|e^{-2e^{2\theta}\frac{u^2}{\Lambda^2}}}|
\notag \\
&\leq |e^{ - \theta (1+ \mu)} | \sqrt{4\pi} \rho_n^{\mu} \rho_n,
\end{align}
and similarly,
\begin{align} \label{tuto2}
\norm{f^\theta/\sqrt \omega}_{\mathfrak{h}^{(n,n+1)}}
\leq |e^{ - \theta (1+ \mu)} | \sqrt{4\pi}\rho_n^{\mu}\rho_n^{\frac{1}{2}}.
\end{align}
From our choice of $r$, it follows that $\sqrt{\frac{\rho_n}{r}}\leq \frac{\sqrt{10}}{\sqrt{\rho \sin (\nu)}} $ and, consequently, we obtain
\begin{align}
\label{eq:keyest22'}
\norm{V^{(n,n+1) ,\theta}\frac{1}{H_f^{(n,n+1)}+r}}
&\leq |e^{ - \theta (1+ \mu)} | \sqrt{4\pi} \left( \frac{\rho_n}{r} +2\sqrt{ \frac{\rho_n}{r}}\right)\rho_n^{\mu}
\leq |e^{ - \theta (1+ \mu)} | \frac{60}{\rho \sin (\boldsymbol \nu)}\rho_n^{\mu} .
\end{align}
Plugging \eqref{eq:keyest21}, \eqref{eq:similar} and \eqref{eq:keyest22'} into \eqref{eq:keyest20} yields (we recall that $ \mu \in (0, 1/2)$)
\begin{align}
\norm{V^{(n,n+1) ,\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}}
\leq \frac{ 2500}{\rho \sin (\boldsymbol \nu)^2}\bold C^{n+1}\rho_n^{\mu}.
\end{align}
This completes the proof.
\end{proof}
\subsubsection{Induction step}
\label{sec:is}
In this section, we apply the results from Section \ref{sec:indkey} in order to show the induction step, i.e., we assume that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\leq n \in\mathbb{N}$, and prove that ($\mathcal P 1$)-($\mathcal P 4$) hold true for $ n+1$. This together with Remark \ref{rem.ib} completes the proof of Theorem \ref{thm:ind}.
We first employ the estimates of Section \ref{sec:indkey} in order to prove Property ($\mathcal P 2$) and ($\mathcal P 3$). After this, we prove ($\mathcal P 1$). Finally, ($\mathcal P 4$) follows again from the results of Section \ref{sec:indkey} together with the maximum modulus principle. \\
\textbf{Proof of ($\mathcal P 2$) and ($\mathcal P 3$) :}
\begin{proposition}
\label{prop:is-resestwithoutdisc}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ with $m\leq n$.
Then,
\begin{align}
\norm{\frac{1}{H^{(n+1),\theta}-z}}\leq\frac{168+ 8 \bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} ,
\end{align}
for all $z\in M_i^{(n)}\setminus \{ \lambda_i^{(n)}\}$ (see Definition \ref{def:regionM}) such that $\left| z- \lambda_i^{(n)}\right| \geq \frac{1}{10}\rho_{n+1} \sin(\nu)$ and for all $i\in\{0,1\}$.
\end{proposition}
\begin{proof}
Let $z\in M_i^{(n)}\setminus \{ \lambda_i^{(n)}\}$ such that $\left| z- \lambda_i^{(n)}\right| \geq \frac{1}{10}\rho_{n+1} \sin(\nu)$ and $i\in\{0,1\}$. Then, it follows from Lemma \ref{lemma:is-keyest2} that
\begin{align} \label{sss}
&\norm{V^{(n,n+1),\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}}
\leq \frac{ 2500}{\rho \sin (\boldsymbol \nu)^2} \bold C^{n+1}\rho_n^{\mu} .
\end{align}
\textbf{Our assumption on $g$ in Definition \ref{gzero} together with
\eqref{sss} yield that (notice that Definition \ref{sequence} implies that
$ C^{n+1}\rho_n^{\mu} \leq 1 $ and Definition \ref{gzero} implies that
$ \frac{ g 2500}{\rho \sin (\boldsymbol \nu)^2} \leq \frac{1}{2} $, see also Remark \ref{recuerda})}
\begin{align}
\label{eq:is-prop-relboundg}
\norm{gV^{(n,n+1),\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}}
& \leq \frac{1}{2} .
\end{align}
This and Lemma \ref{lemma:is-keyest1} imply that
\begin{align}
H^{(n+1),\theta}-z=\left(1+ gV^{(n,n+1),\theta} \frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z } \right)\left( H^{(n),\theta}+H_f^{(n,n+1),\theta}-z \right)
\end{align}
is invertible, and we estimate
\begin{align}
\label{eq-is-propwithoutdisc}
\norm{\frac{1}{H^{(n+1),\theta}-z}}
&\leq 2 \norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z }}
\notag \\
&\leq 2\frac{\norm{P^{(n,n+1)}_i}}{\left| z-\lambda^{(n)}_i \right|}
+2
\norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}\overline{P^{(n,n+1)}_i}}
\notag \\
&\leq 2\frac{\norm{P^{(n)}_i}}{\left| z-\lambda^{(n)}_i \right|}
+
\frac{48+ 8 \bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|},
\end{align}
where we apply Lemma \ref{lemma:is-keyest1}. Moreover, Lemma \ref{lemma:is-normproj} implies that $\norm{P^{(n)}_i}\leq 3$ and it follows from $\left| z- \lambda_i^{(n)}\right| \geq \frac{1}{10}\rho_{n+1} \sin(\nu)$ that
\begin{align}
\frac{1}{\left| z-\lambda^{(n)}_i \right|}\leq \frac{20}{\sin(\boldsymbol \nu)}\frac{1}{\rho_{n+1}+\left| z-\lambda^{(n)}_i \right|} .
\end{align}
Altogether, we obtain
\begin{align}
\norm{\frac{1}{H^{(n+1),\theta}-z}}
&\leq
\frac{168+ 8 \bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} ,
\end{align}
and thereby, complete the proof.
\end{proof}
\begin{lemma}
\label{lemma:is-p3}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ with $m\leq n$. We define
\begin{align}
\label{eq:Phatn+1}
\hat P^{(n+1)}_i := -\frac{1}{2\pi i} \int_{\hat \gamma^{(n)}_i}\mathrm{d}z \, \frac{1}{H^{(n+1),\theta}-z},
\end{align}
where
\begin{align}
\label{eq:gammainhat}
\hat \gamma^{(n)}_i: [0,2\pi ] \to \mathbb{C} , \quad
t \mapsto \lambda^{(n)}_i + \frac{1}{8} \rho_{n+1} \sin (\nu) e^{it} .
\end{align}
Then,
\begin{align}
\norm{\hat P^{(n+1)}_i - P^{(n)}_i\otimes P_{\Omega^{(n,n+1)}}} \leq \frac{|g|}{\rho}\bold C^{2(n+1)+2} \rho^\mu_{n}\leq \frac{|g|}{\rho} \left( \frac{1}{2}
\right)^{n} .
\end{align}
(The last inequality follows from Definition \ref{sequence}.)
\end{lemma}
\begin{proof}
Recall that the definition of $P^{(n)}_i$ is introduced in Eq.\ \eqref{eq:projin}. We notice that the function
\begin{align}
\label{eq:p3map}
B^{(n)}_i\setminus \{ \lambda^{(n)}_i \} \ni z \mapsto \frac{1}{H^{(n),\theta}-z}
\end{align}
is analytic as an operator valued function and the region between $\hat \gamma^{(n)}_i$ and $ \gamma^{(n)}_i$ is contained in the domain of \eqref{eq:p3map}. We obtain from the Cauchy integral theorem that
\begin{align}
P^{(n)}_i = -\frac{1}{2\pi i} \int_{ \gamma^{(n)}_i}\mathrm{d}z \, \frac{1}{H^{(n),\theta}-z}= -\frac{1}{2\pi i} \int_{ \hat \gamma^{(n)}_i}\mathrm{d}z \, \frac{1}{H^{(n),\theta}-z} .
\end{align}
\noindent As in Remark \ref{remp}, it turns out that (see Remark \ref{R})
\begin{align}
P^{(n)}_i\otimes P_{\Omega^{(n,n+1)}} = -\frac{1}{2\pi i} \int_{ \hat \gamma^{(n)}_i}\mathrm{d}z \, \frac{1}{H^{(n),\theta} + H_f^{(n,n+1),\theta} -z} .
\end{align}
We calculate
\begin{align} \label{caraj1}
&\big \| \hat P^{(n+1)}_i - P^{(n)}_i\otimes P_{\Omega^{(n,n+1)}} \big \| =
\frac{1}{2\pi } \Big \| \int_{ \hat \gamma^{(n)}_i}\mathrm{d}z \,
\frac{1}{H^{(n+1),\theta} -z} -
\frac{1}{H^{(n),\theta} + H_f^{(n, n+1),\theta} -z} \Big \|
\notag \\
& =
\frac{1}{2\pi } \Big \| \int_{ \hat \gamma^{(n)}_i}\mathrm{d}z \,
\frac{1}{H^{(n+1),\theta} -z} gV^{(n,n+1),\theta}
\frac{1}{H^{(n),\theta} + H_f^{(n, n+1),\theta} -z} \Big \| .
\end{align}
\noindent Furthermore, Lemma \ref{lemma:is-keyest2} implies that for $z$ in the curve $ \hat \gamma^{(n)}_i $
\begin{align}
\label{caraj2}
\norm{ gV^{(n,n+1),\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}- z}}\leq
|g| \frac{ 2500}{\rho \sin (\boldsymbol \nu)^2} \bold C^{n+1}\rho_n^{\mu}
,
\end{align}
and Proposition \ref{prop:is-resestwithoutdisc} ensures that
\begin{align}
\label{caraj3}
\norm{\frac{1 }{H^{(n+1),\theta}- z}}
&\leq\frac{168+ 8\bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1} } .
\end{align}
Eqs \eqref{caraj1}-\eqref{caraj3} imply
\begin{align}
\norm{\hat P^{(n+1)}_i - P^{(n)}_i\otimes P_{\Omega^{(n,n+1)}}}
\leq |g|
\frac{ 2500}{\rho \sin (\boldsymbol \nu)^2} \bold C^{n+1}\rho_n^{\mu} \frac{168+ 8\bold C^{n+1}}{\sin (\boldsymbol \nu)} \leq \frac{ |g| 2500}{\rho \sin (\boldsymbol \nu)^2} \bold C^{2n+2}\rho_n^{\mu} \frac{200 }{\sin (\boldsymbol \nu)},
\end{align}
\textbf{which together with the definition of $\boldsymbol{C}$ in Definitions \ref{C} imply the desired result (Definition \ref{C} imply that
$ \frac{ 500 \, 000}{ \sin (\boldsymbol \nu)^3} \leq \bold C^2 $, see also Remark \ref{recuerda}). }
\end{proof}
\begin{proposition}[Proof of Properties ($\mathcal P2$) and ($\mathcal P3$)]
\label{rem:1}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ with $m\leq n$, then ($\mathcal P 2$) and ($\mathcal P 3$) hold true for $n+1$.
\end{proposition}
\begin{proof}
Lemma \ref{lemma:is-p3} implies that $\norm{ \hat P^{(n+1)}_i - P^{(n)}_i\otimes P_{\Omega^{(n,n+1)}}} < 1$ \textbf{(see Definition \ref{gzero} and recall Remark \ref{recuerda})}. From the induction hypothesis it follows that $P^{(n)}_i\otimes P_{\Omega^{(n,n+1)}}$ is a rank-one projection. Therefore, $\hat P^{(n+1)}_i $ is also a rank-one projection, and hence, there is only one point in $ M^{(n)}_i $ contained in the spectrum of $H^{(n+1)}_i$. This point is the eigenvalue $\lambda^{(n+1)}_i$ that we introduced above. Lemma \ref{prop:is-resestwithoutdisc} implies that $ | \lambda^{(n+1)}_i - \lambda^{(n)}_i | \leq \frac{1}{10} \rho_{n+1} \sin(\nu) $. This in turn implies that $ B_i^{(n+1)} \subset M_i^{(n)} $. Then, $ \lambda^{(n+1)}_i $ is the only spectral point of $ H_i^{(n+1), \theta} $ in $ B_i^{(n+1)} $, which is Property ($\mathcal P 2$). A deformation in the integration contour in the definitions of
$ \hat P^{(n+1)}_i $ and $ P^{(n+1)}_i $ implies that these projections coincide and, therefore, Property ($\mathcal P 3$) is a consequence of Lemma \ref{lemma:is-p3} .
\end{proof}
\textbf{Proof of Property ($\mathcal P 1$):}
\begin{proposition}[Proof of Property ($\mathcal P1$)]
\label{prop:is-p1}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ with $m\leq n$.
Then, we obtain for $i=0,1$ that
\begin{align}
\left| \lambda^{(n+1)}_i - \lambda^{(n)}_i \right|
\leq |g| \bold C^{(n+1)+1}\rho_n^{1+\mu} \leq |g| \left( \frac{1}{2} \right)^n \rho_n
\end{align}
holds true.
Notice that the last inequality follows from Definition \ref{sequence}.
\end{proposition}
\begin{proof}
In this proof we explicitly emphasize the dependence of $ P^{(n)}_{i} $ on $\theta$ and write $ P^{(n)}_{i} \equiv P^{(n),\theta}_{i}$. We define $\Psi^{(n),\theta}_{i}:= P^{(n),\theta}_{i}\varphi_i\otimes \Omega^{(n+1)} $, see Remark \ref{R}. Proposition \ref{rem:1}, Property $\mathcal P 3$, Definition \ref{compile} (see Remark \ref{lemma:ib-proj2}) \textbf{and the restrictions for $g$ in Definition \ref{gzero} imply that $ \| \Psi^{(n),\theta}_{i} - \varphi_i\otimes \Omega^{(n + 1)} \| \leq \frac{1}{10^2} $. This guarantees that }
\begin{align}
\label{eq:p1proof}
\norm{\Psi^{(n),\theta}_{i}}\leq 2 \qquad \text{and} \qquad \Big | \left\langle \Psi^{(n),\overline \theta}_{i}, P^{(n+1),\theta}_{i} \Psi^{(n),\theta}_{i}\right\rangle \Big | \geq \frac{1}{2} .
\end{align}
Notice that in this work we assume that the imaginary part of $\theta$, $ \nu $, is positive. Then, strictly speaking, we do not have the right to use the symbol $\Psi^{(n), \overline \theta}_{i}:= P^{(n),\overline \theta}_{i}\varphi_i\otimes \Omega^{(n+1)} $. However, the restriction we impose by assuming that $ \nu $ is not negative is irrelevant. This is assumed only for convenience in order to simplify our notation. Of course, the same results hold true if we take $ - \pi/16 < \nu < - \boldsymbol{\nu} $ (we use this fact in the present proof as well as $ P^{(n),\overline \theta}_{i} = \big ( P^{(n),\theta}_{i} \big )^*$).
Then, we obtain
\begin{align}
\label{eq:distlambdas}
\lambda_i^{(n+1)}
&= \frac{\left\langle \Psi^{(n),\overline \theta}_{i}, H^{(n+1), \theta} P^{(n+1),\theta}_i \Psi^{(n),\theta}_{i}\right\rangle}{\left\langle \Psi^{(n),\overline \theta}_{i}, P^{(n+1),\theta}_{i} \Psi^{(n),\theta}_{i}\right\rangle}
= \frac{\left\langle H^{(n+1),\overline \theta}\Psi^{(n),\overline \theta}_{i}, P^{(n+1),\theta}_i \Psi^{(n),\theta}_{i}\right\rangle}{\left\langle \Psi^{(n),\overline \theta}_{i}, P^{(n+1),\theta}_{i} \Psi^{(n),\theta}_{i}\right\rangle}
\notag \\
&=\lambda_i^{(n)}+ g\frac{\left\langle V^{(n,n+1),\overline \theta} \Psi^{(n),\overline \theta}_{i}, P^{(n+1),\theta}_{i} \Psi^{(n),\theta}_{i}\right\rangle}{\left\langle \Psi^{(n),\overline \theta}_{i}, P^{(n+1),\theta}_{i} \Psi^{(n),\theta}_{i}\right\rangle} .
\end{align}
Now we choose $z\in\mathbb{C}$ such that $|z-\lambda_i^{(n)}|=\frac{\rho_{n+1} \sin (\nu)}{10}$. We get that
\begin{align}
&\left| \lambda_i^{(n+1)} -\lambda_i^{(n)}\right|
\leq 2 \norm{gV^{(n,n+1) ,\overline \theta} P^{(n),\overline \theta}_{i} \varphi_i \otimes \Omega^{(n+1)}} \Big \| P^{(n+1),\theta}_{i} \Psi^{(n),\theta}_{i} \Big \| \notag \\ &\leq 54 \left| z-\lambda_i^{(n)} \right| \norm {gV^{(n,n+1),\overline \theta} \frac{1}{H^{(n),\overline \theta} +H_f^{(n,n+1),\overline \theta} -z}}
\notag \\
&\leq g 54 \frac{\rho_{n+1} \sin (\nu) }{10} \frac{2500}{\rho \sin (\boldsymbol{ \nu})^2} \bold C^{n+1}\rho_n^{\mu}
\leq |g| \bold C^{(n+1)+1}\rho_n^{1+\mu} \leq |g| \left( \frac{1}{2} \right)^n \rho_n ,
\end{align}
where we use Lemmas \ref{lemma:is-keyest2} and \ref{lemma:is-normproj} \textbf{and the definition of $ \boldsymbol{C} $ in Definition \ref{C} (it implies that $ 54 \frac{2500}{\sin(\boldsymbol{\nu})^2} \leq \boldsymbol{C} $, see also Remark \ref{recuerda}).}
\end{proof}
\textbf{Proof of Property ($\mathcal P 4$):}
\begin{lemma}
\label{prop:is-resest-fin}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ with $m\leq n$.
Then, for $i\in\{0,1\}$:
\begin{align}
\label{eq:is-resest-fin}
\norm{\frac{1}{H^{(n+1),\theta}-z}\overline{P^{(n+1)}_i}}\leq 3 \frac{168+ 8\bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|},
\qquad \forall z\in M_i^{(m)},
\end{align}
and hence, for all $z \in B_i^{(n+1)}$ (recall $ B_i^{(n+1)} \subset M_i^{(n)}$ from the proof of Proposition \ref{rem:1}).
\end{lemma}
\begin{proof}
Let $z\in M_i^{(n)} $ such that $\left| z- \lambda^{(n)}_i \right|\geq \frac{\rho_{n+1}}{10} \sin(\nu)$ and $i\in\{0,1\}$. Then,
\eqref{eq:is-resest-fin} follows from
Proposition \ref{prop:is-resestwithoutdisc} and the fact that
$\big \| \overline{ P^{(n+1)}_i} \big \| \leq 3$ (see the proof of Lemma \ref{lemma:is-normproj} and Proposition \ref{rem:1}). Furthermore, we observe that $ M_i^{(n)} \ni z \mapsto\frac{1}{H^{(n+1),\theta}-z}\overline{P^{(n+1)}_i}$ is analytic (see the proof of Proposition \ref{rem:1}), and hence, \eqref{eq:is-resest-fin} follows for $\left| z- \lambda^{(n)}_i \right|\leq \frac{\rho_{n+1}}{10} \sin(\nu)$ from the maximum modulus principle of complex analysis.
\end{proof}
\begin{proposition}[Proof of Property ($\mathcal P 4$)]
\label{prop:is-p5}
Suppose that ($\mathcal P 1$)-($\mathcal P 4$) hold true for all $m\in\mathbb{N}$ with $m\leq n$ and take $i\in\{0,1\}$.
Then,
\begin{align}
\label{eq:is-resest-p5}
\norm{\frac{1}{H^{(n+1),\theta}-z}\overline{P^{(n+1)}_i}}
&\leq 4 \frac{168+ 8\bold C^{n+1}}{|\sin (\boldsymbol \nu)|} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n+1)}_i \right|}
\leq \frac{\bold C^{(n+1)+1}}{\rho_{n+1}+\left| z- \lambda^{(n+1)}_i \right|}
\end{align}
for all $z\in B^{(n+1)}_i$.
\end{proposition}
\begin{proof}
Let $i\in\{0,1\}$ and $z\in B^{(n+1)}_i$. It follows from Proposition \ref{prop:is-resest-fin} that
\begin{align}
\label{eq:p5-est0}
\norm{\frac{1}{H^{(n+1),\theta}-z}\overline{P^{(n+1)}_i}}
&\leq 3 \frac{168+ 8C(\nu)^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|}
\end{align}
holds true.
Lemma \ref{prop:is-resestwithoutdisc} implies that $ | \lambda^{(n+1)}_i - \lambda^{(n)}_i | \leq \frac{1}{10} \rho_{n+1} \sin(\nu) \leq \frac{1}{10} \rho_{n+1} $ . Therefore,
\begin{align}
\frac{1}{\rho_{n+1}+\left| z- \lambda^{(n)}_i \right|} \leq \frac{10}{9}\frac{1}{\rho_{n+1}+\left| z- \lambda^{(n+1)}_i \right|} .
\end{align}
This together with \eqref{eq:p5-est0} yields
\begin{align}
\label{eq:p5-est0'}
\norm{\frac{1}{H^{(n+1),\theta}-z}\overline{P^{(n+1)}_i}}
&\leq 4 \frac{168+ 8\bold C^{n+1}}{\sin (\boldsymbol \nu)} \frac{1}{\rho_{n+1}+\left| z- \lambda^{(n+1)}_i \right|}
\leq \frac{\bold C^{(n+1)+1} }{\rho_{n+1}+\left| z- \lambda^{(n+1)}_i \right|} ,
\end{align}
where in the last line \textbf{we use the definition of $ \boldsymbol C $ in Definition \ref{C} (it implies that $ 4 \frac{168+ 8 }{\sin (\boldsymbol \nu)} \leq \bold C $, see also Remark \ref{recuerda}).}
\end{proof}
\section{Resolvent and spectral estimates}
\label{resolvent-estimates}
In this section we assume that Definitions \ref{C}, \ref{sequence} and \ref{gzero} hold true. We additionally assume that $ \boldsymbol{ C} \geq \boldsymbol{D}\sin(\nu/m)^{-1} $, in order to freely apply Corollary \ref{coro:ib-res} and Eq. \eqref{coroeq.ib-resT}.
We fix the Hamiltonians (see Remark \ref{R})
\begin{align}
\tilde H^{(n),\theta} := H^{\theta}_0 +gV^{(n),\theta},
\end{align}
which are densely defined on the Hilbert space $\mathcal H$.
We recall that we already defined
\begin{align} \label{pet1}
\mathfrak{h}^{(n,\infty)}= L^2(\mathcal B_{\rho_n} )
\end{align}
and the corresponding Fock space $ \mathcal F[\mathfrak{h}^{(n, \infty)}] $ (it is defined in \eqref{Fock}), with vacuum state $\Omega^{(n, \infty)}$. We identify, as above,
\begin{align} \label{pet2}
\mathcal H \equiv \mathcal H^{(n)}\otimes \mathcal F[\mathfrak{h}^{(n,\infty )}].
\end{align}
We define the free boson energy operator on $ F[\mathfrak{h}^{(n, \infty)}] $ by restricting the definition in Eq.\ \eqref{h0def} accordingly and denote it by the symbol $ H_f^{(n, \infty), 0} \equiv H_f^{(n, \infty)} $. We set
\begin{align}\label{Hnthetaprima111}
H_f^{(n, \infty), \theta}:= e^{-\theta} H_f^{(n, \infty), 0}.
\end{align}
For every function $ h \in \mathfrak{h}^{(n, \infty)}$, we define the creation and annihilation operators, $a_{n, \infty}(h)$ and $a_{n, \infty}^*(h) $, on $ F[\mathfrak{h}^{(n,\infty)}] $ according to Eq.\ \eqref{aastar}. Again, we use the same notation also for $ h \in \mathfrak{h} $ but then understand $h$ as its restriction to $\mathfrak{h}^{(n,n+1)}$.
\noindent We fix the following operator (defined on $\mathcal K \otimes \mathcal F[\mathfrak{h}^{(n,\infty)}] $, and hence, on $ \mathcal H $ - see Remark \ref{R})
\begin{align} \label{operators333}
V^{(n, \infty), \theta} := \sigma_1 \otimes \left(a_{n, \infty}(f^{\overline \theta})+ a_{n, \infty}(f^{ \theta})^* \right),
\end{align}
and further, we obtain (see Remark \ref{R}):
\begin{align} \label{HnHnmnn}
H^{\theta}= H^{(n),\theta}+H^{(n,\infty),\theta}_f+gV^{(n,\infty),\theta}
= \tilde H^{(n),\theta} +gV^{(n,\infty),\theta} .
\end{align}
\subsection{Resolvent and spectral estimates multi-scale analysis}\label{resspec}
\begin{lemma} Suppose that $|g| \leq \rho \frac{1}{10} \sin(\nu / 2 m)$.
We define for $i=0,1$
\begin{align}
v^{(n)}_i := \lambda^{(n)}_i +
\frac{1}{4} \rho_n e^{-i \nu}, \hspace{1cm} v^{(n, n+1)}_i := \lambda^{(n)}_i +
\Big (\frac{2}{5} - \frac{1}{100} \Big ) \rho_{n+1} e^{-i \nu} .
\end{align}
It follows that
\begin{align} \label{ter0}
| \lambda_i - \lambda_i^{(n)} | \leq & 2 |g| \rho_{n}^{1+ \mu/2}
\end{align}
and
\begin{align}\label{ter1}
\mathcal{C}_m ( v^{(n)}_i ) \subset \mathcal{C}_m ( v^{(n, n+1)}_i ) \subset \mathcal{C}_m ( v^{(n+ 1)}_i ),
\end{align}
where the set $\mathcal C_m(\cdot)$ is defined in \eqref{eq:defcone}.
Moreover,
\begin{align}\label{ladist}
{\rm dist} \Big ( \mathcal{C}_m ( v^{(n, n+1)}_i ) , \mathbb{C} \setminus \mathcal{C}_m ( v^{( n+1)}_i ) \Big ) \geq \sin( \nu / 2 m ) \frac{1}{10}\rho_{n+ 1}.
\end{align}
and
\begin{align}\label{ladistf}
{\rm dist} \Big ( \mathcal{C}_m ( v^{(n)}_i ) , \mathbb{C} \setminus \mathcal{C}_m ( v^{(n, n+1)}_i ) \Big ) \geq \sin( \nu / m ) \frac{1}{10}\rho_{n+ 1}.
\end{align}
\end{lemma}
\begin{proof}
That $ \mathcal{C}_m ( v^{(n)}_i ) \subset \mathcal{C}_m ( v^{(n, n+1)}_i ) $ is immediate.
From Theorem \ref{thm:ind} (Property $\mathcal{P}1$) and Definition \ref{sequence} it follows that
\begin{align}\label{ter2}
| \lambda_i^{(n+1)} - \lambda_i^{(n)} | \leq |g| \big (\bold C^4 \rho_0^{ \mu } \big )^{1/2}
\Big ( \bold (C^2 \rho^{\mu})^{n} \Big )^{1/2} \rho_{n}^{1+ \mu/2} \leq |g| \frac{1}{2 ^{n}} \rho_{n}^{1+ \mu/2}.
\end{align}
This and a geometric series argument prove \eqref{ter0}. We write
\begin{align}
v_i^{(n, n+1)} = v_{i}^{(n+1)} + \xi_1 e^{-i \nu} + \xi_2 i e ^{-i\nu}.
\end{align}
Eq.\ \eqref{ter2} implies that
\begin{align}\label{ter3}
|\xi_2| \leq |g| \rho_{n}, \hspace{2cm} \xi_1 \geq \Big (\frac{2}{5} - \frac{1}{100} - \frac{1}{4} \Big ) \rho_{n+1} - |g| \rho_{n} > \frac{1}{10}\rho_{n+ 1}.
\end{align}
The last step follows for $g>0$ sufficiently small (see Definition \ref{gzero}).
To prove that $ \mathcal{C}_m ( v^{(n, n+1)}_i ) \subset \mathcal{C}_m ( v^{(n+ 1)}_i ) $, it is enough to show that $ v^{(n, n+1)}_i \in \mathcal{C}_m ( v^{(n+ 1)}_i ) $. We shall prove that
\begin{align}\label{ter4}
|\xi_2| / \xi_1 < \tan(\nu / 2 m),
\end{align}
which holds true if $ | g | \leq \rho \frac{1}{10} \sin(\nu / 2 m) \leq \rho \frac{1}{10} \tan(\nu / 2 m) $.
Eq.\ \eqref{ladistf} is is implied by the particular geometry considered here because both cones have the same axis, see also Definition \ref{sequence}.
Eq. \eqref{ter4} implies that the angle between the axis of the cone $ \mathcal{C}_m ( v^{(n+ 1)}_i ) $ and the complex number $ v_i^{(n, n+1)} - v_{i}^{(n+1)} $ is smaller than $ \nu / 2 m $ and, therefore, the angle between this complex number and the closest edge of the cone must be larger than $ \nu / 2 m $. Then, the distance between the referred complex number and the edge is larger than
$$ | v_i^{(n, n+1)} - v_{i}^{(n+1)} | \sin( \nu / 2 m ) \geq \xi_1 \sin( \nu / 2 m ) \geq \sin( \nu / 2 m ) \frac{1}{10}\rho_{n+ 1}, $$
this implies \eqref{ladist}.
\end{proof}
\begin{lemma} \label{mana0}
Assume that for all
$ z \in B^{(1)}_i \setminus \mathcal{C}_m( v_i^{(n)} ) $
\begin{align}
\Big \| \frac{1}{ H^{(n), \theta} - z} \overline{P^{(n)}_i} \Big \| \leq \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z , \mathcal{C}_m( v_i^{(n)} ) \Big ) },
\end{align}
then
\begin{align}\label{haha1}
\Big \| \frac{1}{ H^{(n), \theta} + H_{f}^{(n, n+1), \theta} - z} \overline{P^{(n, n + 1)}_i} \Big \| \leq 4 \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) } ,
\end{align}
and
\begin{align}\label{haha2}
\Big \| \frac{1}{ H^{(n), \theta} + H_{f}^{(n, n+1), \theta} - z} H_{f}^{(n, n+1)} \overline{P^{(n, n + 1)}_i }\Big \|
\leq \frac{100}{\sin(\nu/m)} \boldsymbol{C}^{n+1}
,
\end{align}
for all $ z \in B^{(1)}_i \setminus \mathcal{C}_m( v_i^{(n, n+1)} ) $.
\end{lemma}
\begin{proof}
Take $ z \in B^{(1)}_i \setminus \mathcal{C}_m( v_i^{(n, n+1)} ) $.
We use the spectral theorem and that (see \eqref{eq:lemma-is-Pident})
\begin{align}
\label{eq:lemma-is-Pident-R}
\overline{P^{(n)}_i}+P^{(n)}_i\otimes \overline{P_{\Omega^{(n,n+1)}}}
= \overline{P^{(n,n+1)}_i},
\end{align}
to calculate
\begin{align}
\label{eq:is-keyest10-R}
&\norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}\overline{P^{(n,n+1)}_i}}
\notag \\
&\leq \norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}\overline{P^{(n)}_i}}
+\norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}P^{(n)}_i\otimes \overline{P_{\Omega^{(n,n+1)}}}}
\notag \\
&=\sup_{s\in \{0\} \cup [\rho_{n+1},\infty)}
\norm{\frac{1}{H^{(n),\theta}-(z-e^{-\theta}s)}\overline{P^{(n)}_i}}
+\sup_{s\in [\rho_{n+1},\infty)}\frac{\norm{P^{(n)}_i}}{|\lambda^{(n)}_i-(z-e^{-\theta}s)|} .
\end{align}
Thanks to the geometry, for all $s \geq 0$, we have
\begin{align}\label{coco1}
{\rm dist} \Big ( z - e^{- \theta } s, \mathcal{C}_m( v_i^{(n)} ) \Big ) \geq
{\rm dist} \Big ( z, \mathcal{C}_m( v_i^{(n)} ) \Big ) \geq {\rm dist} \Big ( z, \mathcal{C}_m( v_i^{(n, n+1)} ) \Big ) .
\end{align}
Eq.\ \eqref{coco1}, our hypothesis, Lemma \ref{lemma:resestinAprima} and Definitions \ref{compile} and \ref{C} imply that for $ s \geq 0 $
\begin{align}\label{coco2}
\norm{\frac{1}{H^{(n),\theta}-(z-e^{-\theta}s)}\overline{P^{(n)}_i}}
\leq \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) }.
\end{align}
Notice that for $ s \geq \rho_{n+1}$,
\begin{align}\label{coco3}
z - e^{ - \theta} s \notin \mathcal{C}_{m}( \lambda_i^{(n)} )
\end{align}
and, therefore,
\begin{align}\label{coco4}
|\lambda^{(n)}_i-(z-e^{-\theta}s)| \geq {\rm dist} \Big ( z - e^{- \theta } s, \mathcal{C}_m( \lambda_i^{(n)} ) \Big ) = & {\rm dist} \Big ( z ,
\mathcal{C}_m(\lambda_i^{(n)} + e^{- \theta } s ) \Big )
\\ \notag \geq & {\rm dist} \Big ( z ,
\mathcal{C}_m( v_i^{(n, n+1)} ) \Big ).
\end{align}
Eqs.\ \eqref{coco1}, \eqref{coco2} and \eqref{coco4}, and Lemma \ref{lemma:is-normproj} together with Definition \ref{C} imply Eq.\ \eqref{haha1}.
Now, we prove Eq.\ \eqref{haha2}. As in \eqref{eq:is-keyest10-R} and \eqref{coco2}, we have
\begin{align}
\label{coco5}
& \norm{\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z} H_f^{(n,n+1)} \overline{P^{(n,n+1)}_i}}
\notag \\
& \hspace{.5cm} \leq \sup_{s\in \{0\} \cup [\rho_{n+1},\infty)}\boldsymbol{C}^{n+1} \frac{s }{{\rm dist} \Big ( z - e^{-\theta} s , \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) }
+\sup_{s\in [\rho_{n+1},\infty)}\frac{\norm{P^{(n)}_i} s}{|\lambda^{(n)}_i-(z-e^{-\theta}s)|} .
\end{align}
Notice that
for $z \notin \mathcal{C}_m( v_i^{(n, n+1 )} ) ,$
\begin{align}\label{pisa1}
{\rm dist} \Big ( z - e^{-\theta} s , \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) \geq \frac{1}{2} s \sin(\nu/m) .
\end{align}
Now we argue as in \eqref{coco4} and obtain, for $ s \geq \rho_{n+1} $,
\begin{align}\label{pisa3}
|\lambda^{(n)}_i-(z-e^{-\theta}s)| \geq {\rm dist} \Big ( z - e^{- \theta } s, \mathcal{C}_m( \lambda_i^{(n)} ) \Big ) = & {\rm dist} \Big ( z - \frac{1}{10} e^{- \theta } s,
\mathcal{C}_m(\lambda_i^{(n)} + \frac{9}{10} e^{- \theta } s ) \Big )
\\ \notag \geq & {\rm dist} \Big ( z - \frac{1}{10} e^{- \theta } s ,
\mathcal{C}_m( v_i^{(n, n+1)} ) \Big )
\\ \notag \geq & \frac{1}{2 0 } s \sin(\nu/m) .
\end{align}
Eqs. \eqref{coco5}, \eqref{pisa1} and \eqref{pisa3} together with Lemma \ref{lemma:is-normproj} imply Eq. \eqref{haha2}.
\end{proof}
\begin{lemma}\label{coness}
Let $ \mathcal{C}^{(1)} , \mathcal{C}^{(2)}, \mathcal{C}^{(3)} $ be cones in $\mathbb{C}$, such that $C^{(1)} \subsetneqq \mathcal{C}^{(2)} \subsetneqq \mathcal{C}^{(3)} $, of the form \eqref{eq:defcone} - with the same $m$. Assume that
\begin{align}
\max_{ y \in \partial \mathcal{C}^{(2)}}{\rm dist} \Big ( y, \mathcal{C}^{(1)} \Big ) \leq \frac{1}{2}
{\rm dist} \Big ( \mathbb{C} \setminus \mathcal{C}^{(3)}, \mathcal{C}^{(1)} \Big ).
\end{align}
Then, for every $z \notin \mathcal{C}^{(3)}$:
\begin{align}
{\rm dist} \Big ( z , \mathcal{C}^{(2)} \Big ) \geq \frac{1}{2}{\rm dist} \Big ( z , \mathcal{C}^{(1)} \Big ) .
\end{align}
\end{lemma}
\begin{proof}
We take $z \notin \mathcal{C}^{(3)}$, $y \in \partial \mathcal{C}^{(2)}$,
and $x \in \mathcal{C}^{(1)}$ such that $ |y- x| = {\rm dist} \Big ( y, \mathcal{C}^{(1)} \Big ) $. We calculate
\begin{align}
| z - y | \geq |z - x| - |x - y| & = |z- x| - {\rm dist} \Big ( y, \mathcal{C}^{(1)} \Big )
\\ & \geq |z- x| - \frac{1}{2}
{\rm dist} \Big ( \mathbb{C} \setminus \mathcal{C}^{(3)}, \mathcal{C}^{(1)} \Big )
.
\end{align}
Next, we use that
\begin{align}
{\rm dist} \Big ( \mathbb{C} \setminus \mathcal{C}^{(3)}, \mathcal{C}^{(1)} \Big )
\leq |z - x|
\end{align}
to obtain:
\begin{align}
| z - y | \geq \frac{1}{2} |z- x| \geq \frac{1}{2}{\rm dist} \Big ( z , \mathcal{C}^{(1)} \Big ),
\end{align}
and therefore,
\begin{align}
{\rm dist} \Big ( z , \mathcal{C}^{(2)} \Big ) \geq \frac{1}{2}{\rm dist} \Big ( z , \mathcal{C}^{(1)} \Big ) .
\end{align}
\end{proof}
\begin{lemma}\label{mana}
Assume that $ |g| \leq \frac{ \sin(\nu/2 m)^3 \rho }{10^8} $ and $\rho \leq 10^{-3} \sin(\nu/m)e_1 $ and that for all
$ z \in B^{(1)}_i \setminus \mathcal{C}_m( v_i^{(n)} ) $
\begin{align}\label{difim1}
\Big \| \frac{1}{ H^{(n), \theta} - z} \overline{P^{(n)}_i} \Big \| \leq \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n)} ) \Big ) },
\end{align}
then $ \Big ( B^{(1)}_i \setminus \mathcal{C}_{m}(v_i^{( n+1)})
\Big ) \setminus \{ \lambda_i^{(n + 1)} \} $
is contained in the resolvent set of $ H^{(n+1), \theta} $ and
\begin{align}\label{difim2}
\Big \| \frac{ 1 }{ H^{(n+ 1), \theta} - z } \overline{P_i^{(n+1)}}\Big \| \leq & \frac{ 10^5}{ \sin(\nu/m)^2 } \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) . },
\end{align}
for every $ z \in B^{(1)}_i \setminus \mathcal{C}_{m}(v_i^{( n+1)}) $.
Moreover, assuming that $ \boldsymbol{C} \geq \frac{10^5}{ \sin(\nu/m)^2 }$,
for all
$ z \in B^{(1)}_i \setminus \mathcal{C}_m( v_i^{(n + 1)} ) $
\begin{align}\label{difim3}
\Big \| \frac{1}{ H^{(n + 1), \theta} - z} \overline{P^{(n)}_i} \Big \| \leq \boldsymbol{C}^{n+2} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n + 1)} ) \Big ) }.
\end{align}
\end{lemma}
\begin{proof}
Eq.\ \eqref{difim3} is a consequence of \eqref{difim2} and \eqref{ter1} together with $\boldsymbol{C} \geq \frac{10^5}{ \sin(\nu/m)^2 } $ .
We fix the cones:
\begin{align}
\mathcal{C}^{(1)} & = \mathcal{C}_m(v_i^{(n, n+1)}), \hspace{.5cm}
\mathcal{C}^{(2)} = \mathcal{C}_m(v_i^{(n, n+1)} - \rho_{n+ 1} e^{-i\nu}),
\hspace{.5cm} \\ \notag
\mathcal{C}^{(3)} & = \mathcal{C}_m(v_i^{(n, n+1)} - 2 \frac{1}{\sin(\nu/m)} \rho_{n+ 1} e^{-i\nu}).
\end{align}
Note that the cones we just defined fulfill the hypothesis of Lemma \ref{coness}. They
satisfy the following properties (see Lemma \ref{coness}). For all $z \notin C^{(3)}$ and for all $s \geq 0$:
\begin{align}\label{perra0}
\lambda_i^{(n)} \in \mathcal{C}^{(2)}, \hspace{.5cm} |(z - s e^{-\theta} ) - \lambda_i^{(n)}| \geq
{\rm dist}( (z - s e^{-\theta} ) , \mathcal{C}^{(2)}) \geq {\rm dist}( z , \mathcal{C}^{(2)}) \geq \frac{ 1 }{2 } {\rm dist}(z, \mathcal{C}^{(1)}),
\end{align}
where we use that $ z - s e^{-\theta} \notin C^{(3)} $.
We define $ z_1 = x^{(1)}_1 + i x^{(1)}_2 $ ($x^{(1)}_1, x^{(1)}_2 \in \mathbb{R} $) to be the point in the intersection of $ v_{i}^{(n, n+ 1)} + \frac{1}{100}
\rho_{n+1} e^{-i\nu} + \mathbb{R} $ with $ \partial \mathcal{C}^{(1)}$ with smaller $x^{(1)}_1$, and similarly, $z_2 =x^{(2)}_1 + i x^{(2)}_2 $ the point in the intersection of $ v_{i}^{(n, n+ 1)} + \frac{1}{100}
\rho_{n+1} e^{-i\nu} + i \mathbb{R} $ with $\mathcal{C}^{(1)}$ with bigger $x^{(2)}_2$ . We recall that
\begin{align}\label{perra1}
v_{i}^{(n, n+ 1)} + \frac{1}{100}
\rho_{n+1} e^{-i\nu} = \lambda_i^{(n)} + \frac{2}{5}
\rho_{n+1} e^{-i\nu} \in M^{(n)}_i,
\end{align}
see Definition \ref{def:regionM}, and therefore,
\begin{align}\label{perra2}
z_1, z_2 \in M^{(n)}_i
\end{align}
(the factor $\frac{1}{100}$ is chosen for this reason).
Now, we set
\begin{align}\label{perra3}
\mathcal{U} : = \Big ( \overline{\mathcal{C}^{(3)} \setminus \mathcal{C}^{(1)}} \Big ) \cap
\bigcup_{t \in [0,1]} \big \{ t z_1 + (1- t)z_2 + e^{-i \nu} \mathbb{R} \big \}.
\end{align}
Our restrictions on $\rho$ together with \eqref{perra1} and \eqref{perra2} imply that
\begin{align}\label{perra4}
\mathcal{U} \subset M^{(n)}_i .
\end{align}
It follow from the particular considered geometry at hand that the distance between the boundary of
$ \mathcal{U} $ and $\lambda_i^{(n)}$ is bigger or equal than the distance between the point $ z_2 $ and the line $v_i^{(n, n+1)} + \mathbb{R} e^{- i \nu} $, which equals $\tan( \nu/m) \iota $, where $ \iota $ is the distance between $ v_i^{(n, n+1)} $
and the intersection of the line $ z_2 + i e^{-i \nu}\mathbb{R} $ with $v_i^{(n, n+1)} + \mathbb{R} e^{- i \nu} $. Then $\iota$ is bigger or equal to the the distance between $ v_i^{(n, n+1)} $ and the intersection of the line $ \tilde z_2 + i e^{-i \nu}\mathbb{R} $ with $v_i^{(n, n+1)} + \mathbb{R} e^{- i \nu} $, where $ \tilde z_2 $
is the intersection of $ z_2 + i \mathbb{R} $ with $ v_i^{(n, n+1)} + \mathbb{R} $. Then, $\iota \geq \frac{1}{100}\rho_{n+1} \cos(\nu)\cos(\nu) $. We obtain that
\begin{align}\label{distu}
{\rm dist}(\partial \mathcal U, \lambda_i^{(n)} ) \geq \frac{1}{100}\rho_{n+1} \cos(\nu)\cos(\nu) \tan(\nu/m) \geq \frac{1}{200} \sin(\nu/m) \rho_{n+1},
\end{align}
where we use that $\theta \in \mathcal S $.
For every $ z \in \overline{ \mathcal{C}^{(3)} \setminus \mathcal{C}^{(1)} \setminus \mathcal{U} }$ and $s \geq 0$, we have that
\begin{align}\label{perra5}
| \lambda_{i}^{(n)} - (z - s e^{-\theta} ) | \geq \frac{1}{200} \sin(\nu/m)\rho_{n+1}
\end{align}
and
\begin{equation}\label{perra6}
{\rm dist}( z , \mathcal{C}^{(1)} ) \leq 2 \frac{ 1}{ \sin(\nu/m) }\rho_{n+1}.
\end{equation}
It follows form \eqref{perra5}, \eqref{perra6} together with \eqref{perra0} that
\begin{align}\label{perra7}
\frac{ {\rm dist}( z , \mathcal{C}^{(1)} ) }{| \lambda_{i}^{(n)} - (z - s e^{-\theta} ) |} \leq 400 \frac{1}{ \sin(\nu/m)^2 },
\end{align}
for every $ z \in \Big (B_i^{(1)} \setminus \mathcal C^{(1)} \Big ) \setminus \mathcal{U} $.
This implies, we also use Lemma \ref{mana0} and the spectral theorem that
(actually we only need $s = 0 $ above),
\begin{align}\label{perra8}
\Big \| \frac{1}{ H^{(n), \theta} + H_{f}^{(n, n+1), \theta} - z} \Big \| \leq \frac{ 10^4}{ \sin(\nu/m)^2 } \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) } ,
\end{align}
and for every positive number $r$
\begin{align}\label{perra9}
\Big \| \frac{1}{ H^{(n), \theta} + H_{f}^{(n, n+1), \theta} - z}
( H_{f}^{(n, n+1)} + r) \Big \|
\leq & \frac{ 100}{\sin(\nu/m)} \boldsymbol{C}^{n+1} \\ \notag & + \frac{ 10^4}{ \sin(\nu/m)^2 } \boldsymbol{C}^{n+1} \frac{ r }{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) }
,
\end{align}
where we use that $ H_f^{(n, n+1)} P_i^{(n, n+1)} = 0 $,
for every $ z \in \Big (B^{(1)}_i \setminus \mathcal C^{(1)} \Big ) \setminus \mathcal{U} $. Choosing $r = \rho_{n+1} $, and additionally, $z \notin
\mathcal{C}_m(v_i^{(n+1)}) $, we get from \eqref{ladist} and \eqref{perra9} that
\begin{align}\label{perra10}
\Big \| \frac{1}{ H^{(n), \theta} + H_{f}^{(n, n+1), \theta} - z}
( H_{f}^{(n, n+1)} + r) \Big \|
\leq & \boldsymbol{C}^{n+1} \frac{10^6}{\sin(\nu/2m)^3 }
.
\end{align}
We observe that
\begin{align}
\label{difi1}
&\norm{V^{(n,n+1) ,\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}}
\leq \norm{V^{(n,n+1) ,\theta}\frac{1}{H_f^{(n,n+1)}+r}}
\notag \\
& \hspace{6.4cm} \times \norm{\left(H_f^{(n,n+1)}+r \right)\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z}} .
\end{align}
Then, we have (see also Eq.\ \eqref{eq:keyest22'})
\begin{align}
\label{difi4}
&\big \| g V^{(n,n+1) ,\theta}\frac{1}{H^{(n),\theta}+H_f^{(n,n+1),\theta}-z} \Big \|
\leq |g| \frac{10^6}{\sin(\nu/2m)^3 } \boldsymbol{C}^{n+1}\norm{V^{(n,n+1) ,\theta}\frac{1}{H_f^{(n,n+1)}+r}}
\notag \\
&\leq |g| \frac{10^6}{\sin(\nu/2m)^3 } \boldsymbol{C}^{n+1} |e^{ - \theta (1+ \mu)} | \sqrt{4\pi} \left( \frac{\rho_n}{r} +2\sqrt{ \frac{\rho_n}{r}}\right)\rho_n^{\mu}
\leq \frac{ 10^8 |g| }{ 2 \sin(\nu/2 m)^3 \rho } \boldsymbol{C}^{n+1} \rho_n^{\mu}
\leq \frac{1}{2},
\end{align}
because Definition \ref{sequence} implies that $ \boldsymbol{C}^{n+1} \rho_n^{\mu} \leq 1 $ (we use as well our restrictions in $|g|$).
Eq.\ \eqref{difi4} and a Neumann series argument implies that
$ \Big (B_i^{(1)} \setminus \mathcal{C}_m(v_i^{(n+1)}) \Big ) \setminus \mathcal{U} $ is contained in the
resolvent set of $ H^{(n+1), \theta} $ and for all $z$ in this set (see also \eqref{perra8})
\begin{align}\label{difi5}
\Big \| \frac{1}{H^{(n + 1), \theta} - z} \Big \| \leq & 2 \frac{ 10^4}{ \sin(\nu/m)^2 } \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) }.
\end{align}
Lemma \ref{prop:is-resest-fin} ensures that $\lambda^{(n+1)}_i$ is the only spectral point of $H^{(n+ 1), \theta} $ in $ M^{(n)}_i $. Hence, the function
\begin{align}\label{difi6}
\mathcal U \ni z \mapsto \frac{ 1 }{ H^{(n+ 1), \theta} - z } \overline{P_i^{(n+1)}}
\end{align}
is analytic. The maximum modulus principle implies that it attains its maximum on the boundary of $\mathcal U$, then we have (see Definition \ref{C} and Lemma \ref{prop:is-resest-fin})
\begin{align} \label{difi7}
\Big \| \frac{ 1 }{ H^{(n+ 1), \theta} - z } \overline{P_i^{(n+1)}}\Big \| \leq &
\boldsymbol{C}^{n+1} \frac{1}{\rho_{n+1}}
\end{align}
for every $z \in \mathcal{U}$.
Next, notice that, for $z \in \mathcal{U},$ $ {\rm dist} \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) \leq\frac{2}{\sin(\nu/m)} \rho_{n+1} $.
Then, we obtain
\begin{align} \label{difi8}
\Big \| \frac{ 1 }{ H^{(n+ 1), \theta} - z } \overline{P_i^{(n+1)}}\Big \| \leq &
\frac{1}{\sin(\nu/m)} \boldsymbol{C}^{n+1} \frac{2}{ {\rm dist} \Big ( z, \mathcal{C}_m( v_i^{(n, n+1 )} ) \Big ) } \\ \notag \leq & \frac{1}{\sin(\nu/m)} \boldsymbol{C}^{n+1} \frac{2}{ {\rm dist} \Big ( z, \mathcal{C}_m( v_i^{( n+1 )} ) \Big ) },
\end{align}
Eqs.\ \eqref{difi5} and \eqref{difi8} together with Lemma \ref{lemma:is-normproj} imply the desired result.
\end{proof}
The next theorem is proved inductively using Corollary \ref{coro:ib-res} and Lemma \ref{mana}. This is the main theorem of the present subsection.
\begin{theorem}\label{ResPrinc}
Assume that $ |g| \leq \frac{ \sin(\nu/2 m)^3 \rho }{10^8} $, $\rho \leq 10^{-3} \sin(\nu/m) e_1 $ and $ \boldsymbol{C} \geq \frac{10^5}{ \sin(\nu/m)^2 } $. Then, for all $n\in\mathbb{N}$
and for all
$ z \in B^{(1)}_i \setminus \mathcal{C}_m( v_i^{(n )} ) $:
\begin{align}\label{difim3prima}
\Big \| \frac{1}{ H^{(n ), \theta} - z} \overline{P^{(n)}_i} \Big \| \leq \boldsymbol{C}^{n+1} \frac{1}{\rm dist \Big ( z, \mathcal{C}_m( v_i^{(n )} ) \Big ) }.
\end{align}
\end{theorem}
\subsection{Resolvent estimates}
In this section we assume, in addition to Definitions \ref{C} \ref{sequence} and \ref{gzero} (and $ \boldsymbol{C} \geq \boldsymbol{D} \sin(\nu/m)^{-1} $), that
\begin{align}\label{ggg}
|g| \leq \frac{ \sin(\nu/2 m)^3 \rho }{10^8}, \hspace{1cm} \rho \leq 10^{-3} \sin(\nu/m), \hspace{1cm} \boldsymbol{C} \geq \frac{10^5}{ \sin(\nu/m)^2 } .
\end{align}
\begin{lemma}
\label{mmta}
Let $z\notin \mathcal{C}_m(\lambda_i^{(n)}) $ and $ 0 \leq r \leq |z-
\lambda_i^{(n)}| $, $s \geq 0$. It follows that
\begin{align}\label{nose1}
\Big | \frac{ s+r}{ {\rm dist } \Big ( z - e^{-\theta}s, \mathcal{C}_m(\lambda_i^{(n)}) \Big )} \Big |
& \leq 2 \frac{ 1 }{ \sin (\nu/m) }
+ \frac{r}{ {\rm dist } \Big ( z , \mathcal{C}_m(\lambda_i^{(n)}) \Big ) }
, \\ \notag
\Big | \frac{ s+r}{ e^{-\theta} s -\left(z-
\lambda_i^{(n)}\right)} \Big |
& \leq \frac{ 6 }{ \sin (\nu/m ) }.
\end{align}
\end{lemma}
\begin{proof}
We use coordinates in $\mathbb{C} \equiv \mathbb{R}^2$ with origin at $ \lambda^{(n)}_{i} $, the first coordinate axis with direction $ e^{-i \nu} $ and the second coordinate axis with direction $ i e^{-i \nu} $. Notice that
for every point $ z = \lambda^{(n)}_{i} + \xi_1 e^{- i\nu} + \xi_2 i e^{- i\nu} \notin \mathcal{C}_m(\lambda_i^{(n)}) $ and every $ s \geq 0 $, the following facts are implied by the considered geometry:
\begin{align}
\xi_1 \leq 0 & \Longrightarrow \Big | \lambda_i^{(n)} - ( z- s e^{- \theta} )
\Big | \geq | \lambda_i^{(n)} -z \Big |, \label{pit} \\
\xi_1 > 0 & \Longrightarrow | \xi_2 | \geq | \xi_1 | \tan (\nu/ m ) . \label{pito}
\end{align}
Eq.\ \eqref{pito} implies that for $ \xi_1 > 0 $
\begin{align} \label{pitos}
| z - \lambda_i^{(n)} | \leq |\xi_2| \sqrt{ 1 + \tan(\nu / m)^{-2}},
\end{align}
and because $ | \lambda_i^{(n)} - ( z- s e^{- \theta} )
\Big | \geq |\xi_2| $, we obtain that (we also use \eqref{pit})
\begin{align}\label{pitosn}
\frac{2}{ \sin(\nu /m)} | \lambda_i^{(n)} - ( z- s e^{- \theta} )
\Big | \geq | \lambda_i^{(n)} -z \Big |,
\end{align}
for every $z\notin \mathcal{C}_m(\lambda_i^{(n)}) $ and every $ s \geq 0 $.
Take $z\notin \mathcal{C}_m(\lambda_i^{(n)}) $ to obtain
\begin{align}
\label{eq:helplemma111}
\left| \frac{s+r}{e^{-\theta}s+\lambda_i^{(n)}-z} \right|
&\leq |e^{\theta}| + |e^{\theta}| \left| \frac{e^{-\theta}r-\lambda_i^{(n)}+z}{e^{-\theta}s+\lambda_i^{(n)}-z} \right|
\leq |e^{\theta}| + (1 + |e^{\theta}| ) \frac{\left| z- \lambda_i^{(n)}\right| }{\left| e^{-\theta}s+\lambda_i^{(n)}-z\right| }
\notag \\
&\leq |e^{\theta}| + \frac{ 2 (1 + |e^{\theta}| ) }{ \sin ( \nu/ m )}\leq \frac{ 6 }{ \sin ( \nu/ m)},
\end{align}
which proves the second inequality in \eqref{nose1}. The first inequality of the claim is again ensured thanks to our considered geometry that implies:
\begin{align}
{\rm dist } \Big ( z - e^{-\theta}s, \mathcal{C}_m(\lambda_i^{(n)}) \Big ) \geq \max \Big [ \frac{\sin(\nu/ m)}{2} s , {\rm dist } \Big ( z , \mathcal{C}_m(\lambda_i^{(n)}) \Big ) \Big ] ;
\end{align}
recall that $\theta \in \mathcal{S}$.
\end{proof}
\begin{lemma} \label{EL}
\label{mmtaa}
For $i=0,1$, the set $ B^{(1)}_i \setminus \mathcal C_m(\lambda_i^{(n)} ) $
is contained in the resolvent set of
$ \tilde H^{(n), \theta} $ and for all $z $ this set:
\begin{align}
\norm{ \frac{1}{\tilde H^{(n),\theta}-z}}
\leq 4
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z, \mathcal{C}_m(\lambda_i^{(n)}) )}.
\end{align}
\end{lemma}
\begin{proof}
The spectral theorem, Lemma \ref{lemma:resestinAprima}, Definitions \ref{compile} and \ref{C}, and Theorem \ref{ResPrinc} imply that (we also use Lemma \ref{lemma:is-normproj}, which is valid for every $n$ because Theorem
\ref{thm:ind} is proved above)
\begin{align}
\norm{ \frac{1}{\tilde H^{(n),\theta}-z}}
= & \sup_{ r \geq 0} \norm{ \frac{1}{ H^{(n),\theta} + e^{-\theta } r -z}
\big ( P_i^{(n)} + \overline{P_i^{(n)}} \big) } \\ \notag \leq &
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z- e^{-\theta } r, \mathcal{C}_m(v_i^{(n)}) )} + 3 \frac{1}{ | \lambda_i^{(n)} + e^{-\theta } y - z | }
\\ \notag \leq & 4
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z- e^{-\theta } r, \mathcal{C}_m(\lambda_i^{(n)}) )}
\leq 4
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z, \mathcal{C}_m(\lambda_i^{(n)}) )},
\end{align}
were we use that $\lambda_i^{(n)} \in \mathcal{C}_m(\lambda_i^{(n)}) $ and the geometrical fact that $ {\rm dist} ( z- e^{-\theta } r, \mathcal{C}_m(\lambda_i^{(n)}) ) \geq {\rm dist} ( z, \mathcal{C}_m(\lambda_i^{(n)}) ) $.
\end{proof}
\begin{lemma}
\label{lemma:is-keyesttilde}
For every $ z \in B^{(1)}_i \setminus \mathcal C_m(\lambda_i^{(n)} - e^{-i \nu} \rho_n^{1+ \mu/4} ) $
the following inequality holds true
\begin{align}
\norm{V^{(n,\infty),\theta}\frac{1}{\tilde H^{(n),\theta}-z}}
\leq \frac{10^3}{\sin ( \nu/m)^2 }\bold C^{n+1} \rho_n^{\frac{3\mu}{4}} .
\end{align}
\end{lemma}
\begin{proof}
We compute (see Remark \ref{R}):
\begin{align}
\label{eq:is-keyesttilde0}
\norm{V^{(n,\infty),\theta}\frac{1}{\tilde H^{(n),\theta}-z}}
\leq & \norm{V^{(n,\infty),\theta}\frac{1}{ H^{(n,\infty )}_f+r}}
\norm{( H_f^{(n, \infty )} + r ) \frac{1}{\tilde H^{(n),\theta}-z}}
\end{align}
where we take
\begin{align}
r = {\rm dist } \Big ( \mathbb C \setminus \mathcal{C}_m(\lambda_i^{(n)} - e^{- i \nu} \rho_n^{1+ \mu/4} ) , \mathcal{C}_m(\lambda_i^{(n)}) \Big )
= \sin(\nu/m) \rho_n^{1 + \mu/4} \ \leq \left| z- \lambda_i^{(n)}\right|.
\end{align}
As in Eq.\ \eqref{eq:keyest22'} we obtain
\begin{align}
\label{teto1}
\norm{V^{(n,\infty) ,\theta}\frac{1}{H_f^{(n, \infty )}+r}}
&\leq |e^{ - \theta (1+ \mu)} | \sqrt{4\pi} \left( \frac{\rho_n}{r} +2\sqrt{ \frac{\rho_n}{r}}\right)\rho_n^{\mu}
\leq \frac{50}{\sin(\nu/m)} \rho_n^{ \frac{3\mu}{4}} .
\end{align}
The spectral theorem and Theorem \ref{ResPrinc} and Lemma \ref{mmta} imply that
\begin{align}\label{teto2}
\norm{\left( H_f^{(n, \infty ),\theta} +r \right)\frac{1}{\tilde H^{(n),\theta}-z}} \leq & 4
\boldsymbol{ C}^{n+1} \sup_{s \geq 0}
\Big | \frac{ s+r}{ {\rm dist } \Big ( z - e^{-\theta}s, \mathcal{C}_m(\lambda_i^{(n)}) \Big )} \Big |
\\ \notag \leq & 4
\boldsymbol{ C}^{n+1} \left[ \frac{ 2 }{ \sin (\nu/m) }
+ \frac{r}{ {\rm dist } \Big ( z , \mathcal{C}_m(\lambda_i^{(n)}) \Big ) }
\right]
\\ \notag \leq & 4
\boldsymbol{ C}^{n+1} \left[ \frac{ 2 }{ \sin (\nu/m) }
+ \frac{r}{ \sin(\nu/m) \rho_n^{1+ \mu/4} } \right] \leq \frac{ 12 \boldsymbol{ C}^{n+1} }{ \sin (\nu/m) } .
\end{align}
We conclude the desired result by \eqref{eq:is-keyesttilde0} together with \eqref{teto1} and \eqref{teto2}.
\end{proof}
\begin{theorem} \label{resuno}
The set $ \in B^{(1)}_i \setminus \mathcal C_m(\lambda_i^{(n)} - e^{-i \nu} \rho_n^{1+ \mu/4} ) $
is contained in the resolvent set of
$ H^{ \theta} $ and for all $z\in B^{(1)}_i \setminus \mathcal C_m(\lambda_i^{(n)} - e^{-i \nu} \rho_n^{1+ \mu/4} ) $:
\begin{align}
\norm{ \frac{1}{ H^{\theta}-z}}
\leq 8
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z, \mathcal{C}_m(\lambda_i^{(n)}) )},
\end{align}
and
\begin{align}
\norm{ \frac{1}{ H^{\theta}-z} - \frac{1}{ \tilde H^{(n),\theta}-z} }
\leq | g | \frac{ 10^5}{\sin ( \nu/m)^2 }\bold C^{2n+2}
\rho_n^{\frac{3\mu}{4}} \frac{1}{ {\rm dist} ( z, \mathcal{C}_m(\lambda_i^{(n)}) )}.
\end{align}
\end{theorem}
\begin{proof}
The result is a consequence of Neumann series and Lemmas \ref{EL} and \ref{lemma:is-keyesttilde} and Eq.\ \eqref{ggg}. Notice that our assumptions on $\boldsymbol{C}$ in Definition \ref{C} imply that $ \bold C^{n+1} \rho_n^{\frac{3\mu}{4}} \leq 1 $.
\end{proof}
\begin{theorem}\label{resdos}
For every $n \in \mathbb{N}$ and $i\in\{0,1\}$,
\begin{align} \label{lal}
\mathcal{C}_m ( \lambda_i^{(n+ 1)} - \rho_{n+ 1}^{1+ \mu/4} e^{-i \nu} ) \subset \mathcal{C}_m ( \lambda_i^{(n)} - \rho_n^{1+ \mu/4} e^{-i \nu} ) \subset \mathcal{C}_m ( \lambda_i - 2 \rho_n^{1+ \mu/4} e^{-i \nu}) ,
\end{align}
and thus, $ B^{(1)}_i \setminus \mathcal{C}_m ( \lambda_i - 2 \rho_n^{1+ \mu/4} e^{-i \nu } ) $ is contained in the resolvent set of $H^{\theta}$ (see Theorem \ref{resuno}).
Moreover, $ B^{(1)}_i\setminus \mathcal C_m \left(\lambda_i\right) $ is contained in the resolvent set of $H^{\theta}$. Additionally, the following estimate holds true:
\begin{align}
\norm{ \frac{1}{ H^{\theta}-z}}
\leq 16
\boldsymbol{ C}^{n+1} \frac{1}{ {\rm dist} ( z, \mathcal{C}_m(\lambda_i ) )} , \hspace{.5cm} \forall z \in \mathcal{C}_m ( \lambda_i - 2 \rho_n^{1+ \mu/4} e^{- i \nu} ) . \label{segunda}
\end{align}
\end{theorem}
\begin{proof}
It follows from Eq.\ \eqref{ter2} that
\begin{align}\label{ter2pe}
| \lambda_i^{(n+1)} - \lambda_i^{(n)} | \leq |g| \rho_{n}^{1+ \mu/2}
\end{align}
holds true.
We write, for $ \xi_1, \xi_2 \in \mathbb{R} $,
\begin{align}
\lambda_i^{(n+1)} - \rho_{n+ 1}^{1+ \mu/4} e^{-i \nu} = \lambda_{i}^{(n)} -
\rho_{n}^{1+ \mu/4} e^{-i \nu} + \xi_1 e^{-i \nu} + \xi_2 i e ^{-i\nu}.
\end{align}
Eq.\ \eqref{ter2pe} implies that
\begin{align}\label{ter3pe}
|\xi_2| \leq |g| \rho_{n}^{1+ \mu/2}, \hspace{2cm} \xi_1 \geq \rho_{n}^{1+ \mu/4} - \rho_{n + 1}^{1+ \mu/4} - |g| \rho_{n}^{1+ \mu/2} > \frac{1}{2} \rho_{n}^{1+ \mu/4} ,
\end{align}
see Definition \ref{sequence} and Definition \ref{gzero} (or \eqref{ggg} - notice that
$ |g| \rho_{n}^{1+ \mu/2} \leq \frac{|g|}{\rho} \rho_{n + 1}^{1+ \mu/4} \rho_0^{\mu/4} $).
To prove the first assertion in \eqref{lal} it is enough to prove that $ \lambda^{(n+1)}_i - \rho_{n+1}^{1+ \mu/4} e^{-i\nu} \in \mathcal{C}_m ( \lambda^{(n)}_i - \rho_{n}^{1+\mu/4} e^{-i\nu} ) $. Note that since $ | g | \leq \frac{1}{2} \sin(\nu / m) \leq \frac{1}{2} \tan(\nu / m) $ (which is verified by \eqref{ggg}), we have
\begin{align}\label{ter4pe}
|\xi_2| / \xi_1 < \tan(\nu / m).
\end{align}
This proves the first assertion in \eqref{lal}.
The first part of Eq.\ \eqref{lal} implies that, for all $n$,
\begin{align}\label{ter5pe}
\mathbb{C} \setminus \mathcal{C}_m ( \lambda_i^{(n)} - \rho_{n}^{1+ \mu/4} e^{-i \nu} ) \subset \mathbb{C} \setminus \mathcal{C}_m ( \lambda_i^{(n + 1)} - \rho_{n+1}^{1+ \mu/4}e^{-i \nu})
\end{align}
and
\begin{align}\label{ter6pe}
\bigcup_n \mathbb{C} \setminus \mathcal{C}_m ( \lambda_i^{(n)} - \rho_{n}^{1+ \mu/4}e^{-i \nu} ) = \mathbb{C} \setminus \mathcal{C}_m ( \lambda_i)
\end{align}
belongs to the resolvent set of $H^{\theta}$, see Theorem \ref{resuno}.
In a similar fashion as above we prove that
\begin{align} \label{lal1}
\mathcal{C}_m ( \lambda_i^{(n)} - \rho_{n}^{1+ \mu/4}e^{-i \nu} ) \subset \mathcal{C}_m ( \lambda_i - 2 \rho_n^{1+ \mu/4} e^{-i \nu}) ,
\end{align}
using \eqref{ter0}. For every $ z \notin \mathcal{C}_m ( \lambda_i - 2 \rho_n^{1+ \mu/4}e^{-i \nu})$ and $a \in \mathcal{C}_m ( \lambda^{(n)}_i ) $ we know that (see \eqref{ter0})
\begin{align}\label{casa1}
{\rm dist}(z, \mathcal{C}_m ( \lambda_i ) )\leq &
{\rm dist}(z, a ) +
{\rm dist}( a , \mathcal{C}_m ( \lambda_i ) )
\leq {\rm dist}(z, a ) + 2 |g| \rho_n^{1+ \mu/2},
\end{align}
and hence, we obtain (see Eq. \eqref{ggg})
\begin{align}\label{casa2}
{\rm dist}(z, \mathcal{C}_m ( \lambda_i ) )\leq {\rm dist}(z, \mathcal{C}_m ( \lambda^{(n)}_i ) ) + \sin(\nu/m) \rho_n^{1+ \mu/4}.
\end{align}
Moreover (see \eqref{lal1}),
\begin{align}\label{casa3}
{\rm dist}(z, \mathcal{C}_m ( \lambda^{(n)}_i ) ) \geq
{\rm dist}\Big ( \mathbb{C} \setminus \mathcal{C}_m ( \lambda_i^{(n)} - \rho_{n}^{1+ \mu/4} e^{-i \nu} ) , \mathcal{C}_m ( \lambda^{(n)}_i ) \Big ) \geq \sin(\nu/m) \rho_n^{1+ \mu/4} .
\end{align}
Then, it follows that
\begin{align}\label{casa2prima}
\frac{{\rm dist}(z, \mathcal{C}_m ( \lambda_i ) )}{ {\rm dist}(z, \mathcal{C}_m ( \lambda^{(n)}_i ) ) }\leq \frac{ {\rm dist}(z, \mathcal{C}_m ( \lambda^{(n)}_i ) )}{ {\rm dist}(z, \mathcal{C}_m ( \lambda^{(n)}_i ) ) } + \frac{ \sin(\nu/m) \rho_n^{1+ \mu/4}}{ \sin(\nu/m) \rho_n^{1+ \mu/4} } \leq 2.
\end{align}
This and Theorem \ref{resuno} implies Eq.\ \eqref{segunda}.
\end{proof}
We end this section with a remark about the convergence rate of the projections
$ P^{n}_i\otimes P_{\Omega^{(n,\infty)}} $ that will be used in a forthcoming paper.
\begin{remark}\label{scattering}
It follows from Theorem \ref{thm:ind}, Property ($\mathcal P 3$), that
\begin{align}
\label{eq:P3prima}
\norm{P_i - P^{(n)}_i\otimes P_{\Omega^{(n,\infty)}}} \leq 2 \frac{|g|}{\rho} \frac{1}{2^n} \rho_n^{\mu /2}.
\end{align}
This is a consequence of a geometric series argument and Definition \ref{sequence}, since it implies that
\begin{align}
\boldsymbol{C}^{2(n+1)+2} \rho_{n}^{\mu} = (\boldsymbol{C}^8 \rho_0^{\mu})^{1/2}
(\boldsymbol{C}^4 \rho^{\mu})^{n/2} \rho_n^{\mu /2} \leq \frac{1}{2^n} \rho_n^{\mu /2}.
\end{align}
\end{remark}
\section{Proof of Proposition~\ref{thm:res}}
\label{sec:proof-res}
We define the sequence of vectors (see Remark \ref{R})
\begin{align}
\Psi^{(n)}_{\lambda_i}:=P^{(n)}_i \varphi_i\otimes \Omega , \quad n\in\mathbb{N} .
\end{align}
Due to the Property ($\mathcal P 3$) in Theorem \ref{thm:ind}, we know that the sequence above converges to the non-zero limit $\Psi_{\lambda_i}:=P_i\varphi_i\otimes \Omega\neq 0$ (see the discussion above Eq.\ \eqref{eq:p1proof}).
Note that (see Remark \ref{R})
\begin{align}
H^{\theta}=H^{(n),\theta}+H_f^{(n,\infty),\theta}+gV^{(n,\infty),\theta}
= \tilde H^{(n),\theta} + gV^{(n,\infty),\theta}
\end{align}
and set $z=\lambda^{(n)}_i - 10 \rho_n e^{-i\nu} $. Then,
\begin{align}
H^{\theta}\Psi^{(n)}_{\lambda_i}=\lambda^{(n)}_i\Psi^{(n)}_{\lambda_i} +g (\lambda^{(n)}_i-z)V^{(n,\infty),\theta}\frac{1}{ \tilde H^{(n),\theta} -z}\Psi^{(n)}_{\lambda_i}.
\end{align}
Lemma \ref{lemma:is-keyesttilde} implies that $V^{(n,\infty),\theta}\frac{1}{
\tilde H^{(n),\theta} -z}$ tends to zero as $ n$ tends to infinity. We conclude that
\begin{align}
\lim\limits_{n\to \infty}H^\theta \Psi^{(n)}_{\lambda_i}=\lambda_i\Psi_{\lambda_i} .
\end{align}
As $H^\theta$ is a closed operator, $\Psi_{\lambda_i}$ belongs to its domain and is an eigenvector of $H^\theta$ corresponding to the eigenvalue $\lambda_i$. Furthermore, as $P_i$ is rank-one, $\Psi_{\lambda_i}$ spans its range.
\section{Analyticity}\label{analyticity}
In this section we assume that Definitions \ref{C}, \ref{sequence}, \ref{gzero} hold true. We recall that we use the symbol $c$ to represent any generic (indeterminate) constant
that does not depend on $n$, $g$, $\rho$, $\rho_0$ and dilation parameters (here do not only use $\theta$, but also $\eta$ and $\lambda$). In this section we prove Theorems \ref{thm:anaP} and \ref{thm:anaPg}.
\begin{lemma}\label{ana1}
For every $r > 0$ and $\lambda, \eta \in D(0, \pi/16) $ (and every $n \in \mathbb{N}$):
\begin{align}\label{lis1}
\Big \| \frac{ H_0^{(n), \lambda} + r }{ H_0^{(n), \eta } + r } \Big \| \leq & 10.
\end{align}
Moreover, for large enough $r$ (independent of $n$, $g$, $\rho$, $\rho_0$ and $\eta$ and
$\lambda$) and every $z$ in the resolvent set of $ H^{(n), \eta} $:
\begin{align}\label{lis2}
\Big \| H_0^{(n), \lambda} \frac{ 1 }{ H^{(n), \eta} - z } \Big \|
\leq 20 + 20(|z| + r/2 ) \Big \| \frac{1}{ H^{(n), \eta} - z } \Big \|.
\end{align}
\end{lemma}
\begin{proof}
Notice that for every $\eta, \lambda \in D(0, \pi/16) $
\begin{align}
\Big \| \frac{ H_0^{(n), \lambda} + r }{ H_0^{(n), \eta } + r } \Big \| \leq & \sup_{s \geq 0, i \in {0,1}}
\Big | \frac{ e_i + e^{-\lambda }s + r }{ e_i + e^{-\eta }s + r } \Big | .
\end{align}
For every $s \geq 0$ and $ i \in \{0, 1 \} :$
\begin{align}
\Big | \frac{ e_i + e^{-\lambda } s + r }{ e_i + e^{-\eta }s + r } \Big |
\leq & |e^{\eta - \lambda} | + \Big | e^{\eta - \lambda } \frac{ r + e_i }{ e_i + e^{-\eta }s + r } \Big | +
\Big | \frac{ e_i + r }{ e_i + e^{-\eta }s + r } \Big | \leq 10,
\end{align}
where we use that $\eta, \lambda \in D(0, \pi/16) $ and $e_i \geq 0$. This implies \eqref{lis1}.
It follows from Appendix \ref{app:sa} that there is a constant $c$ that does not depend on $n$, $g$, $\rho$, $\rho_0$ and $\eta$ such that for every $r \geq 1$:
\begin{align}
\Big \| V^{(n), \eta} \frac{1}{ (H_0^{(n), 0} + r)^{1/2}} \Big \| \leq c .
\end{align}
In conclusion,
\begin{align}
\Big \| V^{(n), \eta} \frac{1}{ H_0^{(n), 0} + r} \Big \| \leq \frac{c}{r^{1/2}}
\end{align}
holds true.
It follows that there is a constant
$ C_{\eqref{zee}} $ that does not depend on $n$, $g$, $\rho$, $\rho_0$ and $\eta$ such that for every $r \geq 1$:
\begin{align} \label{zee}
\Big \| V^{(n), \eta} \frac{1}{ H_0^{(n), \eta} + r} \Big \| \leq \frac{C_{\eqref{zee}}}{r^{1/2}} .
\end{align}
Take $ \phi $ in the domain of $ H_0^{(n), \eta} $, $z \in \mathbb{C}$ and $ r \geq 4 C_{\eqref{zee}}^2 $. Then we have
(recall that $|g| \leq 1$):
\begin{align}
\| H_0^{(n), \eta} \phi \| \leq & \| ( H^{(n), \eta} - z ) \phi \| +
\| V^{(n), \eta} \phi \| + |z| \| \phi \| \\ \notag \leq &
\| ( H^{(n), \eta} - z ) \phi \| +
(1/2)\| H_0^{(n), \eta} \phi \| + (|z| + r/2 ) \| \phi \|.
\end{align}
Then, we obtain, for $z$ in the resolvent set of $ H^{(n), \eta} $ and $s > 0$ (we take the term
$ (1/2)\| H_0^{(n), \eta} \phi \| $ in the previous equation to the other side and $ \phi $ of the form
$ \frac{1}{ H^{(n), \eta} - z } \psi $):
\begin{align}
\Big \| ( H_0^{(n), \eta} + s) \frac{ 1 }{ H^{(n), \eta} - z } \Big \|
\leq & 2 + 2 (|z| + (r+ 2 s)/2 ) \Big \| \frac{1}{ H^{(n), \eta} - z } \Big \|.
\end{align}
Using \eqref{lis1}, we find
\begin{align}
&\Big \| ( H_0^{(n), \lambda} + s ) \frac{ 1 }{ H^{(n), \eta} - z } \big \| \leq
\Big \| \frac{ H_0^{(n), \lambda} + s }{ H_0^{(n), \eta} + s } ( H_0^{(n), \eta} + s) \frac{ 1 }{ H^{(n), \eta} - z } \Big \|
\notag \\
&\leq 10 \Big \| ( H_0^{(n), \eta} + s) \frac{ 1 }{ H^{(n), \eta} - z } \Big \|
\leq 20 + 2 0(|z| + (r + 2 s)/2 ) \Big \| \frac{1}{ H^{(n), \eta} - z } \Big \|.
\end{align}
Taking the limit $s $ to zero, we arrive at Eq.\ \eqref{lis2}.
\end{proof}
\begin{lemma} \label{pppp}
For every $\lambda, \eta, \theta \in D(0, \pi/16) $ (and every $n \in \mathbb{N}$),
there is a constant $c$ (independent of $n$, $g$, $\rho$, $\rho_0$, $\eta$, $\theta$ and
$\lambda$) such that for every $z$ in the resolvent set of $ H^{(n), \theta} $:
\begin{align}
\label{Mtama}
\Big \| (H^{(n), \eta} - H^{(n), \lambda} ) \frac{1}{ H^{(n), \theta} - z }
\Big \| \leq c (1+ |z|) |\eta - \lambda | \Big ( \Big \| \frac{1}{ H^{(n), \theta} - z }
\Big \| + 1 \Big ) .
\end{align}
\end{lemma}
\begin{proof}
We take a large enough $r > 0$ such that the results of Lemma \ref{ana1} hold true. We calculate
\begin{align}\label{toto1}
\Big \| (H^{(n), \eta} - H^{(n), \lambda} ) \frac{1}{ H^{(n), \theta} - z }
\Big \| \leq \Big \| (H^{(n), \eta} - H^{(n), \lambda} ) \frac{1}{ H_0^{(n), 0} + r } \Big \|
\Big \|(H_0^{(n), 0} + r )
\frac{1}{ H^{(n), \theta} - z }
\Big \|.
\end{align}
Next, we notice that
\begin{align}\label{toto2}
\Big \| (H_0^{(n), \eta} - H_0^{(n), \lambda} ) \frac{1}{ H_0^{(n), 0} + r } \Big \|
= \sup_{s \geq 0, i \in \{1, 2 \} } \Big \| (e^{- \eta } - e^{-\lambda}) s \frac{1}{ e_i + s + r } \Big \| \leq | e^{- \eta } - e^{-\lambda} |.
\end{align}
Using Appendix \ref{app:sa}, we find a constant
$ c $ (independent of $n$, $g$, $\rho$, $\rho_0$ and $\eta$ and
$\lambda$) such that
\begin{align}\label{toto3}
\Big \| (V^{(n), \eta} - V^{(n), \lambda} ) \frac{1}{ H_0^{(n), 0} + r } \Big \|
\leq c |\eta - \lambda |
\end{align}
Eqs.\ \eqref{toto1}-\eqref{toto3}, together with Lemma \ref{ana1}, imply the desired result.
\end{proof}
\begin{definition}
For every $ \theta \in D(0, \pi/16) $, we set $h^{\theta} = \frac{\partial}{\partial \theta} f^{ \theta} $ and
\begin{align} \label{operators444}
\frac{\partial}{ \partial \theta } V^{(n), \theta} := \sigma_1 \otimes
\left(a_n( h^{\overline \theta})+ a_n( h^{ \theta})^* \right)
\end{align}
and (see Remark \ref{R})
\begin{align}
\frac{\partial}{ \partial \theta } H^{(n), \theta} : = - H_f^{(n), \theta} + g\frac{\partial}{ \partial \theta } V^{(n), \theta} .
\end{align}
\end{definition}
\begin{lemma} \label{penu}
For every $\lambda, \eta, \theta \in D(0, \pi/16) $ (and every $n \in \mathbb{N}$),
there is a constant $c$ (independent of $n$, $g$, $\rho$, $\rho_0$, $\eta$, $\theta$ and
$\lambda$) such that for every $z$ in the resolvent set of $ H^{(n), \theta} $:
\begin{align}
\label{Mtamaba}
\Big \| \Big ( \frac{1}{\eta - \lambda} (H^{(n), \eta} - H^{(n), \lambda}) - \frac{\partial}{ \partial \lambda } H^{(n), \lambda}\Big ) \frac{1}{ H^{(n), \theta} - z }
\Big \| \leq c (1+ |z|) |\eta - \lambda | \Big ( \Big \| \frac{1}{ H^{(n), \theta} - z }
\Big \| + 1 \Big ).
\end{align}
\end{lemma}
\begin{proof}
The proof is very similar to the proof of Lemma \ref{pppp}, and therefore, we omit it.
\end{proof}
\begin{lemma}\label{lemsud}
For every $\lambda, \eta, \theta \in D(0, \pi/16) $ (and every $n \in \mathbb{N}$),
there is a constant $c$ (independent of $n$, $g$, $\rho$, $\rho_0$, $\eta$, $\theta$ and
$\lambda$) such that for every $z$ in the resolvent set of both $ H^{(n), \eta} $ and $ H^{(n), \lambda } $:
\begin{align} \label{sud0}
\Big \| \frac{1}{\eta - \lambda} \Big ( \frac{1}{H^{(n), \lambda} - z } & - \frac{1}{H^{(n), \eta} - z } \Big ) - \frac{1}{H^{(n), \lambda} - z } \frac{\partial }{\partial \lambda } H^{(n), \lambda} \frac{1}{H^{(n), \lambda } - z } \Big \|
\\ \notag & \leq c ( 1 + |z|)^2 | \eta - \lambda | \Big ( \Big \| \frac{1}{H^{(n), \lambda } - z } \Big \| + 1 \Big )^2 \Big ( \Big \| \frac{1}{H^{(n), \eta } - z } \Big \| + 1 \Big ) ,
\end{align}
and
\begin{align} \label{sud01}
\Big \| H_0^{(n), \theta} \Big ( \frac{1}{\eta - \lambda} \Big ( \frac{1}{H^{(n), \lambda} - z } & - \frac{1}{H^{(n), \eta} - z } \Big ) - \frac{1}{H^{(n), \lambda} - z } \frac{\partial }{\partial \lambda } H^{(n), \lambda} \frac{1}{H^{(n), \lambda } - z } \Big ) \Big \|
\\ \notag & \leq c ( 1 + |z|)^2 | \eta - \lambda | \Big ( \Big \| \frac{1}{H^{(n), \lambda } - z } \Big \| + 1 \Big )^2 \Big ( \Big \| \frac{1}{H^{(n), \eta } - z } \Big \| + 1 \Big ) ,
\end{align}
\end{lemma}
\begin{proof}
First, we notice that Lemma \ref{pppp} and the resolvent identity imply
\begin{align}\label{sud1}
\Big \| \frac{1}{H^{(n), \lambda} - z } & ( H^{(n), \eta} - H^{(n), \lambda } ) \Big ( \frac{1}{H^{(n), \eta} - z } - \frac{1}{H^{(n), \lambda} - z } \Big ) \Big \| \\ \notag & \leq c ( 1 + |z|)^2 | \eta - \lambda |^2 \Big ( \Big \| \frac{1}{H^{(n), \lambda } - z } \Big \| + 1 \Big )^2 \Big ( \Big \| \frac{1}{H^{(n), \eta } - z } \Big \| + 1 \Big ).
\end{align}
We use the resolvent identity again and also Eq.\ \eqref{sud1} to obtain
\begin{align}\label{sud2}
\Big \| \Big ( \frac{1}{H^{(n), \lambda} - z } & - \frac{1}{H^{(n), \eta} - z } \Big ) - \Big ( \frac{1}{H^{(n), \lambda} - z } ( H^{(n), \eta} - H^{(n), \lambda } ) \frac{1}{H^{(n), \lambda} - z } \Big ) \Big \|
\\ \notag & \leq c ( 1 + |z|)^2 | \eta - \lambda |^2 \Big ( \Big \| \frac{1}{H^{(n), \lambda } - z } \Big \| + 1 \Big )^2 \Big ( \Big \| \frac{1}{H^{(n), \eta } - z } \Big \| + 1 \Big ) .
\end{align}
Then, Eq.\ \eqref{sud0} follows from \eqref{sud2} and Lemma \ref{penu}. The proof of
\eqref{sud01} follows in a similar fashion as the one of \eqref{sud0}, using Lemma \ref{ana1}. Therefore, we omit it.
\end{proof}
\begin{proposition}\label{propana1}
For every $\eta \in D(0, \pi/16) $, the operator valued functions
\begin{align}
\theta \in \mathcal S \mapsto P_i^{(n)}, \hspace{2cm} \theta \in \mathcal S \mapsto
H_0^{(n), \eta} P_i^{(n)}
\end{align}
are analytic.
\end{proposition}
\begin{proof}
The proof is an obvious consequence of Lemma \ref{lemsud} and the formula for the Riesz projections as line integrals in the complex plane.
\end{proof}
\begin{proposition}\label{propana2}
The complex valued function
\begin{align}
\theta \in \mathcal S \mapsto \lambda_i^{(n)}
\end{align}
is analytic.
\end{proposition}
\begin{proof}
We use the formalism of the proof of Proposition \ref{prop:is-p1} and make explicit the dependence of $ P^{(n)}_{i} $ on $\theta$, i.e., $ P^{(n)}_{i} \equiv P^{(n),\theta}_{i} $. We define $\Psi^{(n),\theta}_{i} = P^{(n),\theta}_{i}\varphi_i\otimes \Omega^{(n)} $ (here we use a slightly different notation from proof of Proposition \ref{prop:is-p1}). Notice that
\begin{align}
\lambda_i^{(n)} = \frac{\langle \Psi^{(n), \overline \theta}_{i}, H^{(n), \theta} \Psi^{(n),\theta}_{i} \rangle}
{ \langle \Psi^{(n), \overline \theta}_{i}, \Psi^{(n),\theta}_{i} \rangle },
\end{align}
and that the denominator does not vanish (this follows as in \eqref{eq:p1proof}).
Then, the result is a consequence of Proposition \ref{propana1}, because it implies that the
functions
\begin{align}
\theta \mapsto \Psi^{(n),\theta}_{i}, \hspace{1cm} \theta \mapsto H^{(n), \theta} \Psi^{(n),\theta}_{i} = H^{(n), \theta} \frac{1}{ H_0^{(n), 0} + 1 } \Big ( ( H_0^{(n), 0} + 1 ) P^{(n),\theta}_{i}\varphi_i\otimes \Omega^{(n)} \Big )
\end{align}
are analytic. Notice that the function $ \theta \mapsto H^{(n), \theta} \frac{1}{ H_0^{(n), 0} + 1 } $ is an operator valued analytic function (the proof of this fact is similar to the proof of Lemma \ref{penu}, but much simpler).
\end{proof}
\begin{proposition}\label{propana3}
The maps
\begin{align}
g \in D(0, g_0 ) \mapsto P_i^{(n)}, \qquad g \in D(0, g_0 ) \mapsto \lambda_i^{(n)}
\end{align}
are analytic.
\end{proposition}
\begin{proof}
The proof follows directly from the proofs of Propositions \ref{propana1} and \ref{propana2}. In this case the proof is much simpler because the coupling constant is only present in the interaction term (and the interaction term depends linearly on the coupling constant).
\end{proof}
\begin{theorem}
\label{thm:anaPp}
The functions
\begin{align}
\mathcal S \ni & \theta \mapsto P_i, \hspace{1cm} \mathcal S \ni \theta \mapsto \lambda_i \\ \notag
D(0, g_0) \ni & g \mapsto P_i, \hspace{1cm} D(0, g_0) \ni g \mapsto \lambda_i
\end{align}
are analytic. Moreover, this implies that $ \lambda_i (\theta) \equiv \lambda_i$ is constant for $\theta\in\mathcal S$ (see \eqref{def:setS}).
\end{theorem}
\begin{proof}
Theorem \ref{thm:ind}, Properties $\mathcal P1$ and $ \mathcal P 3 $, imply that
the convergence rates of $ \lambda^{(n)}_i $ to $ \lambda_i $ and
$ P^{(n)}_i\otimes P_{\Omega^{(n, \infty)}} $ to $ P_i $ do
not depend on $ \theta $ and $g$. Then $ \lambda_i $ and $ P_i $ are uniform limits of analytic functions (see Propositions \ref{propana1}, \ref{propana2}, \ref{propana3}). Therefore, they are analytic. That $\lambda_i$ is constant with respect to $\theta$ follows from the fact that it does not depend of the real part of $\theta$
because a change in the real part of $ H^{\theta} $ produces unitarily equivalent Hamiltonians: if $ \theta $ and $ \tilde \theta $ have the same imaginary part, then $ H^{\theta} $ and $ H^{\tilde \theta} $ are unitarily equivalent (thus, isospectral). Both, $\lambda_i(\theta)$ and $ \lambda_i(\tilde \theta) $ are distinguished points in the spectrum because they are the vertex of the same cone (see Theorem \ref{spectralestimates}), we conclude that $\lambda_i(\theta) = \lambda_i(\tilde \theta) $.
\end{proof}
\begin{appendix}
\section{Closedness of H and standard estimates}
\label{app:sa}
In the following we shall use the well-known standard inequality
\begin{align}
\begin{split}
\|a(h)\Psi\|&\leq \|h/\sqrt\omega\|_2 \, \|H_f^{1/2}\Psi\|
\\
\|a(h)^*\Psi\|&\leq \|h/\sqrt\omega\|_2 \, \|H_f^{1/2}\Psi\| + \|h \|_2 \, \| \Psi\|
\end{split}
\label{eq:st-est}
\end{align}
which holds for all $h , h/\sqrt\omega \in\mathfrak h$ and $\Psi\in\mathcal H$ such that the
left- and right-hand side are well-defined; see \cite[Eq. (13.70)]{spohn_dynamics_2008}.
\begin{lemma}
\label{lemma:standardest1}
Let $h, h/\sqrt{\omega}\in \mathfrak{h}$. Then, we have the following standard estimates
\begin{align}
\norm{a(h)^* (H_f+1)^{-\frac{1}{2}}}\leq \norm{h}_2 +\norm{h/\sqrt{\omega}}_2 \quad \text{and} \quad \norm{a(h) (H_f+1)^{-\frac{1}{2}}}\leq \norm{h/\sqrt{\omega}}_2 .
\label{eq:standartesta}
\end{align}
\end{lemma}
\begin{proof}
Let $\Psi\in \mathcal F [\mathfrak{h}]$ such that $\|\Psi \|_{\mathcal H}=1$. Then, it follows from \eqref{eq:ccr} that
\begin{align}
&\norm{ a(h)^* (H_f+1)^{-\frac{1}{2}}\Psi }^2
\leq \left\langle \Psi, (H_f+1)^{-\frac{1}{2}} a(h)a(h)^* (H_f+1)^{-\frac{1}{2}} \Psi \right\rangle
\notag \\
& = \left\langle \Psi,(H_f+1)^{-\frac{1}{2}} \left(\norm{h}^2_2 + a(h)^*a(h) \right) (H_f+1)^{-\frac{1}{2}} \Psi \right\rangle
\notag \\
&\leq \norm{h}^2_2 +\left\langle \Psi,(H_f+1)^{-\frac{1}{2}} a(h)^*a(h) (H_f+1)^{-\frac{1}{2}} \Psi \right\rangle
\notag \\
&=\norm{h}^2_2 +\norm{ a(h) (H_f+1)^{-\frac{1}{2}}\Psi}^2 .
\end{align}
This implies that
\begin{align}
\label{eq:standest111}
\norm{ a(h)^* (H_f+1)^{-\frac{1}{2}} }\leq \norm{h}_2 +\norm{ a(h) (H_f+1)^{-\frac{1}{2}}}.
\end{align}
Moreover, it follows from the Cauchy-Schwarz inequality that
\begin{align}
&\norm{ a(h) (H_f+1)^{-\frac{1}{2}}\Psi }\leq \int\mathrm{d}^3k\, |h(k)| \norm{a(k) (H_f+1)^{-\frac{1}{2}}\Psi}
\notag \\
&\leq \left( \int\mathrm{d}^3k\, |h(k)|^2/\omega(k)\right)^{\frac{1}{2}}\left( \int\mathrm{d}^3k\, \omega(k) \norm{a(k) (H_f+1)^{-\frac{1}{2}}\Psi}^2 \right)^{\frac{1}{2}}
\notag \\
&\leq \norm{h/\sqrt \omega}_2 \norm{H_f(H_f+1)^{-1}}\leq \norm{h/\sqrt \omega}_2 .
\end{align}
This proves the second estimate, and the first one follows together with \eqref{eq:standest111}.
\end{proof}
\begin{lemma}
\label{lemma:standardest2}
\begin{align}
\norm{ V \left( H_0 +1 \right)^{-\frac{1}{2}}}&\leq
\norm{f}_2 +2\norm{f/\sqrt{\omega}}_2 .
\end{align}
\end{lemma}
\begin{proof}
By definition in \eqref{eq:f}, we have $f, f/\sqrt{\omega}\in \mathfrak{h}$. Then, it follows from Lemma \ref{lemma:standardest1} that
\begin{align}
&\norm{ V \left( H_f +1 \right)^{-\frac{1}{2}}}
\leq \norm{ \sigma_1 \otimes a(f)\left( H_f +1 \right)^{-\frac{1}{2}} }+ \norm{\sigma_1 \otimes a(f)^*\left( H_f +1 \right)^{-\frac{1}{2}} }
\notag \\
& \leq \norm{ a(f) \left( H_f +1 \right)^{-\frac{1}{2}} } +\norm{ a(f)^*\left( H_f +1 \right)^{-\frac{1}{2}} }
\leq \norm{f}_2 +2\norm{f/\sqrt{\omega}}_2 ,
\end{align}
and furthermore, we obtain by the functional calculus together with $|e_i+r+1|\geq |r+1|$ that
\begin{align}
\norm{ \frac{(H_f +1)^{\frac{1}{2}} }{(H_0 +1)^{\frac{1}{2}} }}\leq \norm{ \frac{(H_f +1)^{1/2} }{(H_0 +1)^{1/2} }} &=\sup_{r\in [0,\infty), i=0,1} \frac{(r+1)^{1/2}}{(e_i+r+1)^{1/2}}
\leq 1 .
\end{align}
Then, we conclude
\begin{align}
\norm{ V \left( H_0 +1 \right)^{-\frac{1}{2}}}
\leq \norm{ V \left( H_f +1 \right)^{-\frac{1}{2}}} \norm{ \frac{(H_f +1)^{1/2} }{(H_0 +1)^{1/2} }} \leq \left( \norm{f}_2 +2\norm{f/\sqrt{\omega}}_2\right) .
\end{align}
This completes the proof.
\end{proof}
\begin{proof}[Proof of Proposition \ref{thm:Hsa}]
Lemma \ref{lemma:standardest2} implies that $\| \lim_{s \to \infty} V \frac{1}{H_0 + s} \| = 0 $ and, therefore, $V$ is bounded relatively to $H_0$, with infinitesimal bound.
\end{proof}
\end{appendix}
\section*{Acknowledgement}
D.\ -A.\ Deckert and F.\ H\"anle would like to thank the IIMAS at UNAM and M.\
Ballesteros the Mathematisches Institut at LMU Munich for their hospitality. This
project was partially funded by the DFG Grant DE 1474/3-1, the grants PAPIIT-DGAPA
UNAM IN108818, SEP-CONACYT 254062, and the junior research group ``Interaction
between Light and Matter'' of the Elite Network Bavaria. M.\ B.\ is a
Fellow of the Sistema Nacional de Investigadores (SNI).
F.\ H.\ gratefully acknowledges financial support by the ``Studienstiftung des deutschen Volkes''.
Moreover, the authors express their gratitude for the fruitful discussions with
V.\ Bach, J.\ Faupin, J.\ S.\ M\o ller, A.\ Pizzo and W.\ De Roeck.
\bibliographystyle{amsplain}
|
\section{Introduction and main ideas}
The class of autoregressive moving averages (ARMA) is one of the most popular classes of stochastic processes for modeling time series in discrete time. This class goes back to the thesis of Whittle in 1951 and was popularized in \citet{Box-Jenkins}. The continuous time analogue of an ARMA process is called a CARMA process, and it is the formal solution $(X_t)_{t\in \RR}$ to the equation
\begin{align}\label{MCARMAheureqn}
P(D) X_t = Q(D) DZ_t,\quad t \in \mathbb{R},
\end{align}
where $P$ and $Q$ are polynomials of degree $p$ and $q$, respectively. Furthermore, $D$ denotes differentiation with respect to $t$, and $(Z_t)_{t\in \RR}$ is a L\'evy process, the continuous time analogue of a random walk. In the following we will assume that $p>q$ and $P(z),Q(z)\neq 0$ whenever $\text{Re}(z)\geq 0$. In this case one can give precise meaning to $(X_t)_{t\in \RR}$ as a causal stochastic process through a state-space representation as long as $(Z_t)_{t\in \mathbb{R}}$ has log moments. L\'{e}vy-driven CARMA processes have found many applications, for example, in modeling temperature, electricity and stochastic volatility, cf. \cite{BenthBenthPrice,GarciaKluppelbergMullerElectricity,TodorovStochasticVol}. Moreover, there exists a vast amount of literature on theoretical results for CARMA processes (and variations of these), and a few references are \cite{brockwellRecent,BrockwellMarquardt,BrockwellLevyCARMA,nonNegCARMA,marquardtMFICARMA,MarquardtStelzer,stelzer2011carma}.
It is well-known that any causal CARMA process has a continuous time moving average representation of CMA($\infty$) type
\begin{align*}
X_t = \int_{-\infty}^t g(t-u)\,dZ_u,\qquad t\in\RR,
\end{align*}
see the references above or Section~\ref{CARMArelation}. This representation may be very convenient for studying many of their properties. A main contribution of our work is that we obtain a CAR($\infty$) representation of CARMA processes of the form
\begin{align}\label{ARrep}
R(D)X_t = \int_0^\infty X_{t-u}f(u)\, du + DZ_t,\quad t \in \mathbb{R},
\end{align}
where $R$ is a polynomial of order $p-q$ and $f:\RR\to \RR$ is a deterministic function, both defined through $P$ and $Q$. Since $(X_t)_{t\in \mathbb{R}}$ is $p-q-1$ times differentiable, see \cite[Proposition~3.32]{MarquardtStelzer}, the relation \eqref{ARrep} is well-defined if we integrate both sides once. A heuristic argument for obtaining \eqref{ARrep} from \eqref{MCARMAheureqn} is as follows. If $q=0$, $Q$ is constant and \eqref{ARrep} holds with $R=P$ and $f =0$. If $q\geq 1$, it is convenient to rephrase \eqref{MCARMAheureqn} in the frequency domain:
\begin{align}\label{frequencyEq}
\frac{P(-iy)}{Q(-iy)}\mathcal{F}[X](y) = \mathcal{F}[DL](y),\quad y \in \mathbb{R}.
\end{align}
Using polynomial long division we may choose a polynomial $R$ of order $p-q$ such that
\begin{align*}
S(z):= Q(z) R(z)-P(z),\qquad z\in \CC,
\end{align*}
is a polynomial of at most order $q-1$. Now observe that
\begin{align*}
\frac{P(-iy)}{Q(-iy)}\mathcal{F}[X](y)
&= \biggr(R(-iy) - \frac{S(-iy)}{Q(-iy)} \biggr) \mathcal{F}[X](y)\\
&= \mathcal{F}[R(D)X](y) - \mathcal{F}[f](y) \mathcal{F}[X](y),
\end{align*}
where $f:\mathbb{R}\to \mathbb{R}$ is the $L^2$ function characterized by $\mathcal{F}[f](y) = S(-iy)/Q(-iy)$ for $y\in \mathbb{R}$. (In fact, we even know that $f$ is vanishing on $(-\infty,0)$ and decays exponentially fast at $\infty$, cf. Remark~\ref{fComputation}.) Combining this identity with \eqref{frequencyEq} results in the representation \eqref{ARrep}.
We show in Theorem~\ref{MCARMAasMSDDE} that \eqref{ARrep} does indeed hold true for any invertible (L\'{e}vy-driven) CARMA process. Similar relations are shown to hold for invertible fractionally integrated CARMA (FICARMA) processes, where $(Z_t)_{t\in \mathbb{R}}$ is a fractional L\'{e}vy process, and also for their multi-dimensional counterparts, which we will refer to as MCARMA and MFICARMA processes, respectively. We use these representations to obtain a prediction formula for general CARMA type processes (see Corollary~\ref{CARMAprediction}). A prediction formula for invertible one-dimensional L\'{e}vy-driven CARMA processes is given in \cite[Theorem 2.7]{brockwell2015prediction}, but prediction of MCARMA processes has, to the best of our knowledge, not been studied in the literature.
Autoregressive representations such as \eqref{ARrep} are useful for several reasons. To give a few examples, they separate the noise $(Z_t)_{t\in \RR}$ from $(X_t)_{t\in\RR}$ and hence provide a recipe for recovering increments of the noise from the observed process, they ease the task of prediction (and thus estimation), and they clarify the dynamic behavior of the process. These facts motivate the idea of defining a broad class of processes, including the CARMA type processes above, which all admit an autoregressive representation, and it turns out that a well-suited class to study is the one formed by solutions to multi-dimensional stochastic delay differential equations (MSDDEs). To be precise, for an integrable $n$-dimensional (measurable) process $Z_t=(Z_t^1,\dots,Z_t^n)^T$, $t \in \RR$, with stationary increments and a finite signed $n \times n$ matrix-valued measure $\eta$, concentrated on $[0,\infty)$, a stationary process $X_t = (X^1_t,\dots, X^n_t)^T$, $t \in \mathbb{R}$, is a solution to the associated MSDDE if it satisfies
\begin{align}\label{MultiSDDEcompact}
dX_t = \eta \ast X (t)\, dt + dZ_t.
\end{align}
By equation \eqref{MultiSDDEcompact} we mean that
\begin{align}\label{MSDDE}
X^j_t -X^j_s= \sum_{k=1}^n \int_s^t \int_{[0,\infty)} X^k_{u-v}\, \eta_{jk}(dv)\, du + Z^j_t - Z^j_s , \quad j = 1,\dots, n,
\end{align}
almost surely for each $s<t$. This system of equations is an extension of the stochastic delay differential equation (SDDE) in \cite[Section~3.3]{contARMAframework} to the multivariate case. The overall structure of \eqref{MultiSDDEcompact} is also in line with earlier literature such as \cite{GK,Mohammed} on univariate SDDEs, but here we allow for infinite delay ($\eta$ is allowed to have unbounded support) which is a key property in order to include the CARMA type processes in the framework.
The structure of the paper is as follows: In Section~\ref{prel} we introduce the notation used throughout this paper. Next, in Section~\ref{sdfsdf}, we develop the general theory for MSDDEs with particular focus on existence, uniqueness and prediction. The general results of Section~\ref{sdfsdf} are then specialized in Section~\ref{sectionMSDDE} to various settings. Specifically, in Section~\ref{RegularIntegrator} we consider the case where the noise process gives rise to a reasonable integral, and in Section~\ref{hOrderSection} we demonstrate how to derive results for higher order SDDEs by nesting them into MSDDEs. Finally, in Section~\ref{CARMArelation} we use the above mentioned findings to represent CARMA processes and generalizations thereof as solutions to higher order SDDEs and to obtain the corresponding prediction formulas.
\section{Notation}\label{prel}
Let $f:\mathbb{R}\to \mathbb{C}^{m\times k}$ be a measurable function and $\mu$ a $k \times n$ (non-negative) matrix measure, that is,
\begin{align*}
\mu = \begin{bmatrix} \mu_{11} & \cdots & \mu_{1n} \\ \vdots & \ddots & \vdots \\
\mu_{k1} & \cdots & \mu_{kn}
\end{bmatrix}
\end{align*}
where each $\mu_{jl}$ is a measure on $\mathbb{R}$. Then, we will write $f\in L^p(\mu)$ if
\begin{align*}
\int_\mathbb{R}\vert f_{il}(u) \vert^p \mu_{lj} (du) < \infty
\end{align*}
for $l=1,\dots, k$, $i=1,\dots, m$ and $j=1,\dots, n$. Provided that $f\in L^1(\mu)$, we set
\begin{align}\label{fIntegral}
\int_\mathbb{R} f(u)\, \mu (du) = \sum_{l=1}^k\begin{bmatrix} \int_\mathbb{R} f_{1l}(u)\, \mu_{l1}(du) & \cdots & \int_\mathbb{R} f_{1l}(u)\, \mu_{ln}(du) \\
\vdots & \ddots & \vdots \\
\int_\mathbb{R} f_{ml}(u)\, \mu_{l1}(du) & \cdots & \int_\mathbb{R} f_{ml}(u)\, \mu_{ln}(du)
\end{bmatrix}.
\end{align}
If $\mu$ is the Lebesgue measure, we will suppress the dependence on the measure and write $f \in L^p$, and in case $f$ is measurable and bounded Lebesgue almost everywhere, $f \in L^\infty$. For two (matrix) measures $\mu^+$ and $\mu^-$ on $\mathbb{R}$, where at least one of them are finite, we call the set function $\mu (B) := \mu^+ (B) - \mu^- (B)$, defined for any Borel set $B$, a signed measure (and, from this point, simply referred to as a measure). We may and do assume that the two measures $\mu^+$ and $\mu^-$ are singular. To the measure $\mu$ we will associate its variation measure $\vert \mu\vert := \mu^+ + \mu^-$, and when $\vert \mu \vert (\mathbb{R})< \infty$, we will say that $\mu$ is finite. Integrals with respect to $\mu$ are defined in a natural way from \eqref{fIntegral} whenever $f\in L^1 (\mu):=L^1(\vert \mu \vert)$. If $f$ is one-dimensional, respectively if $\mu$ is one-dimensional, we will write $f\in L^1(\mu)$ if $f \in L^1 (\vert \mu_{ij} \vert)$ for all $i=1,\dots, k$ and $j=1,\dots, n$, respectively if $f_{ij}\in L^1(\vert \mu \vert)$ for all $i=1,\dots,m$ and $j=1,\dots, k$. The associated integral is defined in an obvious manner.
We define the convolution at a given point $t\in \mathbb{R}$ by
\begin{align*}
f \ast \mu (t) = \int_\mathbb{R} f(t-u)\mu (du)
\end{align*}
provided that $f(t-\cdot)\in L^1( \mu )$. In case that $\mu$ is the Lebesgue-Stieltjes measure of a function $g:\mathbb{R}\to \mathbb{R}^{k\times n}$ we will also write $f\ast g (t)$ instead of $f\ast \mu (t)$ (not to be confused with the standard convolution between functions). For a given measure $\mu$ we set
\begin{align*}
D(\mu) = \biggr\{ z \in \mathbb{C}\, :\, \int_\mathbb{R}e^{\text{Re}(z)u}\, \vert \mu_{ij}\vert (du)< \infty \quad \text{ for } i=1,\dots, k\ \text{and}\ j=1,\dots, n \biggr\}
\end{align*}
and define its Laplace transform $\mathcal{L}[\mu]$ as
\begin{align*}
\mathcal{L}[\mu]_{ij}(z) = \int_\mathbb{R}e^{zu}\, \mu_{ij} (du), \quad \text{for} \quad i=1,\dots,k, \: j=1,\dots,n,
\end{align*}
for every $z \in D(\mu)$. If $\mu$ is a finite measure, we will also refer to the Fourier transform $\mathcal{F}[\mu]$ of $\mu$, which is given as $\mathcal{F}[\mu](y) = \mathcal{L}[\mu](iy)$ for $y \in \mathbb{R}$. If $\mu (du) = f(u)\, du$ for some measurable function $f$, we write $\mathcal{L}[f]$ and $\mathcal{F}[f]$ instead. We will also use that the Fourier transform $\mathcal{F}$ extends from $L^1$ to $L^1 \cup L^2$, and it maps $L^2$ onto $L^2$. We will say that $\mu$ has a moment of order $p \in \mathbb{N}_0$ if
\begin{align*}
\int_\mathbb{R}\vert u \vert^p\, \vert\mu_{jk}\vert (du)< \infty
\end{align*}
for all $j,k = 1,\dots, n$. Finally, for two functions $f,g:\mathbb{R}\to \mathbb{R}$ and $a \in [-\infty,\infty]$, we write $f (t) = o(g(t))$, $f(t) \sim g(t)$ and $f(t) = O(g(t))$ as $t \to a$ if
\begin{align*}
\lim_{t\to a} \frac{f(t)}{g(t)}\to 0,\quad \lim_{t\to a}\frac{f(t)}{g(t)} = 1\quad \text{and}\quad \limsup_{t\to a} \biggr\vert\frac{f(t)}{g(t)} \biggr\vert < \infty,
\end{align*}
respectively.
\section{Stochastic delay differential equations}\label{sdfsdf}
Consider the general MSDDE in \eqref{MultiSDDEcompact}, where the noise $(Z_t)_{t\in \mathbb{R}}$ is a measurable process, which is integrable and has stationary increments. The first main result provides sufficient conditions to ensure existence and uniqueness of a solution. To obtain such results we need to put assumptions on the delay measure $\eta$. In order to do so, we associate to $\eta$ the function $h : D(\eta) \to \CC^{n\times n}$ given by
\begin{align}\label{DefOfh}
h(z) = -zI_n - \mathcal{L}[\eta](z).
\end{align}
where $I_n$ is the $n\times n$ identity matrix.
\begin{theorem}\label{existence} Let $h$ be given in \eqref{DefOfh} and suppose that $\det (h(iy)) \neq 0$ for all $y \in \RR$. Suppose further that $\eta$ has second moment. Then there exists a function $g:\mathbb{R}\to \mathbb{R}^{n \times n}$ in $L^2$ characterized by
\begin{align}\label{gKernelChar1}
\mathcal{F}[g](y) = h(iy)^{-1},
\end{align}
the convolution
\begin{align}\label{solutionForm}
g \ast Z (t) := Z_t + \int_\mathbb{R} g \ast \eta (t-u)\, Z_u\, du
\end{align}
is well-defined for each $t \in \mathbb{R}$ almost surely, and $X_t = g \ast Z (t)$, $t \in \mathbb{R}$, is the unique (up to modification) stationary and integrable solution to \eqref{MultiSDDEcompact}. If, in addition to the above stated assumptions, $\det (h(z)) \neq 0$ for all $z \in \CC$ with $\Real (z) \leq 0$ then the solution in \eqref{solutionForm} is casual in the sense that $(X_t)_{t\in \mathbb{R}}$ is adapted to the filtration
\begin{align*}
\{\sigma (Z_t - Z_s\, :\, s<t)\}_{t \in \mathbb{R}}.
\end{align*}
\end{theorem}
The solution $(X_t)_{t\in \mathbb{R}}$ to \eqref{MultiSDDEcompact} will very often take form as a $(Z_t)_{t\in \mathbb{R}}$-driven moving average, that is,
\begin{align}\label{noiseMA}
X_t = \int_\mathbb{R}g(t-u)\, dZ_u
\end{align}
for each $t \in \mathbb{R}$ (cf. Section~\ref{RegularIntegrator}). This fact justifies the notation $g\ast Z$ introduced in \eqref{solutionForm}. In case $n=1$, equation \eqref{MultiSDDEcompact} reduces to the usual first order SDDE, and then the existence condition becomes $h(iy) = -iy - \mathcal{F}[\eta](y) \neq 0$ for all $y\in \mathbb{R}$, and the kernel driving the solution is characterized by $\mathcal{F}[g](y) = 1/h(iy)$. This is consistent with earlier literature (cf. \cite{contARMAframework,GK,Mohammed}).
The second main result concerns prediction of MSDDEs. In particular, the content of the result is that we can compute a prediction of future values of the observed process if we are able to compute the same type of prediction of the noise.
\begin{theorem}\label{Prediction} Suppose that $\det (h(z)) \neq 0$ for all $z \in \CC$ with $\text{Re}(z)\leq 0$ and that $\eta$ has second moment. Furthermore, let $(X_t)_{t\in \mathbb{R}}$ be the stationary and integrable solution to \eqref{MultiSDDEcompact} and let $g$ be given by \eqref{gKernelChar1}. Fix $s<t$. Then, if we set
\begin{align}\label{PredictionOfNoise}
\hat{Z}_u = \mathbb{E}[Z_u-Z_s \mid Z_s -Z_r,\, r< s],\quad u >s,
\end{align}
it holds that
\begin{align*}
\MoveEqLeft \EE [X_t \mid X_u, \, u\leq s] \\
&= g(t-s) X_s + \int_s^t g(t-u) \eta \ast \big\{\mathds{1}_{(-\infty,s]}X\big\} (u) \, du + g \ast \big\{\mathds{1}_{(s,\infty)}\hat{Z} \big\}(t),
\end{align*}
using the notation
\begin{align*}
\big(\eta \ast \{\mathds{1}_{(-\infty,s]}X\}(u)\big)_j &:= \sum_{k=1}^n \int_{[u-s,\infty)} X^k_{u-v}\, \eta_{jk}(dv) \quad \text{and} \\
\big(g \ast \{\mathds{1}_{(s,\infty)}\hat{Z}\}(u)\big)_j &:= \sum_{k=1}^n \int_{[0,u-s)} \hat{Z}^k_{u-v}\, g_{jk}(dv)
\end{align*}
for $u>s$ and $j=1,\dots,n$.
\end{theorem}
\begin{remark}
In case $(Z_t)_{t\in \mathbb{R}}$ is a L\'{e}vy process, the prediction formula in Theorem~\ref{Prediction} simplifies, since $\hat{Z}_u = (u-s)\mathbb{E}[Z_1]$ and thus
\begin{align*}
\MoveEqLeft \EE [X_t \mid X_u, \, u\leq s] \\
&= g(t-s) X_s + \int_s^t g(t-u) \eta \ast \big\{\mathds{1}_{(-\infty,s]}X\big\} (u) \, du + \int_s^tg(t-u)\, du\, \mathbb{E}[Z_1],
\end{align*}
using integration by parts. Obviously, the formula takes an even simpler form if $\mathbb{E}[Z_1]= 0$. If instead we are in a long memory setting and $(Z_t)_{t\in \mathbb{R}}$ is a fractional Brownian motion, we can rely on \cite{GripenbergNorros} to obtain $(\hat{Z}_u)_{s<u\leq t}$ and then use the formula given in Theorem~\ref{Prediction} to compute the prediction $\mathbb{E}[X_t\mid X_u,\, u\leq s]$.
\end{remark}
In Section~\ref{CARMArelation} we use this prediction formula combined with the relation between MSDDEs and MCARMA processes to obtain a prediction formula for any invertible MCARMA process.
\section{Examples and further results}\label{sectionMSDDE}
In this section we will consider several examples of MSDDEs and give some additional results. We begin by defining what we mean by a regular integrator, since this makes it possible to have the compact form \eqref{noiseMA} of the solution to \eqref{MultiSDDEcompact} in most cases. Next, we show how one can nest higher order MSDDEs in the (first order) MSDDE framework. Finally, we show that invertible MCARMA processes (and some generalizations) form a particular subclass of solutions to higher order MSDDEs.
\subsection{Regular integrators and moving average representations}\label{RegularIntegrator}
When considering the form of the solution in Theorem~\ref{existence} it is natural to ask if this can be seen as a moving average of the kernel $g$ with respect to the noise $(Z_t)_{t \in \RR}$, that is, if
\begin{align}\label{movingAverageEntry}
X^j_t = \bigg(\int_\mathbb{R}g (t-u)\, dZ_u\biggr)_j = \sum_{k=1}^n \int_\mathbb{R}g_{jk}(t-u)\, dZ^k_u , \quad t \in \mathbb{R},
\end{align}
for $j=1,\dots, n$. The next result shows that the answer is positive if $(Z^k_t)_{t\in \mathbb{R}}$ is a "reasonable"\ integrator for a suitable class of deterministic integrands for each $k=1,\dots, n$.
\begin{proposition}\label{MArep}
Let $h$ be the function given in \eqref{DefOfh} and suppose that, for all $y \in \RR$, $\det (h(iy)) \neq 0$. Suppose further that $\eta$ has second moment and let $(X_t)_{t\in \mathbb{R}}$ be the solution to \eqref{MultiSDDEcompact} given by \eqref{solutionForm}. Finally assume that, for each $k=1, \dots, n$, there exists a linear map $I_k:L^1 \cap L^2 \to L^1(\mathbb{P})$ which has the following properties:
\begin{enumerate}[(i)]
\item For all $s<t$, $I_k (\mathds{1}_{(s,t]}) = Z^k_t - Z^k_s$.
\item If $\mu$ is a finite Borel measure on $\mathbb{R}$ having first moment then
\begin{align}\label{FubiniRelation}
I_k \biggr( \int_\mathbb{R}f_r(t-\cdot)\, \mu (dr) \biggr) = \int_\mathbb{R}I_k(f_r(t-\cdot))\, \mu (dr)
\end{align}
almost surely for all $t\in \mathbb{R}$, where $f_r= \mathds{1}_{[0,\infty)}(\cdot-r) - \mathds{1}_{[0,\infty)}$ for $r \in \mathbb{R}$.
\end{enumerate}
Then it holds that
\begin{align}\label{MArelation}
X^j_t = \sum_{k=1}^n I_k (g_{jk}(t-\cdot)), \quad j=1,\dots, n,
\end{align}
almost surely for each $t\in \mathbb{R}$. In this case, $(Z_t)_{t \in \RR}$ will be called a regular integrator and we will write $\int \cdot \, dZ^k = I_k$.
\end{proposition}
The typical example of a regular integrator is a multi-dimensional L\'{e}vy process:
\begin{example}\label{LevyIntegrator}
Suppose that $(Z_t)_{ t\in \RR}$ is an $n$-dimensional integrable L\'{e}vy process. Then, in particular, each $(Z^j_t)_{t\in \mathbb{R}}$ is an integrable (one-dimensional) L\'{e}vy process, and in \cite[Lemma~5.3]{contARMAframework} it is shown that the integral $\int_\mathbb{R} f(u)\, dZ^j_u$ is well-defined in the sense of \cite{Rosinski_spec} and belongs to $L^1(\mathbb{P})$ if $f \in L^1 \cap L^2$. Moreover, the stochastic Fubini result given in \cite[Theorem~3.1]{QOU} implies in particular that condition (ii) of Proposition~\ref{MArep} is satisfied, which shows that $(Z_t)_{t\in \mathbb{R}}$ is a regular integrator and that \eqref{movingAverageEntry} holds.
\end{example}
We will now show that a class of multi-dimensional fractional L\'{e}vy processes can serve as regular integrators as well (cf. Example~\ref{fracLevyIntegrator} below). Fractional noise processes are often used as a tool to incorporate (some variant of) long memory in the corresponding solution process. As will appear, the integration theory for fractional L\'{e}vy processes we will use below relies on the ideas of \cite{Tina}, but is extended to allow for symmetric stable L\'{e}vy processes as well. For more on fractional stable L\'{e}vy processes, the so-called linear fractional stable motions, we refer to \cite[p. 343]{Stable}. First, however, we will need the following observation:
\begin{proposition}\label{RiemannReg}
Let $f:\mathbb{R}\to \mathbb{R}$ be a function in $L^1\cap L^{\alpha}$ for some $\alpha \in (1,2]$. Then the right-sided Riemann-Liouville fractional integral
\begin{align}\label{RiemannLiouvilleFrac}
I^{\beta}_- f: t \mapsto \frac{1}{\Gamma (\beta)}\int_t^\infty f(u)(u-t)^{\beta -1}\, du
\end{align}
is well-defined and belongs to $L^\alpha$ for any $\beta \in (0,1-1/\alpha)$.
\end{proposition}
\begin{example}\label{fracLevyIntegrator}
Let $\alpha = (\alpha_1,\dots, \alpha_n)$ with $\alpha_j \in (1,2]$ and $f=(f_{jk}):\mathbb{R}\to \mathbb{R}^{n\times n}$ be a function such that $f_{jk} \in L^1 \cap L^{\alpha_k}$ for $j,k=1,\dots, n$. Consider an $n$-dimensional L\'{e}vy process $(L_t)_{t\in \mathbb{R}}$ where its $j$-th coordinate is symmetric $\alpha_j$-stable if $\alpha_j \in (1,2)$ and mean zero and square integrable if $\alpha_j=2$. Then, for a given vector $\beta = (\beta_1,\dots, \beta_n)$ with $\beta_j \in (0,1-1/\alpha_j)$ for $j=1,\dots, n$ the corresponding fractional L\'{e}vy process $(Z_t)_{t\in \mathbb{R}}$ with parameter $\beta$ is defined as
\begin{align*}
Z^j_t &= \int_\mathbb{R}\big(I^{\beta_j}_- [\mathds{1}_{(-\infty,t]}-\mathds{1}_{(-\infty,0]}]\big)(u)\, dL^j_u \\
&= \frac{1}{\Gamma (1+ \beta_j)}\int_\mathbb{R}\big[(t-u)_+^{\beta_j} - (-u)^{\beta_j}_+\big]\, dL^j_u
\end{align*}
for $t\in \mathbb{R}$ and $j=1,\dots, n$, and where $x_+ = \max \{x,0\}$. In light of Proposition~\ref{RiemannReg}, this definition makes it natural to define the integral of a function $f:\mathbb{R}\to \mathbb{R}$ in $L^1 \cap L^{\alpha_j}$ (particularly in $L^1 \cap L^2$) with respect to $(Z^j_t)_{t\in \mathbb{R}}$ as
\begin{align*}
\int_\mathbb{R} f(u)\, dZ^j_u = \int_\mathbb{R} \big(I^{\beta_j}_-f\big)(u)\, dL^j_u
\end{align*}
for $j=1,\dots, n$. Note that the integral belongs to $L^2(\mathbb{P})$ for $\alpha_j=2$ and to $L^\gamma(\mathbb{P})$ for any $\gamma <\alpha_j$ if $\alpha_j \in (1,2)$. Using Proposition~\ref{RiemannReg} and the stochastic Fubini result given in \cite[Theorem~3.1]{QOU} for $(L^j_t)_{t\in \mathbb{R}}$ it is straightforward to verify that assumption (ii) of Proposition~\ref{MArep} is satisfied as well, and thus $(Z_t)_{t \in \RR}$ is a regular integrator and the solution $(X_t)_{t\in \mathbb{R}}$ to \eqref{MultiSDDEcompact} takes the moving average form \eqref{movingAverageEntry}.
\end{example}
At this point it should be clear that the conditions for being a regular integrator are mild, hence they will, besides the examples mentioned above, also be satisfied for a wide class of semimartingales with stationary increments.
\subsection{Higher order (multivariate) SDDEs}\label{hOrderSection}
An advantage of introducing the multivariate setting \eqref{MultiSDDEcompact} is that we can nest higher order MSDDEs in this framework. Effectively, as usual and as will be demonstrated below, it is done by increasing the dimension accordingly.
Let $\varpi_0,\varpi_1,\dots, \varpi_{m-1}$ be (entrywise) finite $n \times n$ measures concentrated on $[0,\infty)$ which all admit second moment, and let $(Z_t)_{t\in \mathbb{R}}$ be an $n$-dimensional integrable stochastic process with stationary increments. For convenience we will assume that $(Z_t)_{t\in \mathbb{R}}$ is a regular integrator in the sense of Proposition~\ref{MArep}. We will say that an $n$-dimensional stationary, integrable and measurable process $(X_t)_{t\in \mathbb{R}}$ satisfies the corresponding $m$-th order MSDDE if it is $m-1$ times differentiable and
\begin{align}\label{hOrderSDDE}
dX^{(m-1)}_t = \sum_{j=0}^{m-1} \varpi_j\ast X^{(j)} (t)\, dt + dZ_t
\end{align}
where $(X^{(j)}_t)_{t\in \mathbb{R}}$ denotes the entrywise $j$-th derivative of $(X_t)_{t\in\mathbb{R}}$ with respect to $t$. By \eqref{hOrderSDDE} we mean that
\begin{align*}
\big(X^{(m-1)}_t\big)^k-\big(X^{(m-1)}_s\big)^k = \sum_{j=0}^{m-1}\sum_{l=1}^n \int_s^t \int_{[0,\infty)} \big(X^{(j)}_{u-v}\big)^l\, (\varpi_j )_{kl} (dv)\, du +Z^k_t - Z^k_s
\end{align*}
for $k=1,\dots, n$ and each $s<t$ almost surely.
Equation \eqref{hOrderSDDE} corresponds to the $mn$-dimensional MSDDE in \eqref{MultiSDDEcompact} with noise $(0,\dots, 0,Z_t^T)^T \in \mathbb{R}^{m n}$ and
\begin{align}\label{etaHigherOrder}
\eta = \begin{bmatrix*}[c] 0 & I_n\delta_0 & 0 & \cdots & 0 \\
0 & 0 & I_n\delta_0 & \cdots & 0 \\
\vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & 0 & \cdots & I_n\delta_0 \\
\varpi_0 & \varpi_1 & \varpi_2 & \cdots & \varpi_{m-1}
\end{bmatrix*}.
\end{align}
(If $n=1$ then $\eta = \varpi_0$.)
With $\eta$ given by \eqref{etaHigherOrder} it follows that
\begin{align*}
D(\eta) = \bigcap_{j=0}^{m-1} D(\varpi_j)
\end{align*}
and
\begin{align*}
h(z) = -\begin{bmatrix}
I_n z & I_n & 0 & \cdots & 0 \\
0 & I_n z & I_n & \cdots & 0 \\
\vdots & \vdots & \ddots & \ddots & \vdots \\
0 & 0 & \cdots & I_n z & I_n\\
\mathcal{L}[\varpi_0](z) & \mathcal{L}[\varpi_1](z) & \cdots & \mathcal{L}[\varpi_{m-2}](z) & I_n z+\mathcal{L}[\varpi_{m-1}](z)
\end{bmatrix}
\end{align*}
for $z \in D(\eta)$. In general, we know from Theorem~\ref{existence} that a solution to \eqref{hOrderSDDE} exists if $\det (h(iy))\neq 0$ for all $y \in \mathbb{R}$, and in this case the unique solution is given by
\begin{align}\label{solutionTohOrder}
X_t = \int_\mathbb{R}g_{1m}(t-u)\, dZ_u,\quad t \in \mathbb{R},
\end{align}
where $\mathcal{F}[g_{1m}]$ is characterized as entrance $(1,m)$ in the $n \times n$ block representation of $h(i\cdot)^{-1}$. In other words, if $e_j$ denotes the $j$-th canonical basisvector of $\mathbb{R}^m$ and $\otimes$ the Kronecker product,
\begin{align*}
\mathcal{F}[g_{1m}](y) = (e_1\otimes I_n)^T h(iy)^{-1}(e_m\otimes I_n)
\end{align*}
for $y \in \mathbb{R}$. However, due to the particular structure of $\eta$ in \eqref{etaHigherOrder} we can simplify these expressions:
\begin{theorem}\label{simpleFourierKernel} Let the setup be as above. Then it holds that
\begin{align}\label{higherOrderDetCond}
\det (h(z)) = \det \biggr(I_n(-z)^m - \sum_{j=0}^{m-1} \mathcal{L}[\varpi_j](z) (-z)^j \biggr)
\end{align}
for all $z\in D(\eta)$, and if $\det (h(iy))\neq 0$ for all $y \in \mathbb{R}$, there exists a unique solution to \eqref{hOrderSDDE} and it is given as \eqref{solutionTohOrder} where $g:\mathbb{R}\to \mathbb{R}^{n \times n}$ is characterized by
\begin{align}\label{hOrderFourierT}
\mathcal{F}[g_{1m}](y) = \biggr( I_n(-iy)^m - \sum_{j=0}^{m-1} \mathcal{F}[\varpi_j](y) (-iy)^j \biggr)^{-1}
\end{align}
for $y \in \mathbb{R}$. The solution is causal if $\det (h(z))\neq 0$ whenever $\text{Re}(z)\leq 0$.
\end{theorem}
Observe that, as should be the case, we are back to the first order MSDDE when $m=1$ and \eqref{higherOrderDetCond}-\eqref{hOrderFourierT} agree with Theorem~\ref{existence}. As we will see in Section~\ref{CARMArelation} below, one motivation for introducing higher order MSDDEs of the form \eqref{hOrderSDDE} and to study the structure of the associated solutions, is their relation to MCARMA processes. However, we start with the multivariate CAR($p$) process, where no delay term will be present, as an example:
\begin{example}\label{CARprocess}
Let $P(z) = I_nz^p + A_1z^{p-1}+ \cdots + A_p$, $z \in\mathbb{C}$, for suitable $A_1,\dots, A_p\in \mathbb{R}^{n \times n}$. The associated CAR($p$) process $(X_t)_{t\in \mathbb{R}}$ with noise $(Z_t)_{t\in \mathbb{R}}$ can be thought of as formally satisfying $P(D)X_t = DZ_t$, $t\in \mathbb{R}$, where $D$ denotes differentiation with respect to $t$. Integrating both sides and rearranging terms gives
\begin{align}\label{CARrelation}
dX^{(p-1)}_t = - \sum_{j=0}^{p-1}A_{p-j}X^{(j)}_t\, dt + dZ_t,\quad t \in \mathbb{R},
\end{align}
which is of the form \eqref{hOrderSDDE} with $m=p$ and $\varpi_j = -A_{p-j}\delta_0$ for $j=0,1,\dots, p-1$. Proposition~\ref{simpleFourierKernel} shows that a unique solution exists if
\begin{align*}
\det\biggr( I_n(iy)^p + \sum_{j=0}^{p-1} A_{p-j} (iy)^j \biggr) = \det (P(iy))\neq 0
\end{align*}
for all $y \in \mathbb{R}$, and in this case $\mathcal{F}[g_{1m}](y) = P(-iy)^{-1}$ for $y\in \mathbb{R}$. This agrees with the rigorous definition of the CAR($p$) process, see e.g. \cite{MarquardtStelzer}. In case $p=1$, \eqref{CARrelation} collapses to the multivariate Ornstein-Uhlenbeck equation
\begin{align*}
dX_t = - A_1 X_t\, dt + dZ_t, \quad t \in \mathbb{R},
\end{align*}
and if the eigenvalues of $A_1$ are all positive, it is easy to check that $g_{1m}(t) = e^{-A_1t}\mathds{1}_{[0,\infty)}(t)$ so that the unique solution $(X_t)_{t\in \mathbb{R}}$ is causal and takes the well-known form
\begin{align}\label{OUsolution}
X_t = \int_{-\infty}^te^{-A_1(t-u)}\, dZ_u
\end{align}
for $t\in \mathbb{R}$. Lévy-driven multivariate Ornstein-Uhlenbeck processes have been studied extensively in the literature, and the moving average structure \eqref{OUsolution} of the solution is well-known when $(Z_t)_{t \in \RR}$ is a Lévy process. We refer to \cite{BarndorffJensenSoerensen,SatoWatanabeYamazato,SatoYamazato} for further details. The one-dimensional case where $(Z_t)_{t \in \RR}$ is allowed to be a general stationary increment process has been studied in \cite{QOU}.
\end{example}
\subsection{Relations to MCARMA processes}\label{CARMArelation}
Let $p \in \mathbb{N}$ and define the polynomials $P,Q:\mathbb{C}\to \mathbb{C}^{n\times n}$ by
\begin{align}\label{CARMApolyn}
\begin{aligned}
P(z) &= I_nz^p + A_1 z^{p-1}+ \cdots + A_p \quad \text{and}\\
Q(z) &= B_0 + B_1z + \cdots + B_{p-1}z^{p-1}
\end{aligned}
\end{align}
for $z \in \mathbb{C}$ and suitable $A_1,\dots, A_p,B_0,\dots,B_{p-1} \in \RR^{n \times n}$. We will also fix $q \in \mathbb{N}_0$, $q<p$, and set $B_q = I_n$ and $B_j = 0$ for all $q<j<p$. It will always be assumed that $\det (P(iy)) \neq 0$ for all $y \in \mathbb{R}$. Under this assumption there exists a function $\tilde{g}:\mathbb{R}\to \mathbb{R}^{n \times n}$ which is in $L^1\cap L^2$ and
\begin{align}\label{definingMCARMAkernel}
\mathcal{F}[\tilde{g}](y)= P(-iy)^{-1}Q(-iy)
\end{align}
for every $y \in \mathbb{R}$. Consequently, for any regular integrator $(Z_t)_{t\in \mathbb{R}}$ in the sense of Proposition~\ref{MArep}, the $n$-dimensional stationary and integrable process $(X_t)_{t \in \mathbb{R}}$ given by
\begin{align}\label{ZdrivenCARMA}
X_t = \int_\mathbb{R} \tilde{g}(t-u)\, dZ_u, \quad t \in \mathbb{R},
\end{align}
is well-defined. If it is additionally assumed that $\det (P(z)) \neq 0$ for $z \in \CC$ with $\Real (z) \geq 0$ then it is argued in \cite{MarquardtStelzer} that
\begin{align}\label{gMCARMA}
\tilde{g}(t) = \mathds{1}_{[0,\infty)}(t)(e_1^p\otimes I_n)^T e^{At}E
\end{align}
where
\begin{align*}
A = \begin{bmatrix}
0 & I_n & 0 & \cdots & 0 \\
0 & 0 & I_n &\cdots & 0\\
\vdots & \vdots & \ddots & \ddots & \vdots \\
0 & 0 & \cdots & 0 & I_n \\
-A_p & -A_{p-1} & \cdots & -A_{2} & -A_{1}
\end{bmatrix} \quad \text{and}\quad E= \begin{bmatrix}
E_1 \\ \vdots \\ E_p
\end{bmatrix},
\end{align*}
with $E(z) = E_1z^{p-1} + \cdots + E_p$ chosen such that
\begin{align*}
z \mapsto P(z) E(z) -Q(z)z^p
\end{align*}
is at most of degree $p-1$. (Above, and henceforth, we use the notation $e^k_j$ for the $j$-th canonical basis vector of $\RR^k$.) We will refer to the process $(X_t)_{t\in \mathbb{R}}$ as a $(Z_t)_{t\in \mathbb{R}}$-driven MCARMA($p,q$) process. For instance, when $(Z_t)_{t \in \RR}$ is an $n$-dimensional L\'{e}vy process, $(X_t)_{t\in \mathbb{R}}$ is a (L\'{e}vy-driven) MCARMA($p,q$) process as introduced in \cite{MarquardtStelzer}. If $(L_t)_{t\in \mathbb{R}}$ is an $n$-dimensional square integrable L\'{e}vy process with mean zero, and
\begin{align*}
Z^j_t = \frac{1}{\Gamma (1+\beta_j)}\int_\mathbb{R} \big[(t-u)_+^{\beta_j}-(-u)_+^{\beta_j}\big]\, dL^j_u, \quad t\in \mathbb{R},
\end{align*}
for $\beta_j \in (0,1/2)$ and $j=1,\dots, n$, then $(X_t)_{t\in \mathbb{R}}$ is an MFICARMA($p,\beta,q$) process, $\beta = (\beta_1,\dots, \beta_n)$, as studied in \cite{marquardtMFICARMA}. For the univariate case ($n=1$), the processes above correspond to the CARMA($p,q$) and FICARMA($p,\beta_1,q$) process, respectively. The class of CARMA processes has been studied extensively, and we refer to the references in the introduction for details.
\begin{remark} Observe that, generally, L\'{e}vy-driven MCARMA (hence CARMA) processes are defined even when $(Z_t)_{t\in \mathbb{R}}$ has no more than log moments. However, it relies heavily on the fact that $\tilde{g}$ and $(Z_t)_{t\in \mathbb{R}}$ are well-behaved enough to ensure that the process in \eqref{ZdrivenCARMA} remains well-defined. At this point, a setup where the noise does not admit a first moment has not been integrated in a framework as general as that of \eqref{MultiSDDEcompact}.
\end{remark}
In the following our aim is to show that, under a suitable invertibility assumption, the $(Z_t)_{t\in \mathbb{R}}$-driven MCARMA($p,q$) process given in \eqref{ZdrivenCARMA} is the unique solution to a certain (possibly higher order) MSDDE of the form \eqref{hOrderSDDE}. Before formulating the main result of this section we introduce some notation. To $P$ and $Q$ defined in \eqref{CARMApolyn} we will associate the unique polynomial $R(z) = I_n z^{p-q}+C_{p-q-1}z^{p-q-1} + \cdots +C_0$, $z \in \mathbb{C}$ and $C_0,C_1,\dots, C_{p-q-1}\in \mathbb{R}^{n\times n}$, having the property that
\begin{align}\label{residuePolyn}
z \mapsto Q(z)R(z) - P(z)
\end{align}
is a polynomial of at most order $q-1$ (see the introduction for an intuition about why this property is desirable).
\begin{theorem}\label{MCARMAasMSDDE} Let $P$ and $Q$ be given as in \eqref{CARMApolyn}, and let $(X_t)_{t\in \mathbb{R}}$ be the associated $(Z_t)_{t\in \mathbb{R}}$-driven MCARMA($p,q$) process. Suppose that $\det (Q(z))\neq 0$ for all $z\in \CC$ with $\Real (z)\geq0$. Then $(X_t)_{t\in \mathbb{R}}$ is the unique solution to \eqref{hOrderSDDE} with
\begin{align*}
m=p-q, \quad\varpi_0 (du) = -C_0\delta_0(du) + f(u)\, du , \quad \text{and} \quad \varpi_j= -C_j\delta_0,
\end{align*}
for $1 \leq j \leq m-1$ or, written out,
\begin{align}\label{specificRelation}
dX^{(m-1)}_t = -\sum_{j=0}^{m-1} C_j X^{(j)}_t\, dt + \biggr(\int_0^\infty X_{t-u}^Tf(u)^T\, du\biggr)^T\, dt + dZ_t,
\end{align}
where $C_0,\dots, C_{m-1}\in \mathbb{R}^{n \times n}$ are defined as in \eqref{residuePolyn} above, $(X^{(j)}_t)_{t\in \mathbb{R}}$ is the $j$-th derivative of $(X_t)_{t\in \mathbb{R}}$, and where $f : \RR \to \RR^{n\times n}$ is characterized by
\begin{align}\label{fFourier}
\FF[f](y) = R(-iy)-Q(-iy)^{-1}P(-iy).
\end{align}
\end{theorem}
It follows from Theorem~\ref{MCARMAasMSDDE} that $p-q$ is the order of the (possibly multivariate) SDDE we can associate with a (possibly multivariate) CARMA process. Thus, this seems as a natural extension of \cite{contARMAframework}, where the univariate first order SDDE is studied and related to the univariate CARMA($p,p-1$) process.
\begin{remark}\label{CARMAasSDDEassump}
An immediate consequence of Theorem~\ref{MCARMAasMSDDE} is that we obtain an inversion formula for $(Z_t)_{t\in\mathbb{R}}$-driven MCARMA processes. In other words, it shows how to recover the increments of $(Z_t)_{t\in \mathbb{R}}$ from observing $(X_t)_{t\in \mathbb{R}}$. For this reason it seems natural to impose the invertibility assumption $\det (Q(z))\neq 0$ for all $z \in\mathbb{C}$ with $\text{Re}(z)\geq 0$, which is the direct analogue of the one for discrete time ARMA processes (or, more generally, moving averages). It is usually referred to as the minimum phase property in signal processing. The inversion problem for (L\'{e}vy-driven) CARMA processes has been studied in \cite{contARMAframework,brockwellRecent, nonNegCARMA,brockwell2015prediction} and for (L\'{e}vy-driven) MCARMA processes in \cite{BrockwellSchlemmMCARMAinversion}. In both cases a different approach, which does not rely on MSDDEs, is used.
\end{remark}
\begin{remark}\label{fComputation}
Since the Fourier transform $\mathcal{F}[f]$ of the function $f$ defined in Theorem~\ref{MCARMAasMSDDE} is rational, one can determine $f$ explicitly (e.g., by using the partial fraction expansion of $\mathcal{F}[f]$). Indeed, since the Fourier transform of $f$ is of the same form as the Fourier transform of the solution kernel $\tilde{g}$ of the MCARMA process we can deduce that
\begin{align}\label{fMCARMA}
f(t) = (e_1^q \otimes I_n)^T e^{Bt} F, \quad t \geq 0,
\end{align}
with
\begin{align*}
B = \begin{bmatrix}
0 & I_n & 0 & \cdots & 0 \\
0 & 0 & I_n &\cdots & 0\\
\vdots & \vdots & \ddots & \ddots & \vdots \\
0 & 0 & \cdots & 0 & I_n \\
-B_0 & -B_1 & \cdots & -B_{q-2} & -B_{q-1}
\end{bmatrix} \quad \text{and}\quad F= \begin{bmatrix}
F_1 \\ \vdots \\ F_q
\end{bmatrix},
\end{align*}
where $F(z) = F_1z^{q-1} + \cdots + F_q$ is chosen such that
\begin{align*}
z \mapsto Q(z) F(z) - [Q(z)R(z) - P(z)]z^q
\end{align*}
is at most of degree $q-1$ (see \eqref{definingMCARMAkernel} and \eqref{gMCARMA}).
\end{remark}
In Corollary~\ref{CARMAprediction} we formulate the prediction formula in Theorem \ref{Prediction} in the special case where $(X_t)_{t \in \RR}$ is a $(Z_t)_{t \in \RR}$-driven MCARMA process. In the formulation we use the definition
\begin{align*}
\hat{Z}_u = \mathbb{E}[Z_u-Z_s \mid Z_s -Z_r,\, r< s],\quad u >s,
\end{align*}
in line with \eqref{PredictionOfNoise}.
\begin{corollary}\label{CARMAprediction}
Let $(X_t)_{t \in \RR}$ be a $(Z_t)_{t \in \RR}$-driven MCARMA process and set
\begin{align*}
\tilde{g}_j(t) = (e_1^p \otimes I_n)^T e^{At} \sum_{k=j}^{p-q} A^{k-j}EC_k, \quad t \geq 0,
\end{align*}
for $j=1,\dots, p-q$, where $C_0,\dots,C_{p-q-1}$ are given in \eqref{residuePolyn} and $C_{p-q}=I_n$. Suppose that $\det (P(z))\neq 0$ and $\det (Q(z))\neq 0$ for all $z\in \CC$ with $\Real (z)\geq0$. Fix $s<t$. Then the following prediction formula holds
\begin{align*}
\MoveEqLeft \EE [X_t \mid X_u, \, u\leq s] = \sum_{j=1}^{p-q} \tilde{g}_j(t-s) X_s^{(j-1)} \\
&+ \int_{-\infty}^s \int_s^t \tilde{g}(t-u) f(u-v) \, du \, X_v \, dv + \tilde{g} \ast \{\hat{Z}\mathds{1}_{(s,\infty)} \}(t),
\end{align*}
where $\tilde{g}$ and $f$ are given in \eqref{gMCARMA} and \eqref{fMCARMA}, respectively, and
\begin{align*}
\tilde{g} \ast \{\hat{Z}\mathds{1}_{(s,\infty)} \}(t) = \mathds{1}_{\{p=q+1\}} \hat{Z}_u + (e_1^p \otimes I_n)^TAe^{At}\int_s^t e^{-Av}E \hat{Z}_v \, dv.
\end{align*}
\end{corollary}
\begin{example}
To illustrate the results above we will consider an $n$-dimensional $(Z_t)_{t \in \RR}$-driven MCARMA($3$,$1$) process $(X_t)_{t \in \RR}$ with $P$ and $Q$ polynomials given by
\begin{align*}
\begin{aligned}
P(z) &= I_nz^3 + A_1 z^{2}+ A_2 z + A_3,\\
Q(z) &= B_0 + I_nz
\end{aligned}
\end{align*}
for matrices $B_0,A_1, A_2,A_3\in \mathbb{R}^{n \times n}$ such that $\det (P(z)) \neq 0$ and $\det (Q(z))\neq 0$ for all $z \in \mathbb{C}$ with $\text{Re}(z)\geq 0$. According to \eqref{gMCARMA}, $(X_t)_{t\in \mathbb{R}}$ may be written as
\begin{align*}
X_t = \int_{-\infty}^t (e_1^3\otimes I_n)^Te^{A(t-u)} E \, dZ_u
\end{align*}
where $E_1 = 0$, $E_2=I_n$, and $E_3 = B_0 - A_1$. With
\begin{align*}
C_1 = A_1-B_0, \quad C_0 = A_2 + B_0( B_0 - A_1)
\end{align*}
and
\begin{align*}
F = B_0(A_2 - B_0 ( A_1- B_0))-A_3,
\end{align*}
Theorem~\ref{MCARMAasMSDDE} and Remark \ref{fComputation} imply that
\begin{align*}
dX^{(1)}_t =& - C_1 X^{(1)}_t\, dt -C_0 X_t\, dt + \biggr(\int_0^\infty (FX_{t-u})^T e^{-B_0^Tu}\, du\biggr)^T\, dt + dZ_t.
\end{align*}
Moreover, by Corollary~\ref{CARMAprediction}, we have the prediction formula
\begin{align*}
\EE [X_t \mid X_u, \, u\leq s] =& (e_1^3\otimes I_n)^T e^{At}\biggr[ (EC_1 + AE) X_s + E X_s^{(1)} \\
&+\int_s^t e^{-Au} E \biggr(e^{B_0 u}\int_{-\infty}^s e^{-B_0 v} F X_v\, dv + \hat{Z}_u \biggr)\, du\biggr].
\end{align*}
\end{example}
\section{Proofs and auxiliary results}\label{proofs}
We will start this section by discussing some technical results. These results will then be used in the proofs of all the results stated above.
Recall the function $h:D(\eta) \to \mathbb{C}^{n \times n}$ defined in \eqref{DefOfh}. Note that we always have $\{z \in \mathbb{C}\, :\, \text{Re}(z) \leq 0\}\subseteq D(\eta)$ and $h(iy) = -iyI_n-\mathcal{F}[\eta](y)$ for $y\in \mathbb{R}$. Provided that $\eta$ is sufficiently nice, Proposition~\ref{gExistence} below ensures the existence of a kernel $g:\mathbb{R}\to \mathbb{R}^{n \times n}$ which will drive the solution to \eqref{MultiSDDEcompact}.
\begin{proposition}\label{gExistence} Let $h$ be given as in \eqref{DefOfh} and suppose that $\det (h(iy))\neq 0$ for all $y \in \mathbb{R}$. Then there exists a function $g=(g_{jk}):\mathbb{R}\to \mathbb{R}^{n \times n}$ in $L^2$ characterized by
\begin{align}\label{gKernelChar}
\mathcal{F}[g](y) = h(iy)^{-1}
\end{align}
for $y \in \mathbb{R}$. Moreover, the following statements hold:
\begin{enumerate}[(i)]
\item\label{functionRelation} The function $g$ satisfies
\begin{align*}
g(t-r) - g(s-r) = \mathds{1}_{(s,t]}(r) I_n + \int_s^t g \ast \eta (u-r) \, du
\end{align*}
for almost all $r\in \mathbb{R}$ and each fixed $s<t$.
\item\label{gMoments} If $\eta$ has moment of order $p \in \mathbb{N}$, then $g \in L^q$ for all $q \in [1/p,\infty]$, and
\begin{align}\label{gStieltjes}
g(t) = \mathds{1}_{[0,\infty)}(t)I_n + \int_{-\infty}^t g \ast \eta (u)\, du
\end{align}
for almost all $t \in \mathbb{R}$. In particular,
\begin{align}\label{minusIdentity}
\int_\mathbb{R}g \ast \eta (u)\, du = - I_n.
\end{align}
\item\label{gExponential} If $\int_{[0,\infty)}e^{\delta u }\, \vert \eta_{jk}\vert (du) < \infty$ for all $j,k=1,\dots, n$ and some $\delta >0$, then there exists $\varepsilon >0$ such that
\begin{align*}
\sup_{t \in \mathbb{R}}\max_{j,k =1,\dots,n} \vert g_{jk}(t)\vert e^{\varepsilon \vert t \vert} \leq C
\end{align*}
for a suitable constant $C>0$.
\item\label{gCausality} If $\det (h(z)) \neq 0$ for all $z \in \mathbb{C}$ with $\text{Re}(z)\leq 0$ then $g$ is vanishing on $(-\infty,0)$ almost everywhere.
\end{enumerate}
\end{proposition}
\begin{proof
In order to show the existence of $g$ it suffices to argue that
\begin{align}\label{L2requirement}
y \mapsto \big(h(iy)^{-1} \big)_{jk} \text{ is in } L^2\text{ for } j,k=1,\dots, n,
\end{align}
since the Fourier transform $\mathcal{F}$ maps $L^2$ onto $L^2$. (Here $(h(iy)^{-1})_{jk}$ refers to the $(j,k)$-th entry in the matrix $h(iy)^{-1}$.) Indeed, in this case we just set $g_{jk} = \mathcal{F}^{-1}[(h(i\cdot)^{-1})_{jk}]$.
Let $\widehat{h(iy)}$ denote the $n\times n$ matrix which has the same rows as $h(iy)$, but where the $j$-th column is replaced by the $k$-th canonical basis vector (that is, the vector with all entries equal to zero except of the $k$-th entry which equals one). Then it follows by Cramer's rule that
\begin{align*}
\big(h(iy)^{-1} \big)_{jk} = \frac{\det (\widehat{h(iy)})}{\det (h(iy))}.
\end{align*}
Recalling that $h(iy) = -iyI - \mathcal{F}[\eta](y)$ and that $\mathcal{F}[\eta](y)$ is bounded in $y$ we get by the Leibniz formula that $\vert \det (h(iy))\vert \sim \vert y\vert^n$ and $\vert \det ( \widehat{h(iy)})\vert = O( \vert y\vert^{n-1})$ as $\vert y\vert\to \infty$. This shows in particular that
\begin{align}\label{hInvAsymp}
\big\vert \big(h(iy)^{-1} \big)_{jk}\big\vert =O\big( \vert y \vert ^{-1}\big)
\end{align}
as $\vert y \vert \to \infty$. Since $j$ and $k$ were arbitrarily chosen we get by continuity of (all the entries of) $y \mapsto h(iy)^{-1}$ that \eqref{L2requirement} holds, which ensures the existence part. The fact that $\overline{\mathcal{F}[g](-y)} = \mathcal{F}[g](y)$, $y \in \mathbb{R}$, implies that $g$ takes values in $\mathbb{R}^{n \times n}$.
To show \eqref{functionRelation}, we fix $s<t$ and apply the Fourier transform to obtain
\begin{align*}
\MoveEqLeft\mathcal{F}\biggr[g(t-\cdot) -g(s-\cdot) - \int_s^t g\ast \eta (u-\cdot)\, du\biggr] (y) \\
&= (e^{ity}-e^{isy}) \mathcal{F}[g] (-y) - \mathcal{F}[\mathds{1}_{(s,t]}](y)\mathcal{F}[g](-y)\mathcal{F}[\eta](-y) \\
&= \mathcal{F}[\mathds{1}_{(s,t]}](y) h(-iy)^{-1}(iyI-\mathcal{F}[\eta](-y)) \\
&= \mathcal{F}[\mathds{1}_{(s,t]}](y)I_n,
\end{align*}
which verifies the result.
We will now show \eqref{gMoments} and for this we suppose that $\eta$ has a moment of order $p \in \mathbb{N}$. Then it follows that $\tilde{h}:y \mapsto h(iy)$ is (entry-wise) $p$ times differentiable with the $m$-th derivative given by
\begin{align*}
-\biggr(i\delta_0(\{m-1\}) + i^m\int_{[0,\infty)}e^{iuy}u^m\eta_{jk}(du)\biggr), \quad m=1,\dots, p,
\end{align*}
and in particular all the the entries of $(D^m\tilde{h})(y)$ are bounded in $y$. Observe that, clearly, if a function $A:\mathbb{R}\to \mathbb{C}^{n\times n}$ takes the form
\begin{align}\label{boundedFunc}
A(t) = B(t)C(t) D(t), \quad t \in \mathbb{R},
\end{align}
where all the entries of $B,D:\mathbb{R}\to \mathbb{C}^{n \times n}$ decay at least as $\vert y \vert^{-1}$ as $\vert y \vert \to \infty$ and all the entries of $C:\mathbb{R}\to \mathbb{C}^{n \times n}$ are bounded, then all the entries of $A$ decay at least as $\vert y \vert^{-1}$ as $\vert y \vert \to \infty$. Using the product rule for differentiation and the fact that
\begin{align*}
\big(D\tilde{h}^{-1}\big)(y) = -\tilde{h}(y)^{-1}(D\tilde{h})(y)\tilde{h}(y)^{-1}, \quad y \in \mathbb{R},
\end{align*}
it follows recursively that $D^m\tilde{h}^{-1}$ is a sum of functions of the form \eqref{boundedFunc}, thus all its entries decay at least as $\vert y \vert^{-1}$ as $\vert y \vert \to \infty$, for $m=1,\dots, p$. Since the entries of $D^m\tilde{h}^{-1}$ are continuous as well, they belong to $L^2$, and we can use the inverse Fourier transform $\mathcal{F}^{-1}$ to conclude that
\begin{align*}
\mathcal{F}^{-1}[D^p\tilde{h}] (t) = (it)^p \mathcal{F}^{-1}[\tilde{h}](t) = (it)^p g(t), \quad t \in \mathbb{R},
\end{align*}
is an $L^2$ function. This implies in turn that $t \mapsto g_{jk}(t)(1+ \vert t \vert)^p \in L^2$ and, thus,
\begin{align*}
\int_\mathbb{R}\vert g_{jk}(t)\vert^q\, dt\leq \biggr(\int_\mathbb{R}\big(g_{jk}(t)(1+ \vert t \vert)^p \big)^2\, dt \biggr)^{\tfrac{q}{2}} \biggr(\int_\mathbb{R} (1+ \vert t \vert)^{-\tfrac{2pq}{2-q}}\, dt \biggr)^{1-\tfrac{q}{2}}< \infty
\end{align*}
for any $q \in [1/p,2)$ and $j,k=1, \dots , n$. By using the particular observation that $g \in L^1$ and \eqref{functionRelation} we obtain that
\begin{align}\label{gFunctionalEq}
g(t) = \mathds{1}_{[0,\infty)}(t)I + \int_{-\infty}^t g\ast \eta (u)\, du
\end{align}
for (almost) all $t\in \mathbb{R}$. This shows that
\begin{align*}
\vert g_{jk}(t)\vert \leq 1 + \int_\mathbb{R} \vert (g\ast \eta (u))_{jk} \vert\, du
\leq 1 + \sum_{l=1}^n \int_\mathbb{R} \vert g_{jl}(u)\vert\, du\, \vert \eta_{lk}\vert ([0,\infty))
\end{align*}
for all $t\in \mathbb{R}$ and for every $j,k=1,\dots, n$ which implies $g \in L^\infty$ and, thus, $g\in L^q$ for all $q \in [1/p,\infty]$. Since $g(t) \to 0$ entrywise as $t\to \infty$, we get by \eqref{gFunctionalEq} that
\begin{align*}
\int_\mathbb{R} g \ast \eta (u)\, du = - I_n,
\end{align*}
which concludes the proof of \eqref{gMoments}.
Now suppose that $\int_{[0,\infty)}e^{\delta u}\, \vert \eta_{jk}\vert (du) < \infty$ for all $j,k = 1,\dots, n$ and some $\delta >0$. In this case, $\mathcal{S}_\delta :=\{z \in \mathbb{C}\, :\, \text{Re}(z) \in [-\delta,\delta]\}\subseteq D(\eta)$ and
\begin{align*}
z \mapsto \det (h(z)) = \det \biggr(-zI-\int_{[0,\infty)} e^{zu}\, \eta (du)\biggr)
\end{align*}
is strictly separated from $0$ when $\vert z \vert$, $z\in \mathcal{S}_\delta$, is sufficiently large. Indeed, the dominating term in $\det (h(z))$ is $(-1)^nz^n$ when $\vert z \vert$ is large, since
\begin{align*}
\biggr\vert\biggr(\int_{[0,\infty)} e^{zu}\, \eta (du) \biggr)_{jk}\biggr\vert\leq \max_{l,m=1,\dots, n} \int_{[0,\infty)} e^{\delta u }\, \vert \eta_{lm}\vert (du)
\end{align*}
for $j,k=1,\dots, n$. Using this together with the continuity of $z \mapsto \det (h(z))$ implies that there exists $\tilde{\delta} \in (0,\delta]$ so that $z \mapsto \det (h(z))$ is strictly separated from $0$ on $\mathcal{S}_{\tilde{\delta}}:= \{z \in \mathbb{C}\, :\, \text{Re}(z)\in [-\tilde{\delta},\tilde{\delta}]\}$. In particular, $z \mapsto (h(z)^{-1})_{jk}$ is bounded on any compact set of $\mathcal{S}_{\tilde{\delta}}$, and by using Cramer's rule and the Leibniz formula as in \eqref{hInvAsymp} we get that $\vert (h(z)^{-1})_{jk}\vert =O( \vert z \vert^{-1})$ as $\vert z \vert \to \infty$ provided that $z \in \mathcal{S}_{\tilde{\delta}}$. Consequently,
\begin{align*}
\sup_{x \in [-\tilde{\delta},\tilde{\delta}]} \int_\mathbb{R}\big\vert \big(h(x+iy)^{-1}\big)_{jk} \big\vert^2\, dy <\infty,
\end{align*}
and this implies by \cite[Lemma~5.1]{contARMAframework} that $t\mapsto g_{jk}(t)e^{\varepsilon t}\in L^1$ for all $\varepsilon \in (-\tilde{\delta},\tilde{\delta})$. Fix any $\varepsilon \in (0,\tilde{\delta})$ and $j,k \in \{1, \dots, n\}$, and observe from \eqref{gFunctionalEq} that $g_{jk}$ is absolutely continuous on both $[0,\infty)$ and $(-\infty,0)$ with density $(g \ast \eta)_{jk}$. Consequently, for fixed $t >0$, integration by parts yields
\begin{align}\label{IBPpositive}
\vert g_{jk}(t)\vert e^{\varepsilon t} \leq \vert g_{jk}(0)\vert + \int_\mathbb{R} \vert (g\ast \eta (u))_{jk} \vert e^{\varepsilon u}\, du + \varepsilon \int_\mathbb{R} \vert g_{jk}(u)\vert e^{\varepsilon u}\, du.
\end{align}
Since
\begin{align*}
\int_\mathbb{R} \vert (g\ast \eta (u))_{jk} \vert e^{\varepsilon u}\, du \leq \sum_{l=1}^n \int_\mathbb{R}\vert g_{jl}(u)\vert e^{\varepsilon u}\, du\, \int_{[0,\infty)} e^{\varepsilon u}\, \vert \eta_{lk}\vert (du)
\end{align*}
it follows from \eqref{IBPpositive} that
\begin{align*}
\max_{j,k = 1, \dots , n} \vert g_{jk} (t)\vert \leq C e^{-\varepsilon t}
\end{align*}
for all $t >0$ with
\begin{align*}
C&:= 1 \\
&+ \max_{j,k = 1 ,\dots , n} \biggr(\sum_{l=1}^n\int_\mathbb{R} \vert g_{jl}(u)\vert e^{\varepsilon \vert u\vert }\, du\, \int_{[0,\infty)} e^{\varepsilon u }\, \vert \eta_{lk} \vert (du) + \varepsilon \int_\mathbb{R}\vert g_{jk}(u)\vert e^{\varepsilon \vert u \vert}\, du \biggr).
\end{align*}
By considering $-\varepsilon$ rather than $\varepsilon$ in the above calculations one reaches the conclusion that
\begin{align*}
\max_{j,k = 1, \dots , n} \vert g_{jk} (t)\vert \leq C e^{\varepsilon t}, \quad t<0,
\end{align*}
and this verifies \eqref{gExponential}.
Finally, suppose that $\det (h(z))\neq 0$ for all $z \in \mathbb{C}$ with $\text{Re}(z) \leq 0$. Then it holds that $h$, and thus $z\mapsto h(z)^{-1}$, is continuous on $\{z\in \mathbb{C}\, :\, \text{Re}(z)\leq 0\}$ and analytic on $\{z\in \mathbb{C}\, :\, \text{Re}(z)< 0\}$. Moreover, arguments similar to those in \eqref{hInvAsymp} show that $\vert (h(z)^{-1})_{jk}\vert = O(\vert z\vert^{-1})$ as $\vert z \vert \to \infty$, and thus we may deduce that
\begin{align*}
\sup_{x<0}\int_\mathbb{R} \vert (h(x+iy)^{-1})_{jk}\vert\, dy <\infty.
\end{align*}
From the theory on Hardy spaces, see \cite[Lemma~5.1]{contARMAframework}, \cite[Section~2.3]{Doetsch} or \cite{Dym:gaussian}, this implies that $g$ is vanishing on $(-\infty,0)$ almost everywhere, which verifies \eqref{gCausality} and ends the proof.
\end{proof}
From Proposition~\ref{gExistence} it becomes evident that we may and, thus, do choose the kernel $g$ to satisfy \eqref{gStieltjes} pointwise, so that the function induces a finite Lebesgue-Stieltjes measure $g(du)$. We summarize a few properties of this measure in the corollary below.
\begin{corollary}\label{gMeasure} Let $h$ be the function introduced in \eqref{DefOfh} and suppose that $\det (h(iy))\neq 0$ for all $y \in \mathbb{R}$. Suppose further that $\eta$ has first moment. Then the kernel $g:\mathbb{R}\to \mathbb{R}^{n\times n}$ characterized in \eqref{gKernelChar} induces an $n \times n$ finite Lebesgue-Stieltjes measure, which is given by
\begin{align}\label{specMeasure}
g(du) = I_n\delta_0 (du) + g\ast \eta (u)\, du.
\end{align}
A function $f=(f_{jk}):\mathbb{R}\to \mathbb{C}^{m \times n}$ is in $L^1(g(du))$ if
\begin{align*}
\int_\mathbb{R} \vert f_{jl}(u) (g\ast \eta)_{lk}(u)\vert \, du < \infty, \quad l = 1, \dots, n,
\end{align*}
for $j=1,\dots, m$ and $k = 1, \dots , n$. Moreover, the measure $g(du)$ has $(p-1)$-th moment whenever $\eta$ has $p$-th moment for any $p \in \mathbb{N}$.
\end{corollary}
\begin{proof
The fact that $g$ induces a Lebesgue-Stieltjes measure of the form \eqref{specMeasure} is an immediate consequence of \eqref{gStieltjes}. For a measurable function $f=(f_{jk}):\mathbb{R}\to \mathbb{C}^{m\times n}$ to be integrable with respect to $g(du) = (g_{jk}(du))$ we require that $f_{jl} \in L^1(\vert g_{lk}(du)\vert)$, $l=1,\dots, n$, for each choice of $j=1,\dots, m$ and $k=1,\dots, n$. Since the variation measure $\vert g_{lk}\vert (du)$ of $g_{lk}(du)$ is given by
\begin{align*}
\vert g_{lk}\vert(du) = \delta_0(\{l-k\})\delta_0(du) + \vert (g \ast \eta (u))_{lk}\vert\, du,
\end{align*}
we see that this condition is equivalent to the statement in the result. Finally, suppose that $\eta$ has $p$-th moment for some $p \in \mathbb{N}$. Then, for any $j,k \in \{1,\dots, n\}$, we get that
\begin{align*}
\int_\mathbb{R} \vert u \vert^{p-1}\, \vert g_{jk}\vert (du) \leq \sum_{l=1}^n\bigg( &\vert \eta_{lk}\vert ([0,\infty)) \int_\mathbb{R} \vert u^{p-1} g_{jl}(u)\vert\, du \\
&+ \int_{[0,\infty)} \vert v \vert^{p-1}\, \vert \eta_{lk} \vert (dv)\, \int_\mathbb{R} \vert g_{jl}(u)\vert\, du\biggr).
\end{align*}
From the assumptions on $\eta$ and Proposition~\ref{gExistence}(\ref{gMoments}) we get immediately that $\vert \eta_{lk}\vert ([0,\infty))$, $\int_{[0,\infty)}\vert v \vert^{p-1}\, \vert \eta_{lk} \vert (dv)$ and $\int_\mathbb{R} \vert g_{jl}(u)\vert\, du$ are finite for all $l=1,\dots, n$. Moreover, for any such $l$ we compute that
\begin{align*}
\MoveEqLeft\int_\mathbb{R}\vert u^{p-1} g_{jl}(u)\vert\, du \\
&\leq \int_{\{\vert u \vert \leq 1\}}\vert u^{p-1}g_{jl}(u)\vert\, du +\biggr(\int_{\{\vert u \vert >1\}} u^{-2}\, du \biggr)^{\tfrac{1}{2}}\biggr( \int_{\{\vert u \vert >1\}}( u^p g_{jl}(u))^2 \, du\biggr)^{\tfrac{1}{2}}
\end{align*}
which is finite since $u \mapsto u^p g_{jl}(u)\in L^2$, according to the proof of Proposition~\ref{gExistence}(\ref{gMoments}), and hence we have shown the last part of the result.
\end{proof}
We now give a result that both will be used to prove the uniqueness part of Theorem~\ref{existence} and Theorem~\ref{Prediction}.
\begin{lemma}\label{variationOfConstants} Suppose that $\det (h(iy)) \neq 0$ for all $y \in \mathbb{R}$ and that $\eta$ is a finite measure with second moment, and let $g$ be given by \eqref{gKernelChar1}. Furthermore, let $(X_t)_{t\in \mathbb{R}}$ be a measurable process, which is bounded in $L^1(\mathbb{P})$ and satisfies \eqref{MSDDE} almost surely for all $s<t$. Then, for each $s\in \mathbb{R}$ and almost surely,
\begin{align}\label{varOfConstants}
\begin{aligned}
X_t =&\, g(t-s) X_s + \int_s^\infty g(t-u) \, \eta \ast \big\{\mathds{1}_{(-\infty,s]}X\big\} (u)\, du\\
&+ g \ast \big\{\mathds{1}_{(s,\infty)}(Z-Z_s) \big\}(t)
\end{aligned}
\end{align}
for Lebesgue almost all $t>s$, using the notation
\begin{align*}
(\eta \ast \{\mathds{1}_{A}X\})_j (t) &:= \sum_{k=1}^n \int_{[0,\infty)} \mathds{1}_A (t-u)X^k_{t-u}\, \eta_{jk}(du) \quad \text{and} \\
(g\ast \{\mathds{1}_{(s,\infty)}(Z-Z_s) \})_j(t) &:=\sum_{k=1}^n \int_\mathbb{R}\mathds{1}_{(s,\infty)}(t-u)\big(Z^k_{t-u}-Z^k_s \big)\, g_{jk}(du)
\end{align*}
for $j=1,\dots, n$ and $t\in \mathbb{R}$.
\end{lemma}
\begin{proof
By arguments similar to those in the proof of Proposition~\ref{gExistence}(\ref{gExponential}) we get that the assumption $\det (h(iy)) \neq 0$ for all $y \in \mathbb{R}$ implies that we can choose $\delta\in (0,\varepsilon)$, such that $\det (h(z))\neq 0$ for all $z \in \mathbb{C}$ with $-\delta<\text{Re}(z) \leq 0$ and
\begin{align*}
\sup_{x \in (-\delta ,0)} \int_\mathbb{R}\big\vert \big(h(x+iy)^{-1}\big)_{jk}\big\vert^2\, dy<\infty.
\end{align*}
for all $j,k=1,\dots, n$. Thus, \cite[Lemma~5.1]{contARMAframework} ensures that $\mathcal{L}[g](z) = h(z)^{-1}$ when $\text{Re}(z) \in (-\delta,0)$. From this point we will fix such $z$ and let $s\in \mathbb{R}$ be given. Since $(X_t)_{t\in\mathbb{R}}$ satisfies \eqref{MultiSDDEcompact},
\begin{align*}
\mathds{1}_{(s,\infty)}(t)X_t = \mathds{1}_{(s,\infty)}(t)X_s + \int_{-\infty}^t\mathds{1}_{(s,\infty)}(u)\, \eta \ast X (u)\, du + \mathds{1}_{(s,\infty)}(t)(Z_t-Z_s)
\end{align*}
for Lebesgue almost all $t\in \mathbb{R}$ outside a $\mathbb{P}$-null set (which is a consequence of Tonelli's theorem). In particular, this shows that
\begin{align*}
\MoveEqLeft -z\mathcal{L}[\mathds{1}_{(s,\infty)}X](z) \\
=& -z\biggr\{ X_s\mathcal{L}[\mathds{1}_{(s,\infty)}](z) + \mathcal{L}\biggr[\int_{-\infty}^\cdot \mathds{1}_{(s,\infty)}(u)\, \eta \ast X (u)\, du \biggr](z)\\
&+ \mathcal{L}[\mathds{1}_{(s,\infty)} (Z-Z_s)](z) \biggr\}\\
=& \mathcal{L}[X_s\delta_0(\cdot - s)](z) + \mathcal{L} [\mathds{1}_{(s,\infty)}\, \eta \ast X](z) - z\mathcal{L}[\mathds{1}_{(s,\infty)} (Z-Z_s)](z).
\end{align*}
By noticing that
\begin{align*}
\mathcal{L}[\mathds{1}_{(s,\infty)}\, \eta \ast X](z) &= \mathcal{L} \big[\mathds{1}_{(s,\infty)}\, \eta \ast \big\{\mathds{1}_{(-\infty,s]}X\big\}\big](z) + \mathcal{L} \big[ \eta \ast \big\{\mathds{1}_{(s,\infty)}X\big\}\big](z)\\
&= \mathcal{L} \big[\mathds{1}_{(s,\infty)}\, \eta \ast \big\{\mathds{1}_{(-\infty,s]}X\big\}\big](z) + \mathcal{L}[ \eta](z)\mathcal{L}[\big\{\mathds{1}_{(s,\infty)}X\big\}\big](z)
\end{align*}
it thus follows that
\begin{align*}
\MoveEqLeft h(z) \mathcal{L}[\mathds{1}_{(s,\infty)} X](z) \\
=& \mathcal{L}\big[X_s\delta_0(\cdot - s) + \mathds{1}_{(s,\infty)}\, \eta \ast \big\{\mathds{1}_{(-\infty,s]}X\big\} \big](z) - z \mathcal{L}[\mathds{1}_{(s,\infty)}(Z-Z_s)](z).
\end{align*}
(The reader should observe that since both $(X_t)_{t\in \mathbb{R}}$ and $(Z_t)_{t\in \mathbb{R}}$ are bounded in $L^1(\mathbb{P})$, the Laplace transforms above are all well-defined almost surely. We refer to the beginning of the proof of Theorem~\ref{existence} where details for a similar argument are given.) Now, using that $\mathcal{L}[g](z) = h(z)^{-1}$, we notice
\begin{align*}
-zh(z)^{-1}\mathcal{L}[\mathds{1}_{(s,\infty)}(Z-Z_s)](z) &= \mathcal{L}[g(du)](z) \mathcal{L}[\mathds{1}_{(s,\infty)}(Z-Z_s)](z) \\
&= \mathcal{L}\big[g \ast \big\{\mathds{1}_{(s,\infty)}(Z-Z_s) \big\}\big] (z),
\end{align*}
and thus
\begin{align*}
X_t =&\, g(t-s) X_s + \int_s^\infty g(t-u) \, \eta \ast \big\{\mathds{1}_{(-\infty,s]}X\big\} (u)\, du + g \ast \big\{\mathds{1}_{(s,\infty)}(Z-Z_s) \big\}
\end{align*}
for Lebesgue almost all $t>s$ with probability one.
\end{proof}
With Lemma~\ref{variationOfConstants} in hand we are now ready to prove the general result, Theorem~\ref{existence}, for existence and uniqueness of solutions to the MSDDE \eqref{MultiSDDEcompact}.
\begin{proof}[Proof of Theorem~\ref{existence}]
Fix $t\in \mathbb{R}$. The convolution in \eqref{solutionForm} is well-defined if $u \mapsto Z^T_{t-u}$ is $g^T$-integrable (by Corollary~\ref{gMeasure}) which means that $u \mapsto Z^k_{t-u}$ belongs to $L^1(\vert g_{jk}\vert(du))$ for all $j,k = 1,\dots, n$. Observe that, since $(Z^k_u)_{u \in\mathbb{R}}$ is integrable and has stationary increments, \cite[Corollary~A.3]{QOU} implies that there exists $\alpha,\beta>0$ such that $\mathbb{E}[\vert Z_u^k\vert]\leq \alpha + \beta \vert u \vert$ for all $u \in \mathbb{R}$. Consequently,
\begin{align*}
\mathbb{E}\biggr[\int_\mathbb{R}\vert Z_{t-u}^k \vert\, \mu (du) \biggr] \leq (\alpha + \beta \vert t \vert) \mu (\mathbb{R}) + \beta\int_{\mathbb{R}} \vert u \vert\, \mu (du) <\infty
\end{align*}
for any (non-negative) measure $\mu$ which has first moment. This shows that $u \mapsto Z^k_{t-u}$ will be integrable with respect to such measure almost surely, in particular with respect to $\vert g_{jk}\vert (du)$, $j=1,\dots ,n$, according to Corollary~\ref{gMeasure} as $\eta$ has second moment.
We will now argue that $(X_t)_{t\in \mathbb{R}}$ defined by \eqref{solutionForm} does indeed satisfy \eqref{MultiSDDEcompact}, and thus we fix $s<t$. Due to the fact that
\begin{align*}
\int_s^t X^T\ast \eta^T (u)\, du = \int_s^t Z^T\ast \eta^T (u)\, du + \int_s^t \biggr(\int_\mathbb{R} g\ast \eta (r) Z_{\cdot - r}\, du\biggr)^T\ast \eta^T (u)\, du
\end{align*}
it is clear by the definition of $(X_t)_{t\in \mathbb{R}}$ that it suffices to argue that
\begin{align*}
\MoveEqLeft\int_s^t \biggr(\int_\mathbb{R} g\ast \eta (r) Z_{\cdot - r}\, du\biggr)^T\ast \eta^T (u)\, du \\
&= \int_\mathbb{R} Z^T_r[g\ast \eta (t-r) - g\ast \eta (s-r)]^T\, dr - \int_s^tZ^T\ast \eta^T(r)\, dr.
\end{align*}
We do this componentwise, so we fix $i \in \{1, \dots, n\}$ and compute that
\begin{align*}
\MoveEqLeft\biggr(\int_s^t \biggr(\int_\mathbb{R}g \ast \eta (r)Z_{\cdot - r}\, dr \biggr)^T\ast \eta^T(u)\, du \biggr)_i\\
=& \sum_{j=1}^n\sum_{k=1}^n\sum_{l=1}^n \int_s^t \biggr( \int_\mathbb{R}g_{jl}\ast \eta_{lk}(v)Z^k_{\cdot - r}\, dr \biggr)\ast \eta_{ij}(u)\, du\\
=& \sum_{j=1}^n\sum_{k=1}^n\sum_{l=1}^n\int_\mathbb{R}Z^k_r\int_{[0,\infty)}\int_s^t \int_{[0,\infty)} g_{jl}(u-v - r - w)\, \eta_{ij}(dv)\, du\, \eta_{lk}(dw)\, dr\\
=& \sum_{k=1}^n\sum_{l=1}^n \int_\mathbb{R} Z_r^k \int_{[0,\infty)}\int_s^t (g\ast \eta)_{il}(u-r-w)\, du\, \eta_{lk}(dw)\, dr\\
=& \sum_{k=1}^n\sum_{l=1}^n \biggr(\int_\mathbb{R} Z^k_r \int_{[0,\infty)}[g_{il}(t-r-w) - g_{il}(s-r-w)]\, \eta_{lk}(dw)\, dr\\
& -\int_\mathbb{R} Z^k_r \int_{[0,\infty)}\delta_0(\{i-l\})\mathds{1}_{(s,t]}(r+w)\, \eta_{lk}(dw)\, dr \biggr)\\
=& \sum_{k=1}^n \biggr(\int_\mathbb{R}Z^k_r [(g \ast \eta)_{ik}(t-r)-(g \ast \eta)_{ik}(s-r)]\, dr- \int_s^t Z^k\ast \eta_{ik} (r)\, dr \biggr)\\
=& \biggr(\int_\mathbb{R} Z_r^T [g\ast \eta (t-r)-g \ast \eta (s-r)]^T\, dr -\int_s^t Z^T\ast \eta^T (r)\, dr \biggr)_i
\end{align*}
where we have used \eqref{functionRelation} in Proposition~\ref{gExistence} and the fact that $g$ and $\eta$ commute in a convolution sense, $g\ast \eta = (g^T \ast \eta^T)^T$ (compare the associated Fourier transforms).
Next, we need to argue that $(X_t)_{t\in\mathbb{R}}$ is stationary. Here we will use \eqref{minusIdentity} to write the solution as
\begin{align*}
X_t = \int_\mathbb{R}g\ast \eta (u)\, [Z_{t-u}-Z_t]\, du
\end{align*}
for each $t\in \mathbb{R}$. Fix $m \in \mathbb{R}$. Let $-m=t^k_0<t^k_1<\dots<t^k_k = m$ be a partition of $[-m,m]$ with $\max_{j=1,\dots, k}( t^k_j-t^k_{j-1})\to 0$, $k \to \infty$, and define the Riemann sum
\begin{align*}
X^{m,k}_t = \sum_{j=1}^k g\ast \eta (t^k_{j-1})\, [Z_{t-t^k_{j-1}}-Z_t]\, (t^k_j-t^k_{j-1}).
\end{align*}
Observe that $(X^{m,k}_t)_{t\in \mathbb{R}}$ is stationary. Moreover, the $i$-th component of $X_t^{m,k}$ converges to the $i$-th component of
\begin{align*}
X^m_t= \int_{-m}^m g \ast \eta (u)\, [Z_{t-u}-Z_t]\, du
\end{align*}
in $L^1(\mathbb{P})$ as $k\to \infty$. To see this, we start by noting that
\begin{align*}
\mathbb{E}\big[\big\vert \big(X^m_t\big)_i - \big(X^{m,k}_t\big)_i \big\vert\big] \leq &
\sum_{j=1}^n \int_\mathbb{R}\sum_{l=1}^k\mathds{1}_{(t^k_{l-1},t^k_l]}(u)\mathbb{E}\Big[\big\vert (g\ast \eta)_{ij}(u) \big[Z^j_{t-u}-Z^j_t \big]\\&-(g\ast \eta)_{ij}\big(t_{l-1}^k\big)\big[Z^j_{t-t_{l-1}^k}-Z_t^j \big] \big\vert \Big]\, du.
\end{align*}
Then, for each $j \in \{1,\dots, n\}$,
\begin{align*}
\max_{l=1,\dots, k}\mathds{1}_{(t^k_{l-1},t_l^k]}(u)\MoveEqLeft\mathbb{E}\Big[\big\vert (g\ast \eta)_{ij}(u) \big[Z^j_{t-u}-Z^j_t \big]-(g\ast \eta)_{ij}\big(t_{l-1}^k\big)\big[Z^j_{t-t_{l-1}^k}-Z_t^j \big] \big\vert \Big]\\
\leq & \max_{l=1,\dots, k}\mathds{1}_{(t_{l-1}^k,t_l^k]}(u) \Big(\vert(g\ast \eta)_{ij}(u)\vert\, \mathbb{E}\big[\big\vert Z^j_{t-u} - Z^j_{t-t_{l-1}^k}\big\vert\big]\\
&+ \mathbb{E}\big[\big\vert Z^j_{t-t_{l-1}^k} - Z^j_t \big\vert \big]\, \big\vert(g\ast \eta)_{ij}(u) - (g\ast \eta)_{ij}\big(t_{l-1}^k\big)\big\vert \Big) \to 0
\end{align*}
as $k \to \infty$ for almost all $u \in\mathbb{R}$ using that $(Z^j_t)_{t\in \mathbb{R}}$ is continuous in $L^1(\mathbb{P})$ (cf. \cite[Corollary~A.3]{QOU}) and that $(g\ast \eta)_{ij}$ is càdlàg. Consequently, Lebesgue's theorem on dominated convergence implies that $X^{m,k}_t \to X^m_t$ entrywise in $L^1(\mathbb{P})$ as $k\to \infty$, thus $(X^m_t)_{t\in \mathbb{R}}$ inherits the stationarity property from $(X^{m,k}_t)_{t\in \mathbb{R}}$. Finally, since $X^m_t \to X_t$ (entrywise) almost surely as $m \to\infty$, we obtain that $(X_t)_{t\in \mathbb{R}}$ is stationary as well.
To show the uniqueness part, we let $(U_t)_{t \in \mathbb{R}}$ and $(V_t)_{t \in \mathbb{R}}$ be two stationary, integrable and measurable solutions to \eqref{MultiSDDEcompact}. Then $X_t := U_t - V_t$, $t\in \mathbb{R}$, is bounded in $L^1(\mathbb{P})$ and satisfies an MSDDE without noise. Consequently, Lemma~\ref{variationOfConstants} implies that
\begin{align*}
X_t = g(t-s) X_s + \int_s^\infty g(t-u) \eta \ast \{ \mathds{1}_{(-\infty,s]}X\}(u)\, du
\end{align*}
holds for each $s \in \RR$ and Lebesgue almost all $t >s$. For a given $j$ we thus find that
\begin{align*}
\mathbb{E}\big[\big\vert X^j_t\big\vert\big] &\leq C\sum_{k=1}^n\biggr(\vert g_{jk}(t-s) \vert
+ \sum_{l=1}^n \int_s^\infty \vert g_{jk}(t-u)\vert\, \vert \eta_{kl}\vert ([u-s,\infty))\, du
\biggr)
\end{align*}
where $C:= \max_k \mathbb{E}[\vert U_0^k\vert + \vert V_0^k\vert]$. It follows by Proposition~\ref{gExistence}(ii) that $g(t)$ converges as $t \to \infty$, and since $g \in L^1$ it must be towards zero. Using this fact together with Lebesgue's theorem on dominated convergence it follows that the right-hand side of the expression above converges to zero as $s$ tends to $-\infty$, from which we conclude that $U_t = V_t$ almost surely for Lebesgue almost all $t$. By continuity of both processes in $L^1(\mathbb{P})$ (cf. \cite[Corollary~A.3]{QOU}), we get the same conclusion for all $t$.
Finally, under the assumption that $\det(h(z)) \neq 0$ for $z \in \CC$ with $\Real(z) \leq 0$ it follows from Proposition~\ref{gExistence}(\ref{gCausality}) that $g\ast \eta$ is vanishing on $(-\infty,0)$, and hence we get that the solution $(X_t)_{t\in \mathbb{R}}$ defined by \eqref{solutionForm} is causal since
\begin{align*}
X_t = Z_t + \int_0^\infty g \ast \eta (u)\, Z_{t-u}\, du = - \int_0^\infty g\ast \eta (u) [Z_t - Z_{t-u}]\, du
\end{align*}
for $t\in \mathbb{R}$ by \eqref{minusIdentity}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{Prediction}]
Since $(X_t)_{t \in \RR}$ is a solution to an MSDDE,
\begin{align*}
\sigma (X_u : u \leq s) = \sigma (Z_s -Z_u : u \leq s)
\end{align*}
and the theorem therefore follows by Lemma~\ref{variationOfConstants}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{MArep}]
We start by arguing why \eqref{FubiniRelation} is well-defined. To see that this is the case, note initially that $I_k(f_r(t-\cdot)) = Z_t^k-Z_{t-r}^k$ and thus, since $(Z^k_t)_{t\in \mathbb{R}}$ is integrable and has stationary increments, there exists $\alpha,\beta >0$ such that $\mathbb{E}[\vert I_k(f_r(t-\cdot))\vert]\leq \alpha + \beta \vert r \vert$ for all $r \in \mathbb{R}$ (see, e.g., \cite[Corollary~A.3]{QOU}). In particular
\begin{align*}
\mathbb{E}\biggr[\int_\mathbb{R}\vert I_k(f_r(t-\cdot))\vert\, \vert \mu \vert (dr) \biggr] \leq \alpha \vert \mu \vert (\mathbb{R}) + \beta \int_\mathbb{R} \vert r \vert \, \vert \mu \vert (dr) < \infty,
\end{align*}
which shows that $I_k(f_r(t-\cdot))$ is integrable with respect to $\mu$, thus the right-hand side of \eqref{FubiniRelation} is well-defined, almost surely for each $t \in \mathbb{R}$. To show that the left-hand side is well-defined, it suffices to note that $u \mapsto \int_\mathbb{R} f_r(u)\, \mu (dr)$ belongs to $L^1 \cap L^2$ by an application of Jensen's inequality and Tonelli's theorem.
To show \eqref{MArelation} we start by fixing $t \in \mathbb{R}$ and $j,k \in \{1,\dots, n\}$, and by noting that $\mu (dr) = (g\ast \eta)_{jk}(r)\, dr$ is a finite measure with having first moment according to Corollary~\ref{gMeasure}. Consequently, we can use assumptions (i)-(ii) on $I_k$ to get
\begin{align*}
\int_\mathbb{R} (g\ast \eta)_{jk}(r)\big[Z^k_{t-r}-Z^k_t \big]\, dr &= \int_\mathbb{R} I_k(\mathds{1}_{(t,t-r]})(g\ast \eta)_{jk}(r)\, dr\\
&=I_k \biggr(\int_\mathbb{R} \mathds{1}_{(t,t-r]} (g\ast \eta)_{jk}(r)\, dr \biggr)\\
&= I_k\biggr(\delta_0(\{j-k\})\mathds{1}_{[0,\infty)}(t-\cdot)+\int_{-\infty}^{t-\cdot} (g\ast \eta)_{jk}(u)\, du \biggr)\\
&= I_k (g_{jk}(t-\cdot))
\end{align*}
using \eqref{gStieltjes} and the convention that $\mathds{1}_{(a,b]} = -\mathds{1}_{(b,a]}$ when $a>b$. By combining this relation with \eqref{minusIdentity} and \eqref{solutionForm} we obtain
\begin{align*}
X^j_t = \sum_{k=1}^n \int_\mathbb{R} (g\ast \eta)_{jk}(r) [Z^k_{t-r}-Z^k_t]\, dr = \sum_{k=1}^n I_k (g_{jk}(t-\cdot)).
\end{align*}
\end{proof}
\begin{proof}[Proof of Proposition \ref{RiemannReg}]
Let $\alpha \in (1,2]$ and $\beta \in (0,1-1/\alpha)$, and consider a function $f:\mathbb{R}\to \mathbb{R}$ in $L^1\cap L^{\alpha}$. We start by noticing that
\begin{align*}
\int_t^\infty \vert f(u) \vert (u-t)^{\beta -1}\, du = \int_0^1 \vert f(t+u) \vert u^{\beta -1}\, du + \int_1^\infty \vert f(t+u) \vert u^{\beta -1}\, du.
\end{align*}
For the left term we find that
\begin{align*}
\MoveEqLeft \int_\RR \biggr( \int_0^1 \vert f(t+u)\vert u^{\beta -1}\, du \biggr) ^\alpha\, dt \\
&\leq \biggr( \int_0^1 u^{\beta -1}\, du\biggr)^{\alpha-1} \int_\RR \int_0^1 \vert f(t+u)\vert ^\alpha u^{\beta -1}\, du\, dt\\
& = \biggr(\int_0^1 u^{\beta -1}\, du\biggr)^\alpha \int_\RR \vert f(t) \vert^ \alpha\, dt < \infty.
\end{align*}
For the right term we find
\begin{align*}
\MoveEqLeft \int_\RR \biggr(\int_1^\infty \vert f(t+u) \vert u^{\beta -1}\, du \biggr)^\alpha\, dt \\
&\leq \biggr( \int_\RR f(u) du \biggr)^{\alpha-1} \int_\RR \int_1^\infty \vert f(t+u) \vert u^{\alpha(\beta -1)}\, du\, dt \\
& = \biggr( \int_\RR f(u) du \biggr)^{\alpha} \int_1^\infty u^{\alpha(\beta -1)}\, du <\infty.
\end{align*}
We conclude that $\big(I^{\beta}_-f\big)(u) \in L^{\alpha}$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{simpleFourierKernel}]
The identity \eqref{higherOrderDetCond} is just a matter of applying standard computation rules for determinants. For instance, one may prove the result when $z\neq 0$ by induction using the block representation
\begin{align}\label{blockH}
-h(z) = \begin{bmatrix} A & B \\ C & D \end{bmatrix}
\end{align}
with $A=I_nz$, $B= (e_1 \otimes I_n)^T \in \mathbb{R}^{n \times (m-1)n}$, $C= e_{m-1}\otimes \mathcal{L}[\varpi_0](z) \in \mathbb{R}^{(m-1)n \times n}$, and
\begin{align*}
D= \begin{bmatrix}
I_nz & I_n & 0 & \cdots & 0 \\
0 & I_n z & I_n & \cdots & 0 \\
\vdots & \vdots & \ddots & \ddots & \vdots \\
0 & 0 & \cdots & I_n z & I_n \\
\mathcal{L}[\varpi_1](z) & \mathcal{L}[\varpi_2](z) & \cdots & \mathcal{L}[\varpi_{m-2}](z) & I_nz +\mathcal{L}[\varpi_{m-1}](z)
\end{bmatrix}.
\end{align*}
Here $e_1$ and $e_{m-1}$ refer to the firs and last canonical basis vector of $\mathbb{R}^{m-1}$, respectively. The case where $z=0$ follows directly from the Leibniz formula. In case $\det (h(iy))\neq 0$ for all $y \in \mathbb{R}$, we may write $h(iy)^{-1}$ as an $m \times m$ matrix, where each element $(h^{-1}(iy))_{jk}$ is an $n \times n$ matrix. We then know from Theorem~\ref{existence} that the unique solution to \eqref{hOrderSDDE} is a $(Z_t)_{t\in \mathbb{R}}$-driven moving average of the form \eqref{solutionTohOrder} with $\mathcal{F}[g_{1m}](y) = (h^{-1}(iy))_{1m}$. Similar to the computation of $\det (h(z))$, when $h(z)$ is invertible, block $(1,m)$ of $h(z)^{-1}$ can inductively be shown to coincide with
\begin{align*}
\biggr( I_n (-z)^m- \sum_{j=0}^{m-1} \mathcal{L}[\varpi_j] (z) (-z)^j\biggr)^{-1}
\end{align*}
using the representation \eqref{blockH} and standard rules for inverting block matrices. This means in particular that \eqref{hOrderFourierT} is true.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{MCARMAasMSDDE}]
We start by arguing that that there exists a function $f$ with the Fourier transform in \eqref{fFourier}. Note that, since $z \mapsto \det (Q(z))$ is just a polynomial (of order $nq$), the assumption that $\det (Q(z))\neq 0$ whenever $\text{Re}(z)\geq 0$ implies in fact that
\begin{align*}
H(z) := R(-z)-Q(-z)^{-1}P(-z) = Q(-z)^{-1}[Q(-z)R(-z) - P(-z)]
\end{align*}
is well-defined for all $z \in \mathcal{S}_\delta := \{x+iy\, :\, x \leq \delta,\, y \in \mathbb{R}\}$ and a suitably chosen $\delta>0$. According to \cite[Lemma~5.1]{contARMAframework} it suffices to argue that there exists $\varepsilon\in (0,\delta]$ such that
\begin{align}\label{keyfExist}
\sup_{x <\varepsilon} \int_\mathbb{R} \vert H(x+iy)_{jk}\vert^2\, dy< \infty
\end{align}
for all $j,k=1,\dots, n$. Let $\lVert \cdot \rVert$ denote any sub-multiplicative norm on $\mathbb{C}^{n\times n}$ and note that $\vert H(z)_{jk} \vert \leq \lVert Q(-z)^{-1} \rVert \lVert Q(-z)R(-z) - P(-z) \rVert$. Thus, since $\lVert Q(z)R(z) - P(z) \rVert \sim c_1\vert z \vert^{q-1}$ and $\lVert Q(z)^{-1} \rVert \sim c_2\vert z\vert^{-q}$ as $\vert z \vert \to \infty$ for some $c_1,c_2\geq 1$ (the former by the choice of $R$ and the latter by Cramer's rule), $\vert H(z)_{jk}\vert = O(\vert z \vert^{-1})$. Consequently, the continuity of $H$ ensures that \eqref{keyfExist} is satisfied for a suitable $\varepsilon\in (0,\delta]$, and we have established the existence of $f$ with the desired Fourier transform. This also establishes that the $n\times n$ measures $\varpi_0,\varpi_1,\dots, \varpi_{p-q-1}$ defined as in the statement of the theorem are finite and have moments of any order. Associate to these measures the $n(p-q)\times n(p-q)$ measure $\eta$ given in \eqref{etaHigherOrder}. Then it follows from \eqref{higherOrderDetCond} that
\begin{align*}
\det (h(iy)) = \det\biggr(I_n(-iy)^{p-q} + \sum_{j=0}^{p-q-1} R_j (-iy)^j - \mathcal{F}[f](y)\biggr) = \frac{\det (P(-iy))}{\det (Q(-iy))},
\end{align*}
and hence is non-zero for all $y \in \mathbb{R}$. In light of Proposition~\ref{simpleFourierKernel}, in particular \eqref{hOrderFourierT}, we may therefore conclude that the unique solution to \eqref{hOrderSDDE} is a $(Z_t)_{t\in \mathbb{R}}$-driven moving average, where the driving kernel has Fourier transform
\begin{align*}
\biggr(I_n (-iy)+ \sum_{j=0}^{p-q-1} R_j (-iy)^j - \mathcal{F}[f](y) \biggr)^{-1} = P(-iy)^{-1}Q(-iy)
\end{align*}
for $y \in \mathbb{R}$. In other words, the unique solution is the $(Z_t)_{t\in \mathbb{R}}$-driven MCARMA($p,q$) process associated to the polynomials $P$ and $Q$.
\end{proof}
Before giving the proof of Corollary \ref{CARMAprediction} we will need the following lemma:
\begin{lemma}\label{FirstRowOfg}
Let $C_0,\dots,C_{p-q-1}$ be given in \eqref{residuePolyn} and $C_{p-q} =I_n$. Define
\begin{align*}
R_j(z) = \sum_{k = j}^{p-q} C_k z^{k-j} , \quad j=1,\dots,p-q-1.
\end{align*}
Then $\tilde{g}$ is $p-q-2$ times differentiable and $D^{p-q-2}\tilde{g}$ has a density with respect to the Lebesgue measure which we denote $D^{p-q-1}\tilde{g}$. Furthermore, we have that
\begin{align}\label{Firstthingtoprove}
(e_1^{p-q} \otimes I_n)^T g = (\tilde{g} R_1(D),\dots,\tilde{g} R_{p-q-1}(D),\tilde{g} )
\end{align}
where
\begin{align}\label{Secondthingtoprove}
\begin{aligned}
\tilde{g} R_j(D)(t) &= \sum_{k = j}^{p-q} D^{k-j} \tilde{g} (t) C_k \\
& = \mathds{1}_{[0,\infty)}(t) (e_1^p\otimes I_n)^T e^{At} \sum_{k=j}^{p-q} A^{k-j} E C_k
\end{aligned}
\end{align}
for $j=1,\dots,p-q-1$ and $g : \RR \to \RR^{n\times n}$ is characterized by $\FF[g](y)=h(iy)^{-1}$ with $h: \CC \to \CC^{n(p-q)\times n(p-q)}$ given by
\begin{align*}
h(-z) = \begin{bmatrix}
I_n z & -I_n & 0 & \cdots & 0 \\
0 & I_n z & -I_n & \cdots & 0 \\
\vdots & \vdots & \ddots & \ddots & \vdots \\
0 & 0 & \cdots & I_n z & -I_n\\
Q^{-1}(z)P(z) -zR_1(z) & C_1 & \cdots & C_{p-q-2} & I_n z+C_{p-q-1}
\end{bmatrix}.
\end{align*}
\end{lemma}
\begin{proof}
That $\tilde{g}$ is $p-q-2$ times differentiable and $D^{p-q-2}\tilde{g}$ has a density with respect to the Lebesgue measure follows form the relation in \eqref{gStieltjes}. Furthermore, by Theorem~\ref{MCARMAasMSDDE} we know that $\FF[\tilde{g} ](y) = P^{-1}(-iy)Q(-iy)$. Consequently, \eqref{Firstthingtoprove} follows since
\begin{align*}
\MoveEqLeft (P^{-1}(-iy)Q(-iy)R_1(-iy),\\
&\dots,P^{-1}(-iy)Q(-iy)R_{p-q-1}(-iy),P^{-1}(-iy)Q(-iy) )h(z) = (e_1^{p-q} \otimes I_n)^T.
\end{align*}
The relation in \eqref{Secondthingtoprove} follows by the representation of $\tilde{g}$ given in \eqref{gMCARMA}.
\end{proof}
\begin{proof}[Proof of Corollary \ref{CARMAprediction}]
The prediction formula is a consequence of Lemma~\ref{FirstRowOfg} combined with Theorem~\ref{Prediction} and Theorem~\ref{MCARMAasMSDDE}. Furthermore, to get the expression for $\tilde{g} \ast \{\hat{Z}\mathds{1}_{(s,\infty)} \}$, note that
\begin{align*}
\tilde{g} (dv) = \mathds{1}_{\{ p=q+1 \} } \delta_0 (dv) +(e_1^p \otimes I_n)^Te^{Av}AE \, dv,
\end{align*}
which follows from the representation of $\tilde{g}$ in \eqref{gMCARMA}
\end{proof}
\subsection*{Acknowledgments}
This work was supported by the Danish Council for Independent Research (Grant DFF - 4002 - 00003).
\bibliographystyle{chicago}
|
\section{Introduction}
The purpose of this article is to study regularity properties of solutions to degenerate-parabolic equations in non-divergence form, whose prototypical example is given by
\begin{equation}\label{prototype}
\mathcal{K} := \mathrm{tr}\left(A(v,y,t)D^2_v\right)+\left\langle v, \nabla_y \right\rangle - \partial_t, \quad\mbox{for }(v,y,t)\in\mathbb{R}^d\times\mathbb{R}^d\times\mathbb{R},
\end{equation}
where the $d\times d$ matrix $A(v,y,t)$ is uniformly positive definite. The case $A = \mathbb{I}_d$ corresponds to the well-known Kolmogorov equation \cite{Kol}, which governs the probability density of a system with $2d$ degrees of freedom. The Kolmogorov operator is one of the key examples of hypoelliptic operators studied by H\"ormander in his seminal work \cite{Ho}. Operators like \eqref{prototype} also appear naturally in mathematical finance and various other stochastic models \cite{Ba, Ch, LPP}. Most notably perhaps, the operator $\mathcal{K}$ (and its divergence form counterpart) arises in the kinetic theory of gases as the leading order term in the spatially inhomogeneous Landau equation, which can be interpreted as the limit of the Boltzmann equation when only grazing collisions are taken into account \cite{AV04, Li}. For more on recent progress in the regularity theory of kinetic equations, we refer the reader to the works \cite{CSS, GIMV, GG, IS} and references therein.
In the Landau equation, the coefficients of the principal part depend on the solution itself in a nonlocal manner. This motivates the study of regularity properties of $\mathcal{K}$ with minimal assumptions on the smoothness of $A$. For the divergence-form version of $\mathcal{K}$,
$$\text{div}_v\left(A(v,y,t)\nabla_v\right)+\left\langle v, \nabla_y \right\rangle - \partial_t,$$
with bounded measurable coefficients $A$, a Moser-type $L^2$-to-$L^{\infty}$ iteration was obtained in \cite{PP04}, a H\"older regularity result for the solutions was shown in \cite{WZ}, and the Harnack inequality has been proved recently in \cite{GIMV}. Related regularity estimates for a more general class of divergence-form operators with rough coefficients can be found in \cite{CPP, aLP, aLPP}.
For the non-divergence form operator \eqref{prototype} with $A$ assumed to be merely bounded and measurable, the analogue of the Krylov-Safonov Harnack inequality \cite{KS80} is still unknown. This is primarily due to the lack of a suitable version of the Aleksandrov-Bakelman-Pucci maximum principle; we refer to \cite[Chapter VII]{Lieberman} for the uniformly parabolic case. On the other hand, with $A$ assumed to be H\"older continuous, the regularity theory is well-settled and several results have been obtained, even for operators with more general drift terms: we mention, among others, the results concerning the existence of the fundamental solution via Levi-parametrix methods, two-sided Gaussian-type bounds, and also Harnack inequalities \cite{DM, DiP, Poli}.
In this work, we prove Harnack's inequality for non-negative solutions to $\mathcal{K}u = 0$ under either a Cordes-Landis condition or a continuity assumption on the coefficient matrix $A$ (see subSection \ref{results}, hypotheses \hyperref[H1]{H1} and \hyperref[H2]{H2}). Similar results have been obtained for other H\"ormander type operators, namely for non-divergence form operators structured on Heisenberg vector fields \cite{AGT, Gutierrez-Tournier-Harnack, Tralli-Critical-Density}.
The techniques we employ in the present work are inspired by the insightful contributions of Landis from the '60s \cite{Landisell}, where he obtained what is nowadays referred to as the \emph{growth lemma} for nonnegative subsolutions of uniformly elliptic equations, assuming that the eigenvalue ratio is close to 1. Glagoleva \cite{glago} established analogous results for uniformly parabolic equations. We refer the reader to the book \cite{Landis} for an exposition of these ideas. In accordance with the literature on ultraparabolic equations, we present our results for operators more general than $\mathcal{K}$ which enjoy invariance properties with respect to a homogeneous Lie group structure. We proceed to describe these operators in more detail and state our main results.
\subsection{Main Results}\label{results}
Fix $N\in\mathbb{N}$. Throughout the paper we denote by $z=(x,t)\in\mathbb{R}^N\times\mathbb{R}$ a generic point in $\mathbb{R}^{N+1}$. The spatial differential operators will be denoted $\nabla=\nabla_x$, $D^2=D^2_x$. Fix $p_0, n \in \mathbb{N}$, with $1\leq p_0< N$ and $n\geq 1$. Let $\mathbb{I}_{p_0}$ denote the $p_0 \times p_0$ identity matrix. For some open set $\Omega \subseteq \mathbb{R}^{N+1}$, we consider the class of operators
\begin{equation}\label{varintroL}
\mathcal{L}_A=\mathrm{tr}\left( A(z) D^2\right) +\left\langle x, B\nabla\right\rangle - \partial_t \qquad z \in \Omega,
\end{equation}
where $A(z)\in\mathbb{R}^{N\times N}$ is a symmetric nonnegative definite matrix which takes the block form
\begin{equation}\label{A}
A(z)=
\begin{bmatrix}
\mathbb{A}(z) & 0\\
0 & 0
\end{bmatrix}\quad\mbox{ with }\quad \mathbb{A}(z) \in \mathbb{R}^{p_0 \times p_0},
\end{equation}
and
\begin{equation}\label{B}
B=
\begin{bmatrix}
0 & \mathbb{B}_{1} & 0 & \ldots & 0 \\
0 & 0 & \mathbb{B}_{2} & \ldots & 0 \\
\vdots & \vdots & \ddots & \ddots & \vdots\\
0 & 0 & 0 & \ldots & \mathbb{B}_{n} \\
0 & 0 & \ldots & 0 & 0
\end{bmatrix} \in \mathbb{R}^{N \times N}
\end{equation}
where, for $j = 1, \ldots, n$, $\mathbb{B}_{j}$ is a $p_{j-1}\times p_{j}$ block of rank $p_{j}$, $p_{0}\geq p_{1}\geq \ldots \geq p_{n}\geq1$ and $p_{0}+p_{1}+ \ldots +p_{n}=N$. The matrix $\mathbb{A}(z)$ is assumed to be uniformly positive definite; that is, there exist constants $\lambda, \Lambda > 0$ such that
\begin{equation}\label{ellipticity}
0<\lambda\mathbb{I}_{p_0}\leq \mathbb{A}(z) \leq \Lambda\mathbb{I}_{p_0} \qquad\mbox{ for all }z \in \Omega.
\end{equation}
Notice that the class of operators \eqref{prototype} corresponds to the choices $N=2d, p_0=d, n=1, \mathbb{B}_1=\mathbb{I}_d$.
The conditions on $A(\cdot)$ and $B$ endow the operators $\mathcal{L}_A$ with rich algebraic properties. As a matter of fact, in the case of a constant matrix $\mathbb{A}$, the operator is of H\"ormander type and the fundamental solution can be written explicitly \cite{Ho, K82}. Moreover, it has been shown in \cite{LP} that this operator is invariant under the action of a homogeneous Lie group, with homogeneous dimension
\begin{equation}\label{defQ}
Q+2:=p_{0}+ 3 p_{1}+...+ (2n +1 ) p_{n} + 2.
\end{equation}
We remark that the presence of homogeneity is tied to the upper triangular form \eqref{B} of the matrix $B$. The group structure allows one to define a homogeneous norm and corresponding cylinder-like sets
$$\mathcal{Q}^{t_1,t_2}_{r}(z_0), \quad\mbox{for }z_0\in \mathbb{R}^{N+1},\, t_1,t_2\in\mathbb{R},\, r>0.$$
We refer to Section \ref{notcon} for a precise description of all these notions.
To establish Harnack's inequality for the aforementioned operators \eqref{varintroL}, we will assume that the matrix coefficients $A(\cdot)$ satisfy either one of the following hypotheses:
\begin{itemize}
\item[(H1)]\label{H1} {\bf Cordes-Landis assumption:} The coefficients $A(\cdot)$ satisfy the condition \eqref{ellipticity} with
$$\frac{\Lambda}{\lambda}<1+\frac{2}{Q}.$$
\item[(H2)]\label{H2} {\bf Uniform continuity in $\Omega$:} The coefficients $A(\cdot)$ admit a uniform modulus of continuity $\omega$ in $\Omega$ (see Definition \ref{modulusofcontinuity}).
\end{itemize}
We can now state our main results. Any constant that depends solely on $B, Q, n, \lambda, \Lambda$ will henceforth be referred to as a \emph{structural} constant.
\begin{theorem}\label{harnackineqH1}(Harnack Inequality under \hyperref[H1]{H1}) Suppose $\mathcal{L}_A$ satisfies the Cordes-Landis condition \hyperref[H1]{H1}. There exist structural constants $b_B, K, \sigma_0, C > 0$ with $K>\sigma_0$ such that, for all $\mathcal{Q}_{Kr}^{-b_B r^2,0}(z_0)\Subset\Omega$ and $u \in C^2(\Omega)$ satisfying
$$u\geq 0\quad \text{ and } \quad \mathcal{L}_A u = 0\quad\mbox{ in }\,\,\, \mathcal{Q}_{Kr}^{-b_B r^2,0}(z_0),$$
we have
\begin{equation}\label{harnackH1}
\sup_{\mathcal{Q}_r^-} u \leq C \inf_{\mathcal{Q}_r^+} u,
\end{equation}
where $\mathcal{Q}_r^- := \mathcal{Q}_{\frac{\sigma_0}{2}r}^{-\frac{3b_B}{4}r^2,-\frac{b_B}{2}r^2}(z_0)$ and $\mathcal{Q}_r^+ := \mathcal{Q}_{\frac{\sigma_0}{2}r}^{-\frac{b_B}{4}r^2,0}(z_0)$.
\end{theorem}
\begin{theorem}\label{harnackineqH2}(Harnack Inequality under \hyperref[H2]{H2}) Suppose $\mathcal{L}_A$ satisfies the uniform continuity assumption \hyperref[H2]{H2} in $\Omega$, with modulus of continuity $\omega$. There exist positive constants $b_B, K, \sigma, C > 0$ depending on $\omega$ and on structural constants such that, for all $\mathcal{Q}_{Kr}^{-b_B r^2,0}(z_0)\Subset\Omega$ with $0<r\leq 1$ and $u \in C^2(\Omega)$ satisfying
$$u\geq 0\quad \text{ and } \quad \mathcal{L}_A u = 0\quad\mbox{ in }\,\,\, \mathcal{Q}_{Kr}^{-b_B r^2,0}(z_0),$$
we have
\begin{equation}\label{harnackH2}
\sup_{\mathcal{Q}_r^-} u \leq C \inf_{\mathcal{Q}_r^+} u,
\end{equation}
where $\mathcal{Q}_r^- := \mathcal{Q}_{\sigma r}^{-\frac{3b_B}{4}r^2,-\frac{5b_B}{8}r^2}(z_0)$ and $\mathcal{Q}_r^+ := \mathcal{Q}_{\sigma r}^{-\frac{b_B}{8}r^2,0}(z_0)$.
\end{theorem}
We point out that Theorem \ref{harnackineqH1} is, to the best of our knowledge, the first regularity result for non-divergence form operators like $\mathcal{K}$ that is independent of the smoothness of the coefficients. Theorem \ref{harnackineqH2} also generalizes, in the case of the homogeneous operators \eqref{varintroL}, the Harnack's inequality obtained in \cite{DiP} assuming H\"older continuity of the coefficients.
The essential ingredients in the proof of Theorems \ref{harnackineqH1} and \ref{harnackineqH2} are, respectively, Theorem \ref{growthlemma} and Theorem \ref{growthlemmaH2}. These are the analogues of the classical growth lemma of Landis, and they establish pointwise-to-measure estimates for nonnegative subsolutions to $\mathcal{L}_A$ in a quantitative manner. In order to establish these key estimates, we construct barriers using the potentials generated by kernels resembling the fundamental solution for constant coefficient operators. This involves a careful estimate of the aforementioned kernels in terms of the length scale of the cylinders. It is only in the construction of these barriers where we use the hypotheses \hyperref[H1]{H1} and \hyperref[H2]{H2}. Once the required pointwise-to-measure estimates are established, there are, by now, standard ways in the literature to proceed with the proof of Harnack's inequality. In this work, we have chosen to follow the general approach outlined by Landis in \cite{Landis}. For this strategy to succeed, we must deal with the non-standard nature of the cylinder-like sets $\mathcal{Q}^{t_1,t_2}_{r}(z_0)$.
The outline of this paper is as follows. In Section \ref{notcon}, we set up our notation and recall some properties of the relevant geometric objects. In Section \ref{secest}, we establish upper and lower bounds for the kernels \eqref{defbar}. We then use these kernels in Section \ref{barriers} to construct barriers for $\mathcal{L}_A$ under the hypotheses \hyperref[H1]{H1} and \hyperref[H2]{H2} (see respectively subSections \ref{cordeslandiscondition} and \ref{continuityassumption}). In Section \ref{sec5}, we prove the growth lemmas (Theorems \ref{growthlemma} and \ref{growthlemmaH2}), and provide as application the oscillation decay and the H\"older continuity of solutions to $\mathcal{L}_A u = 0$. Finally, in Section \ref{harn} we complete the proofs of Theorems \ref{harnackineqH1} and \ref{harnackineqH2}.
\section{Preliminaries}\label{notcon}
The assumptions \eqref{B} on the matrix $B$ imply (see \cite[Section 2]{LP}) that the following H\"ormander's rank condition \cite{Ho} is satisfied;
$${\rm{rank\,\, Lie\, }}\left\{\partial_{x_1},\ldots,\partial_{x_{p_0}},\left\langle x, B\nabla\right\rangle-\partial_t\right\}(z)=N+1\quad\forall\,z\in\mathbb{R}^{N+1}.$$
In particular, for any constant matrix $A_0\in\mathbb{R}^{N\times N}$ with the block structure
\begin{equation}\label{azero}
A_0=
\begin{bmatrix}
\mathbb{A}_{0} & 0\\
0 & 0
\end{bmatrix}\quad\mbox{ with }\quad 0<\lambda\mathbb{I}_{p_0}\leq \mathbb{A}_{0}\leq \Lambda\mathbb{I}_{p_0},
\end{equation}
the operator
$$\mathcal{L}_0=\mathrm{tr}\left( A_0 D^2\right) +\left\langle x, B\nabla\right\rangle - \partial_t$$
is hypoelliptic. The operator $\mathcal{L}_0$ is the infinitesimal generator of a Gaussian process with covariance matrix given by
$$C_0(t)=\int_{0}^{t} E(\sigma)A_0E^T(\sigma)\,d\sigma,$$
where
\begin{equation}\label{matrixexponential}
E(\sigma) :=\exp\left(-\sigma B^T\right), \qquad \sigma\in\mathbb{R}.
\end{equation}
The H\"ormander rank condition is actually equivalent (see \cite{Ho, LP}) to the following Kalman-type condition
$$C_0(t)>0\qquad\forall\,t>0.$$
Throughout the paper we will use the notation
\begin{equation}\label{Kalma0}
I_0=\begin{bmatrix}
\mathbb{I}_{p_0} & 0\\
0 & 0
\end{bmatrix}\in \mathbb{R}^{N \times N} \quad\mbox{ and }\quad C(t)=\int_{0}^{t} E(\sigma)I_0E^T(\sigma)\,d\sigma .
\end{equation}
Then, the assumption \eqref{ellipticity} for the coefficient matrix $A(z)$ of the operators $\mathcal{L}_A$ in \eqref{varintroL}-\eqref{A} is clearly equivalent to assuming
\begin{equation}\label{ellipticityofA}
0<\lambda I_0\leq A(z) \leq \Lambda I_0 \qquad \text{ for all }z \in \Omega.
\end{equation}
Let us now describe the group structure mentioned in the Introduction. We refer the reader to \cite[Section 1]{LP} for a complete exposition. Recalling \eqref{matrixexponential}, the group law is given by
$$z \circ\zeta =\left( \xi+E(\tau) x,t+ \tau\right),\quad \mbox{for }z=\left( x,t\right), \zeta=\left( \xi,\tau\right)\in\mathbb{R}^{N+1}.$$
Moreover, recalling the $p_i$'s coming from the structure of $B$ in \eqref{B}, we can denote any $x \in \mathbb{R}^N$ as
$$x=\left(x^{(p_0)},x^{(p_1)},\ldots,x^{(p_n)}\right)\in \mathbb{R}^{p_0}\times\mathbb{R}^{p_1}\times\cdots\times\mathbb{R}^{p_n} = \mathbb{R}^N,$$
and we can define the family of group automorphisms $\left(\delta_r\right)_{r>0}$ as
\begin{eqnarray*}
\delta_r&:&\mathbb{R}^{N+1}\longrightarrow\mathbb{R}^{N+1}\\
\delta_r(x,t)&=&\left(r
x^{(p_0)},r^3
x^{(p_1)},\ldots,r^{2n+1}x^{(p_n)},r^2t\right).
\end{eqnarray*}
These will play the role of homogeneous dilations. For convenience, we also denote the spatial dilations by
$$D_r:\mathbb{R}^N\longrightarrow\mathbb{R}^N,\quad D_r(x) = \left(r
x^{(p_0)},r^3
x^{(p_1)},\ldots,r^{2n+1}x^{(p_n)}\right).$$
The fact that $\delta_r$ are automorphisms with respect to $\circ$ is encoded in the following commutation property (see \cite[equation (2.20)]{LP}, see also \cite{K82})
\begin{equation}\label{commu}
E(r^2\sigma)=D_rE(\sigma)D_{\frac{1}{r}}\quad\mbox{ for any }r>0\,\mbox{ and }\,\sigma\in\mathbb{R}.
\end{equation}
From this, one can deduce that the covariance matrix $C_0(t)$ satisfies the commutation relation
\begin{equation}\label{split}
C_0(t)= D_{\sqrt{t}} C_0(1) D_{\sqrt{t}}.
\end{equation}
If $Q$ is the number defined in \eqref{defQ} and $|\cdot|$ denotes Lebesgue measure (both in $\mathbb{R}^{N+1}$ and $\mathbb{R}^N$), then we have
\begin{equation}\label{invmeasure}
|\delta_r(E)|=r^{Q+2}|E|,\quad |D_r(F)|=r^{Q}|F|,\quad |z_0\circ E|=|E|
\end{equation}
for all $z_0\in\mathbb{R}^{N+1}$, $r>0$, and for any Lebesgue measurable sets $E\subset\mathbb{R}^{N+1}$, $F\subset\mathbb{R}^N$. In \cite{LP} it is shown that the vector fields $\left\{\partial_{x_1},\ldots,\partial_{x_{p_0}},\left\langle x, B\nabla\right\rangle-\partial_t\right\}$ are left-translation invariant and $\delta_r$-homogeneous (respectively of degree $1$ and $2$). Consequently, the operators $\mathcal{L}_0$ are left-translation invariant and $\delta_r$-homogenous of degree $2$.
One can associate to this homogeneous structure a family of cylinder-like sets. Denoting also the Euclidean norms in $\mathbb{R}^N$, $\mathbb{R}^{p_k}$ or $\mathbb{R}$ by $\left|\cdot \right|$, we can define the norms $\left|\cdot \right|_B:\mathbb{R}^N\longrightarrow\mathbb{R}^+$ and $\left\|\cdot\right\|_B:\mathbb{R}^{N+1}\longrightarrow\mathbb{R}^+$ by
$$|x|_B=\sum_{i=0}^n\left|x^{(p_i)}\right|^{\frac{1}{2i+1}},\qquad\mbox{for }x=\left(x^{(p_0)},\ldots,x^{(p_n)}\right)\in\mathbb{R}^{p_0}\times\cdots\times\mathbb{R}^{p_n}=\mathbb{R}^N,$$
$$\left\|z\right\|_B=|x|_B + |t|^{1/2}, \qquad\mbox{for } z = (x,t) \in \mathbb{R}^{N+1}.$$
The subscript $B$ is used to distinguish the homogeneous norm $\left\|\cdot \right\|_B$ from the matrix norm $\left\|\cdot \right\|$.
Note that $\left|\cdot \right|_B$ and $\left\|\cdot\right\|_B$ are respectively $D_r$-homogeneous and $\delta_r$-homogeneous functions of degree $1$.
The homogeneous ball of radius $r>0$ centered at $0$ is the set
$$B_{r}(0): =\left\{x\in\mathbb{R}^N\,:\,|x|_B < r\right\}=D_{r}\left(B_1(0)\right).$$
The cylinder-like sets centered at $0$ are defined as
$$\mathcal{Q}^{t_1, t_2}_{r} = B_{r}(0) \times (t_1, t_2)$$
where $r>0$ and $t_1<t_2\in\mathbb{R}$. Cylinder-like sets centered at an arbitrary point $z_0 \in \mathbb{R}^{N+1}$ are defined as
$$\mathcal{Q}^{t_1, t_2}_{r}(z_0): =z_0\circ \mathcal{Q}^{t_1, t_2}_{r}.$$
It is clear from \eqref{invmeasure} and the composition and dilation laws that, for any $b>0$,
$$\left|\mathcal{Q}^{t_1, t_1+br^2}_{r}(z_0)\right|=r^{Q+2}\left|\mathcal{Q}^{0, b}_{1}\right|\quad\mbox{ for all }z_0\in\mathbb{R}^{N+1}, t_1\in\mathbb{R}, r>0.$$
The notion of parabolic boundary of a cylinder can be naturally extended to this setting, and is defined as
$$\partial_p\mathcal{Q}^{t_1, t_2}_{r}:=\left(B_r(0)\times\{t_1\}\right)\cup \left(\partial B_r(0)\times [t_1,t_2] \right)\quad\mbox{ and }\quad\partial_p\mathcal{Q}^{t_1, t_2}_{r}(z_0):=z_0\circ \partial_p\mathcal{Q}^{t_1, t_2}_{r}.$$
It is easy to check that $\partial_p\mathcal{Q}^{t_1, t_2}_{r}(z_0)=\overline{\partial\mathcal{Q}^{t_1, t_2}_{r}(z_0)\cap\{t<t_2+t_0\}}$.
We can now state the analogue of the parabolic weak maximum principle for the operators $\mathcal{L}_A$ in \eqref{varintroL}, whose proof is, by now, classical for degenerate-parabolic equations.
\begin{center} Let $T\in\mathbb{R}$ and let $D\subset\mathbb{R}^{N+1}$ be a bounded open set; if $v\in C^2(D)\cap C\left(\overline{D}\right)$ satisfies
\begin{equation}\label{MP}
\begin{cases}
\mathcal{L}_A v\geq 0 & \text{ in }\,D\cap\{t<T\},\\
v\leq 0 & \text{ on }\,\partial D\cap\{t<T\},
\end{cases}
\qquad\mbox{ then }\qquad v\leq 0 \,\,\text{ in }\,\, D\cap\{t<T\}.
\end{equation}
\end{center}
We recall a number of essential relations between the homogeneous norm $|\cdot|_B$ and the Euclidean norm that will be used throughout the paper. Some of these can already be found in \cite{KLT, Manfr, Poli}; we collect and prove them in the following lemma for the reader's convenience.
\begin{lemma}\label{normsarecomparable} The following properties hold:
\begin{itemize}
\item[(i)] The triangle inequality holds in the norm $|\cdot|_B$:
\begin{equation}\label{triangleB}
|x+\xi|_B\leq |x|_B + |\xi|_B\qquad \forall\,x,\xi\in\mathbb{R}^N.
\end{equation}
\item[(ii)] Denoting $\sigma_0=\min_{\left|x\right|=1}{|x|_B}$ and $\bar{\sigma}=\max_{\left|x\right|=1}{|x|_B}$ we have
\begin{equation}\label{confrhomnonhom}
\sigma_0\min{\left\{\left|x\right|, \left|x\right|^{\frac{1}{2n+1}}\right\}}\leq|x|_B\leq \bar{\sigma}\max{\left\{\left|x\right|, \left|x\right|^{\frac{1}{2n+1}}\right\}}\qquad \forall\,x\in\mathbb{R}^N.
\end{equation}
\item[(iii)] There exists a structural constant $c(n,B)>0$ such that
\begin{equation}\label{benne}
\left|\left(E(t)-\mathbb{I}_N\right) x\right|_B\leq c(n,B)\max\left\{|x|^{\frac{1}{3}},|x|^{\frac{1}{2n+1}}\right\}\max\left\{|t|^{\frac{1}{2n+1}},|t|^{\frac{n}{2n+1}}\right\} \qquad \forall\,x\in\mathbb{R}^N,\,\,\forall\, t\in\mathbb{R}.
\end{equation}
\end{itemize}
\end{lemma}
\begin{proof}
\begin{itemize}
\item[(i)] This follows from the subadditivity of $|\cdot|^p$ for $0 < p < 1$.
\item[(ii)] For $x=0$ is trivial. For any $x\neq 0$, we have
$$\frac{|x|_B}{\max{\left\{\left|x\right|,\left|x\right|^{\frac{1}{2n+1}}\right\}}}\leq\sum_{i=0}^n{\frac{\left|x^{(p_i)}\right|^{\frac{1}{2i+1}}}{\left|x\right|^{\frac{1}{2i+1}}}}=\sum_{i=0}^n{\left|\left(\frac{x}{\left|x\right|}\right)^{(p_i)}\right|^{\frac{1}{2i+1}}}=\left|\frac{x}{\left|x\right|}\right|_B \leq\bar{\sigma}.$$
while on the other side,
$$\frac{\left|x\right|_B}{\min{\left\{\left|x\right|,\left|x\right|^{\frac{1}{2n+1}}\right\}}}\geq\sum_{i=0}^n{\frac{\left|x^{(p_i)}\right|^{\frac{1}{2i+1}}}{\left|x\right|^{\frac{1}{2i+1}}}}=\sum_{i=0}^n{\left|\left(\frac{x}{\left|x\right|}\right)^{(p_i)}\right|^{\frac{1}{2i+1}}}=\left|\frac{x}{\left|x\right|}\right|_B \geq\sigma_0.$$
\item[(iii)] Fix any $x\in\mathbb{R}^N$, $t\in\mathbb{R}$. By the upper triangular form of $B$, we have $\left(E(t)x\right)^{(p_0)}=x^{(p_0)}$ and for any $i\in\{1,\ldots,n\}$
$$\left(E(t)x\right)^{(p_i)}=x^{(p_i)}+\sum_{k=1}^i\frac{(-t)^k}{k!}\left(\mathbb{B}^T_i\mathbb{B}^T_{i-1}\cdots\mathbb{B}^T_{i-k+1}\right)x^{(p_{i-k})}.$$
Hence, by denoting $M_B=\max_{i}\left\|\mathbb{B}^T_i\right\|$, we get
\begin{eqnarray*}
\left|\left(E(t)-\mathbb{I}_N\right) x\right|_B&=&\sum_{i=1}^n\left|\sum_{k=1}^i\frac{(-t)^k}{k!}\left(\mathbb{B}^T_i\mathbb{B}^T_{i-1}\cdots\mathbb{B}^T_{i-k+1}\right)x^{(p_{i-k})}\right|^{\frac{1}{2i+1}}\\
&\leq& \sum_{i=1}^n\sum_{k=1}^i|t|^{\frac{k}{2i+1}}M_B^{\frac{k}{2i+1}}\left|x^{(p_{i-k})}\right|^{\frac{1}{2i+1}}\\
&\leq&\frac{n(n+1)}{2}\max\left\{M_B^{\frac{1}{2n+1}},M_B^{\frac{n}{2n+1}}\right\}\max\left\{|x|^{\frac{1}{3}},|x|^{\frac{1}{2n+1}}\right\}\max\left\{|t|^{\frac{1}{2n+1}},|t|^{\frac{n}{2n+1}}\right\}.
\end{eqnarray*}
\end{itemize}
\end{proof}
It is known \cite{Ho, K82} that the fundamental solution of $\mathcal{L}_0$ with pole at the origin is given by
\begin{equation}\label{fondfrozen}
\Gamma_0 \left(x,t\right) =\left\{
\begin{tabular}
[c]{lll}
$0$ & $\text{for }t\leq0,$\\ \mbox{}\\
$\frac{(4\pi)^{-\frac{N}{2}}}{\sqrt{\mathrm{det}(C_0(t))}} \exp \left( -\frac{1}{4} \langle C_0^{-1} (t) x, x \rangle \right)=\frac{c_0}{t^{\frac{Q}{2}}} \exp \left( -\frac{1}{4} \langle C_0^{-1} (1) D_{\frac{1}{\sqrt{t}}} x, D_{\frac{1}{\sqrt{t}}} x \rangle \right) $ & $\text{for } t>0,$
\end{tabular}
\ \right.
\end{equation}
where $c_0=(4\pi)^{-\frac{N}{2}}\left(\mathrm{det}(C_0(1))\right)^{-\frac{1}{2}}$. By the translation invariance of $\mathcal{L}_0$, one can relocate the pole to any desired point. Note also that $\Gamma_0$ is $\delta_r$-homogeneous of degree $-Q$. The fundamental solution $\Gamma_0$ and its level sets play an essential role in the proof of Harnack's inequality for the operator $\mathcal{L}_0$ established in \cite{GL90, LP}. In the sequel, it will be necessary for us to have good estimates on the quadratic form $\langle C_0^{-1} (t) x, x \rangle$. We begin by defining a relevant structural constant $b_B$. Since
$E(\sigma)=e^{-\sigma B^T}\rightarrow\mathbb{I}_N$ as $\sigma\rightarrow 0$,
we can define the constant $b_B$ such that
\begin{equation}\label{bibi}
0<b_B\leq\left(\frac{\sigma_0}{\bar{\sigma}}\right)^2\qquad\mbox{and}\qquad \left\|E(\sigma)\right\|\leq 2\,\,\,\,\mbox{ for all }|\sigma|\leq b_B.
\end{equation}
Here the constants $\sigma_0\leq\bar{\sigma}$ are the ones from \eqref{confrhomnonhom}.
\begin{lemma}\label{evalsofC}
There exist structural constants $\Lambda_1, \lambda_1$ such that
\begin{equation}\label{bici}
\frac{1}{\Lambda_1 t}\mathbb{I_N}\leq C_0^{-1}(t) \leq \frac{1}{\lambda_1 t^{2n+1}} \mathbb{I_N}\quad\mbox{ for all }0<t\leq b_B.
\end{equation}
\end{lemma}
\begin{proof} Fix an arbitrary $v\in\mathbb{R}^N$ with $|v|=1$. If $0<t\leq b_B$, then it follows from \eqref{confrhomnonhom} and \eqref{bibi}
$$\min{\left\{\left|D_{\sqrt{t}} v\right|, \left|D_{\sqrt{t}} v\right|^{\frac{1}{2n+1}}\right\}}\leq\frac{1}{\sigma_0}\left|D_{\sqrt{t}} v\right|_B=\frac{\sqrt{t}}{\sigma_0}\left|v\right|_B\leq \sqrt{t}\frac{\bar{\sigma}}{\sigma_0}\leq 1.$$
This says in particular that $\left|D_{\sqrt{t}} v\right|\leq \left|D_{\sqrt{t}} v\right|^{\frac{1}{2n+1}}$, and so
\begin{equation}\label{upv}
\left|D_{\sqrt{t}} v\right|\leq \sqrt{t}\frac{\bar{\sigma}}{\sigma_0}.
\end{equation}
On the other side, we can use \eqref{confrhomnonhom} again to obtain
\begin{equation}\label{dowv}
\left|D_{\sqrt{t}} v\right|^{\frac{1}{2n+1}}=\max{\left\{\left|D_{\sqrt{t}} v\right|, \left|D_{\sqrt{t}} v\right|^{\frac{1}{2n+1}}\right\}}\geq\frac{1}{\bar{\sigma}}\left|D_{\sqrt{t}} v\right|_B=\frac{\sqrt{t}}{\bar{\sigma}}\left|v\right|_B \geq \sqrt{t}\frac{\sigma_0}{\bar{\sigma}}.
\end{equation}
We can now employ the commutation relation \eqref{split} and the hypothesis \eqref{azero} on $A_0$ to uniformly bound from above and below the quadratic form $\left\langle C_0(t)v,v\right\rangle$. Denote by $\Lambda_I$ and $\lambda_I$ respectively the maximum and the minimum eigenvalue of $C(1)$. Then by \eqref{upv} and \eqref{dowv}, we get
\begin{eqnarray*}
&&\left\langle C_0(t) v,v \right\rangle\leq \Lambda \left\langle C(1) D_{\sqrt{t}} v,D_{\sqrt{t}} v \right\rangle \leq\Lambda
\Lambda_I|D_{\sqrt{t}} v|^2\leq \Lambda\L_I\left(\frac{\bar{\sigma}}{\sigma_0}\right)^2 t=:\Lambda_1 t,\\
&&\left\langle C_0(t) v,v \right\rangle \geq \lambda \left\langle C(1) D_{\sqrt{t}} v,D_{\sqrt{t}} v \right\rangle
\geq \lambda\l_I|D_{\sqrt{t}} v|^2\geq \lambda\l_I\left(\frac{\sigma_0}{\bar{\sigma}}\right)^{4n+2} t^{2n+1}=:\lambda_1 t^{2n+1},
\end{eqnarray*}
for every $v\in\mathbb{R}^N$ with $|v|=1$. In other words, we have just shown that
$$\lambda_1 t^{2n+1}\mathbb{I_N}\leq C_0(t) \leq \Lambda_1 t \mathbb{I_N} \quad\mbox{ for all }0<t\leq b_B$$
for some structural constants $\lambda_1,\Lambda_1$. This implies
\begin{equation*}
\frac{1}{\Lambda_1 t}\mathbb{I_N}\leq C_0^{-1}(t) \leq \frac{1}{\lambda_1 t^{2n+1}} \mathbb{I_N}\quad\mbox{ for all }0<t\leq b_B.
\end{equation*}
\end{proof}
\section{Pointwise estimates for Gaussian Kernels}\label{secest}
In this section, we initiate the construction of explicit barriers which will be used to prove the growth lemma. These barriers are modeled after the fundamental solution $\Gamma_0$ \eqref{fondfrozen}. To this end, for $s, \beta > 0$, we consider the function
\begin{equation}\label{defbar}
\Gamma_{s,\beta}(z ) =\left\{
\begin{tabular}
[c]{lll}
$0$ & $\text{for }t\leq0,$\\\mbox{}\\
$\frac{1}{t^{s\frac{Q}{2}} }
\exp\left( -\frac{1}{4\beta}\left\langle C_0^{-1}( 1)D_{\frac{1}{\sqrt{t}}} x,D_{\frac{1}{\sqrt{t}}} x\right\rangle
\right) $ & $\text{for }t>0.$
\end{tabular}
\ \right.
\end{equation}
Note that $\Gamma_{s,\beta}$ is $\delta_r$-homogeneous of degree $-sQ$. We devote the rest of this section to establishing the necessary pointwise estimates for $\Gamma_{s,\beta}$.
\begin{lemma}\label{upperboundlemma} (Upper Bound for $\Gamma_{s,\beta}$)
Let $s,\beta$ be positive numbers. There exist $c_1>0$ and $K_1> \sigma_0$ depending just on $s, \beta$, and structural constants such that, for every $r>0$ and $K\geq K_1$, if we consider the cylinders
$$\mathcal{Q}^1_r:=\mathcal{Q}_{K r}^{-b_B r^2, 0},
\qquad S^1_r:=\partial B_{Kr}(0)\times [-b_B r^2,0],\qquad \mathcal{Q}^3_r:=\mathcal{Q}_{\sigma_0 r}^{-b_B r^2, -\frac{1}{2}b_B r^2},
$$
then we have
\begin{equation}\label{upperbound}
\Gamma_{s,\beta}(\zeta^{-1}\circ z)\leq \frac{1}{(b_B r^2)^{s\frac{Q}{2}} }
\exp\left(\frac{-c_1 K^2}{b_B} \right)\quad\forall\,\zeta=(\xi,\tau)\in \mathcal{Q}^3_r, \ z=(x,t)\in S^1_r.
\end{equation}
\end{lemma}
\begin{proof} Recall that $\Gamma_{s,\beta}(\zeta^{-1}\circ z)=0$ if $t\leq \tau$. Therefore, it suffices to assume $z\in S^1_r$ and $\zeta\in \mathcal{Q}^3_r$ with $t>\tau$. In this case, we have
$$\Gamma_{s,\beta}(\zeta^{-1}\circ z)=\frac{1}{(t-\tau)^{s\frac{Q}{2}} }
\exp\left( -\frac{1}{4\beta}\left\langle C_0^{-1}(t-\tau)\left(x-E(t-\tau)\xi\right),\left(x-E(t-\tau)\xi\right)\right\rangle\right).$$
Let us deal with the term inside the exponential. By \eqref{split}, we have
\begin{eqnarray*}
&&\left\langle C_0^{-1}(t-\tau)\left(x-E(t-\tau)\xi\right),\left(x-E(t-\tau)\xi\right)\right\rangle\\
&&=\left\langle C_0^{-1}\left(\frac{t-\tau}{r^2}\right)D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right),D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right)\right\rangle \qquad \mbox{ for all } r>0.
\end{eqnarray*}
By definition we have $0<\frac{t-\tau}{r^2}\leq \frac{-\tau}{r^2}\leq b_B$. Therefore,
\begin{eqnarray*}
&&\left\langle C_0^{-1}\left(\frac{t-\tau}{r^2}\right)D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right),D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right)\right\rangle \\
&\geq& \frac{r^2}{(t-\tau)\Lambda_1}|D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right)|^2 \qquad \mbox{ by } \eqref{bici} \\
&=&\frac{r^2}{(t-\tau)\Lambda_1}\left(|D_{\frac{1}{r}} x|^2-2\left\langle D_{\frac{1}{r}} x, D_{\frac{1}{r}} E(t-\tau)\xi \right\rangle + |D_{\frac{1}{r}} E(t-\tau)\xi|^2\right)\\
&\geq& \frac{r^2}{(t-\tau)\Lambda_1}\left(|D_{\frac{1}{r}} x|^2-2\left\langle D_{\frac{1}{r}} x, D_{\frac{1}{r}} E(t-\tau)\xi \right\rangle\right)\\
&=& \frac{r^2}{(t-\tau)\Lambda_1}\left(|D_{\frac{1}{r}} x|^2-2\left\langle D_{\frac{1}{r}} x, E\left(\frac{t-\tau}{r^2}\right)D_{\frac{1}{r}}\xi \right\rangle\right) \qquad \mbox{ by } \eqref{commu} \\
&\geq&\frac{r^2}{(t-\tau)\Lambda_1}\left(|D_{\frac{1}{r}} x|^2-2\left\|E\left(\frac{t-\tau}{r^2}\right)\right\| |D_{\frac{1}{r}} x| |D_{\frac{1}{r}}\xi|\right)\\
&\geq& \frac{r^2}{(t-\tau)\Lambda_1}|D_{\frac{1}{r}} x|\left(|D_{\frac{1}{r}} x|-4 |D_{\frac{1}{r}}\xi|\right) \qquad \mbox{ by } \eqref{bibi}.
\end{eqnarray*}
Summing up, we have just proved
\begin{equation}\label{belowexp}
\left\langle C_0^{-1}(t-\tau)\left(x-E(t-\tau)\xi\right),\left(x-E(t-\tau)\xi\right)\right\rangle\geq \frac{r^2}{(t-\tau)\Lambda_1}|D_{\frac{1}{r}} x|\left(|D_{\frac{1}{r}} x|-4 |D_{\frac{1}{r}}\xi|\right).
\end{equation}
We now need a bound from below for $|D_{\frac{1}{r}} x|$ and a bound from above for $|D_{\frac{1}{r}} \xi|$. By \eqref{confrhomnonhom} and the definition of $S^1_r$, we have
$$K_1\leq K=\left|D_{\frac{1}{r}} x\right|_B\leq \bar{\sigma}\max{\left\{\left|D_{\frac{1}{r}}x\right|, \left|D_{\frac{1}{r}}x\right|^{\frac{1}{2n+1}}\right\}}.$$
If we choose $K_1\geq \bar{\sigma}$, this yields
\begin{equation}\label{drx}
|D_{\frac{1}{r}} x|=\max{\left\{\left|D_{\frac{1}{r}}x\right|, \left|D_{\frac{1}{r}}x\right|^{\frac{1}{2n+1}}\right\}}\geq \frac{K}{\bar{\sigma}}.
\end{equation}
On the other hand, by \eqref{confrhomnonhom} and the definition of $\mathcal{Q}^3_r$, we have
$$\sigma_0\geq \left|D_{\frac{1}{r}} \xi\right|_B\geq \sigma_0\min{\left\{\left|D_{\frac{1}{r}} \xi\right|, \left|D_{\frac{1}{r}} \xi\right|^{\frac{1}{2n+1}}\right\}},$$
which says in particular that
\begin{equation}\label{drxi}
|D_{\frac{1}{r}} \xi|=\min{\left\{\left|D_{\frac{1}{r}} \xi\right|, \left|D_{\frac{1}{r}} \xi\right|^{\frac{1}{2n+1}}\right\}}\leq 1.
\end{equation}
Using \eqref{drx} and \eqref{drxi} in \eqref{belowexp}, and choosing $K_1\geq 8\bar{\sigma}$, we get
$$\left\langle C_0^{-1}(t-\tau)\left(x-E(t-\tau)\xi\right),\left(x-E(t-\tau)\xi\right)\right\rangle\geq \frac{r^2}{t-\tau}\frac{K^2}{2\Lambda_1(\bar{\sigma})^2}.$$
Hence
$$\Gamma_{s,\beta}(\zeta^{-1}\circ z)\leq\frac{1}{(t-\tau)^{s\frac{Q}{2}} }
\exp\left( \frac{-r^2}{t-\tau}\frac{K^2}{8\Lambda_1(\bar{\sigma})^2\beta} \right).$$
The function $f(y)=y^{-s\frac{Q}{2}}\exp{\left(\frac{-C}{y}\right)}$ (for $C>0$) is monotone increasing for $y\in(0,\frac{2C}{sQ}]$. If we finally fix
$$K_1=\max\left\{8\bar{\sigma}, 2\bar{\sigma}\sqrt{b_B\Lambda_1\beta s Q}\right\},$$
then $0<t-\tau\leq b_Br^2\leq \frac{r^2K^2}{4\Lambda_1(\bar{\sigma})^2\beta s Q }$ by construction and we get
$$\Gamma_{s,\beta}(\zeta^{-1}\circ z)\leq\frac{1}{(b_B r^2)^{s\frac{Q}{2}} }
\exp\left(\frac{-K^2}{8\Lambda_1(\bar{\sigma})^2\beta b_B} \right).$$
\end{proof}
\begin{lemma}\label{lowerboundlemma} (Lower Bound for $\Gamma_{s,\beta}$)
Let $s,\beta, r$ be positive numbers. Consider the cylinders
$$\mathcal{Q}^2_r:=\mathcal{Q}_{\sigma_0 r}^{-\frac{1}{4}b_B r^2, 0},\qquad
\mathcal{Q}^3_r:=\mathcal{Q}_{\sigma_0 r}^{-b_B r^2, -\frac{1}{2}b_B r^2}.$$
There exists $c_2>0$ depending on $\beta$ and structural constants such that
\begin{equation}\label{lowerbound}
\Gamma_{s,\beta}(\zeta^{-1}\circ z)\geq \frac{1}{(b_B r^2)^{s\frac{Q}{2}} }
\exp\left(\frac{-c_2}{b_B^{2n+1}}\right)\quad\forall\,\zeta=(\xi,\tau)\in \mathcal{Q}^3_r,\,\ z=(x,t)\in \mathcal{Q}^2_r.
\end{equation}
\end{lemma}
\begin{proof}
Fix $z\in \mathcal{Q}^2_r$ and $\zeta\in \mathcal{Q}^3_r$. Then
\begin{equation}\label{tmentau}
b_Br^2\geq t-\tau\geq \frac{1}{4}b_Br^2.
\end{equation}
We argue similarly to the proof of Lemma \ref{upperboundlemma} to estimate the quadratic form from below:
\begin{eqnarray*}
&&\left\langle C_0^{-1}(t-\tau)\left(x-E(t-\tau)\xi\right),\left(x-E(t-\tau)\xi\right)\right\rangle=\\
&=&\left\langle C_0^{-1}\left(\frac{t-\tau}{r^2}\right)D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right),D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right)\right\rangle \qquad \mbox{ by } \eqref{split}\\
&\leq&\frac{1}{\lambda_1}\left(\frac{r^2}{t-\tau}\right)^{2n+1}|D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right)|^2 \qquad \mbox{ by } \eqref{bici}\\
&\leq& \frac{1}{\lambda_1}\left(\frac{4}{b_B}\right)^{2n+1}|D_{\frac{1}{r}} \left(x-E(t-\tau)\xi\right)|^2 \qquad \mbox{ by } \eqref{tmentau}\\
&\leq& \frac{2}{\lambda_1}\left(\frac{4}{b_B}\right)^{2n+1}\left(|D_{\frac{1}{r}}x|^2 + |D_{\frac{1}{r}}E(t-\tau)\xi|^2\right)\\
&=& \frac{2}{\lambda_1}\left(\frac{4}{b_B}\right)^{2n+1}\left(|D_{\frac{1}{r}}x|^2 + \left|E\left(\frac{t-\tau}{r^2}\right)D_{\frac{1}{r}}\xi\right|^2\right) \qquad \mbox{ by } \eqref{commu}\\
&\leq&\frac{2}{\lambda_1}\left(\frac{4}{b_B}\right)^{2n+1}\left(|D_{\frac{1}{r}}x|^2 + \left\|E\left(\frac{t-\tau}{r^2}\right)\right\|^2|D_{\frac{1}{r}}\xi|^2\right)\\
&\leq&\frac{2}{\lambda_1}\left(\frac{4}{b_B}\right)^{2n+1}\left(|D_{\frac{1}{r}}x|^2 + 4|D_{\frac{1}{r}}\xi|^2\right) \qquad \mbox{ by } \eqref{tmentau}\mbox{ and }\eqref{bibi}.
\end{eqnarray*}
Since $x,\xi\in B_{\sigma_0 r}(0)$, we can argue as in \eqref{drxi} to conclude
$$|D_{\frac{1}{r}}x|, \ |D_{\frac{1}{r}}\xi|\leq 1.$$
Hence
$$\left\langle C_0^{-1}(t-\tau)\left(x-E(t-\tau)\xi\right),\left(x-E(t-\tau)\xi\right)\right\rangle\leq\frac{10}{\lambda_1}\left(\frac{4}{b_B}\right)^{2n+1}.$$
Recalling once more \eqref{tmentau}, we finally obtain
$$\Gamma_{s,\beta}(\zeta^{-1}\circ z)\geq \frac{1}{(b_B r^2)^{s\frac{Q}{2}} }
\exp\left(\frac{-5}{2\lambda_1\beta} \left(\frac{4}{b_B}\right)^{2n+1} \right).$$
\end{proof}
\section{Construction of Barriers}\label{barriers}
Our aim in this section is to construct potentials using the kernels $\Gamma_{s,\beta}$ \eqref{defbar}. The eventual goal is to use these potentials as barriers for the operators $\mathcal{L}_A$ under the the hypotheses \hyperref[H1]{H1} and \hyperref[H2]{H2}. The pointwise bounds from Section \ref{secest} will then allow us to successfully use comparison principle arguments in the proof of the growth lemma given in Section \ref{sec5}.
For a fixed Lebesgue-measurable set $E\subset\mathbb{R}^{N+1}$, consider the function
\begin{equation}\label{defUE}
U_E(z):=\int_E{\Gamma_{s,\beta}(\zeta^{-1}\circ z)\,{\rm d}\zeta}, \qquad z\in\mathbb{R}^{N+1}.
\end{equation}
In order for $U_E$ to be well-defined, we must impose a bound on the parameter $s$, as shown in the following lemma.
\begin{lemma}\label{lemdd}
Fix $s,\beta>0$ and assume $s<1+\dfrac{2}{Q}$. Then, for any $T_1<T_2$ there exists a constant $C = C(T_1,T_2,s,\beta) > 0$ such that
\begin{equation}\label{boundint}
\varphi(z) := \int_{\mathbb{R}^N\times [T_1,T_2]} \Gamma_{s,\beta}(\zeta^{-1}\circ z)\,{\rm d}\zeta \leq C\qquad\mbox{ for all }z\in\mathbb{R}^{N+1}.
\end{equation}
Moreover, for all Lebesgue-measurable sets $E\subset\mathbb{R}^{N+1}$ and for all $r>0$, we have
\begin{equation}\label{maxr}
\sup_{\mathbb{R}^{N+1}}{U_{\left(\delta_r E\right)}}=r^{Q+2-sQ}\sup_{\mathbb{R}^{N+1}}{U_{E}}.
\end{equation}
\end{lemma}
\begin{proof}
We first prove \eqref{boundint}. Fix $z=(x,t)\in\mathbb{R}^{N+1}$. Suppose $t> T_1$, since the other possibility is trivial. Note that $\varphi(z)$ can be written as
$$\int_{\mathbb{R}^N\times [T_1,\min{\{t,T_2\}})}{\frac{1}{(t-\tau)^{s\frac{Q}{2}} }
\exp\left( -\frac{1}{4\beta}\left\langle C_0^{-1}( 1)D_{\frac{1}{\sqrt{t-\tau}}} \left(x-E(t-\tau)\xi\right),D_{\frac{1}{\sqrt{t-\tau}}} \left(x-E(t-\tau)\xi\right)\right\rangle
\right)\,{\rm d}\xi{\rm d}\tau}.$$
By performing the change of variables $\xi'=\xi-E(\tau-t)x$ and using the commutation property \eqref{commu}, we get
\begin{eqnarray*}
&&\varphi(z) = \int_{\mathbb{R}^N\times [T_1,\min{\{t,T_2\}})}{\frac{1}{(t-\tau)^{s\frac{Q}{2}} }
\exp\left( -\frac{1}{4\beta}\left\langle C_0^{-1}( 1)D_{\frac{1}{\sqrt{t-\tau}}} E(t-\tau)\xi',D_{\frac{1}{\sqrt{t-\tau}}} E(t-\tau)\xi'\right\rangle
\right)\,{\rm d}\xi'{\rm d}\tau}\\
&=& \int_{\mathbb{R}^N\times [T_1,\min{\{t,T_2\}})}{\frac{1}{(t-\tau)^{s\frac{Q}{2}} }
\exp\left( -\frac{1}{4\beta}\left\langle E^T(1)C_0^{-1}( 1)E(1)D_{\frac{1}{\sqrt{t-\tau}}} \xi',D_{\frac{1}{\sqrt{t-\tau}}} \xi'\right\rangle
\right)\,{\rm d}\xi'{\rm d}\tau}.
\end{eqnarray*}
We can now change $\xi=D_{\frac{1}{\sqrt{t-\tau}}} \xi'$ and get that the last integral is equal to
$$\left(\int_{\mathbb{R}^N}{\exp\left( -\frac{1}{4\beta}\left\langle E^T(1)C_0^{-1}( 1)E(1)\xi,\xi\right\rangle
\right)\,{\rm d}\xi}\right)\left(\int_{[T_1,\min{\{t,T_2\}})}{\frac{(t-\tau)^{\frac{Q}{2}}}{(t-\tau)^{s\frac{Q}{2}}}\,{\rm d}}\tau\right).$$
The second integral is finite if $s<1+\frac{2}{Q}$. The first integral can be easily bounded, and it can even be computed explicitly. Indeed, one can see that $\int_{\mathbb{R}^N}\Gamma_0(\zeta^{-1}\circ z)\,{\rm d}\xi=1$ for every $z$ with $t>\tau$. By choosing $x=0$ and $t=\tau+1$, we infer that $\int_{\mathbb{R}^N}c_0\exp\left( -\frac{1}{4}\left\langle E^T(1)C_0^{-1}( 1)E(1)\xi,\xi\right\rangle
\right)\,{\rm d}\xi=1$. This implies, by \eqref{azero},
$$\int_{\mathbb{R}^N}{\exp\left( -\frac{1}{4\beta}\left\langle E^T(1)C_0^{-1}( 1)E(1)\xi,\xi\right\rangle
\right)\,{\rm d}\xi}=\frac{1}{c_0\beta^{\frac{N}{2}}}\leq \left(\frac{4\pi\Lambda}{\beta}\right)^{\frac{N}{2}}\sqrt{\mathrm{det}(C(1))}.$$
This proves \eqref{boundint}. The proof of \eqref{maxr} follows by homogeneity and the properties of the group automorphisms $\delta_r$. In fact, for $z\in\mathbb{R}^{N+1}$ and $r>0$, we have
$$U_{\left(\delta_r E\right)}(\delta_r z)=r^{Q+2}\int_{E}{\Gamma_{s,\beta}((\delta_r \zeta)^{-1}\circ (\delta_r z))\,{\rm d}\zeta}=r^{Q+2-sQ}\int_{E}{\Gamma_{s,\beta}(\zeta^{-1}\circ z)\,{\rm d}\zeta}=r^{Q+2-sQ} U_E(z).$$
\end{proof}
We will assume from here onward that the set $E$ in the definition of $U_E$ is contained in some fixed strip $\mathbb{R}^N\times (T_1,T_2)$. In the remainder of this section, we will determine conditions on the parameters $s, \beta$ that are necessary for $U_E$ to be a subsolution for the class of variable coefficient operators \eqref{varintroL}. To this end, we compute $\mathcal{L}_A \Gamma_{s,\beta}$. Recalling the definition of $\Gamma_{s,\beta}$ \eqref{defbar}, for fixed $s,\beta>0$ and for all $\Omega\ni z\neq 0$, we can compute
\begin{eqnarray}
\nabla\Gamma_{s,\beta}(z)&=&-\frac{1}{2\beta}\Gamma_{s,\beta}(z)C_0^{-1}(t) x\label{nabG}\\
D^2\Gamma_{s,\beta}(z)&=&\frac{1}{2\beta}\Gamma_{s,\beta}(z)\left(-C_0^{-1}(t) +\frac{1}{2\beta}\left(C_0^{-1}(t) x\right)\otimes\left(C_0^{-1}(t) x\right)\right)\label{HG}.
\end{eqnarray}
To compute the $t$-derivative, we use the following identities for invertible matrices $M(t)$:
\begin{itemize}
\item[$\cdot$] $(M^{-1}(t))'=-M^{-1}(t)M'(t)M^{-1}(t)$;
\item[$\cdot$] $(\mathrm{det}(M(t)))'=\mathrm{tr}(M'(t)M^{-1}(t))\,\mathrm{det}(M(t))$.
\end{itemize}
This yields
\begin{eqnarray*}
\partial_t\Gamma_{s,\beta}(z)&=&\Gamma_{s,\beta}(z)\left(-\frac{s}{2}\frac{(\mathrm{det}(C_0(t)))'}{\mathrm{det}(C_0(t))}+\frac{1}{4\beta}\left\langle C_0'(t)C_0^{-1}(t) x,C_0^{-1}(t) x\right\rangle \right)=\\
&=&\Gamma_{s,\beta}(z)\left(-\frac{s}{2}\mathrm{tr}\left(C_0'(t)C_0^{-1}(t)\right) +\frac{1}{4\beta}\left\langle C_0'(t)C_0^{-1}(t) x,C_0^{-1}(t) x\right\rangle\right).
\end{eqnarray*}
We have by definition $C_0'(t)=E(t)A_0E^T(t)$. On the other hand, the following identity also holds
$$E(t)A_0E^T(t) = A_0-B^TC_0(t)-C_0(t)B\qquad\forall\,t>0.$$
To see this, note that the r.h.s. and the l.h.s. agree at $t=0$ and they have the same derivative. Consequently,
\begin{equation}\label{forgotten}
C_0'(t)=A_0-B^TC_0(t)-C_0(t)B\qquad\forall\,t>0.
\end{equation}
Multiplying by $C_0^{-1}(t)$ and taking the trace (recall $\mathrm{tr}(B)=0$), we get
$$\mathrm{tr}\left(C_0'(t)C_0^{-1}(t)\right) =\mathrm{tr}\left(A_0C_0^{-1}(t)\right),$$
which says
\begin{equation}\label{tderG}
\partial_t\Gamma_{s,\beta}(z)=\frac{1}{2\beta}\Gamma_{s,\beta}(z)\left(-s\beta \,\mathrm{tr}\left(A_0C_0^{-1}(t)\right)+\frac{1}{2}\left\langle C_0'(t)C_0^{-1}(t) x,C_0^{-1}(t) x\right\rangle\right).
\end{equation}
Recalling the definition of $\mathcal{L}_A$ and gathering the information in \eqref{nabG}, \eqref{HG} and \eqref{tderG}, we obtain
\begin{eqnarray}\label{primadihip}
\mathcal{L}_A\Gamma_{s,\beta}(z)=\frac{1}{2\beta}\Gamma_{s,\beta}(z)&&\hspace{-0.6cm}\left(-\mathrm{tr}\left(A(z)C_0^{-1}(t)\right)+\frac{1}{2\beta}\left\langle A(z)C_0^{-1}(t) x,C_0^{-1}(t) x\right\rangle - \left\langle BC_0^{-1}(t) x, x\right\rangle+\right. \nonumber\\
&&\left. +s\beta \,\mathrm{tr}\left(A_0C_0^{-1}(t)\right)-\frac{1}{2}\left\langle C_0'(t)C_0^{-1}(t) x,C_0^{-1}(t) x\right\rangle\right)=\nonumber\\
=\frac{1}{2\beta}\Gamma_{s,\beta}(z)&&\hspace{-0.6cm}\left(s\beta \,\mathrm{tr}\left(A_0C_0^{-1}(t)\right) -\mathrm{tr}\left(A(z)C_0^{-1}(t)\right)+\vphantom{\frac{1}{2}}\right.\\
&&\left. +\frac{1}{2}\left(\frac{1}{\beta}\left\langle A(z)C_0^{-1}(t) x,C_0^{-1}(t) x\right\rangle-\left\langle A_0C_0^{-1}(t) x,C_0^{-1}(t) x\right\rangle\right)\right)\nonumber
\end{eqnarray}
where in the last equality we exploited again the expression for $C_0'(t)$ in \eqref{forgotten}.
Using \eqref{primadihip}, we now show that the parameters $s, \beta$ can be chosen appropriately under each of the hypotheses (H1) and (H2) so that $U_E$ is a subsolution for $\mathcal{L}_A$ outside $E$.
\subsection{Cordes-Landis Condition \hyperref[H1]{H1}}\label{cordeslandiscondition}
Consider the kernel $\Gamma_{s,\beta}$ with the choice
$$A_0=I_0,$$
where $I_0$ is defined in \eqref{matrixexponential}. Note that with this choice of $A_0$ we have $C_0 = C$, where $C$ is defined in \eqref{Kalma0}. Fix also
\begin{equation}\label{defh1bs}
\beta=\lambda\qquad\mbox{ and }\qquad s=\frac{\Lambda}{\lambda}.
\end{equation}
Having fixed these quantities, we can prove the following
\begin{lemma}\label{lemuu}
The function $U_E$ is continuous in $\mathbb{R}^{N+1}$, $C^2$ outside $\overline{E}$, and for all $A$ satisfying \hyperref[H1]{H1}, we have
$$\mathcal{L}_A U_E(z)\geq 0 \quad\mbox{ for }z\in\Omega\smallsetminus \overline{E}.$$
\end{lemma}
\begin{proof}
The choice of $s$ and the hypothesis \hyperref[H1]{H1} allow us to invoke Lemma \ref{lemdd} and conclude that $U_E<+\infty$. Moreover, with such choices, $U_E$ is continuous by the dominated convergence theorem and smooth away from the poles in $E$.
Let us now prove the $\mathcal{L}_A$-subsolution property. By \eqref{defh1bs}, and using \eqref{ellipticityofA} in \eqref{primadihip}, we have
\begin{eqnarray*}
\mathcal{L}_A\Gamma_{s,\beta}(z)=\frac{1}{2\beta}\Gamma_{s,\beta}(z)&&\hspace{-0.6cm}\left(s\beta \,\mathrm{tr}\left(I_0C^{-1}(t)\right) -\mathrm{tr}\left(A(z)C^{-1}(t)\right)+\vphantom{\frac{1}{2}}\right.\\
&&\left. +\frac{1}{2}\left(\frac{1}{\beta}\left\langle A(z)C^{-1}(t) x,C^{-1}(t) x\right\rangle-\left\langle I_0C^{-1}(t) x,C^{-1}(t) x\right\rangle\right)\right) \\
\geq\frac{1}{2\beta}\Gamma_{s,\beta}(z)&&\hspace{-0.6cm}\left((s\beta-\Lambda) \,\mathrm{tr}\left(I_0C^{-1}(t)\right) +\frac{1}{2}\left(\frac{\lambda}{\beta}-1\right)\left\langle I_0C^{-1}(t) x,C^{-1}(t) x\right\rangle\right)=0.
\end{eqnarray*}
To complete the proof, we have only to remember that the vector fields $\partial_{x_1}, \ldots, \partial_{x_N}$ and $\left\langle x, B\nabla\right\rangle - \partial_t$ are left-invariant with respect to the group law $\circ$. Thus, the function $\Gamma_{s,\beta}(\zeta^{-1}\circ \cdot)$ is $\mathcal{L}_A$-subharmonic for any $\zeta \in \mathbb{R}^{N+1}$, and the lemma follows.
\end{proof}
\subsection{Uniform Continuity Assumption \hyperref[H2]{H2}}\label{continuityassumption}
We make precise here the notion of uniform continuity of the coefficients as stipulated in condition \hyperref[H2]{H2}.
\begin{definition}\label{modulusofcontinuity}
Let $\omega : [0, 1) \rightarrow [0,1)$ be a non-decreasing function satisfying $\lim\limits_{s \rightarrow 0^+} \omega(s) = \omega(0) = 0$. We say the matrix $A(\cdot)$ admits a uniform modulus of continuity $\omega$ if
$$\sup\left\{\left\|\mathbb{A}(z)-\mathbb{A}(z_0)\right\|\,:\,z \in\Omega\cap \mathcal{Q}_\epsilon^{-\epsilon^2,\epsilon^2}(z_0) \right\} \leq \omega(\epsilon)\quad\mbox{ for all }0<\epsilon <1\,\, \mbox{ and for all }z_0 \in \Omega.$$
\end{definition}
Assume now that $A(\cdot)$ admits a uniform modulus of continuity $\omega$. Fix any $z_0\in\Omega$ and choose
$$A_0=A(z_0).$$
Define
$$C_{z_0}(t):=C_0(t)=\int_{0}^{t} E(\sigma)A(z_0)E^T(\sigma)\,d\sigma.$$
Let
\begin{equation}\label{pick}
s=1+s_0,\quad\beta=\frac{1}{1+\frac{s_0}{2}}=\frac{2}{2+s_0}\qquad\mbox{for some }s_0>0 \text{ to be determined}.
\end{equation}
Let $\Gamma_{s,\beta}$ be the kernel corresponding to the above choices. We want the associated potential $U_E$ to be a subsolution in a neighborhood of $z_0$ for $0<s_0<\frac{2}{Q}$. To do this, we exploit the continuity of $A(\cdot)$.
\begin{lemma}\label{barriersforH2} For every $0<s_0<\frac{2}{Q}$ there exists $0<\epsilon_0<1$ depending on $s_0$ and $\omega(\cdot)$ such that
$$\mathcal{L}_A U_E(z) \geq 0 \qquad\mbox{ for all } z \in \Omega \cap \mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0)\smallsetminus\overline{E}.$$
\end{lemma}
\begin{proof}
It suffices to show that for all $\zeta\in E$, we have
$$\left(\mathcal{L}_A \Gamma_{s,\beta}(\zeta^{-1}\circ\cdot)\right)(z)\geq 0\qquad\mbox{ for all }z\in\Omega\cap \mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0)\smallsetminus\{\zeta\}.$$
Fix $0<s_0<\frac{2}{Q}$. By Definition \ref{modulusofcontinuity}, there exists $\epsilon_0>0$ such that
$$\left\|\mathbb{A}(z)-\mathbb{A}(z_0)\right\|\leq\omega(\epsilon_0)\leq \frac{s_0}{2+s_0}\lambda\qquad\mbox{ for all }z\in\Omega\cap \mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0).$$
In particular, this implies
\begin{equation}\label{conscont}
-\frac{s_0}{2}\lambda I_0 \leq A(z)-A(z_0)\leq \frac{s_0}{2+s_0}\lambda I_0\qquad\mbox{ for all }z\in \Omega\cap \mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0).
\end{equation}
Let us now fix $\zeta\in\mathbb{R}^{N+1}$. Arguing as in \eqref{primadihip}, we have for any $z\neq\zeta$
\begin{eqnarray*}
\left(\mathcal{L}_A \Gamma_{s,\beta}(\zeta^{-1}\circ\cdot)\right)(z)&=& \frac{1}{2\beta}\Gamma_{s,\beta}(\zeta^{-1}\circ z)\left(s\beta \,\mathrm{tr}\left(A(z_0)C^{-1}_{z_0}(t-\tau)\right) -\mathrm{tr}\left(A(z)C^{-1}_{z_0}(t-\tau)\right)+\vphantom{\frac{1}{2}}\right.\\
&&+\frac{1}{2}\left(\frac{1}{\beta}\left\langle A(z)C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi),C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi)\right\rangle+\right.\\
&&-\left.\left.\vphantom{\frac{1}{2}}\left\langle A(z_0)C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi),C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi)\right\rangle\right)\right).
\end{eqnarray*}
Let us bound from below separately the trace-terms and the quadratic-terms. Consider any $z\in \Omega\cap \mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0)$. With our choice of $\beta$ in \eqref{pick} we have $\frac{1}{\beta}=1+\frac{s_0}{2}$. Thus, using \eqref{ellipticityofA} and \eqref{conscont}, we get
$$M_1(z):=\frac{1}{\beta}A(z)-A(z_0)=A(z)-A(z_0)+\frac{s_0}{2}A(z)\geq A(z)-A(z_0)+\frac{s_0}{2}\lambda I_0\geq 0$$
which implies
$$\left\langle M_1(z) C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi),C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi)\right\rangle\geq 0.$$
On the other hand, by \eqref{pick} we have $s\beta=1+\frac{s_0}{2+s_0}$. Again, by \eqref{ellipticityofA} and \eqref{conscont}, we get
$$M_2(z):=s\beta A(z_0)-A(z)=A(z_0)-A(z)+\frac{s_0}{2+s_0} A(z_0)\geq A(z_0)-A(z)+\frac{s_0}{2+s_0}\lambda I_0\geq 0$$
which implies
$$\mathrm{tr}\left(M_2(z) C^{-1}_{z_0}(t-\tau)\right)\geq 0.$$
Hence
\begin{eqnarray*}
&&\left(\mathcal{L}_A \Gamma_{s,\beta}(\zeta^{-1}\circ\cdot)\right)(z)= \frac{1}{2\beta}\Gamma_{s,\beta}(\zeta^{-1}\circ z)\left(\mathrm{tr}\left(M_2(z)C^{-1}_{z_0}(t-\tau)\right) +\vphantom{\frac{1}{2}}\right.\\
&&\left.+\frac{1}{2}\left\langle M_1(z)C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi),C^{-1}_{z_0}(t-\tau) (x-E(t-\tau)\xi)\right\rangle\right)\geq 0
\end{eqnarray*}
for every $z\in \Omega\cap \mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0)\smallsetminus\{\zeta\}$.
\end{proof}
\begin{remark}
In lieu of the hypothesis \hyperref[H1]{H1}, we could have assumed the existence of a fixed matrix $A_0$ of the form \eqref{azero} such that
\begin{equation}\label{generalcordes}
\tilde{\lambda} A_0\leq A(z) \leq \tilde{\Lambda} A_0 \quad\mbox{ with }\quad \frac{\tilde{\Lambda}}{\tilde{\lambda}}<1+\frac{2}{Q}.
\end{equation}
The proof of Lemma \ref{lemuu} can then be carried out in exactly the same way. This is in contrast with operators in groups of Heisenberg type considered in \cite{AGT} and \cite{Tralli-Critical-Density}, where it is not clear how to establish the analogue of Lemma \ref{lemuu} under the more general condition \eqref{generalcordes} without making additional structural assumptions on the coefficient matrix. A similar obstruction arises when attempting to prove the analogue of Lemma \ref{barriersforH2} (see \cite[Section 3]{AGT}).
\end{remark}
\section{Growth Lemma and Applications}\label{sec5}
In this section, we establish the Landis growth lemma for the operators $\mathcal{L}_A$ under the hypotheses \hyperref[H1]{H1} and \hyperref[H2]{H2}. It is well known that such growth lemmas are the starting point for proving oscillation decay, H\"older continuity and Harnack's inequality for solutions.
Let us recall the definitions of the cylinder-like sets considered in Lemma \ref{upperboundlemma} and \ref{lowerboundlemma}
$$\mathcal{Q}^1_r(z_0)=\mathcal{Q}_{K r}^{-b_B r^2, 0}(z_0) \qquad \mathcal{Q}^2_r(z_0)=\mathcal{Q}_{\sigma_0 r}^{-\frac{1}{4}b_B r^2, 0}(z_0) \qquad \mathcal{Q}^3_r(z_0)=\mathcal{Q}_{\sigma_0 r}^{-b_B r^2, -\frac{1}{2}b_B r^2}(z_0).$$
In Lemma \ref{upperboundlemma}, it was shown that there exists a constant $K_1 > \sigma_0$ depending only on the structure such that for all $K > K_1$, we have the upper bound on $\Gamma_{s,\beta}$ given in \eqref{upperbound}. We can choose the constant $K$ large enough so that the bound in \eqref{lowerbound} is greater than the bound in \eqref{upperbound}. To this end, we fix $K > 0$ satisfying
\begin{equation}\label{choiceofK}
K^2 > \max \left\{\frac{c_2}{c_1 b_B^{2n}}, K_1^2 \right\}.
\end{equation}
\begin{theorem}\label{growthlemma} (``Growth Lemma'' under \hyperref[H1]{H1}) Let $z_0 \in \Omega$, and consider an open set $D \subseteq \mathcal{Q}^1_r(z_0)\subset \Omega$ such that $D\cap \mathcal{Q}^2_r(z_0)\neq \emptyset$. Suppose $u\in C^2(D) \cap C\left(\overline{D}\right)$ is nonnegative in $D$, vanishes on $\partial D\cap \mathcal{Q}^1_r(z_0)$, and satisfies $\mathcal{L}_A u \geq 0$ in $D$. Assume, moreover, that the Cordes-Landis condition \hyperref[H1]{H1} holds for the operator $\mathcal{L}_A$. Then there exists a structural constant $\eta > 0$ such that
$$\sup\limits_D u \geq \left(1 + \eta \dfrac{|\mathcal{Q}^3_r(z_0) \backslash D|}{|\mathcal{Q}^3_r(z_0)|} \right) \sup\limits_{D \cap \mathcal{Q}^2_r(z_0)} u.$$
\end{theorem}
\begin{proof}
By translation invariance of the class of operators under consideration, we may assume $z_0 = 0\in\Omega$. We may also assume $u$ is non-trivial, and so $D$ has limit points on $S^1_r \cup \left(B_{Kr}(0) \times \left\{-b_Br^2 \right\} \right)$ by the weak maximum principle \eqref{MP}. Recall from Section \ref{secest} that $S^1_r=\partial B_{Kr}(0)\times [-b_Br^2 ,0]$. Let $E = \mathcal{Q}^3_r \backslash D$ and consider the function $U_E$ defined in \eqref{defUE} with the choice of $\Gamma_{s,\beta}$ as in subSection \ref{cordeslandiscondition} (recall \eqref{defh1bs}). If we call $C$ the positive structural constant given in \eqref{boundint} such that $\sup U_{\mathcal{Q}^3_1}\leq C$, then we have by \eqref{maxr}
\begin{equation}\label{maxrE}
\sup{U_E}\leq \sup{U_{\mathcal{Q}^3_r}}\leq C r^{Q+2-sQ}.
\end{equation}
Moreover, by the bounds \eqref{upperbound} and \eqref{lowerbound}, we have
\begin{equation}\label{boundsonsides}
\sup_{S^1_r} U_E \leq \frac{e^{-\mu_2}}{(b_B r^2)^{s\frac{Q}{2}}}|E|\qquad\mbox{with }\mu_2:=\frac{c_2}{b_B^{2n+1}},
\end{equation}
and
\begin{equation}\label{boundsontop}
\inf_{\mathcal{Q}^2_r} U_E \geq \frac{e^{-\mu_1}}{(b_B r^2)^{s\frac{Q}{2}}}|E|\qquad\mbox{with }\mu_1:=\frac{c_1 K^2}{b_B}.
\end{equation}
Consider the auxiliary function
$$v(z) = \sup_D u \left(1 - \frac{U_E(z)}{C r^{Q+2-sQ}} + \frac{|E|}{C r^{Q+2}}\frac{e^{-\mu_2}}{(b_B)^{s\frac{Q}{2}} }\right).$$
Then $v$ is non-negative everywhere by \eqref{maxrE}. Since $\mathcal{L}_A U_E(z)\geq 0$ for all $z\in\Omega\smallsetminus \overline{E}$ by Lemma \ref{lemuu}, we have $\mathcal{L}_A v \leq 0\leq \mathcal{L}_A u$ in $D$. We now want to compare $v$ and $u$ on the portion of $\partial D$ required to apply the weak maximum principle. For this purpose, we define the sets $\gamma:=\partial D\cap \mathcal{Q}^1_r$, $\gamma_1 := \overline{D} \cap (B_{Kr}(0) \times \left\{-b_Br^2 \right\})$, and $\gamma_2 := \overline{D} \cap S^1_r$. Since $u = 0$ on $\gamma$, $v \geq u$ on $\gamma$. Recall also that $U_E$ is a continuous function. Since $U_E(z) = 0$ for $z \in B_{Kr}(0) \times \left\{-b_B r^2 \right\}$, we then have $v(z) \geq \sup_D u \geq u(z)$ for all $z \in \gamma_1$. Finally, for $z \in \gamma_2$, we have by \eqref{boundsonsides}
$$ v(z) \geq \sup_D u \left(1 - \dfrac{\sup_{S^1_r} U_E}{C r^{Q+2-sQ}} + \frac{|E|}{C r^{Q+2}}\frac{e^{-\mu_2}}{(b_B)^{s\frac{Q}{2}}}\right)\geq \sup_D u. $$
Thus $v \geq u$ on $\gamma_2$. By the weak maximum principle, it follows that $v \geq u$ in $D$. Hence, for $z \in D \cap \mathcal{Q}^2_r\neq \emptyset$, we have by \eqref{boundsontop}
\begin{align*}
u(z) & \leq \sup_D u \left(1 - \frac{\inf_{\mathcal{Q}^2_r} U_E}{C r^{Q+2-sQ}} + \frac{|E|}{C r^{Q+2}}\frac{e^{-\mu_2}}{(b_B)^{s\frac{Q}{2}}}\right)
\leq \sup_D u \left(1 - \frac{|E|}{C r^{Q+2}}\frac{e^{-\mu_1}}{ (b_B)^{s\frac{Q}{2}}} + \frac{|E|}{C r^{Q+2}}\frac{e^{-\mu_2}}{(b_B)^{s\frac{Q}{2}}}\right) \\
& = \sup_D u \left(1 - \frac{e^{-\mu_1} - e^{-\mu_2}}{C \, (b_B )^{s\frac{Q}{2}}} \frac{|E|}{r^{Q+2}} \right).
\end{align*}
By \eqref{choiceofK}, we have $\mu_1 < \mu_2$. Hence, we can define $\bar{\eta} := \dfrac{e^{-\mu_1} - e^{-\mu_2}}{C\,b_B^{sQ/2}} > 0$ and we can write
$$\bar{\eta}\, \frac{|E|}{r^{Q+2}}=\bar{\eta}\,\frac{|\mathcal{Q}^3_r \backslash D|}{r^{Q+2}}=\left(\bar{\eta}|\mathcal{Q}^3_1|\right)\frac{|\mathcal{Q}^3_r \backslash D|}{|\mathcal{Q}^3_r|}=:\eta\,\frac{|\mathcal{Q}^3_r \backslash D|}{|\mathcal{Q}^3_r|}.$$
This completes the proof.
\end{proof}
The proof of Theorem \ref{growthlemma} also allows us to obtain the following version of the growth lemma under the condition \hyperref[H2]{H2}. More specifically, we assume the continuity assumption \hyperref[H2]{H2} holds for the operator $\mathcal{L}_A$ and we let $\epsilon_0$ be the constant from Lemma \ref{barriersforH2} corresponding to the choice $s_0 := \frac{1}{Q}$, which we fix from here onwards.
\begin{theorem}\label{growthlemmaH2} (``Growth Lemma'' under \hyperref[H2]{H2}) Let $z_0 \in \Omega$, and suppose $0 < r \leq \frac{\epsilon_0}{K}$. Consider an open set $D \subseteq \mathcal{Q}^1_r(z_0)\subset \Omega$ such that $D\cap \mathcal{Q}^2_r(z_0)\neq \emptyset$. Suppose $u\in C^2(D) \cap C\left(\overline{D}\right)$ is nonnegative in $D$, vanishes on $\partial D\cap \mathcal{Q}^1_r(z_0)$, and satisfies $\mathcal{L}_A u \geq 0$ in $D$. Assume, moreover, that the continuity condition \hyperref[H2]{H2} holds for the operator $\mathcal{L}_A$. Then there exists a constant $\eta > 0$ such that
$$\sup\limits_D u \geq \left(1 + \eta \dfrac{|\mathcal{Q}^3_r(z_0) \backslash D|}{|\mathcal{Q}^3_r(z_0)|} \right) \sup\limits_{D \cap \mathcal{Q}^2_r(z_0)} u.$$
\end{theorem}
\begin{proof} The proof is essentially that of Theorem \ref{growthlemma}. The only modification is that the function $U_E$ is now constructed with the choice of $\Gamma_{s,\beta}$ as in subSection \ref{continuityassumption}, see \eqref{pick}. By Lemma \ref{barriersforH2}, $U_E$ is a subsolution only inside a cylinder $\mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0)$ of size $\epsilon_0$ depending on the modulus of continuity $\omega$ for the coefficients $A(\cdot)$. The assumption $0 < r \leq \frac{\epsilon_0}{K}$ (recall that $K\geq 1 \geq\sqrt{b_B}$) yields $D \subseteq \mathcal{Q}^1_r(z_0)\subset \Omega\cap \mathcal{Q}_{\epsilon_0}^{-\epsilon_0^2,\epsilon_0^2}(z_0)$.
\end{proof}
We provide an immediate application of Theorem \ref{growthlemma} and Theorem \ref{growthlemmaH2} by showing oscillation decay and H\"older continuity of solutions to $\mathcal{L}_A u = 0$. Recall that the oscillation of a function $u$ over a set $E$ is defined to be $\mathrm{osc}_{E} u := \sup_E u - \inf_E u$.
\begin{corollary}\label{oscillationdecayandHolder}(Oscillation Decay and Local H\"older Continuity under \hyperref[H1]{H1}) Suppose the operator $\mathcal{L}_A$ satisfies the hypothesis \hyperref[H1]{H1}. There exists a structural constant $\theta > 1$ such that if $u$ is a classical solution of $\mathcal{L}_A u = 0$ in an open set $D\subseteq \Omega$ and $\mathcal{Q}^1_r(z_0)\Subset D$, then
\begin{equation}\label{oscdecay}
\mathrm{osc}_{\mathcal{Q}^1_{r}(z_0)} u \geq \left(1 + \frac{\eta}{4}\right) \ \mathrm{osc}_{\mathcal{Q}^1_{r/\theta}(z_0)} u.
\end{equation}
Consequently, there exists a structural constant $\alpha$ and, for any $\rho>0$, a positive constant $C_\rho$ such that
\begin{equation}\label{BHolderestimate}
|u(z) - u(\zeta)|\leq C_\rho\left\|\zeta^{-1}\circ z\right\|^{\alpha}_B \left\|u\right\|_{L^{\infty}(D)}\quad\mbox{ for all }z,\zeta\mbox{ such that }\mathcal{Q}^1_\rho(z),\mathcal{Q}^1_\rho(\zeta)\Subset D.
\end{equation}
\end{corollary}
\begin{proof} We first prove \eqref{oscdecay}. Let $\theta := \frac{K}{\sigma_0}$. Recalling that $\theta\geq 2$, we have $\mathcal{Q}^1_{r/\theta}(z_0) \subseteq \mathcal{Q}^2_r(z_0)$ and then $\mathrm{osc}_{\mathcal{Q}^1_{r/\theta}(z_0)} u \leq \mathrm{osc}_{\mathcal{Q}^2_{r}(z_0)} u$. Consider the function
$$v = 2 u - \left(\sup\limits_{\mathcal{Q}^2_r(z_0)} u + \inf\limits_{\mathcal{Q}^2_r(z_0)} u\right),$$
and let $D^+$ be the set of points in $\mathcal{Q}^1_r(z_0)$ where $v > 0$. We may assume, without loss of generality, that $|\mathcal{Q}^3_r(z_0) \backslash D^+| \geq \frac{1}{2}|\mathcal{Q}^3_r(z_0)|$; otherwise, consider $-v$ instead of $v$. Applying Theorem \ref{growthlemma} with $D = D^+$, we obtain
\begin{eqnarray*}
2\mathrm{osc}_{\mathcal{Q}^1_r(z_0)} u - \mathrm{osc}_{\mathcal{Q}^2_r(z_0)} u&\geq& \sup\limits_{\mathcal{Q}^1_r(z_0)} v \geq \left(1 + \eta \dfrac{|\mathcal{Q}^3_r(z_0) \backslash D^+|}{|\mathcal{Q}^3_r(z_0)|} \right) \sup\limits_{D^+ \cap \mathcal{Q}^2_r(z_0)} v \geq \left(1 + \dfrac{\eta}{2} \right) \sup\limits_{\mathcal{Q}^2_r(z_0)} v \\
&=& \left(1 + \dfrac{\eta}{2} \right)\mathrm{osc}_{\mathcal{Q}^2_r(z_0)} u.
\end{eqnarray*}
which implies
\begin{equation}\label{oscdec}
\mathrm{osc}_{\mathcal{Q}^1_{r}(z_0)} u \geq P \,\mathrm{osc}_{\mathcal{Q}^2_{r}(z_0)} u
\geq P \,\mathrm{osc}_{\mathcal{Q}^1_{r/\theta}(z_0)} u\quad\,\mbox{with }\,\,P := 1 + \dfrac{\eta}{4}.
\end{equation}
To prove the estimate \eqref{BHolderestimate}, fix $\rho>0$ and let $z,\zeta$ be arbitrary points in $D$ such that $\mathcal{Q}^1_\rho(z),\mathcal{Q}^1_\rho(\zeta)\Subset D$. With no loss of generality, we may assume $t\leq \tau$. We have two cases: either $z\in \mathcal{Q}^1_\rho(\zeta)$ or $z\notin \mathcal{Q}^1_\rho(\zeta)$.\\
If $z\in \mathcal{Q}^1_\rho(\zeta)$, choose $m_0 \in \mathbb{N}\cup\{0\}$ such that $z \in \mathcal{Q}^1_{\frac{\rho}{\theta^{m_0}}}(\zeta)$ and $z \notin \mathcal{Q}^1_{\frac{\rho}{\theta^{m_0+1}}}(\zeta)$. Hence, $\left\|\zeta^{-1}\circ z\right\|_B \geq \min\left\{K, \sqrt{b_B} \right\} \frac{\rho}{\theta^{m_0+1}}=\sqrt{b_B}\frac{\rho}{\theta^{m_0+1}}$. Applying \eqref{oscdec} recursively, we obtain
$$\mathrm{osc}_{\mathcal{Q}^1_{\frac{\rho}{\theta^{m_0}}}(\zeta)} u \leq \frac{1}{P^{m_0}} \mathrm{osc}_{\mathcal{Q}^1_\rho} u \leq \frac{2 P \left\|u\right\|_{L^{\infty}(D)}}{P^{m_0+1}}.$$
Writing $P^{m_0+1} = (\theta^{\log_{\theta} P})^{m_0 + 1} = (\theta^{m_0+1})^{\log_{\theta} P}$ and letting $\alpha:=\log_{\theta} P$, we get
$$|u(z) - u(\zeta)| \leq \text{osc}_{\mathcal{Q}^1_{\frac{\rho}{\theta^{m_0}}}(\zeta)} u \leq \frac{2 P \left\|u\right\|_{L^{\infty}(D)}}{b_B^{\frac{\alpha}{2}}\rho^{\alpha}}\left(\sqrt{b_B}\frac{\rho}{\theta^{m_0+1}}
\right)^{\alpha} \leq \frac{2 P \left\|u\right\|_{L^{\infty}(D)}}{b_B^{\frac{\alpha}{2}}\rho^{\alpha}} \left\|\zeta^{-1}\circ z\right\|^{\alpha}_B.$$
On the other hand, if $z\notin \mathcal{Q}^1_\rho(\zeta)$ we simply have $\left\|\zeta^{-1}\circ z\right\|_B \geq \sqrt{b_B}\rho$ and then
$$|u(z) - u(\zeta)| \leq 2 \left\|u\right\|_{L^{\infty}(D)}\leq \frac{2 \left\|u\right\|_{L^{\infty}(D)}}{b_B^{\frac{\alpha}{2}}\rho^{\alpha}} \left\|\zeta^{-1}\circ z\right\|^{\alpha}_B.$$
Combining the two possibilities, we obtain the desired estimate \eqref{BHolderestimate} with the choice $C_\rho=2 P b_B^{-\frac{\alpha}{2}}\rho^{-\alpha}$.
\end{proof}
In order to establish the corresponding version of Corollary \ref{oscillationdecayandHolder} under the hypothesis \hyperref[H2]{H2}, we notice that the proof can be carried out in the same manner simply by considering $\mathcal{Q}^1_r(z_0)\subset D$ with $0<r\leq\frac{\epsilon_0}{K}$. The constant $C_{\rho}$ will now depend additionally on $\epsilon_0$; however, the constant $\alpha$ remains independent of $\epsilon_0$.
\begin{remark}
The regularity estimate \eqref{BHolderestimate} is equivalent to local H\"older continuity in the standard sense. One can see this by comparing the $|\cdot|_B$-norm with the Euclidean norm as in \eqref{triangleB}--\eqref{benne} (see also \cite[Definition 1.2 and Proposition 2.1]{Poli}). \end{remark}
\section{Harnack Inequality}\label{harn}
In this final section, we prove the Harnack inequality for non-negative solutions to $\mathcal{L}_A u = 0$ using the growth lemma. We follow closely the approach outlined by Landis in \cite[Lemma 9.1 and Theorem 9.1]{Landis} and make a number of necessary modifications to adapt his proof to our setting.
\begin{lemma}\label{incylinder}
There exist structural constants $C_1,C_2>0$ such that:
\begin{itemize}
\item[(i)] for any $R>0$ and any $0\leq\delta_1<\delta_2\leq\frac{1}{2}$, if $\rho\leq C_1 R (\delta_2-\delta_1)^{n+\frac{1}{2}}$ then
$$\mathcal{Q}_{K \rho}^{-b_B\rho^2,0}(z_0)\subseteq \mathcal{Q}_{R\left(\frac{1}{2}+\delta_2\right)}^{-b_B R^2\left(\frac{1}{2}+\delta_2\right),0}\qquad \forall\,z_0\in \overline{\mathcal{Q}_{R\left(\frac{1}{2}+\delta_1\right)}^{-b_B R^2\left(\frac{1}{2}+\delta_1\right),0}}; $$
\item[(ii)] for any $R>0$ and any $0\leq\delta_1<\delta_2\leq1$, if $\rho\leq C_2 R (\delta_2-\delta_1)^{n+\frac{1}{2}}$ then
$$\mathcal{Q}_{\rho}^{-b_B\rho^2,0}(z_0)\subseteq \mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\delta_2\right)}^{-\frac{b_B}{4} R^2\left(3+\delta^2_2\right),-\frac{b_B}{2} R^2}\smallsetminus \mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\delta_1\right)}^{-\frac{b_B}{4} R^2\left(3+\delta^2_1\right),-\frac{b_B}{2} R^2}\qquad \forall\,z_0\in \partial_p \mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\frac{\delta_1+\delta_2}{2}\right)}^{-\frac{b_B}{4} R^2\left(3+\frac{(\delta_1+\delta_2)^2}{4}\right),-\frac{b_B}{2} R^2}. $$
\end{itemize}
\end{lemma}
We postpone the proof of this lemma to the end of the section. For now, we use it to prove the following important consequence of the growth lemma.
\begin{lemma}\label{lemma91} Let $\bar{z} \in \Omega$, and consider an open set $D \subseteq \mathcal{Q}^{-b_B R^2,0}_R(\bar{z})\subset \Omega$ such that $D^+:=D\cap \mathcal{Q}^{-\frac{1}{2}b_B R^2,0}_{\frac{1}{2}R}(\bar{z})\neq \emptyset$. Suppose $u\in C^2(D)\cap C\left(\overline{D}\right)$, nonnegative in $D$, vanishes on $\partial D\cap \mathcal{Q}^{-b_B R^2,0}_R(\bar{z})$, and satisfies $\mathcal{L}_A u = 0$ in $D$. Assume, moreover, that the Cordes-Landis condition \hyperref[H1]{H1} holds for the operator $\mathcal{L}_A$. Then, for any $M>1$, there exists $\delta > 0$ (depending on $M$ and on structural constants) such that, if $|D|\leq \delta R^{Q+2}$, we have
$$\sup\limits_D{u} \geq M \sup\limits_{D^+}{u}\geq M u(\bar{z}).$$
\end{lemma}
\begin{proof} By translation invariance we can assume $\bar{z}=0\in\Omega$. Let $\eta$ be the constant in Theorem \ref{growthlemma}. For any $M>1$, let $m$ be the smallest natural number such that $\left(1+\frac{\eta}{2}\right)^m>M$. For $i\in\{0,\ldots,m\}$, denote
$$\mathcal{Q}^{(i)}=\mathcal{Q}^{-\frac{1}{2}b_B R^2\left(1 + \frac{i}{m}\right),0}_{\frac{1}{2}R\left(1 + \frac{i}{m}\right)}.$$
For any $i\in\{0,\ldots,m-1\}$ we consider the point $z^i=(x^i,t^i)$ in the parabolic boundary of $\mathcal{Q}^{(i)}$ such that $u(z^i)=\sup_{D\cap \mathcal{Q}^{(i)}}{u}$. Let us now denote
$$\mathcal{Q}^{1,(i)}_\rho=\mathcal{Q}_{K \rho}^{-b_B \rho^2, 0}(z^i).$$
By recalling that $K\geq 1$ and by exploiting Lemma \ref{incylinder} (item $(i)$, with $\delta_1=\frac{i}{2m}$, $\delta_2=\frac{i+1}{2m}$) we know that
$$\rho=\frac{C_1}{(2m)^{n+\frac{1}{2}}} R \,\,\,\,\mbox{ yields }\,\,\,\, \mathcal{Q}^{1,(i)}_\rho\subseteq \mathcal{Q}^{(i+1)}.$$
We are going to prove the statement of the lemma with the choice
$$\delta= \frac{1}{2}C_1^{Q+2}\frac{\left|\mathcal{Q}^{-b_B,-\frac{1}{2}b_B}_{\sigma_0}\right|}{(2m)^{\left(n+\frac{1}{2}\right)(Q+2)}}.$$
In fact, defining $\mathcal{Q}^{2,(i)}_\rho:=\mathcal{Q}_{\sigma_0 \rho}^{-\frac{1}{4}b_B \rho^2, 0}(z^i)$ and $\mathcal{Q}^{3,(i)}_\rho:=\mathcal{Q}_{\sigma_0 \rho}^{-b_B \rho^2, -\frac{1}{2}b_B \rho^2}(z^i)$, the assumption $|D|\leq \delta R^{Q+2}$ implies
$$|D\cap \mathcal{Q}^{3,(i)}_\rho|\leq \delta R^{Q+2}=\frac{1}{2} |\mathcal{Q}^{3,(i)}_\rho|\quad\mbox{ and hence }\quad |\mathcal{Q}^{3,(i)}_\rho\smallsetminus D|= |\mathcal{Q}^{3,(i)}_\rho| - |D\cap \mathcal{Q}^{3,(i)}_\rho|\geq \frac{1}{2}|\mathcal{Q}^{3,(i)}_\rho|.$$
Applying Theorem \ref{growthlemma} in the cylinder $\mathcal{Q}^{1,(i)}$, and using the inclusion $\mathcal{Q}^{1,(i)}_\rho\subseteq \mathcal{Q}^{(i+1)}$, we get
$$\sup\limits_{D\cap \mathcal{Q}^{(i+1)}}{u} \geq \sup\limits_{D\cap \mathcal{Q}^{1,(i)}_\rho}{u} \geq \left(1 + \frac{\eta}{2} \right) \sup\limits_{D \cap \mathcal{Q}^{2,(i)}_\rho}{u}\geq \left(1 + \frac{\eta}{2} \right) u(z^i) = \left(1 + \frac{\eta}{2} \right) \sup\limits_{D \cap \mathcal{Q}^{(i)}}{u}. $$
This holds true for every $i\in\{0,\ldots,m-1\}$. Therefore, since $ \mathcal{Q}^{(0)}= \mathcal{Q}^{-\frac{1}{2}b_B R^2,0}_{\frac{1}{2}R}$ and $\mathcal{Q}^{(m)}=\mathcal{Q}^{-b_B R^2,0}_R$, we finally obtain
$$\sup\limits_D{u} \geq \left(1+\frac{\eta}{2}\right)^m \sup\limits_{D^+}{u} \geq M \sup\limits_{D^+}{u}\geq M u(\bar{z}).$$
\end{proof}
We are finally ready to show the proof of the Harnack inequality. We begin with the proof of Theorem \ref{harnackineqH1}
\begin{proof}[Proof of Theorem \ref{harnackineqH1}]
Assume, without loss of generality, that $z_0 = 0$ and $\sup_{\mathcal{Q}_r^-} u = 2$. The aim is to find a structural lower bound for $u$ on $\mathcal{Q}_r^+$. Let us recall the definitions of the cylinders
$$\mathcal{Q}^2_r=\mathcal{Q}_{\sigma_0 r}^{-\frac{1}{4}b_B r^2, 0} \qquad \mathcal{Q}^3_r=\mathcal{Q}_{\sigma_0 r}^{-b_B r^2, -\frac{1}{2}b_B r^2}.$$
Notice that $\mathcal{Q}_r^+ \subset \mathcal{Q}^2_r$ and $\mathcal{Q}_r^- \subset \mathcal{Q}^3_r$. Consider the set $G := \left\{z \in \mathcal{Q}^3_r : u(z) > 1 \right\}$. Let $\delta > 0$ be the number from Lemma \ref{lemma91} corresponding to the choice of $M = 2^{1+(n+\frac{1}{2})(Q+2)}$, and define the structural constant
$$\epsilon_0 := \left(\frac{C_2}{2^{n+\frac{1}{2}}}\right)^{Q+2} \delta,$$
where $C_2$ is the constant appearing in Lemma \ref{incylinder}, item $(ii)$. We are faced with two possibilities:
\begin{itemize}
\item[-] Case 1: $|G| \geq \epsilon_0 r^{Q+2}$, or
\item[-] Case 2: $|G| < \epsilon_0 r^{Q+2}$.
\end{itemize}
For Case 1, we consider the function $w = 1 - u$. With the intent of applying Theorem \ref{growthlemma}, we define the set $D := \left\{z \in \mathcal{Q}^1_r : w(z) > 0 \right\}$. We may assume $D \cap \mathcal{Q}^2_r \neq \emptyset$, for otherwise $u\geq 1$ in $\mathcal{Q}^2_r\supset \mathcal{Q}_r^+$. Since $u$ is non-negative, we have $w \leq 1$ in $\mathcal{Q}^1_r$. Furthermore, $G \subset \mathcal{Q}^3_r \backslash D$, and so $|\mathcal{Q}^3_r \backslash D| \geq \epsilon_0 r^{Q+2}$. It follows from Theorem \ref{growthlemma} applied to $w$ that
$$1 \geq \sup_{\mathcal{Q}^1_r} w \geq \left(1 + \eta\frac{\epsilon_0 }{|\mathcal{Q}^3_1|}\right) \sup_{\mathcal{Q}^2_r \cap D} w \geq \left(1 + \eta\frac{\epsilon_0 }{|\mathcal{Q}^3_1|}\right) \sup_{\mathcal{Q}_r^+} w.$$
Thus,
$$\inf_{\mathcal{Q}_r^+} u \geq \frac{\hat{C}}{1 + \hat{C}} \qquad \text{ where } \hat{C} := \eta\frac{\epsilon_0 }{|\mathcal{Q}^3_1|}.$$
Consequently, \eqref{harnackH1} follows when Case 1 holds.
\noindent For Case 2, we carry out an iteration procedure, which we describe in the following steps:
{\bf Step 1:} Set
$$\mathcal{Q}^{(s)}:=\mathcal{Q}_{r\frac{\sigma_0}{2}(1+s)}^{-\frac{b_B}{4}r^2(3+s^2),-\frac{b_B}{2}r^2}, \qquad s > 0.$$
Notice that $\mathcal{Q}^{(0)} = \mathcal{Q}_r^-$, while $\mathcal{Q}^{(1)} = \mathcal{Q}^3_r$. Consider the family of sets
$$G^{(0)}_s:=G\cap(\mathcal{Q}^{(s)} \backslash \mathcal{Q}^{(0)}), \qquad 0 < s < 1.$$
Observe that
\begin{equation}\label{valueatonehalf}
\left|G^{(0)}_{1/2}\right| \leq |G| < \epsilon_0 r^{Q+2} = \left(C_2 r\left(\frac{1}{2} \right)^{n+\frac{1}{2}}\right)^{Q+2} \delta.
\end{equation}
We claim
\begin{equation}\label{claimbigO}
\left|G^{(0)}_s\right| \gtrsim s^2 \quad \text{ as } s \rightarrow 0^+
\end{equation}
To see this, consider the point $\bar{\zeta}=(\bar{\xi}, \bar{\tau})\in\partial_p \mathcal{Q}^{(0)}$ such that $u(\bar{\zeta})=2=\sup_{\mathcal{Q}^{(0)}} u$. By continuity of $u$, there exists a small neighborhood $U_{\bar{\xi}}\times(\bar{\tau}-\delta^2,\bar{\tau}+\delta^2)$ of $\bar{\zeta}$ in which $u>1$. We face two possibilities: either $\bar{\zeta}$ is on the ``base'' of the cylinder $\mathcal{Q}^{(0)}$, in which case $\bar{\xi}\in \overline{B_{r\frac{\sigma_0}{2}}(0)}$ and $\bar{\tau}=-\frac{3}{4}b_Br^2$, or $\bar{\zeta}$ is on the ``lateral side'' of $\mathcal{Q}^{(0)}$, in which case $|\bar{\xi}|_B=r\frac{\sigma_0}{2}$ and $\bar{\tau}\in(-\frac{3}{4}b_Br^2,-\frac{1}{2}b_Br^2]$. In the first case, it suffices to notice that, up to restricting $U_{\bar{\xi}}$ and for $s$ small enough with respect to $\delta$ and $r$, we have $U_{\bar{\xi}}\times(\bar{\tau}-\frac{1}{4}b_Br^2s^2,\bar{\tau})\subset G^{(0)}_s$ and thus $\left|G^{(0)}_s\right| \gtrsim s^2$. In the second case, up to restricting $\delta$ and for $s$ small enough, we have instead a sector $C_s\subset B_{r\frac{\sigma_0}{2}(1+s)}(0)\smallsetminus B_{r\frac{\sigma_0}{2}}(0)$ such that $C_s\times (\bar{\tau}-\delta^2,\bar{\tau})\subset G^{(0)}_s$. Since $|B_{r\frac{\sigma_0}{2}(1+s)}(0)\smallsetminus B_{r\frac{\sigma_0}{2}}(0)|\sim s$ by the dilation properties, we deduce that $\left|G^{(0)}_s\right| \gtrsim s\geq s^2$. This completes the proof of the claim \eqref{claimbigO}.
By \eqref{valueatonehalf} and \eqref{claimbigO}, there exists $s_1 \in (0,1/2)$ such that $$\left|G^{(0)}_{s_1}\right| = \left(C_2 \cdot r \cdot s_1^{n+\frac{1}{2}}\right)^{Q+2} \delta.$$
Let $\zeta_0 \in \partial_p \mathcal{Q}^{(s_1/2)}$ be such that $u(\zeta_0) \geq 2$. Using Lemma \ref{incylinder}, item $(ii)$ with $\delta_1 = 0$, $\delta_2 = s_1$, we obtain the existence of a cylinder $\mathcal{Q}_{\rho}^{-b_B\rho^2,0}(\zeta_0) \subseteq \mathcal{Q}^{(s_1)} \backslash \mathcal{Q}^{(0)}$, where
$$\rho = C_2 \cdot r \cdot s_1^{n+\frac{1}{2}}.$$
Define $\mathcal{Q}_{(0)} := \mathcal{Q}_{\rho}^{-b_B\rho^2,0}(\zeta_0)$ and $D_{(0)}:=G \cap \mathcal{Q}_{(0)}$. Notice that $\zeta_0 \in D_{(0)}$ and
$$|D_{(0)}| \leq \left|G^{(0)}_{s_1}\right| = \delta \rho^{Q+2}.$$
Consider the function $v := u - 1$. The measure estimate for $D_{(0)}$ above allows us to apply Lemma \ref{lemma91} to $v$. Noticing that $v(\zeta_0) \geq 1$, we thus conclude
$$\sup_{D_{(0)}} u \geq \sup_{D_{(0)}} v \geq M.$$
This implies, by the weak maximum principle \eqref{MP}, that
$$\sup_{\partial_p \mathcal{Q}^{(s_1)}} u \geq M.$$
{\bf Step 2:} For any index $\mathbb{N}\ni \ell \geq 1$, define $G_{\ell} := \left\{z \in \mathcal{Q}^3_r : u(z) > \left(\frac{M}{2}\right)^{\ell}\right\}$. Let $\mathcal{Q}^{(s_{\ell})}$ be the cylinder satisfying
$$\sup_{\partial_p \mathcal{Q}^{(s_{\ell})}} u \geq 2 \left(\frac{M}{2}\right)^{\ell}.$$
If $s_{\ell} \geq 1/2$, we proceed directly to step 3. Otherwise, we have $1 > 1 - s_{\ell} > 1/2$. For any $s > 0$, define the family of sets
$$G^{(\ell)}_s:=G_{\ell} \cap\left(\mathcal{Q}^{(s + s_{\ell})} \backslash \mathcal{Q}^{(s_{\ell})}\right), \qquad 0 < s < 1 - s_{\ell}.$$
Since $G_{\ell} \subset G$, we have
$$\left|G^{(\ell)}_{1/2} \right| \leq |G| \leq \left(C_2 r\left(\frac{1}{2} \right)^{n+\frac{1}{2}}\right)^{Q+2} \delta.$$
Arguing as in \eqref{claimbigO}, we also have $\left|G^{(\ell)}_s\right| \gtrsim s^2$ as $s \rightarrow 0^+$. Hence, there exists $\rho_{\ell } \in (0, \frac{1}{2}) $ such that
$$\left|G^{(\ell)}_{\rho_{\ell}}\right| = \left(C_2 \cdot r \cdot \rho_{\ell}^{n+\frac{1}{2}} \right)^{Q+2} \delta.$$
Let $\zeta_{\ell} \in \partial_p \mathcal{Q}^{(s_{\ell} + \frac{\rho_{\ell}}{2})}$ be such that $u(\zeta_{\ell}) \geq 2 \left(\frac{M}{2}\right)^{\ell}$. Defining $s_{\ell+1} := s_{\ell} + \rho_{\ell}$ and using Lemma \ref{incylinder}. item $(ii)$ with $\delta_1 = s_{\ell}$, $\delta_2 = s_{\ell+1}$, we obtain the existence of a cylinder $\mathcal{Q}_{\rho}^{-b_B\rho^2,0}(\zeta_{\ell}) \subseteq \mathcal{Q}^{(s_{\ell + 1})} \backslash \mathcal{Q}^{(s_{\ell})}$, where
$$\rho = C_2 \cdot r \cdot \rho_{\ell}^{n+\frac{1}{2}}.$$
Define $\mathcal{Q}_{(\ell)} := \mathcal{Q}_{\rho}^{-b_B\rho^2,0}(\zeta_{\ell})$ and $D_{(\ell)}:= G \cap \mathcal{Q}_{(\ell)}$. Notice that $\zeta_{\ell} \in D_{(\ell)}$ and
$$|D_{(\ell)}| \leq \left|G^{(\ell)}_{\rho_{\ell}}\right| = \left(C_2 \cdot r \cdot \rho_{\ell}^{n+\frac{1}{2}} \right)^{Q+2} \delta = \delta \rho^{Q+2}.$$
Consider the function $v := u - \left(\frac{M}{2}\right)^{\ell}$. The measure estimate for $D_{(\ell)}$ above allows us to apply Lemma \ref{lemma91} to $v$. Noticing that $v(\zeta_{\ell}) \geq \left(\frac{M}{2}\right)^{\ell}$, we thus conclude
$$\sup_{D_{(\ell)}} u \geq \sup_{D_{(\ell)}} v \geq M \cdot \left(\frac{M}{2}\right)^{\ell} = 2 \left(\frac{M}{2}\right)^{\ell +1}.$$
This implies, by the weak maximum principle, that
$$\sup_{\partial_p \mathcal{Q}^{(s_{\ell+1})}} u \geq 2 \left(\frac{M}{2}\right)^{\ell +1}.$$
{\bf Step 3:} There must exist a smallest integer $k \geq 1$ such $s_{k+1} \geq 1/2$, for otherwise the function $u$ would be unbounded on $\mathcal{Q}^3_r$. This implies Step 2 must terminate after finitely many iterations. By denoting $\rho_0:=s_1$ and recalling the definition of $s_{k+1}$, we have $\rho_0 + \rho_1 + \cdots + \rho_k \geq \frac{1}{2}$ and $\rho_0 + \cdots + \rho_{k-1} < \frac{1}{2}$. For each $\ell \in \left\{0, \ldots, k \right\}$, we know that the corresponding set $G^{(\ell)}_{\rho_{\ell}}$ satisfies
$$\left|G^{(\ell)}_{\rho_{\ell}}\right| = \left(C_2 \cdot r \cdot \rho_{\ell}^{n+\frac{1}{2}} \right)^{Q+2} \delta,$$
and that $u > \left(\frac{M}{2}\right)^{\ell}$ on $G^{(\ell)}_{\rho_{\ell}}$ by definition. Since $\rho_0 + \cdots + \rho_k \geq \frac{1}{2}$, there must exist at least one index $i_0 \in\left\{0, \ldots, k \right\}$ such that
$$\rho_{i_0} \geq \left(\frac{1}{2}\right)^{i_0 + 2}.$$
Therefore, we have
$$\left|G^{(i_0)}_{\rho_{i_0}}\right| \geq \left(C_2 \cdot r \cdot 2^{-(i_0 + 2)(n+\frac{1}{2})} \right)^{Q+2} \delta,$$
and
$$u > \left(\frac{M}{2}\right)^{i_0} \ \text{on} \ G^{(i_0)}_{\rho_{i_0}}.$$
We now make one final use of Theorem \ref{growthlemma}. Consider the function $v := \left(\frac{M}{2}\right)^{i_0} -u$. Then $\mathcal{L}_A v = 0$ and $v \leq \left(\frac{M}{2}\right)^{i_0}$ since $u$ is non-negative on $\mathcal{Q}^1_r$. Define $D := \left\{z \in \mathcal{Q}^1_r : v(z) > 0 \right\}$. Then $G_{i_0} \subset \mathcal{Q}^3_r \backslash D$. Since $G^{(i_0)}_{\rho_{i_0}} \subset G_{i_0}$, we have from the measure estimate above that
$$\dfrac{|\mathcal{Q}^3_r \backslash D|}{|\mathcal{Q}^3_r|} \geq \dfrac{\left(C_2 \cdot r \cdot 2^{-(i_0 + 2)(n+\frac{1}{2})} \right)^{Q+2} \delta}{|\mathcal{Q}^3_1| r^{Q+2}} = \frac{C_2^{Q+2}\delta}{|\mathcal{Q}^3_1|} 2^{-(n+\frac{1}{2})(Q+2)(i_0+2)}.$$
Finally, we may assume that $\left\{v \geq 0 \right\} \cap \mathcal{Q}^2_r \neq \emptyset$; for otherwise, we would have $u \geq \left(\frac{M}{2}\right)^{i_0} \geq 1$ on $\mathcal{Q}^2_r\supset \mathcal{Q}_r^+$, and \eqref{harnackH1} would automatically follow. Thus, we may apply Theorem \ref{growthlemma} to $v$ and obtain
$$\left(\frac{M}{2}\right)^{i_0} \geq \sup_{\mathcal{Q}^1_r} v \geq \left(1 + \eta \frac{C_2^{Q+2}\delta}{|\mathcal{Q}^3_1|} 2^{-(n+\frac{1}{2})(Q+2)(i_0+2)} \right) \sup_{\mathcal{Q}^2_r\cap D} v \geq \left(1 + \eta \frac{C_2^{Q+2}\delta}{|\mathcal{Q}^3_1|} 2^{-(n+\frac{1}{2})(Q+2)(i_0+2)} \right) \sup_{\mathcal{Q}_r^+}v .$$
Inserting the definition of $v$ and recalling that $M = 2^{1+(n+\frac{1}{2})(Q+2)}$, we have
$$\left(\frac{M}{2}\right)^{i_0} \geq \left(1 + \eta \frac{C_2^{Q+2}\delta}{|\mathcal{Q}^3_1|} 2^{-2(n+\frac{1}{2})(Q+2)}\left(\frac{M}{2}\right)^{-i_0} \right) \left(\left(\frac{M}{2}\right)^{i_0} - \inf_{\mathcal{Q}_r^+} u \right).$$
Denoting $\hat{c} := \eta \frac{C_2^{Q+2}\delta}{|\mathcal{Q}^3_1|} 2^{-2(n+\frac{1}{2})(Q+2)}$, we get
$$\left(1+ \hat{c} \left(\frac{M}{2}\right)^{-i_0}\right) \inf_{\mathcal{Q}_r^+} u \geq \hat{c}.$$
Since $\left(\frac{M}{2}\right)^{-i_0} \leq 1$, we conclude that
$$\inf_{\mathcal{Q}_r^+} u \geq \frac{\hat{c}}{1 + \hat{c}}.$$
This establishes \eqref{harnackH1} when Case 2 holds, and finishes the proof of the theorem. \end{proof}
We next show the proof of Theorem \ref{harnackineqH2}.
\begin{proof}[Proof of Theorem \ref{harnackineqH2}]
For small radii we can follow the proof of Theorem \ref{harnackineqH1}. In fact, if
$$ r\leq r_0:=\frac{\epsilon_0}{K}\min\left\{1,\frac{1}{C_1C_2}4^{n+\frac{1}{2}}\right\}, $$
then we can invoke Theorem \ref{growthlemmaH2} and argue exactly as in the proofs of Lemma \ref{lemma91} and Theorem \ref{harnackineqH1}. This gives the existence of a structural constant $C_H$ such that, for all $\mathcal{Q}^1_r=\mathcal{Q}_{Kr}^{-b_B r^2,0}(z_0)\Subset\Omega$ and any $u \in C^2(\Omega)$ which is a nonnegative solution to $\mathcal{L}_A u=0$ in $\mathcal{Q}^1_r$, we have
\begin{equation}\label{toiterate}
\sup_{\mathcal{Q}_{\frac{\sigma_0}{2}r}^{-\frac{3b_B}{4}r^2,-\frac{b_B}{2}r^2}(z_0)} u \leq C_H \inf_{\mathcal{Q}_{\frac{\sigma_0}{2}r}^{-\frac{b_B}{4}r^2,0}(z_0)} u,\qquad\mbox{ for any }0<r\leq r_0.
\end{equation}
For $r_0 < r \leq 1$, one can use \eqref{toiterate} along with the existence of Harnack chains established in this context by Polidoro \cite[Section 3]{PoliA}. This proves \eqref{harnackH2} with some constants $\sigma<\frac{\sigma_0}{2}$ and $C>C_H$ depending on the modulus of continuity of the coefficients $\omega$ (i.e. on $\epsilon_0$).
\end{proof}
Finally, we provide the proof of Lemma \ref{incylinder}, as promised.
\begin{proof}[Proof of Lemma \ref{incylinder}] We first prove $(i)$. Fix $z_0=(x_0, t_0)$ in the closure of the cylinder $\mathcal{Q}_{R\left(\frac{1}{2}+\delta_1\right)}^{-b_B R^2\left(\frac{1}{2}+\delta_1\right),0}$, and fix any point $\bar{\zeta}=(\bar{\xi},\bar{\tau})\in \mathcal{Q}_{K\rho}^{-b_B\rho^2,0}(z_0)$. This means that there exists $\zeta=(\xi,\tau)\in \mathcal{Q}_{K\rho}^{-b_B\rho^2,0}$ such that $\bar{\zeta}=z_0\circ \zeta = (\xi + E(\tau)x_0, \tau + t_0)$. By definition we have
$$0>\tau + t_0>-b_B\rho^2 -b_B R^2\left(\frac{1}{2}+\delta_1\right) \geq -b_B R^2\left(\frac{1}{2}+\delta_2\right),$$
where the last inequality holds provided that
\begin{equation}\label{euno}
\rho\leq R\sqrt{\delta_2-\delta_1}.
\end{equation}
We have also to find conditions ensuring that $|\xi + E(\tau)x_0|_B<R\left(\frac{1}{2}+\delta_2\right)$. This is trivial if $x_0=0$. So, suppose $x_0\neq 0$. We start by noticing that
$$\max{\left\{\left|D_{\frac{\bar{\sigma}}{|x_0|_B}} x_0\right|, \left|D_{\frac{\bar{\sigma}}{|x_0|_B}} x_0\right|^{\frac{1}{2n+1}}\right\}}\geq \frac{1}{\bar{\sigma}}\left|D_{\frac{\bar{\sigma}}{|x_0|_B}} x_0\right|_B=1,$$
which yields
\begin{equation}\label{eee}
\left|D_{\frac{\bar{\sigma}}{|x_0|_B}} x_0\right|^{\frac{1}{2n+1}}=\min{\left\{\left|D_{\frac{\bar{\sigma}}{|x_0|_B}} x_0\right|, \left|D_{\frac{\bar{\sigma}}{|x_0|_B}} x_0\right|^{\frac{1}{2n+1}}\right\}}\leq\frac{1}{\sigma_0}\left|D_{\frac{\bar{\sigma}}{|x_0|_B}} x_0\right|_B=\frac{\bar{\sigma}}{\sigma_0}.
\end{equation}
By \eqref{commu}, \eqref{benne}, and \eqref{eee}, we get
\begin{eqnarray*}
\left|\left(E(\tau)-\mathbb{I}_N\right) x_0\right|_B&=& \frac{|x_0|_B}{\bar{\sigma}} \left| D_{\frac{\bar{\sigma}}{|x_0|_B}} \left(E(\tau)-\mathbb{I}_N\right) x_0\right|_B = \frac{|x_0|_B}{\bar{\sigma}} \left|\left(E\left(\frac{\bar{\sigma}^2\tau}{|x_0|^2_B}\right)-\mathbb{I}_N\right) D_{\frac{\bar{\sigma}}{|x_0|_B}}x_0\right|_B\\
&\leq&c(n,B)\frac{|x_0|_B}{\bar{\sigma}} \max\left\{\left|D_{\frac{\bar{\sigma}}{|x_0|_B}}x_0\right|^{\frac{1}{3}},\left|D_{\frac{\bar{\sigma}}{|x_0|_B}}x_0\right|^{\frac{1}{2n+1}}\right\}\max\left\{\left|\frac{\bar{\sigma}^2\tau}{|x_0|^2_B}\right|^{\frac{1}{2n+1}},\left|\frac{\bar{\sigma}^2\tau}{|x_0|^2_B}\right|^{\frac{n}{2n+1}}\right\}\\
&\leq& R\frac{c(n,B)}{\bar{\sigma}}\left(\frac{\bar{\sigma}}{\sigma_0}\right)^{\frac{2n+1}{3}} \frac{|x_0|_B}{R} \max\left\{\left(\frac{R}{|x_0|_B}\right)^{\frac{2}{2n+1}}\left|\frac{\bar{\sigma}^2\tau}{R^2}\right|^{\frac{1}{2n+1}},\left(\frac{R}{|x_0|_B}\right)^{\frac{2n}{2n+1}}\left|\frac{\bar{\sigma}^2\tau}{R^2}\right|^{\frac{n}{2n+1}}\right\}\\
&\leq&R\frac{c(n,B)}{\bar{\sigma}}\left(\frac{\bar{\sigma}}{\sigma_0}\right)^{\frac{2n+1}{3}} \left(\frac{|x_0|_B}{R}\right)^{\frac{1}{2n+1}} \max\left\{\left|\frac{\bar{\sigma}^2\tau}{R^2}\right|^{\frac{1}{2n+1}},\left|\frac{\bar{\sigma}^2\tau}{R^2}\right|^{\frac{n}{2n+1}}\right\},
\end{eqnarray*}
where we have used $|x_0|_B\leq R $ (since $\delta_1\leq \frac{1}{2}$) and $n\geq 1$. If in addition
\begin{equation}\label{edue}
\rho\leq \frac{R}{\bar{\sigma}\sqrt{b_B}},
\end{equation}
then we also have $\left|\frac{\bar{\sigma}^2\tau}{R^2}\right|\leq \frac{\bar{\sigma}^2 b_B\rho^2}{R^2}\leq 1$, and so
\begin{equation}\label{valbe}
\left|\left(E(\tau)-\mathbb{I}_N\right) x_0\right|_B\leq R\frac{c(n,B)}{\bar{\sigma}}\left(\frac{\bar{\sigma}}{\sigma_0}\right)^{\frac{2n+1}{3}} \left(\bar{\sigma}^2 b_B\frac{\rho^2}{R^2}\right)^{\frac{1}{2n+1}}.
\end{equation}
Hence, by the definition of $\xi, x_0$, and by \eqref{triangleB}, \eqref{valbe}, we get
$$|\xi + E(\tau)x_0|_B\leq |\xi|_B + |x_0|_B + \left|\left(E(\tau)-\mathbb{I}_N\right) x_0\right|_B< K\rho + R\left(\frac{1}{2}+\delta_1\right) + R\frac{c(n,B)}{\bar{\sigma}}\left(\frac{\bar{\sigma}}{\sigma_0}\right)^{\frac{2n+1}{3}} \left(\bar{\sigma}^2 b_B\frac{\rho^2}{R^2}\right)^{\frac{1}{2n+1}}.$$
Finally, if we have
\begin{equation}\label{etre}
\rho\leq R \frac{\delta_2-\delta_1}{2K}\qquad\mbox{ and }
\end{equation}
\begin{equation}\label{equattro}
\rho\leq R\frac{1}{\bar{\sigma} \sqrt{b_B}}\left(\frac{\delta_2-\delta_1}{2}\frac{\sigma_0^{\frac{2n+1}{3}}}{c(n,B)\bar{\sigma}^{\frac{2n-2}{3}}}\right)^{n+\frac{1}{2}},
\end{equation}
then
$$|\xi + E(\tau)x_0|_B< R(\delta_2-\delta_1) + R\left(\frac{1}{2}+\delta_1\right)=R\left(\frac{1}{2}+\delta_2\right)$$
as desired. The conditions which $\rho$ has to satisfy are \eqref{euno}, \eqref{edue}, \eqref{etre}, and \eqref{equattro}. Since $\delta_2-\delta_1<1$, these conditions are satisfied if
$$\rho\leq C_1 R (\delta_2-\delta_1)^{n+\frac{1}{2}}$$
for a suitable $C_1$ depending on $b_B, \sigma_0, \bar{\sigma}, K, c(n,B)$.\\
Let us now prove $(ii)$. Proceeding verbatim as in the first part, we can prove the existence of a structural constant $\tilde{C}_2$ such that, for any $\rho\leq \tilde{C}_2 R \left(\delta_2-\delta_1\right)^{n+\frac{1}{2}}$ and any $z_0$ in the closure of $\mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\frac{\delta_1+\delta_2}{2}\right)}^{-\frac{b_B}{4} R^2\left(3+\frac{(\delta_1+\delta_2)^2}{4}\right),-\frac{b_B}{2} R^2}$, we have
$$z_0\circ\zeta \in \mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\delta_2\right)}^{-\frac{b_B}{4} R^2\left(3+\delta^2_2\right),-\frac{b_B}{2} R^2}\quad\mbox{ for every fixed }\zeta\in \mathcal{Q}_{\rho}^{-b_B\rho^2,0}.$$
In particular, for such $\rho$ and using $\delta_1,\delta_2\leq 1$, we also have \eqref{valbe}. If we assume, in addition, that $z_0\in\partial_p\mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\frac{\delta_1+\delta_2}{2}\right)}^{-\frac{b_B}{4} R^2\left(3+\frac{(\delta_1+\delta_2)^2}{4}\right),-\frac{b_B}{2} R^2}$, then either $t_0=-\frac{b_B}{4} R^2\left(3+\frac{(\delta_1+\delta_2)^2}{4}\right)$ or $|x_0|_B=R\frac{\sigma_0}{2}\left(1+\frac{\delta_1+\delta_2}{2}\right)$. If the first possibility occurs, we have
$$\tau + t_0<t_0<-\frac{b_B}{4} R^2\left(3+\delta_1^2\right) \quad\mbox{ which implies }\quad z_0\circ\zeta \notin \mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\delta_1\right)}^{-\frac{b_B}{4} R^2\left(3+\delta^2_1\right),-\frac{b_B}{2} R^2}.$$
On the other hand, if the second possibility occurs, then by \eqref{triangleB} and \eqref{valbe} we get
\begin{eqnarray*}
|\xi + E(\tau)x_0|_B&\geq& |x_0|_B - |\xi|_B - \left|\left(E(\tau)-\mathbb{I}_N\right) x_0\right|_B\\
&>& R\frac{\sigma_0}{2}\left(1+\frac{\delta_1+\delta_2}{2}\right) - \rho - R\frac{c(n,B)}{\bar{\sigma}}\left(\frac{\bar{\sigma}}{\sigma_0}\right)^{\frac{2n+1}{3}} \left(\bar{\sigma}^2 b_B\frac{\rho^2}{R^2}\right)^{\frac{1}{2n+1}}.
\end{eqnarray*}
If $\rho$ also satisfies
$$\rho\leq R\frac{\sigma_0}{2} \frac{\delta_2-\delta_1}{4}\quad\mbox{ and }\quad\rho\leq R\frac{1}{\bar{\sigma} \sqrt{b_B}}\left(\frac{\sigma_0}{2}\frac{\delta_2-\delta_1}{4}\frac{\sigma_0^{\frac{2n+1}{3}}}{c(n,B)\bar{\sigma}^{\frac{2n-2}{3}}}\right)^{n+\frac{1}{2}},$$
then
$$|\xi + E(\tau)x_0|_B>R\frac{\sigma_0}{2}\left(1+\delta_1\right) \quad\mbox{ which implies }\quad z_0\circ\zeta \notin \mathcal{Q}_{R\frac{\sigma_0}{2}\left(1+\delta_1\right)}^{-\frac{b_B}{4} R^2\left(3+\delta^2_1\right),-\frac{b_B}{2} R^2}.$$
Therefore, up to modifying the constant $\tilde{C}_2$ to a suitable structural constant $C_2$, we have the desired conclusion.
\end{proof}
\section*{Acknowledgments}
\noindent F.A. wishes to thank Prof. Brian Rider for providing financial support through his NSF grant DMS--1406107. G.T. has been partially supported by the Gruppo Nazionale per l'Analisi Matematica, la Probabilit\`a e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM). The authors would like to thank Prof. Luis Silvestre for suggesting this problem at the 2017 Chicago Summer School in Analysis.
\bibliographystyle{amsplain}
|
\section{Introduction}
\label{sec:introduction}
Generalized linear geostatistical models (GLGMs) are a class of mixed models where, conditional on a realisation of a Gaussian process $\mathcal{S} = \{S(x) : x \in A \subset \mathbb{R}^2\}$ in a study area $A$, the outcome of interest $Y_{i}$, for $i=1,\dots,n$, follows a classical generalized linear model (GLM) \citep{GLM1989}. Hence, the following properties hold.
\begin{itemize}
\item The $Y_{i}$, conditional on $\mathcal{S}$, are a set of mutually independent variables with mean
$$E[Y_{i}| S(x_{i})]=m_{i}\mu_{i}=m_{i}g^{-1}(\eta_{i})$$
and variance
$$\text{Var}[Y_{i}| S(x_{i})]=m_{i}V(\mu_{i}),$$
where: $m_{i}$ is an offset (e.g. number of trials for a Binomial response); $\eta_{i}$ is the linear predictor; $g(\cdot)$ is the link function; and $V(\cdot)$ the variance function.
\item $\eta_{i} = d(x_{i})^\top \beta + S(x_i)$ where $d(x_i)$ is a vector of explanatory variables associated with location $x_i$ and $\beta$ is a vector of regression coefficients.
\item The conditional distribution of $Y_{i}$ belongs to the exponential family.
\end{itemize}
In this technical note, we address the following question: how should we assess the contribution of the explanatory variables $d(x_{i})$ and of the random effects $S(x_{i})$ to our predictive inferences? \par
To answer this question, we propose a generalization of the coefficient of determination proposed by \citet{zhang2017} to GLGMs and show its application to a geostatistical data-set on river-blindness. For classical GLMs, Zhang's coefficient is defined as
\begin{equation}
\label{eq:r2_glm}
R^2_{GLM} = 1-\frac{\sum_{i=1}^n c_{V}(y_{i}, \hat{y}_{i}\{d(x_{i})\})}{\sum_{i=1}^n c_{V}(y_{i}, \hat{y}_0)},
\end{equation}
where: $\hat{y}_{i}\{d(x_{i})\}$ is the prediction for $Y_{i}$ based on $d(x_{i})$ by plugging-in the estimated regression coefficients via maximum likelihood; $\hat{y}_{0}$ is the prediction from a GLM with an intercept only; and
$$
c_{V}(a,b) = \left\{\int_{a}^{b} \sqrt{1+[V'(u)]^2} \: du \right\}^2, \quad a, b \in \mathbb{R}
$$
which measures the change in the variance function $V(\cdot)$ for a change in the mean from $a$ to $b$. When $Y_{i}$ is Gaussian and $g(\cdot)$ is the identity function, the numerator in \eqref{eq:r2_glm} reduces to the residual sum of squares, i.e. $c_{V}(y_{i}, \hat{y}_{i}\{d(x_{i})\}) = \sum_{i=1}^n (y_{i}-\hat{y}_{i}\{d(x_{i})\})^2$. \citet{zhang2017} also shows that \eqref{eq:r2_glm} does not overstate the proportion of explained variance by the explanatory variables compared to other generalizations of the coefficient of determination to GLMs that are based on the likelihood ratio \citep{maddala1983, cox1989, nagelkerke1991}. Furthermore, unlike the generalization by \citet{cameron1997} based on the Kullback-Leibler divergence, Zhang's coefficient of determination is also defined for quasi-models \citep{wedderburn1972} and, therefore, does not require the full specification of the likelihood function.
\section{A generalization of Zhang's coefficient of determination to GLGMs}
\label{sec:r2_glgm}
Our generalization of Zhang's coefficient of determination is based on the intuitive interpretation of random effects as accounting for the effect of unmeasured variables. \par
For simplicity, consider a GLM with two explanatory variables $D_{1}(x)$ and $D_{2}(x)$, hence
$$
\eta_{i} = \beta_{0} + \beta_{1}D_{1}(x_{i}) + \beta_{2}D_{2}(x_{i}), \quad \text{for }i=1,\dots,n.
$$
Note that the two explanatory variables appear in the above equation in upper-case letters because we have not yet conditioned on them. Under such model, conditioning only on one of the two covariates might induce residual spatial correlation in the outcome $Y_{i}$. Hence, if, for example, we condition on $D_{1}(x)=d_{1}(x)$, a natural model for the data would be a GLGM where $d_{1}(x)$ is used as an explanatory variable and $S(x)$ is used to account for the residual effect $\beta_{2}D_{2}(x)$. This argument can also be easily extended to any number of measured and unmeasured variables. \par
It follows that, conditionally on a realisation of $S^\top=(S(x_{1}),\ldots,S(x_{n}))$, a natural approach to quantify the total variation in $Y^\top=(Y_{1}, \ldots, Y_{n})$ is through
\begin{equation}
\label{eq:tot_variation}
\sum_{i=1}^n c_{V}(y_{i},\hat{y}\{d(x_{i}),S(x_{i})\}),
\end{equation}
where $\hat{y}\{d(x_{i}),S(x_{i})\}$ is the prediction for $Y_{i}$ based on the vector of explanatory variables $d(x_{i})$ and the realisation of $S(x_{i})$. Since $S$ is not observed, we can use its predictive distribution, defined as the distribution of $S$ conditional on $y^\top=(y_{1},\ldots,y_{n})$ and the covariates $d^\top=(d(x_{1}),\ldots,d(x_{n}))$ (henceforth $S|(y,d)$), to compute \eqref{eq:tot_variation}. More specifically, we average \eqref{eq:tot_variation} over the distribution of $S|(y,d)$, which leads to
\begin{eqnarray}
\label{eq:r2_glgm}
R^2_{GLGM} &=& 1-\frac{E_{S|(y,d)}[\sum_{i=1}^n c_{V}(y_{i},\hat{y}\{d(x_{i}),S(x_{i})\})]}{\sum_{i=1}^n c_{V}(y_{i}, \hat{y}_0)}
\end{eqnarray}
In the case of a linear geostatistical model, obtained by setting $m_i=1$, $g^{-1}(\eta_i)=\eta_i$ and $V(\mu_i)=\tau^2$ for all $i$, the expectation of \eqref{eq:tot_variation} reduces to
$$
(y-D\beta)^\top(y-D\beta)+\xi^\top [\xi-2(y-D\beta)]+\text{tr}(\Omega),
$$
where: $D$ is a matrix of covariates;
$
\xi = \Sigma(\Sigma+I\tau^2)^{-1}(y-D\beta),
$
with $\Sigma$ and $I$ denoting the covariance matrix of the marginal distribution of $S$ and the identity matrix, respectively; and, finally,
$
\Omega = \Sigma-\Sigma(\Sigma+I\tau^2)^{-1}\Sigma.
$
For non-Gaussian responses, the expectation of \eqref{eq:tot_variation} is generally not available in closed form. We then propose to use a Monte Carlo Markov chain (MCMC) algorithm to simulate from $S | (y, d)$ and approximate \eqref{eq:r2_glgm} with
\begin{equation}
\label{eq:r2_glgm_mc}
R^2_{GLGM} \approx 1-\frac{\frac{1}{B}\sum_{j=1}^B\sum_{i=1}^n c_{V}(y_{i},\hat{y}\{d(x_{i}),s_{(j)}(x_{i})\})}{\sum_{i=1}^n c_{V}(y_{i},\hat{y}_{0})}
\end{equation}
where $s_{(j)}(x_{i})$ is the $j$-th out of $B$ Monte Carlo samples for the $i$-th component of $S|(y,d)$.\par
We can also define the coefficient of partial determination for the vector of explanatory variables $d$ given $S$ as
\begin{equation}
\label{eq:partial_r2_glgm}
\tilde{R}^2_{GLGM} = 1-\frac{E_{S|(y,d)}\left[\sum_{i=1}^n c_{V}(y_{i},\hat{y}\{d(x_{i}),S(x_{i})\}) \right]}{E_{S|(y,1)}\left[\sum_{i=1}^n c_{V}(y_{i},\hat{y}\{1,S(x_{i})\}) \right]},
\end{equation}
where $\hat{y}_{i}\{1,S(x_{i})\}$ is the prediction for $Y_{i}$ based on $S(x_{i})$ but excluding the explanatory variables $d(x_{i})$ from the model. We interpret \eqref{eq:partial_r2_glgm} as the fraction of explained variation in the response $Y$ by the explanatory variables $d$ but unexplained by the spatial random effects $S$. \par
In the next example, we compute \eqref{eq:r2_glgm} and \eqref{eq:partial_r2_glgm} by plugging-in the maximum likelihood estimates of the regression coefficients. These are obtained using the Monte Carlo likelihood method \citep{christensen2004} implemented in the R package PrevMap \citep{giorgi2017}. We simulate from $S|(y,d)$ using a Laplace sampling technique described in detail in Section 2.1 of \citet{giorgi2017}.
\section{Example: River-blindness mapping in Liberia}
\label{sec:rb_liberia}
River-blindness is an infectious disease caused by the parasite \textit{Onchocerca volvulus} and is transmitted by a black fly of the genus \textit{Simulium}. We analyse data from 90 communities in Liberia, where people were tested by palpation for the presence of skin nodules caused by the disease; for an Africa-wide analysis of these data, see \citet{zoure2014}. \par
Let $x_{i}$ be the location of the $i$-th sampled community, where $y_{i}$ out of $n_{i}$ randomly selected individuals tested positive. Our model for the data is a GLGM, where the $Y_{i}$ conditionally on $S(x_{i})$ are mutually independent Binomial variables with number of trials $n_{i}$ and probability of having skin nodules $p(x_{i})$, such that
\begin{equation}
\label{eq:lin_pred}
\log\left\{\frac{p(x_{i})}{1-p(x_{i})}\right\} = \beta_0 + \beta_1 x_{i,1}+ \beta_2 x_{i,2} + S(x_{i}),
\end{equation}
where $x_{i,1}$ and $x_{i,2}$ are the abscissa and ordinate components of the geographical location $x_{i}$. The reason for using a linear trend in $x_{i}$ is shown in Figure \ref{fig:Liberia_rb}, where the map of the empirical nodule prevalence shows an increase in the values as we move further from the coast in the north-east direction. Finally, we model $S(x)$ as a zero-mean Gaussian process with isotropic exponential covariance function having variance $\sigma^2$ and scale parameter $\phi$. \par
The maximum likelihood estimates of the model parameters and their 95$\%$ confidence intervals are reported in Table \ref{tab:rb_liberia}. We observe that the use of the explanatory variables leads to a remarkable reduction in the values of the estimated $\sigma^2$ and $\phi$. The fitted GLGM explains about $59\%$ of the variation in nodule prevalence compared to $27\%$ from a classical GLM where $S(x)=0$ for all $x$. However, the small value of $1\%$ for the coefficient of partial determination, $\tilde{R}^2$, indicates that the point estimates from the GLGM with covariates, given by \eqref{eq:lin_pred}, are strongly similar to a model without covariates, where $\beta_{1}=\beta_{2}=0$. Nonetheless, Figure \ref{fig:std_errors_lib} shows that the standard errors for the estimated nodule prevalence (computed using Monte Carlo samples from $S|(y,d)$ while pugging-in the Monte Carlo maximum likelihood estimates) at the observed locations from the model with covariates are smaller almost everywhere than those from the model with only the intercept. More precisely, the largest relative reduction in the standard errors is of about $10\%$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{liberia_rb.pdf}
\caption{Map of the empirical nodule prevalence. The radius of each point is proportional the quintile class within which the associated prevalence falls, with larger radiuses corresponding to higher quintile clasees. \label{fig:Liberia_rb}}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.5]{std_errors_lib.pdf}
\caption{Standard errors for the estimated nodule prevalence from a Binomial geostatistical model without covariates against one with covariates as defined in Section \ref{sec:rb_liberia}. The solid line is the identity line. \label{fig:std_errors_lib}}
\end{center}
\end{figure}
\begin{table}[ht]
\centering
\caption{Monte Carlo maximum likelihood estimates with associated 95$\%$ confidence intervals for the regression coefficients of the model with and without covariates defined in Section \ref{sec:rb_liberia}. \label{tab:rb_liberia}}
\begin{tabular}{rrcrc}
\hline
& \multicolumn{2}{c}{Without covariates} & \multicolumn{2}{c}{With covariates} \\
Term & Estimate & 95$\%$ CI & Estimate & 95$\%$ CI \\
\hline
$\beta_{0}$ & -1.941 & (-3.312, -0.571) & -6.327 & (-9.126, -3.528) \\
$\beta_{1} \times 10^3$ & & & 2.761 & (0.223, 5.299) \\
$\beta_{2} \times 10^3$ & & & 4.784 & (2.208, 7.360) \\
$\sigma^2$ & 0.791 & (0.075, 8.295) & 0.145 & (0.055, 0.384) \\
$\phi$ & 395.050 & (32.608, 4786.143) & 68.526 & (20.438, 229.755) \\
\hline
\vspace{-0.2cm} \\
\multicolumn{5}{c}{$R^2_{GLM}= 27\%; R^2_{GLGM} = 59\%; \tilde{R}^2_{GLGM} = 1\%$} \\
\vspace{-0.2cm} \\
\hline
\end{tabular}
\end{table}
\section{Discussion}
\label{sec:discussion}
We have introduced a generalization of Zhang's coefficient of determination to quantify the proportion of explained variation in the outcome of interest by the covariates and/or the residual spatial random effects. This has a more intuitive interpretation than other measures of predictive performance, such as mean square errors, and also allows to quantify the individual contribution of each component of the linear predictor to spatial prediction. Although our focus was on geostatistical models, the developed methodology can be applied to any generalized linear mixed model. \par
Through an example on river-blindness mapping, we have quantified the impact of the adopted explanatory variables on the spatial estimates of prevalence. The proposed generalization of the coefficient of partial determination, $\tilde{R}^2_{GLGM}$, indicated that the impact of these on the point estimates of prevalence was negligible. We have also shown that the reduction in the standard errors, albeit small, was however more tangible than the change in the point estimates after adjusting for the north-east trend in disease prevalence. Hence, our recommendation is that $\tilde{R}^2_{GLGM}$ should not be used as a stand-alone tool but should be complemented with other measures that assess the impact on the accuracy of the spatial estimates. \par
Future research will aim to extend the methods of Section \ref{sec:r2_glgm} to point process models, including log-Gaussian Cox processes.
\section*{Acknowledgements}
Emanuele Giorgi holds an MRC fellowship in Biostatistics (MR/M015297/1).
\bibliographystyle{biometrika}
|
\section{Introduction}
This article is the first in a series of articles that are meant to introduce and study a new category that encodes the simple rational characters of a reductive algebraic group in characteristics above the Coxeter number. This category is ``combinatorial'' in the sense that it is defined in terms of the underlying root system without reference to the group itself.
The following should give the reader a first idea on the nature of this category. Let $R$ be a root system and denote by ${\mathcal A}$ the associated set of (affine) alcoves. This set carries a partial order (sometimes called the {\em generic Bruhat order} or the {\em Bruhat order at $-\infty$}) and we obtain a topology on ${\mathcal A}$ with the order ideals as open sets. Now fix a field $k$ that satisfies the GKM-property with respect to $R$ (i.e., the characteristic is not $2$ and not $3$ if $R$ contains a component of type $G_2$). Denote by $S$ the symmetric algebra over the $k$-vector space associated with the coweight lattice of $R$. The action of the finite Weyl group on the set of ${\mathbb Z} R$-orbits in ${\mathcal A}$ gives rise to a commutative $S$-algebra ${\mathcal Z}$, which is non-canonically isomorphic to what is called the {\em structure algebra} of $R$.
The category that we are proposing is a full subcategory of the category of sheaves\footnote{(in the most ordinary sense)} of ${\mathcal Z}$-modules on the topological space ${\mathcal A}$. We denote it by ${\mathbf S}$. Apart from some minor technical assumptions there are two main properties that we stipulate on objects in ${\mathbf S}$. The first is the following. Let $x$ be a ${\mathbb Z} R$-orbit in ${\mathcal A}$ and let ${\mathscr M}$ be a presheaf of ${\mathcal Z}$-modules on ${\mathcal A}$. There are two constructions, associated with $x$, that we can perform on ${\mathscr M}$. One is algebraic, the other topological. By the definition of the structure algebra, the ${\mathbb Z} R$-orbit $x$ in ${\mathcal A}$ gives rise to a ${\mathcal Z}$-module homomorphism ${\mathcal Z}\to{\mathcal Z}^x$, where ${\mathcal Z}^x$ is free as an $S$-module of (graded) rank $1$. We obtain the presheaf ${\mathscr M}^x$ of ${\mathcal Z}$-modules on ${\mathcal A}$ by composing ${\mathscr M}$ with the functor ${\mathcal Z}^x\otimes_{{\mathcal Z}}\cdot$. On the other hand, as $x$ is a subset of ${\mathcal A}$, we can consider the inclusion $i_x\colon x\to{\mathcal A}$ and we obtain a natural morphism ${\mathscr M}^x\to i_{x\ast}i_x^\ast{\mathscr M}^x$. We say that ${\mathscr M}$ {\em satisfies the support condition} if this is an isomorphism for all ${\mathbb Z} R$-orbits $x$ in ${\mathcal A}$.
The second condition that we want the objects in ${\mathbf S}$ to satisfy is that they should ``behave well under base change''. Again, our definition of a base change functor is based on an algebraic and a topological construction. Note that the structure algebra is an algebra over the symmetric algebra $S$ over the coroot lattice. If $T$ is a flat $S$-algebra, then we can extend scalars an obtain ${\mathcal Z}_T={\mathcal Z}\otimes_S T$. On the other hand, note that the generic Bruhat order is generated by relations between an alcove and its mirror image with respect to a reflection at an affine root hyperplane. We can coarsen this relation by considering only reflections at hyperplanes corresponding to coroots that are not invertible in $T$, together with translations by arbitrary positive
roots. This yields a finer topology on ${\mathcal A}$. We then define a base change functor $(\cdot)\boxtimes_S T$ that maps a presheaf of ${\mathcal Z}$-modules to a presheaf of ${\mathcal Z}_T$-modules on the finer topology. It is constructed in such a way that its images satisfy the support condition mentioned above.
Now a sheaf ``behaves well under base change'' if it is still a sheaf, and not only a presheaf, after all flat base changes $S\to T$.
Once the category ${\mathbf S}$ is defined, we introduce a wall crossing functor $\vartheta_s$ on ${\mathbf S}$ with respect to each wall reflection $s$. It is constructed on the level of presheaves of ${\mathcal Z}$-modules by a simple universal property, but in general it does not preserve the category of sheaves. However, we show that it preserves the category ${\mathbf S}$.
In the companion article \cite{FieLanModRep} we consider the exact structure on ${\mathbf S}$ that is inherited via its inclusion in the abelian category of sheaves of ${\mathcal Z}$-modules on ${\mathcal A}$. We then show that ${\mathbf S}$ contains enough projectives, and we prove that the ranks of the stalks of the indecomposable projective objects in ${\mathbf S}$ encode the characters of the simple rational representations of the reductive algebraic groups with root system $R$ if the characteristic of $k$ is larger than the Coxeter number. For this we construct a functor $\Psi$ into the Andersen--Jantzen--Soergel category ${\mathbf K}$ defined in \cite{AJS} and we show that the indecomposable projective objects are mapped to the indecomposable ``special'' objects in ${\mathbf K}$\footnote{We should point out right away that our functor $\Psi$ is not fully faithful, not even when restricted to the projective objects in ${\mathbf S}$. We will show in a forthcoming article that ${\mathbf S}$ is a filtered category and that the image of $\Psi\colon{\mathbf S}\to{\mathbf K}$ can be thought of as the ``associated graded category''.}. From this one obtains the irreducible characters of $G$ via known results.
The advantage of ${\mathbf S}$ over ${\mathbf K}$ is that the definition of the category ${\mathbf K}$ is very ad-hoc and technical as it is a collection of subgeneric, i.e. local data, without a corresponding global object. Moreover, the ``special'' objects are defined by applying wall crossing functors to certain base objects and they do not have an intrinsic categorical characterization such as projectivity. This makes working with ${\mathbf K}$ rather difficult, and simple looking statements often require sophisticated and technical arguments\footnote{An example is the self-duality of special objects studied in \cite{S}.}. We believe that our category is much easier to work with, as its objects are nothing but sheaves on a topological space, and the objects that are important for representation theoretic applications are defined intrinsically, i.e. without reference to wall crossing functors.
We have further hopes with respect to ${\mathbf S}$. Note that Lusztig's formula for the irreducible characters of $G$ is only valid if the characteristic of $k$ is large enough. Ever since Williamson provided a huge list of examples for characteristics in which the formula fails (called the {\em torsion primes}), one tries to understand the phenomenon of torsion primes in modular representation theory. As ${\mathbf S}$ encodes the characters for all characteristics above the Coxeter number, it might be a helpful tool for this. Moreover, we believe that it is possible to simplify the work of Andersen, Jantzen and Soergel by constructing a functor from the category of (deformed) $G_1T$-modules into the associated graded of ${\mathbf S}$ directly, without refering to ${\mathbf K}$. This functor then hopefully makes sense also for restricted critical level representations of the affine Kac-Moody algebra associated with $R$ (so here we should take $k={\mathbb C}$), which might lead to a calculation of the irreducible critical characters, i.e. a proof of the Feigin--Frenkel conjecture.
\subsection*{Acknowledgements}
This material is based upon work supported by the National Science Foundation under Grant No. 0932078 000 while the first author was in residence at the Mathematical Sciences Research Institute in Berkeley, California, during the Fall 2014 semester. The second author would like to thank the University of Edinburgh, that supported her research during part of this project.
Both authors were partially supported by the DFG grant SP1388.
\section{(Pre-)Sheaves on partially ordered sets}\label{sec-sheavesparord}
This section provides some basic results on the topology of partially ordered sets and their theory of (pre-)sheaves. Fix a partially ordered set $({\mathcal X},\preceq)$. For an element $A$ of ${\mathcal X}$ we will use the short hand notation $\{\preceq A\}=\{B\in{\mathcal X}\mid B\preceq A\}$. The notations $\{\succeq A\}$, $\{\prec A\}$, etc. have an analogous meaning.
\subsection{A topology on $({\mathcal X},\preceq)$}\label{subsec-TopPar}
The following clearly yields a topology on the set ${\mathcal X}$.
\begin{definition} \label{def-topspac}
A subset ${\mathcal J}$ of ${\mathcal X}$ is called {\em open}, if $A\in {\mathcal J}$ and $B\preceq A$ imply $B\in {\mathcal J}$, i.e. if
$
{\mathcal J}=\bigcup_{A\in{\mathcal J}}\{\preceq A\}.
$
\end{definition}
The following statements are easy to check.
\begin{remark}\label{rem-top}
\begin{enumerate}
\item A subset ${\mathcal I}$ of ${\mathcal X}$ is closed if and only if ${\mathcal I}=\bigcup_{A\in{\mathcal I}}\{\succeq A\}$.
\item Arbitrary unions and intersections of open sets are open. The same holds for closed sets.
\item A subset ${\mathcal K}$ of ${\mathcal X}$ is locally closed (i.e. an intersection of an open and a closed subset) if and only if $A,B\in{\mathcal K}$ and $A\preceq C\preceq B$ imply $C\in{\mathcal K}$.
\end{enumerate}
\end{remark}
For any subset $ {\mathcal T} $ of ${\mathcal X}$ we define
$$ {\mathcal T}_{\preceq}:=\bigcup_{A\in {\mathcal T} }\{\preceq A\}.
$$
This is the smallest open subset of ${\mathcal X}$ that contains $ {\mathcal T} $.
\subsection{Presheaves on partially ordered sets}\label{subsec-PosetSheaves}
Now suppose ${\mathbf A}$ is an abelian category that has arbitrary products. For a presheaf ${\mathscr M}$ on ${\mathcal X}$ with values in ${\mathbf A}$ and an open subset ${\mathcal J}$ of ${\mathcal A}$ we denote by ${\mathscr M}{({\mathcal J})}$ the object of sections of ${\mathscr M}$ over ${\mathcal J}$. For an inclusion ${\mathcal J}^\prime\subset{\mathcal J}$ we denote by $r_{{\mathcal J}}^{{\mathcal J}^\prime}\colon {\mathscr M}({\mathcal J})\to{\mathscr M}{({\mathcal J}^\prime)}$ the restriction morphism. Sometimes we write $m|_{{\mathcal J}^\prime}$ instead of $r_{\mathcal J}^{{\mathcal J}^\prime}(m)$. Recall that one calls a presheaf ${\mathscr M}$ {\em flabby} if for any inclusion ${\mathcal J}^\prime\subset{\mathcal J}$ of open sets the restriction homomorphism $r_{{\mathcal J}}^{{\mathcal J}^\prime}$ is surjective.
For a subset ${\mathcal T}$ of ${\mathcal X}$ we denote by $i_{\mathcal T}\colon {\mathcal T}\to{\mathcal X}$ the inclusion.
\begin{definition}\label{def-supppre} Let ${\mathscr M}$ be a presheaf on ${\mathcal X}$. \begin{enumerate}\item ${\mathscr M}$ is said to be {\em supported on ${\mathcal T}\subset{\mathcal X}$} if the natural morphism ${\mathscr M}\to i_{{\mathcal T}\ast}i_{\mathcal T}^\ast{\mathscr M}$ is an isomorphism. \item ${\mathscr M}$ is said to be {\em finitary} if ${\mathscr M}(\emptyset)=0$ and if there exists a finite subset ${\mathcal T}$ of ${\mathcal X}$ such that ${\mathscr M}$ is supported on ${\mathcal T}$. \end{enumerate}
\end{definition}
Here, $i_{{\mathcal T}\ast}$ and $i_{\mathcal T}^\ast$ are the push-forward and pull-back functors for presheaves. More explicitely, ${\mathscr M}$ is supported on ${\mathcal T}$ if for any open subset ${\mathcal J}$ of ${\mathcal X}$ the restriction homomorphism ${\mathscr M}({\mathcal J})\to{\mathscr M}(({\mathcal J}\cap{\mathcal T})_{\preceq})$ is an isomorphism.
The partially ordered sets that we will be interested in in this article will be the set of alcoves associated with a finite root system endowed with various partial orders depending on the choice of a base ring.
\section{Alcove Geometry}\label{sec-alcoves}
In the following we review the basic features of the alcove geometry associated with a root system. As a reference one might consult \cite{Hum}\footnote{In \cite{Hum}, what we refer to as the \emph{affine Weyl group associated with a root system}, would be rather associated with the dual root system. We hope the reader will not be confused by this choice of terminology.}. We then endow the set of alcoves with a topology. This depends on the choice of a {\em base ring} $T$. We then study the decomposition into connected components. The main result is that each connected component is an orbit under a subgroup of the affine Weyl group (this subgroup depends on $T$). We finally introduce {\em admissible} families of open subsets. We prove in a later section that the sheaves that we are interested in are determined by their restriction to an arbitrary admissible family.
\subsection{Roots, reflections and alcoves}\label{subsec-Alc}
Fix a finite irreducible root system $R$ in a real finite dimensional vector space $V$. For any $\alpha\in R$ denote by $\alpha^\vee\in V^\ast={\operatorname{Hom}}_{\mathbb R}(V,{\mathbb R})$ the corresponding coroot. Let $\langle\cdot,\cdot\rangle\colon V\times V^\ast\to {\mathbb R}$ be the natural pairing and fix a system $R^+\subset R$ of positive roots. Let
\begin{align*}
X&:=\{\lambda\in V\mid \langle \lambda,\alpha^\vee\rangle\in{\mathbb Z}\text{ for all $\alpha\in R$}\},\\
X^\vee&:=\{v\in V^\ast\mid \langle\alpha,v\rangle\in{\mathbb Z}\text{ for all $\alpha\in R$}\}
\end{align*}
be the weight and the coweight lattice, resp.
For $\alpha\in R^+$ and $n\in{\mathbb Z}$ define
$$
H_{\alpha,n}:=\{\mu\in V\mid \langle \mu,\alpha^\vee\rangle = n\},
$$
the {\em affine reflection hyperplane} associated with $\alpha$ and $n$, and
\begin{align*}
H_{\alpha,n}^+&:=\{\mu\in V\mid \langle \mu, \alpha^\vee\rangle>n\},\\
H_{\alpha,n}^-&:=\{\mu\in V\mid \langle \mu,\alpha^\vee\rangle<n\},
\end{align*}
the corresponding positive and the negative half-space, resp.
\begin{definition}
The connected components of $V\setminus\bigcup_{\alpha\in R^{+},n\in{\mathbb Z}}H_{\alpha,n}$ are called {\em alcoves}. We denote by ${\mathcal A}$ the set of alcoves.
\end{definition}
Let $\alpha\in R^+$ and $n\in{\mathbb Z}$. The {\em affine reflection} with fixed point hyperplane $H_{\alpha,n}$ is
$$
s_{\alpha,n}\colon V\to V,\quad \lambda\mapsto \lambda-(\langle \lambda,\alpha^\vee\rangle-n)\alpha.
$$
It maps $X$ into $X$ and preserves the set $\{H_{\beta,m}\}$ of affine hyperplanes, hence induces a bijection on the set ${\mathcal A}$ that we denote by $s_{\alpha,n}$ as well.
For $\gamma\in X$ we denote by $t_\gamma\colon V\to V$ the affine translation $\mu\mapsto \mu+\gamma$. Again this preserves the set of hyperplanes and induces a bijection $t_\gamma\colon{\mathcal A}\to{\mathcal A}$.
Easy calculations yield:
\begin{lemma} \label{lemma-easypeasy}
\begin{enumerate}
\item For $\alpha\in R^+$ and $m,n\in{\mathbb Z}$ we have $s_{\alpha,n}\circ s_{\alpha,m}=t_{(n-m)\alpha}$.
\item For $\alpha\in R^+$, $n\in{\mathbb Z}$ and $\lambda\in X$ we have $s_{\alpha,n}\circ t_\lambda= t_{s_{\alpha,0}(\lambda)}\circ s_{\alpha,n}$.
\end{enumerate}
\end{lemma}
Denote by ${\widehat\CW}$ the affine Weyl group, i.e. the group of affine transformations on $V$ generated by the set $\{s_{\alpha,n}\mid \alpha\in R^+, n\in{\mathbb Z}\}$. Lemma \ref{lemma-easypeasy} implies that $t_{\gamma}$ is contained in ${\widehat\CW}$ for $\gamma\in{\mathbb Z} R$. The affine Weyl group acts on the set ${\mathcal A}$, and ${\mathcal A}$ is a principal homogeneous set for this ${\widehat\CW}$-action (cf. \cite[Section 4.5]{Hum}).
\subsection{Base rings}\label{subsec-basering} The topology that we introduce on ${\mathcal A}$ will depend on the choice of an algebra. Let $k$ be a field. We denote by $X^\vee_k=X^\vee\otimes_{\mathbb Z} k$ the $k$-vector space associated with the lattice $X^\vee$. For $v\in X^\vee$ we often denote by $v$ its canonical image $v\otimes 1$ in $X^\vee_k$.
We always assume that the following holds:
\begin{GKM} The characteristic of $k$ is $\ne 2$ and for $\alpha,\beta\in R^+$, $\alpha\ne\beta$ we have $\alpha^\vee\not\in k\beta^\vee$ (as elements in $X^\vee_k$).
\end{GKM}
Note that this excludes characteristic $2$ in all cases, and characteristic $3$ in case $G_2$. Note that for the representation theoretic applications we need the stronger assumption that the characteristic of $k$ is larger than the Coxeter number of $R$. From now on we assume that $k$ satisfies the above.
Let $S=S(X^\vee_k)$ be the symmetric algebra of the $k$-vector space $X^\vee_k$. We consider $S$ as a ${\mathbb Z}$-graded algebra with degree 2 component $S_2=X^\vee_k\subset S$. Let $T$ be a unital, commutative $S$-algebra that is flat as an $S$-module. Sometimes we assume in addition that $T$ is a ${\mathbb Z}$-graded $S$-algebra. In this case, a $T$-module is assumed to be graded, and a homomorphism between graded modules is assumed to respect the grading, i.e. it is of degree $0$. For a ${\mathbb Z}$-graded object $M=\bigoplus_{n\in{\mathbb Z}} M_n$ and $l\in{\mathbb Z}$ we write $M[l]$ for the object that we obtain from $M$ by shifting the grading in such a way that $M[l]_n=M_{l+n}$ for all $n\in{\mathbb Z}$.
Let $T$ be a (not necessarily graded) commutative, unital, flat $S$-algebra. Again we often write $\alpha^\vee$ for the image of $\alpha^\vee$ in $T$.
Note that the flatness implies that the structure homomorphism $S\to T$, $f\mapsto f1_T$, is injective. In particular, for two positive roots $\alpha\ne\beta$, the images in $T$ of $\alpha^\vee$ and $\beta^\vee$ are $k$-linearly independent.
\begin{definition}
\begin{enumerate}
\item Denote by $I_T\subset R^+$ the set of all $\alpha$ such that the left multiplication with $\alpha^\vee$ on $T$ is {\em not} a bijection (i.e. $\alpha^\vee$ is not invertible in $T$).
\item Denote by ${\widehat\CW}_T$ the subgroup of ${\widehat\CW}$ that is generated by $t_\gamma$ with $\gamma\in{\mathbb Z} R$ and by $s_{\alpha,m}$ for all $\alpha\in I_T$ and $m\in{\mathbb Z}$.
\item Denote by $R_T^+\subset R^+$ the set of all $\alpha$ with the property that ${\widehat\CW}_T$ contains $s_{\alpha,m}$ for some $m\in{\mathbb Z}$.
\end{enumerate}
\end{definition}
Note that ${\widehat\CW}_T$ contains ${\mathbb Z} R$, hence Lemma \ref{lemma-easypeasy} implies that $s_{\alpha,m}\in{\widehat\CW}_T$ for some $m\in{\mathbb Z}$ if and only if this condition holds for all $m\in{\mathbb Z}$.
By definition $I_T\subset R_T^+$. But in general this is not a bijection. For example, if $T=S[\alpha^{\vee-1}\mid\text{ $\alpha\in R^+$ is not simple}]$, then $I_T\subset R^+$ is the set of simple roots. It follows that ${\widehat\CW}_T={\widehat\CW}$, so $R_T^+=R^+$
.
\begin{definition} Call $T$ {\em saturated} if $I_T=R_T^+$.
\end{definition}
\begin{remark} \begin{enumerate}
\item For $T=S$ we have $I_S=R^+$, ${\widehat\CW}_T={\widehat\CW}$ and $R_S^+=R^+$. Hence $S$ is saturated.
\item Suppose that $I_T=\emptyset$. Then ${\widehat\CW}_T={\mathbb Z} R$ and $R_T^+=\emptyset$. In particular, $T$ is saturated in this case.
\item Suppose that $I_T=\{\alpha\}$ for some $\alpha\in R^+$. Then ${\widehat\CW}_T$ is generated by $s_{\alpha,m}$ with $m\in{\mathbb Z}$ and $t_\gamma$ with $\gamma\in{\mathbb Z} R$. Then ${\widehat\CW}_T=\{{\operatorname{id}},s_{\alpha,0}\}\ltimes{\mathbb Z} R$ and $R_T^+=\{\alpha\}$. Again, $T$ is saturated.
\end{enumerate}
\end{remark}
\begin{definition}
A {\em base ring} is an $S$-algebra $T$ that satisfies the following:
\begin{enumerate}
\item $T$ is unital and commutative.
\item $T$ is flat as an $S$-module.
\item $T$ is saturated.
\end{enumerate}
A base ring $T$ is called {\em generic}, if $R_T^+=\emptyset$, and {\em subgeneric}, if $R_T^+=\{\alpha\}$ for some $\alpha\in R^+$.
\end{definition}
\subsection{Partial orders and topologies on ${\mathcal A}$}
Fix a base ring $T$.
We now endow the set ${\mathcal A}$ with a partial order that depends on $T$, or rather on the set $I_T=R_T^+$.
\begin{definition}
Denote by $\preceq_T$ the partial order on ${\mathcal A}$ that is generated by $A\preceq_T B$ if either $B=t_\gamma(A)$ with $\gamma\in{\mathbb Z}_{\ge0} R^+$, or $B=s_{\alpha,n}(A)$ with $\alpha\in R_T^+$ and $n\in{\mathbb Z}$ such that $A\subset H_{\alpha,n}^-$.
\end{definition}
\begin{remark}\label{rem-ordexp} If $A\preceq_T B$, then there exists a finite sequence $A=A_0$, $A_1$,\dots,$A_n=B$ such that for all $i=1,\dots,n$ we either have $A_{i+1}=t_\gamma(A_i)$ for some $\gamma\in{\mathbb Z}_{\ge 0}R^+$, or $A_{i+1}=s_{\alpha,n}(A_i)$ for some $\alpha\in R_T^+$, $n\in{\mathbb Z}$ with $A_i\subset H_{\alpha,n}^-$.
\end{remark}
We denote by ${\mathcal A}_T$ the topological space associated with $({\mathcal A},\preceq_T)$. Sometimes we call a subset ${\mathcal J}$ of ${\mathcal A}$ {\em $T$-open} if it is open in ${\mathcal A}_T$.
\begin{remark}\label{rem-ordorb}
Let $T\to T^\prime$ be a homomorphism of base rings. If $\alpha^\vee$ is invertible in $T$, then it is also invertible in $T^\prime$. Hence $I^+_{T^\prime}\subseteq I^+_T$ and ${\widehat\CW}_{T^\prime}\subseteq{\widehat\CW}_T$, so $R_{T^\prime}^+\subseteq R_T^+$. Then $A\preceq_{T^\prime} B$ implies $A\preceq_T B$, hence a $T$-open subset is also $T^\prime$-open. That means that the identity $i_{T^\prime}^T\colon {\mathcal A}_{T^\prime}\to{\mathcal A}_T$ is continuous.
\end{remark}
\begin{lemma} \label{lemma-gentop} Let $A\in{\mathcal A}$ and $\gamma\in{\mathbb Z} R$. Then $A\preceq_T A+\gamma$ if and only if $\gamma\in{\mathbb Z}_{\ge 0}R^+$. In particular, the topology induced on a ${\mathbb Z} R$-orbit in ${\mathcal A}_T$ via the inclusion is independent of $T$. \end{lemma}
\begin{proof} For $A\in{\mathcal A}$ denote by $\lambda_A\in V$ the barycenter of $A$. We claim that $A\preceq_T B$ implies $\lambda_B-\lambda_A\in {\mathbb R}_{\ge 0} R^+$. In the case $B=t_\gamma(A)$ with $\gamma\in {\mathbb Z}_{\ge 0}R^+$ we have $\lambda_B-\lambda_A=\gamma\in{\mathbb R}_{\ge 0}R^+$. If $A=s_{\alpha,n}(B)$ with $A\subset H_{\alpha,n}^-$ we have $\lambda_B=s_{\alpha,n}\lambda_A$ and $s_{\alpha,n}(\lambda_A)-\lambda_A=-(\langle\lambda_A,\alpha^\vee\rangle-n)\alpha\in{\mathbb R}_{\ge 0}\alpha$. The general case follows from Remark \ref{rem-ordexp}.
Now $\lambda_{A+\gamma}-\lambda_A=\gamma$. Hence $A\preceq_T A+\gamma$ implies that $\gamma\in{\mathbb Z} R\cap{\mathbb R}_{\ge 0}R^+={\mathbb Z}_{\ge 0}R^+$. The converse is true by definition.
\end{proof}
In this article we denote ${\mathbb Z} R$-orbits by lower case Latin letters like $x,y,$ etc
\subsection{Connected components of ${\mathcal A}_T$} \label{subsec-conncomp}
In the following we denote connected components of ${\mathcal A}_T$ by upper case Greek letters like $\Lambda$, $\Omega$, etc.
\begin{lemma}\label{lemma-concomp} The connected components of ${\mathcal A}_T$ coincide with the ${\widehat\CW}_T$-orbits in ${\mathcal A}_T$.
\end{lemma}
\begin{proof} Let $A,B\in{\mathcal A}$. Then $A$ and $B$ are contained in the same connected component if and only if there is a sequence $A=A_0$, $A_1$, \dots, $A_n=B$ in ${\mathcal A}$ such that for all $i=1,\dots, n$ we have either $A_{i}\preceq_T A_{i+1}$ or $A_{i+1}\preceq_TA_i$. From Remark \ref{rem-ordexp} it follows that this is the case if and only if there is a sequence $A=A_0$, $A_1$,\dots, $A_n=B$ where $A_{i+1}=t_{\gamma}(A_i)$ for some $\gamma\in{\mathbb Z} R$, or $A_{i+1}=s_{\alpha,m}(A_i)$ for some $\alpha\in R_T^+$ and $m\in{\mathbb Z}$. This is the case if and only if $A$ and $B$ are in the same ${\widehat\CW}_T$-orbit.
\end{proof}
Denote by $A_e$ the unique alcove that is contained in the dominant Weyl chamber $\{\lambda\in V\mid \langle \lambda,\alpha^\vee\rangle>0\text{ for all $\alpha\in R^+$}\}$ and contains $0$ in its closure. Then the map
$$
{\widehat\CW}\to{\mathcal A},\quad w\mapsto A_w:=w(A_e)
$$
is a bijection as ${\mathcal A}$ is a principal homogeneous ${\widehat\CW}$-space (\cite[Section 4.5]{Hum}).
This allows us to construct the following right action of ${\widehat\CW}$ on ${\mathcal A}$. For $A=A_x\in{\mathcal A}$ and $w\in{\widehat\CW}$ let $Aw:=A_{xw}$. Clearly, this right action commutes with the left action.
This action is not continuous, but it preserves the connected components, as we show in part (1) of the following result.
\begin{lemma}\label{lemma-comp} \begin{enumerate}
\item Let $\Lambda\subset{\mathcal A}_T$ be a connected component. Then $\Lambda w$ is again a connected component.
\item Let $T\to T^\prime$ be a homomorphism of base rings. Then every connected component of ${\mathcal A}_T$ is a union of connected components of ${\mathcal A}_{T^\prime}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Claim (1) follows from Lemma \ref{lemma-concomp} and the fact that the right action of ${\widehat\CW}$ on ${\mathcal A}$ commutes with the left action. Claim (2) follows from Lemma \ref{lemma-concomp} and the fact that ${\widehat\CW}_{T^\prime}$ is a subgroup of ${\widehat\CW}_T$.
\end{proof}
\subsection{The finite Weyl group and ${\mathbb Z} R$-orbits}\label{subsec-finWeyl}
Denote by ${\mathcal V}={\mathcal A}/{\mathbb Z} R$ the set of ${\mathbb Z} R$-orbits in ${\mathcal A}$ (under the left action), and let $\pi\colon{\mathcal A}\to{\mathcal V}$ be the orbit map. We often denote by $\overline A$ or $\overline\Lambda$ the image of $A\in{\mathcal A}$ or $\Lambda\subset{\mathcal A}$ in ${\mathcal V}$. Note that every connected component $\Lambda$ of ${\mathcal A}_T$ is stable under the action of ${\mathbb Z} R$ (by Lemma \ref{lemma-concomp}). Hence $\Lambda=\pi^{-1}\pi(\Lambda)$.
Denote by ${\mathcal W}\subset{\widehat\CW}$ the finite Weyl group, i.e the subgroup generated by the ${\mathbb R}$-linear transformations $s_{\alpha,0}\colon V\to V$ for $\alpha\in R^+$. Lemma \ref{lemma-easypeasy} implies that the ${\widehat\CW}$-action on ${\mathcal A}$ induces a ${\widehat\CW}$-action on ${\mathcal V}$. It has the property that for all $\alpha\in R^+$, $m,n\in{\mathbb Z}$ and $x\in{\mathcal V}$ we have $s_{\alpha,n}(x)=s_{\alpha,m}(x)$. As ${\widehat\CW}={\mathcal W}\ltimes{\mathbb Z} R$ and as ${\mathcal A}$ is a principal homogeneous ${\widehat\CW}$-set, ${\mathcal V}$ is a principal homogeneous ${\mathcal W}$-set. For convenience, we abbreviate $s_\alpha:=s_{\alpha,0}$ for all $\alpha\in R^+$.
\begin{lemma} Suppose that $T$ is subgeneric with $R_T^+=\{\alpha\}$. Let $\Lambda$ be a connected component of ${\mathcal A}_T$. Then $\Lambda$ is the union of two distinct ${\mathbb Z} R$-orbits $x$ and $y$ that satisfy $y=s_\alpha x$.
\end{lemma}
\begin{proof} $\Lambda$ is an orbit of the subgroup ${\widehat\CW}_T$ of ${\widehat\CW}$, hence it is a principal homogeneous ${\widehat\CW}_T$-set. As ${\widehat\CW}_T$ is generated by $s_{\alpha,n}$ for $n\in{\mathbb Z}$ and ${\mathbb Z} R$ we have ${\widehat\CW}_T=\{{\operatorname{id}},s_{\alpha}\}\ltimes{\mathbb Z} R$ and the statement follows.
\end{proof}
\subsection{Admissible families}\label{subsec-admfam}
We will soon consider sheaves on the topological space ${\mathcal A}_T$. The condition that we assume on those sheaves assure that they are already determined once their restriction to ``$T$-admissible'' families of open subsets are known. Here is the definition of this notion.
Suppose that ${\mathbb T}$ is a family of subsets in ${\mathcal A}$.
\begin{definition}\label{def-Tadm} We say that ${\mathbb T}$ is a {\em $T$-admissible family}, if it satisfies the following assumptions.
\begin{enumerate}
\item ${\mathcal A}\in{\mathbb T}$.
\item Each element in ${\mathbb T}$ is $T$-open.
\item ${\mathbb T}$ is stable under taking finite intersections.
\item For any $T$-open subset ${\mathcal J}$ and any ${\mathbb Z} R$-orbit $x$ in ${\mathcal A}$ there is some ${\mathcal J}^\prime\in{\mathbb T}$ with ${\mathcal J}\cap x={\mathcal J}^\prime\cap x$.
\end{enumerate}
\end{definition}
Here is an example of an admissible family. Let $T\to T^\prime$ be a homomorphism of base rings.
\begin{lemma}\label{lemma-Top} The family ${\mathbb T}$ of $T$-open subsets in ${\mathcal A}$ is $T^\prime$-admissible.
\end{lemma}
\begin{proof} Clearly ${\mathcal A}$ is $T$-open, and any $T$-open subset is $T^\prime$-open, and the set of $T$-open subsets is stable under taking (arbitrary) intersections. Property (4) follows from Lemma \ref{lemma-gentop}. \end{proof}
In Lemma \ref{lemma-rigsinv} we introduce another admissible family.
\section{The structure algebra}\label{sec-strucalg}
Again we fix a root system $R$. In this section we define the {\em structure algebra} ${\mathcal Z}_S$ associated with $R$ over the field $k$. This is a commutative, associative, unital $S$-algebra that is ubiquitous in algebraic Lie theory. For example, it occurs as the torus equivariant cohomology of flag varieties, or as the center of deformed blocks of the category ${\mathcal O}$ of a semisimple Lie algebra (for $k={\mathbb C}$).
For any base ring $T$ we obtain the $T$-algebra ${\mathcal Z}_T={\mathcal Z}_S\otimes_ST$. We show that ${\mathcal Z}_T$ has a canonical decomposition with the direct summands parametrized by the connected components of the topological space ${\mathcal A}_T$. Then we study the category of ${\mathcal Z}_T$-modules. This category will be the target category for the (pre-)sheaves that we are interested in. We introduce the notion of ${\mathcal Z}$-support for ${\mathcal Z}_T$-modules. Later we will define a {\em support condition} for presheaves of ${\mathcal Z}_T$-modules on ${\mathcal A}_T$. The main idea of this condition is that the ${\mathcal Z}$-support and the sheaf theoretic support should be compatible.
\subsection{The structure algebra} Fix a base ring $T$.
Recall that we denote by ${\mathcal V}$ the principal homogeneous ${\mathcal W}$-set of ${\mathbb Z} R$-orbits in ${\mathcal A}$.
\begin{definition}\label{def-strucalg} Define
$$
{\mathcal Z}_S:=\left\{(z_x)\in\bigoplus_{x\in {\mathcal V}}S\left|\begin{matrix}\, z_x\equiv z_{s_{\alpha} x}\mod\alpha^\vee \\ \text{ for all $x\in{\mathcal V}$ and $\alpha\in R^+$ }\end{matrix}\right\}\right.
$$
and ${\mathcal Z}_T:={\mathcal Z}_S\otimes_ST$.
\end{definition} We call ${\mathcal Z}_T$ the {\em structure algebra} over our data.
Since $T$ is flat as an $S$-module we obtain a canonical inclusion ${\mathcal Z}_T\subset\bigoplus_{x\in{\mathcal V}} T$. So we can write every element $z\in{\mathcal Z}_T$ as a ${\mathcal V}$-tuple $z=(z_x)_{x\in{\mathcal V}}$ with $z_x\in T$. We simplify notation and set ${\mathcal Z}:={\mathcal Z}_T$ if $T$ can be determined from the context.
\subsection{The canonical decomposition of the structure algebra}
For a subset ${\mathcal L}$ of ${\mathcal V}$ we define
$$
{\mathcal Z}_S({\mathcal L}):=\left\{(z_x)\in\bigoplus_{x\in {\mathcal L}}S\left|\begin{matrix}\, z_x\equiv z_{s_{\alpha} x}\mod\alpha^\vee \\ \text{ for all $x\in{\mathcal L}$ and $\alpha\in R^+$ } \\ \text{ with $s_{\alpha} x\in{\mathcal L}$}\end{matrix}\right\}\right.
$$
and ${\mathcal Z}_T({\mathcal L})={\mathcal Z}_S({{\mathcal L}})\otimes_ST$.
Denote by $p^{\mathcal L}\colon\bigoplus_{x\in{\mathcal V}}S\to\bigoplus_{x\in{\mathcal L}}S$ the projection along the decomposition. Clearly $p^{{\mathcal L}}({\mathcal Z}_S)\subset{\mathcal Z}_S({\mathcal L})$, but in general, this is a proper inclusion. Denote by $C({\mathcal A}_T)$ the set of connected components of ${\mathcal A}_T$. Recall that we denote by $\overline\Lambda$ the image of $\Lambda$ in ${\mathcal V}$. As each connected component is stable under the action of ${\mathbb Z} R$, we obtain a decomposition ${\mathcal V}=\dot\bigcup_{\Lambda\in C({\mathcal A}_T)}\overline\Lambda$. In particular, the direct sum of the homomorphisms $p^{\overline\Lambda}$ yields an injective homomorphism $ p\colon {\mathcal Z}_S\to\bigoplus_{\Lambda\in C({\mathcal A}_T)}{\mathcal Z}_S(\overline\Lambda)$.
\begin{lemma} \label{lemma-decZcc}
The homomorphism $p\otimes_ST\colon{\mathcal Z}_T\to\bigoplus_{\Lambda\in C({\mathcal A}_T)}{\mathcal Z}_T(\overline\Lambda)$ is an isomorphism.
\end{lemma}
\begin{proof}
As $T$ is a flat $S$-algebra, $p\otimes_ST$ is injective. Now let $\rho$ be the product of all coroots that are invertible in $T$. Let $(z_{\overline\Lambda})\in\bigoplus_{\Lambda\in C({\mathcal A}_T)}{\mathcal Z}_S(\overline\Lambda)$. If $x\in\overline\Lambda$ and $\alpha\in R^+$ are such that $s_{\alpha}(x)\not\in\overline\Lambda$, then $\alpha\not\in R_T^+$, as $\Lambda$ is a ${\widehat\CW}_T$-orbit. Hence $\alpha^\vee$ must be invertible in $T$. It follows that $(\rho z_{\overline\Lambda})$ defines an element in ${\mathcal Z}_S$, so $(z_{\overline\Lambda})=(\rho z_{\overline\Lambda})\otimes\rho^{-1}$ is contained in ${\mathcal Z}_T={\mathcal Z}_S\otimes_ST$.
\end{proof}
\subsection{The connection to moment graphs}
Let $\Lambda$ be a connected component of ${\mathcal A}_T$. We denote by ${\mathcal G}_\Lambda$ the following moment graph over the lattice $X^\vee$. Its set of vertices is $\overline\Lambda$ (i.e. the set of ${\mathbb Z} R$-orbits in $\Lambda$), and $x,y\in\overline\Lambda$ are connected by an edge if and only if $y=s_{\alpha}x$ for some $\alpha\in R^+$. This edge is then labeled by $\alpha^\vee$. Note that $x,s_\alpha x\in\overline\Lambda$ imply that $\alpha\in R_T^+$. Then ${\mathcal Z}_S(\Lambda)$ is, by definition, the structure algebra over the field $k$ of the moment graph ${\mathcal G}_\Lambda$.
The category of ${\mathcal Z}_S(\Lambda)$-modules is intimately connected to the theory of sheaves on the moment graph ${\mathcal G}$ (cf. \cite{FieAdv}). In this article we do not refer to sheaves on moment graphs in order to avoid unnecessary confusion. In a forthcoming work we relate the sheaf category ${\mathbf S}$ that we are about to define to the category of moment graph sheaves not on ${\mathcal G}_\Lambda$, but on an affine version that is called the {\em stable moment graph} (cf. \cite{LanTAMS}).
\subsection{Quotients of ${\mathcal Z}$}
Let ${\mathcal L}\subset{\mathcal V}$ be an arbitrary subset and define ${\mathcal Z}_S^{\mathcal L}\subset{\mathcal Z}_S({\mathcal L})$ as the image of $p^{{\mathcal L}}$. Set ${\mathcal Z}^{{\mathcal L}}:={\mathcal Z}_T^{{\mathcal L}}={\mathcal Z}_S^{{\mathcal L}}\otimes_ST$. This is a sub-algebra of ${\mathcal Z}_T({\mathcal L})$ by the flatness of $T$ as an $S$-algebra, and it is a commutative, associative, unital $T$-algebra. We denote by $p^{\mathcal L}$ also the induced homomorphism ${\mathcal Z}\to{\mathcal Z}^{{\mathcal L}}$. If ${\mathcal L}$ containes a single element $x$, we write ${\mathcal Z}^x$ instead of ${\mathcal Z}^{\{x\}}$. Clearly ${\mathcal Z}^x=T$ as a $T$-module and the direct sum of the homomorphisms $p^x$ identifies with the inclusion ${\mathcal Z}\subset \bigoplus_{x\in{\mathcal V}}T$.
\begin{lemma}\label{lemma-subgensur}
Let $x\in{\mathcal V}$, $\alpha\in R^+$ and ${\mathcal L}=\{x,s_\alpha x\}$. Then
$$
{\mathcal Z}^{{\mathcal L}}=\{(z_x,z_{s_\alpha x})\in T\oplus T\mid z_x\equiv z_{s_\alpha x}\mod \alpha^\vee\}.
$$
\end{lemma}
\begin{proof}
Denote the right hand side of the alleged equation by $L$. Clearly, ${\mathcal Z}^{{\mathcal L}}$ is a subset of $L$. As the characteristic of $k$ is not $2$, $L$ is generated by $(1,1)$ and $(\alpha^\vee,-\alpha^\vee)$. Clearly $(1,1)$ is contained in ${\mathcal Z}^{\mathcal L}$. Consider the bijection ${\mathcal W}\xrightarrow{\sim}{\mathcal V}$, $w\mapsto w(x)$. The group ${\mathcal W}$ acts on $X^\vee$ in such a way that $s_{\beta}(h)=h-\langle\beta,h\rangle\beta^\vee$ for all $\beta\in R^+$ and $h\in X^\vee$. For $y\in{\mathcal V}$ define $z_{y}=w(\alpha^\vee)$ for $y=w(x)$ with $w\in{\mathcal W}$. Then $z_{s_\beta w(x)}-z_{w(x)}=-\langle\beta,w(v)\rangle\beta^\vee\equiv 0\mod\beta^\vee$ for all $w\in{\mathcal W}$ and $\beta\in R^+$, hence $z=(z_y)\in{\mathcal Z}_S$. We have $z_x=\alpha^\vee$ and $z_{s_\alpha x}=-\alpha^\vee$. Hence also the second generator of $L$ is contained in ${\mathcal Z}^{{\mathcal L}}$.
\end{proof}
\subsection{${\mathcal Z}$-modules} \label{subsec-Zmod}
Let $M$ be a ${\mathcal Z}$-module. For a subset ${\mathcal L}$ of ${\mathcal V}$ set $$M^{{\mathcal L}}:={\mathcal Z}^{{\mathcal L}}\otimes_{{\mathcal Z}} M.$$ We denote the canonical homomorphism $M\to M^{{\mathcal L}}$, $m\mapsto 1\otimes m$, by $p^{{\mathcal L}}$. For ${\mathcal L}^\prime\subset{\mathcal L}$ we have $(M^{{\mathcal L}})^{{\mathcal L}^\prime}=M^{{\mathcal L}^\prime}$. For us, a very important case is the case that ${\mathcal L}$ contains a unique element of ${\mathcal V}$. We then write $M^x$ instead of $M^{\{x\}}$. We call this the {\em stalk of $M$ at $x$}.
\begin{definition} \label{def-rtf} We say that $M$ is {\em root torsion free} if the left multiplication with $\alpha^\vee$ on $M$ is injective for all $\alpha\in R^+$.
\end{definition}
Denote by $T^{\emptyset}$ the localization of $T$ at the multiplicative set generated by $\{\alpha^{\vee}\}_{\alpha\in R^+}$. Then $T\to T^{\emptyset}$ is injective, and $I_{T^\emptyset}=\emptyset$. Hence $T^{\emptyset}$ is a generic base ring, so a connected component of ${\mathcal A}_{T^\emptyset}$ is a ${\mathbb Z} R$-orbit. In particular, Lemma \ref{lemma-decZcc} implies ${\mathcal Z}_{T^\emptyset}=\bigoplus_{x\in{\mathcal V}}T^{\emptyset}$.
Let $M$ be a root torsion free ${\mathcal Z}$-module. Then the canonical map $M\to M\otimes_TT^{\emptyset}$, $m\mapsto m\otimes 1$, is injective. As $M\otimes_TT^{\emptyset}$ is a ${\mathcal Z}_{T^\emptyset}$-module, we obtain a canonical decomposition $M\otimes_TT^{\emptyset}=\bigoplus_{x\in{\mathcal V}}(M\otimes_TT^{\emptyset})^x$ such that $(z_x)\in {\mathcal Z}_{T^\emptyset}$ acts on $(M\otimes_TT^{\emptyset})^x$ as multiplication with $z_x$.
More generally, the canonical decomposition ${\mathcal Z}=\bigoplus_{\Lambda\in C({\mathcal A}_T)} {\mathcal Z}^{\overline\Lambda}$ of Lemma \ref{lemma-decZcc} induces a direct sum decomposition $M=\bigoplus_{\Lambda\in C({\mathcal A}_T)}M^{\overline\Lambda}$. This we call the {\em canonical decomposition}.
\begin{lemma}\label{lemma-stalksZmod}
Suppose that $M$ is a root torsion free ${\mathcal Z}$-module. \begin{enumerate} \item For any ${\mathcal L}\subset{\mathcal V}$, the ${\mathcal Z}$-module $M^{{\mathcal L}}$ is root torsion free again. \item If $\{{\mathcal L}_i\}_{i\in I}$ is a finite set of subsets of ${\mathcal V}$ with ${\mathcal V}=\bigcup_{i\in I}{\mathcal L}_i$, then the direct sum of the homomorphisms $p^{{\mathcal L}_i}$ yields an injective homomorphism $M\to\bigoplus_{i\in I}M^{{\mathcal L}_i}$.\end{enumerate}
\end{lemma}
\begin{proof} (1) Let ${\mathcal L}$ be a subset of ${\mathcal V}$ and consider the composition $M\subset \bigoplus_{x\in{\mathcal V}}(M\otimes_TT^{\emptyset})^x\to\bigoplus_{x\in{\mathcal L}} (M\otimes_TT^{\emptyset})^x$, in which the map on the right is the projection along the decomposition. It is easy to check that this composition factors over the quotient $p^{\mathcal L}\colon M\to M^{{\mathcal L}}$ and induces an injective homomorphism $M^{{\mathcal L}}\subset \bigoplus_{x\in{\mathcal L}}(M\otimes_TT^{\emptyset})^x$. The module on the right is, as a $T^{\emptyset}$-module, root torsion free, hence so is $M^{{\mathcal L}}$.
(2) If $m$ is contained in the kernel of $p^{{\mathcal L}_i}$ for any $i$, then the arguments above show that $m$ must be in the kernel of the inclusion $M\subset\bigoplus_{x\in{\mathcal V}}(M\otimes_TT^{\emptyset})^x$, hence $m=0$.
\end{proof}
In particular we obtain a canonical injective homomorphim $M\to \bigoplus_{x\in{\mathcal V}}M^x$. Note that $M^x$ is the image of $M$ in $(M\otimes_TT^{\emptyset})^x$. This allows us to write each element $m$ of $M$ as a ${\mathcal V}$-tuple $m=(m_x)_{x\in{\mathcal V}}$ with $m_x\in M^x$.
\begin{definition}\begin{enumerate}\item We say that $M$ is {\em ${\mathcal Z}$-supported on ${\mathcal L}\subset{\mathcal V}$} and we write ${\operatorname{supp}}_{\mathcal Z} M\subset{\mathcal L}$ if $p^{\mathcal L}\colon M\to M^{{\mathcal L}}$ is an isomorphim.\item Let $m\in M$. We say that $m$ is {\em ${\mathcal Z}$-supported on ${\mathcal L}$} and we write ${\operatorname{supp}}_{\mathcal Z} m\subset{\mathcal L}$ if $m_x\ne 0$ implies $x\in{\mathcal L}$. \end{enumerate} \end{definition}
Clearly, $M$ is ${\mathcal Z}$-supported on ${\mathcal L}$ if and only if all its elements are ${\mathcal Z}$-supported on ${\mathcal L}$.
\begin{lemma}\label{lemma-Zsupp} Let $f\colon M\to N$ be a homomorphism of root torsion free ${\mathcal Z}$-modules and let ${\mathcal L}$ be a subset of ${\mathcal V}$. Then the following are equivalent.
\begin{enumerate}
\item $\ker f$ is ${\mathcal Z}$-supported on ${\mathcal L}\subset{\mathcal V}$.
\item For all $w\in{\mathcal V}\setminus{\mathcal L}$ the induced homomorphism $f^w\colon M^w\to N^w$ is injective.
\end{enumerate}
\end{lemma}
\begin{proof} Let us show that (1) implies (2). So let $w\in{\mathcal V}\setminus{\mathcal L}$ and let $m\in M$ be such that $m_w\ne 0$, but $f^w(m_w)=0$. Let $\delta_w=(\delta_{w,x})\in\bigoplus_{x\in{\mathcal V}}T$ be defined by $\delta_{w,w}=\prod_{\alpha\in R^+}\alpha^\vee$ and $\delta_{w,x}=0$ for all $x\ne 0$. Then $\delta_{w}\in{\mathcal Z}$. As $M^w$ is root torsion free, $(\delta_wm)_w=\prod_{\alpha\in R^+}\alpha^\vee m_w\ne 0$. But $\delta_wm$ is supported on $\{w\}$, so it must be in the kernel of $f$ as $f^w(\prod_{\alpha\in R^+}\alpha^\vee m_w)=0$, which contradicts the assumption in (1). Conversely, suppose that (2) holds. Let $m\in M$ be in the kernel of $f$. We have $f^x(m_x)=0$ for all $x\in{\mathcal V}$, hence $m_x\ne 0$ implies $x\in{\mathcal L}$. So $m$ is supported on ${\mathcal L}$. \end{proof}
\subsection{Root reflexive ${\mathcal Z}$-modules}
For $\alpha\in R^+$ define
$$T^{\alpha}:=T[\beta^{\vee-1}\mid\beta\in R^+,\beta\ne\alpha]
$$
and, as defined before,
$$
T^\emptyset:=T[\beta^{\vee-1}\mid \beta\in R^+].
$$
Then $R^+_{T^\emptyset}=\emptyset$ and $R^+_{T^\alpha}=\{\alpha\}$ or $R_{T^\alpha}^+=\emptyset$. So a connected component of ${\mathcal A}_{T^{\emptyset}}$ is a ${\mathbb Z} R$-orbit, and a connected component of ${\mathcal A}_{T^\alpha}$ is either a ${\mathbb Z} R$-orbit, or a union of two ${\mathbb Z} R$-orbits $x$ and $y$ with $y=s_\alpha x$. If $M$ is a root torsion free ${\mathcal Z}$-module, then we can view $M$ as a subset in $\bigcap_{\alpha\in R^+}(M\otimes_TT^{\alpha})\subset M\otimes_TT^{\emptyset}$. The following definition will become relevant later.
\begin{definition} We say that $M$ is {\em root reflexive} if it is root torsion free and $M=\bigcap_{\alpha\in R^+}M\otimes_TT^{\alpha}$.
\end{definition}
\section{Sheaves on the alcoves} \label{sec-sheavesA}
In this section we study presheaves on the topological space ${\mathcal A}_T$ with values in the category of ${\mathcal Z}_T$-modules. The sheaves that are relevant for us are finitary and flabby, and their local sections are root torsion free or even root reflexive. But we need two more conditions. The first connects the sheaf theoretic notion of support with the ${\mathcal Z}$-support defined earlier and is called the {\em support condition}. Once we defined this, we introduce a functor $(\cdot)^+$ that associates with a presheaf ${\mathscr M}$ its maximal quotient satisfying the support condition. We then prove a ``rigidity'' result: A presheaf satisfying the support condition is already determined by its restriction to an arbitrary admissible family of open subsets in ${\mathcal A}_T$. The category ${\operatorname{pSh}}_T$ contains presheaves of ${\mathcal Z}$-modules on ${\mathcal A}_T$ that satisfy the support condition (and some minor technical assumptions).
In order to state the second condition we construct a base change functor $(\cdot)\boxtimes_TT^{\prime}$ for any flat homomorphism $T\to T^\prime$ of base rings. It incorporates both the extension of scalars functor $(\cdot)\otimes_TT^{\prime}$ and the topological pull-back along $i\colon {\mathcal A}_{T^\prime}\to{\mathcal A}_T$, but also the functor $(\cdot)^+$. For us, the most relevant category is the category ${\mathbf S}_T$ of sheaves of ${\mathcal Z}_T$-modules on ${\mathcal A}_T$ that satisfy the support condition and have the property that they remain sheaves after base change along flat homomorphisms of base rings (again, we add some minor technical assumptions).
\subsection{The support condition} We fix a base ring $T$. We say that a presheaf ${\mathscr M}$ of ${\mathcal Z}$-modules on ${\mathcal A}_T$ is {\em root torsion free} if every ${\mathcal Z}$-module of local sections is root torsion free. Analogously we define \emph{root reflexive} presheaves.
Let ${\mathscr M}$ be a presheaf of ${\mathcal Z}$-modules on ${\mathcal A}_T$ and let ${\mathcal L}$ be a subset of ${\mathcal V}$. We define a new presheaf ${\mathscr M}^{\mathcal L}$ by composing with the functor $(\cdot)^{\mathcal L}$, i.e. for an open subset ${\mathcal J}$ of ${\mathcal A}_T$ we set
$${\mathscr M}^{\mathcal L}({\mathcal J}):={\mathscr M}({\mathcal J})^{\mathcal L}={\mathcal Z}^{{\mathcal L}}\otimes_{{\mathcal Z}}{\mathscr M}({\mathcal J})$$ with restriction homomorphism ${\operatorname{id}}\otimes r_{{\mathcal J}}^{{\mathcal J}^\prime}$. Again we write ${\mathscr M}^x$ instead of ${\mathscr M}^{\{x\}}$ in the case ${\mathcal L}=\{x\}\subset{\mathcal V}$.
Note that the natural homomorphisms ${\mathscr M}({\mathcal J})\to{\mathscr M}({\mathcal J})^{\mathcal L}$ combine and yield a morphism $p^{{\mathcal L}}\colon{\mathscr M}\to{\mathscr M}^{\mathcal L}$ of presheaves. If ${\mathscr M}$ is root torsion free, then Lemma \ref{lemma-stalksZmod} implies that the direct sum of the $p^x$ yields an injective morphism ${\mathscr M}\to\bigoplus_{x\in{\mathcal V}}{\mathscr M}^x$ of presheaves on ${\mathcal A}_T$.
Let $x$ be a ${\mathbb Z} R$-orbit in ${\mathcal A}$ and denote by $i_x\colon x\to{\mathcal A}_T$ the inclusion. Consider the presheaf $i_{x\ast}i_x^\ast{\mathscr M}$. Recall that for an open set ${\mathcal J}$ we denote by $({\mathcal J}\cap x)_{\preceq_T}$ the smallest $T$-open subset that contains ${\mathcal J}\cap x$. Then $$(i_{x\ast}i_x^\ast{\mathscr M})({\mathcal J})={\mathscr M}(({\mathcal J}\cap x)_{\preceq_T}),$$ and the restriction homomorphism $(i_{x\ast}i_x^\ast{\mathscr M})({\mathcal J})\to (i_{x\ast}i_x^\ast{\mathscr M})({\mathcal J}^\prime)$ is the restriction homomorphism ${\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})\to{\mathscr M}(({\mathcal J}^\prime\cap x)_{\preceq_T})$. The restriction homomorphisms ${\mathscr M}({\mathcal J})\to{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})$ combine and yield a morphism ${\mathscr M}\to i_{x\ast}i_x^\ast{\mathscr M}$ of presheaves. Clearly, the functors $(\cdot)^x$ and $i_{x\ast}i_x^\ast(\cdot)$ naturally commute.
The following definition is central for our approach.
\begin{definition}
We say that ${\mathscr M}$ {\em satisfies the support condition} if for any ${\mathbb Z} R$-orbit $x$ in ${\mathcal A}$ the natural morphism ${\mathscr M}^x\to i_{x\ast}i_x^\ast{\mathscr M}^x$ is an isomorphism.
\end{definition}
Explicitely, ${\mathscr M}$ satisfies the support condition if for any open subset ${\mathcal J}$ and any $x\in{\mathcal V}$ the natural homomorphism ${\mathscr M}({\mathcal J})^x\to{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x$ is an isomorphism.
Here is an equivalent notion of the support condition.
\begin{lemma} \label{lemma-suppcond} Let ${\mathscr M}$ be a flabby, root torsion free presheaf of ${\mathcal Z}$-modules on ${\mathcal A}_T$. Then the following are equivalent.\begin{enumerate}\item ${\mathscr M}$ satisfies the support condition.\item For any inclusion ${\mathcal J}^\prime\subset{\mathcal J}$ of open sets in ${\mathcal A}_T$ the kernel of the restriction homomorphism ${\mathscr M}({\mathcal J})\to{\mathscr M}{({\mathcal J}^\prime)}$ is ${\mathcal Z}$-supported on $\pi({\mathcal J}\setminus{\mathcal J}^\prime)$. \end{enumerate}\end{lemma}
\begin{proof} Suppose (1) and let ${\mathcal J}^\prime\subset{\mathcal J}$ be open. Suppose that $x\in{\mathcal V}\setminus\pi({\mathcal J}\setminus{\mathcal J}^\prime)$. We can identify the homomorphism ${\mathscr M}({\mathcal J})^x\to{\mathscr M}({\mathcal J}^\prime)^x$ with ${\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x\to {\mathscr M}(({\mathcal J}^\prime\cap x)_{\preceq_T})^x$. Now $x\not\in\pi({\mathcal J}\setminus{\mathcal J}^\prime)$ is equivalent to
${\mathcal J}\cap x={\mathcal J}^\prime\cap x$, hence the latter homomorphism is an isomorphism. So ${\mathscr M}({\mathcal J})^x\to{\mathscr M}({\mathcal J}^\prime)^x$ is an isomorphism for all $x\in{\mathcal V}\setminus\pi({\mathcal J}\setminus{\mathcal J}^\prime)$. By Lemma \ref{lemma-Zsupp} the kernel of ${\mathscr M}({\mathcal J})\to{\mathscr M}({\mathcal J}^\prime)$ is ${\mathcal Z}$-supported on $\pi({\mathcal J}\setminus{\mathcal J}^\prime)$. Hence (2) holds.
Conversely, suppose that (2) holds for ${\mathscr M}$. Let ${\mathcal J}$ be open and $x\in{\mathcal V}$. We need to show that ${\mathscr M}({\mathcal J})^x\to{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x$, i.e. the $x$-stalk of the restriction homomorphism, is an isomorphism. As ${\mathscr M}$ is flabby and the functor $(\cdot)^x$ is right exact, this homomorphism is surjective. Condition (2) implies, via Lemma \ref{lemma-Zsupp}, that it is injective, since $x\not\in\pi({\mathcal J}\setminus({\mathcal J}\cap x)_{\preceq_T})$. \end{proof}
\begin{definition}
Denote by ${\operatorname{pSh}}={\operatorname{pSh}}_T$ the full subcategory of the category of presheaves of ${\mathcal Z}$-modules on ${\mathcal A}_T$ that contains all objects ${\mathscr M}$ that are flabby, finitary, root torsion free and satisfy the support condition.
\end{definition}
\subsection{The canonical decomposition of objects in ${\operatorname{pSh}}$}
Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$. The canonical decomposition of ${\mathcal Z}$-modules studied in Section \ref{subsec-Zmod} implies that ${\mathscr M}$ splits into a direct sum ${\mathscr M}=\bigoplus_{\Lambda\in C({\mathcal A}_T)}{\mathscr M}^{\overline\Lambda}$, where each ${\mathscr M}^{\overline\Lambda}$ is a presheaf of ${\mathcal Z}^{\overline\Lambda}$-modules. In general, a presheaf on a topological space does not split into direct summands supported on connected components. The support condition in our situation, however, ensures that objects in ${\operatorname{pSh}}$ do.
\begin{lemma}\label{lemma-supp} Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$. For any $\Lambda\in C({\mathcal A}_T)$, ${\mathscr M}^{\overline\Lambda}$ is supported, as a presheaf, on $\Lambda\subset{\mathcal A}_T$.\end{lemma}
\begin{proof}We need to show that for any open subset ${\mathcal J}$ of ${\mathcal A}_T$, the restriction homomorphism ${\mathscr M}^{\overline\Lambda}({\mathcal J})\to{\mathscr M}^{\overline\Lambda}({\mathcal J}\cap\Lambda)$ is an isomorphism. Since $\pi({\mathcal J}\setminus({\mathcal J}\cap\Lambda))\subset{\mathcal V}\setminus{\overline\Lambda}$ Lemma \ref{lemma-suppcond} implies that the kernel of the above restriction is ${\mathcal Z}$-supported on ${\mathcal V}\setminus{\overline\Lambda}$. As ${\mathscr M}^{\overline\Lambda}({\mathcal J})$ is, by definition, ${\mathcal Z}$-supported on ${\overline\Lambda}$, the kernel of the restriction is trivial. The restriction is surjective as ${\mathscr M}$ is flabby. \end{proof}
\subsection{The maximal quotient satisfying the support condition}
Let ${\mathscr M}$ be a flabby and root torsion free presheaf of ${\mathcal Z}$-modules on ${\mathcal A}_T$. For any $x\in{\mathcal V}$ and any open subset ${\mathcal J}$ consider the composition ${\mathscr M}({\mathcal J})\to{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})\to{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x=i_{x\ast}i_x^\ast{\mathscr M}^x({\mathcal J})$, where the first homomorphism is the restriction and the second the natural homomorphism onto the $x$-stalk. These homomorphisms combine and yield a morphism $\gamma^x\colon{\mathscr M}\to i_{x\ast}i_x^\ast{\mathscr M}^x$. Denote by $\gamma\colon{\mathscr M}\to\bigoplus_{x\in{\mathcal V}}i_{x\ast}i_x^\ast{\mathscr M}^x$ the direct sum. Define ${\mathscr M}^+$ as the image (in the category of presheaves) of $\gamma$, i.e. ${\mathscr M}^+({\mathcal J})$ is the image of ${\mathscr M}({\mathcal J})$ in $\bigoplus_{x\in{\mathcal V}}{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x$. This is a flabby and root torsion free presheaf again (cf. Lemma \ref{lemma-stalksZmod}). We now show that it also satisfies the support condition.
\begin{lemma} \label{lemma-inj} Let ${\mathscr M}$ be a flabby, root torsion free presheaf.
\begin{enumerate}
\item The presheaf ${\mathscr M}^+$ satisfies the support condition.
\item ${\mathscr M}$ satisfies the support condition if and only if $\gamma$ induces an isomorphism ${\mathscr M}\cong{\mathscr M}^+$, i.e. if and only if $\gamma$ is injective (in the category of presheaves).
\end{enumerate}
\end{lemma}
\begin{proof} (1) Note that for any $x\in{\mathcal V}$ we obtain an injective morphism ${\mathscr M}^{+x}\to i_{x\ast}i_x^{\ast}{\mathscr M}^x$ (as ${\mathscr M}^+\subset\bigoplus_{y\in{\mathcal V}}i_{y\ast}i^\ast_y{\mathscr M}^y$ and as each local section of $i_{y\ast}i_y^{\ast}{\mathscr M}^y$ is ${\mathcal Z}$-supported on $y$). As ${\mathscr M}$ is flabby, this morphism is also surjective, hence an isomorphism. This implies ${\mathscr M}^{+x}\cong i_{x\ast}i_x^{\ast}{\mathscr M}^{+x}$. Hence ${\mathscr M}^+$ satisfies the support condition.
(2) If ${\mathscr M}\cong{\mathscr M}^+$, then ${\mathscr M}$ satisfies the support condition by (1). Conversely, suppose that ${\mathscr M}$ satisfies the support condition. The direct sum of the canonical homomorphisms onto the stalks ${\mathscr M}\to\bigoplus_{x\in{\mathcal V}}{\mathscr M}^x$ is injective. Hence the composition ${\mathscr M}\to\bigoplus_{x\in{\mathcal V}}{\mathscr M}^x\xrightarrow{\sim}\bigoplus_{x\in{\mathcal V}} i_{x\ast}i_x^\ast{\mathscr M}^x$ is injective. Hence ${\mathscr M}\cong{\mathscr M}^+$.
\end{proof}
Let us denote by $\gamma\colon{\mathscr M}\to{\mathscr M}^+$ the induced morphism as well. The following is a universal property for $\gamma$.
\begin{proposition} Let $f\colon {\mathscr M}\to{\mathscr N}$ be a morphism of flabby, root torsion free presheaves and suppose that ${\mathscr N}$ satisfies the support condition. Then there is a unique morphism $g\colon {\mathscr M}^{+}\to{\mathscr N}$ such that $f=g\circ \gamma$.
\end{proposition}
\begin{proof} Note that $f$ induces a morphism $\bigoplus_{x\in{\mathcal V}} i_{x\ast}i_x^\ast{\mathscr M}^x\to \bigoplus_{x\in{\mathcal V}} i_{x\ast}i_x^\ast{\mathscr N}^x$ such that the diagram
\centerline{
\xymatrix{
{\mathscr M}\ar[d]_{\gamma^{\mathscr M}}\ar[r]^f&{\mathscr N}\ar[d]^{\gamma^{\mathscr N}}\\
\bigoplus_{x\in{\mathcal V}} i_{x\ast}i_x^\ast{\mathscr M}^x\ar[r]& \bigoplus_{x\in{\mathcal V}} i_{x\ast}i_x^\ast{\mathscr N}^x
}
}
\noindent
commutes. By Lemma \ref{lemma-inj}, the morphism $\gamma^{\mathscr N}$ is injective. Hence the kernel of $\gamma^{\mathscr M}$ is contained in the kernel of $f$, so $f$ factors over the image of $\gamma^{\mathscr M}$. \end{proof}
\subsection{Rigidity}
Recall the definition of a $T$-admissible family in Definition \ref{def-Tadm}.
We now show that a morphism in ${\operatorname{pSh}}$ is already determined by its restriction to elements in a $T$-admissible family.
\begin{proposition}\label{prop-rig} Suppose that ${\mathbb T}$ is a $T$-admissible family. Let ${\mathscr M}$ and ${\mathscr N}$ be objects in ${\operatorname{pSh}}$. Suppose we are given a homomorphism $f^{({\mathcal J})}\colon {\mathscr M} {({\mathcal J})}\to {\mathscr N} {({\mathcal J})}$ of ${\mathcal Z}$-modules for each ${\mathcal J}\in{\mathbb T}$ in such a way that for any inclusion ${\mathcal J}^\prime\subset{\mathcal J}$ of sets in ${\mathbb T}$ the diagram
\centerline{
\xymatrix{
{\mathscr M} {({\mathcal J})}\ar[rr]^{f^{({\mathcal J})}}\ar[d]_{r_{\mathcal J}^{{\mathcal J}^\prime}}&&{\mathscr N} {({\mathcal J})}\ar[d]^{r_{{\mathcal J}}^{{\mathcal J}^\prime}}\\
{\mathscr M} {({\mathcal J}^\prime)}\ar[rr]^{f^{({\mathcal J}^\prime)}}&&{\mathscr N} {({\mathcal J}^\prime)}
}
}
\noindent
commutes. Then there is a unique morphism $f\colon {\mathscr M} \to {\mathscr N} $ of presheaves such that $f{({\mathcal J})}=f^{({\mathcal J})}$ for all ${\mathcal J}\in{\mathbb T}$.
\end{proposition}
\begin{proof} Note that the uniqueness follows immediately from the flabbiness of ${\mathscr M}$ together with the fact that ${\mathcal A}\in{\mathbb T}$. It remains to show the existence. In a first step we construct a morphism $f^x\colon{\mathscr M}^{ x}\to{\mathscr N}^{ x}$ for any $x\in{\mathcal V}$. For ${\mathcal J}$ open in ${\mathcal A}_T$ let ${\widetilde{\CJ}}\in{\mathbb T}$ be the smallest set such that ${\mathcal J}\cap x={\widetilde{\CJ}}\cap x$ (the definition of a $T$-admissible family ensures the existence of such a ${\mathcal J}^\prime$). Then ${\mathscr M}^{ x}({\mathcal J})={\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x={\mathscr M}(({\widetilde{\CJ}}\cap x)_{\preceq_T})^x={\mathscr M}^x({\widetilde{\CJ}})$ and, similarly, ${\mathscr N}^{ x}({\mathcal J})\cong{\mathscr N}^x({\widetilde{\CJ}})$. So there is a unique homomorphism $f^{ x}({\mathcal J})\colon {\mathscr M}^{ x}{({\mathcal J})}\to{\mathscr N}^{ x}({\mathcal J})$ induced by $(f^{({\widetilde{\CJ}})})^x$ using these identifications. Clearly this defines a morphism $f^{ x}\colon{\mathscr M}^{ x}\to{\mathscr N}^{ x}$ of presheaves on ${\mathcal A}_T$.
Again let ${\mathcal J}$ be an arbitrary open subset of ${\mathcal A}_T$. Consider the following commutative diagram:
\centerline{
\xymatrix{
{\mathscr M}({\mathcal A})\ar[d]^{f^{({\mathcal A})}}\ar[r]^{r_{\mathcal A}^{\mathcal J}}&{\mathscr M}({\mathcal J})\ar[r]&\bigoplus_{x\in{\mathcal V}}{\mathscr M}^{ x}({\mathcal J})\ar[d]^{f^{ x}({\mathcal J})}\\
{\mathscr N}({\mathcal A})\ar[r]^{r_{\mathcal A}^{\mathcal J}}&{\mathscr N}({\mathcal J})\ar[r]&\bigoplus_{x\in{\mathcal V}}{\mathscr N}^{ x}({\mathcal J}).\\
}
}
\noindent
Note that the two horizontal maps on the right are injective as ${\mathscr M}$ and ${\mathscr N}$ are root torsion free. As ${\mathscr M}$ is flabby, the upper left horizontal map is surjective. It follows that there is a unique homomorphism $f({\mathcal J})\colon{\mathscr M}({\mathcal J})\to{\mathscr N}({\mathcal J})$ completing the diagram. The uniqueness statement also ensures that all homomorphisms $f({\mathcal J})$ thus constructed yield a morphism $f\colon {\mathscr M}\to{\mathscr N}$ of presheaves. If ${\mathcal J}$ is in ${\mathbb T}$ and if we put $f^{({\mathcal J})}$ in the middle of the above diagram, then it commutes by the construction of the $f^{ x}$. Hence $f({\mathcal J})=f^{({\mathcal J})}$.
\end{proof}
\subsection{Base change}
Suppose that $T\to T^\prime$ is a flat homomorphism of base rings. In this section we want to construct a base change functor $\cdot\boxtimes_TT^\prime$ from ${\operatorname{pSh}}_T$ to ${\operatorname{pSh}}_{T^\prime}$. Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}_T$. As a first step define ${\mathscr M}\otimes_TT^\prime$ as the presheaf of ${\mathcal Z}_{T^\prime}$-modules on ${\mathcal A}_T$ with sections $({\mathscr M}\otimes_TT^\prime)({\mathcal J})={\mathscr M}({\mathcal J})\otimes_TT^\prime$ for any $T$-open subset ${\mathcal J}$ and with the obvious restriction homomorphisms.
Denote by $i\colon{\mathcal A}_{T^\prime}\to{\mathcal A}_T$ the (continuous) identity map. Again we denote by $i^{\ast}$ the presheaf-theoretic pullback functor. Note that it naturally commutes with base change functors. But in general the functor $i^\ast(\cdot)\otimes_TT^\prime$ maps an object in ${\operatorname{pSh}}_T$ only to a presheaf of ${\mathcal Z}_{T^\prime}$-modules on ${\mathcal A}_{T^\prime}$, not necessarily to an object in ${\operatorname{pSh}}_{T^\prime}$. However, every object in the image clearly is finitary, flabby and root torsion free. Hence the only missing property is the support condition. The following statement defines and characterizes the base change functor.
\begin{proposition}\label{prop-defbox}
There is an up to isomorphism unique pair ($\cdot\boxtimes_TT^\prime$, $\tau$), where $\cdot\boxtimes_TT^\prime\colon {\operatorname{pSh}}_T\to{\operatorname{pSh}}_{T^\prime}$ is a functor and $\tau\colon i^\ast(\cdot)\otimes_TT^\prime\to(\cdot)\boxtimes_TT^\prime$ is a natural transformation with the following property. For any $T^\prime$-open subset ${\mathcal J}$ that is also $T$-open $\tau$ induces an isomorphism $i^\ast{\mathscr M}\otimes_TT^\prime({\mathcal J})\xrightarrow{\sim}{\mathscr M}\boxtimes_TT^\prime({\mathcal J})$.
\end{proposition}
\begin{proof} As the set of $T$-open subsets is $T^\prime$-admissible by Lemma \ref{lemma-Top}, the uniqueness follows from Proposition \ref{prop-rig}. We need to show the existence. For an object ${\mathscr M}$ in ${\operatorname{pSh}}_T$ set
$$
{\mathscr M}\boxtimes_TT^\prime:=(i^\ast({\mathscr M}\otimes_TT^{\prime}))^{+}.
$$
More explicitely, if ${\mathcal J}$ is $T^\prime$-open, then $({\mathscr M}\boxtimes_TT^\prime)({\mathcal J})$ is the image of ${\mathscr M}({\mathcal J}_{\preceq_T})\otimes_TT^\prime$ in $\bigoplus_{x\in{\mathcal V}}{\mathscr M}(({\mathcal J}_{\preceq_T}\cap x)_{\preceq_T})^{x}\otimes_TT^\prime$. The natural transformation $\tau\colon i^\ast({\mathscr M})\otimes_TT^\prime=i^{\ast}({\mathscr M}\otimes_TT^\prime)\to {\mathscr M}\boxtimes_TT^\prime$ is induced by the natural transformation $\gamma\colon {\operatorname{id}}\to(\cdot)^+$ defined earlier. We already observed that ${\mathscr M}\boxtimes_TT^{\prime}$ is flabby, finitary and root torsion free. As it satisfies the support condition by construction, it is an object in ${\operatorname{pSh}}_{T^\prime}$. It remains to show that the pair $(\cdot\boxtimes_TT^{\prime},\tau)$ has the required property.
Suppose that ${\mathcal J}$ is $T$-open. Then $i^{\ast}{\mathscr M}({\mathcal J})={\mathscr M}({\mathcal J})$. As ${\mathscr M}$ satisfies the support condition, ${\mathscr M}({\mathcal J})\to\bigoplus_{x\in{\mathcal V}}{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x$ is injective. As $T\to T^\prime$ is flat, the induced homomorphism ${\mathscr M}({\mathcal J})\otimes_TT^\prime\to \bigoplus_{x\in{\mathcal V}}{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x\otimes_TT^\prime$ is injective. Hence ${\mathscr M}({\mathcal J})\otimes_TT^\prime$ is isomorphic to the image of this homomorphism, which, by definition, is $({\mathscr M}\boxtimes_TT^\prime)({\mathcal J})$. (Note that the extension of scalars functor $\otimes_TT^{\prime}$ commutes with $i^\ast$ and with $(\cdot)^+$.
\end{proof}
\begin{remarks}\label{rem-propbox} \begin{enumerate}
\item For flat homomorphisms $T\to T^\prime$ and $T^\prime\to T^{\prime\prime}$ it follows from the defining property that we have an isomorphism $({\mathscr M}\boxtimes_TT^\prime)\boxtimes_{T^\prime}T^{\prime\prime}\cong {\mathscr M}\boxtimes_TT^{\prime\prime}$.
\item Suppose that $T\to T^\prime$ is a flat homomorphism of base rings with the property that the identity ${\mathcal A}_{T^\prime}\to{\mathcal A}_T$ is a homeomorphism of topological spaces. Then ${\mathscr M}\boxtimes_TT^\prime\cong{\mathscr M}\otimes_TT^\prime$.
\end{enumerate}
\end{remarks}
If ${\mathscr M}$ is a sheaf on ${\mathcal A}_T$ and $T\to T^\prime$ is a flat homomorphism of base rings, then ${\mathscr M}\boxtimes_TT^\prime$ is not necessarily a sheaf on ${\mathcal A}_{T^\prime}$. This leads us to the following definition.
\begin{definition}\label{def-C} Denote by ${\mathbf S}={\mathbf S}_T$ the full subcategory of the category ${\operatorname{pSh}}_T$ that contains all objects ${\mathscr M}$ that satisfy the following properties: \begin{enumerate} \item ${\mathscr M}$ is a root reflexive sheaf on ${\mathcal A}_T$. \item For all flat homomorphisms $T\to T^\prime$ of base rings the presheaf ${\mathscr M}\boxtimes_TT^\prime$ is a root reflexive sheaf on ${\mathcal A}_{T^\prime}$. \end{enumerate}
\end{definition
\section{Reflections on alcove walls} As a preparation for the wall crossing functors, we return in this section to basic alcove geometry. We consider reflections at alcove walls (i.e. the right action of simple reflections in the affine Weyl group). The first result is about the partial order and it is well-known. We use it to deduce some topological statements. Then we discuss subsets that are invariant under alcove wall reflections on the set level and also invariants in the structure algebra. We finish this section with twisting functors for certain presheaves
that are associated with wall reflections.
\subsection{Reflections at alcove walls} Fix a base ring $T$.
Denote by ${\widehat\CS}$ the set of reflections in ${\widehat\CW}$ at hyperplanes that have a codimension $1$ intersection with the closure of the fundamental alcove $A_e$. The elements in ${\widehat\CS}$ are called the {\em simple affine reflections}. Fix $s\in{\widehat\CS}$. Recall the right action $A\mapsto As$ that we defined in Section \ref{subsec-conncomp}.
For any connected component $\Lambda$ of ${\mathcal A}_T$ the set $\Lambda s$ is a connected component again by Lemma \ref{lemma-comp}.
\begin{lemma}\label{lemma-tops} Let $\Lambda$ be a connected component of ${\mathcal A}_T$.
\begin{enumerate}
\item Suppose that $\Lambda=\Lambda s$ and let $A\in \Lambda$. Then $A,As$ are
$\preceq_{T}$-comparable and $\{A,As\}$ is a
$\preceq_{T}$-interval. Moreover, for all $B\in\Lambda$ the following holds.
\begin{enumerate}
\item If $As\preceq_{T} A$ and $B\preceq_{T}A$ then
$Bs\preceq_{T}A$,
\item If $As\preceq_{T} A$ and $As\preceq_{T}B$ then
$As\preceq_{T}Bs$.
\end{enumerate}
\item Suppose that $\Lambda\ne\Lambda s$. Then the map $\Lambda\to\Lambda s$, $A\mapsto As$, is a homeomorphism of topological spaces.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) We have $As=s_{\alpha,n}(A)$ for some $\alpha\in R^+$ and $n\in{\mathbb Z}$. As ${\mathcal A}$ is a principal homogeneous ${\widehat\CW}$-space and as $A,As$ are contained in the same ${\widehat\CW}_T$-orbit $\Lambda$, this implies $s_{\alpha,n}\in{\widehat\CW}_T$, i.e. $\alpha\in R^+_T=I_T$, hence $A$ and $As$ are $\preceq_T$-comparable. Now note that in the case $T=S$ we have $I_T=R_T^+=R^+$ and ${\widehat\CW}_T={\widehat\CW}$. The partial order $\preceq_S$ hence coincides with the partial order on ${\mathcal A}$ considered in \cite{LusAdv}. In this case, the statements (1), (1a) and (1b) are proven in \cite[Proposition 3.2, Corollary 3.3]{LusAdv}. Moreover, as $T$ is an $S$-algebra, $C\preceq_T D$ implies $C\preceq_S D$. As $\{A,As\}$ is a $\preceq_S$-interval, it must be a $\preceq_T$-interval as well.
(1a) We can assume $B\ne A$ and $B\preceq_T Bs$. Note that by the above arguments, the corresponding result in \cite{LusAdv} implies $Bs\preceq_S A$. Since $B\prec_T A$, there is a sequence
$$
B=B_0\preceq_T B_1\preceq_T\ldots \preceq_T B_{n-1}\preceq_T B_n=A,
$$
where either $B_i=s_{\alpha_i, m_i} B_{i-1}$ for $\alpha_i\in R^+_T$ and $m_i\in{\mathbb Z}$, or $B_i=B_{i-1}+\gamma$, for $\gamma\in{\mathbb Z}_{\geq 0}R^+$. We prove that $Bs\preceq_T A$ by induction on $n$. Suppose $n=1$. If $A=t_{\gamma}(B)$, then $Bs\preceq_T As$ as $t_\gamma$ preserves $\preceq_T$. Hence $Bs\preceq_T A$. If $A=s_{\alpha,n}(B)$ for some $\alpha\in R^+_T$, then $As=s_{\alpha,n}(Bs)$. Hence $As$ and $Bs$ are $\preceq_T$-comparable. If $As\preceq_T Bs$ then $As\preceq_S Bs$. As $Bs\preceq_S A$ and as $\{A,As\}$ is a $\preceq_S$-interval, this implies $As=Bs$, i.e. $A=B$ contrary to our assumption. Hence $Bs\preceq_T As$, so $Bs\preceq_T A$.
So suppose $n>1$. If $B_1s\preceq_T B_1$, then our arguments above prove $Bs\preceq_T B_1$, hence $Bs\preceq_T A$. If $B_1\preceq_T B_1s$ the induction hypothesis yields $Bs\preceq_T B_1 s$ and $B_1s\preceq_T A$, hence $Bs\preceq_T A$.
(1b) is proven with similar arguments.
(2) As $\Lambda$ and $\Lambda s$ play symmetric roles, it suffices to show that for $A,B\in\Lambda$ with $A\preceq_T B$ it follows that $As\preceq_T Bs$. So suppose $A\preceq_T B$. We can assume that $B=A+\gamma$ for some $\gamma\ge 0$ or $A=s_{\alpha,n}(B)$ for some $\alpha\in R^+_T$ with $A\subset H_{\alpha,n}^-$. In the first case we have $Bs=As+\gamma$, hence $As\preceq_T Bs$. In the second case, $As=s_{\alpha,n}(Bs)$, hence $As$ and $Bs$ are comparable. We need to show $As\subset H_{\alpha,n}^-$. If this were not the case, then the straight line between the barycenters of $A$ and $As$ would pass through $H_{\alpha,n}$. As $A$ and $As$ share a wall, this means that this wall would be contained in $H_{\alpha,n}$, i.e. $As=s_{\alpha,n}(A)$. So $A$ and $As$ would be in the same ${\widehat\CW}_T$-orbit, as $\alpha\in R_T^+$, hence in the same connected component which contradicts the assumption $\Lambda\ne\Lambda s$.
\end{proof}
\subsection{$s$-invariant subsets}
A subset ${\mathcal T}$ of ${\mathcal A}$ is called {\em $s$-invariant} if ${\mathcal T}={\mathcal T} s$. For any subset ${\mathcal T}$ set ${\mathcal T}^\sharp={\mathcal T}\cup{\mathcal T} s$ and ${\mathcal T}^\flat={\mathcal T}\cap{\mathcal T} s$. These are the smallest $s$-invariant subset of ${\mathcal A}$ containing ${\mathcal T}$ and the smallest $s$-invariant subset of ${\mathcal T}$, resp.
\begin{lemma}\label{lemma-Js} Suppose that ${\mathcal J}$ is open in ${\mathcal A}_T$. Then ${\mathcal J}^\sharp$ and ${\mathcal J}^\flat$ are open as well.
\end{lemma}
\begin{proof} We can assume that ${\mathcal J}$ is contained in a connected component $\Lambda$ of ${\mathcal A}_T$. If $\Lambda\ne\Lambda s$ then the statement follows immediately from Lemma \ref{lemma-tops}. Now assume $\Lambda=\Lambda s$. Let $A\in{\mathcal J}^\sharp$ and suppose that $B\preceq_T A$. If $A\in{\mathcal J}$, then $B\in{\mathcal J}$ as ${\mathcal J}$ is open, and hence $B\in{\mathcal J}^\sharp$. So suppose that $A\in{\mathcal J} s$. If $A\preceq_T As$, then $B\preceq_T As$ and we can argue as before and obtain $B\in{\mathcal J}^\sharp$. So assume $As\preceq_T A$. Then $Bs\preceq_T As$ by Lemma \ref{lemma-tops}, hence, by the above, $Bs\in{\mathcal J}^\sharp$, hence $B\in{\mathcal J}^\sharp$.
If $A\in{\mathcal J}^\flat$ and $B\preceq_T A$, then $B\in{\mathcal J}$. If $Bs\preceq_T B$, then $Bs\in{\mathcal J}$, hence $B\in{\mathcal J}^\flat$. If $B\preceq_T Bs$, then either $Bs\preceq_T A$ or $Bs\preceq_T As$ (or both), by Lemma \ref{lemma-tops}. In either case, $Bs\in{\mathcal J}$. Hence $B\in{\mathcal J}^\flat$.
\end{proof}
\begin{lemma} \label{lemma-sinvset} Let $\Lambda$ be a connected component of ${\mathcal A}_T$ and suppose that $\Lambda=\Lambda s$.
\begin{enumerate}
\item Let ${\mathcal J}$ be an open subset of $\Lambda$. Then $A\in{\mathcal J}^\sharp\setminus{\mathcal J}$ implies $As\preceq_T A$ and $A\in{\mathcal J}\setminus{\mathcal J}^\flat$ implies $A\preceq_T As$.
\end{enumerate}
Let ${\mathcal J}_i$ be a family of open subsets of $\Lambda$.
\begin{enumerate}\setcounter{enumi}{1}
\item We have $\bigcup_{i\in I}{\mathcal J}_i^\sharp=(\bigcup_{i\in I}{\mathcal J}_i)^\sharp$ and $\bigcap_{i\in I}{\mathcal J}_i^\sharp=(\bigcap_{i\in I}{\mathcal J}_i)^\sharp$.
\item We have $\bigcup_{i\in I}{\mathcal J}_i^\flat=(\bigcup_{i\in I}{\mathcal J}_i)^\flat$ and $\bigcap_{i\in I}{\mathcal J}_i^\flat=(\bigcap_{i\in I}{\mathcal J}_i)^\flat$.
\end{enumerate}
\end{lemma}
\begin{proof}
The assumption $\Lambda=\Lambda s$ implies that $A$ and $As$ are comparable for all $A\in\Lambda$ by Lemma \ref{lemma-tops}.
(1) Let $A\in{\mathcal J}^\sharp\setminus{\mathcal J}$. Then $As\in{\mathcal J}$, hence $A\preceq_T As$ would imply $A\in{\mathcal J}$, as ${\mathcal J}$ is open, which contradicts our assumption. If $A\in{\mathcal J}\setminus{\mathcal J}^\flat$, then $As\not\in{\mathcal J}$, so $As\preceq_T A$ would contradict the fact that ${\mathcal J}$ is open.
(2) The first identity $(\bigcup_{i\in I}{\mathcal J}_i)^\sharp=\bigcup_{i\in I}{\mathcal J}_i^\sharp$ is clear. It is also clear that $(\bigcap_{i\in I}{\mathcal J}_i)^\sharp\subset \bigcap_{i\in I}{\mathcal J}_i^\sharp$. So assume $A\in\bigcap_{i\in I}{\mathcal J}_i^\sharp$. This implies $As\in\bigcap_{i\in I} {\mathcal J}_i^\sharp$. As $(\bigcap_{i\in I}{\mathcal J}_i)^\sharp$ is $s$-invariant, it suffices to show that it contains either $A$ or $As$. So it suffices to show that it contains $A$ in the case $A\preceq_TAs$. But in this case $A\in{\mathcal J}_i^\sharp$ implies $A\in{\mathcal J}_i$ by (1), hence $A$ is even contained in $\bigcap_{i\in I}{\mathcal J}_i$.
(3) is proven with similar arguments. \end{proof}
\subsection{More admissible families}
Recall the definition of an admissible family in Section \ref{subsec-admfam}.
\begin{lemma} \label{lemma-rigsinv} The family of $s$-invariant $T$-open subsets in ${\mathcal A}$ is $T$-admissible. \end{lemma}
\begin{proof} Properties (1), (2) and (3) of an admissible family are clear. It remains to check property (4). Let ${\mathcal J}\subset{\mathcal A}_T$ be open and let $x$ be a ${\mathbb Z} R$-orbit in ${\mathcal A}$. We can replace ${\mathcal J}$ by the smallest open subset containing ${\mathcal J}\cap x$ and hence assume ${\mathcal J}=({\mathcal J}\cap x)_{\preceq_T}$. Then ${\mathcal J}$ and $x$ are contained in the same connected component of ${\mathcal A}_T$. We claim that $({\mathcal J}\cup{\mathcal J} s)\cap x={\mathcal J}\cap x$. As ${\mathcal J}\cup{\mathcal J} s$ is open by Lemma \ref{lemma-Js} and clearly $s$-invariant, this serves our purpose.
If $x$ and $xs$ are not contained in the same component of ${\mathcal A}_T$, then ${\mathcal J} s\cap x=\emptyset$ and the statement follows. If $x$ and $xs$ are contained in the same component, then either $A\preceq_T As$ for all $A\in x$ or $As\preceq_T A$ for all $A\in x$ as the ${\mathbb Z} R$-action preserves the partial order. In the second case, the set ${\mathcal J}$ is $s$-invariant since ${\mathcal J}=\bigcup_{A\in{\mathcal J}\cap x}\{\preceq_T A\}$ and since, by Lemma \ref{lemma-tops}, the set $\{\preceq_T A\}$ is $s$-invariant for all $A\in x$, so $({\mathcal J}\cup{\mathcal J} s)\cap x={\mathcal J}\cap x$. So assume that $A\preceq_T As$ for all $A\in x$. Let $B\in{\mathcal J} s\cap x$. Then $B\preceq_T Bs$ and $Bs\in{\mathcal J}$, hence $B\in{\mathcal J}$ as ${\mathcal J}$ is open. So ${\mathcal J} s\cap x\subset{\mathcal J}\cap x$ and hence $({\mathcal J}\cup{\mathcal J} s)\cap x={\mathcal J}\cap x$.
\end{proof}
\begin{lemma} \label{lemma-admtwice} Let $T\to T^\prime$ be a homomorphism of base rings. Then the family of $s$-invariant $T$-open subsets of ${\mathcal A}$ is $T^\prime$-admissible. \end{lemma}
\begin{proof} Again, the properties (1), (2) and (3) of an admissible family are easily checked. Let ${\mathcal J}$ be a $T^\prime$-open set and $x\in{\mathcal V}$. By Lemma \ref{lemma-Top} there is a $T$-open set ${\mathcal J}^\prime$ with ${\mathcal J}\cap x={\mathcal J}^\prime\cap x$. By Lemma \ref{lemma-rigsinv} there is a $T$-open $s$-invariant subset ${\mathcal J}^{\prime\prime}$ with ${\mathcal J}^{\prime\prime}={\mathcal J}^\prime\cap x$. This is what we wanted to show.
\end{proof}
\subsection{$s$-invariants in ${\mathcal Z}$} \label{subsec-sinvinZ} Recall that we have a right action of $s$ on the set ${\mathcal V}$ of ${\mathbb Z} R$-orbits in ${\mathcal A}$. Suppose ${\mathcal L}\subset{\mathcal V}$ is $s$-invariant (i.e. ${\mathcal L}={\mathcal L} s$).
Define $\eta_s\colon\bigoplus_{x\in{\mathcal L}}T\to\bigoplus_{x\in{\mathcal L}}T$ by $\eta_s(z_x)=(z^\prime_x)$, where $z^\prime_x=z_{xs}$ for all $x\in{\mathcal L}$.
\begin{lemma} $\eta_s$ preserves the subalgebra ${\mathcal Z}^{\mathcal L}\subset\bigoplus_{x\in{\mathcal L}}T$.
\end{lemma}
\begin{proof} This follows directly from the definition in the case $T=S$ and ${\mathcal L}={\mathcal V}$, as the right action of $s$ on ${\mathcal V}$ commutes with the left action of ${\mathcal W}$. It follows for arbitrary $T$ and ${\mathcal L}={\mathcal V}$. Clearly, the projection $\bigoplus_{x\in{\mathcal V}} T \to\bigoplus_{x\in{\mathcal L}} T$ intertwines the action of $\eta_s$ on both spaces for an arbitrary $s$-invariant ${\mathcal L}$. The claim follows. \end{proof}
\begin{definition}
We denote by ${\mathcal Z}^{{\mathcal L},s}$ the sub-$T$-algebra of $\eta_s$-fixed elements, and by ${\mathcal Z}^{{\mathcal L},-s}\subset{\mathcal Z}^{{\mathcal L}}$ the sub-${\mathcal Z}^{{\mathcal L},s}$-module of $\eta_s$-antiinvariants (i.e. the set of elements $z$ with $\eta_s(z)=-z$). We write ${\mathcal Z}^s$ and ${\mathcal Z}^{-s}$ in the case ${\mathcal L}={\mathcal V}$.
\end{definition}
As $2$ is invertible in $k$ we have ${\mathcal Z}^{{\mathcal L}}={\mathcal Z}^{{\mathcal L},s}\oplus{\mathcal Z}^{{\mathcal L},-s}$.
\begin{lemma}\label{lemma-lstrucfree} We have ${\mathcal Z}^{{\mathcal L},-s}\cong{\mathcal Z}^{{\mathcal L},s}[-2]$ as a ${\mathcal Z}^s({\mathcal L})$-module. In particular, ${\mathcal Z}^{{\mathcal L}}\cong{\mathcal Z}^{{\mathcal L},s}\oplus{\mathcal Z}^{{\mathcal L},s}[-2]$ as a ${\mathcal Z}^{{\mathcal L},s}$-module. \end{lemma}
Note that here the grading shift only makes sense, of course, if $T$ is a graded $S$-algebra. Otherwise, the statement holds without the shift. This remark applies to each occurrence of grading shifts in the following.
\begin{proof} As ${\mathcal Z}_S$ is the structure algebra associated with the root system $R$ the statement translates into \cite[Lemma 5.1 \& Proposition 5.3]{FieTAMS}.
\end{proof}
\begin{remark}\label{rem-trans} For a ${\mathcal Z}$-module $M$ that is ${\mathcal Z}$-supported on an $s$-invariant set ${\mathcal L}$ it follows that ${\mathcal Z}^{{\mathcal L}}\otimes_{{\mathcal Z}^{{\mathcal L},s}}M\cong M\oplus M[-2]$ as a ${\mathcal Z}^{{\mathcal L},s}$-module, and hence as a $T$-module. So the functor ${\mathcal Z}^{{\mathcal L}}\otimes_{{\mathcal Z}^{{\mathcal L},s}} (\cdot)$ is exact as an endofunctor on the category of ${\mathcal Z}^{{\mathcal L}}$-modules and preserves root torsion freeness and root reflexivity.
\end{remark}
\begin{lemma}\label{lemma-transsinv} Let $M$ be a root torsion free ${\mathcal Z}$-module that is ${\mathcal Z}$-supported in ${\mathcal L}$. Then there is a canonical isomorphism ${\mathcal Z}\otimes_{{\mathcal Z}^s} M={\mathcal Z}^{{\mathcal L}}\otimes_{{\mathcal Z}^{{\mathcal L},s}}M$. In particular, ${\mathcal Z}\otimes_{{\mathcal Z}^s} M$ is ${\mathcal Z}$-supported in ${\mathcal L}$ as well.
\end{lemma}
\begin{proof} Note that the ${\mathcal Z}$-action on $M$ factors over the homomorphism ${\mathcal Z}\to{\mathcal Z}^{{\mathcal L}}$, hence we obtain a homomorphism ${\mathcal Z}\otimes_{{\mathcal Z}^s}M\to {\mathcal Z}^{{\mathcal L}}\otimes_{{\mathcal Z}^{{\mathcal L},s}}M$ that clearly is surjective. Remark \ref{rem-trans} now implies that this is an isomorphism.
\end{proof}
\begin{lemma} \label{lemma-decZ} Suppose that $\Lambda\subset C({\mathcal A}_T)$ is a union of connected components with the property $\Lambda\cap \Lambda s=\emptyset$. The composition ${\mathcal Z}^{\overline{\Lambda\cup\Lambda s},s}\subset{\mathcal Z}^{\overline{\Lambda\cup\Lambda s}}={\mathcal Z}^{\overline\Lambda}\oplus{\mathcal Z}^{\overline{\Lambda s}}\xrightarrow{pr_\Lambda}{\mathcal Z}^{\overline\Lambda}$ is an isomorphism of $T$-algebras.
\end{lemma}
\begin{proof} It follows from the definition that the isomorphism $\gamma\colon\bigoplus_{x\in{\overline\Lambda}}T\to\bigoplus_{y\in\overline{\Lambda s}} T$ that maps $(z_x)$ into $(z^\prime_y)$ with $z^\prime_y=z_{ys}$ induces an isomorphism ${\mathcal Z}^{\overline\Lambda}\xrightarrow{\sim} {\mathcal Z}^{\overline{\Lambda s}}$ of $T$-algebras. As ${\mathcal Z}^{\overline{\Lambda\cup\Lambda s},s}\subset{\mathcal Z}^{\overline\Lambda}\oplus{\mathcal Z}^{\overline{\Lambda s}}$ is the subset of elements $(z,\gamma(z))$, the claim follows.
\end{proof}
\subsection{$s$-invariant elements in ${\mathcal Z}\otimes_{{\mathcal Z}^s}M$} \label{subsec-sinv}
Let $M$ be a ${\mathcal Z}$-module. Then the map $\eta_s^M\colon {\mathcal Z}\otimes_{{\mathcal Z}^s} M\to {\mathcal Z}\otimes_{{\mathcal Z}^s} M$, $z\otimes m\mapsto \eta_s(z)\otimes m$, is a ${\mathcal Z}^s$-linear automorphism of ${\mathcal Z}\otimes_{{\mathcal Z}^s} M$.
\begin{definition}
An element $m\in{\mathcal Z}\otimes_{{\mathcal Z}^s}M$ is called {\em $s$-invariant} if $\eta_s^M(m)=m$.
\end{definition}
\begin{lemma} \label{lemma-sinvel} Let ${\mathcal L}\subset{\mathcal V}$ and let $M$ be a root torsion free ${\mathcal Z}$-module. \begin{enumerate}
\item $\eta_s^M$ induces an involutive isomorphism between $({\mathcal Z}\otimes_{{\mathcal Z}^s} M)^{{\mathcal L}}$ and $({\mathcal Z}\otimes_{{\mathcal Z}^s} M)^{{\mathcal L} s}$.
\item If ${\mathcal L}\cap{\mathcal L} s=\emptyset$, then an element $(m_{\mathcal L},m_{{\mathcal L} s})\in ({\mathcal Z}\otimes_{{\mathcal Z}^s} M)^{{\mathcal L}\cup{\mathcal L} s}\subset ({\mathcal Z}\otimes_{{\mathcal Z}^s}M)^{{\mathcal L}}\oplus({\mathcal Z}\otimes_{{\mathcal Z}^s}M)^{{\mathcal L} s}$ is $s$-invariant if and only if $m_{{\mathcal L} s}=\eta_s^M(m_{\mathcal L})$.
\end{enumerate}
\end{lemma}
\begin{proof} As $({\mathcal Z}\otimes_{{\mathcal Z}^s}M)^{{\mathcal I}}={\mathcal Z}^{{\mathcal I}}\otimes_{{\mathcal Z}^s} M$ the statement (1) follows from the fact that $\eta_s$ induces an involutive isomorphism between ${\mathcal Z}^{{\mathcal L}}$ and ${\mathcal Z}^{{\mathcal L} s}$ (of ${\mathcal Z}^s$-algebras). Part (1) implies part (2). \end{proof}
The following result is a rather technical statement that is used later on.
\begin{lemma}\label{lemma-decm} Let ${\mathcal L}=\{x,xs\}\subset{\mathcal V}$ and let $M$ be a root torsion free ${\mathcal Z}$-module that is ${\mathcal Z}$-supported on ${\mathcal L}$. For each $m\in{\mathcal Z}\otimes_{{\mathcal Z}^s} M$ there is a unique $s$-invariant element $m^\prime\in{\mathcal Z}\otimes_{{\mathcal Z}^s} M$ such that $m_x=m^\prime_x$.
\end{lemma}
\begin{proof} As $M$ is ${\mathcal Z}$-supported on ${\mathcal L}$ we have ${\mathcal Z}\otimes_{{\mathcal Z}^s}M={\mathcal Z}^{\mathcal L}\otimes_{{\mathcal Z}^{{\mathcal L}, s}} M$ by Lemma \ref{lemma-transsinv}. Let $\alpha\in R^+$ be such that $xs=s_{\alpha} x$. By Lemma \ref{lemma-subgensur} we have $$
{\mathcal Z}^{{\mathcal L}}=\{(z_x,z_{xs})\in T\oplus T\mid z_x\equiv z_{xs}\mod \alpha^\vee\}
$$
and hence ${\mathcal Z}^{{\mathcal L},s}=T(1,1)$.
The elements $(1,1)$ and $(0,\alpha^\vee)$ form a $T$-basis of ${\mathcal Z}^{{\mathcal L}}$. So we can write $m=(1,1)\otimes a+(0,\alpha^\vee)\otimes b$ with $a,b\in M$. Clearly, $(1,1)\otimes a$ is $s$-invariant and $((0,\alpha^\vee)\otimes b)_{x}=0$. Hence $m_x=((1,1)\otimes a)_x$. This shows the existence.
In order to show uniqueness it suffices to show that if $n\in ({\mathcal Z}\otimes_{{\mathcal Z}^s} M)^{{\mathcal L}}$ is $s$-invariant and such that $n_x=0$, then $n=0$. This is a consequence of Lemma \ref{lemma-sinvel}.
\end{proof}
\subsection{Twisting (pre-)sheaves} Suppose that $\Lambda$ is a union of connected components of ${\mathcal A}_T$ that satisfies $\Lambda\cap\Lambda s=\emptyset$. The map $\gamma_s\colon \Lambda\to\Lambda s$, $A\mapsto As$ is then homeomorphism of topological spaces by Lemma \ref{lemma-tops}. Hence the pull-back functor $\gamma_s^{\ast}$ is an equivalence between the categories of (pre-)sheaves of ${\mathcal Z}$-modules on $\Lambda s$ and (pre-)sheaves of ${\mathcal Z}$-modules on $\Lambda$. For a ${\mathcal Z}$-module $M$ denote by $M^{s-tw}$ the ${\mathcal Z}$-module that we obtain from $M$ by twisting the ${\mathcal Z}$-action with the automorphism $\eta_s\colon{\mathcal Z}\to{\mathcal Z}$ (cf. Section \ref{subsec-sinvinZ}), i.e. $M^{s-tw}=M$ as an abelian group and $z\cdot_{M^{s-tw}} m=\eta_s(z)\cdot_M m$ for $z\in{\mathcal Z}$ and $m\in M$.
Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$ supported on $\Lambda s$ and denote by $\gamma_s^{[\ast]}{\mathscr M}$ the presheaf of ${\mathcal Z}$-modules on $\Lambda\subset{\mathcal A}_T$ that we obtain from the pull-back $\gamma_s^\ast{\mathscr M}$ by applying the functor $(\cdot)^{s-tw}$ to any local section. Hence, for any open subset ${\mathcal J}$ of $\Lambda$ we have
$$
(\gamma_s^{[\ast]}{\mathscr M})({\mathcal J})={\mathscr M}({\mathcal J} s)^{s-tw}.
$$
\begin{lemma}\label{lemma-gamma} Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$ supported on $\Lambda s$.
\begin{enumerate}
\item The presheaf $\gamma_s^{[\ast]}{\mathscr M}$ is an object in ${\operatorname{pSh}}$ supported on $\Lambda$.
\item If ${\mathscr M}$ is a sheaf, then $\gamma_s^{[\ast]}{\mathscr M}$ is a sheaf as well.
\end{enumerate}
Let $T\to T^\prime$ be a flat homomorphism of base rings.
\begin{enumerate}\setcounter{enumi}{2}
\item There is a functorial isomorphism $(\gamma_s^{[\ast]}{\mathscr M})\boxtimes_TT^\prime\cong \gamma_s^{[\ast]}({\mathscr M}\boxtimes_TT^\prime)$.
\item If ${\mathscr M}$ is an object in ${\mathbf S}$, then $\gamma_s^{[\ast]}{\mathscr M}$ is an object in ${\mathbf S}$ as well.
\end{enumerate}
\end{lemma}
\begin{proof} (1) Clearly $\gamma_s^{[\ast]}{\mathscr M}$ is finitary, root torsion free, flabby and supported on $\Lambda$. We need to show that it satisfies the support condition. Let $x\in{\overline\Lambda}$. Note that $(\gamma_s^{[\ast]}{\mathscr M})^x=(\gamma_s^\ast{\mathscr M})^{xs}=\gamma_s^{\ast}({\mathscr M}^{xs})$. As $i_{x}\circ\gamma_s=\gamma_s\circ i_{xs}$, we can identify the morphism $(\gamma_s^{[\ast]}{\mathscr M})^{x}({\mathcal J})\to i_{x\ast}i_x^\ast(\gamma_s^{[\ast]}{\mathscr M})^{x}({\mathcal J})$ with the morphism ${\mathscr M}^{xs}({\mathcal J} s)\to i_{xs\ast}i_{xs}^\ast{\mathscr M}^{xs}({\mathcal J} s)$ on the level of ${\mathcal Z}^s$-modules for any open subset ${\mathcal J}$ of $\Lambda$. As ${\mathscr M}$ satisfies the support condition, the latter is an isomorphism.
(2) This follows from the fact that $\gamma_s$, by Lemma \ref{lemma-tops}, is a homeomorphism of topological spaces.
(3) Note that $\Lambda$ is also union of connected components of ${\mathcal A}_{T^\prime}$ by Lemma \ref{lemma-comp}. The presheaves $(\gamma_s^{[\ast]}{\mathscr M})\boxtimes_TT^\prime$ and $\gamma_s^{[\ast]}({\mathscr M}\boxtimes_TT^\prime)$ are both supported on $\Lambda$. So let ${\mathcal J}\subset\Lambda$ be a $T$-open subset. Then ${\mathcal J} s\subset\Lambda s$ is $T$-open as well, and we have functorial identifications
\begin{align*}
((\gamma_s^{[\ast]}{\mathscr M})\boxtimes_TT^\prime)({\mathcal J})&=(\gamma_s^{[\ast]}{\mathscr M})({\mathcal J})\otimes_TT^\prime\\
&={\mathscr M}({\mathcal J} s)^{s-tw}\otimes_TT^\prime\\
&=({\mathscr M}\otimes_TT^\prime)({\mathcal J} s)^{s-tw}\\
&=({\mathscr M}\boxtimes_TT^\prime)({\mathcal J} s)^{s-tw}\\
&=\gamma_s^{[\ast]}({\mathscr M}\boxtimes_TT^\prime)({\mathcal J})
\end{align*}
that are compatible with the restriction homomorphisms.
As the set of $T$-open subsets in ${\mathcal A}$ is $T^\prime$-admissible by Lemma \ref{lemma-Top}, the statement follows from Proposition \ref{prop-rig}.
(4) Clearly $\gamma_s^{[\ast]}{\mathscr M}$ is root reflexive. Now the claim follows from this fact together with (2) and (3).
\end{proof}
\section{Wall crossing functors}\label{sec-wcf}
In this section we construct wall crossing functors on ${\operatorname{pSh}}_T$ and show that they preserve the subcategory ${\mathbf S}_T$ (they do not preserve the sheaf property in general). Wall crossing functors are of major importance in the companion article \cite{FieLanModRep}, where they are used to show that there are enough projective objects in ${\mathbf S}_T$ and to link ${\mathbf S}_T$ to the category defined by Andersen, Jantzen and Soergel.
\subsection{Characterization of the wall crossing functor}
Fix a base ring $T$ and $s\in{\widehat\CS}$.
Recall that for a $T$-open subset ${\mathcal J}$ the set ${\mathcal J}^\sharp={\mathcal J}\cup{\mathcal J} s$ is $T$-open again by Lemma \ref{lemma-Js}. Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$. Denote by $\epsilon_s{\mathscr M}$ the presheaf of ${\mathcal Z}$-modules on ${\mathcal A}_T$ given by
$$
(\epsilon_s{\mathscr M})({\mathcal J}):={\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J}^\sharp)
$$
for ${\mathcal J}$ open in ${\mathcal A}_T$, with restriction homomorphism ${\operatorname{id}}_{{\mathcal Z}}\otimes r^{{\mathcal J}^{\prime\sharp}}_{{\mathcal J}^\sharp}$ for an inclusion ${\mathcal J}^\prime\subset{\mathcal J}$.
Then $\epsilon_s{\mathscr M}$ is root torsion free by Remark \ref{rem-trans} and it is flabby as the functor ${\mathcal Z}\otimes_{{\mathcal Z}^s}\cdot$ is right exact on the category of ${\mathcal Z}$-modules. Clearly $\epsilon_s{\mathscr M}$ is finitary, but in general it does not satisfy the support condition. Define $$\vartheta_s{\mathscr M}:=(\epsilon_s{\mathscr M})^+.$$ This is now an object in ${\operatorname{pSh}}$ and we obtain a functor $\vartheta_s\colon{\operatorname{pSh}}\to{\operatorname{pSh}}$ which is called the {\em wall crossing functor} associated with $s$. The natural transformation ${\operatorname{id}}\to(\cdot)^+$ induces a natural transformation $\rho_s\colon\epsilon_s\to\vartheta_s$.
\begin{theorem}\label{thm-main} Let $s\in{\widehat\CS}$. The functor $\vartheta_s\colon{\operatorname{pSh}}\to{\operatorname{pSh}}$ is the up to isomorphism unique functor that admits a natural transformation $\rho_s\colon \epsilon_s\to\vartheta_s$ such that for every object ${\mathscr M}$ in ${\operatorname{pSh}}$ and any $s$-invariant open subset ${\mathcal J}$, the homomorphism $\rho_s^{\mathscr M}({\mathcal J})\colon \epsilon_s{\mathscr M}({\mathcal J})\to\vartheta_s{\mathscr M}({\mathcal J})$ is an isomorphism of ${\mathcal Z}$-modules.
\end{theorem}
\begin{proof} Note that the uniqueness statement follows from Proposition \ref{prop-rig} and the fact that the set of $s$-invariant open subsets in ${\mathcal A}_T$ is $T$-admissible by Lemma \ref{lemma-rigsinv}. It remains to show the following: Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$ and ${\mathcal J}$ an $s$-invariant open subset of ${\mathcal A}_T$. Then $\rho_s^{\mathscr M}({\mathcal J})\colon\epsilon_s{\mathscr M}({\mathcal J})\to\vartheta_s{\mathscr M}({\mathcal J})$ is an isomorphism. Using Lemma \ref{lemma-inj} this translates into the statement that the natural map
$$
\epsilon_s{\mathscr M}({\mathcal J})\to\bigoplus_{x\in{\mathcal V}}\epsilon_s{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x$$ (direct sum of restrictions and surjections onto stalks) is injective.
Let $x\in{\mathcal V}$. Then ${\mathcal J}\cap xs=({\mathcal J}\cap x)s$ and $(({\mathcal J}\cap x)_{\preceq_T})^\sharp=({\mathcal J}\cap x)_{\preceq_T}\cup({\mathcal J}\cap xs)_{\preceq_T}$. Let us denote this latter set by ${\widetilde{\CJ}}_x$. This is an $s$-invariant open subset of ${\mathcal J}$ and ${\widetilde{\CJ}}_x={\widetilde{\CJ}}_{xs}$. By definition,
$
\epsilon_s{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x=\left({\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\widetilde{\CJ}}_x)\right)^x.
$
Now note that $\pi({\mathcal J}\setminus{\widetilde{\CJ}}_x)\subset{\mathcal V}\setminus\{x,xs\}$. By Lemma \ref{lemma-suppcond}, the kernel of the restriction ${\mathscr M}({\mathcal J})\to{\mathscr M}({\widetilde{\CJ}}_x)$ is hence ${\mathcal Z}$-supported on ${\mathcal V}\setminus\{x,xs\}$. So the kernel of its composition ${\mathscr M}({\mathcal J})\to{\mathscr M}({\widetilde{\CJ}}_x)\to{\mathscr M}({\widetilde{\CJ}}_x)^{x,xs}$ is ${\mathcal Z}$-supported on ${\mathcal V}\setminus\{x,xs\}$ as well. Taking the direct sum of all these homomorphisms yields hence an {\em injective} homomorphism
$$
{\mathscr M}({\mathcal J})\to\bigoplus_{\{x,xs\}\in{\mathcal V}/s}{\mathscr M}({\widetilde{\CJ}}_x)^{x,xs}.
$$
As the functor ${\mathcal Z}\otimes_{{\mathcal Z}^s}(\cdot)$ is left exact on the category of ${\mathcal Z}$-modules, by Remark \ref{rem-trans}, also the homomorphism
$$
{\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J})\to\bigoplus_{\{x,xs\}\in{\mathcal V}/s}{\mathcal Z}\otimes_{{\mathcal Z}^s}({\mathscr M}({\widetilde{\CJ}}_x)^{x,xs})
$$
is injective. As $\{x,xs\}$ is an $s$-invariant subset of ${\mathcal V}$, Lemma \ref{lemma-transsinv} implies that ${\mathcal Z}\otimes_{{\mathcal Z}^s}({\mathscr M}({\widetilde{\CJ}}_x)^{x,xs})=({\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\widetilde{\CJ}}_x))^{x,xs}$. The canonical homomorphism $N^{x,xs}\to N^x\oplus N^{xs}$ is injective for any root torsion free ${\mathcal Z}$-module $N$, so the composition
$$
{\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J})\to \bigoplus_{\{x,xs\}\in{\mathcal V}/s}({\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\widetilde{\CJ}}_x))^{x,xs}\to \bigoplus_{x\in{\mathcal V}}({\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\widetilde{\CJ}}_x))^{x}
$$
is injective. But this is the homomorphism $\epsilon_s{\mathscr M}({\mathcal J})\to\bigoplus_{x\in{\mathcal V}}\epsilon_s{\mathscr M}(({\mathcal J}\cap x)_{\preceq_T})^x$.
\end{proof}
\subsection{Wall crossing and base change}
Suppose that $T\to T^\prime$ is a flat homomorphism of base rings. We have two wall crossing functors $\vartheta_{T,s}\colon {\operatorname{pSh}}_T\to{\operatorname{pSh}}_T$ and $\vartheta_{T^\prime,s}\colon{\operatorname{pSh}}_{T^\prime}\to{\operatorname{pSh}}_{T^\prime}$ and a base change functor $\cdot\boxtimes_TT^\prime\colon{\operatorname{pSh}}_T\to{\operatorname{pSh}}_{T^\prime}$.
\begin{lemma} \label{lemma-wcbc} The wall crossing functors commute naturally with base change, i.e. $\vartheta_{T,s}(\cdot)\boxtimes_TT^\prime\cong\vartheta_{T^\prime,s}(\cdot\boxtimes_TT^\prime)\colon{\operatorname{pSh}}_T\to{\operatorname{pSh}}_{T^\prime}$.
\end{lemma}
\begin{proof} Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}_T$ and let ${\mathcal J}$ be an $s$-invariant $T$-open subset. Then
\begin{align*}
((\vartheta_{T,s}{\mathscr M})\boxtimes_TT^\prime)({\mathcal J})&=(\vartheta_{T,s}{\mathscr M})({\mathcal J})\otimes_TT^\prime\\
&=({\mathcal Z}_T\otimes_{{\mathcal Z}_T^s}{\mathscr M}({\mathcal J}))\otimes_TT^\prime\\
&={\mathcal Z}_{T^\prime}\otimes_{{\mathcal Z}_{T^\prime}^s}({\mathscr M}({\mathcal J})\otimes_TT^\prime)\\
&={\mathcal Z}_{T^\prime}\otimes_{{\mathcal Z}_{T^\prime}^s}({\mathscr M}\boxtimes_TT^\prime)({\mathcal J})\\
&=\vartheta_{T^\prime,s}({\mathscr M}\boxtimes_TT^\prime)({\mathcal J}).
\end{align*}
The set of $s$-invariant open subsets in ${\mathcal A}_T$ is $T^{\prime}$-admissible by Lemma \ref{lemma-admtwice}. Hence the statement of the lemma follows from Proposition \ref{prop-rig}.
\end{proof}
\subsection{Wall crossing functors in the case $\Lambda\cap\Lambda s=\emptyset$}
Let $\Lambda$ be a union of connected components of ${\mathcal A}_T$ and suppose $\Lambda\cap \Lambda s=\emptyset$.
\begin{lemma}\label{lemma-decgen} Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$. Then
$$
(\vartheta_s{\mathscr M})^{\overline\Lambda}\cong{\mathscr M}^{\overline\Lambda}\oplus \gamma_s^{[\ast]}({\mathscr M}^{\overline{\Lambda s}}).
$$
\end{lemma}
\begin{proof} As $\vartheta_s{\mathscr M}$, ${\mathscr M}$ and $\gamma_s^{[\ast]}({\mathscr M}^{\overline{\Lambda s}})$ are objects in ${\operatorname{pSh}}$, it is sufficient, by Proposition \ref{prop-rig}, to establish an isomorphism
$$
(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J})\cong{\mathscr M}^{\overline\Lambda}({\mathcal J})\oplus\gamma_s^{[\ast]}({\mathscr M}^{\overline{\Lambda s}})({\mathcal J})
$$
for any $s$-invariant $T$-open subset ${\mathcal J}$ in a way tha is compatible restrictions.
So let ${\mathcal J}$ be $s$-invariant. We have functorial identifications
\begin{align*}
(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J})&=(\vartheta_s{\mathscr M})({\mathcal J})^{\overline\Lambda}=(\epsilon_s{\mathscr M})({\mathcal J})^{\overline\Lambda}\\%\quad\text{ by Theorem \ref{thm-main} }\\
&=({\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J}))^{\overline\Lambda}\\%\quad\text{ by Theorem \ref{thm-main} }\\
&={\mathcal Z}^{\overline\Lambda}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J})\\
&={\mathcal Z}^{\overline\Lambda}\otimes_{{\mathcal Z}^{\overline{\Lambda\cup\Lambda s},s}}({\mathscr M}^{\overline\Lambda}({\mathcal J})\oplus{\mathscr M}^{\overline{\Lambda s}}({\mathcal J}))\\
&={\mathscr M}^{\overline\Lambda}({\mathcal J})\oplus{\mathscr M}^{\overline{\Lambda s}}({\mathcal J})^{s-tw}\\
&={\mathscr M}^{\overline\Lambda}({\mathcal J})\oplus\gamma_s^{[\ast]}({\mathscr M}^{\overline{\Lambda s}})({\mathcal J}).
\end{align*}
In the sixth equation we used Lemma \ref{lemma-decZ} (note that $\eta_s$ induces an isomorphism ${\mathcal Z}^{\overline\Lambda}\xrightarrow{\sim}{\mathcal Z}^{\overline{\Lambda s}}$ of $T$-algebras, so ${\mathscr M}^{\overline{\Lambda s}}({\mathcal J})^{s-tw}$ is a ${\mathcal Z}^{\overline\Lambda}$-module).
\end{proof}
\begin{lemma}\label{lemma-subgensheaf}
If ${\mathscr M}$ is an object in ${\mathbf S}$ that is supported on a union $\Lambda$ of connected components that satisfies $\Lambda\cap\Lambda s=\emptyset$, then $\vartheta_s{\mathscr M}$ is an object in ${\mathbf S}$ again.
\end{lemma}
\begin{proof}
Note that ${\mathscr M}$ is an object in ${\mathbf S}$ if and only if ${\mathscr M}^{\overline\Lambda}$ and ${\mathscr M}^{\overline{\Lambda s}}$ are objects in ${\mathbf S}$. The statement then follows from Lemma \ref{lemma-decgen} and Lemma \ref{lemma-gamma}.
\end{proof}
\subsection{Wall crossing functors and $s$-invariant open sets}
It is not difficult to control the local sections on images of the wall crossing functors on $s$-invariant open sets.
\begin{lemma}\label{lemma-wcrf} Let ${\mathscr M}$ be an object in ${\operatorname{pSh}}$. \begin{enumerate}
\item Suppose ${\mathscr M}$ is root reflexive, and let ${\mathcal J}$ be an $s$-invariant open subset of ${\mathcal A}_T$. Then $\vartheta_s{\mathscr M}({\mathcal J})$ is root reflexive.
\item Suppose that ${\mathscr M}$ is a sheaf on ${\mathcal A}_T$. Then $\vartheta_s{\mathscr M}$ satisfies the sheaf property with respect to families of $s$-invariant open subsets in ${\mathcal A}_T$.
\end{enumerate}
\end{lemma}
\begin{proof} Recall that ${\mathcal Z}\cong{\mathcal Z}^s\oplus {\mathcal Z}^s[-2]$ as ${\mathcal Z}^s$-modules by Lemma \ref{lemma-lstrucfree}. Fix such an isomorphism.
For any $s$-invariant open subset ${\mathcal J}$ we then have an isomorphism $\vartheta_s{\mathscr M}({\mathcal J})={\mathscr M}({\mathcal J})\oplus {\mathscr M}({\mathcal J})[-2]$ (of ${\mathcal Z}^s$-modules). From this, part (1) follows. Note that the above isomorphism is compatible with restrictions for inclusions ${\mathcal J}^\prime\subset{\mathcal J}$ of $s$-invariant open subsets.
In order to prove claim (2) we need to show the following. Let ${\mathcal J}_i\subset{\mathcal A}_T$ for $i\in I$ be $s$-invariant open subsets, and $m_i\in(\vartheta_s{\mathscr M}){({\mathcal J}_i)}$ sections that satisfy $m_i|_{{\mathcal J}_i\cap{\mathcal J}_j}=m_j|_{{\mathcal J}_i\cap{\mathcal J}_j}$ for all pairs $i,j\in I$. Then there exists a unique section $m\in (\vartheta_s{\mathscr M})(\bigcup_i {\mathcal J}_i)$ that satisfies $m|_{{\mathcal J}_i}=m_i$ for all $i\in I$.
By the above we have an identification $\vartheta_s{\mathscr M}({\mathcal J}_i)={\mathscr M}({\mathcal J}_i) \oplus{\mathscr M}({\mathcal J}_i)[-2]$ of ${\mathcal Z}^s$-modules that is compatible with the restriction homomorphisms. Hence we can write $m_i=m_i^\prime+m_i^{\prime\prime}$ with $m_i^\prime\in{\mathscr M}{({\mathcal J}_i)}$ and $m_i^{\prime\prime}\in{\mathscr M}({\mathcal J}_i)[-2]$ for all $i\in I$. As ${\mathscr M}$ is a sheaf, there are unique sections $m^\prime\in{\mathscr M}({\bigcup_{i}{\mathcal J}_i})$ and $m^{\prime\prime}\in{\mathscr M}(\bigcup_i{\mathcal J}_i)[-2]$ restricting to $m_i^\prime$ and $m_i^{\prime\prime}$ on ${\mathcal J}_i$, resp., for all $i\in I$. Hence $m:=m^\prime+m^{\prime\prime}$ is the unique extension of the $m_i$.
\end{proof}
\subsection{$s$-invariant sections}
If ${\mathcal J}$ is an $s$-invariant open subset, then the natural transformation $\rho_s\colon\vartheta_s\to\epsilon_s$ yields a natural identification $\vartheta_s{\mathscr M}({\mathcal J})={\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J})$ which is compatible with the restriction homomorphisms associated to an inclusion of $s$-invariant open subsets. So we can consider the endomorphism $\eta_s=\eta_s^{{\mathcal Z}}\otimes{\operatorname{id}}$ on ${\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}^{{\mathcal J}}$ constructed in Section \ref{subsec-sinv}.
\begin{definition} For an $s$-invariant open subset ${\mathcal J}$ and an object in ${\mathscr M}$ in ${\operatorname{pSh}}$ we say that a section $m$ in $\vartheta_s{\mathscr M}({\mathcal J})$ is {\em $s$-invariant} if $\eta_s(m)=m$, i.e. if $m$ is contained in ${\mathcal Z}^s\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J})\subset{\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J})$.
\end{definition}
Let $\Lambda\subset {\mathcal A}_T$ be a union of connected components. To simplify the notation, we write
$$
{\mathcal J}_\Lambda:={\mathcal J}\cap\Lambda
$$
for an open subset ${\mathcal J}$ of ${\mathcal A}_T$.
Here is an easy characterization of $s$-invariant sections on ${\mathcal J}\subset \Lambda\cup\Lambda s$ in the case $\Lambda\cap\Lambda s=\emptyset$.
\begin{lemma}\label{lemma-sinvgamma} Let $\Lambda$ be a union of components of $C({\mathcal A}_T)$ with $\Lambda\cap\Lambda s=\emptyset$. For any open subset ${\mathcal J}$ of ${\mathcal A}_T$ there is an isomorphism
$$
\eta_s^{\overline\Lambda}\colon(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J}_\Lambda)\xrightarrow{\sim}(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J}_\Lambda s)
$$
of ${\mathcal Z}^s$-modules that is functorial in ${\mathscr M}$ and compatible with restriction homomorphisms. Moreover, if ${\mathcal J}$ is $s$-invariant, then ${\mathcal J}_{\Lambda s}={\mathcal J}_\Lambda s$ and an element $m=(m_{\overline\Lambda},m_{{\overline{\Lambda s}}})\in(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J}_\Lambda)\oplus(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J}_{\Lambda s})$ is $s$-invariant if and only if $m_{{\overline{\Lambda s}}}=\eta^{\overline\Lambda}_s(m_{\overline\Lambda})$.
\end{lemma}
\begin{proof} Suppose that ${\mathcal J}\subset\Lambda$. Then ${\mathcal J}^{\sharp}={\mathcal J}\dot\cup{\mathcal J} s$ and, by the support condition, $\vartheta_s{\mathscr M}^{\overline\Lambda}({\mathcal J})=\vartheta_s{\mathscr M}^{\overline\Lambda}({\mathcal J}^{\sharp})$. As $\vartheta_s{\mathscr M}({\mathcal J}^{\sharp})={\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J}^{\sharp})$ we can define $\eta_s^{\overline\Lambda}$ as in Lemma \ref{lemma-sinvel}. If ${\mathcal J}$ is $s$-invariant, then clearly ${\mathcal J}_{\Lambda s}={\mathcal J}_\Lambda s$ and the remaining claim follows from Lemma \ref{lemma-sinvel} as well. \end{proof}
The next result is the main technical result of this article and the most important ingredient in the proof that the wall crossing functors preserve the category ${\mathbf S}$.
\begin{proposition} \label{prop-sinvpreim} Let ${\mathscr M}$ be an object in ${\mathbf S}$, let ${\mathcal J}\subset{\mathcal A}_T$ be open and let $m\in\vartheta_s{\mathscr M}({\mathcal J})$ be a section such that $m^\flat:=m|_{{\mathcal J}^\flat}$ is $s$-invariant. Then there exists a unique $s$-invariant section
$m^\sharp\in\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$ with $m^\sharp|_{\mathcal J}=m$.
\end{proposition}
\begin{proof
Suppose we have proven the statement of the proposition in the cases that $T$ is either generic or subgeneric. For general $T$ we can then view $m$ as an element in $\vartheta_s{\mathscr M}({\mathcal J})\otimes_TT^{\ast}=\vartheta_s({\mathscr M}\boxtimes_TT^{\ast})({\mathcal J})$ for any $\ast\in R^+\cup\{\emptyset\}$ and obtain unique preimages $m^{\sharp,\ast}$ in $\vartheta_s{\mathscr M}({\mathcal J}^\sharp)\otimes_TT^{\ast}=\vartheta_s({\mathscr M}\boxtimes_TT^\ast)({\mathcal J}^\sharp)$ with $m^{\sharp\ast}|_{\mathcal J}=m$. The uniqueness statement implies that we have actually found an element in $\bigcap_{\alpha\in R^+}\vartheta_s{\mathscr M}({\mathcal J}^\sharp)\otimes_TT^{\alpha}$. By Lemma \ref{lemma-wcrf}, $\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$ is root reflexive, so this is an element $m^\sharp\in\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$ with $m^\sharp|_{\mathcal J}=m$. The uniqueness of $m^\sharp$ is implied by the uniqueness statement for $\ast=\emptyset$.
So we can now assume that $T$ is either generic or subgeneric. Consider the canonical decomposition $\vartheta_s{\mathscr M}=\bigoplus_{\Lambda\in C({\mathcal A}_{T})}(\vartheta_s{\mathscr M})^{\overline\Lambda}$. We can assume that $m$ is supported on $\Lambda\cup\Lambda s$ for some $\Lambda\in C({\mathcal A}_T)$. Moreover, by the support condition we can also assume that ${\mathcal J}$ is a subset of $\Lambda\cup\Lambda s$. We now distinguish the cases $\Lambda=\Lambda s$ and $\Lambda\ne\Lambda s$.
First assume that $\Lambda=\Lambda s$.
Then $T$ must be subgeneric and $\overline\Lambda=\{x,xs\}$ for some ${\mathbb Z} R$-orbit $x$. Lemma \ref{lemma-sinvset} implies that we can fix the notation in such a way that $\pi(({\mathcal J}^\sharp
\setminus{\mathcal J})\cap\Lambda)\subset\{xs\}$ and $\pi(({\mathcal J}\setminus{\mathcal J}^\flat)\cap\Lambda)\subset\{x\}$. Let $\tilde m^\sharp\in\vartheta_s{\mathscr M}({\mathcal J}^\sharp)={\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J}^\sharp)$ be an arbitrary preimage of $m$. Lemma \ref{lemma-decm} allows us to write $\tilde m^\sharp=m_1+m_2$ with an $s$-invariant element $m_1$ and an element $m_2$ that is ${\mathcal Z}$-supported on $\{xs\}$. Now $m|_{{\mathcal J}^\flat}=m_1|_{{\mathcal J}^\flat}+m_2|_{{\mathcal J}^\flat}\in\vartheta_s{\mathscr M}({\mathcal J}^\flat)={\mathcal Z}\otimes_{{\mathcal Z}^s}{\mathscr M}({\mathcal J}^\flat)$ is $s$-invariant, and the uniqueness statement in Lemma \ref{lemma-decm} implies that $m_2|_{{\mathcal J}^\flat}=0$. In particular, $m_2|_{{\mathcal J}}$ is contained in the kernel of $\vartheta_s{\mathscr M}({\mathcal J})\to\vartheta_s{\mathscr M}({\mathcal J}^\flat)$. As $\vartheta_s{\mathscr M}$ satisfies the support condition, Lemma \ref{lemma-suppcond} implies that ${\operatorname{supp}}_{{\mathcal Z}}\,m_2|_{{\mathcal J}}\subset\{x\}$. But by construction ${\operatorname{supp}}_{{\mathcal Z}}\,m_2|_{{\mathcal J}}\subset\{xs\}$. This is only possible if $m_2|_{\mathcal J}=0$. It follows that $m_1|_{\mathcal J}=\tilde m^\sharp|_{\mathcal J}=m$, and so $m^\sharp:=m_1$ is an $s$-invariant preimage of $m$. The uniqueness of $m^\sharp$ again follows from Lemma \ref{lemma-decm}.
Now suppose that $\Lambda\ne\Lambda s$. Then we can write $m=m_{\overline\Lambda}+m_{{\overline{\Lambda s}}}$ with $m_{\overline\Lambda}\in(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J})$ and $m_{{\overline{\Lambda s}}}\in(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J})$. We have $(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J})\cong(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J}_{\Lambda})$ and $(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J})\cong(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J}_{\Lambda s})$. Lemma \ref{lemma-sinvgamma} gives us isomorphisms $(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J}_\Lambda)\xrightarrow{\eta_s^{\overline\Lambda}}(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J}_{\Lambda} s)$ and $(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J}_{\Lambda s})\xrightarrow{\eta_s^{\overline{\Lambda s}}}(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J}_{\Lambda s}s)$. Now $m_{\overline\Lambda}$ and $\eta_s^{\overline{\Lambda s}}(m_{{\overline{\Lambda s}}})$ are sections of $(\vartheta_s{\mathscr M})^{\overline\Lambda}$ over ${\mathcal J}_\Lambda$ and ${\mathcal J}_{\Lambda s}s$, resp. Note that ${\mathcal J}_\Lambda\cap {\mathcal J}_{\Lambda s}s=({\mathcal J}^\flat)_\Lambda$ and ${\mathcal J}_{\Lambda}\cup{\mathcal J}_{\Lambda s}s=({\mathcal J}^{\sharp})_\Lambda$. As $m_{\overline\Lambda}$ and $\eta_s^{\overline{\Lambda s}}(m_{{\overline{\Lambda s}}})$ agree on ${\mathcal J}^\flat$ (as $m|_{{\mathcal J}^\flat}$ is $s$-invariant), and as $(\vartheta_s{\mathscr M})^{\overline\Lambda}$ is a sheaf by Lemma \ref{lemma-subgensheaf} the two sections glue and yield a section $m_{\overline\Lambda}^\sharp\in(\vartheta_s{\mathscr M})^{\overline\Lambda}(({\mathcal J}^\sharp)_\Lambda)=(\vartheta_s{\mathscr M})^{\overline\Lambda}({\mathcal J}^\sharp)$. Similarly, with the roles of $\Lambda$ and $\Lambda s$ interchanged, we construct a section $m_{{\overline{\Lambda s}}}^\sharp$ in $(\vartheta_s{\mathscr M})^{\overline{\Lambda s}}({\mathcal J}^\sharp)$ that restricts to $m_{{\overline{\Lambda s}}}$ and $\eta_s^{\overline\Lambda}(m_{\overline\Lambda})$ on ${\mathcal J}_{\Lambda s}$ and ${\mathcal J}_\Lambda s$, resp.
Now
\begin{align*}
\eta_s^{\overline\Lambda}(m_{\overline\Lambda}^\sharp)|_{{\mathcal J}_{\Lambda} s}=\eta_s^{\overline\Lambda}(m_{\overline\Lambda}^\sharp|_{{\mathcal J}_\Lambda})=\eta_s^{\overline\Lambda}(m_{\overline\Lambda})=m^\sharp_{\Lambda s}|_{{\mathcal J}_\Lambda s}.
\end{align*}
Analogously we obtain $\eta_s^{\overline{\Lambda s}}(m_{{\overline{\Lambda s}}}^\sharp)|_{{\mathcal J}_{\Lambda s}s}=m_{\overline\Lambda}^\sharp|_{{\mathcal J}_{\Lambda s}s}$, from which we deduce $\eta_s^{\overline\Lambda}(m_{\overline\Lambda}^\sharp)|_{{\mathcal J}_{\Lambda s}}=m^{\sharp}_{\Lambda s}|_{{\mathcal J}_{\Lambda s}}$. As $\vartheta_s{\mathscr M}$ is a sheaf and since ${\mathcal J}_\Lambda s\cup{\mathcal J}_{\Lambda s}=({\mathcal J}^\sharp)_{\Lambda s}$, we deduce $\eta_s^{\overline\Lambda}(m_{\overline\Lambda}^\sharp)=m_{{\overline{\Lambda s}}}^\sharp$. From Lemma \ref{lemma-sinvgamma} we deduce that $m^\sharp:=(m_{\overline\Lambda}^\sharp,m_{{\overline{\Lambda s}}}^\sharp)$ is an $s$-invariant section in $\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$. By construction, $m^\sharp|_{\mathcal J}=m_{\overline\Lambda}+m_{{\overline{\Lambda s}}}=m$.
In order to show uniqueness also in the case $\Lambda\ne\Lambda s$ it suffices to show that if $m^\sharp\in\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$ is $s$-invariant and $m^\sharp|_{\mathcal J}=0$, then $m^\sharp=0$. But $m^\sharp|_{{\mathcal J}_\Lambda}=0$ implies $m^{\sharp}|_{{\mathcal J}_\Lambda\cup{\mathcal J}_{\Lambda} s}=0$ by $s$-invariance. Analogously $m^{\sharp}|_{{\mathcal J}_{\Lambda s}\cup{\mathcal J}_{\Lambda s} s}=0$. As ${\mathcal J}^\sharp=({\mathcal J}_\Lambda\cup{\mathcal J}_{\Lambda} s)\cup({\mathcal J}_{\Lambda s}\cup{\mathcal J}_{\Lambda s} s)$ and as both sets in brackets are $s$-invariant, Lemma \ref{lemma-wcrf} implies that $m^\sharp=0$. \end{proof}
\subsection{Wall crossing preserves the category ${\mathbf S}$}
Now we state and prove one of the main results in this article.
\begin{theorem} Suppose that ${\mathscr M}$ is an object in ${\mathbf S}$. Then $\vartheta_s{\mathscr M}$ is an object in ${\mathbf S}$ as well.
\end{theorem}
\begin{proof} We can assume that ${\mathscr M}$ is supported on the connected component $\Lambda$. The case $\Lambda\ne\Lambda s$ is already proven in Lemma \ref{lemma-subgensheaf}. So now suppose that $\Lambda=\Lambda s$. Then $\vartheta_s{\mathscr M}$ is supported on $\Lambda$ as well. First we show that $\vartheta_s{\mathscr M}$ is a sheaf.
Let $\{{\mathcal J}_i\}_{i\in I}$ be a family of open subsets in $\Lambda$ and set ${\mathcal J}=\bigcup_{i\in I}{\mathcal J}_i$. Let $m_i\in\vartheta_s{\mathscr M}({\mathcal J}_i)$ be sections with $m_i|_{{\mathcal J}_i\cap{\mathcal J}_j}=m_j|_{{\mathcal J}_i\cap{\mathcal J}_j}$ for all $i,j\in I$. Using Lemma \ref{lemma-wcrf} we see that the sections $m_i|_{{\mathcal J}_i^\flat}$ glue and yield a section $m^\flat\in\vartheta_s{\mathscr M}({\mathcal J}^\flat)$ (note that ${\mathcal J}^\flat=\bigcup_{i\in I}{\mathcal J}_i^\flat$ by Lemma \ref{lemma-sinvset}). As $\vartheta_s{\mathscr M}$ is flabby, there is a preimage $m^\prime\in\vartheta_s{\mathscr M}({\mathcal J})$ of $m^\flat$. Subtracting $m^\prime|_{{\mathcal J}_i}$ from $m_i$ shows that we can from now on assume that $m_i|_{{\mathcal J}_i^\flat}=0$ for all $i\in I$. In particular, each $m_i|_{{\mathcal J}_i^\flat}$ is $s$-invariant. For any $i\in I$ let $m_i^\sharp\in\vartheta_s{\mathscr M}({\mathcal J}_i^\sharp)$ be the unique $s$-invariant preimage of $m_i$ (cf. Proposition \ref{prop-sinvpreim}). We claim that $m_i^\sharp|_{{\mathcal J}_i^\sharp\cap{\mathcal J}_j^\sharp}=m_j^\sharp|_{{\mathcal J}_i^\sharp\cap{\mathcal J}_j^\sharp}$ for all $i,j\in I$. Note that ${\mathcal J}_i^\sharp\cap{\mathcal J}_j^\sharp=({\mathcal J}_i\cap{\mathcal J}_j)^\sharp$ by Lemma \ref{lemma-sinvset}, and $m_i^\sharp|_{{\mathcal J}_i\cap{\mathcal J}_j}=m_i|_{{\mathcal J}_i\cap{\mathcal J}_j}=m_j|_{{\mathcal J}_i\cap{\mathcal J}_j}=m_j^\sharp|_{{\mathcal J}_i\cap{\mathcal J}_j}$. From the uniqueness statement in Proposition \ref{prop-sinvpreim} we can now deduce $m_i^\sharp|_{({\mathcal J}_i\cap{\mathcal J}_j)^\sharp}=m_j^\sharp|_{({\mathcal J}_i\cap{\mathcal J}_j)^\sharp}$. Using Lemma \ref{lemma-wcrf} again shows that the $m_i^\sharp$ glue and yield a section $m^\sharp\in\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$. Then $m:=m^\sharp|_{{\mathcal J}}$ is the section we are looking for.
Now we show that $\vartheta_s{\mathscr M}$ is root reflexive.
We need to show that $\vartheta_s{\mathscr M}({\mathcal J})$ is root reflexive for any open subset ${\mathcal J}$. Let $m$ be an element in $\bigcap_{\alpha\in R^+}\vartheta_s{\mathscr M}({\mathcal J})\otimes_TT^{\alpha}$. Then $m|_{{\mathcal J}^\flat}\in\vartheta_s{\mathscr M}({\mathcal J}^\flat)$ by Lemma \ref{lemma-wcrf}. By subtracting from $m$ a preimage of $m|_{{\mathcal J}^\flat}$ in $\vartheta_s{\mathscr M}({\mathcal J})$ we can assume that $m|_{{\mathcal J}^\flat}=0$. Proposition \ref{prop-sinvpreim} now shows that $m$ has a unique preimage $m^{\ast\sharp}$ in $\vartheta_s{\mathscr M}({\mathcal J}^\sharp)\otimes_TT^{\ast}$ for all $\ast\in R^+\cup\{\emptyset\}$. By uniqueness, the elements $m^{\ast\sharp}$ agree as elements in $\vartheta_s{\mathscr M}({\mathcal J}^\sharp)\otimes_TT^{\emptyset}$, hence define an element in $\bigcap_{\alpha\in R^+}\vartheta_s{\mathscr M}({\mathcal J}^\sharp)\otimes_TT^{\alpha}$. Again by Lemma \ref{lemma-wcrf}, $\bigcap_{\alpha\in R^+}\vartheta_s{\mathscr M}({\mathcal J}^\sharp)\otimes_TT^{\alpha}=\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$. So $m$ has a preimage $m^\sharp$ in $\vartheta_s{\mathscr M}({\mathcal J}^\sharp)$, so $m$ must be contained in $\vartheta_s{\mathscr M}({\mathcal J})$. Hence $\vartheta_s{\mathscr M}({\mathcal J})$ is root reflexive.
To finish the proof, we need to show that $(\vartheta_s{\mathscr M})\boxtimes_TT^\prime$ is a root reflexive sheaf for any flat homomorphism $T\to T^\prime$ of base rings. As $(\vartheta_s{\mathscr M})\boxtimes_TT^\prime=\vartheta_{T^\prime,s}({\mathscr M}\boxtimes_TT^\prime)$ this follows from what we have already proven and the fact that ${\mathscr M}\boxtimes_TT^\prime$ is a root reflexive sheaf (on ${\mathcal A}_{T^\prime}$).
\end{proof}
|
\section*{Introduction}
In this paper, we study and present, in a uniform fashion, results from two closely related topics. The first topic is the representation of the additive and multiplicative decomposition of the Az\'ema supermartingale associated with a finite honest time or last passage times (see \autoref{honest}), which has received interest in credit risk modeling, the study of asymmetric information and enlargement of filtration, see for example Aksamit et al. \cite{ACDJ}, Fontana et al. \cite{FJS} and Zwierz \cite{Z}. The second topic is on semimartingales of class-$(\Sigma)$, which was first introduced for positive continuous submartingales in Yor \cite{Y1} and later extended to semimartingales in Yor \cite{Y2}, Nikeghbali \cite{N1}, Cheridito et al. \cite{CNP}. In Cheridito et al \cite{CNP}, the authors have shown that the introduction of class-$(\Sigma)$ allows for a martingale proof of the Madan-Roynette-Yor formula \cite{MRY}. The beauty of the Madan-Roynette-Yor formula lies in that it illustrates a deep connection between the last passage time of zero of a semimartinagale of class-$(\Sigma)$ and the price of an European call/put option.
The goal of the current paper is to revisit and extend results in these two closely related areas and our investigation is staged in two parts. In the first part of the paper, we complete the study on representations of the Az\'ema supermartingale associated with a finite honest time by obtaining representations of it's additive and multiplicative decomposition in terms of some process and its running supremum, namely, the draw-down and relative draw-down of the process. In the current literature, results of this type was first shown by Nikeghbali and Yor \cite{NY2} under the assumptions that all martingales are continuous and the given honest time $\tau$ avoids all stopping times (see \autoref{A}). More explicitly, the following additive and multiplicative representation of the Az\'ema supermartingale $Z_t:={\mathbb P}(\tau > t\,|\,{\cal F}_t)$ holds,
\begin{align}
Z_t &= 1+ m_t - \sup_{s\leq t}m_s\label{eqa}\\
Z_t & = \frac{M_t}{\sup_{s\leq t} M_s}\label{eqm}
\end{align}
where $m$ and $M$ are local martingales with continuous supremum and $\lim_{t\rightarrow \infty} M_t = 0$. The above multiplicative representation was later extended in Kardaras \cite{K} and Acciaio and Penner \cite{BI} to the case where the finite honest time $\tau$ avoids all (predictable) stopping times. To illustrate the extend of their results, Acciaio and Penner \cite{BI} gave a counter example from \cite{ACDJ} showing that for a honest time which does not avoid all stopping times, the multiplicative representation \eqref{eqm} do not necessarily hold. The short explanation for this observation is that in \cite{BI, K}, the class of honest times considered are the end of predictable sets, while the counter example and in general, finite honest times are the end of optional sets. This led to the work of Song \cite{S} where the author derived the necessary and sufficient conditions for decomposition of the form \eqref{eqm} to hold for any Az\'ema supermartingale. The open question, which we answer, is whether \eqref{eqa} and \eqref{eqm} holds for any finite honest time. It appears that the main obstacle in removing the assumption that $\tau$ avoids all stopping times is that one is too focused on the supermartingale $Z$ and have insisted that $m$ and $M$ should be kept martingales.
From recent developments in the theory of enlargement of filtration, it is becoming evident that the l\`agl\`ad process $\widetilde Z_t :={\mathbb P}(\tau \geq t\, |\,{\cal F}_t) $ is easier to work with compared to $Z$ and one can retrieve $Z$ by computing the right limit $\widetilde Z_+$. In view of this, we study the additive and multiplicative decomposition of the process $\widetilde Z$. This switch from $Z$ to $\widetilde Z$ is crucial in extending the results of \cite{BI,K, NY2}. By studying $\widetilde Z$, we are able to remove the assumption that $\tau$ avoids all stopping times and obtain representations of the additive and multiplicative decomposition of $\widetilde Z$ in the form given in \eqref{eqa} and \eqref{eqm}, with the key difference been that the martingales $m$ and $M$ are now replaced by {\it optional supermartingales} with continuous supremum (see \autoref{d1.1}). We must point out that the convenience of working with $\widetilde Z$ does not come for free. The main difficulty faced in this study compared to \cite{BI,K, NY2} is that, if we do not assume that the honest time $\tau$ avoids all stopping times then $\widetilde Z$ is not a c\`adl\`ag process and $\tau$ is the end of an {\it optional set}. The main techniques employed, such as Doob-maximal equality and Skorokhod reflection lemma, in \cite{BI,K, NY2} cannot be applied in the current setting and one needs to seek alternative methods.
To obtain our first main result on the representation of the additive and multiplicative decomposition of $\widetilde Z$ in \autoref{t1}, we
first derive a multiplicative decompositions of $\widetilde Z$, which makes use of the strongly predictable process of finite variation in the Doob-Meyer-Merten decomposition of the strong optional supermartingale $\widetilde Z$.
Our approach is inspired by the works of Az\'ema, Meyer and Y\oe urp in \cite{A2,M, M0, YM} on multiplicative decomposition of positive submartingales, and relies strongly on the finer properties of honest times from the theory of enlargement of filtration obtained in Jeulin \cite{J2} and stochastic calculus for optional semimartingales, which was systematically developed in Gal'chuk \cite{G1,G2,G3}.
In the second part of the paper, we place our results on Az\'ema supermartingales in a general framework and extend the notion of semimartingales of class-$(\Sigma)$. The notion of class-($\Sigma$) was first introduced for positive continuous submartingales in Yor \cite{Y1} and later extended in \cite{CNP, N1,NY2, Y1,Y2} to semimartingales and more recently examined in Eyi-Obiang et al. \cite{EOM, EOMT} in the context of signed measures. Roughly speaking this class of processes is closely related to processes which are reflected at the boundary point zero. Typical examples include continuous martingales, absolute value of a continuous martingales, the positive part of a continuous martingale, the draw-down of a continuous martingale and the Az\'ema supermartingale of finite honest times which avoids all stopping times.
In the current literature, given a semimartingale of class-$(\Sigma)$, the assumption that the process of finite variation in the semimartingale decomposition is continuous is build into the definition. This effectively implies that such processes can only reflect off of zero in a continuous fashion. In the context of Az\'ema's supermartingales associated with honest times, this continuity assumption is equivalent to the assumption that the honest time avoids all stopping times. In \autoref{os}, we extend the definition of class-$(\Sigma)$ by allowing for jumps in the process of finite variation. Under the new definition, we will generalise existing results for semimartingales of class-$(\Sigma)$ given in Cheridito et al. \cite{CNP}. In particular, we show in \autoref{l3.1} and \autoref{l3.3} that if $X$ and $Y$ are processes of class-($\Sigma$) then (under some conditions) $X^+$, $X^-$, $|X|$ and $XY$ are again processes of class-$(\Sigma)$. The presentation is similarly to those found in Yor \cite{Y2} and Cheridito et al. \cite{CNP}, but we will relax the no negative or no positive jump assumption that is imposed in \cite{CNP}. As our second main result, we obtain the generalisation of the Madan-Roynette-Yor type formula in \autoref{t3.1} and \autoref{c3.1}.
As an application of the results in \autoref{os}, we illustrate in \autoref{max} and \autoref{e3.1}, a method to construct non-trivial examples of finite honest times for which the additive and multiplicative representation of the Az\'ema's supermartingale can be retrieved from \autoref{t1}, but not from the results of \cite{BI, K, NY2}.
To conclude, in \autoref{soms}, for positive optional submartingales of class-$(\Sigma)$, we revisit the Madan-Roynette-Yor type formula in \autoref{l3.1} (see also Madan et al. \cite{MRY} or Profeta et al. in \cite{PRY}). We illustrate that the Madan-Roynette-Yor type formula can be viewed as a special result from the theory of {\it multiplicative system associated with a positive submartingale}, which was first introduced in the predictable case by Meyer \cite{M} and extend to the optional case in Li and Rutkowski \cite{LR1}. We show in \autoref{cornnew} that given any positive optional submartingale $X$ there exist an optional multiplicative system associated with the positive submartingale $X_+$.
The practical applications of (optional) multiplicative systems are of two folds. Firstly, our result shows that, in essence, the Madan-Roynette-Yor type formula always holds for stock prices whose dynamic follows a positive martingale, and one just have to replace last passage times with optional multiplicative systems. Secondly, given a c\`adl\`ag increasing process $B$ satisfying the condition $\mathbb{E}(B_\infty - B_{t-}\,|\,{\cal F}_t) \leq 1$, then one can construct, using the multiplicative system associated with $1-\mathbb{E}((B_\infty- B_{t-})\,|\,{\cal F}_t)$, a random time $\tau$ such that $\widetilde Z_t= \mathbb{E}(B_\infty- B_{t-}\,|\,{\cal F}_t)$. This construction is used in credit risk modelling as it gives the possibility of constructing default times with a given survival process and do not satisfy the immersion property.
For the reader's convenience, we collect in the appendix useful definitions and results from the theory of enlargement of filtrations and stochastic calculus for optional semimartingales. We also give results on the {\it local time} of optional semimartingales in \autoref{tanaka2} and \autoref{tanaka3}. To the best of our knowledge the {\it local time} of optional semimartingales is not defined/studied in the literature.
\section{Notations and Terminologies}\label{tool}
We work on a filtered probability space $(\Omega,\mathcal A,\mathbb F,\mathbb P)$, where $\mathbb{F}:=({\cal F}_t)_{t\geq 0}$ denotes a filtration satisfying the usual conditions, we set ${\cal F}_\infty := \bigvee_{t\geq 0} {\cal F}_t \subset \mathcal{A}$ and all martingales are taken to be c\`adl\`ag. Given a real valued stochastic process $X$, by convention, we set $X_{0-} = 0$ and $X_\infty = \lim_{t\rightarrow \infty} X_t$ a.s, if it exists. The running supremum of $X$ is denoted by $\overline{X}_t := \sup_{s\leq t} X_s$.
Given a c\`adl\`ag increasing process $A$, we will denote its left support by $S^g(A)$ (see page 61, Chapter IV of Jeulin \cite{J2}). We say that the random measure $dA$ is carried on a set $G$ if $\int_{[0,\infty[}\mathds{1}_{G^c}(s) dA_s = 0$. We denote by $\mathcal{T}$ the set of all stopping times and for $0\leq s< t<\infty$, $\mathcal{T}_{[s,t]}$ the set of all stopping times $T$ such that $s\leq T\leq t$. A stochastic process $X$ is said to be of class-$(D)$ if the family $\seq{X_T\mathds{1}_\seq{T<\infty}, T \in \mathcal{T}}$ is uniformly integrable, and it is said to be of class-$(DL)$ if for every $0<t <\infty$, the family $\seq{X_{T}, T \in \mathcal{T}_{[0,t]}}$ is uniformly integrable.
For any locally integrable variation process $V$, we denote the ${\mathbb F}$-optional (predictable) projection of $V$ by $^{o}V$ ($^{p}V$) and the ${\mathbb F}$-dual optional (predictable) projection of $V$ by $V^{o}$ ($V^{p}$).
It is known that the process $\,^oV - V^o$ is a uniformly integrable ${\mathbb F}$-martingale starting at zero and the equality $\,^o(\Delta V) = \Delta V^o$ holds.
\begin{defi} \rm \label{honest}
A random time $\tau$ is a honest time, if for all $t\geq 0$, there exist a ${\cal F}_t$-measurable random variable $\tau_t$ such that $\tau_t = \tau$ on the set $\seq{\tau < t}$.
\end{defi}
The main tool used in the current work is the stochastic calculus for optional semimartingales, which was systematically developed under the {\it unusual conditions} in Gal'chuk \cite{G1, G2, G3}. However we do not make use of the full power of the calculus and needed only the integration by parts or It\^o's formula and the solution to the stochastic exponential for optional semimartingales. In fact our situation is much simpler as the filtration ${\mathbb F}$ is assumed to satisfy the {\it usual conditions} and all martingales are taken to be c\`adl\`ag. In this setting, an optional martingale is a uniformly integrable martingale, an optional local martingale is a local martingale and any optional semimartingale $X$ takes the form $X = X_0 + M + A$, where $M$ is a local martingale, which we denote by $X^m$, and $A$ is a l\`agl\`ad process of finite variation. These observations are very convenient as stochastic integrals for optional semimartingales essentially reduces to the usual stochastic integral and one need only to take care in counting the jumps of the integral against the process of finite variation $A$.
In the rest of this paper, unless otherwise stated, all stochastic processes in concern are optional semimartingales which is known to exhibit finite left and right limits. Given any l\`agl\`ad process $X$, following the notation given in \cite{G1, G2,G3}, we denote by $X_-$ and $X_+$ the left and right limits of $X$. The left and right jumps of $X$ are denoted by $\Delta X= X - X_-$ and $\Delta^+ X = X_+-X$ respectively. The process $X$ can be decomposed into its right continuous part and left continuous part by setting $X^g = \sum_{s< \cdot} \Delta^+ X_s$ and $X^r := X - X^g$. The right continuous part $X^r$ can be further decomposed into $X^r = X^c + X^d$, where $X^d = \sum_{s\leq \cdot} \Delta X_s$ and $X^c := X^r - X^d$. This gives us the decomposition,
\begin{align}
X = X^c + X^d+ X^g. \label{Xg}
\end{align}
Finally, we mention that prior to the work of Gal'chuk \cite{G1}, Merten \cite{MJ} introduced under the usual conditions the notion of strong optional supermartingale and generalized the Doob-Meyer decomposition to this setting. However, here we work with optional supermartingales as defined by Gal'chuk (see \autoref{d1.1}), and point out that the process $\widetilde Z$ is both an optional supermartingale and a strong optional supermartingale. Therefore, in the rest of the paper, we will abuse slightly the terminology and call the Doob's decomposition for optional supermartingales, the Doob-Meyer-Merten-Gal'chuk decomposition.
For more details on the general theory for stochastic processes the reader is referred to He et al. \cite{HWY}, for results from the theory of enlargement of filtrations to Jeulin \cite{J2}. The reader can also refer to the recent book of Aksamit and Jeanblanc \cite{AJ} for a modern english exposition of the results from the theory of enlargement of filtration. In general, we will follow as closely as possible the notations established in \cite{G1,G2, G3, HWY,J2}.
\section{The Additive and Multiplicative Representations}\label{ad}
In this section, given a finite honest time $\tau$, we study the additive and multiplicative representation of the Az\'ema supermartingale associated with $\tau$.
Before proceeding, we first introduce some quantities that are specific to the study of random times. For an arbitrary random time $\tau$, we set $H:=\mathds{1}_{[\![\tau,\infty[\![}$
and define \hfill\break
\noindent $\bullet$ the supermartingale $Z$ associated with $\tau$, $Z :=\,^{o}(\mathds{1}_{[\![0,\tau[\![})= 1- \,^{o}H$,\hfill\break
\noindent $\bullet$ the supermartingale $\widetilde Z$ associated with $\tau$, $\widetilde Z :=\,^{o}(\mathds{1}_{[\![0,\tau]\!]})= 1- \,^{o}(H_-)$,\hfill\break
\noindent $\bullet$ the martingale $m :=1-\left ( \,^{o}H-H^o\right )$.
In the literature, the process $Z$ is often termed the Az\'ema supermartingale. For simplicity, we shall name the process $1-Z$, the Az\'ema submartingale. From \cite{J2}, we know that the above processes are linked through the following relationships:
\begin{equation}
\label{relation}
Z=m-H^o \quad \textrm{and} \quad \widetilde Z=m- (H^o)_-
\end{equation}
and we have $\widetilde Z - Z = \Delta H^o$, $\widetilde Z_+ = Z$ and $\widetilde Z_- = Z_-$. From Theorem 5.22 \cite{HWY}, the dual optional projection $H^o$ is of integrable variation since $H$ is of integrable variation. At time equal to zero, we have $1-\widetilde Z_0 =0$ and $1-\widetilde Z_{0+} = 1-Z_0 = 1-(\Delta H^o)_0 = 1-(H^o)_0$. We set $R := \inf \seq{s: Z_s = 0}$ and for a random time $\tau$, it is well know that $\tau \leq R$ and the support of $H^o$ is contained in $\llbracket\, 0, R\,\rrbracket$.
The process $\widetilde Z$ is a bounded positive (strong) optional supermartingale with it's Doob-Meyer-Merten-Gal'chuk decomposition given by $\widetilde Z = m- (H^{o})_-$. For notational simplicity and to be consistent with the notation later used in the additive decomposition of optional semimartingales, we will set $A:= (H^o)_-$. Note that $A$ is a left continuous process and one can apply the decomposition in \eqref{Xg} to obtain the additive decomposition $\widetilde Z = m - A^c - A^g$.
\begin{defi} \rm \label{A}
A random time $\tau$ is said to avoids all ${\mathbb F}$-stopping times, that is ${\mathbb P}(\tau = T<\infty) = 0$ for all ${\mathbb F}$-stopping times $T$, then $\Delta H^o = 0$ and $\widetilde Z = Z$.
\end{defi}
The main difficulty we face here is that without the assumption that $\tau$ avoids all stopping times, the process $\widetilde Z$ is not c\`adl\`ag and the standard Skorokhod reflection lemma can not be applied to obtain the additive representation as done in \cite{NY2} and the Doob-maximal equality cannot be applied to obtain the multiplicative representation as done in \cite{BI,K, NY2}. Also unlike the Az\'ema supermartingale $Z$, the multiplicative decomposition of $\widetilde Z$ is not available in the literature.
Therefore, we first obtain below the multiplicative decomposition of $\widetilde Z$. The main obstacle in constructing the multiplicative decomposition of $\widetilde Z$ is that it is not bounded from below and this is dealt with using the finer properties of honest times given in \autoref{l1.2} and \autoref{p1.1}. We start by considering the additive decomposition of $\widetilde Z = m - A^c - A^{g}$ and we set
\begin{align*}
Y_t &:= \int_{]0,t]} \widetilde Z^{-1}_sdA^c_s + \int_{[0,t[} \widetilde Z^{-1}_{s+} dA^{g}_{s+}= :Y^c_t + Y^g_t.
\end{align*}
Both the processes $Y^c$ and $Y^g$ are increasing (possibly non-finite) and stopped after $\tau$. For both $Y^c$ and $Y^g$ to be well defined finite increasing processes, we need to check that the they are finite before $\tau$ and have finite left limits at $\tau$.
Note that $Y^c = A^c$ since the support of $A$ is carried on $\seq{\widetilde Z= 1}$. Recall that $R := \inf \seq{s: Z_s = 0}$ and on the set $\seq{\tau < R}$, it is clear that both $Y^c$ and $Y^g$ are well defined, since the support of $A$ is contained in the set $\llbracket\, 0, \tau \,\rrbracket\subset \seq{\widetilde Z > 0}$. On the set $\seq{\tau = R}$, for $Y^c$ we make use of the fact that $\widetilde Z_{\tau-} = Z_{\tau-} > 0$ since $\llbracket\, 0, \tau \,\rrbracket \subset \seq{Z_- > 0}$. For $Y^g$, we see that
\begin{gather*}
\int_{[0,\tau[} \widetilde Z^{-1}_{s+} dA^{g}_{s+} = \sum_{0\leq s< \tau}Z^{-1}_s\Delta^+ A^{g}_s
\end{gather*}
which is finite since $\llbracket\, 0, \tau \,\llbracket\, \subset \seq{Z >0}$.
We define $\widetilde D$ to be the optional stochastic exponential of $Y$, that is $\widetilde D$ is the unique solution to the following equation
\begin{align}
\widetilde D_t &= 1 + \int_{]0,t]}\widetilde D_{s-} dY^{c}_s + \int_{[0,t[}\widetilde D_s dY^{g}_{s+}\nonumber \\
&= 1+ \int_{]0,t]}\widetilde D_{s-}\widetilde Z^{-1}_s dA^c_s + \int_{[0,t[}\widetilde D_s\widetilde Z^{-1}_{s+} dA^{g}_{s} \label{Dcg}\\
&=: 1+ \widetilde D^c_t + \widetilde D^g_t\nonumber .
\end{align}
Using \autoref{stochexp}, the unique solution to the optional stochastic exponential is given by
\begin{align*}
\widetilde D = e^{Y^c}e^{Y^g}\prod_{0\leq s< \cdot} (1+\Delta^+ Y_s)e^{\Delta^+ Y_s}.
\end{align*}
The process $\widetilde D$ can be further decomposed multiplicatively into $\widetilde D = D^cD^g$, where
\begin{align*}
D^c &:= e^{Y^c} \quad \mathrm{and}\quad D^g := e^{Y^g}\prod_{0\leq s< \cdot} (1+\Delta^+ Y_s)e^{\Delta^+ Y_s}.
\end{align*}
From the form of $Y^c$ and $Y^g$, we see that $D^c$ and $D^g$ are strictly positive increasing processes and the support of $D^c$ and $D^g_+$ are contained in the support of $dA_+$ and is carried on the set $\seq{\widetilde Z = 1}$.
\begin{lem}\label{mdz}
The process $\widetilde M = \widetilde D\widetilde Z$ is a local martingale.
\begin{proof}
By an application of It\^o's formula given in \autoref{ito},
\begin{align*}
& \widetilde D_t\widetilde Z_t - \widetilde D_0\widetilde Z_0 \\
& = \int_{]0,t]}\widetilde D_{s-}d\widetilde Z^r_s + \int_{[0,t[}\widetilde D_s d\widetilde Z^g_{s+} + \int_{]0,t]}\widetilde Z_{s-}d\widetilde D^r_s + \int_{[0,t[}\widetilde Z_{s} d\widetilde D^g_{s+} + \sum_{0\leq s<t} \Delta \widetilde D^g_s \Delta^+\widetilde Z_s + \sum_{0< s\leq t} \Delta \widetilde D^r_s \Delta\widetilde Z_s\\
& = \int_{]0,t]}\widetilde D_{s-}d\widetilde Z^r_s + \int_{[0,t[}\widetilde D_s d\widetilde Z^g_{s+} + \int_{]0,t]}\widetilde Z_sd\widetilde D^c_s + \int_{[0,t[}\widetilde Z_{s+} d\widetilde D^g_{s+}\\
& = \int_{]0,t]}\widetilde D_{s-}dm_s
\end{align*}
where in the last equality, we have used the fact that $\widetilde Z^r = m - A^c$, $\widetilde Z^g = A^{g}$ and \eqref{Dcg}.
\end{proof}
\end{lem}
\brem
For a honest time, the process $\widetilde D$ is well defined on $[0,\infty)$. The difference between honest times and general random times is that the support of $dA_+$ is contained in $\llbracket\, 0, \tau\,\rrbracket$ rather than $\llbracket\, 0, R\,\rrbracket$, this allows us to handle the case where $Z$ go to zero continuously at $R$ on the set $\seq{\tau < R}$.
\erem
The key idea of this section is that we consider, instead of the uniformly integrable martingale $m$ and $\widetilde M$, the optional supermartingales $n : = m - A^{g}$ and $N: = \widetilde M/D^g$. In general, the process $n$ and $N$ not necessarily c\`adl\`ag on $\llbracket\, 0 ,\tau \,\rrbracket$ and due to \autoref{p1.1} it is only c\`adl\`ag on $\,\rrbracket \tau ,\infty \llbracket\,$. We note that since $A^{g}$ has only positive jumps, the process $\overline{n}_t = \sup_{s\leq t} n_s$ and $\overline{N}_t = \sup_{s\leq t} N_s$ must be c\`adl\`ag and hence optional.
\begin{lem}\label{l2.1}
For a finite honest time, we have $\seq{n = \overline{n}} = \seq{\widetilde Z = 1}$.
\begin{proof}
From \autoref{p1.1},
\begin{equation}\label{teq1}
\tau = \sup\seq{s: \widetilde Z_s = 1} = \sup\seq{s: n_s=1+ A^c_{s} }.
\end{equation}
From the inequality $\widetilde Z \leq 1$, we deduce that $n \leq 1+ A^c$ and $\seq{n=1+ A^c} \subseteq \seq{n=\overline n}$. Using \eqref{teq1} and the fact that the process $1+A$ is constant after $\tau$ (from \autoref{l1.2} (i) and \autoref{l1.3}), the process $\overline{n}$ must equal to the constant process $1+A^c$ after $\tau$. This together with the fact that $\seq{\widetilde Z = 1}$ is contained on $\llbracket\, 0, \tau \,\rrbracket$, we have
\begin{displaymath}
\seq{n=\overline{n}}\,\cap \,\,\rrbracket \tau ,\infty \llbracket\, = \seq{\widetilde Z=1}\,\cap \,\,\rrbracket \tau ,\infty \llbracket\, = \phi.
\end{displaymath}
This implies $\seq{n=1+ A^c} \subseteq \seq{n=\overline n} \subseteq \llbracket\, 0,\tau\,\rrbracket$. By \autoref{l1.2}, the set $\seq{\widetilde Z = 1}$ is the largest optional set contained in $\llbracket\, 0, \tau\,\rrbracket$ from which we conclude that $\seq{n= \overline {n}} = \seq{\widetilde Z= 1}$.
\end{proof}
\end{lem}
The set equality $\seq{n = \overline{n}} = \seq{n = 1+A^c}$ implies that $\overline{n} = 1+A^c$ on the set $\seq{\widetilde Z = 1}$. The equality $\overline{n} = 1+A^c$ should also hold everywhere, since they have the same initial condition and the support of $\overline{n}$ and $1+A^c$ are both carried on the set $\seq{\widetilde Z = 1}$. To show this we make use of the following time change process,
\begin{align*}
g_t &= \sup\seq{s \leq t: \widetilde Z_s = 1}.
\end{align*}
Note that $\widetilde Z_\tau = 1$, but in general, it is not true that $\widetilde Z_{g_t} = 1$.
\begin{theorem}\label{t1}
Let $\tau$ be a finite honest time random time then
\begin{align*}
\widetilde Z &= 1+ n - \overline{n}\\
\widetilde Z & = N/\overline{N}
\end{align*}
where $n = m - A^{g}$ and $1+A^c = \overline {n}$, and $N = \widetilde ZD^c = \widetilde M/D^g$ and $D^c =e^{A^c}= \overline {N}$.
\begin{proof}
The process $\overline{n}$ and $1+A^c$ have the same initial condition and it is clear that $\overline{n} \leq 1+A^c$. To show the reverse inequality, we must consider two cases. Given a stopping time $T$, suppose $g_T(\omega) \in \seq{t:\widetilde Z_t(\omega)=1}$ then the set equality $\seq{n = \overline{n}} = \seq{n = 1+A^c} = \seq{\widetilde Z = 1}$ shows that
\begin{align*}
1+ A^c_T = 1+ A^c_{g_T} = \overline{n}_{g_T} \leq \overline{n}_T.
\end{align*}
The other case we must consider is $g_T(\omega)\not \in \seq{t:\widetilde Z_t(\omega) = 1}$ but belongs to the closure $\seq{t:\widetilde Z_t = 1}$, i.e. there exists a sequence $(t_n(\omega))_{n\in{\mathbb N}}$ in $\seq{t:\widetilde Z_t(\omega)= 1}$ converging to $g_T(\omega)$ from the left. Then
\begin{align*}
1+ A^c_T = 1+ A^c_{g_T} = (\overline{n}_{-})_{g_T} \leq \overline{n}_T,
\end{align*}
which implies $1+ A^c = \overline{n}$.
To prove the multiplicative representation, we first make the observation that the support of $dD^{c}$ and $dD^{g}_+$ are both carried on the set $\seq{\widetilde Z= 1} = \seq{N= D^c}$. Then one can repeat the arguments of \autoref{l2.1} and the proof of the additive representation with $N$ and $D^c$ in place of $n$ and $1+ A^c$ to show that $\seq{\widetilde Z = 1} = \seq{N = D^c} = \seq{N = \overline{N}}$ and $\overline{N}$ is equal to $D^c$.
\end{proof}
\end{theorem}
\brem
The counter example given by Acciaio and Penner \cite{BI} to demonstrate that \eqref{eqm} does not necessarily hold is an example of a honest time in the Poisson filtration for which $A^c = 0$. In this particular example, the multiplicative representation given in \autoref{t1} holds trivially in the sense that $N = \widetilde Z$ and $\overline{N} = 1$.
To obtain non-trivial examples, we need to find finite honest times for which both $A^c$ and $A^g$ are non-zero. Honest times with such property can be constructed by taking known examples of finite honest times $\tau$ from the Brownian filtration and consider the honest time $\tau \vee \sigma$ where $\sigma$ is any finite stopping time. We will discuss this type of construction in more detail in \autoref{e3.1} once we have developed some generic tools.
\erem
\section{Optional Semimartingales of Class-$(\Sigma)$}\label{os}
In this part of the paper, we study the Az\'ema supermartingale of finite honest times in a general context and extend the notion of semimartingales of class-$(\Sigma)$, which was first introduced for continous positive submartingales in Yor \cite{Y2} and subsequently extended in Yor \cite{Y1}, Nikeghabali \cite{N1} and Cheridito et al. \cite{CNP}. More recently, processes of class-$(\Sigma)$ was also studied in the context of signed measures in Eyi-Obiang et al. \cite{EOM, EOMT}. Typical examples of semimartingales of class-$(\Sigma)$ in the literature includes continuous martingales, the absolute value of a continuous martingale, the draw-down process of a continuous martingale and the Az\'ema submartingale $1-Z$ associated with finite honest times which avoids stopping time. From these examples we make the observe that the behaviour of this class of processes is characterised by how it reflects off of the boundary point zero and is therefore closely related to the Skorokhod reflection conditions.
\begin{defi} \rm \label{d3.1}
An optional semimartingale $X$ with decomposition $X = X_0+ M + A$ where $M$ is a local martingale with $M_0= 0$ and $A = A^c + A^d + A^g$ is a l\`agl\`ad process of finite variation with $A_0 =0$ is said to satisfy the Skorokhod minimal reflection condition at zero if
\begin{gather*}
\int_{[0,\infty[} \mathds{1}_\seq{X_s \neq 0 } (dA^c_s+dA^g_{s+}) = 0 \quad \mathrm{and} \quad \int_{[0,\infty[}\mathds{1}_\seq{X_{s-} \neq 0 } dA^d_s= 0.
\end{gather*}
\end{defi}
\begin{defi} \rm \label{d3.2}
An optional semimartingale $X$ is said to be of class-$(\Sigma)$ if it satisfies the Skorokhod minimial reflection condition and $X_0 + A^d = 0$.
\end{defi}
We point out that the condition $X_0 + A^d= 0$ together with the convention that $A_0= 0$ implies $X_0 = 0$ and $X_{0+} = A^g_{0+}$. This is consistent with the behaviour of the submartingales $1-\widetilde Z$ and $1-Z$ at time zero. In the rest of the paper, given an optional semimartingale $X$, the process $M$ and $A$ will denote the local martingale and the l\`agl\`ad process of finite variation in the optional semimartingale decomposition of $X$. The left jumps are given by $\Delta X = \Delta M$, the right jumps are given by $\Delta^+ X = \Delta A_+ = \Delta A^g_+$ and $\seq{\Delta^+ X \neq 0} \subset \seq{X= 0}$.
In the previously mentioned works \cite{BI,EOM, EOMT, K,N1, NY2, Y1,Y2}, the process of finite variation $A$ in the decomposition of $X$ is continuous by definition. The main contribution of this section of the paper is that we extend the definition and previous results on semimartingales of class-$(\Sigma)$ to optional semimartingales for which the finite variation part $A$ can have jumps. This is a non-trival extension, since recent study on honest times in the Poisson filtration have given us explicit examples of positive optional submartingales of class-$(\Sigma)$ whose finite variation part $A$ is purely discontinuous. In fact it is proven that in Theorem 4.8 of Aksamit et al. \cite{ACJ} that in any discrete/jumping filtration (in particular, the Poisson filtration), the finite variation part $A$ in the Doob-Meyer-Merten-Gal'chuk decomposition of $\widetilde Z$ must be purely discontinuous.
We now explore some general property of optional semimartingales of class-$(\Sigma)$ defined in \autoref{d3.2} and demonstrate that under conditions `certain' some conditions, if $X$ and $Y$ are processes of class-($\Sigma$) then $X^+$, $X^-$, $|X|$ and $XY$ are again processes of class-$(\Sigma)$. In particular, we relax the no negative or no positive jump assumption that is imposed in Cheridito et al. \cite{CNP} and replace them with the following conditions.
\begin{defi} \rm \label{nojump}
An optional semimartingale $X$ is said to \hfill\break
(i) cross zero continuously from the positive side if $\sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} > 0 }(X_{s})^- =0$.\hfill\break
(ii) cross zero continuously from the negative side if $\sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} \leq 0 }(X_{s})^+ =0.$\hfill\break
(iii) cross zero continuously if it cross zero continuously from the positive and the negative side.
\end{defi}
\begin{lem} \label{l3.1}
Let $X$ be an optional semimartingale of class-$(\Sigma)$ then:\hfill\break
(i) The processes $X^+$, $X^-$ and $|X|$ are local optional submartingales.\hfill\break
(ii) If $X$ crosses zero continuous from the positive side then $X^+$ is of class-$(\Sigma)$.\hfill\break
(iii) If $X$ crosses zero continuous from the negative side then $X^-$ is of class-$(\Sigma)$.\hfill\break
(iv) If $X$ crosses zero continuous then $|X|$ is of class-$(\Sigma)$.
\begin{proof}
$(i)$ The fact that $X^+$, $X^-$ and $|X|$ are local submartingales follows directly from the Tanaka formula in \autoref{tanaka2} and the fact that $dA$ is carried on $\seq{X= 0}$.
\noindent $(ii)$ We proof only $(ii)$ as the proof of $(iii)$ and $(iv)$ are similar. By \autoref{tanaka2}
\begin{align*}
X_t^+ &=
\int_{]0,t]} \mathds{1}_\seq{X_{s-} > 0}d(A^c_s + M_s) + \int_{[0,t[} \mathds{1}_\seq{X_{s} > 0}\, dA^g_{s+} + \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} > 0 }(X_{s})^- \\
& \quad + \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} \leq 0 }(X_{s})^+ + \sum_{0\leq s < t} \mathds{1}_\seq{X_s > 0 }(X_{s+})^- + \sum_{0\leq s < t} \mathds{1}_\seq{X_s \leq 0 }(X_{s+})^+ + \frac{1}{2}L^0_t(X).\\
&= \int_{]0,t]} \mathds{1}_\seq{X_{s-} > 0}dM_s + \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} \leq 0 }(X_{s})^+ + \sum_{0\leq s < t} \mathds{1}_\seq{X_s > 0 }(X_{s+})^- \\
& \quad + \sum_{0\leq s < t} \mathds{1}_\seq{X_s \leq 0 }(X_{s+})^+ + \frac{1}{2}L^{0}_t(X)
\end{align*}
where in the last equality, we have used the fact that the process $X$ enters zero continuously from the positive side. By \autoref{tanaka3}, the support of $dL^0(X)$ is carried on the set $\seq{X= 0} \subseteq \seq{X^+ = 0}$. The right hand jumps
$$\sum_{0\leq s < t} \mathds{1}_\seq{X_s > 0 }(X_{s+})^- + \mathds{1}_\seq{X_s \leq 0 }(X_{s+})^+,$$
are supported on $\seq{X^+ = 0}$ because $\seq{\mathrm{sign}(X) \neq \mathrm{sign}(X_+)} \subseteq \seq{\Delta^+X > 0}$ and
\begin{align*}
\seq{\Delta^+X > 0} & \subseteq \seq{X = 0} \subseteq \seq{X^+ = 0}.
\end{align*}
The l\`agl\`ad process of finite variation $A$ in the optional semimartingale decomposition of $X$ is left continuous and therefore predictable. This implies that there exists a localising sequence of stopping times $(T_n)_n$ such that $X^+ = (M + A)^+$ is integrable and $M^{T_n}$ is a uniformly integrable martingale. This implies that the increasing process
\begin{align*}
V_t = \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} \leq 0 }(X_{s})^+
\end{align*}
stopped at $T_n$ is of integrable variation and the dual predictable projection $V^p$ of $V$ exists and is locally of integrable variation.
To show that $V^p$ is continuous, we note that on the set $\seq{\Delta V>0}$, the jump $\Delta V$ is bounded by $\Delta M$. Also the set $\seq{\Delta V>0}$ is equal to the left support of $V$ and from Jeulin \cite{J2} Chapter IV, Lemma 4.2, we have $S^g(V) \subseteq S^g(V^p)$ which is a predictable set. This shows that
\begin{align*}
\Delta V^p = \,^p(\Delta V) & \leq \,^p(\Delta M\mathds{1}_\seq{\Delta V > 0}) \\
& \leq \,^p(\Delta M) \mathds{1}_{S^g(V^p)},
\end{align*}
and from the predictable sampling theorem, $\,^p(\Delta M)_T = 0$ for all predictable stopping times $T$. Using the continuity of $V^p$ we obtain
\begin{align*}
\mathbb{E}\big(\int_{[0,T_n[} \mathds{1}_\seq{X^+_s > 0} dV^p_s\,\big) & = \mathbb{E}\big(\int_{[0,{T_n}[} \mathds{1}_\seq{X^+_{s-} > 0} dV^p_s\,\big)\\
& = \mathbb{E}\big(\int_{[0,{T_n}[} \mathds{1}_\seq{(X^+)_{s-} > 0}\mathds{1}_\seq{X_{s-}\leq 0} dV_s \,\big) = 0.
\end{align*}
Finally, by monotone convergence theorem, we let $n\rightarrow \infty$ to show that $V^p$ is supported on $\seq{X^+=0}$.
From similarly arguments we can also conclude that $X^-$ and $|X|$ are also positive optional submartingales of class-$(\Sigma)$.
\end{proof}
\end{lem}
\begin{lem} \label{l3.3}
Given two optional semimartingales $X$ and $Y$ of class-$(\Sigma)$, suppose $[X^{m},Y^{m}]= 0$ then $XY$ is an optional semimartingale of class-$(\Sigma)$.
\begin{proof}
By application of the It\^o formula in \autoref{ito}
\begin{align*}
& X_tY_t - X_0Y_0 \\
& = \int_{]0,t]}Y_{s-}dX^r_s + \int_{[0,t[}Y_s d X^g_{s+} + \int_{]0,t]} X_{s-}dY^r_s + \int_{[0,t[} X_{s} dY^g_{s+} + [X^{m}, Y^{m}]_t + \sum_{0\leq s<t} \Delta Y^g_s \Delta^+ X_s \\
& = \int_{]0,t]}Y_{s-}dX^r_s + \int_{[0,t[}Y_{s} d X^g_{s+} + \int_{]0,t]} X_{s-}dY^r_s + \int_{[0,t[} X_{s+} dY^g_{s+}.
\end{align*}
To see that the finite variation part only moves on the set $\seq{XY = 0}$, it is sufficient to note that $\seq{XY \neq 0} = \seq{X\neq 0}\cap\seq{Y \neq 0}$.
\end{proof}
\end{lem}
For simplicity, we present the following lemma for $C^1$-functions rather than bounded measurable, since unlike Nikeghbali \cite{N1} or Cheridito et al. \cite{CNP}, we do not attempt to solve the Skorokhod embedding problem for optional semimartingales of class-$(\Sigma)$ and we only include the following lemma to illustrate that class-$(\Sigma)$ is closed under this transform. In fact, we point out that the continuity of the process $A$ was crucial in their solution of the Skorokhod embedding problem for semimartingales of class-$(\Sigma)$.
\begin{lem}\label{l3.2} Let $X$ be an optional semimartingale of class-$(\Sigma)$ then for any $C^1$-function $f$
\begin{gather*}
f(A_{t})X_t = f(0)X_0 + \int_{]0,t]} f(A_{s})dM_s + \int_{[0,t[} f(A_{s+})dA_{s+}
\end{gather*}
is an optional semimartingale of class-$(\Sigma)$.
\end{lem}
\begin{proof}
We first note that the process $A$ is left continuous and by an application of the It\^o formula in \autoref{ito} we obtain
\begin{align*}
f(A_{t})X_t & = f(0)X_0 + \int_{]0,t]} f(A_{s})dX^r_s + \int_{[0,t[} f(A_{s})dX^g_s \\
& \quad + \int_{]0,t]} X_{s-}df(A_s)^r + \int_{[0,t[} X_{s}df(A_s)^g + \sum_{s<t} (f(A_{s+})- f(A_s))\Delta A^+_s.
\end{align*}
By applying the It\^o's formula to $f(A)$,
\begin{align*}
df(A_t) & = \int_{]0,t]}f'(A_s)dA^r_s + \int_{[0,t[}f'(A_s)dA^g_{s+} + \sum_{s<t} f(A_{s+}) - f(A_s) - f'(A_s)\Delta^+ A_s\\
& = \int_{]0,t]}f'(A_s)dA^c_s + \int_{[0,t[}f'(A_s)dA^g_{s+} + \sum_{s<t} f(A_{s+}) - f(A_s) - f'(A_s)\Delta^+ A_s.
\end{align*}
The support of $df(A)^g_+$ is contained in the support of $dA^g_+$ which is contained in $\seq{X= 0}$, therefore
\begin{align*}
f(A_t)X_t & = f(0)X_0 + \int_{]0,t]} f(A_s)dX^r_s + \int_{[0,t[} f(A_s)dX^g_s + \sum_{s<t} (f(A_{s+})- f(A_s))\Delta A^+_s\\
& = f(0)X_0 + \int_{]0,t]} f(A_s)d(X^r - A^c)_s + \int_{]0,t]} f(A_s)dA^c_{s} + \int_{[0,t[} f(A_{s+})dA^g_{s+}\\
& = f(0)X_0 + \int_{]0,t]} f(A_s)dM_s + \int_{[0,t[} f(A_{s+})dA_{s+}.
\end{align*}
It is sufficient to note that since the support of $A_+$ is carried on $\seq{X= 0 }$
\begin{gather*}
\int_{[0,t[} \mathds{1}_\seq{X_sf(A_{s}) \neq 0}f(A_{s+})dA_{s+} = \int_{[0,t[} \mathds{1}_\seq{X_s\neq 0}\mathds{1}_\seq{f(A_{s})\neq 0}f(A_{s+})dA_{s+} = 0.
\end{gather*}
\end{proof}
\subsection{Last Passage Times and Optional Semimartingales of Class-$(\Sigma)$}
From the works of Madan et al. \cite{MRY}, Profeta et al, \cite{PRY} and generalizations of Cheridito et al. \cite{CNP} we know that, under some continuity assumptions, there is a deep connection between semimartingales of class-$(\Sigma)$ and their last passage time of zero, given by
\begin{align*}
\tau & : =\sup\seq{s:X_s = 0}.
\end{align*}
In the following, this connection is re-discovered for optional semimartingale of class-$(\Sigma)$. Deviating from the works of \cite{CNP, EOM, EOMT, N1}, we will re-examine this connection for processes of class-$(DL)$ rather than class-$(D)$. To do this, consider
\begin{align*}
g_t &= \sup\seq{s \leq t: X_s = 0}\\
k_t &= \inf\seq{s > t: X_s = 0}
\end{align*}
where $\seq{g_t \leq u} = \seq{t< k_u}$ and $\tau = g_\infty = \sup\seq{s: X_s = 0}$. The right jumps $\Delta^+ X = \Delta A^g$ only are non-negative on $\seq{X= 0}$ which implies $X_{k_t} = 0$, but $X_{g_t}$ may or may not be zero. Note that by definition $X_0 = 0$, which allows us to circumvent the issue that $g_t$ can be undefined.
We first derive below a balayage type formula which provides us with the martingale that will underpin later computations. We point out that the balayage type formula presented below cannot be obtained from the standard balayage formula for c\`adl\`ag processes given in for example \cite{Y1}, since $X_{+}$ at $k_t$ might not be zero.
\begin{lem}\label{l3.4} Let $X$ be an optional semimartingale of class-$(\Sigma)$ and fix $u\geq 0$. Then for $t\geq u$
\begin{align*}
(X_{+})_t\mathds{1}_\seq{g_{t} \leq u} & = M_{t\wedge k_{u}} + A^c_{u} + (A^g_+)_{u}
\end{align*}
and for $t>u$,
\begin{align*}
X_t\mathds{1}_\seq{g_{t} \leq u} & = M_{t\wedge k_{u}} + A^c_{u} + (A^g_+)_{u}
\end{align*}
\end{lem}
\begin{proof}
The process $X_{t+}\mathds{1}_\seq{g_t \leq u} = X_{t+}\mathds{1}_\seq{t < k_u}$ is right continuous in $t$ and by the usual It\^o's formula
\begin{align*}
X_{t+}\mathds{1}_\seq{g_t \leq u} & = X_{t+}\mathds{1}_\seq{t < k_u}\\
& = X_{u+} + M_{t\wedge k_u} - M_u + \int_{(u,t]} \mathds{1}_\seq{s \leq k_u} dA^c_s + \int_{(u,t]} \mathds{1}_\seq{s \leq k_u} dA^g_{s+} - X_{k_u+}\mathds{1}_\seq{t\geq k_u}\\
& = M_{t\wedge k_u} + A^c_{t\wedge u} + (A^g_{+})_{t\wedge k_u} - \Delta A^g_{k_u}\mathds{1}_\seq{t\geq k_u}
\end{align*}
where in the last equality, we used the fact that $X_{k_u} = 0$ and $\Delta^+ X = \Delta A^g_+$. On the set $\seq{t\geq k_u}$
\begin{align*}
(A^g_{+})_{t\wedge k_u} - \Delta A^g_{k_u}\mathds{1}_\seq{t\geq k_u} & = (A^g)_{k_u} \mathds{1}_\seq{t\geq k_u}\\
& = (A^g_+)_{u} \mathds{1}_\seq{t\geq k_u}.
\end{align*}
On the complement $\seq{t < k_u}$, we have $(A^g_{+})_{t\wedge k_u} = (A^g_{+})_{t} = (A^g_{+})_{u}$, where the last equality comes from the fact that $u\leq t <k_u$ and $A^g_+$ does not increase on $\llbracket\, u, k_u\llbracket\,$. This shows that
\begin{align*}
(X_{+})_t\mathds{1}_\seq{g_t \leq u} &= X_0 + M_{t\wedge k_{u}} + A^c_{u} + (A^g_+)_{u}
\end{align*}
and $(X_{+})_t\mathds{1}_\seq{g_t \leq u}$ is a local martingale for $t\geq u$. On the other hand, for fixed $u\geq 0 $ and all $t>u$,
$$(X_{+})_t\mathds{1}_\seq{g_t \leq u} = X_t\mathds{1}_\seq{g_t \leq u},$$
since $\Delta (A^g_{+})_t= 0$ on the set $\seq{g_t\leq u< t}$.
\end{proof}
The first observation we make from \autoref{l3.4} is that for fixed $0\leq u$
\begin{align}
X_{t+}\mathds{1}_\seq{g_t\leq u} - X_{u+}\mathds{1}_\seq{g_u \leq u} & = M_{t}^{k_{u}} - M_{u} \qquad \,\,\,u\leq t,\label{bala}\\
X_{t}\mathds{1}_\seq{g_t\leq u} - X_{s}\mathds{1}_\seq{g_s \leq u} & = M_{t}^{k_{u}} - M^{k_u}_{s} \qquad u<s\leq t.\label{bala1}
\end{align}
By examining \eqref{bala}, we note that if the local martingale $M^{k_u}-M^u$ is a true martingale on $[u,\infty)$ then one can take the conditional expectation with respect to ${\cal F}_u$ and eliminate the right hand side using optional sampling theorem (see for example Theorem 2.58 of \cite{HWY}). The second observation we make is that the integrability properties of $M^{k_{u}}_s - M_{s}^{u}$ for $s \in [u,\infty)$ can be derived from the integrability properties of $X_{s+}\mathds{1}_\seq{g_s\leq u} - X_{u+}\mathds{1}_\seq{g_u\leq u}$ for $s\in [u,\infty)$.
Of course one can in the definition of class-$(\Sigma)$ restrict ourselves to optional semimartingales for which $M$ is a martingale, however we wish to look for some sufficient conditions on the process $X$ or $X_+$. The assumption that $X$ and $X_+$ are of class-$(D)$ is likely too strong for problems that is on a finite horizon. It is more practical to assume that $M$ is a martingale rather than an uniformly integrable martingale. For example, one can take the Brownian motion, which is of class-$(\Sigma)$, but does not belong to class-$(D)$.
\begin{theorem}\label{t3.1}
(i) Suppose $X_+$ is of class-$(DL)$ then for $0\leq u\leq s\leq t < \infty$,
\begin{align*}
\mathbb{E}(X_{t+}\mathds{1}_\seq{g_{t} \leq u}|{\cal F}_s) & = X_{s+}\mathds{1}_\seq{g_{s} \leq u}.
\end{align*}
(ii) Suppose $X$ is of class-$(DL)$ then for $0\leq u< s\leq t < \infty$,
\begin{align*}
\mathbb{E}(X_t\mathds{1}_\seq{g_t < u}|{\cal F}_s) & = X_s\mathds{1}_\seq{g_s < u}.
\end{align*}
\begin{proof}
(i) The result can be deduce from \eqref{bala} given that we have shown for fixed $u\geq 0$, the process $(M^{k_{u}}_s-M^{u}_s)_{s\geq u}$ is a martingale. Inspired by the fact that a local martingale is a martingale if and only if it is of class-$(DL)$. We show that a sufficient condition for $(M^{k_{u}}_s-M^{u}_s)_{s\geq u}$ to be a martingale is that for all $t\geq u$ the family $\mathcal{M}_{[u,t]} :=\seq{M^{k_{T\wedge u}}_{T}- M^u_{T}\,: \,T\in \mathcal{T}_{[0,t]}}$ is uniformly integrable.
Let $(T_n)_{n\in {\mathbb N}}$ be a localizing sequence which makes $M^{T_n}$ a martingale, then for $t\geq s> u$
\begin{displaymath}
\mathbb{E}[M^{k_{T_n\wedge u}}_{T_n\wedge t}-M^u_{T_n\wedge t}\,|\,{\cal F}_s] = M^{k_{T_n\wedge u}}_{T_n\wedge s}-M_{T_n\wedge u}
\end{displaymath}
where we applied optional sampling theorem to the bounded stopping times $k_{T_n\wedge u}\wedge T_n\wedge t$ and $s$. As $n\rightarrow \infty$, the right hand side in the above converges almost surely to $M^{k_{u}}_{s}-M_{u}$ and by uniform integrability of the family $\mathcal{M}_{[u,t]}$, the left hand side converges to $\mathbb{E}[M^{k_{u}}_{t}-M_{u}|{\cal F}_s]$. This shows that $(M^{k_{u}}_s-M^{u}_s)_{s\geq u}$ is a martingale.
On the other hand, by taking $t = t\wedge T_n$, $u = u\wedge T_n$ in \eqref{bala}, we obtain
\begin{align*}
|M_{T_n\wedge t}^{k_{T_n\wedge u}} - M^u_{T_n}| & = |(X_+)_{T_n\wedge t}\mathds{1}_\seq{g_{T_n\wedge t}\leq {T_n\wedge u}} - (X_{+})_{T_n\wedge u}|\\
& \leq |X_+|_{T_n\wedge t}+ |X_{+}^u|_{T_n\wedge t}
\end{align*}
from which one deduce that if $X_+$ is of class-$(DL)$ then $\mathcal{M}_{[u,t]}$ is a uniformly integrable family.
(ii) From similar arguments to (i), we deduce from \eqref{bala1} that if $X$ is of class-$(DL)$ then
\begin{displaymath}
\mathbb{E}(X_{t}\mathds{1}_\seq{g_t\leq u} |{\cal F}_s) = X_{s}\mathds{1}_\seq{g_s \leq u}, \quad t>u>0
\end{displaymath}
and the claim for $u>0$ follows from monotone convergences theorem applied in $u$. At $u=0$ the equality holds trivially since $(g_s)_{s\geq 0}$ is positive.
\end{proof}
\end{theorem}
In the following, we will obtain similar results at $t = \infty$. In general, to establish the desired result for $X$ at infinity, one will have to first establish convergence results for non-right continuous submartingales which are not available in the literature. We avoid this issue by supposing that the honest time $\tau := \sup\seq{s:X_s = 0}$ is finite almost surely. This is a convenient assumption, because $dA_+$ is carried on the set $\seq{X=0}$ and the process $A$ is flat on $\,\rrbracket \tau, \infty \,\rrbracket$. Therefore if $X_+$ converges almost surely to a random variable $X_\infty$, then $X$ must also converge almost surely to $X_\infty$.
\begin{theorem}\label{c3.1}
(i) Suppose $X_+$ is of class-(D) then for all $s\geq 0$
$$\mathbb{E}(X_\infty\mathds{1}_\seq{\tau \leq s}|{\cal F}_s) = X_{s+}.$$
(ii) Suppose $X$ and $X_+$ are both of class-$(D)$ and $\tau := \sup\seq{s:X_s = 0} <\infty$ a.s, then for all $s\geq 0$,
$$\mathbb{E}(X_\infty\mathds{1}_\seq{\tau < s}|{\cal F}_s) = X_{s}.$$
\end{theorem}
\begin{proof}
(i) From \autoref{l3.1}, we know that $X_+$ can be written as the difference of two submartingale, that is $X_+ = (X^+)_+ - (X^-)_+$ and $|X_+| = (X^+)_+ + (X^-)_+$. This implies that both $(X^+)_+$ and $(X^-)_+$ are right continuous positive submartingales of class-$(D)$ and there exists a random variable
$$X_\infty := (X^+)_\infty-(X^-)_\infty \in L^1.$$
One can now take $s= u$ in \autoref{t3.1} (i). The result then follows by taking the limit as $t\rightarrow \infty$ and using uniform integrability of $X_+$, the left continuity of the process $\mathds{1}_{\llb0,s\,\rrbracket}$ and the fact that $\lim_{t\uparrow \infty} g_t = \tau$ almost surely.
(ii) Given that $X_+$ is of class-$(D)$ and $\tau < \infty$, we have $\lim_{t\rightarrow \infty} X_t = \lim_{t\rightarrow \infty} X_{t+} = X_\infty$ almost surely. From \autoref{l3.1} (i), $X^+$ and $X^-$ are local optional submartingales. Using the fact that $X$ is of class-$(D)$ we can deduce that $X^+$ and $X^-$ are submartingales of class-$(D)$. The process $X^+$ is an optional submartingales because the right continuous positive submartingale $(X^+)_+$ is of class-$(D)$ and by optional sampling theorem we have for all stopping times $T$
\begin{align*}
\mathbb{E}(X_\infty^+ \,|\,{\cal F}_T) \geq (X^+)_{T+} \geq (X^+)_{T}.
\end{align*}
Similar arguments shows that $X^-$ is an optional submartingale of class-$(D)$. From the Doob-Meyer-Merten-Gal'chuk decomposition we can write $X^+ = m + a$ and $X^- = u+v$, where $m$ and $u$ are optional martingales and $a$ and $v$ are strongly predictable increasing process of integrable variation. This shows that $M + A = m-u + a-v$. Since $X$ is of class-$(\Sigma)$, the process $A$ is left continuous and therefore strongly predictable. This shows that $M-(m-u)= (a-v)-A$, which a (c\`adl\`ag) predictable optional local martingale of finite variation. This implies $M = m-u$ and $M$ is an optional martingale and therefore uniformly integrable.
We set $\sigma_s = \inf\seq{u\geq s: X_u= 0}$ and observe that for $s\geq 0$
\begin{align*}
X_{\sigma_s} & = X_\infty\mathds{1}_\seq{\tau< s} + X_{\sigma_s}\mathds{1}_\seq{\tau\geq s}\\\
& = X_\infty\mathds{1}_\seq{\tau< s}.
\end{align*}
The left hand side above is $X_{\sigma_s} = M_{\sigma_s} - A_{s}$, because $A$ is left continuous and $dA$ is carried on $\seq{X= 0}$. Since $M$ is uniformly integrable, the result follows from optional sampling theorem.
\end{proof}
\subsection{Positive Optional Submartingale of Class-$(\Sigma)$}
In this subsection, we let $X$ be a positive optional submartingale and take as convention $X_0 = 0$. This implies that $X$ and $X_+$ are both of class-$(D)$, since the uniform integrability of the families $\seq{X_T\mathds{1}_\seq{T<\infty}, T\in \mathcal{T}}$ and $\seq{X_{T+}\mathds{1}_\seq{T<\infty}, T\in \mathcal{T}}$ follows from the definition of optional submartingale given by Gal'chuk in \autoref{d1.1}.
By \autoref{dm1}, the optional submartingale $X$ can be decomposed into $X= X_0 + M+A$ where $M$ is an optional martingale and $A$ is a strongly predictable increasing process of integrable variation. Again we say that $X$ is of class-$(\Sigma)$ if $dA^c$ and $dA^g_+$ are carried on the set $\seq{X=0}$ and $X_0 + A^d = 0$. In the following, we again decompose the process $A$ into $A = A^c + A^g$, set $n := M + A^g$ and denote the Skorokhod reflection map by $R$, that is given function $x:{\mathbb R}\rightarrow {\mathbb R}$, $R(x) = (\overline{-x})\vee 0$.
To avoid technicalities, we again assume that the honest time $\tau = \sup\seq{s:X_s = 0}$ is almost surely finite. In this setting, as time goes to infinity, both the processes $X$ and $X_+$ both converges almost surely to a random variable $X_\infty \in L^1$ (see for example Theorem 2.5 of \cite{HWY}) and the assumptions of \autoref{t3.1} and \autoref{c3.1} are always satisfied. That is the Madan-Roynette-Yor type formula always holds for positive optional submartingales of class-$(\Sigma)$.
\begin{lem}
Suppose $X$ is a positive optional submartingale of class-$(\Sigma)$ then $A^c =R(n)$, where $n := M + A^g$.
\end{lem}
\begin{proof}
From the fact that $X\geq 0$, we have $-n\leq A^c$ and
\begin{align*}
\seq{X = 0} = \seq{-n = A^c} \subseteq \seq{-n = \overline{-n}}.
\end{align*}
This shows that $A^c = \overline{-n}$ on the set $\seq{X = 0}$. It is also evident that the processes $R(n) = (\overline{-n})\vee 0$ and $A^c$ have the same initial condition. Using the the inequality $-n \leq A^c$, we can conclude that $\overline{-n} \leq A^c$ and $(\overline{-n})\vee 0 \leq A^c$. To show the reverse inequality, we need to consider two cases. That is for any $t>0$, $g_t = 0$ or $g_t>0$. In the case where $g_t > 0$, we have by continuity of $A^c$
\begin{align*}
A^c_t = A^c_{g_t} = (\overline{-n})_{g_t} = (\overline{-n})_{g_t} \vee 0 \leq (\overline{-n})_t \vee 0.
\end{align*}
In the case where $g_t = 0$, we have $A^c_t = A^c_0 = 0 \leq (\overline{-n})_t \vee 0$. This gives us $A^c = (\overline{-n})\vee 0$.
\end{proof}
In the following, we apply \autoref{l3.3} to construct examples of processes of class-$(\Sigma)$ where both $A^c$ and the left continuous purely discontinuous part $A^g$ are non-zero. The idea is to look into filtrations which are the product of a continuous filtration and a discrete/jumping filtration. The fact that in continuous filtrations, all martingale are continuous is well known, while the case of discrete/jumpting filtration it was shown in Theorem 1 of Jacod and Skorokhod \cite{JS} that all martingales are almost surely of locally finite variation. By combining Theorem 1 of Jacod and Skorokhod \cite{JS} and \autoref{l3.3}, one can produce non-trivial example of positive optional submartingales of class-$(\Sigma)$ as defined in \autoref{d3.2} by taking the products of known examples in the Brownian (see Mansuy and Yor \cite{MY}) and Poisson filtrations (see Aksamit et al. \cite{ACJ}).
We shall describe one such example in \autoref{e3.1}, which is also gives an example where the multiplicative representation derived in \autoref{t1} is non trivial.
\begin{defi} \rm
A honest time is said to be from a continuous (discrete) filtration if the martingale in the Doob-Meyer-Merten-Gal'chuk decomposition of $1-\widetilde Z$ is continuous (locally of finite variation).
\end{defi}
\begin{pro}\label{max}
Let $\tau^c$ be a finite honest time from a continuous filtration and $\tau^d$ be a finite honest time from a discrete filtration, then
\begin{align*}
{\mathbb P}(\tau^c\vee \tau^d \leq t\,|\,{\cal F}_t) &= \widetilde F^c_{t+} \widetilde F^d_{t+}\\
{\mathbb P}(\tau^c\vee \tau^d < t\,|\,{\cal F}_t) &= \widetilde F^c_t \widetilde F^d_t.
\end{align*}
where $\widetilde F^c:= \,^o(\mathds{1}_{\,\rrbracket \tau^c , \infty \llbracket\,} )$ and $\widetilde F^d:= \,^o(\mathds{1}_{\,\rrbracket \tau^d , \infty \llbracket\,} ).$
\end{pro}
\begin{proof}
Both the processes $\widetilde F^c$ and $\widetilde F^d$ are positive optional submartingales of class-$(\Sigma)$. From \autoref{honest} the random time $\tau^c\vee \tau^d$ is a honest time and from \autoref{l3.3} the process $\widetilde F^c\widetilde F^d$ is a positive submartingale of class-$(\Sigma)$. We observe that
\begin{align*}
\tau^c\vee \tau^d = \sup\seq{s:\widetilde F^d_s = 0} \vee \sup\seq{s:\widetilde F^c_s = 0} = \sup\seq{s:\widetilde F^c_s \widetilde F^d_s = 0}
\end{align*}
and the result follows by applying \autoref{c3.1} to $\widetilde F^c\widetilde F^d$.
\end{proof}
The representation obtained in \autoref{max} is interesting in that the Az\'ema submartingale of the maximum of two honest times is expressed as the product of the Az\'ema submartingale associated with the each individual honest time. To the best of our knowledge, this type of representation has not previously appeared in the literature. We conjecture that it is possible to express any finite honest time as the maximum of a finite honest time from a continuous filtration and a finite honest time from a discrete/jumping filtration.
\begin{ex} \rm \label{e3.1}
Example of honest times in Brownian filtration are widely available, we consider the following taken from Mansuy and Yor \cite{MY}. Let $B$ be a Brownian motion and
\begin{align*}
T &:= \inf\seq{t: B_t = 1} \\
\tau^c &:= \sup\seq{u\leq T : B_u = 0}.
\end{align*}
The random time $\tau^c$ is a honest time and it's Az\'ema's submartingale is given $F^c = M^c + A^c$ where
\begin{align*}
M^c_t &= B^+_{t\wedge T}-\frac{1}{2}L^0_{t\wedge T} \\
A^c_t &= \frac{1}{2}L^0_{t\wedge T}
\end{align*}
where $L^0$ is the Brownian local time at zero. In this example both $M^c$ and $A^c$ are continuous.
For examples of honest time in a jumping filtrations, we consider the example studied in Proposition 4.12 of \cite{ACJ}. Let $X$ be a compound Poisson process with intensity $\mu$. Given $a \geq 0$, we set
$$\tau^d := \sup\seq{t : \mu t - X_t \leq a}.$$
Under some conditions on the intensity and the distribution of the jump size, it is known that $\tau^d$ is a finite honest time. We denote by $\Psi(x)$ the ruin probability associated with the process $\mu t - X_t$, i.e., for every $x \geq 0$, $\Psi(x) := {\mathbb P}(t^x < \infty)$ with $t^x := \inf\seq{t : x + \mu t - X_t < 0}$. It was shown in Proposition 4.12 \cite{ACJ} that the Az\'ema submartingale of $\tau^d$ admits the decomposition $F^d = M^d + A^d$,
\begin{align*}
M^d_t & = 1- (1 - \Psi(0))\sum_n \mathds{1}_\seq{t\geq T_n} - \Psi(\mu t - X_t - a)\mathds{1}_\seq{\mu t-X_t\geq a} - \mathds{1}_\seq{\mu t-X_t< a}\\
A^d_t & = (1 - \Psi(0))
\sum_n \mathds{1}_\seq{t\geq T_n}.
\end{align*}
In this case, the martingale $M^d$ is of finite variation and $A^d$ is predictable purely discontinuous process with jump times given by $(T_n)_{n\in {\mathbb N}}$.
\end{ex}
\section{Optional Multiplicative Systems}\label{soms}
To conclude, we present some results on the multiplicative system associated with a positive submartingales. Given a positive optional submartingales $X$ which is of class-$(\Sigma)$, the random field $C_{u,t} = \mathds{1}_\seq{\tau_{t} \leq u}$ for $0\leq u\leq t\leq \infty$ is a {\it optional multiplicative system} associated with $X_+$ and as a special case, \autoref{t3.1} and \autoref{c3.1} or the Madan-Roynette-Yor type formula can be obtained from the theory of multiplicative system associated with a positive submartingale.
The notion of a predictable multiplicative system was first introduced in Meyer \cite{M} and later extended to the optional case in Li and Rutkowski \cite{LR1}. However, in \cite{LR1}, the existence of an optional multiplicative system was only established for the special case of the Az\'ema submartingales. In the following, we show that given any positive optional submartingale $X$ the existence of an optional multiplicative system associated with $X_+$. As an additional application, we use the optional multiplicative system to construct random times with given dual optional projection.
Recall that any positive optional submartingale $X$ and it's right continuous modification $X_+$ are of class-$(D)$, therefore $X_\infty = \lim_{t\uparrow \infty} X_{t+}$ exists and is in $L^1$. For the readers convenience we recall below the definition of a multiplicative system.
\begin{defi} \rm \label{defms}
A {\it multiplicative system} is a positive random field $(C_{u,t})_{u,t\in [0,\infty]}$ satisfying the following conditions: \hfill\break
(i) for all $u \leq s \leq t$ the equality $C_{u,s}C_{s,t} = C_{u,t}$ holds; moreover, $C_{u,t} = 1$ for $u\geq t$, \hfill\break
(ii) for any fixed $u \in {\mathbb R}_+$, the process $(C_{u,t})_{t\in [0,\infty]}$ is adapted and decreasing, \hfill\break
(iii) for any fixed $t \in {\mathbb R}_+$, the process $(C_{u,t})_{u\in [0,\infty]}$ is right-continuous and increasing. \hfill \break
A multiplicative system is called {\it predictable} ({\it optional}, resp.) when for each $u$, the process $(C_{u,t})_{t\in [0,\infty]}$ is predictable (optional, resp.).
\end{defi}
\begin{defi} \rm \label{defmsy}
Suppose $X$ is a positive optional submartingale, we say that $(C_{u,t})_{u,t\in [0,\infty]}$ is a
{\it multiplicative system associated with} $X_+$ if, in addition to conditions (i)-(iii) of Definition \ref{defms},
we have, for all $t \in [0,\infty] $,
\begin{equation}
\cE{{\mathbb P}}{C_{t,\infty}X_\infty}{{\cal F}_t} = X_{t+} . \label{msdef}
\end{equation}
\end{defi}
\begin{ex} \rm
Given the strike price $K>0$ and the stock process $S$ which is a continuous uniformly integrable martingale. The Madan-Roynette-Yor formula can be retrieved from \eqref{msdef} as soon as one set $X_t = (K-S_t)^+$ and $C_{u,t} = \mathds{1}_\seq{g_t < u}$ where $g_t = \sup \seq{s\leq t: X_s = 0}$.
\end{ex}
The main tool used to obtain the existence of an optional multiplicative system associated with a submartingale is the Doob-Meyer-Merten-Gal'chuk decomposition given in \autoref{dm1}. Given any positive optional submartingale $X$, we can write $X = X_0 + M+ A$, where $M$ is an uniformly integrable martingale and $A$ is a strongly predictable process of integrable variation and we can again decompose $A$ into $A^r + A^g$, where $A^r$ and $A^g$ are strongly predictable.
\begin{lem} \label{oms}
Given a positive optional submartingale $X$ which is bounded below by a strictly positive constant.
Let the random field $\bar C_{u,t}$ be defined
by $\bar C_{u,t} = 1$ for all $u\geq t$ and satisfies the equation
\begin{equation} \lab{mseq1xy}
\bar C_{u,t} = 1 - \int_{]u,t]}\bar C_{u,s-} (\,^pX_{s})^{-1} \, dA^r_{s} - \int_{[u,t[}\bar C_{u,s} (X_{s+})^{-1} \, dA^g_{s+} , \quad \forall \, t \geq u .
\end{equation}
Then the process $(\bar Q_{u,t}:=\bar C_{u,t} X_{t})_{t\in [u,\infty]}$
is a positive uniformly integrable martingale and it satisfies
\begin{equation} \label{xmueq1x}
d\bar Q_{u,t} = -\bar C_{u,t-}\, dM_t
\end{equation}
and $C_{u,t} := \bar C_{u+,t+}$ is a multiplicative system associated with $X_+$.
\end{lem}
\begin{proof}
We will first show that, for any fixed $u$, the process $(\bar C_{u,t})_{t \in [0,\infty ]}$ is positive and bounded by one.
To this end, it suffices to observe that it is a decreasing process with $\bar C_{u,t} = 1$ for $t \leq u$ and from the \eqref{mseq1xy}, the left jump satisfies $\bar C_{u,t} = \bar C_{u,t-} (1- X^{-1}_{t} \Delta A^r_{t})$ for $t > u$.
Since $X_{t-} = (^pX)_{t}- \Delta A^r_t$, we obtain
\begin{displaymath}
0<(^pX)_{t-}X^{-1}_{t-} = 1- (^pX)_{t-}^{-1} \Delta A^r_{t} \leq 1,
\end{displaymath}
While the right jump satisfies $\bar C_{u,t+} = \bar C_{u,t} (1- X^{-1}_{t+} \Delta A^g_{t+})$ for $t \geq u$. Since $X_{t} = X_{t+} - \Delta A^g_t$, we obtain similarly
\begin{displaymath}
0< X_{t+}X^{-1}_{t} = 1- X^{-1}_{t+} \Delta A^g_{t+} \leq 1.
\end{displaymath}
Thus we conclude that the process $(\bar C_{u,t})_{t \in [u,\infty ]}$ is positive and bounded by one.
Therefore the process $(\bar Q_{u,t}=\bar C_{u,t}X_t)_{t\in [u,\infty]}$ is positive. Next, we show that
the process $(\bar Q_{u,t})_{t\in [u,\infty]}$ is a uniformly integrable martingale.
To this end, from the It\^o formula we obtain
\begin{displaymath}
d\bar Q_{u,t} = -\bar C_{u,t-}\, dM_t
\end{displaymath}
which is a uniformly martingale since $\bar C_{u,t}$ is bounded by one and $M$ is uniformly integrable. For ease of notation we set
\begin{align*}
Y_t:= 1- \int_{(0,t]}(\,^pX_{s})^{-1} \, dA^r_{s} - \int_{[0,t)}(X_{s+})^{-1} \, dA^g_{s+}.
\end{align*}
From \autoref{stochexp} we have
\begin{displaymath}
\bar C_{u,t} = \exp\left\{ Y_t - Y_u - \frac{1}{2}\int_{(u,t]}d\left< Y^c,Y^c\right>_s \right\}\prod_{u<s\leq t} (1+\Delta Y_s)e^{-\Delta Y_s}\prod_{u\leq s< t} (1+\Delta^+ Y_s)e^{-\Delta^+ Y_s}.
\end{displaymath}
From the above, it is clear that $\bar C_{u,t} = \bar C_{0,t}\bar C^{-1}_{0,u}$ and $\bar C_{u,u} = 1$.
Strictly speaking the random field $\bar C_{u,t}$ is not a multiplicative system associated with $X$ since it is not right continuous in $u$. Therefore we need to regularise the random field by considering for all $u\leq t$ and $\epsilon> 0$
\begin{gather*}
\bar C^{-1}_{0,u+\epsilon}\bar C_{0,t+\epsilon}X_{t+\epsilon} = X_{u+\epsilon} + \bar C^{-1}_{0,u+\epsilon}\int_{(0,t+\epsilon]} \bar C_{0,s-} dM_s - \bar C^{-1}_{0,u+\epsilon}\int_{(0,u+\epsilon]} \bar C_{0,s-} dM_s.
\end{gather*}
The right continuity of the stochastic integral gives for all $t\geq u$
\begin{align*}
\bar C^{-1}_{0,u+}\bar C_{0,t+}X_{t+} & = X_{u+} - \bar C^{-1}_{0,u+}\int_{(0,t]} \bar C_{0,s-} dM_s + \bar C^{-1}_{0,u+}\int_{(0,u]} \bar C_{0,s-} dM_s\\
& = X_{u+} - \int_{(u,t]} \bar C^{-1}_{0,u+}\bar C_{0,s-} dM_s.
\end{align*}
We note that for $s> u$, the term $\bar C_{u+, s-} = \bar C^{-1}_{0,u+}\bar C_{0,s-}$ is positive and bounded by one. Therefore $C_{u,t} := \bar C^{-1}_{0,u+}\bar C_{0,t+}$ is a multiplicative system associated with the submartingale $X_+$.
\end{proof}
\begin{cor} \lab{cornnew}
For any optional submartingale $X$ then there exists an optional multiplicative system associated with $X_+$.
\end{cor}
\begin{proof}
We set $C_{u,t} := \lim_{\epsilon \downarrow 0} C^\epsilon_{u,t}$ where $C^{\epsilon}_{u,t}$ is defined in Lemma \ref{oms}
for $X^{\epsilon} = X+\epsilon $. It is obvious that the process $C^\epsilon_{u,t}$ is decreasing in $t$ and increasing in $u$. Moreover,
\begin{align*}
\mathbb{E}_{\mathbb{P}} \big( X_\infty C_{u,\infty}\,\big|\, {\cal F}_u \big) & = \mathbb{E}_{\mathbb{P}} \Big( \lim_{\epsilon \downarrow 0} \, (X_\infty + \epsilon) C^\epsilon_{u,\infty} \,\big|\, {\cal F}_u \Big)
= \lim_{\epsilon \downarrow 0} \mathbb{E}_{\mathbb{P}} \Big( X^{\epsilon}_\infty C^\epsilon_{u,\infty}\,\big|\, {\cal F}_u \Big) \\
& = \lim_{\epsilon \downarrow 0} \, (X_{u+} + \epsilon )
=X_{u+}
\end{align*}
where the second equality holds by the monotone convergence theorem and the third equality is a consequence of Lemma \ref{oms}.
\end{proof}
\noindent {\bf Application of multiplicative systems to the construction of random times:}
\noindent Given an adapted integrable c\`adl\`ag increasing process $B$ such that $X_t:=1-\mathbb{E}(B_\infty - B_{t-}\,|\,{\cal F}_t)$ is positive, we present a method to construct a random time $\tau$ such that the dual optional projection of $H:=\mathds{1}_{\llbracket\, \tau, \infty \llbracket\,}$ is given by $B$ (which implies that $\widetilde Z = X$). This construction is an extension of the Cox-construction used in credit risk modeling and a similar construction was given in \cite{LR1} for the dual predictable projection. In view of this, we will not present the full motivation and will refer the reader to \cite{JS1,JS2,LR1} for potential applications.
As in the Cox-construction of random times, we suppose that there exist an uniformly distributed random variable $U$ on $[0,1]$ which is independent of ${\cal F}_\infty$. We note that since $B_-$ is left continuous we have $B_- = B^c + B^g$. Let $X^\epsilon: =X+\epsilon$, one can construct a random field $\bar C^\epsilon_{u,t}$ as done in \eqref{mseq1xy} and by applying monotone convergence theorem as done in \autoref{cornnew} to show that there exists a random field $\bar C_{u,t}$ such that $X_t\bar C_{u,t}$ is a martingale for $u\leq t$. Furthermore for a fixed $t$, $\bar C_{u,t}$ is left continuous in $u$, since $\bar C^\epsilon_{u,t}$ is left continuous in $u$ and this property is preserved in the limit.
We define $\tau : = \inf \seq{u\geq 0: X_\infty\bar C_{u,\infty} > U}$ and for every $t\geq 0$, we have $\seq{\tau < t} = \seq{X_\infty\bar C_{t,\infty} > U}$. From \eqref{xmueq1x} and independence of $U$, we have
\begin{align*}
{\mathbb P}(\tau < t\, |\, {\cal F}_t) &= \mathbb{E}(X_\infty\bar C_{t,\infty} > U \,|\,{\cal F}_t)\\
&= \mathbb{E}(X_\infty \bar C_{t,\infty}|\,{\cal F}_t) = X_t.
\end{align*}
We deduce from the uniqueness of the Doob-Meyer-Merten-Gal'chuk decomposition, $B_{-} = (H^o)_{-}$ and since both $B$ and $H^o$ are c\`adl\`ag (or one can apply Theorem 5.30 in \cite{HWY}) we obtain $B = H^o$ .
\section{Appendix}
\subsection{Theory of Enlargement of Filtration}
We refer the reader to Aksamit and Jeanblanc \cite{AJ} for an english exposition of the following results.
\begin{lem}[Lemma 4.3, Chapter IV of Jeulin \cite{J2}] \label{l1.2}
Given a random time $\tau$, we have that\hfill\break
(i) the sets $\seq{\widetilde Z_- = 1}$ and $\seq{\widetilde Z = 1}$ are the largest predictable and optional set contained in the stochastic interval $\llbracket\, 0,\tau\,\rrbracket$,\hfill\break
(ii) the stochastic interval $\llbracket\, 0,\tau\,\rrbracket$ is contained in the sets $\seq{\widetilde Z > 0}$ and $\seq{\widetilde Z_- > 0}$,\\
(iii) the stochastic interval $\llbracket\, 0,\tau\,\llbracket\,$ is contained in the sets $\seq{Z > 0}$ and $\seq{Z_- > 0}$.
\end{lem}
\begin{pro}[Proposition 5.1, Chapter V of Jeulin \cite{J2}] \label{p1.1}
The following are equivalent;\hfill\break
(i) a random time $\tau$ is a finite honest time.\\
(iii) $\tau = \sup \seq{s: \widetilde Z_s = 1}$, i.e. it is the end of an optional set.\\
(ii) $H^o_t = H^o_{t\wedge \tau}$ for $t\geq 0$.\\
(v) $\widetilde Z_\tau = 1$.
\end{pro}
\begin{lem}\label{l1.3}
Let $\tau$ be a finite honest time then the support of $dH^o$ is carried on $\seq{\widetilde Z=1}$.
\begin{proof}
It enough to note that $\mathbb{E}(\int_{[0,\infty[} \mathds{1}_\seq{\widetilde Z_s< 1} dH^o_s) = {\mathbb P}(\widetilde Z_\tau< 1) = 0$
\end{proof}
\end{lem}
\subsection{Stochastic Calculus for Optional Semimartingales}
In the following we recall some definition and results for stochastic calculus for optional semimartingales. Note that in our setting, an optional martingale is a uniformly integrable martingale.
\begin{defi} \rm \label{d1.1}
A stochastic process $X$ is said to be an optional (super or sub)martingale if it is a (super or sub)martingale and (i) $X$ is an optional process, (ii) for any stopping time $T$, $X_T\mathds{1}_\seq{T<\infty}$ is integrable, (iii) there exists an integrable random variable $\zeta$ such that for any stopping time $T$, $X_T = \mathbb{E}(\zeta|{\cal F}_T)$ ($X_T \geq \mathbb{E}(\zeta|{\cal F}_T)$ or $X_T \leq \mathbb{E}(\zeta|{\cal F}_T)$) a.s. on the set $\seq{T<\infty}$.
\end{defi}
\begin{defi} \rm \label{d1.3}
A l\`agl\`ad stochastic process $X$ is said to be strongly predictable if $X$ is predictable and the right limit $X_+$ is optional.
\end{defi}
\begin{defi} \rm
A stochastic process $X$ is called an optional semimartingale if it can be written as
\begin{displaymath}
X = X_0 + M + A, \quad M_0 =0,\quad A_0 = 0,
\end{displaymath}
where $M$ is a local martingale and $A$ is an (l\`agl\`ad) adapted process of finite variation.
\end{defi}
\begin{theorem}[It\^o formula. Theorem 8.2 \cite{G1}]\label{ito} Let $X = (X^1,\dots, X^k)$ be an optional semimartingale and $X^k = X^k_0 + M^k + A^k$ for $k= 1,\dots,n$. Let $F(x) = F(x_1,\dots, x_n)$ be a continuously twice differentiable function on ${\mathbb R}^n$ then for $t\in {\mathbb R}_+$,
\begin{align*}
F(X_t) &= F(X_0) + \sum_{k=1}^n \int_{]0,t]} D^kF(X_{s-})d(A^{k,r} + M^{k,r})_s + \frac{1}{2} \sum_{k,l=1}^n\int_{]0,t]} D^kD^lF(X_{s-}) d\left<M^{k,c},M^{l,c}\right>_s\\
&\quad + \sum_{0<s\leq t} \left[F(X_s) - F(X_{s-}) - \sum^n_{k=1} D^kF(X_{s-})\Delta X^k_s \right] +\sum_{k=1}^n \int_{[0,t[} D^kF(X_s)dA^{k,g}_{s+} \\
& \quad + \sum_{0\leq s< t} \left[F(X_{s+}) - F(X_{s}) - \sum^n_{k=1} D^kF(X_{s})\Delta^+ X^k_s \right]
\end{align*}
where $D^k$ is the partial derivative with respect to the $k$-th coordinate and $M^r = M^c + M^d$.
\end{theorem}
\begin{lem}\label{tanaka}(Tanaka formula. Lemma 5.7 \cite{GIOQ}) Let $X$ be a (real-valued) optional semimartingale
with decomposition $X = X_0 + M + A^r + A^g$. Let $f : {\mathbb R} \rightarrow {\mathbb R}$ be a convex
function. Then $f(X)$ is an optional semimartingale. Moreover, denoting by $f'$ the left-hand
derivative of $f$, then we have
\begin{align*}
f(X_t) &= f(X_0) +
\int_{]0,t]} f'(X_{s-})d(A^r_s + M_s) + \int_{[0,t[}f'(X_s)dA^g_{s+}\\
& \quad + \sum_{0<s\leq t} f(X_s) - f(X_{s-}) - f'(X_{s-})\Delta X_s + \sum_{0\leq s < t} f(X_{s+}) - f(X_s) - f'(X_s)\Delta^+X_s + C^f_t
\end{align*}
where $C^f$ is a continuous increasing process.
\end{lem}
\begin{theorem}[Dol\'ean-exponential. Theorem 5.1 \cite{G3}] \label{stochexp}
Let $X$ be an optional semimartingale. There exists a unique (to within indistinguishably) optional semimartingale $S$ such that
\begin{displaymath}
S_t = S_0 + \int_{]0,t]} {S_{s-}}dX^r_s + \int_{[0,t[} S_s dX^g_{s+}
\end{displaymath}
The process $S$ is given by the formula
\begin{displaymath}
S = S_0\exp\left\{ X - \frac{1}{2}\left< X^c,X^c\right> \right\}\prod_{0<s\leq \cdot} (1+\Delta X_s)e^{-\Delta X_s}\prod_{0<s< \cdot} (1+\Delta^+ X_s)e^{-\Delta^+ X_s}
\end{displaymath}
and is termed the optional stochastic exponential $X$ which we shall denote by $\mathcal{E}(X)$.
\end{theorem}
\begin{theorem}[Doob-Meyer-Merten-Gal'chuk decomposition, \cite{G2} \cite{MJ}]\label{dm1}
An optional supermartingale $X$ admits a decomposition $X = M-A$, where $M$ is a (local) optional martingale and $A$ is an increasing strongly predictable (locally) integrable process with $A_0 = 0$ if and only if $X$ belongs to the class-(D) (class-(DL)). This decomposition is unique to within indistinguishably.
\end{theorem}
To the best of our knowledge, there is no results in the literature on the {\it local time} of an optional semimartingales. By following similar arguments to section 6, Chapter IX of \cite{HWY}, we define the local time of an optional semimartingale at $a \in {\mathbb R}$ and show that its support is carried on $\seq{X=a}$.
\begin{lem}\label{tanaka2}
Let $X$ be an optional semimartingale and $a\in {\mathbb R}$. Then
\begin{align*}
(X_t-a)^+ &= (X_0-a)^+ +
\int_{]0,t]} \mathds{1}_\seq{X_{s-} > a}d(A^c_s + M_s) + \int_{[0,t[} \mathds{1}_\seq{X_{s} > a}\, dA^g_{s+} \\
& \quad + \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} > a }(X_{s}-a)^- + \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} \leq 0 }(X_{s}-a)^+ \\
& \quad + \sum_{0\leq s < t} \mathds{1}_\seq{X_s > a }(X_{s+}-a)^- + \sum_{0\leq s < t} \mathds{1}_\seq{X_s \leq a }(X_{s+}-a)^+ + \frac{1}{2}L^a_t(X),\\
(X_t-a)^- &= (X_0-a)^- +
\int_{]0,t]} \mathds{1}_\seq{X_{s-} \leq a}d(A^c_s + M_s) + \int_{[0,t[} \mathds{1}_\seq{X_{s} \leq a}\, dA^g_{s+} \\
& \quad + \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} > a }(X_{s}-a)^- + \sum_{0< s \leq t} \mathds{1}_\seq{X_{s-} \leq a }(X_{s}-a)^+\\
& \quad + \sum_{0\leq s < t} \mathds{1}_\seq{X_s > a }(X_{s+}-a)^- + \sum_{0\leq s < t} \mathds{1}_\seq{X_s \leq a }(X_{s+}-a)^+ + \frac{1}{2}L^a_t(X)
\end{align*}
where $L^a(X)$ is a continuous adapted increasing process with $L^a_0(X) = 0$. The process $L^a(X)$ is called the local time of $X$ at $a\in{\mathbb R}$.
\end{lem}
\begin{proof}
By applying the Tanaka formula in \autoref{tanaka} to $f(x) = (x-a)^+$ and $g(x) = (x-a)^-$ and taking the difference, we obtain
\begin{align*}
(X_t-a)^+ - (X_t-a)^- = (X_0-a)^+ - (X_0-a)^- + \int_{]0,t]} dX^r_s + \int_{[0,t[} dX^g_{s+} + C^f_t-C^g_t.
\end{align*}
This gives $C^f = C^g$ and we denote them by $\frac{1}{2}L^a(X)$.
\end{proof}
\begin{theorem}\label{tanaka3}
Let $X$ be an optional semimartingale and $a \in {\mathbb R}$ then almost surely the support of $dL^a(X)$ is contained in $\seq{X = a}$.
\end{theorem}
\begin{proof}
The proof is similar to that of Theorem 9.44 of \cite{HWY} in the c\`adl\`ag case. We start by supposing $0<S\leq T$ and $\llbracket\, S,T\llbracket\, \subset \seq{X < 0}$. This implies that $\,\rrbracket S,T\,\rrbracket \subset \seq{X_- \leq 0}$ and $\llbracket\, S,T\llbracket\, \subset \seq{X_+ \leq 0}$. By applying the Tanaka formula in \autoref{tanaka},
\begin{align*}
(X_T-a)^+ - (X_S-a)^+ &= \sum_{S< s \leq T} \mathds{1}_\seq{X_{s-} \leq a }(X_{s}-a)^+ + \frac{1}{2}L^a_T(X) - \frac{1}{2}L^a_S(X)\\
&= (X_{T}-a)^+ + \frac{1}{2}L^a_T(X) - \frac{1}{2}L^a_S(X)
\end{align*}
which implies that $L^a_T(X) = L^a_S(X)$. Let $r$ be a rational number and set
\begin{align*}
S(r) &=
\begin{cases}
r & X_{r} < a\\
\infty & X_{r} \geq a
\end{cases}\\
T(r) &= \inf \seq{t>S(r): X_{t} \geq a}\\
H &= \bigcup_{r>0} \,\rrbracket S(r), T(r)\llbracket\,
\end{align*}
For each $\omega \in \Omega$, the section $H(\omega)$ is the interior of the $\seq{t:X_t(\omega)<0}$. We see that the process $L^a(X)$ does not increase in the interior of $\seq{t:X_t(\omega)<0}$ and by similar arguments $L^a(X)$ does not increase on the interior of $\seq{t:X_t(\omega)>0}$.
We conclude by noticing that the set $\seq{t:X_t(\omega)\neq 0}$ differs from its interior by a countable set, since the boundary set is contained in the set of jumps of the optional semimartingale $X$ which for each $\omega \in \Omega$ is countable (see Theorem 1.14 \cite{G1}).
\end{proof}
|
\section{Introduction}
\noindent The study of random growth models of rooted phylogenetic trees and the statistical properties of the shapes of the phylogenetic trees they produce was initiated almost one century ago by Yule \cite{Yule} and it has gained momentum in the last 20 years: see, for instance, \cite{Ald1,BF,Ford1,Egui,Kirk,Po16,SV}. The final goal of this line of research is to understand the relationship between the forces that drive evolution and the topological properties of ``real-life'' phylogenetic trees \cite{BF,MH}; see also \cite[Chap. 33]{fel:04}. One of the most popular such models is Ford's $\alpha$-model for rooted binary trees \cite{Ford1}, a parametric model that generalizes both the uniform model (where new leaves are added equiprobably to any arc, giving rise to the uniform probability distribution on the sets of binary rooted phylogenetic trees, or \emph{cladograms}, with a fixed set of taxa) and Yule's model (where new leaves are added equiprobably only to \emph{pendant} arcs, that is, to arcs ending in leaves) by allowing a different probability for the addition of new leaves to pendant or to {internal} (i.e., non-pendant) arcs.
When models like Ford's are used to contrast topological properties of phylogenetic trees contained in databases like TreeBase (\url{https://treebase.org}), only
their general properties (moments, asymptotic behavior) are employed. But, in the course of a research where we have needed to compute the probabilities of several specific cladograms under this model, we have noticed that the explicit formulas that Ford gives in \cite[\S 3.5]{Ford1} for the probabilities of cladograms and (unlabeled rooted binary) trees (see Props.\ 29 and 32 therein, with the definition of $\widehat{q}$ given in page 30) are wrong, failing for some trees with $n\geq 8$ leaves.
So, to help the future user of Ford's model, in this paper we give the correct explicit formulas for these probabilities. This paper is accompanied by the GitHub page \url{https://github.com/biocom-uib/prob-alpha} where the interested reader can find a SageMath \cite{sage} module to compute these probabilities, and their
explicit values on the sets $\mathcal{T}_n$ of cladograms with $n$ leaves labeled $1,\ldots,n$, for every $n$ from 2 to 8.
\section{Preliminaries}
\subsection{Definitions, notations, and conventions}
\noindent Throughout this paper, by a \emph{tree} $T$ we mean a rooted binary tree. We shall denote its root and its set of \emph{internal nodes} (i.e., those nodes that are not leaves) by $r_T$ and $V_{int}(T)$, respectively.
We understand $T$ as a directed graph, with its arcs pointing away from the root. The \emph{children} of an internal node $u$
are those nodes $v$ such that $(u,v)$ is an arc in $T$, and they form
the set $\mathrm{child}(u)$. A node $x$ is a \emph{descendant} of a node $v$ when there exists a directed path from $v$ to $x$ in $T$. For every node $v$, the \emph{subtree $T_v$ of $T$ rooted at $v$} is the subgraph of $T$ induced on the set of descendants of $v$.
A tree $T$ is \emph{ordered} when it is endowed with an \emph{ordering} $\prec_v$ on every $\mathrm{child}(v)$.
An \emph{isomorphism} of ordered trees is an isomorphism of rooted trees that moreover preserves these orderings.
A \emph{cladogram} (respectively, an \emph{ordered cladogram}) on a set $\Sigma$ is a tree (resp., an {ordered tree}) with its leaves bijectively labeled in $\Sigma$. An \emph{isomorphism} of cladograms (resp., of ordered cladograms) is an isomorphism of trees (resp., of ordered trees) that preserves the leaves' labels. We shall always identify a tree, an ordered tree, a cladogram, or an ordered cladogram, with its isomorphism class, and in particular we shall make henceforth the abuse of language of saying that two of these objects, $T,T'$, \emph{are the same}, in symbols $T=T'$, when they are (only) isomorphic.
We shall denote trees and cladograms by means of their Newick format \cite{Newick}. In this representation of unlabeled trees, we shall denote the leaves with a symbol $*$.
Let $\mathcal{T}^*_n$, $\mathcal{OT}^*_n$, $\mathcal{T}_n$, and $\mathcal{OT}_n$, respectively, stand for the sets of trees with $n$ leaves, of ordered trees with $n$ leaves, of cladograms on $[n]=\{1,2,\ldots,n\}$, and of ordered cladograms on $[n]$.
There exist natural isomorphism-preserving forgetful mappings
\begin{center}
\begin{tikzpicture}[thick,>=stealth,scale=0.25]
\draw(0,0) node (lbt) {$\mathcal{OT}_n$};
\draw(4,-4) node (bt) {$\mathcal{T}_n$};
\draw(-4,-4) node (lbtp) {$\mathcal{OT}^*_n$};
\draw(0,-8) node (btp) {$\mathcal{T}^*_n$};
\draw[->] (lbt)--(bt);
\draw(3,-1.8) node {$\pi_o$};
\draw(-3,-1.8) node {$\pi_*$};
\draw(3,-6.2) node {$\pi$};
\draw(-3,-6.2) node {$\pi_{o,*}$};
\draw[->] (lbt)--(lbtp);
\draw[->] (bt)--(btp);
\draw[->] (lbtp)--(btp);
\end{tikzpicture}
\end{center}
that ``forget'' the orderings or the labels of the trees. In particular, we shall say that two cladograms \emph{have the same shape} when they have the same image under $\pi$.
Let us introduce some more notations. For every node $v$ in a tree $T$, $\kappa_T(v)$ is its number of descendant leaves.
For every internal node $v$ in an ordered tree $T$, with children $v_1\prec_v v_2$, its \emph{numerical split} is
the ordered pair
$NS_T(v)=(\kappa_T(v_1), \kappa_T(v_2))$. If, instead, $T$ is unordered and if $\mathrm{child}(v)=\{v_1,v_2\}$ with
$\kappa_T(v_1)\leq \kappa_T(v_2)$, then $NS_T(v)=(\kappa_T(v_1), \kappa_T(v_2))$.
In both cases, the \emph{multiset of numerical splits} of $T$ is
$NS(T)=\{NS_T(v)\mid v\in V_{int}(T)\}$.
A \emph{symmetric branch point} in a tree $T$ is an internal node $v$ such that, if $v_1$ and $v_2$ are its children, then
the subtrees $T_{v_1}$ and $T_{v_2}$ of $T$ rooted at them are isomorphic.
Given two cladograms $T$ and $T'$ on $\Sigma$ and $\Sigma'$, respectively, with $\Sigma\cap\Sigma'=\emptyset$, their \emph{root join} is the cladogram $T\star T'$ on $\Sigma\cup\Sigma'$ obtained by connecting the roots of $T$ and $T'$ to a (new) common root $r$.
If $T,T'$ are ordered cladograms, $T\star T'$ is ordered by inheriting the orderings on $T$ and $T'$ and ordering the children of the new root $r$ as $r_T\prec_r r_{T'}$. If $T$ and $T'$ are unlabeled trees, a similar construction yields a tree $T\star T'$; if they are moreover ordered, then $T\star T'$ becomes an ordered tree as explained above.
\subsection{The $\alpha$-model}
\noindent Ford's $\alpha$-model \cite{Ford1} defines, for every $n\geq 1$, a family of probability density functions $P^{(*)}_{\alpha,n}$ on $\mathcal{T}^*_n$ that depend on one parameter $\alpha\in [0,1]$, and then it translates this family into three other families of probability density functions
$P_{\alpha,n}$ on $\mathcal{T}_n$, $P^{(o,*)}_{\alpha,n}$ on $\mathcal{OT}^*_n$, and $P^{(o)}_{\alpha,n}$ on $\mathcal{OT}_n$, by imposing that the probability of a tree is equally distributed among its preimages under $\pi$, $\pi_{o,*}$, or $\pi\circ\pi_o=\pi_{o,*}\circ \pi_*$, respectively.
It is well known \cite{CS} that every $T\in \mathcal{T}_n$ can be obtained in a unique way by adding recurrently to a single node labeled 1, new leaves labeled $2,\ldots,n$ to arcs (i.e., splitting an arc $(u,v)$ into two arcs $(u,w)$ and $(w,v)$ and then adding a new arc from the inserted node $w$ to a new leaf) or to a new root (i.e., adding a new root $w$ and new arcs from $w$ to the old root and to a new leaf).
The value of $P^{(*)}_{\alpha,n}(T^*)$ for $T^*\in \mathcal{T}_n^*$ is determined through all possible ways of constructing dendrograms with shape $T^*$ in this way. More specifically:
\begin{enumerate}
\item Start with the tree $T_1\in \mathcal{T}_1$ consisting of a single node labeled 1. Let $P'_{\alpha,1}(T_1)=1$.
\item For every $m=2,\ldots,n$, let $T_{m}\in \mathcal{T}_{m}$ be obtained by adding a new leaf labeled $m$ to $T_{m-1}$. Then:
$$
\hspace*{-1ex}P'_{\alpha,m}(T_{m})=\left\{\hspace*{-1ex}\begin{array}{ll}
\dfrac{\alpha}{m-1-\alpha}\cdot P'_{\alpha,m-1}(T_{m-1}) & \mbox{if the new leaf is added to}\\[-1ex] &\mbox{an internal arc or to a new root}\\[2ex]
\dfrac{1-\alpha}{m-1-\alpha}\cdot P'_{\alpha,m-1}(T_{m-1}) & \mbox{if the new leaf is added to}\\[-1ex] &\mbox{a pendant arc}
\end{array}\right.
$$
\item When the desired number $n$ of leaves is reached, the probability of every tree $T_n^*\in \mathcal{T}_n^*$ is defined as
$$
P^{(*)}_{\alpha,n}(T_n^*)=\sum_{\pi(T_n)=T_n^*} P'_{\alpha,n}(T_n).
$$
\end{enumerate}
Once $P^{(*)}_{\alpha,n}$ is defined on $\mathcal{T}^*_n$, it is transported to $\mathcal{T}_n$, $\mathcal{OT}_n^*$, and $\mathcal{OT}_n$ by defining the probability of an object in one of these sets as the probability of its image in $\mathcal{T}^*_n$ divided by the number of preimages of this image:
\begin{itemize}
\item For every $T\in \mathcal{T}_n$, if $\pi(T)=T^*\in \mathcal{T}^*_n$ and it has $k$ symmetric branch points, then
\begin{equation}
P_{\alpha,n}(T)=\frac{2^k}{n!}\cdot P^{(*)}_{\alpha,n}(T^*),
\label{s->d}
\end{equation}
because $|\pi^{-1}(T^*)|=n!/2^k$ (see, for instance, \cite[Lem. 31]{Ford1}).
\item For every $T_o\in \mathcal{OT}_n$, if $\pi_o(T_o)=T\in \mathcal{T}_n$, then
\begin{equation}
P^{(o)}_{\alpha,n}(T_o)=\frac{1}{2^{n-1}}\cdot P_{\alpha,n}(T),
\label{od->d}
\end{equation}
because $|\pi_o^{-1}(T)|=2^{n-1}$ ($T$ has $2^{n-1}$ different preimages under $\pi_o$, obtained by taking all possible different combinations of orderings on the $n-1$ sets $\mathrm{child}(v)$, $v\in V_{int}(T^*)$).
\item For every $T_o^*\in \mathcal{OT}^*_n$, if $\pi_{o,*}(T^*_o)=T^*\in \mathcal{T}^*_n$ and it has $k$ symmetric branch points, then
\begin{equation}
P^{(o,*)}_{\alpha,n}(T^*_o)=\frac{1}{2^{n-k-1}}\cdot P^{(*)}_{\alpha,n}(T^*),
\label{s->os}
\end{equation}
because $|\pi_{o,*}^{-1}(T^*)|=2^{n-1-k}$ (from the $2^{n-1}$ possible preimages of $T^*$ under $\pi_{o,*}$, defined by all possible different combinations of orderings on the $n-1$ sets $\mathrm{child}(v)$, $v\in V_{int}(T^*)$, those differing only on the orderings on the children of the $k$ symmetric branch points are actually the same ordered tree).
\end{itemize}
The family $(P^{(o,*)}_{\alpha,n})_{n}$ satisfies the following useful Markov branching recurrence (in the sense of \cite[\S 4]{Ald1}).
\begin{proposition}\label{prop:1}
Let $\Gamma_\alpha:\mathbb{Z}^+\to \mathbb{R}$ be the mapping defined by $\Gamma_\alpha(1)=1$ and, for every $n\geq 2$,
$\Gamma_{\alpha}(n)=(n-1-\alpha)\cdot\Gamma_{\alpha}(n-1)$.
For every $a,b\in\mathbb{Z}^+$, let
$$
q_{\alpha}(a,b)=\frac{\Gamma_\alpha(a)\Gamma_\alpha(b)}{\Gamma_\alpha(a+b)}\cdot \varphi_{\alpha}(a,b),
$$
where
$$
\varphi_{\alpha}(a,b)=\frac{\alpha}{2}\binom{a+b}{a}+(1-2\alpha)\binom{a+b-2}{a-1}.
$$
Then, for every $0<m<n$ and for every $T_m^* \in \mathcal{OT}_m^*$ and $T_{n-m}^*\in \mathcal{OT}_{n-m}^*$,
$$
P^{(o,*)}_{\alpha,n}(T_n^*\star T_{n-m}^*)=q_{\alpha}(m,n-m)P^{(o,*)}_{\alpha,m}(T_m^*)P^{(o,*)}_{\alpha,n-m}(T_{n-m}^*).
$$
\end{proposition}
This recurrence, together with the fact that $P^{(o,*)}_{\alpha,1}$ of a single node is $1$, implies that, for every $T_o^*\in \mathcal{OT}_n^*$,
\begin{equation}
P^{(o,*)}_{\alpha,n}(T_o^*)=\prod_{(a,b)\in NS(T_o^*)} q_{\alpha}(a,b).
\label{explLBT}
\end{equation}
For proofs of Proposition \ref{prop:1} and equation (\ref{explLBT}), see Lemma 27 and Proposition 28 in \cite{Ford1}, respectively.
\section{Main results}
\noindent Our first result is an explicit formula for $P_{\alpha,n}(T)$, for every $n\geq 1$ and $T\in \mathcal{T}_n$:
\begin{proposition}\label{prop:main}
For every $T\in \mathcal{T}_n$, its probability under the $\alpha$-model is
$$
P_{\alpha,n}(T)=\frac{2^{n-1}}{n!\cdot \Gamma_{\alpha}(n)}\prod_{(a,b)\in NS(T)} \varphi_{\alpha}(a,b).
$$
\end{proposition}
\begin{proof}
Given $T\in \mathcal{T}_n$, let $T_o$ be any ordered cladogram such that $\pi_o(T_o)=T$, and let $T_o^*=\pi_{*}(T_o)\in \mathcal{OT}_n^*$ and $T^*=\pi(T)=\pi_{o,*}(T_o^*)$. If $T^*$ has $k$ symmetric branch points, then, by equations (\ref{s->d}), (\ref{s->os}) and (\ref{explLBT}),
$$
P_{\alpha,n}(T)=\frac{2^k}{n!}\cdot P^{(*)}_{\alpha,n}(T^*)=\frac{2^k}{n!}\cdot 2^{n-k-1}\cdot P^{(o,*)}_{\alpha,n}(T_o^*)
=\frac{2^{n-1}}{n!} \prod_{(a,b)\in NS(T_o^*)} q_{\alpha}(a,b).
$$
Now, on the one hand, it is easy to check that
$NS(T)=\big\{(\min\{a,b\},\max\{a,b\})\mid (a,b)\in NS(T_0^*)\big\}$,
and therefore, since $q_\alpha$ is symmetric,
$$
P_{\alpha,n}(T)=\frac{2^{n-1}}{n!} \prod_{(a,b)\in NS(T)} q_{\alpha}(a,b).
$$
It remains to simplify this product. If, for every $v\in V_{int}(T)$, we denote its children by $v_1$ and $v_2$, then
$$
\prod_{(a,b)\in NS(T)} q_{\alpha}(a,b)=
\prod_{v\in V_{int}(T)}\frac{\Gamma_{\alpha}(\kappa_T(v_1))\Gamma_{\alpha}(\kappa_T(v_2))}{\Gamma_{\alpha}(\kappa_T(v))} \varphi_{\alpha}(NS(v)).
$$
For every $v\in V_{int}(T)\setminus\{r_T\}$, the term $\Gamma_{\alpha}(\kappa_T(v))$ appears twice in this product: in the denominator of the factor corresponding to $v$ itself and in the numerator of the factor corresponding to its parent. Therefore, all terms $\Gamma_\alpha(\kappa_T(v))$ in this product cancel except $\Gamma_{\alpha}(\kappa_T(r_T))=\Gamma_\alpha(n)$ (that appears in the denominator of its factor) and every $\Gamma_\alpha(\kappa_T(v))=\Gamma_\alpha(1)=1$ with $v$ a leaf. Thus,
$$
P_{\alpha,n}(T)=\frac{2^{n-1}}{n!}\cdot \frac{1}{\Gamma_{\alpha}(n)} \cdot
\prod_{v\in V_{int}(T)}\varphi_{\alpha}(NS(v))
$$
as we claimed.\qed
\end{proof}
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}[thick,>=stealth,scale=0.3]
\draw(-2,0) node [tre] (0) {}; \draw (0) node {\tiny $1$};
\draw(0,0) node [tre] (1) {}; \draw (1) node {\tiny $2$};
\draw(2,0) node [tre] (2) {}; \draw (2) node {\tiny $3$};
\draw(4,0) node [tre] (3) {}; \draw (3) node {\tiny $4$};
\draw(6,0) node [tre] (4) {}; \draw (4) node {\tiny $5$};
\draw(8,0) node [tre] (5) {}; \draw (5) node {\tiny $6$};
\draw(10,0) node [tre] (6) {}; \draw (6) node {\tiny $7$};
\draw(12,0) node [tre] (7) {}; \draw (7) node {\tiny $8$};
\draw(11,1) node[tre] (a) {};
\draw(7,1) node[tre] (b) {};
\draw(9,3) node[tre] (c) {};
\draw(1,3) node[tre] (d) {};
\draw(0,2) node[tre] (e) {};
\draw(-1,1) node[tre] (f) {};
\draw(5,7) node[tre] (r) {};
\draw (r)--(d);
\draw (d)--(e);
\draw (d)--(3);
\draw (e)--(f);
\draw (e)--(2);
\draw (f)--(0);
\draw (f)--(1);
\draw (r)--(c);
\draw (c)--(a);
\draw (c)--(b);
\draw (b)--(4);
\draw (b)--(5);
\draw (a)--(6);
\draw (a)--(7);
\end{tikzpicture}
\caption{The cladogram with Newick string $((1,(2,(3,4))),((5,6),(7,8)));$. }
\label{tree1}
\end{center}
\end{figure}
\begin{remark}
Ford states (see \cite[Prop. 32 and page 30]{Ford1}) that if $T\in \mathcal{T}_n$, then
$$
P_{\alpha,n}(T)=\frac{2^k}{n!}\prod_{(a,b)\in NS(T)} \widehat{q}_{\alpha}(a,b)
$$
where $k$ is the number of symmetric branching points in $T$ and
$$
\widehat{q}_{\alpha}(a,b)=\left\{\begin{array}{ll}
2q_{\alpha}(a,b) & \mbox{if $a\neq b$}\\
q_{\alpha}(a,b) & \mbox{if $a=b$}
\end{array}\right.
$$
If we simplify $\prod_{(a,b)\in NS(T)} \widehat{q}_{\alpha}(a,b)$ as in the proof of Proposition \ref{prop:main}, this formula for $P_{\alpha,n}(T)$ becomes
\begin{equation}
P_{\alpha,n}(T)=\frac{2^{k+m}}{n! \cdot\Gamma_{\alpha}(n)} \cdot\prod_{(a,b)\in NS(T)} \varphi_{\alpha}(a,b)
\label{eq:ford}
\end{equation}
where
$m$ is the number of internal nodes whose children have different numbers of descendant leaves.
This formula does not agree with the one given in Proposition \ref{prop:main} above, because
$$
\begin{array}{rl}
k+m=n-1-\big|\{v\in V_{int}(T)\mid &\ \mathrm{child}(v)=\{v_1,v_2\}\mbox{ and } \kappa_T(v_1)= \kappa_T(v_2)\\
& \mbox{ but }
\pi(T_{v_1})\neq \pi(T_{v_2})\}\big|
\end{array}
$$
and, hence, it may happen that $k+m<n-1$. The first example of a cladogram with this property (and the only one, up to relabelings, with at most 8 leaves) is the tree $\widetilde{T}\in \mathcal{T}_8$ with Newick string $((1,(2,(3,4))),((5,6),(7,8)));$ depicted in Fig. \ref{tree1}. For this tree, our formula gives (see (8.22) in the document \url{https://github.com/biocom-uib/prob-alpha/blob/master/ProblsAlpha.pdf})
$$
P_{\alpha,8}(\widetilde{T})=\frac{(1-\alpha)^2(2-\alpha)}{126(7-\alpha)(6-\alpha)(5-\alpha)(3-\alpha)}
$$
while expression (\ref{eq:ford}) assigns to $\widetilde{T}$ a probability of half this value:
\begin{equation}
\frac{(1-\alpha)^2(2-\alpha)}{252(7-\alpha)(6-\alpha)(5-\alpha)(3-\alpha)}.
\label{wrong}
\end{equation}
This last figure cannot be right, for several reasons.
Firstly, by \cite[\S 3.12]{Ford1}, when $\alpha=\frac{1}{2}$, Ford's model is equivalent to the uniform model, where every dendrogram in $\mathcal{T}_n$ has the same probability
has the same probability
$$
\frac{1}{|BT_n|}=\frac{1}{(2n-3)!!}
$$
and when $\alpha=0$, Ford's model gives rise to the Yule model \cite{Harding71,Yule}, where the probability of every $T\in\mathcal{T}_n$ is
$$
P_Y(T)=\frac{2^{n-1}}{n!}\prod_{v\in V_{int}(T)}\frac{1}{\kappa_T(v)-1}.
$$
In particular, $P_{\frac{1}{2},8}(\widetilde{T})$ should be equal to $1/135135$ and $P_{0,8}(\widetilde{T})$ should be equal to $1/19845$.
Both values are consistent with our formula, while expression (\ref{wrong}) yields half these values.
As a second reason, which can checked using a symbolic computation program, let us mention that if we take expression (\ref{wrong}) as the probability of $\widetilde{T}$ and we assign to all other dendrograms in $\mathcal{T}_8$ the probabilities computed with Proposition \ref{prop:main}, which agree on them with the values given by equation (\ref{eq:ford}) (they are given in the aforementioned supplementary file), these probabilities do not add up 1.
\end{remark}
Now, the family of density functions $(P_{\alpha,n})_n$ satisfies the following Markov branching recurrence.
\begin{corollary}\label{cor:sm}
For every $0<m<n$ and for every $T_m\in \mathcal{T}_m$ and $T_{n-m}\in \mathcal{T}_{n-m}$,
$$
P_{\alpha,n}(T_m\star T_{n-m})=\frac{2q_{\alpha}(m,n-m)}{\binom{n}{m}}P_{\alpha,m}(T_m)P_{\alpha,n-m}(T_{n-m}).
$$
\end{corollary}
\begin{proof}
If $T_m\in \mathcal{T}_m$ and $T_{n-m}\in \mathcal{T}_{n-m}$, then
$$
\begin{array}{l}
\displaystyle P_{\alpha,m}(T_m)=\frac{2^{m-1}}{m!\Gamma_\alpha(m)}\prod_{(a,b)\in NS(T_m)}\varphi_{\alpha}(a,b)\\[1ex]
\displaystyle P_{\alpha,n-m}(T_{n-m})=\frac{2^{n-m-1}}{(n-m)!\Gamma_\alpha(n-m)}\prod_{(a,b)\in NS(T_{n-m})}\varphi_{\alpha}(a,b)
\end{array}
$$
and
$$
\begin{array}{l}
P_{\alpha,n}(T_m\star T_{n-m}) \displaystyle =\frac{2^{n-1}}{n!\Gamma_\alpha(n)}\prod_{(a,b)\in NS(T_m\star T_{n-m})}\varphi_{\alpha}(a,b)\\
\qquad \displaystyle =\frac{2^{n-1}}{n!\Gamma_\alpha(n)}\varphi_{\alpha}(m,n-m)\Big(\prod_{(a,b)\in NS(T_m)}\varphi_{\alpha}(a,b)\Big)\Big(\prod_{(a,b)\in NS(T_{n-m})}\varphi_{\alpha}(a,b)\Big)\\
\qquad \displaystyle =\frac{2^{n-1}}{n!\Gamma_\alpha(n)}\varphi_{\alpha}(m,n-m)\frac{m!\Gamma_\alpha(m)}{2^{m-1}}P_{\alpha,m}(T_m)\frac{(n-m)!\Gamma_\alpha(n-m)}{2^{n-m-1}}P_{\alpha,n-m}(T_{n-m})
\\
\qquad \displaystyle = \frac{2q_{\alpha}(m,n-m)}{\binom{n}{m}}P_{\alpha,m}(T_m)P_{\alpha,n-m}(T_{n-m})
\end{array}
$$
as we claimed. \qed\end{proof}
Combining Proposition \ref{prop:main} and equation (\ref{s->d}) we obtain the following result:
\begin{corollary}\label{cor:shape}
For every $T^*\in \mathcal{T}_n^*$ with $k$ symmetric branch points,
$$
P^{(*)}_{\alpha,n}(T^*)=\frac{2^{n-k-1}}{\Gamma_{\alpha}(n)}\prod_{(a,b)\in NS(T^*)} \varphi_{\alpha}(a,b).
$$
\end{corollary}
This formula does not agree, either, with the one given in \cite[Prop. 29]{Ford1}: the difference lies again in the same factor of 2 to the power of the number of internal nodes that are not symmetric branch points but whose children have the same number of descendant leaves.
\begin{figure}[htb]
\begin{center}
\begin{tikzpicture}[thick,>=stealth,scale=0.3]
\draw(2,0) node [tre] (2) {};
\draw(4,0) node [tre] (3) {};
\draw(6,0) node [tre] (4) {};
\draw(8,0) node [tre] (5) {};
\draw(10,0) node [tre] (6) {};
\draw(12,0) node [tre] (7) {};
\draw(11,1) node[tre] (a) {};
\draw(10,2) node[tre] (b) {};
\draw(5,1) node[tre] (c) {};
\draw(8,4) node[tre] (d) {};
\draw(7,5) node[tre] (e) {};
\draw (e)--(2);
\draw (e)--(d);
\draw (d)--(b);
\draw (d)--(c);
\draw (c)--(3);
\draw (c)--(4);
\draw (b)--(5);
\draw (b)--(a);
\draw (a)--(6);
\draw (a)--(7);
\end{tikzpicture}\qquad
\begin{tikzpicture}[thick,>=stealth,scale=0.3]
\draw(2,0) node [tre] (2) {};
\draw(4,0) node [tre] (3) {};
\draw(6,0) node [tre] (4) {};
\draw(8,0) node [tre] (5) {};
\draw(10,0) node [tre] (6) {};
\draw(12,0) node [tre] (7) {};
\draw(11,1) node[tre] (a) {};
\draw(10,2) node[tre] (b) {};
\draw(9,3) node[tre] (c) {};
\draw(3,1) node[tre] (d) {};
\draw(7,5) node[tre] (e) {};
\draw (e)--(c);
\draw (e)--(d);
\draw (d)--(2);
\draw (d)--(3);
\draw (c)--(b);
\draw (c)--(4);
\draw (b)--(5);
\draw (b)--(a);
\draw (a)--(6);
\draw (a)--(7);
\end{tikzpicture}
\caption{The trees with Newick strings $(*,((*,*),(*,(*,*))));$ (left) and $((*,*),(*,(*,(*,*))));$ (right). }
\label{tree2}
\end{center}
\end{figure}
\begin{remark}
Against what is stated in \cite{Ford1}, $P^{(*)}_{\alpha,n}$ does not satisfy any Markov branching recurrence. Indeed, let $T_m,T_m'\in \mathcal{T}^*_m$ be any two different trees with the same numbers $m$ of leaves and $k$ of symmetric branch points:
for instance, the trees in $\mathcal{T}^*_6$ with Newick strings $(*,((*,*),(*,(*,*))));$ and $((*,*),(*,(*,(*,*))));$ depicted in Fig. \ref{tree2}.
Then,
$$
\begin{array}{l}
\displaystyle P_{\alpha,m}^{(*)}(T_m)=\frac{2^{m-k-1}}{\Gamma_{\alpha}(m)}\prod_{(a,b)\in NS(T_m)}\varphi_{\alpha}(a,b)\\[2ex]
\displaystyle P_{\alpha,m}^{(*)}(T'_m)=\frac{2^{m-k-1}}{\Gamma_{\alpha}(m)}\prod_{(a,b)\in NS(T'_m)}\varphi_{\alpha}(a,b).
\end{array}
$$
In this case, $T_m\star T_m\in \mathcal{T}_{2m}^*$ has $2k+1$ symmetric branch points and therefore
$$
\begin{array}{l}
\displaystyle P_{\alpha,m}^{(*)}(T_m\star T_m) =\frac{2^{2m-2k-2}}{\Gamma_{\alpha}(2m)}\prod_{(a,b)\in NS(T_m\star T_m)} \varphi_{\alpha}(a,b)\\
\qquad\qquad \displaystyle=\frac{2^{2m-2k-2}}{\Gamma_{\alpha}(2m)}\varphi_{\alpha}(m,m)\Big(\prod_{(a,b)\in NS(T_m)} \varphi_{\alpha}(a,b)\Big)^2\\
\qquad\qquad \displaystyle= \frac{2^{2m-2k-2}}{\Gamma_{\alpha}(2m)}\varphi_{\alpha}(m,m)\Big(\frac{\Gamma_{\alpha}(m)}{2^{m-k-1}}P_{\alpha,m}^{(*)}(T_m)\Big)^2\\
\qquad\qquad \displaystyle= q_{\alpha}(m,m)P_{\alpha,m}^{(*)}(T_m)P_{\alpha,m}^{(*)}(T_m)
\end{array}
$$
while $T_m\star T'_m\in \mathcal{T}_{2m}^*$ has $2k$ symmetric branch points and therefore
$$
\begin{array}{l}
\displaystyle P_{\alpha,m}^{(*)}(T_m\star T'_m) =\frac{2^{2m-2k-1}}{\Gamma_{\alpha}(2m)}\prod_{(a,b)\in NS(T_m\star T'_m)} \varphi_{\alpha}(a,b)\\
\qquad \displaystyle=\frac{2^{2m-2k-1}}{\Gamma_{\alpha}(2m)}\varphi_{\alpha}(m,m)\Big(\prod_{(a,b)\in NS(T_m)} \varphi_{\alpha}(a,b)\Big)\Big(\prod_{(a,b)\in NS(T'_m)} \varphi_{\alpha}(a,b)\Big)\\
\qquad \displaystyle= \frac{2^{2m-2k-1}}{\Gamma_{\alpha}(2m)}\varphi_{\alpha}(m,m)\cdot \frac{\Gamma_{\alpha}(m)}{2^{m-k-1}}P_{\alpha,m}^{(*)}(T_m)\cdot \frac{\Gamma_{\alpha}(m)}{2^{m-k-1}}P_{\alpha,m}^{(*)}(T'_m)\\
\qquad \displaystyle= 2q_{\alpha}(m,m)P_{\alpha,m}^{(*)}(T_m)P_{\alpha,m}^{(*)}(T'_m)
\end{array}
$$
and $q_{\alpha}(m,m)\neq 2q_{\alpha}(m,m)$.
\end{remark}
\noindent\textbf{Acknowledgments}.
This research was supported by Spanish Ministry of Economy and Competitiveness and European Regional Development Fund project DPI2015-67082-P (MINECO/FEDER). We thank G. Cardona and G. Riera for several comments on the SageMath module companion to the paper.
|
\section{Introduction}
In this paper we study $H_i(\bMgn)$, the homology of the Deligne--Mumford moduli space of stable marked curves, from the point of view of representation stability. The space $\bMgn$ is a natural compactification of the moduli space of smooth curves with $n$ marked points, obtained by allowing families of smooth curves to degenerate to singular curves with double points. The symmetric group $\bS_n$ acts on $\bMgn$ by relabeling the marked points, so that if we fix $i$ and $g$ we obtain a sequence of symmetric group representations $n \mapsto H_i(\bMgn, \bQ)$.
Our goal is to understand the asymptotic behavior of these $\bS_n$ representations for $n \gg 0$.
The following theorem gives applications of our main result.
\begin{thm}\label{hilbert}
Let $i,g \in \bN$. There is a constant $C = p(g,i)$, where $p$ is a polynomial in $g$ and $i$ of order $O(g^2i^2)$, such that the following hold.
\begin{enumerate}
\item The generating function for the dimension of $H_i(\bMgn)$ is rational and takes the form $$\sum_n \dim H_i(\bMgn) t^n = \frac{ f(t)}{\prod_{j = 1}^{C} (1 - j t)^{d_j} }$$
for some polynomial $f(t)$ and $d_j \in \bN$.
In particular, there exist polynomials $f_1(n), \dots, f_{C}(n)$ such that for $n \gg 0$ we have $$ \dim H_i(\bMgn) = \sum_{j=1}^C f_{j}(n) j^n$$
\item Let $\lambda$ be an integer partition of $n$. If the irreducible $\bS_n$ representation $M_\lambda$ occurs in the decomposition of $H_i(\bMgn, \bQ)$, then $\lambda$ has length $\leq C$. (The Young diagram of $\lambda$ has $\leq C$ rows).
\item Let $\lambda = \lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_C $ be an integer partition of $k$, and $\lambda + n$ be the partition $\lambda_1 + n \geq \lambda_2 \geq \dots \geq \lambda_C$. The multiplicity of $\lambda + n$ in $H_i(\widebar \cM_{g, n+k})$, $$n \mapsto \dim \Hom_{\bS_{n+k}}(M_{\lambda + n}, H_i(\widebar \cM_{g, n + k})),$$ is bounded by a polynomial of degree $C - 1$.
\end{enumerate}
\end{thm}
To establish Theorem \ref{hilbert}
we use techniques from the field of representation stability. We extend the action of the symmetric groups to the action of a category, and we prove that the homology groups are finitely generated under this action. Finite generation then constrains the behavior of $H_i(\bMgn)$ for $n \gg 0$.
\subsection{Main Result}\label{actiondef}
Let $\FSop$ be the opposite of the category of finite sets and surjections. We give $n \mapsto H_\bdot(\bar \cM_{g,n})$ the structure of an $\FSop$ module.
More concretely, let $[n]$ be the set $\{1, \dots, n\}$. For every surjection $f: [n] \to [m]$, we define a map $$f^*: H_\bdot(\bMgn) \leftarrow H_\bdot(\bar \cM_{g,m}),$$ such that $(f \circ g)^* =g^* f^*$. We describe $f^*$ in two special cases, which suffice to determine it in general. In these cases, $f^*$ is the map on homology induced by a map of spaces, $F^* : \bMgn \leftarrow \bar \cM_{g,m}$.
When $f$ is a bijection, $F^*$ is the map that takes a stable marked curve $C$ and permutes its marked points by precomposing with $f$.
When $f: [n+1] \to [n]$ is the surjection defined by $f(n+1) = n$ and $f(i) = i$ otherwise, $F^*$ is the map that glues a copy of $\bP^1$ to the point $n \in C$ and marks $ \bP^1 \sqcup_{\{n\}} C $ by keeping the marked points $\{1, \dots, n-1\} \subset C$ and marking two new points $\{n, n+1\} \subset (\bP^1 \sqcup_{\{n\}} C) - C$.
Our main theorem states that this $\FSop$ module is finitely generated.
\begin{thm}\label{Mgnbar}
Let $g,i \in \bN$. Then the $\FS \op$ module $$n \mapsto H_i(\bMgn, \bQ)$$ is a subquotient of an $\FS \op$ module that is finitely generated in degree $\leq p(g,i)$ where $p(g,i)$ is a polynomial in $g$ and $i$ of order $O(g^2 i^2)$.
\end{thm}
Theorem \ref{hilbert} is implied by Theorem \ref{Mgnbar} and results on finitely generated $\FSop$ modules due to Sam and Snowden \cite{sam2017grobner}. The polynomial of Theorem \ref{Mgnbar} is the same as the polynomial of Theorem \ref{hilbert}.
\begin{rmk}
The category $\FSop$ acts on the homology of $\bMgn$ through maps that glue on copies of $\bP^1$ with two marked points. These maps are a very small part of the full operadic structure on $H_\bdot(\bMgn)$ generated by all gluing maps. The \emph{tautological ring} is the subring of $H^\bdot(\bMgn)$ generated by the image of all of the fundamental classes $[\bMgn]$ under gluing maps and cup products.
In some sense, Theorem \ref{Mgnbar} says that for $i,g$ fixed, all of the classes in $H_i(\bMgn)$ are tautological ``relative'' to a finite list of classes, using only maps that glue on copies of $\bP^1$ with $2$ marked points.
\end{rmk}
\subsection{Stability}
Although the dimensions $\dim H_i(\bMgn)$ grow exponentially in $n$, and therefore do not stabilize in the naive sense, Theorem \ref{Mgnbar} implies that there exists a constant $N$ such that the $\bS_n$ representations $H_{i}(\bMgn)$ are completely determined by the vector spaces $\{H_i(\mathcal \bar\cM_{g,m})\}_{m \leq N}$ and the algebraic structure they inherit from surjections $[m] \onto [m']$.
For $r \in \bN$, let $\FS_r \op$ be the full subcategory of $\FSop$ spanned by sets of size $\leq r$. We may restrict an $\FS \op$ module $M$ to an $\FS_r \op$ module, denoted $\Res_r M$. The functor $\Res_r$ has a left adjoint $\Ind_r$, which takes an $\FS_r \op$ module to the $\FS \op$ module freely generated by it modulo relations from $\FS_r \op$.
\begin{thm}\label{stability}
Let $i, g \in \bN$. There exists $N \in \bN$ such that the natural map of $\FSop$ modules $$\Ind_N \Res_N H_i(\bar \cM_{g,-}) \to H_i(\bar \cM_{g,-})$$ is an isomorphism. In particular, any presentation of the $\FS_N \op$ module $\Res_N H_i(\bar \cM_{g,-}) $ gives a presentation of the $\FSop$ module $ H_i(\bar \cM_{g,-})$.
\end{thm}
Since $\FS_N$ is a finite category and $\Ind_N$ can be described explicitly as a colimit, Theorem \ref{stability} gives a procedure to compute $H_i(\bar \cM_{g,n})$ from a finite amount of algebraic data.
Theorem \ref{stability} follows from Theorem \ref{Mgnbar} and a Noetherianity result due to Sam and Snowden \cite{sam2017grobner}.
\subsection{Relation to other work}
Our work is motivated by the approach to representation stability introduced by Church, Ellenberg and Farb, which uses modules over $\FI$, the category of finite sets and injections \cite{CEF}.
The theory of $\FI$ modules has been used by Jim\'enez Rolland to study the homology of $\Mgn$ \cite{rolland2013cohomology}, and by Maya Duque and Jim\'enez Rolland to study the real locus of $\bMon$ \cite{rolland2015representation}. Because the homology of $\bMgn$ grows at an exponential rate, it cannot admit the structure of a finitely generated $\FI$ module, and so a larger category is needed to control the homology of the full compactification.
Using an explicit presentation of the cohomology ring $H^\bdot(\bMon)$ given in \cite{etingof2005cohomology}, Sam defined an action of $\FS \op$ on the cohomology of $\bMon$, and proved that it was finitely generated. Our work was motivated by his suggestion that there could exist a finitely generated $\FSop$ action on the cohomology of $\bMgn$ for general $g$.
Sam and Snowden showed that $\FSop$ is noetherian (submodules of finitely generated modules are finitely generated), and described the Hilbert series of finitely generated $\FSop$ modules \cite{sam2017grobner}. We use their results to deduce concrete implications from Theorem \ref{Mgnbar}.
Proudfoot and Young have used $\FSop$ modules to study the intersection cohomology of a space closely related to $\bMon$ \cite{proudfoot2017configuration}. The $\FSop$ module they construct is closely related to our construction in the case $g = 0$. The statement of our Theorem \ref{hilbert} parallels their Theorem 4.3.
In order to produce non-tautological classes, Faber and Pandharipande \cite{faber2011tautological} established restrictions on the $\bS_n$ representations that appear in the tautological ring, which resemble the restrictions on $\bS_n$ representations we obtain in Theorem \ref{hilbert}. Our restrictions on representations are weaker, but they hold for all cohomology classes.
Kapranov--Manin \cite{kapranov2001modules} observed that $\bigoplus_{i,n} H_i(\bMgn)$ is a right module over the hypercommutative opeard. This algebraic structure extends the action of $\FSop$ on $H_i(\bMgn)$ for fixed $i$.
\subsection{Structure of the Paper} In the next section, \S \ref{action}, we describe the action of $\FSop$ on $H_i(\bMgn)$ in more detail. Then we sketch the proof of Theorem \ref{Mgnbar} and describe how the other sections of the paper fit into the proof.
In the final section, \S \ref{furtherquestions}, we ask further questions.
\subsection{Acknowledgements}
We thank Steven Sam for sharing his work on the cohomology of $\bMon$.
We also thank Dan Petersen, John Wiltshire-Gordon and Andrew Snowden for helpful conversations.
\section{The Action of $\FSop$}\label{action}
Let $\FS$ be the category of finite sets and surjections. The objects of $\FS$ are the sets $[n] = \{1, \dots, n\}$ for $n \geq 1$, and the morphisms are surjections. An \emph{$\FSop$ module}, or an \emph{action of $\FSop$} on a sequence of vector spaces $V_n$, is a functor from $\FSop$ to the category of vector spaces $n \mapsto V_n$.
Let $\bMgn$ be the Deligne--Mumford stack of stable genus $g$ curves with $n$ marked points. The stack $\bMgn$ parameterizes genus curves $C$ and markings $p_1, \dots, p_n \in C$, such that all of the singularities of $C$ are double points, each marked point $p_i$ is smooth, each genus $0$ component of $C$ contains at least $3$ marked or singular points, and each genus $1$ component contains at least one marked or singular point.
\begin{rmk} Since our results concern homology with rational coefficients, we may work with either the homology of the coarse moduli space or the homology of the Deligne--Mumford stack. For definiteness, we will work with the moduli stack defined over the complex numbers, but our methods are algebraic and should apply over other algebraically closed fields. \end{rmk}
\subsection{Action of binary trees}
Before passing to homology, the category $\FSop$ does not naturally act on $\bMgn$. Instead, we construct an action of a category of binary trees.
We define $\bB \bT$ to be the category whose objects are natural numbers, and whose morphisms $m \to n$ are rooted binary forests with the leaves labelled by $[m]$ and the roots labelled by $[n]$. (Morphisms $n \to n$ are binary forests such that each tree is a single root). Composition is defined by gluing the roots to the leaves. For a forest $F$, the function that takes each leaf to its root is a surjection $h_F: [n] \to [m]$. The assignment $F \mapsto h_F$ defines a functor $\bB \bT \to \FS$ that realizes $\FS$ as a quotient of $\BT$.
The category $\bB \bT \op$ acts on $\bMgn$ by gluing on trees of marked projective lines. For each labelled binary forest $F \in \bB \bT(m,n)$, there is a corresponding variety $L_F$ which is a disjoint union of trees and projective lines. The variety $L_F$ has a component for each binary tree $T$ of $F$: if $T$ consists only of a root, then the component is a point; otherwise the component is a genus zero stable curve with dual graph isomorphic to $T$.\footnote{See Section \ref{stratification} for definition of dual graph.} The variety $L_F$ is naturally marked by $\{1, \dots, m\}$ and $\{1, \dots n\}$ in a way that is unique up to unique isomorphism.
From a stable curve $C \in \bMgn$ and a labelled rooted forest $F \in \bB \bT(m,n)$, we obtain a new stable curve $F^* C \in \widebar \cM_{g,m}$ by gluing $L_F$ to $C$ along the marked points $\{1, \dots, m\}$, and using the marking of $L_F$ by $\{1, \dots, n\}$ to mark $F^*C$. This defines a gluing map $F^*: \widebar \cM_{g,m} \to \bMgn$ for each $F \in \BT(n,m)$, and these maps define an action of $\BT \op$ on $\bMgn$.
To see that the representation of $\BT \op$, $n \mapsto H_i(\bMgn)$, factors through $\FSop$, let $F_1$ and $F_2$ be two forests inducing the same surjection $h: [n] \to [m]$. There is a family of gluing maps from $\widebar \cM_{g,m}$ to $\bMgn$, $$ \widebar \cM_{g,m} \times \left( \prod_{i \in [m], ~ h \inv(i) > 1} \widebar \cM_{0, \# h \inv(i) + 1 } \right) \to \bMgn.$$ The maps $F_1^*$ and $F_2^*$ correspond to evaluating at the points in the second factor defined by the components of $L_{F_1}$ and $L_{F_2}$ respectively. Since the second factor is connected, $F_1^*$ and $F_2^*$ induce the same map on homology.
\subsection{Sketch of proof of Theorem \ref{Mgnbar}}
We stratify $\bMgn$ by dual graphs $G$. To prove finite generation of $H_i(\bMgn)$, it suffices to show that only finitely many strata $\cM_G$ contribute $\FSop$ generators. The stratum $\cM_G$ is a quotient of a product of moduli spaces $\prod_{v \in G} \cM_{g(v),n(v)}$.
Since the homology of $\bMgn$ is pure, it suffices to consider the pure Borel--Moore homology of the strata: $\rW_{-i} H_i^{\rB\rM}(\cM_G)$. By fibering $\Mgn$ over $\cM_{g,1}$ we show that the Borel--Moore homology of $\Mgn$ vanishes for $n > i +3$, so only strata $\cM_G$ for which $\sum_v {\rm val}(v) - 3 \leq i$ contribute to $H_i(\bMgn)$.
Thus for $G$ ranging over all graphs that contribute to $H_i(\bMgn)$, the number of vertices of $G$ that have valence $>3$ and genus $> 0$ is bounded by a function of $g$ and $i$. So as $n \to \infty$, the number of trivalent genus $0$ vertices of $G$ must increase.
Say that a stable graph $H$ \emph{has an external $\rY$} if it has a genus $0$ trivalent vertex $v$ which is adjacent to two external edges. Since $\FSop$ acts on graphs by gluing on trivalent vertices\footnote{More accurately, $\bB \bT \op$ acts}, if $H$ has an external $\rY$ then the classes from $\cM_H$ are pushed forward from lower degree. If $G$ has two adjacent trivalent genus $0$ vertices $v_1, v_2$ such that each $v_i$ has an external edge, the cross ratio relation shows that the classes from $\cM_G$ are homologous to classes from $\cM_H$, where $H$ has an external $\rY$, and thus are also pushed forward from lower degree.
Therefore, to prove finite generation, it is enough to show that when the number of trivalent genus $0$ vertices of $G$ is large then either $(1)$ $G$ has an external $\rY$, or $(2)$ $G$ has two adjacent trivalent genus $0$ vertices $v,v'$, each with an external edge. Each trivalent vertex with \emph{no} external edges contributes $1/2$ to $-\chi(G)$. The bound $-\chi(G) \leq g-1$ implies that, as the number of trivalent vertices increases, one of the two possibilities must occur.
\subsection{Structure of the proof} Unfortunately, in formalizing the above argument, we encounter the problem that $\FSop$ does not act on the Borel--Moore homology spectral sequence for the stable graph stratification, and the category of binary trees, $\bB \bT \op$, which does act, is not known to be noetherian.
Accordingly, in \S \ref{stratification}, we define a coarsening of the stable graph stratification for which $\FSop$ does act on the associated Borel--Moore homology spectral sequence. In \S \ref{graphcombinatorics}, we prove two combinatorial lemmas about stable graphs. In \S \ref{proof}, we combine the results from previous sections to prove Theorem \ref{Mgnbar}.
\section{Stable graph stratification}\label{stratification}
A \emph{stable graph $G$ of genus $g$, with $n$ external edges} is a connected graph $G$: a subset of the univalent vertices of G are marked as external and are put in bijection with $\{1, \dots, n\}$, and the other vertices $v$ are assigned a genus $g(v) \in \bN$. We call the edges adjacent to external vertices external: every external vertex corresponds to a unique external edge. For this reason we do not refer to external vertices directly: when we say that $v \in G$ is a vertex of $G$, we will always mean an internal vertex. This data is subject to the condition that $\dim H^1(G) + \sum_{v \in G} g(v) = g$, each genus $0$ vertex is at least trivalent, and each genus $1$ vertex is at least $1$-valent. We write $g = g(G)$ and $n = n(G)$, and $n(v)$ for the valence of an vertex.
If $G$ is a stable graph and $G'$ is a quotient graph, such that none of the external edges of $G$ are identified, then $G'$ inherits a stable structure: the external vertices are the images of the external vertices under $\pi: G \to G'$ with the induced labelling, and the genus of a vertex $w$ of $G'$ is $\dim H^1(\pi\inv(G)) + \sum_{v \in \pi\inv (w)} g(v)$. We say that the stable graph $G'$ is a quotient of $G$.
Write $\Stab(g,n)$ for the set of isomorphism classes of stable graphs of genus $g$ with $n$ external edges. Then $\Stab(g,n)$ is naturally a poset: we define $G \geq G'$ if $G$ is a quotient of $G'$. And we say that $G \prec G'$ if $G'$ can be obtained from $G$ by quotienting a single internal edge.
It is standard that the Deligne-Mumford compactification $\bMgn$ admits a stratification by the poset of stable graphs $\Stab(g,n)$: the stratum $\cM_G$ for $G \in \Stab(g,n)$ is the locus of stable curves $C$ such that the dual graph of $C$ is isomorphic to $G$.
Here the \emph{dual graph} of $C$ is the stable graph that has an vertex $v$ for each irreducible component of $C$, an edge between $v$ and $w$ for every a double point joining their irreducible components, and an external edge for each marked point of $C$ adjacent to the vertex of the irreducible component that the point lies on. The genus $g(v)$ of a vertex is the genus of its irreducible component.
We use a coarsening of the standard stratification by $\Stab(g,n)$, which has the property that $\FSop$ still acts on the corresponding Borel--Moore homology spectral sequence, to prove Theorem \ref{Mgnbar}. This stratification is defined in terms of the operation in the next proposition:
\begin{prop}\label{closureprop}
Let $G$ be a stable graph. Then there is a unique stable quotient $\widebar G \geq G$ such that no two distinct genus $0$ vertices of $\widebar G$ are connected by an edge, and $\widebar G \to G$ only identifies edges between genus $0$ vertices.
\end{prop}
\begin{proof}
Consider the subgraph spanned by the genus $0$ vertices of $G$ and the edges between them. Choose a minimal spanning tree $T_i$ for each connected component, and define $H$ to be the quotient of $G$ such that each $T_i$ is identified to a point.
Any other quotient of $G$ satisfying the hypotheses of the proposition must collapse a minimal spanning tree, and all quotients of minimal spanning trees are isomorphic.
\end{proof}
Notice that $\widebar{ \widebar H} = \widebar H$. Warning: It is not true that if $G \leq H$, then $\widebar G \leq \widebar H$.
Let $\rQ(g,n)$ be the set of stable graphs $G \in \Stab(g,n)$ such that that no distinct genus $0$ vertices are connected by an edge. The next proposition gives the stratification of $\bMgn$ by $\rQ(g,n)$ that we use:
\begin{prop}
There is a poset structure on $\rQ(g,n)$ such that the map $H \mapsto \widebar H$ is a surjection of posets $\Stab(g,n) \to \rQ(g,n)$ . This surjection induces a stratification of $\bMgn$, where the stratum corresponding $G \in \rQ(g,n)$ is $S(G) = \bigcup_{H,~ \widebar H = G} \cM_H.$
\end{prop}
\begin{proof}
Let $G, H \in \rQ(g,n)$. We say that $G \leq_\rQ H \iff$ there is sequence of stable graphs $G = G_0, G_1, \dots, G_n = H$ such that either $G_i \prec G_{i+1}$ or $G_i \succ G_{i+1}$ and $G_i$ is obtained from $G_{i+1}$ by collapsing an edge between two genus $0$ vertices.
By definition, $\leq_\rQ$ is reflexive and transitive. To prove antisymmetry, suppose $G \leq_\rQ H$ and $H \leq_\rQ G$, and let $G = G_0, G_1, \dots, G_n = H$ be as sequence exhibiting that $G \leq H$.
For a stable graph $J$, we define $s(J) \in \bN^{\oplus \infty}$ to be the vector whose $0$th entry is the number of edges between genus $0$ vertices and genus $\geq 1$ vertices, and whose $i$th entry for $i \geq 1$ is the number of vertices of genus $i$. Totally order these vectors reverse lexicographically. If $J$ and $J'$ are related by collapsing an edge between genus distinct genus $0$ vertices, then $s(J) = s(J')$. If $J$ and $J'$ are related by collapsing a self edge of a genus $0$ vertex, an edge between two genus $\geq 1$ vertices, or an edge from a genus $\geq 1$ vertex to a genus $0$ vertex, then $s(J) < s(J')$. Thus $s(G_i) \leq s(G_{i+1})$ for all $i$, and we have $s(G) \leq s(H)$ and $s(H) \leq s(G)$.
For every $i$, the graphs $G_i$ and $G_{i+1}$ are related by collapsing edges between distinct genus zero vertices. Otherwise, we would have $s(G_i) < s(G_{i+1})$, contradicting $s(G) =s(H)$. Therefore $\widebar G_i = \widebar G_{i+1}$ for all $i$, and so $G = \bar G = \bar H = H$, establishing antisymmetry.
The map $H \mapsto \widebar H$ is map of posets since $H \leq G$ if and only if $H$ and $G$ are related by a sequence of edge collapses, and $H$, $\widebar H$ and $G, \widebar G$ are related by a sequence of collapses between distinct genus zero vertices. Composing these three sequences proves that $\widebar H \leq_Q \widebar G$.
Finally, the surjection $\Stab(g,n) \to \rQ(g,n)$ shows that $$Z(G) := \bigcup_{H \in \rQ(g,n),~ J \in \Stab(g,n)~|~ H \leq_\rQ G ,~\widebar J = H} \cM_J$$ is a closed subset of $\bMgn$ and $$S(G) := Z(G) - \left( \bigcup_{H \in \rQ(g,n),~ H < G} Z(H) \right) = \bigcup_{\widebar J = G} \cM_G. $$
\end{proof}
\begin{defn} For $G \in \rQ(g,n)$ define $\widetilde S(G)$ to be the space
\[ \widetilde S(G) := \prod_{v \text{ vertex}, ~ g(v) = 0} \widebar \cM_{e(v),n(v), 0} \times \prod_{v \text{ vertex}, ~ g(v) \geq 1} \cM_{g(v), n(v)},\]
where $e(v)$ is the number of self edges of $v$ and $\widebar \cM_{e ,n,0}$ is the moduli space of stable curves $C$ of genus $e$ and $n$ marked points, such that all of the irreducible components of $C$ have genus $0$. \qed
\end{defn}
There is a canonical map $\widetilde S(G) \to S(G)$ given by gluing a product of curves along the points joined by the graph.
The data of the graph $G$ can be encoded as a set of half edges, a partition that records which half edges are adjacent to the same vertex, and an involution that records which half edges are glued together. Let $A$ be the group of permutations of half edges that do not belong to self edges between genus $0$ vertices of $G$. Let $A_G \subset A$ be the subgroup which preserves the partition and involution (i.e. respects the graph structure). The group $A_G$ acts on $\tilde S(G)$ by relabeling.
\begin{prop} \label{quotient}
For $G \in \rQ(g,n)$, the map $\widetilde S(G) \to S(G)$ induces an isomorphism $\widetilde S(G)/A_G \iso S(G)$.
\end{prop}
\begin{proof}
We directly check the map is an isomorphism on $\bC$ points, then in formal neighborhoods by deformation theory, and then show this suffices by covering by the Hilbert scheme.
On $\bC$ points $\widetilde S(G) (\bC)$ is the groupoid of collections of marked curves $\{ C_v\}_{v \in {\rm Vert}(G)}$, such such that if $g(v) \geq 1$ then $C_v$ is smooth of genus $g(v)$ and if $g(v) = 0$, then $C_{v}$ is genus $e(v)$ and each irreducible component of $C_{v}$ has genus $0$. Similarly, $S(G)(\Spec \bC)$ is the groupoid of stable marked curves $C$ whose dual graph $H$ has $\widebar H = G$.
The group $A_G$ acts strictly on the groupoid $\widetilde S(G) (\Spec \bC)$, by taking $\sigma( \{C_v\}) = \{ C_{\sigma v}\}$ and relabeling marked points according to the action of $\sigma$ on half edges. There are canonical isomorphisms $\sigma: {\rm glue}(\{C_v\}) \simto {\rm glue}( \{C_{\sigma v}\})$; any curve in $S(G)(\bC)$ is glued from a curve in $\tilde S(G)(\bC)$; and any isomorphism ${\rm glue}(\{C_v\}) \to {\rm glue}(\{C'_v\})$ factors uniquely as a relabeling of the components of $C_v$ by $\sigma \in A_g$, and isomorphisms between the components $f_v: C_{\sigma v} \to C'_{\sigma v}$. This shows that $S(G)(\bC)$ is the groupoid quotient of $\widetilde S(G)(\bC)$ by $G$.
Next we check that the map is an isomorphism in formal neighborhoods. More precisely, let $\{C_v\}_{v\in G} \in \tilde S(G)( \bC)$ be a point, let $C := \glue(\{C_v\})$ have dual graph $H$, and let $B$ be a local artinan $\bC$ algebra. We claim that $$\tilde S(G)(B) /A_G \times_{\tilde S(G)(\bC) } \{C_v\} \to S(G)(B) \times_{ S(G)(\bC)} C$$ is an equivalence.
From the deformation theory of marked stable curves, we have that the groupoid $ S(G)(B) \times_{ S(G)(\bC)} C$ is naturally equivalent to $\Hom(\Spec B, X)$ where $$X \subset \Ext^1(\Omega_C(\sum_i p_i), \cO_C) $$ is a union of linear subspaces (here $\{p_i\}_{i= 1}^n$ are the marked points of $C$). There is one subspace for each spanning forest of the subgraph $H_0 \subset H$ spanned by genus $0$ vertices. The subspace corresponding to a spanning forest $F$ is $$ \ker \left( \Ext^1(\Omega_C(\sum_i p_i), \cO_C) \to \prod_{e \in \rm {edge}(H - F)} \Ext^1(\Omega_{\hat C_e}, \cO_{\hat C_e})\right) $$ where $\hat C_e$ is the formal neighborhood of the double point $e \in C$ corresponding to $e$.
Similarly we have $\tilde S(G)(B) /A_G \times_{\tilde S(G)(\bC) } \{C_v\} = \tilde S(G)(B) \times_{\tilde S(G)(\bC) } \{C_v\} $ is equivalent to $\Hom(\Spec B , \tilde X)$, where $$\tilde X \subset \prod_{v \in {\rm vert}(G)} \Ext^1( \Omega_{C_v}(\sum_{u_v } q_{u_v}), \cO_{C_v})$$ is a union of linear subspaces (the sum is over half edges $u_v$ adjacent to $v$). Again, $\tilde X$ is a union of subspaces, one for each spanning forest of $H_0$. The subspace corresponding to $F$ is $$ \ker\left( \prod_{v \in {\rm vert}(G) } \Ext^1(\Omega_{C_v}(\sum_{u_v} q_{u_v}), \cO_{C_v}) \to \prod_{e \in {\rm edge}(H - H_0)}\Ext^1(\Omega_{\hat C_e}, \cO_{\hat C_e}) \right).$
From these identifications, and the fact that $$\prod_{v \in {\rm vert}(G)} \Ext^1(\Omega_{C_v}(\sum_{u_v} q_{u_v}), \cO_{C_v}) = \ker\left(\Ext^1(\Omega_C(\sum_i p_i), \cO_C) \to \prod_{e \in {\rm edge}(G - H_0)} \Ext^1(\Omega_{\hat C_e}, \cO_{\hat C_e})\right)$$ it follows that the map is an isomorphism on formal neighborhoods.
Now, let $\cH$ be the Hilbert scheme of tricanonically embedded stable curves of genus $g(G)$ with $n(G)$ marked points. Let $\cH(G)$ be the stratum consisting of curves with dual graph $J$ such that $\bar J = G$. Then $\cH(G)$ is a finite type scheme which surjects onto $S(G)$, and the fibre product $\tilde \cH(G) := \cH(G) \times_{S(G)} \tilde S(G)/A_G$ is a locally closed subscheme of a product of hilbert schemes, hence also of finite type. Base changing, we see that $f: \tilde \cH(G)(\bC) \to \cH(G)(\bC)$ is an equivalence and $$\tilde \cH(G)(B) \times_{\tilde \cH(G)(\bC) }x \to \cH(G)(B) \times_{ \cH(G)(\bC)} f(x)$$ is an equivalence for all $x \in \tilde\cH(G)(\bC)$ and $B$ as above. A map between finite type schemes which is an isomorphism on $\bC$ points and formal neighborhoods is an isomorphism, thus $\tilde \cH(G) \to \cH(G)$ is an isomorphism, and so $\tilde S(G)/A_G \to S(G)$ is an isomorphism.
\end{proof}
We extend the action of $\BT \op$ $\bMgn$ to an action on $\sqcup_{G \in \rQ(g,n)} S(G)$ and $\sqcup_{G \in \rQ(g,n)}\tilde S(G)$, as follows. The category $\BT \op$ acts on the poset $\rQ(g,n)$ by gluing on trees of stable graphs, and the gluing map $f: \bar \cM_{g,n} \to \bMgn$ corresponding to $f \in \BT([m], [n])$ induces maps $S(G) \to S(f^*G)$ and $\tilde S(G) \to \tilde S(f^*G)$, which give maps on the disjoint union.
From the stratification of $\bMgn$ by $\rQ(g,n)$, we obtain a Borel--Moore homology spectral sequence. The next proposition records how the isomorphism of Proposition \ref{quotient} is compatible with the action of $\FSop$, and the spectral sequence of this stratification.
\begin{prop}\label{compatible}
The actions of $\BT \op$ on $\sqcup_{G \in \rQ(g,n)} S(G)$ and $\sqcup_{G \in \rQ(g,n)}\tilde S(G)$ induce an $\FSop$ module structure on Borel--Moore homology such that $$\bigoplus_{G \in \rQ(g,n)} H_\bdot^{\BM}( S(G)) \to \bigoplus_{G \in \rQ(g,n)} H_\bdot^{\BM}( \tilde S(G))$$ is a map of $\FSop$ modules, and the Borel--Moore homology spectral sequence $$\bigoplus_{G \in \rQ(g,n)} H_\bdot^{\BM}( S(G)) \implies H_{\bdot}(\bMgn)$$ is a spectral sequence of $\FSop$ modules.
\end{prop}
\begin{proof}
By definition, the action of $\BT \op$ preserves the stratification and acts through closed embeddings, and so induces maps of spectral sequences. On Borel--Moore homology, the action factors through $\FSop$ because the gluing map for a tree $T$ with $n$ leaves $T: \tilde S(G) \to \tilde S( T^* G)$ extends to a map $S(G) \times \bMon \to S(T^* G)$, the maps corresponding to different trees correspond to evaluating on different boundary strata of $\bMon$. Since $\bMon$ is smooth and proper, $H_0^{\BM}(\bMon) = \bQ$, and so every tree induces the same map on Borel-Moore homology.
\end{proof}
\section{Bounding Graphs} \label{graphcombinatorics}
The two lemmas in this subsection are combinatorial: the first says that there are only finitely many graphs in $\rQ(g,n)$ with $\leq i+1$ valent genus one vertices, and genus zero vertices obeying certain bounds.
\begin{lem}\label{finitelymanygraphs}
Let $f(i,e,a)$ be a linear function $f(i,e,a) = r i + s e + t a + u$ with $r,s,t \geq 0$.
If $G$ is a stable genus $g$ graph such that:
\begin{enumerate}
\item There are no edges between distinct genus $0$ vertices
\item Every genus $\geq 1$ vertex has valence $\leq i +1$
\item Every genus $0$ vertex with $e$ self edges and $a$ edges to other vertices, has $ \leq f(i,e,a)$ external edges. \end{enumerate}
Then the total number of external edges of $G$ is bounded by: $$n(G) \leq (i+1) g + (i+1) g f(i, g, (i+1)g)$$
\end{lem}
\begin{proof}
For a vertex $v$, let $e(v)$ be the number of self edges and $a(v)$ be the number of edges to a distinct vertex.
By $2)$ and $3)$ we have $$n(G) \leq \sum_{v, g(v) \geq 1} (i+1) + \sum_{v, ~g(v) = 0} f(i,e(v),a(v)). $$
For every genus $0$ vertex $v$, we have $e(v) \leq g$, because the genus is $\leq g$, and $a(v) \leq (i+1) g$ because any genus $0$ vertex is adjacent to a vertex of genus $\geq 1$. Thus $f(i,e(v), a(v)) \leq f(i, g, (i+1)g)$.
Now either $G$ has only one genus $0$ vertex, or every genus $0$ vertex must be adjacent to a genus $\geq 1$ vertex. In the first case $n(G) \leq f(i, g, 0)$, and in the second we have $$ \# \{\text{ genus $0$ vertices }\} \leq (i+1) g $$ So either $n(G) \leq f(i,e,0)$ or $$n(G) \leq (i+1) g + (i+1) g f(i, g, (i+1)g) $$ which proves the claim.
\end{proof}
The next lemma says that stable graphs of genus $g$ that correspond to strata of dimension $\leq i$ and only have genus $0$ vertices must contain certain subgraphs when $n(H) \gg 0$.
\begin{lem}\label{forcing}
Let $J$ be the stable graph consisting of a single genus $0$ vertex, $e$ self edges, and $n-e$ labelled external edges, with a partition of the $[n-e]$ sets $[a]$ and $[b]$ of size $a$ and $b$. If $b > 13a + 16i + 8e - 7$ then every stable graph $H$ with $\widebar H = J$ satisfies at least one of the following:
\begin{enumerate}
\item The sum $\sum_{v \in H} (n(v) - 3) $ is $> i$
\item There is a trivalent vertex $v$ adjacent to $2$ external edges in $[b]$
\item There are two adjacent trivalent vertices $v, v'$ such that both $v$ and $v'$ are adjacent to external edges in $[b]$
\end{enumerate}
\end{lem}
\begin{proof}
We prove the contrapositive. Let $a(v)$, $b(v)$, and $e(v)$ denote the number of edges adjacent to $v$ that are respectively external in $[a]$, external in $[b]$, and self edges. Let $H$ be a stable graph with $\widebar H = G$ and which satisfies:
\begin{enumerate}
\item $\sum_{v \text{ vertex}} (n(v) - 3) \leq i $
\item Every trivalent vertex has $b(v) \leq 1$
\item There are no two adjacent trivalent vertices $v, v'$ with $b(v), b(v') \geq 1$.
\end{enumerate}
First, by $(1)$, $H$ has $\leq i$ vertices of valence $> 3$.
Next by retracting external edges and computing the (negative) Euler characteristic of $H$, we have $$ e - 1 = \sum_{v \in H} ( 1/2 ~\#\{ \text{non external edges of $v$} \} - 1).$$ We have that $\#\{ \text{non external edges of $v$} \} = n(v) - a(v) - b(v)$. Then breaking up the sum by $n(v)$ and $a(v)$, we have:
\[\begin{split} e-1 ~= ~& \sum_{v, ~n(v) > 3} ( 1/2 ( n(v) - a(v) - b(v) ) - 1) ~~ + \sum_{v, ~n(v) = 3, ~a(v) \geq 1} ( 1/2 ( n(v) - a(v) - b(v) ) - 1) \\ +~ &\sum_{v, ~n(v) = 3, ~a(v) = 0, ~b(v) = 1} 0 ~~ + \sum_{v, ~n(v) = 3, ~a(v) = 0,~ b(v) = 0} 1/2 \\ \geq~ &~ -i - a + 0 ~+~ \sum_{v, ~n(v) = 3,~ a(v) = 0,~ b(v) = 0} 1/2. \end{split} \] Here we used $2)$ and the bound on vertices of valence $> 3$. Thus rearranging this inequality \[ \#\{\text{trivalent vertices $v$, s.t. $a(v) = b(v) = 0$}\} \leq 2(i +a + e - 1) \]
Since there are at most $a$ trivalent vertices with $a(v) > 0$, there are at most $\leq 2(i +a + e -1) +a$ trivalent vertices with $b(v) = 0$.
Next by $(3)$ every trivalent vertex with $b(v) = 1$ satisfies at least one of the following: \begin{itemize} \item[I] $v$ has an edge in $[a]$ \item[II]$v$ is adjacent to a trivalent vertex with $b(v) = 0$ \item[III] $v$ is adjacent to a vertex of valence $> 3$ \item[IV] $v$ has a self edge and is the only vertex. \end{itemize}
We have the following bounds the number of trivalent vertices satsifying each of these cases
\begin{itemize} \item[I] $\leq a$ \item[II] $\leq 3 \#\{ \text{ genus } 0 \text{ trivalent vertices with } b(v) = 0\} \leq3( 2(i +a + e -1) +a)$ \item[III] $$\leq \sum_{v, ~ n(v) > 3} n(v) \leq i + 3~ \#\{\text{ $> 3$-valent vertices}\} \leq i + 3i $$ \item[IV] $\leq 1$, \end{itemize}
where in III we have rearranged (1) and used the bound on vertices of valence $>3$. Combining these bounds, we have
\[ \#\{\text{trivalent vertices with $b(v) = 1$ } \} \leq a+ 3(2(i +a + e - 1)+a) + 4i + 1 = 10 a + 10 i + 6e - 5\]
Thus in total, the valence $3$ vertices contribute at most $13 a + 12 i + 8 e - 7$ to $b$.
Lastly we have $$\sum_{v,~n(v) > 3} b(v) \leq \sum_{v,~n(v) > 3} b(v) \leq i + 3i,$$ as above. So the vertices of valence $> 3$ contribute at most $4i$ to $b$. In total we see that $b \leq 13a + 16i + 8e - 7$, completing the proof.
\end{proof}
\section{Proof of Finite Generation} \label{proof}
Before proving finite generation, we record the following vanishing statement.
\begin{prop}\label{noclasses}
We have that $H_i^{\rB \rM}(\cM_{g,n}) = 0$ for $i < n -1$.
\end{prop}
\begin{proof}
We will show $H^i_c(\cM_{g,n}) = 0$ for $i < n+1$. Consider the map map $f: \Mgn \to \cM_{g,1}$ which forgets the last $n-1$ points. By the Leray sequence for compact support and base change, it suffices to show that $H^i_c(f\inv (x))$ vanishes for $i < n-1$. The fibers are topologically isomorphic to $\Conf_{n-1} (\Sigma_g - p)$ where $\Sigma_g - p$ is the genus $g$ surface with one point removed. This vanishing follows from Getzler's \cite{getzler1999resolving} spectral sequence converging to the compactly supported cohomology of configuration space, or its generalization by Petersen \cite{petersen2017spectral}. As a graded vector space, the $E_2$ page is $$\bigoplus_{0 \leq k \leq n-2} ( \bQ[-1] \oplus \bQ^{\oplus {2g}} [-2] )^{\otimes n -1 - k} \otimes \bQ^{\oplus c_{n,k}}[-k],$$ where $c_{n,k}$ is a certain unsigned Stirling number. Since the lowest degree term is in degree $n-1$ , the result follows.
\end{proof}
Throughout this section, we let $f(i,e,a) = 13a + 16i + 8e - 7$ and $p(i,g) = (i+1) g + (i+1) g f(i, g, (i+1)g)$. Expanding, we have $p(i,g) = 8 g^2 i^2 + 29 g^2 i + 16 g i^2 + 21 g^2 + 10 g i - 6 g$.
We now show that the $\FS \op$ module $n \mapsto H_i(\bMgn, \bQ)$ is a subquotient of one that is finitely generated in degrees $\leq p(i,g)$. In particular, since Sam and Snowden\cite{sam2017grobner} proved that submodules of finitely generated $\FSop$ modules are finitely generated, we see that the homology is finitely generated.
\begin{proof}[Proof of Theorem \ref{Mgnbar}]
We use the Borel--Moore homology spectral sequence for the stratification of $\bMgn$ by $\rQ(g,n)$ defined in Section \ref{stratification}. Write $S(G)$ for the stratum corresponding to $G$ and $\widetilde S (G) = \prod_{v \in G, ~ g(v) = 0} \widebar \cM_{0,n(v)} \times \prod_{v \in G, ~ g(v) \geq 1} \cM_{g(v), n(v)}$ for the cover of $S(G)$. Then by Proposition \ref{compatible}, this is a spectral sequence of $\FSop$ modules. And there is a homomorphism of $\FSop$ modules $$\bigoplus_{G \in \rQ(g,n)} H_\bdot^{\rB\rM}\left( S(G) \right) \to \bigoplus_{G \in \rQ(g,n)} H_\bdot^{\rB\rM}\left( \widetilde S(G) \right).$$ By Proposition \ref{quotient}, the left hand side is identified with the $A_G$ invariants of the right hand side.
We use Deligne's theory of weights \cite{ deligne1974poids, deligne1971theorie}. Since the homology of $\bMgn$ is pure, to show that $H_i(\bMgn)$ is finitely generated, it suffices to show that $\bigoplus_{G \in \rQ(g,n)} W_{-i} H_i^{\rB\rM}\left( \widetilde S(G) \right)$ is a finitely generated $\FSop$ module.
By Lemma \ref{finitelymanygraphs}, all graphs $G \in \rQ(g,n)$ with \begin{enumerate} \item no genus $1$ vertices of valence $> i+1$ \item no genus $0$ vertices with $e$ self edges, $a$ edges to adjacent vertices and $> f(i,e,a)$ external edges. \end{enumerate} have $n \leq p(i,g)$. Thus for the rest of the proof, we show that the $\FS \op$ module defined by \[ \bigoplus_{G \in \rQ(g,n)} W_{-i} H_i^{\rB\rM}(\widetilde S(G))\] is generated by the groups $W_{-i} H_i^{\rB\rM}(\tilde S(L))$ for graphs $L$ satisfying $1$ and $2$.
Accordingly, let $c$ be a pure class in $W_{-i}H_i^{\rB\rM}(\widetilde S(G))$, where $G$ is a graph that does not satisfy $1$ and $2$. We want to show that $c$ is a linear combination of classes pushed forward from lower degrees. (In other words $c$ is a linear combination of classes of the form $f^* (d)$ for some surjection $f: [n] \to [n-1]$).
We have that \[ W_{-i}H_i^{\rB\rM}(\widetilde S(G))= \bigoplus_{j: {\rm Vert }(G) \to \bN, ~ \sum_v j(v) = i} \left( \bigotimes_{v, ~g(v) = 0} W_{-j(v)} H_{j(v)}^{\rB\rM}( \widebar \cM_{0}(v) )\otimes \bigotimes_{v,~ g(v) \geq 1} W_{-{j(v)}} H_{j(v)}^{\rB\rM}(\cM_{n(v), g(v)}) \right).\]
If $G$ has a genus $\geq 1$ vertex with valence $> i+1$, then by Proposition \ref{noclasses} we have that $W_{-{j}} H_{j}(\cM_{g(v), n(v)}) = 0$ for all $j \leq i$. Thus $W_{-i} H_i^{\rB\rM}(\widetilde S(G)) = 0$ and the claim follows.
So assume that $w$ is a genus $0$ vertex of $G$ with $e$ self edges, $a$ edges to distinct vertices, and $> f(i,e,a)$ external edges.
Taking linear combinations it suffices to show $c = f_*(d)$ for classes $c$ of the form $c =c_w \otimes c'$, where $c_w \in H_{j(w)}( \widebar \cM_0(w))$ is a pure class for some $j(w) \leq i$ and \[ c \in W_{j(w) -i} H_{i - j(w)}\left( \prod_{v \neq w, ~ g(v) = 0} \widebar \cM_{e(v),n(v), 0} \times \prod_{v , ~ g(v) \geq 1} \cM_{g(v), n(v)} \right) \]
Let $J$ be the graph spanned by edges adjacent to $v_0$ and $b > 11 i + 7 e + 9 a - 5$ external edges. The stable graph stratification $\widebar \cM_{e(v),n(v), 0}$ has strata given by finite quotients of copies of $\Mon$, so the Borel--Moore homology spectral sequence shows that the pure Borel--Moore homology of $\widebar \cM_{e(v),n(v), 0}$ is spanned by fundamental classes of the strata: $[\widebar{ \cM_H}]$ for stable graphs $H$ with $\widebar H = J$ and $\dim \cM_H \leq i$. So it suffices to show that $[\widebar{\cM_H}] \otimes c' = f^*(d)$ for some surjection $f: [n] \to [n-1]$. By Lemma \ref{forcing} at least one of the following holds:
\begin{enumerate}
\item $\sum_{v \text{ vertex}} (n(v) - 3) $ is $> i$, this contradicts the dimension of $\cM_H$ and so does not hold.
\item There a trivalent vertex $v$ adjacent to $2$ external edges, labelled by $s < t \in [n]$. Let $f$ be the surjection $[n] \to [n]/s \sim t \to [n-1]$, where we have fixed a bijection $b$ between $[n]/s \sim t$ and $[n-1]$. Let $K$ be the graph obtained from $H$ by merging the edges labelled by $s,t$, labelled by $[n-1]$ via the bijection $b$. Then $f^* ([\widebar{ \cM_K}] \otimes c') = [\widebar {\cM_H}] \otimes c'$.
\item There are two adjacent trivalent vertices $v_1, v_2$ in $H$, such that $v_1, v_2$ both have external edges. Then applying the cross ratio relation to $H$ at $v_1, v_2$ we have that $[\widebar{ \cM_H]} = [\widebar{ \cM_{H'}}]$ where $H'$ has a vertex which is adjacent to two external edges. Now the same argument as $2)$ applies.
\end{enumerate}
Therefore we have $[\widebar \cM_H] \otimes c' = f^*(d)$ for some surjection $f: [n] \to [n-1]$, and we are done.
\end{proof}
\section{Further Questions}\label{furtherquestions}
\subsection{Extensions of Theorem \ref{Mgnbar}}
From Theorem
\ref{hilbert}, we see that there exists $M(g,i) \in \bN$ such that the function $n \mapsto \dim H_i(\bMgn)$ agrees with a sum of polynomials times exponentials for $n \geq M(g,i)$. What is an effective bound on $M(g,i)$? Equivalently, what is a bound on the degree of the numerator of the generating function $\sum_n \dim H_i(\cMgn) t^n$ in terms of $g$ and $i$?
We show that $H_i(\bMgn)$ is a subquotient of a module generated in degree $O(g^2 i^2)$. We expect that this bound can be improved to one linear in $g$ and $i$. What is a tight bound?
What is an effective bound on the constant $N$ of Theorem \ref{stability}?
\subsection{Representation theory of $\FSop$} Finitely generated $\FSop$ modules are not yet as well understood as finitely generated $\FI$ modules. Any new results on finitely generated $\FSop$ modules will in particular apply to $H_i(\bMgn)$.
Let $M_\lambda$ be an irreducible representation of $\bS_n$. What is the decomposition of the projective $\FSop$ representation $M_\lambda \otimes_{\bS_n} \bQ \FS(-,n)$ in degree $m$? Equivalently, what is the degree $m$ term in the plethysm of symmetric functions $s_\lambda[\sum_{k \geq 1} h_k]$?
The dimension functions of $\FI$ modules are eventually polynomial, and there has been much work in studying the invariants of $\FI$ modules which control when this occurs. (For recent applications see \cite{church2017linear}). Which invariants of $\FSop$ modules control when the Hilbert series of a finitely generated $\FSop$ module agrees with a sum of polynomials times exponential? Which invariants control its presentation degree?
\subsection{Other operadic actions on $H_\bdot(\bMgn)$}
Operads give a conceptual interpretation of the $\FSop$ action on $H_i(\bMgn)$. The homology of moduli space $\bigoplus_{g,n} H_\bdot(\bMgn)$ forms a (modular) operad, and $\bigoplus_{g,n} H_i(\bMgn)$ is a right module over the suboperad generated by $[\bar \cM_{0,3}]$. This operad is isomorphic to the commutative operad, thus we obtain an right action of the commutative operad, which corresponds to a representation of its opposite wiring category, $\FSop$. Using this approach, we can define other right actions of wiring categories on the homology of moduli space. For example, for any $j$, the operad generated by $[\bMgn]$ acts on $\bigoplus_{j,g,n|~ 3g - 3 + n - j = i } H_j(\bMgn)$. Is this right module finitely generated?
Similarly, as first observed by Kapranov--Manin \cite{kapranov2001modules}, for $g$ fixed, the vector space $\bigoplus_{i,n} H_i(\bMgn)$ is a right module over the \emph{hypercommutative operad}, the operad spanned by the fundamental classes of $\bMon$. Is this right module finitely generated?
For the answers to these questions to have concrete implications, we need to understand the structure of right modules over these operads (equivalently representation theory of their opposite wiring categories). If a sequence of graded vector spaces is a finitely generated module over the hypercommutative operad, is its Hilbert series restricted?
Is the category $\bB\bT \op$, the opposite of the wiring category of the free operad on a binary commutative operation, Noetherian? If so, it would be possible to simplify the proof of Theorem \ref{Mgnbar}. In \cite{barter2015noetherianity}, Barter established the Noetherianity of a category of trees, but we do not know of a relationship between this category and $\bB \bT \op$.
\nocite{*}
\printbibliography
\end{document}
|
\section{Object-level analysis}\label{Sec:Modularity}
While the ECF property is capable of detecting unwanted executions which do not satisfy it,
it can be further used for modular analysis of objects.
We will show that in environments in which objects are encapsulated,
we can consider ECF for executions of a single object only, to help simplify object-level analysis.
We define the notion of a \emph{most general client (MGC)} in our model.
The most general client for an object $\objectid$, $MGC_\objectid$, is an external program that works on a system that includes a single object in the store.
The store $\store$ of $MGC_\objectid$ contains a single object $NoCB(\objectid)$, which is built based on the original object $\objectid$.
Every invocation of an object of the form $\SCallC{x}{\objectid'}{e}$ in $\objectid$ is replaced with a non-deterministic choice of the value of $x$ as returned by the call, in correspondence with the definition of havoc transitions in \Cref{Sec:CorrectnessConditions}.
Furthermore, $MGC_\objectid$ is allowed to repeatedly call $NoCB(\objectid)$ with any parameter and in any order.
As such, the semantics of $MGC_\objectid$ soundly approximate all executions of the object $\objectid$ (see, e.g.,~\citet{GotsmanYangICALP11}), while every execution in $MGC_\objectid$ is in fact a projected, callback-free execution of $\objectid$.
We show that, if the object $\objectid$ is ECF in the general model, then any object-level assertion can be soundly verified on $MGC_\objectid$.
This is because, all reachable states of an object $\objectid$ in any arbitrary code context $\ocode$ that does not change $\objectid$'s code, are reachable in the system containing only $NoCB(\objectid)$.
\begin{theorem}\label{Thm:ObjLevelAnalysisSoundness}
Let $R$ be the set of all states of the object $\objectid$ in a quiescent state,
and let $R_0$ be the set of all states of the object $\objectid$ in a run of $MGC_\objectid$.
If $\objectid$ is $\SECF[\subFS]$ then $R_0\supset R$.
\end{theorem}
Clearly, the theorem holds if the object is shown to be $\SECF[\subC]$, since this implies it is also $\SECF[\subFS]$ (see \Cref{Thm:CECFImpliesFSECF}).
This analysis is not overly imprecise, since in real environments, such as Ethereum, we could simulate such behaviors.
This is particularly correct since in Ethereum the store of the objects is updateable, and new objects may be added to the system.
Importantly, the analysis simply assumes ECF, and does not require to prove it:
An alternative formulation of \Cref{Thm:ObjLevelAnalysisSoundness} is assuming that the runtime system enforces $\DECF$ on all executions.
In that case, the analysis is still sound.
It is not unreasonable to assume such a dynamic analysis of $\DECF$,
because we found out an efficient method to verify it, presented in \Cref{Sec:Dynamic}.
We illustrate \Cref{Thm:ObjLevelAnalysisSoundness} using the example shown in \Cref{Fi:DaoContract}.
We implemented \code{Dao} as a class in Dafny~\cite{Leino2010}
with two methods: \code{deposit} and \code{withdrawAll},
whose pre- and post- conditions capture the object invariant which should be valid after every execution of the \code{Dao} object.
Primarily, we wish to ensure that the data elements in the \code{credit} map are not negative, and that the sum of all these elements is equal to the balance of the \code{Dao}.
We model the \code{pay} method without the recursive call to \code{withdrawAll}, but annotate it as possibly modifying (any field of) the \code{Dao} object.
This annotation generalizes the possible behaviors of the \code{Dao} object without the ECF property.
With such weak assumptions, and perhaps unsurprisingly, Dafny fails to verify the postconditions for both the original and the fixed versions of the \code{Dao} object.
In contrast, when we assume that the ECF property holds, technically by adding a postcondition to \code{pay} which ensures that the previously read fields of the \code{Dao} object, i.e., \code{balance} and \code{credit[o]}, are not modified, Dafny is able to establish the post condition.
\Cref{Thm:ObjLevelAnalysisSoundness} implies the fixed \code{Dao} contract respects the given specification when executed using the original runtime system and the original DAO object respects the specification if it is executed on a runtime system which enforces ECF.
\section{Conclusion}\label{Sec:Conclusion}
In this paper we have presented a simple generic correctness
condition for callbacks called Effective Callback Freedom and
studied its usefulness.
We have shown that it enables modular reasoning in environments with local-only mutable states like Ethereum.
We have also shown that in Ethereum it can be used to prevent bugs without drastically limiting programming style,
and that it can be checked dynamically with low runtime overhead.
In the future, we expect to apply the concept of ECF and prove its usefulness in other environments such as Microservices and Amazon~$\lambda$.
\section{Correctness conditions}\label{Sec:CorrectnessConditions}
In this section we give a formal definition for two notions of the ECF property: \emph{final-state ECF} and \emph{conflict ECF}.
We start by formally defining callbacks and callback-freedom in executions.
\newcommand{\cb}[2]{\mathbb{CB}_{#2}(#1)}
\newcommand{\ncb}[2]{{\mathbb{CB}^{f}_{#2}(#1)}}
A stack frame $\stk$ \emph{is a callback frame (to object $\SelObject(\stk)$)} in a stack $\Stk$
if there exist stack frames $\stk'$ and $\stk''$ such that
$\Stk =\_ \stk'' \_\,\stk' \_\, \stk \_$ and $\SelObject(\stk)=\SelObject(\stk'')$, but $\SelObject(\stk)\neq\SelObject(\stk')$.
A stack $\Stk$ contains a \emph{callback (to $\objectid$)}, denoted by $\cb{\Stk}{\objectid}$, if it contains a callback frame.
A state $\sstate$ contains a \emph{callback (to $\objectid$)}, denoted by $\cb{\sstate}{\objectid}$, if its stack does, and
an execution $\exec$ contains a \emph{callback (to $\objectid$)}, denoted by $\cb{\exec}{\objectid}$, if it contains a state $\sstate$ such that $\cb{\sstate}{\objectid}$.
A stack resp. state resp. execution is \emph{callback-free (for $\objectid$)}, denoted by $\ncb{\Stk}{\objectid}$ resp. $\ncb{\sstate}{\objectid}$ resp. $\ncb{\exec}{\objectid}$,
if it does not contain a callback.
We now define what it means for an execution to be \emph{effectively} callback free, or \emph{ECF}, with respect to a given object $\objectid$.
Note that ECF is a property of both an execution $\exec$ and some object $\objectid$.
Specfically, we are interested in the case where $\exec$ has a callback to $\objectid$.
\begin{definition
A well-formed complete execution $\ocode \inctx \exec$ is \emph{equivalently effectively callback-free for an object $\objectid$}, denoted by $\DECFo[\subFS]{\objectid}$, if there is a well-formed callback-free run $\ocode' \inctx \exec'$, which is final-state equivalent for $\objectid$ to $\exec$:
$$
\ocode \inctx \exec \vDash \DECFo[\subFS]{\objectid} \iff \exists \ocode',\exec'. \,
\ocode \inctx\exec \fseqo{\objectid} \ocode' \inctx\exec'
\land
\ncb{\exec'}{\objectid}
$$
\end{definition}
We say that the execution $\ocode' \inctx \exec'$ is a \emph{witness} for $\exec$ being a $\DECFo[\subFS]{\objectid}$ execution.
Checking $\DECFo[\subFS]{\objectid}$ is difficult in practice, and undecidable in general for models with an infinite state.
We describe a stronger definition of ECF, based on conflict-equivalence, called $\CECF$,
which permits an efficient algorithm for checking it.
Interestingly, even though executions in our model do not allow for concurrency, callbacks can be thought of as allowing to express a limited subset of concurrent executions.
In fact, the ECF property in our model is analogous to serializability in models that permit concurrency.
Using this analogy, \emph{invocations} are analogous to \emph{transactions}.
We show what this means to reorder invocations in the sequential semantics of $\PLname$.
In general terms, $\CECF$ requires to find a callback-free execution which is conflict-equivalent to the execution with the callbacks.
Conflict-equivalence requires that the trace of the callback-free execution is a permutation of the trace of the original execution.
It is thus useful to start with a characterization of the legal permutations of an execution.
Firstly, the permutation may not break program order of contract code.
That is, the permutation must retain the ordering of events whose transitions are part of the same invocation $\exec$,
and their state have the same depth as $\execdepth(\exec)$.
Secondly, we want to allow permutations that remove callback invocations from their original call location, and sets them to execute in a quiescent state, and still receive a modular well-formed execution.
\input{extraction-fig}
When we permute a trace such that a callback invocation is removed from its original place, we replace the call transition leading to the callback with a havoc transition.
An example can be seen in \Cref{fig:CallbackExtraction}, showing all legal permutations of a trace that has a callback-free execution with havoc transitions.
\changed{
\begin{remark}
A havoc transition is not part of the permuted trace; it is merely used to justify the execution which is no longer well-formed as we defined earlier.
\end{remark}
}
Using modular well-formed executions, we can formally define the $\CECF$ property for executions, $\DECF[\subC]$:
\begin{definition
A well-formed complete execution $\ocode \inctx \exec$ is \emph{conflict-equivalently effectively callback-free for an object $\objectid$},
denoted by $\ocode \inctx \exec \vDash \DECFo[\subC]{\objectid}$,
if there is a modular well-formed callback-free run $\ocode \inctx \exec'$ which is object-conflict-equivalent to $\exec$ for $\objectid$:
$$
\ocode \inctx \exec \vDash \DECFo[\subC]{\objectid}
\iff
\exists \exec'. \,
\ocode \inctx \exec \ceq^{\objectid} \ocode \inctx \exec'
\land
\ncb{\exec'}{\objectid}
\,.
$$
\end{definition}
It is easy to prove that conflict equivalence implies final-state equivalence (see, e.g., \cite{BOOK:BHG87}).
Thus, it can be concluded that $\CECF$ implies $\FSECF$ \changed{as, by using \Cref{Prop:ModularToWellFormed},}{} we can use the same witness for $\exec$ being $\DECF[\subC]$ to prove that $\exec$ is also $\DECF[\subFS]$.
\begin{theorem}\label{Thm:CECFImpliesFSECF}
Let $\exec$ be a well-formed complete execution.
If $\exec \vDash \DECFo[\subC]{\objectid}$ then $\exec \vDash \DECFo[\subFS]{\objectid}$.
\end{theorem}
\changed{
\begin{proof}
Since $\ocode \inctx \exec \vDash \DECFo[\subC]{\objectid}$, there is a callback-free modular well-formed complete execution $\ocode \inctx \exec'$
such that $\ocode \inctx \exec \ceq^{\objectid} \ocode \inctx \exec'$.
It is easy to see that since $\ncb{\exec'}{\objectid}$ then $\exec'|_{\objectid}$ is too a callback-free and modular well-formed run.
From an immediate generalization of \Cref{Prop:ModularToWellFormed} to runs,
we conclude that for $\ocode \inctx \exec'|_{\objectid}$ there is
a context $\ocode'$
and a well-formed run $\ocode' \inctx \hat{\exec}$
such that $\ocode'(\objectid)=\ocode(\objectid)$,
and $\trace(\exec'|_{\objectid})=\trace(\hat{\exec}|_{\objectid})$.
It should be noted that the equality of the trace projected on $\objectid$ implies callback-freedom for $\objectid$ is retained:
namely, $\ncb{\hat{\exec}}{\objectid}$.
Conflict equivalence implies final-state equivalence (\cite{BOOK:BHG87}),
hence, $\src(\exec)=\src(\exec')$ and $\trg(\exec)=\trg(\exec')$.
Also, because conflict ordering in primitive commands of $\objectid$ is retained between $\exec'|_{\objectid}$ and $\hat{\exec}|_{\objectid}$, we conclude that $\exec' \fseq^{\objectid} \hat{\exec}$, therefore $\src(\exec')=\src(\hat{\exec})$ and $\trg(\exec')=\trg(\hat{\exec})$.
It can be concluded then that since (i) $\src(\exec)=\src(\hat{\exec})$ and $\trg(\exec)=\trg(\hat{\exec})$,
(ii) $\ocode(\objectid)=\ocode'(\objectid)$, and (iii) $\hat{\exec}$ is well-formed, then:
$\exec \fseq^{\objectid} \hat{\exec}$.
making $\ocode' \inctx \hat{\exec}$ a witness proving $\exec \vDash \DECFo[\subFS]{\objectid}$.
\end{proof}
}
Finally, as we are also interested in ECF as a property of objects (\SECF{}),
we extend the definitions of $\FSECF$ and $\CECF$ to objects (\SECF{\subFS} and \SECF{\subC}) instead of executions (\DECF{\subFS} and \DECF{\subC}),
which we refer to as \emph{static ECF}.
\ignore{\emph{projected executions}, which include only transitions that pertain to a single object.
\begin{definition
Let $\exec$ be an execution and $\objectid$ be an object which is the active object of a state $\sstate\in\execstates(\exec)$.
The \emph{projected execution of $\objectid$} is an execution $\exec_o$ whose trace contains only events $\event\in\trace(\exec)$ such that $\objectid(\event)=\objectid$.
\end{definition}
\changed{Projected executions are used to define the two notions of static ECF (\SECF{}):}
}
\begin{definition
An object $\objectid$ is $\SECF[\subFS]$
if for every complete execution $\ocode \inctx \exec$ it holds that \mbox{$\ocode \inctx \exec \vDash \DECFo[\subFS]{\objectid}$}.
$\objectid$ is $\SECF[\subC]$
if for every complete execution $\ocode \inctx \exec$ it holds that $\ocode \inctx \exec \vDash \DECFo[\subC]{\objectid}$ .
\end{definition}
\section{Dynamic Verification}\label{Sec:Dynamic}
We describe a sound procedure for verifying the \DECF[\subC] property dynamically.
More precisely, for each execution, it checks for every object that participates in the execution, if the subsequence of the transitions that pertain only to that object (the \emph{projected execution}) is $\CECF$.
We assume the existence of an interpreter or virtual machine implementing the semantics defined in \Cref{Sec:Preliminaries}.
Below is a description of the data structures used by the algorithm, as well as the instrumentation of the object code to maintain these data structures. We then present a higher-level description of the algorithm, followed with pseudo-code and a complexity analysis.
We use the example presented in the overview section in \Cref{Fi:DaoContract} to explain the procedure.
The general structure of the procedure is that the instrumentation step starts every time we exit a quiescent state, and ends when we reach the next quiescent state.
Once instrumentation has completed,
the algorithm runs on the instrumented structures and returns whether all projected executions derived from the execution are ECF.
The procedure repeats each time we enter an active state.
\subsection{Data Structures}
A \emph{segment} is a data structure that captures metadata about a portion of the execution's states.
This portion consists of a sequence of adjacent transitions, whose top stack frames have the same active object.
That is, an invocation of a different object marks the beginning of a new segment, as well as returning from an invocation to a caller invocation which is executed in the context of a different object.
However, a call from one object to itself does not break the current segment (This is motivated by the definition of callbacks in \Cref{Sec:CorrectnessConditions}).
In simpler terms, a new segment is defined each time the active object changes,
either when we push a stack frame with a different object,
or pop a stack frame such that the new top frame has a different active object.
We show how segments are determined in the instrumentation in \Cref{fig:Instrumentation}, using hooks on calls and returns.
\begin{example}\label{Exmp:Execution}
In the example DAO contract in \Cref{Se:Overview}, an attack execution consists of 6 segments:
(1) the first invocation of \code{withdrawAll}, lines 1-3;
(2) an invocation of \code{pay}, lines 3-5;
(3) the second invocation of \code{withdrawAll}, lines 1-3;
(4) a full invocation of \code{pay}, lines 3-4,7;
(5) the second invocation of \code{withdrawAll}, line 5;
(6) the first invocation of \code{withdrawAll}, line 5;
\end{example}
\begin{definition}[Segments]
A \emph{segment} $\segment$ is representative of a maximal sequence of adjacent transitions pertaining to the same object.
A segment $\segment=\ST{\readset{},\writeset{},\depth{},\indexInExec{}}$ contains information about fields accessed in the segment,
denoted $\readset{\segment}$ and $\writeset{\segment}$ for the read- and write- sets, respectively.
In addition, a segment contains information about the depth of the invocation (denoted $\depth{\segment}$), which is equal to the depth of the transitions' states.
Last, the index in the execution (denoted $\indexInExec{\segment}$), is strictly increasing according to order of creation of the segments.
\end{definition}
The primary metadata saved in each segment is the read and write sets of the fields of the object that were accessed by commands executed in the transitions that pertain to the segment.
Other metadata includes the depth of the invocations in the stack, and an index to maintain the order of the segments in the execution.
\begin{example}\label{Exmp:Segments}
We write down the segments that pertain to the \code{DAO} object in the overview example of the attack execution,
in the same order as they appear in the execution:
$$\begin{array}{ll}
\segment_1=\ST{\SET{\icredit[\code{Attacker}],\ibalance},\SET{\ibalance},0,1} &
\segment_2=\ST{\SET{\icredit[\code{Attacker}],\ibalance},\SET{\ibalance},1,2} \\
\segment_3=\ST{\SET{},\SET{\icredit[\code{Attacker}]},1,3} &
\segment_4=\ST{\SET{},\SET{\icredit[\code{Attacker}]},0,4}
\end{array}
$$
\end{example}
An execution can be represented as a linear sequence of segments.
Furthermore, from these segments
we can determine
the invocations that the execution contains.
\begin{remark}
Segments can be used as an alternative representation of executions and invocations,
that generalize data saved by a sequence of transitions.
In this section only, we redefine the notions of executions and invocations to refer to segments instead of transitions.
\end{remark}
\begin{definition}[Executions, Invocations, and Callbacks]
An \emph{execution} can be represented using a sequence of its instrumented segments $\exec=\ST{\segment_1,\ldots,\segment_n}$.
We can access the j'th segment of the execution using $\exec(j)=\segment_j$. We trivially have that $\indexInExec{\segment_j}=j$.
An \emph{invocation} is a sequence of segments $\inv=\ST{\segment^{\inv}_1,\ldots,\segment^{\inv}_k}$
such that there is a number $d$ for which all of the following holds:
\begin{enumerate}
\item $\forall \segment\in\inv. \depth{\segment}=d$ (all segments of the invocation are in the same depth).
\item $d > \depthNoArg(\exec(\indexInExecNoArg(\segment^{\inv}_1)-1))$ (the first segment before the first segment in the invocation has lower depth, proving it is indeed the beginning of an invocation).
\item $d > \depthNoArg(\exec(\indexInExecNoArg(\segment^{\inv}_k)+1))$ (the first segment after the last segment in the invocations has lower depth, proving it is indeed the end of an invocation).
\item $\forall j. \indexInExec{\segment^{\inv}_1} < j < \indexInExec{\segment^{\inv}_k} \implies \depth{\exec(j)}\geq d$ (the invocation does not end before the last segment, that is all segments of depth $d$ in the given range belong to the same invocation).
\end{enumerate}
As all segments included in the invocation has the same depth $d$, we denote the \emph{depth of an invocation} by $\depthNoArg(\inv)=d$.
We say that an invocation $\inv$ is a \emph{callback} in another invocation $\inv'$ (denoted $\inv\subexec\inv'$) if $\indexInExec{\inv(1)}>\indexInExec{\inv'(1)} \land \indexInExec{\inv(1)}<\indexInExec{\inv'(|\inv'|)}$.
\end{definition}
\begin{remark}
Unlike the definition of invocations in \Cref{Sec:Preliminaries},
here invocations capture only the transitions in the same depth as the depth of its first transition, and not transitions in higher depth.
This allows to define $\depthNoArg(\inv)$ for an invocation $\inv$.
\end{remark}
\begin{example}\label{Exmp:Invocations}
In the attack execution presented in \Cref{Se:Overview},
the first invocation of \code{withdrawAll} is $\inv_{wd_1}=\ST{\segment_1,\segment_4}$, and the second invocation is $\inv_{wd_2}=\ST{\segment_2,\segment_3}$.
$\inv_{wd_2}$ is a callback of $\inv_{wd_1}$: $\inv_{wd_2}\subexec\inv_{wd_1}$.
\end{example}
We associate with each segment in depth $>1$ a \emph{prefix-set} and \emph{suffix-set} of all segments in the caller that precede, or respectively, proceed it:
\begin{definition}[Prefix and Suffix segments]
Let a set of segments representing an invocation $\inv=\ST{\segment^i_{\it{caller}}}$,
and a single segment $\segment_{\it{cb}}$ with $\depthNoArg(\segment_{\it{cb}})>\depthNoArg(\inv)$
and $\indexInExecNoArg(\segment_{\it{cb}}) \in \SET{\indexInExec{\inv(1)},\ldots,\indexInExec{\inv(|\inv|)}}$.
We define for $\segment_{\it{cb}}$ its \emph{prefix and suffix sets relatively to a caller $\inv$} by partitioning the segment in $\inv$ to segments whose index in the execution is smaller than the index of the callback segment $\segment_{\it{cb}}$ (prefix), and segments whose index in the execution is larger than it (suffix):
$$
\begin{array}{lll}
\prefix{\inv,\segment_{\it{cb}}} & = & \{\segment_{\it{caller}}\in\inv \mid \indexInExec{\segment_{\it{caller}}} < \indexInExec{\segment_{\it{cb}}} \}
\\
\suffix{\inv,\segment_{\it{cb}}} & = & \{\segment_{\it{caller}}\in\inv \mid \indexInExec{\segment_{\it{caller}}} > \indexInExec{\segment_{\it{cb}}} \}
\end{array}
$$
\end{definition}
\begin{example}\label{Exmp:PrefixSuffixSegments}
The prefix and suffix segments of $\segment_2$ and $\segment_3$ with respect to $\inv_{wd_1}$ are:
$$
\begin{array}{rl}
\prefix{\inv_{wd_1},\segment_2} = \prefix{\inv_{wd_1},\segment_3} & =\SET{\segment_1} \\
\suffix{\inv_{wd_1},\segment_2} = \suffix{\inv_{wd_1},\segment_3} & =\SET{\segment_4}
\end{array}
$$
\end{example}
The instrumentation process creates the segments and the invocations.
We show pseudo-code of the instrumentation procedure in \Cref{fig:Instrumentation}.
\begin{figure}
\begin{lstlisting}[mathescape=true,basicstyle=\ttfamily\scriptsize]
Segment { Obj, Caller, R, W, D, I }
Invocation { Caller, Obj }
Init():
execution := ()
curSegment := $\bot$
invocations := Map<Invocation -> $\overline{\texttt{Segment}}$>
UponInvocation(object):
if fromQuiescent // Procedure starts
Init()
if object != curSegment.Obj
caller := fromQuiescent ? TopInvocation : curSegment.Caller
inv := Invocation(caller, object)
AddSegment(object, inv, curSegment.D+1, curSegment.I+1)
UponReturn(object):
caller := toQuiescent ? TopInvocation : curSegment.Caller.Caller
if caller.Obj != object
AddSegment(object, caller, curSegment.D-1, curSegment.I+1)
if caller == TopInvocation // End of instrumentation step
CheckECFForAllObjects() // Run the algorithm, and finish procedure
AddSegment(object, caller, D, I):
segment := Segment(object, caller, {}, {}, D, I)
Append(execution, segment)
Append(invocations[caller], segment)
curSegment := segment
UponObjectVarRead(object, F):
curSegment.R[F] := 1
UponObjectVarWrite(object, F):
curSegment.W[F] := 1
\end{lstlisting}
\caption{
Instrumentation procedures, implemented as hooks called upon call commands, return commands, and object variable read/write access command. Generates the \code{execution}, which is a list of segments, and \code{invocations}, a map of invocation identifiers to an invocation object keeping the caller of an invocation and the list of segments that are part of the invocation in the same depth.
The top-level invocation is identified as \code{TopInvocation}.
}\label{fig:Instrumentation}
\end{figure}
The basic check on segments is the commutativity check. We define segment commutativity using read and write sets.
We will show that we actually check commutativity of a segment with either a prefix or suffix segment.
As the prefix/suffix segments are sets of segments,
the read and write sets of prefix/suffix segments are a union of the respective read and write sets of all the segments contained in the prefix or suffix segment.
\begin{definition}[Commutative Segments]\label{def:CommutativeSegments}
Segments $\segment_1$ and $\segment_2$ commute, denoted by $\Commute{\segment_1}{\segment_2}$, if:
$$
\Commute{\segment_1}{\segment_2} \eqdef \readset{\segment_1}\cap\writeset{\segment_2}=\emptyset \land \readset{\segment_2}\cap\writeset{\segment_1}=\emptyset \land \writeset{\segment_1}\cap\writeset{\segment_2}=\emptyset
$$
If segments $\segment_1$ and $\segment_2$ do not commute, we denote this by $\nCommute{\segment_1}{\segment_2}$.
\end{definition}
\begin{example}\label{Exmp:Commutativity}
In the attack execution presented in \Cref{Se:Overview}, indeed we have that $\nCommute{\segment_2}{\segment_1}$,
as $\readset{\segment_1}\cap\writeset{\segment_2}=\SET{\ibalance}$.
Similarly, $\nCommute{\segment_3}{\segment_1}$ because of $\icredit[o]$,
as $\readset{\segment_1}\cap\writeset{\segment_3}=\SET{\icredit[o]}$,
and therefore also $\nCommute{\segment_2}{\segment_4}$.
However, $\segment_3$ does commute with $\segment_4$: $\Commute{\segment_3}{\segment_4}$.
\end{example}
\subsection{Algorithm}
We start with a high-level description of the algorithm.
The algorithm is called every time the system reaches a quiescent state, working on the last complete execution.
The algorithm generates a relation of invocations that defines
constraints on the ordering of invocations in different stack depths, similar to a `happens-before'~\cite{Lamport:1978:TCO:359545.359563} relation.
We name this relation the \emph{invocation order constraint (IOC) graph}.
For example, if a segment $\segment$ of a callback invocation $\inv_{\it{cb}}$ is not commuting with its prefix with respect to one of its calling invocations $\segment_{\it{caller}}$ (i.e., $\nCommute{\segment}{\prefix{\inv_{\it{caller}},\segment}}$), then we add the constraint that the invocation of the caller has to occur before the callback: $\hb{\inv_{\it{caller}}}{\inv_{\it{cb}}}$.
The IOC relation of invocations is thus defined as:
$$
\begin{array}{lll}
\hb{\inv}{\inv'} & \eqdef & (\inv'\subexec\inv \land \exists \segment\in\inv'. \nCommute{\segment}{\prefix{\inv,\segment}}) \\
& & \lor (\inv\subexec\inv' \land \exists \segment\in\inv. \nCommute{\segment}{\suffix{\inv',\segment}})
\end{array}
$$
\begin{example}\label{Exmp:IOC}
The IOC relation of the attack execution in \Cref{Se:Overview} can be easily calculated with the previous metadata given in examples \ref{Exmp:PrefixSuffixSegments} and \ref{Exmp:Commutativity}.
We have that $\hb{\inv_{wd_1}}{\inv_{wd_2}}$ as $\inv_{wd_2}\subexec\inv_{wd_1}$ and for $\segment_2\in\inv_{wd_2}$, $\nCommute{\segment_2}{\prefix{\inv_{wd_1},\segment_2}}$.
Similarly, $\hb{\inv_{wd_2}}{\inv_{wd_1}}$ as for $\segment_2\in\inv_{wd_2}$, $\nCommute{\segment_2}{\suffix{\inv_{wd_1},\segment_2}}$.
\end{example}
After the IOC relation is defined, the algorithm considers the graph induced by this relation,
and checks it has no cycles.
A cycle in the graph could appear if, for example, there is a callback invocation and some caller invocation that contains it, for which there is both (1) a segment that does not commute with its prefix with respect to the caller; and (2) a segment that does not commute with its suffix with respect to the caller.
As each vertex in this graph represents an invocation, the topological sorting returns an ordering of the invocations, which is $\CECF$.
We are merely interested if there is such a topological sorting, that is, if the IOC relation does not contain a cycle.
\begin{theorem
Let $\exec$ be an execution and let $\it{Inv}$ be a map of the instrumented invocations to their segments.
We denote by $\hb{}{}$ the IOC on $\it{Inv}$.
If $\hb{}{}$ has no cycle, then $\exec$ is $\CECF$.
\end{theorem}
\begin{proof}
We assume $\hb{}{}$ has no cycle.
We take a total order $\hb{}{^t}$ of $\it{Inv}$ induced by the transitive closure on $\hb{}{}$.
From $\hb{}{^t}$ we build a run $\exec'$ such that every invocation in $\it{Inv}$ starts in a quiescent state in the order determined by $\hb{}{^t}$.
$\exec'$ is conflict-equivalent to $\exec$.
To show this, we consider two transitions $\transition_1$ and $\transition_2$ which conflict in $\exec$.
If $\transition_1$ and $\transition_2$ are both captured in the same segment during instrumentation, then their ordering is kept in $\exec'$ which only reorders invocations. In particular, the program order of invocations is kept.
The same argument applies when $\transition_1$ and $\transition_2$ are not captured by the same segment, but their respective segments are both part of the same invocation.
In the general case, $\transition_1$ and $\transition_2$ each belong to different segments, pertaining to different invocations.
In that case, their ordering in $\exec'$ is kept as $\hb{}{^t}$ respects that conflict.
\end{proof}
\begin{example}
In continuation to our running example, it is immediate that the IOC relation of the attack execution on the DAO object has a cycle: $\hb{\inv_{wd_1}}{\hb{\inv_{wd_2}}{\inv_{wd_1}}}$.
Therefore, the algorithm cannot determine the attack execution is $\CECF$.
But indeed, the attack execution is not ECF, thus it cannot be $\CECF$.
\end{example}
We already saw in \Cref{Sec:Modularity} that due to state encapsulation, ECF is a modular property.
Therefore, the procedure may either check ECF for the entire execution, by searching for a cycle in the full IOC relation, or to check ECF for one object at a time.
A modular ECF check can be done by projecting the relation only on invocations of the object under examination.
To align with the actual implementation of the algorithm,
we chose to present it in its modular version here as well.
We give the complete pseudo-code of the algorithm in \Cref{fig:DynamicMonitorAlgorithm}.
It begins with an additional step of preprocessing which is calculating the commutativity matrix of all segments against all prefix and suffix segments of all their enclosing invocations (invocations that directly or indirectly call the invocation in which the segment is included).
The commutativity matrix assists in calculating the IOC relation.
We then iterate over all objects encountered in the execution, project the IOC relation on a single object in each iteration, and check if it has a cycle.
If the check returns that it is a DAG, then we verified the projected execution is ECF. Otherwise, the cycle describes the invocations which cannot be moved, and helps identify the callbacks that cause the violation of ECF.
\begin{figure}
\begin{lstlisting}[basicstyle=\ttfamily\scriptsize]]
CheckECFForAllObjects():
commute_matrix := CalculateCommutativityMatrix()
hbRelation := CalculateIOCRelation(commute_matrix)
for each unique object in execution:
if not CheckECF(object):
Print "Object " object " is not ECF"
CheckECF(object):
// It is guaranteed that IOC applies only to invocations of the same object
hbRelationO := project hbRelation on invocations of object only
return isDAG(hbRelation)
CalculateCommutativityMatrix():
matrix := new Map<Invocation, Segment -> Bool, Bool>
for each inv in invocations, segment in execution
if encloses(inv, segment)
prefix := (s for s in inv where s.I < segment.I)
suffix := (s for s in inv where s.I > segment.I)
prefixRS, prefixWS := Union(s.R for s in prefix), Union(s.W for s in prefix)
suffixRS, suffixWS := Union(s.R for s in suffix), Union(s.W for s in suffix)
prefixCommute := isCommutative(prefixRS, prefixWS, segment.R, segment.W)
suffixCommute := isCommutative(suffixRS, suffixWS, segment.R, segment.W)
if prefixCommute == False && suffixCommute == False
Abort("Not ECF")
matrix[inv, segment] := prefixCommute, suffixCommute
return matrix
encloses(inv, segment):
return inv.Obj = segment.Obj
&& segment.I between first segment and last segment in inv
CalculateIOCRelation(commute_matrix):
rel := new Map<Invocation, Invocation -> Bool>
for each inv1, inv2 in invocations
if encloses(inv1, first segment in inv2)
for each segment in inv2
if commute_matrix[inv1, segment] == False, True
rel[inv1, inv2] := True
if encloses(inv2, first segment in inv1)
for each segment in inv1
if commute_matrix[inv2, segment] == True, False
rel[inv1, inv2] := True
\end{lstlisting}
\caption{Algorithm for verifying ECF of an execution.
The code of \code{isDAG} and \code{isCommutative} is not given. The definition of \code{isCommutative} is given according to Definition \ref{def:CommutativeSegments}.
}\label{fig:DynamicMonitorAlgorithm}
\end{figure}
\subsection{Complexity}
\subsubsection{Time.}
The instrumentation step adds a constant factor of work to the runtime.
To analyze the algorithm, we begin by looking at the preprocessing steps first.
Let $n$ denote the number of invocations and $m$ the number of segments ($n<m$).
In addition, let $k$ denote the maximal number of object variables accessed in an object participating in the execution ($k<m$).
The \code{CalculateCommutativityMatrix} procedure loops on all invocations and all segments.
For each pair of an invocation and a segment, the \code{encloses} predicate can be implemented to take constant time.
The calculation of the prefix set and suffix set is taking time linear in the number of segments in an invocation, bounded by $m$.
The time to calculate the read and write sets of the prefix and the suffix set is linear in $k$.
Commutativity check, which involves checking set intersection, where our sets are implemented as associative arrays, is linear in $k$.
Thus the time of \code{CalculateCommutativityMatrix} is $O(nm(m+k))=O(nm^2)$.
For \code{CalculateIOCRelation}, we have a loop over pairs of invocations, and another pair of non-nested loops over segments in an invocation, giving $O(mn^2)$.
Projecting the IOC relation is linear in its size which is $O(n^2)$.
The \code{isDAG} check is linear in the size of the projected relation, which is bounded by $O(n^2)$. (The graph it represents has $O(n)$ vertices and $O(n^2)$ edges, and checking for a graph to be a DAG is $O(|V|+|E|)$).
In total, we have $O(nm^2)$.
\subsubsection{Space.}
The instrumentation adds $O(m)$ space for keeping the segments, and $O(nm)$ for keeping the invocations.
The commutativity matrix takes $O(nm)$ space,
and the IOC relation takes $O(n^2)$ space.
Therefore, the space complexity of the algorithm is $O(nm)$.
\section{Evaluation}\label{Sec:Evaluation}
We developed a prototype implementation for a dynamic monitor verifying ECF for \emph{Ethereum}.\footnote{The source code is available at \url{https://github.com/shellygr/ECFChecker}.}
For each execution, it checks if any of the participating contracts has a non-ECF (projected) execution, and outputs all detections of non-ECF executions.\footnote{The monitor was implemented on top of the Go~\cite{GoProgrammingLanguage} client for Ethereum, called \emph{geth}~\cite{Eth:Geth}, version 1.5.9.}
We ran our experiments by importing the entire blockchain from its inception on July 30, 2015 until March 30, 2017. \footnote{Without delving into the specifics of the blockchain paradigm, executions are organized in a structure called \emph{blocks}. Our primary experiment was to import the first 3,444,354 blocks of the main Ethereum blockchain.}
The host we used is a 64-bit Ubuntu 16.04 with two 2.2~GHz Intel Xeon E5-2699 processors (22 cores each with 2 threads per core) and 256~GB of RAM.
Both the instrumentation and the algorithm were integrated directly into the \emph{Ethereum Virtual Machine (EVM)} module using hooks, as described in \Cref{Sec:Dynamic}.
Our monitor operated in ``'detect-mode'' to avoid affecting the results, and for statistics gathering only.
However, it is trivial to change it to ``'prevent-mode'', that actively invalidates and reverts complete executions which are not ECF.
Had all the Ethereum clients used such a monitor by design, the DAO incident would have been avoided, along with the controversial hard fork.
As our experiments prove, the false-positive rate of the monitor is minuscule: only 10 executions out of about 100 million were legitimately non-ECF. There is also no concern of performance impact, as the measured overhead of running the monitor was less than $3.5\%$. Furthermore, the benchmarks of the monitor were performed in an ideal environment that actually makes the overhead larger than it is in a normal environment. The reason behind it is that normal environments have additional overheads such as networking and disk accesses, which we disabled in order to scale our experiments.
\paragraph{Experiments.}
In \Cref{fig:ECFGeneralExperiment}, we show a short list of experiments conducted.
We also included the number of contracts created as an additional metric of the blockchain.
The primary experiment was checking for ECF in all executions since the creation of blockchain until March 30, 2017.
Of note is that less than $0.01\%$ of the executions were non-ECF.
In the second experiment, we processed all executions starting from March 30, 2017 until June 23, 2017
\footnote{The second experiment processed all blocks from block no. 3,444,355 to block no. 3,918,380.}.
It is interesting to compare the results of the first experiment, conducted on a snapshot of the Ethereum blockchain taken in Mar. 30th, 2017, which we used for benchmarks, to the second experiment, in which we let our modified client to process all newer executions, until June 23rd, 2017.
The number of non-ECF contracts decreased in both absolute quantity and in percentage of total executions.
The newer executions expose the maturity of the network, expressed in both the number of total contracts created (almost $150\%$ more contracts created in less than 3 months than in the entire existence of the blockchain, from August 2015 till the snapshot date), and the number of executions.\footnote{Assuming a new block is generated at an almost constant rate, there were 32 executions per block on average in the second experiment, compared with 23 executions in the first.}
Moreover, the number of executions with callbacks increased significantly, indicating more complex contracts.
In the first experiment there were $128,670$ executions containing callbacks, and in the second experiment there were $155,668$ executions with callbacks.
This amounts to a $641\%$ increase in the number of callbacks in the later period compared with the earlier period.
While the percentage of executions with callbacks is still only $1\%$ of all executions,
the absolute number of executions with callbacks is large enough to indicate that callbacks are inevitable,
either because they are useful, or necessary.
This means that contracts show an increasing use of callbacks, and thus more complex code, that may be prone to bugs resulting from unintended interaction between contracts.
In both experiments, the overall percentage of non-ECF executions out of the executions with callbacks, was $1.17\%$, and less than $0.01\%$ out of all executions.
\begin{figure}
$$
\begin{array}{lrrrrr}
\textbf{Blockchain} & \textbf{Date} & \textbf{Contracts} & \textbf{Executions} & \textbf{Callbacks} & \textbf{Non-ECF (\%)} \\
\text{Ethereum} & \text{30.VII.2015-30.III.2017} & $138,457$ & $81,097,421$ & $128,670$ & $3,315$\, ($0.004\%$) \\
\text{Ethereum} & \text{30.III.2017-23.VI.2017} & $203,859$ & $15,311,650$ & $155,662$ & $6$\, ($<0.001\%$) \\
\text{Eth. Classic} & \text{30.VII.2015-29.VI.2017} & $91,191$ & $32,494,464$ & $81,731$ & $2,288$ \, ($0.007\%$) \\
\end{array}
$$
\caption{
Experimental results.
We use dates to mark the portion of the blockchain checked in the experiment.
The Contracts column shows how many contracts were created (but not necessarily executed) in the relevant time period.
The Executions column records the number of method invocations and the
Callbacks column shows how many of these invocations were callbacks.
Non-ECF column counts how many non-ECF executions were detected, and their percentage out of the total number of executions.
}
\label{fig:ECFGeneralExperiment}
\end{figure}
\paragraph{Discussion of non-ECF examples.}
We present a list of all contracts that demonstrated non-ECF executions in \Cref{fig:ECFContracts}.
Contracts C2, C4 are related to the DAO. C2 is the original DAO~\cite{Eth:DAOAttack}. C4 is known as `The Dark DAO'~\cite{Eth:DarkDAO}, an object containing a copy of the DAO's code, as created by the attack (The mechanism of the DAO was such that, every withdrawal of funds, manifested in the form of a new object whose code is a copy, or `split', of the DAO code).
Contract C1 is an unrelated contract which suffered a vulnerability very similar to the DAO's. The vulnerability, also stemming from non-ECF behavior, was discovered during a security audit and disclosed shortly before the attack on the DAO~\cite{Eth:MakerDAO,Eth:NonECFMention}.
Contract C5 is an exercise published on the blockchain to demonstrate the DAO attack~\cite{Eth:DAOExercise}, and indeed a non-ECF execution was detected.
In some contracts, it is difficult to pinpoint the exact cause of the existence of non-ECF executions, as the only available code is EVM bytecode, which is not trivial to analyze and reverse.
We tried to connect these incognito contracts with their creators or users.
With this approach, we found evidence that C3 is also related to the DAO~\cite{Eth:0x34a}.
The contract at C7~\cite{Eth:0xbf7Creator} was traced back to Validity Labs~\cite{Validity}.
We contacted the authors and they provided us with the Solidity source-code of the contract~\cite{Eth:Validity}.
It was deliberately designed to have a DAO-style callback exploit, and was used in their training workshops to demonstrate its dangers.
\begin{figure}
\begin{minipage}{0.5in}
\begin{footnotesize}
\begin{alltt}
\begin{tabbing}
Object C \\
Object Sender \\
Method call(data, sender) \\
if (Sender != nil) throw \\
Sender = sender; ret = this.do(data); Sender = nil \\
\\
Method do(data) \\
... // read Sender
\end{tabbing}
\end{alltt}
\end{footnotesize}
\end{minipage}
\caption{Pattern used by contracts C6, C8 and C9.
\code{Sender} is initialized to \code{nil}.
\code{call} is a method that throws when \code{Sender} is not \code{nil}, and otherwise sets it, calls method \code{do}, and nullifies \code{Sender} afterwards.
}\label{fig:Ambisafe}
\end{figure}
\begin{figure}
$$
\begin{array}{llrrrrr}
\textbf{Name} & \textbf{Contract address} & \textbf{Execs.} & \textbf{Execs. w. cbs.} & \textbf{Non-ECF} & \textbf{Stack depth} \\
\multicolumn{7}{c}{\textit{Ethereum Network (ETH)}} \\
\text{C1} & $\small{0xd654bdd32fc99471455e...}$ & $924$ & $143$ & $10$ & $3$ \\
\text{C2} & $\small{0xbb9bc244d798123fde78...}$ & $274,820$ & $103,064$ & $3,296$ & $2-146$ \\
\text{C3} & $\small{0x34a5451ef61a567ee088...}$ & $91$ & $8$ & $1$ & $46$ \\
\text{C4} & $\small{0x304a554a310c7e546dfe...}$ & $13,223$ & $2,812$ & 1 & $3$ \\
\text{C5} & $\small{0x59752433dbe28f5aa59b...}$ & $15$ & $6$ & $1$ & $3$ \\
\text{C6} & $\small{0x97361ea911d6348cf2af...}$ & $44$ & $42$ & $6$ & $2$ \\
\text{C7} & $\small{0xbf78025535c98f4c605f...}$ & $25$ & $22$ & $3$ & $3-9$ \\
\text{C8} & $\small{0x232f3a7723137ced12bc...}$ & $144$ & $142$ & $1$ & $2$ \\
\text{C9} & $\small{0x7c525c4e3b273a3afc4b...}$ & $35$ & $33$ & $2$ & $2$ \\
\multicolumn{7}{c}{\textit{Ethereum Classic Network (ETC)}} \\
\text{C1} & $\small{0xd654bdd32fc99471455e...}$ & $850$ & $143$ & $10$ & $3$ \\
\text{C2} & $\small{0xbb9bc244d798123fde78...}$ & $195,428$ & $86,573$ & $805$ & $2-146$\\
\text{C3} & $\small{0x34a5451ef61a567ee088...}$ & $18$ & $9$ & $1$ & $46$ \\
\text{C4} & $\small{0x304a554a310c7e546dfe...}$ & $14,150$ & $3,064$ & $1$ & $3$\\
\text{C10} & $\small{0xf4c64518ea10f995918a...}$ & $428$ & $177$ & $11$ & $42-122$\\
\text{C11} & $\small{0xb136707642a4ea12fb4b...}$ & $2,582$ & $305$ & $201$ & $17-20$\\
\text{C12} & $\small{0x0e0da70933f4c7849fc0...}$ & $5,330$ & $3,992$ & $1,259$ & $12-57$
\end{array}
$$
\vspace*{-3mm}
\caption{A sample of interesting Non-ECF contracts in Ethereum.
Contracts are given a name $\text{C1},\ldots,\text{C12}$, and are ordered chronologically, by the date of the first non-ECF execution.
The Executions and Executions with callbacks columns show statistics on usage style.
The Non-ECF column shows how many executions were detected as non-ECF.
Stack depth column indicates the range of the depths of the non-ECF subexecutions.}
\label{fig:ECFContracts}
\end{figure}
\begin{figure}
$$
\begin{array}{llll}
& \textbf{Time} & \textbf{Memory (max)} & \textbf{Memory (end)} \\
\text{Monitor off} & $16h 17m$ & $5.5GB$ & $803MB$ \\
\text{Monitor on} & $16h 50m$ \, ($3.38\%$ \text{ overhead}) & $5.5GB$ \, ($0\%$) & $940MB$ \, ($17\%$ \text{ overhead}) \\
\end{array}
$$
\vspace*{-3mm}
\caption{Performance statistics. Benchmark experiment was importing the Ethereum main network blockchain, from its creation in July 30, 2015 until March 30, 2017. Compares the import with monitor on or off.}
\label{fig:Performance}
\end{figure}
The same high-level Solidity code of contracts C6, C8, and C9, were provided to us by their creators at Ambisafe~\cite{Eth:Ambisafe}.
The pattern used by these contracts gives rise to behaviors that are purposefully non-ECF.
We show a snippet illustrating the pattern in \Cref{fig:Ambisafe}.
This pattern is inherently non-ECF.\footnote{In the formal definition, it actually is ECF, because a call of a contract to itself is not a callback.
The contracts under examination were discovered due to a deviation of our monitor's implementation from the full definition of ECF. However, this example can be fitted into a slightly modified pattern which is not-ECF even according to the full definition, by adding an intermediary contract between \code{call} and \code{do}.}
The method \code{do} assumes the value of \code{Sender} is not \code{nil}, but this only occurs in the context of an invocation of \code{call}.
The purpose behind this behavior, is to have \code{Sender} act as a lock, protecting against unexpected callbacks.
Such a design may be avoided in presence of a monitor that allows only ECF executions.
The bottom part of the table in \Cref{fig:ECFContracts} shows non-ECF contracts found in~\citet{Eth:ETHClassic}.
Ethereum Classic (or ETC) is the continuation of the original Ethereum blockchain following the controversy of the hard-fork due to the DAO bug.
Until July 20, 2016, both blockchains, Ethereum and Ethereum Classic, contain the same executions, and thus the same non-ECF executions.
Our result and investigation show that all non-ECF executions discovered in the Ethereum Classic network are of copies of the DAO~\cite{Eth:DAOinETC}.
Generally, it is important to stress that: (1) there may be other non-ECF contracts, as crafting and deploying contracts that exploit non-ECF entails investment of real money, thus requires a strong incentive to do so; (2) attacking is harder as Ethereum employs (not bullet-proof) heuristics to limit callbacks; (3) a better playground may be the Ethereum TestNet on which we did not run the experiment, but may provide insight as a future work.
The actual overhead measured by enabling the ECF monitor is given in \Cref{fig:Performance}.
We used the first experiment, where the blocks were imported, as a benchmark.
Normally, there is an additional overhead of network download times, which can vary significantly.
The measured overhead is about $3.5\%$, when calculating the difference in time of importing the blockchain with the monitor off, and importing it with the monitor on.
We believe the actual overhead is even smaller in most realistic scenarios.
First, most clients import the blockchain using the network, which may cause unexpected latencies, unrelated to the monitor.
Additionally, the process was pointed to a directory created on a 200~GB RAM disk to improve the scalability of the experiment.\footnote{The Ethereum blockchain suffered a DoS attack~\cite{Eth:DoSAttack,Eth:DoSAttack2} affecting the blockchain in the range of block numbers 2.2M-2.7M, causing all peers participating in the blockchain to make frequent accesses to disk. Running on a RAM disk was necessary to minimize the runtime of experiments.}
Most clients use a physical disk and not a RAM disk.
Even if the physical disk is an SSD drive, the experiment slows down significantly, and takes about 20h (18\% more than with a RAM disk).
The additional memory footprint measured in the end of the import is about $140MB$, or $17\%$.
It should be noted, that as the implementation is written in Go, which includes automatic garbage collection, the memory consumption varies between tests. The relative difference with the monitor on or off was consistent across repeated tests.
The maximal memory used by the process is $5.5GB$ and is not related to the monitor.
High memory consumption occurred during the processing of one of the DoS attacks on the blockchain.
\section{Introduction}
The theme of this paper is enabling modular reasoning about the
correctness of objects with encapsulated state.
This is inspired by platforms like Ethereum~\cite{Eth:EthYellowPaperGavWood}
that facilitate execution of \emph{Smart Contracts}~\cite{Szabo1997formalizing} on top of a blockchain-based distributed ledger~\cite{nakamoto2008bitcoin}.
A key property in Ethereum Smart Contracts is the lack of global
mutable shared state, in contrast to common standard programming
environments such as C and Java.
A smart contract is analogous to an object with encapsulated state.
However, the Ethereum blockchain, and many other dynamic environments, implement event-driven
programming using callbacks.
These callbacks are necessary for functionality, but can compromise security.
For example, the famous bug in the DAO contract exploited callbacks to steal \$150M~\cite{Dao}.
Indeed, callbacks may break modularity which is essential for good programming style and extendibility.
In the context of Blockchain, modularity is even more important since contracts are contributed by different sources,
some of which may be malicious.
Accordingly, the bug in the DAO allowed an adversarially crafted contract to mutate the DAO's state by calling back to it.
The DAO contract, that implemented a crowd-funding platform, was attacked by a `callback loop-hole' (to be precisely described below).
This attack, the recovery from which required a controversial hard-fork of the blockchain,\footnote{A hard fork can be thought of as taking an agreed history of transactions, and manually change it.}
exhibits a vulnerability that is peculiar to
decentralized consensus systems, like Ethereum:
in such systems, a buggy contract cannot be updated or fixed
(except for extreme measures like hard-forking),
which makes validation and verification of smart contracts
of even greater importance for this application.
\paragraph{Effectively Callback-Free Objects}
We identify a natural generic correctness criteria for objects
which enables modular reasoning in environments with local-only mutable states,
and expect most correct objects to satisfy this requirement.
Informally, if an object $o$ calls another object $o'$, and the execution of $o'$ calls $o$ again, this second call to $o$ is defined as a callback.
The main idea is to allow callbacks in $o$ only when they cannot affect the serial
non-interruptible behavior of $o$.
Thus, such callbacks can be considered harmless and do not affect the set of local reachable states
of the object $o$.
In particular, the behavior of such objects is independent of the client environments and of other objects.
It is possible to reproduce all behaviors of the object using a most general client and without analyzing external objects.
We say that an execution is
\textbf{Dynamically Effectively Callback Free} (\DECF{}) when there exists ``an
equivalent'' execution without callbacks which
starts in the same state and reaches the same final state.
By \emph{equivalent}, we refer to the behavior of a particular object
as an external observer may perceive.
We say that an object is
\textbf{Statically Effectively Callback Free} (\SECF{}) when all its possible
executions are dynamically ECF.
We do not distinguish between dynamic and static ECF when the context is clear.
Both definitions are useful.
Dynamic ECF in particular is applicable to the blockchain environment,
since static ECF is undecidable in the general case.
We ran experiments on Ethereum, proving that checking dynamic ECF is inexpensive,
and thus can be done efficiently in-vivo.
This, combined with Ethereum's built-in rollback feature,
would have allowed to prevent the DAO bug from occurring,
without invalidating legitimate executions.
(In fact, we found just one such legitimate non-ECF contract, discussed in \Cref{Sec:Evaluation}).
We show that the
vulnerable DAO contract is non-ECF while no non-ECF executions are detected
after applying the suggested corrections to it.
Notice that the ECF notion is similar to the notion of atomic
transactions in concurrent systems.
Indeed, despite the fact that contract languages do not usually
support concurrency, modularity and callbacks require similar kind
of reasoning.
The ECF property's usefulness is not limited to bug-finding;
once ECF is established, it can be served to simplify reasoning on the object in isolation of other objects:
We show that the set of reachable local states in ECF objects can
be determined without considering the code of other objects and
thus enable modular reasoning. This modular reasoning can be
performed automatically using abstract interpretation e.g., as
suggested in \citet{LogozzoCLSS09} or by using deductive verification
which is supported by Dafny~\cite{Leino2010}.
We demonstrate this by
verifying an interesting invariant of the DAO contract. (See \Cref{Se:Overview}).
\paragraph{Online Detection of ECF executions}
A na{\"i}ve detection of \DECF{} may be costly because of the need to
enumerate subexecution traces.
Therefore, we develop an effective polynomial online algorithm for checking if an
execution is ECF.
The main idea is to detect conflicting memory accesses and utilize commutativity in an effective manner.
We integrated the algorithm into the \emph{Ethereum Virtual Machine (EVM)}~\cite{Eth:EthYellowPaperGavWood}.
We ran the algorithm on all executions kept in the Ethereum blockchain until 23 June 2017, and demonstrate that:
(i)~the vulnerable DAO contract and other buggy contracts are non-ECF.
(ii)~very few correct contracts are non-ECF,
(iii)~callbacks are not esoteric and are used in many contracts,
and (iv)~the runtime overhead of our implementation is negligible and
thus can be integrated as an online check.
This online detection can thus be used to prevent incidents like the theft from the DAO
at the cost of slightly more restricted form of programming.
As far as we are aware, our tool is also the most precise and effective tool
for finding such vulnerable behaviors due to callbacks.
We compared it to the Oyente tool~\cite{RW:LuuCCS16,Eth:OyenteOnline}, by giving it both ECF and non-ECF contracts based on the DAO object (\Cref{Fi:DaoContract}).
We found that it has false positives, as it detects a `reentrancy bug' (the common name of the DAO vulnerability in the blockchain community) for any one of the fixes that render our example contract ECF.
\paragraph{Decidability of \SECF{} for objects}
We also consider the problem of checking \SECF{} algorithmically.
Obviously, since modern contract languages, such as Solidity~\cite{Eth:Solidity}, are
Turing complete, checking if a contract is ECF
is undecidable.
We show that checking that a contract is \SECF{} in a language with finite local states
is decidable.
This is interesting since many contracts only use small local states or maps with uniform data
independent accesses.
Technically, this result is non-trivial since the nesting of contract calls is unbounded,
and since ECF requires reasoning about permutations of nested invocations.
The reason for the decidability is that non-ECF executions which occur in high depth of nesting
must also occur in depth $2$.
\paragraph{Main Results}
Our results can be summarized as follows:
\begin{enumerate}
\item We define a general safety property, called ECF, for objects (\SECF{}) and executions (\DECF{}).
Our definition is inspired by the Blockchain environment but it may also be useful for other environments with encapsulated states, such as Microservices.
\item We show that objects with encapsulated data, under the assumption that they satisfy ECF, can be verified using modular reasoning in a sound manner.
\item A stronger notion of ECF, based on conflict-equivalence,
enabling efficient verification of \DECF{} in real-life environments,
and for which \SECF{} is decidable for programs with finite state and unbounded stack.
\item A polynomial time and space algorithm for online checking of \DECF{}
and prototype implementation of it as a dynamic monitor of \DECF{}, built on top of an Ethereum client.
\item Evaluation of the algorithm on the entire history of the Ethereum blockchain (both main and `Classic' forks, see \Cref{Sec:Evaluation}).
The monitor detects true bad executions (the infamous DAO and others) as non-ECF, and has near-zero false positives.
Based on this result, it can be inferred that, in practice, most non-ECF executions correspond to bad executions.
We also show that our monitor has a very small runtime overhead.
By retroactively running the \DECF{} monitor on the available history,
we were able to prove its effectiveness in preventing the exploitation of the vulnerability in the DAO,
and even more importantly,
the feasibility of leveraging it in other applications, e.g., simplifying
modular contract verification.{}
\end{enumerate}
\section{Execution Equivalence}\label{Se:EqExec}
We define two notions of equivalence of \emph{runs} (sequences of \emph{complete} executions) with respect to a given (arbitrary) object $\objectid$: \emph{final-state equivalence} and \emph{conflict equivalence}.
\subsection{Object-Final-State Equivalence}
\ignore{
\begin{definition
Executions $\exec_1$ and $\exec_2$ are \emph{final-state equivalent} if they start in the same state, and finish in the same state:
$$\exec_1 \fseq \exec_2 \iff \src(\exec_1) = \src(\exec_2) \land \trg(\exec_1)=\trg(\exec_2)$$
\end{definition}
}
\begin{definition}\label{De:obj-fin-eq
Runs $\ocode_1 \inctx \exec_1$ and $\ocode_2 \inctx \exec_2$ are \emph{object-final-state equivalent for an object $\objectid$},
denoted by $\ocode_1 \inctx \exec_1 \fseqo{\objectid} \ocode_2 \inctx \exec_2$,
if their respective first and last states have the same store for $\objectid$, and their code contexts agree on~$\objectid$:
$$
\begin{array}{l}
\ocode_1 \inctx \exec_1 \fseqo{\objectid} \ocode_2 \inctx \exec_2 \iff {}\\
\qquad \SelStore(\src(\exec_1))(\objectid) = \SelStore(\src(\exec_2))(\objectid) \land \SelStore(\trg(\exec_1))(\objectid)=\SelStore(\trg(\exec_2))(\objectid) \land \ocode_1(\objectid) = \ocode_2(\objectid)
\end{array}
$$
\end{definition}
\begin{remark}
Note that in \Cref{De:obj-fin-eq} it may be that $\ocode_1 \neq \ocode_2$ as long as $\ocode_1$ and $\ocode_2$ map $\objectid$ to the same code.
\end{remark}
\subsection{Object Conflict-Equivalence}
\paragraph{Conflicts}
Two primitive commands $\action$ and $\action'$ \emph{conflict},
denoted by $\Conflict(\action,\action')$,
if both access the same field and at least one of these accesses is a write.
\begin{remark}
Recall that we assume that object states are comprised of integer-typed data members (fields); thus it is possible to detect the heap locations they access from the (syntactic) primitive commands that they execute.
Furthermore, recall that different objects contain different fields and an object cannot directly access the fields of another; thus if two primitive commands conflict then they must be executed by the same active object.
\end{remark}
\paragraph{Modular Well-Formed Executions}
We extend the definition of well-formed executions to \emph{modular} well-formed executions, which allow to
replace the subexecution resulting from a method invocation $\SCallC{x}{\objectid'}{e}$ by an assignment of arbitrary value to $x$.
We refer to such an (implicit) transition as a \emph{havoc transition}.
Intuitively, a havoc transition allows to safely overapproximate the only effect that an object $\objectid$ may observe from the invocation of a method on an object $\objectid'$.
\begin{definition
A \emph{havoc transition}, denoted by $\sstate\rightsquigarrow\sstate'$, is a pair of states $\B{\sstate,\sstate'} \in \State \times \State $ such that
$\sstate=\defostate{\objectid}{\SCallC{x}{\objectid'}{e}}{\reg}{\store}$ and
$\sstate'=\defostate{\objectid}{\cdone}{\reg[\res \mapsto n]}{\store}$ for any values of $\objectid$, $\objectid'$, $e$, $\reg$, $\Stk$, $\store$ and $n$.
\end{definition}
\begin{definition
A \emph{modular well-formed execution} is a finite sequence of transitions coming from $\Tr$ such that
for any consecutive transitions $\transition_1\transition_2$ it contains,
the target state of $\transition_1$ and the source state of $\transition_2$ are either
(i) equal, i.e., $\trg(\transition_1)=\src(\transition_2)$,
or (ii) induce a havoc transition, $\trg(\transition_1) \rightsquigarrow \src(\transition_2)$.
By abuse of notations, we use $\exec$ to denote modular well-formed executions too.
A sequence of transitions $\exec$ is a \emph{complete modular well-formed run} if it is modular well-formed and its first and last states are quiescent.
Such a run is a \emph{complete modular well-formed execution} if, in addition, all other states are active.
\end{definition}
\paragraph{Projected Executions}
\changed{We define what is the projection of an execution on an object, called \emph{projected execution}.
The definition readily generalizes to any sequence of transitions.}
\begin{definition
Let $\exec$ be an execution.
The \emph{projected execution of $\exec$ on $\objectid$}, denoted by $\exec|_\objectid$,
is the subsequence of $\exec$ comprised of the transitions whose active object is $\objectid$.
\end{definition}
\begin{lemma}
For any execution $\exec$ and object $\objectid$ it holds that $\trace(\exec|_\objectid)=\trace(\exec)|_\objectid$.
\end{lemma}
\changed{
It will be useful in the following sections to consider modular well-formed executions yielded by a projection on an object $\objectid$.
The following proposition states that it is possible to reverse that projection, namely, to find a minimal well-formed execution that, when projected, yields the modular well-formed execution we started with.
}
\begin{proposition}\label{Prop:ModularToWellFormed
Let $\ocode \inctx \exec$ be a modular well-formed execution where all transitions have the same active object $\objectid$.
Then there is a context $\ocode'$ such that $\ocode'(\objectid)=\ocode(\objectid)$
and an execution $\ocode' \inctx \exec'$
such that $\trace(\exec)$ is a subsequence in $\trace(\exec')$
and $\ocode' \inctx \exec'$ is well-formed.
\changed{In addition, such an execution $\ocode' \inctx \exec'$ is minimal in the sense that $\trace(\exec|_{\objectid})=\trace(\exec'|_{\objectid})$}
\end{proposition}
\changed{
Finally, we present the definition for execution conflict equivalence with respect to an object $\objectid$.
}
\begin{definition
Let $\ocode_1 \inctx \exec_1$ and $\ocode_2 \inctx \exec_2$ be modular well-formed runs and
$\trace_1= \trace(\exec_1|_\objectid)$ and $\trace_2= \trace(\exec_2|_\objectid)$ be their the traces of their respective projections on $\objectid$.
$\ocode_1 \inctx \exec_1$ and $\ocode_2 \inctx \exec_2$ are \emph{object conflict-equivalent for an object $\objectid$}, denoted $\exec_1 \ceq^{\objectid} \exec_2$, if:
\begin{enumerate}[(i)]
\item
$\ocode_1(\objectid) = \ocode_2(\objectid)$,
\item
$|\trace_1| = |\trace_2|$,
\item
there exists a permutation
$\cep:\{ 1..|\trace_1|\}\rightarrow\{ 1..|\trace_1|\}$
such that:
\begin{enumerate}
\item for any $i\in\{1..|\trace_1|\}$ it holds that $\trace_1(i)=\trace_2(\cep(i))$,
and
\item for any $i,j\in\{1..|\trace_1|\}$, if $\Conflict(\action(\trace_1(i)), \action(\trace_1(j)))$
then $i<j \iff \cep(i)<\cep(j)$.
\end{enumerate}
\end{enumerate}
\end{definition}
\section{Overview}
\label{Se:Overview}
\newcommand{\icredit}{\textit{credit}}
\newcommand{\ibalance}{\textit{balance}}
This section provides some necessary background and an informal overview of our approach.
\subsection{The DAO Bug}
\Cref{Fi:DaoContract} shows pseudocode illustrating the vulnerability in the DAO.\footnote{\emph{DAO} is acronym for \emph{decentralized autonomous organization}, and its purpose is to facilitate voting on proposals and on investments by the owners of the DAO.}
The contract stores a \code{credit} for each object, as well as the current \code{balance}.\footnote{In
programming languages like Solidity, balance is a predefined field of every contract, maintained by the runtime system. We write it explicitly for clarity.}
The \code{credit} represents individual investments per object.
To align with the Ethereum terminology, the unit of currency represented by \code{credit} and \code{balance} is called \emph{ether}.
The contract maintains a representation invariant, where the sum of the credits equals to the current balance,
i.e.,
\begin{equation}
\Sigma_{o\in\dom(\icredit)} \icredit[o] = \ibalance
\label{eq:DaoInvariant}
\end{equation}
The contract offers two methods for manipulating states:
\texttt{deposit} for depositing money and \texttt{withdrawAll} for
withdrawing all available funds of a specific object.
\begin{figure}
\begin{minipage}{1in}
\begin{footnotesize}
\begin{alltt}
\begin{tabbing}
XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=\kill
Object DAO \+\\
Map<Object, int> credit\\
int balance\\
\textbf{Invariant} \textit{(sum o: credit[o]) = balance}\\
\\
Method withdrawAll(Object o) Method deposit(Object o, int amount) \+\\
2: if (oCredit > 0) 6: credit[o] += amount \+\\
// 2.5: credit[o] = 0 7: this.balance += amount\\
3: this.balance -= oCredit \\
4: o.pay(oCredit) \\
5: credit[o] = 0 \-\\
\end{tabbing}
\end{alltt}
\end{footnotesize}
\end{minipage}
\vspace*{-5mm}
\caption{\label{Fi:DaoContract
A contract illustrating the DAO bug.
The representation invariant may be violated by callbacks from malicious contracts.
Line $2.5$ fixes the bug.}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[t]{2.5in}
\centering
\begin{footnotesize}
\begin{alltt}
\begin{tabbing}
Object GoodClient \\
Object Dao, int balance \\
Method init(Object dao) \\
1: this.Dao = dao \\
Method pay(int profit) \\
2: this.balance += profit \\
Method depositCredit(\=Object dao, int amount) \\
3: Dao.deposit(this, amount) \\
Method getCredit(Object dao) \\
4: Dao.withdrawAll(this)
\end{tabbing}
\end{alltt}
\end{footnotesize}
\caption{\label{Fi:DaoGoodClient
An innocent client using the \code{DAO} object without violating its representation
invariant.}
\end{subfigure}
\quad
\begin{subfigure}[t]{2.5in}
\centering
\begin{footnotesize}
\begin{alltt}
\begin{tabbing}
Object Attacker \\
Object Dao, bool stop, int balance\\
Method init(Object dao) \\
1: Dao = dao \\
2: stop = false \\
Method pay(int profit) \\
3: this.balance += profit\\
4: if (!stop) \\
5: stop = true \\
6: Dao.withdrawAll(this) \\
7: stop = false
\end{tabbing}
\end{alltt}
\end{footnotesize}
\caption{\label{Fi:DaoAttack
A snippet of an \code{Attacker} object. It is stealing money from the \code{DAO} object by violating its representation
invariant.}
\end{subfigure}
\caption{An innocent and a malicious client using the \code{DAO} object}\label{Fi:DaoClients}
\end{figure}
\input{daoTrace}
\Cref{Fi:DaoGoodClient} shows a simple client illustrating the expected usages of the \code{DAO} object.
\Cref{Fi:DaoAttack} shows a simple attack on the \code{DAO} object.
The code callbacks to the \code{DAO} method \texttt{withdrawAll} to steal money.
\Cref{Fi:StealTrace} depicts a concrete trace of attacking the \code{DAO} assuming that the \code{DAO}'s initial balance is $200$ ether.
We reached that state after a \code{GoodClient} object and an \code{Attacker} object deposited each $100$ ether.
In the first call to \code{withdrawAll}, the attacker will get the amount he invested originally in the attack ($100$ ether).
The \code{DAO} then calls to the \code{Attacker} object's \code{pay} method, which increases the attacker's balance by $100$ ether, and calls \code{withdrawAll} again.
The \code{pay} method is designed to call \code{withdrawAll} at most once in a trace by updating the \code{stop} variable, and avoid infinite recursion.\footnote{For clarity, we avoid technical discussion of the semantics of executions and exceptions in Ethereum/Solidity, to allow us to focus on the ECF property.}
The code of \code{withdrawAll} in the second run will transfer an additional $100$ ether from the \code{DAO} object to the attacker.
In the end of the trace, the \code{DAO} was depleted of its funds completely, and the attacker managed to illegitimately receive the funds that belonged to \code{GoodClient}.
\subsection{Effectively Callback Free Contracts}
In principle, semi-automatic program verification and abstract interpretation can be used to verify the absence of malicious attacks
like the one in the \texttt{Attacker} object.
However, this requires reasoning about the whole code.\footnote{In the case of Ethereum, it is in fact impossible to reason about the whole code, as new contracts can be added at any time, and these contracts could interact with the contract being checked.}
This paper advocates a different solution by exploring modularity.
The idea is to require stronger conditions from the contracts which prevent the need to reason about other objects at all.
Specifically, we define the notion of \textbf{effectively
callback free} (ECF) objects.
Our definition is inspired by Blockchain contracts but is applicable to enforce modularity in other environments with local states.
We say that an execution of an object with an initial state $s_0$ and final state $s$ is
\textbf{Dynamically Effectively Callback Free} (\DECF{}) when there exists ``an
equivalent'' execution of the contract without callbacks which
starts in the same initial state $s_0$ and
reaches the final state $s$.
We say that an object is
\textbf{Statically Effectively Callback Free} (\SECF{}) when all its possible
executions are effectively callback free.
The \texttt{DAO} object is not ECF. For example, the trace depicted in \Cref{Fi:StealTrace} cannot be reproduced without callbacks to reach the same state.
In contrast, the fix to the \texttt{DAO} object by uncommenting line $2.5$ and deleting line $5$ makes the contract ECF.
This contract is now ECF since all its traces can be reordered to avoid callbacks.
For example, \Cref{Fi:FailedStealInECF} shows a trace of an attempt to perform an attack similar to the attack in \Cref{Fi:StealTrace} and its corresponding reordering that avoids callbacks.
Note, that in the reordered trace, \code{withdrawAll} did not execute line $4$.
Omitting calls is allowed for the sake of proving an execution is ECF,
as our goal is to be able to reproduce, assuming there are no callbacks, the same behaviors that are feasible with callbacks.
\subsection{Online ECF Detection}
It is possible to check in a na{\"i}ve way that an execution is ECF by recording the trace and checking the ECF property at the end of the execution, by enumerating all possible permutations.
However, this is costly both in space and in time, since the number of permutations grows exponentially with the size of the trace.
In particular, it is hard to see if such a solution can be integrated into a virtual machine.
In order to obtain a feasible online algorithm, we check a stronger requirement than ECF, which is inspired by conflict serializability of database transactions.
The main idea is to explore commutativity of operations for efficient online checking of a correctness condition which guarantees that the callback-free trace results in the same state as the original trace.
Consider a trace $\exec$ with potential callbacks and a reordered trace $\exec'$ which does not include callbacks.
$\exec'$ is not necessarily feasible, unless we permit to ignore external calls by objects and force the clients to perform these calls instead.
We say that $\exec$ and $\exec'$ are
conflict equivalent
if every pair of conflicting read/write operations in $\exec$ appear in the same order in $\exec'$.
Operations conflict when they are not commutative.
Commutativity is mechanically checked by comparing the read and the write sets of operations, and forbidding intersection of read/write conflicts.
For example, in \Cref{Fi:StealTrace}, the read operation of \code{D.c[A]} in the \code{withdrawAll} action labeled 1 (lines 1-3) does not commute with the write operation of \code{D.c[A]} in the \code{withdrawAll} action labeled 5 (line 5).
However, the top trace in \Cref{Fi:FailedStealInECF}, depicting a trace of the ECF version of \code{DAO}, the operations in the \code{withdrawAll} action labeled 3 (lines 1-2) commute with the operations in the \code{withdrawAll} action labeled 6, which has an empty read and write sets (no code was executed since line 5 was deleted).
The information regarding commutativity of different subtraces is used to build a constraint graph on the ordering of object invocations.
When this constraint graph contains no cycles, it is possible to perform topological sort to find a concrete callback-free trace.
A full description of the algorithm and its complexity is available in \Cref{Sec:Dynamic}.
We integrated this algorithm into the EVM (the \emph{Ethereum Virtual Machine}) and applied it to all available executions in the blockchain.
The results are summarized in \Cref{Sec:Evaluation}.
They indicate that the vast majority of non-ECF executions come from erroneous contracts.
They also indicate that the runtime overhead of our instrumentation is neglectable.
From these encouraging results, we concluded that if the ECF check was part of the Ethereum protocol,
it could have prevented the vulnerability in the \code{DAO} from being exploited.
Its clearly beneficiary for an environment like Ethereum, which handles sensitive financial transactions,
and in which code is virtually impossible to upgrade.
\subsection{Deciding ECF Contracts}
We also investigated the possibility to verify at compile-time that a contract is ECF (the \SECF{} property).
In general, this is undecidable, since languages such as \emph{Solidity}, a high-level front-end to EVM bytecode, are Turing-complete.
However, we show that for contracts with finite local states, checking ECF is decidable.
This result is non-trivial as the model allows for an unbounded stack length.
The decision procedure devised provides insight on additional techniques for checking ECF in practice.
\subsection{Verifying Properties of ECF Contracts}
In this paper, we show that reasoning about ECF contracts can be performed
in a modular fashion.
The local reachable states of an ECF contract are only affected by the
code of the contract,
and cannot be changed by external contracts.
This is useful for program verification and program analysis, treating
external calls
as
non-deterministic operations that may return an arbitrary value, but cannot change the local state.
We utilized this property using Dafny~\cite{Leino2010}, to verify correctness of the revised \code{DAO} object from \Cref{Fi:DaoContract} (including line 2.5, excluding line 5).
When doing so, we ignored the call in line 4, because the return value was not used.
We provide a deeper discussion on verifying this example using Dafny in \Cref{Sec:Modularity}.
\subsection{Summary of the Rest of This Paper}
The paper is organized as follows:
In \Cref{Sec:Preliminaries} we formally present the syntax and semantics of our programming language for contracts,
called $\PLname$.
Notions of equivalence are presented in \Cref{Se:EqExec}.
The ECF property and its different `flavors' (dynamic vs. static, and for different equivalence notions) are presented in \Cref{Sec:CorrectnessConditions}.
We discuss decidability results for ECF in \Cref{Sec:Decidability}.
\Cref{Sec:Modularity} shows the application of the ECF property to achieve modular object-level analysis.
The algorithm for online verification of \DECF{} is given in full in \Cref{Sec:Dynamic}.
We discuss our experimental results obtained by running the algorithm on the Ethereum blockchain in \Cref{Sec:Evaluation}.
Related work is provided in \Cref{Sec:RelatedWork} and we conclude in \Cref{Sec:Conclusion}.
\section{Programming Language}\label{Se:PL}\label{Sec:Preliminaries}
We formalize our results for \(\PLname\), a simple imperative object-based programming language with
pass-by-value parameters with integer-typed local variables and data members (fields).
For simplicity, and without loss of generality, every method has a single formal parameter named $\inarg$ and
returns a value by assigning it to a designated variable $\ret$.
Even though we present our theoretical development for contracts in $\PLname$, for readability we use
a Java-like notation in our examples, which can be easily desugared.
\begin{figure}
\[
\begin{array}{lclcl}
c & \in & \PCmd & \eqdef & \SAssign{x}{e} \mid \SAssign{F}{x} \mid \SAssign{x}{F} \mid \SAssert{b} \mid
\SCallC{x}{\oid}{e} \mid \SSkip \mid \SEnterV \mid \SReturnV
\\
C & \in & \Cmd & \eqdef & c \mid C \, ; \, C \mid \code{if}\,\,\,b\,\,\,\code{then}\,\,C \,\,\code{else}\,\, C \mid \code{while}\,b\,\code{do}\,\,C
\\
K & \in & \Contract & \eqdef & \cid \colon \overline{f} \, \,\, \, \SEnterV\, \,\code{var}\,\,\overline{x}\,\,\, \, C \,\,\,\, \SReturnV
\end{array}
\]
\caption{Syntax.
}
\label{fig:Syntax}
\end{figure}
\subsection{Syntax}\label{Sec:Syntax}
\Cref{fig:Syntax} defines the syntax of $\PLname$.
We assume infinite syntactic domains of
$\cid\in\CID$, $f \in \FID$, and $x\in\VID$
\emph{contract identifiers}, \emph{field names}, and \emph{variable identifiers}, respectively.
A contract $K$ is identified by a (unique) \emph{contract identifier} $\cid$,
and contains a sequence of \emph{field} definitions $\overline{f}$
and a (single) nameless \emph{contract method}.
The contract method is comprised of a sequence of \emph{local variable} definitions $\overline{x}$ and a command $C\in \Cmd$.
$C$ may be a \emph{primitive command} $c \in \PCmd$
or a \emph{compound command}, i.e., a sequential composition of commands, a conditional, or a loop.
A primitive command $c\in\PCmd$
may be either an assignment of an expression $e$ to a local variable $x$ ($\SAssign{x}{e})$,
an assignment of the value of a local variable $x$ to a field $F$ ($\SAssign{F}{x}$),
an assignment of the value of a field $F$ to a local variable $x$ ($\SAssign{x}{F}$),
an \emph{assert} command ($\SAssert{b}$),
a call to a contract method with a single argument $e$,
keeping the returned value in a local variable $x$ ($\SCallC{x}{\oid}{e}$),
or a $\SSkip$ command.
Each contract has a single method, thus methods are not named, and may be colloquially referred to using the name of their contract.
Without loss of generatility, we assume that no two contracts contain a field with the same name.
In the following, we use the terms `contract' and `object' interchangeably.
\begin{figure}
$
\centering
\begin{array}{ll}
\begin{array}{l@{\,}l@{\,}l@{\,}c@{\,}ll}
\reg & \in & \Reg & = & \VID \partialto_{\mathit{fin}} \mathbb{Z} & \text{Local states} \\
\oovar & \in & \ObjVar & = & \FID \rightarrow_{\mathit{fin}} \mathbb{Z} & \text{Object states}\\
\stk & \in & \Frame & = & \CID \times \Cmd \times \Reg & \text{Frames}\\
\end{array}
&
\begin{array}{l@{\,}l@{\,}l@{\,}c@{\,}ll}
\Stk & \in & \Stack & = & \overline{\Frame}& \text{Stacks}\\
\store & \in & \Store & = & \CID \partialto_{\mathit{fin}} \ObjVar & \text{Stores} \\
\sstate & \in & \State & = & \Stack \times \Store & \text{States}\\
\end{array}
\end{array}
$
\caption{Semantic domains.\label{fig:Semantic-domains}}
\end{figure}
\subsection{Semantics}\label{SubSec:Semantics}
$\PLname$ has a rather mundane stack-based operational semantics, which handle method calls using a \emph{stack} of activation records (\emph{frames}), and uses a \emph{store}
to record the values stored in object fields.
We refer to a state in which the stack is empty as a \emph{quiescent} state and to a non-quiescent state as an \emph{active} state.
Once the execution reaches a \emph{quiescent} state, \emph{any} object method may start running.
We refer the reader's attention to three important points:
\begin{inparaenum}[(i)]
\item contract states are encapsulated: A contract $\oid$ can only access its own fields,
\item local variables are private to their invocation, and
\item \label{I:deterministic} once a contract method is invoked, the semantics is deterministic.
\end{inparaenum}
We denote the \emph{code context} (\emph{context} for short) which provides the code of every called object by $\ocode\in\CID\partialto_{fin}\Cmd$, i.e., $\ocode(\objectid)$ denotes the code of object $\objectid$.
\paragraph{States}
\Cref{fig:Semantic-domains} defines the semantic domains.
A \emph{state} $\sstate=\ostatedefault$ is a pair comprised of a (possibly empty) \emph{stack} of \emph{frames} $\Stk\in\Stack$
and a \emph{store} $\store \in \Store$,
denoted by $\SelStack(\sstate)=\Stk$ and $\SelStore(\sstate)=\store$, respectively.
The \emph{depth} of a state $\sstate$, denoted by $\statedepth(\sstate)$, is the number of elements in its stack, i.e., $\statedepth(\sstate)=|\SelStack(\sstate)|$.
We denote the \emph{top} of the stack in an active state $\sstate=\ostatedefault$ by $\topstk(\sstate)=\Stk(1)$.
Intuitively, $\topstk(\sstate)$ contains the \emph{local state} of the \emph{active} (i.e., currently executing) contract method, while the other frames record the locals states of \emph{pending} calls to contract methods.
A \emph{frame} $\stk=\frame{\objectid}{\cmd}{\reg}$ records the \emph{local state} of
(a call to the contract method of) an object.
Formally, $\stk$ is a triple comprised of an \emph{object identifier}, denoted by $\SelObject(\stk)=\objectid$, a command, denoted by $\SelCmd(\stk)=\cmd$, which the method needs to execute,
and a \emph{local environment} $\reg\in\Reg$, denoted by $\SelEnv(\stk)=\reg$, which assigns values to the invocation's local variables.
A \emph{store} $\store\in\Store$ is a mapping from a finite number of object identifiers to their \emph{object state}.
\begin{figure}
\centering
$
\begin{array}{rclll}
\B{\epsilon,\store} & \Rightarrow & \B{\frame{\oid}{\ocode(o)}{[\inarg \mapsto n]}, \, \store}
\\
\B{\frame{\oid}{\SReturnV}{\reg},\, \store} & \Rightarrow & \B{\epsilon,\store} &
\\[1ex]
\B{\frame{\oid}{\SCallC{x}{\oid'}{e}}{\reg} \cdot \stk,\, \store} & \Rightarrow & \B{\ST{o',\ocode(o'),[\inarg \mapsto \reg(e)]} \cdot \frame{\oid}{\SAssign{x}{\res}}{\reg}\cdot \stk, \store}
\\
\B{\frame{\oid'}{\SReturn}{\reg'} \cdot \frame{\oid}{\cmd}{\reg} \cdot \Stk, \store}
& \Rightarrow &
\B{\frame{\oid}{\cmd}{\reg[\res \mapsto \reg'(\ret)]} \cdot \Stk, \store}
\end{array}
$
\caption{Operational semantics with a \emph{context} $\ocode$.
$\reg$ is naturally extended for expressions over variables in $\Var$.
We denote $o \in \dom(\store), n \in\mathbb{N}$.
\label{Fig:CallOS}
}
\end{figure}
\paragraph{Transition relations}
We formalize the semantics of our programming language using a \emph{transition relation}.
A \emph{transition} is a pair $\transition=(\sstate,\sstate') \in \Tr \subseteq \State \times \State$
comprised of
a \emph{source state} $\sstate$, denoted by $\src(\transition)$,
and \emph{target} state $\sstate'$, denoted by $\trg(\transition)$.
For clarity, we sometimes write a $\transition=(\sstate,\sstate')$ as $\sstate \Rightarrow \sstate'$.
We denote the active object of the transition by $\objectid(\transition)=\objectid(\topstk(\src(\transition)))$, or $\oid_{\it{main}}$ if it starts in a quiescent state.
We denote by $\cmd(\transition)\in\PCmd$
the primitive command that justifies the transition.
The meaning of primitive and compound commands is standard, and thus omitted.
We mention that primitive commands can only use local variables taken from the top stack frame,
and that only the fields of the \emph{active object} can be accessed.
\Cref{Fig:CallOS} defines meaning of method calls and returns.
When an object $\objectid$ is called from a quiescent state, a new stack frame is pushed to the currently empty stack.
The frame determines that the active object is $\objectid$, the command executing is the code $\ocode(\objectid)$ of $\objectid$, and the local environment for the invocation is the assignment of the value of $n$ to $\inarg$.
The last command in $\ocode(\objectid)$ is always a $\SReturn$, after which the frame is popped, leading to a quiescent state.
When a call $\SCallC{x}{\objectid'}{e}$ is made from an active state, a new stack frame is pushed as in the previous case.
We note that the local environment is initialized by assigning to $\inarg$ the value of $e$ in the local environment belonging to the caller, $\reg(e)$.
To handle retrieval of the return value from the callee, the command in the caller is modified to assign to $x$ the value of a specially designated variable $\res$.
When the callee invocation of $\objectid'$ finishes, the command in the top frame is $\SReturn$ and we let $\reg'$ denote the local environment of the callee.
The control transfers back to the caller object $\objectid$, and the value of $\res$ is set to be the value of $\ret$ in $\reg'$.
The assigned value of $\res$ is then automatically assigned to $x$, as determined by the operational semantics of the call.
The primitive command associated with a call is $\SEnterV$, and with a return is $\SReturnV$.
\paragraph{Executions}
An \emph{execution} $\exec = \exec(1) \ldots \exec(|\exec|)$ is a finite sequence of transitions coming from $\Tr$.
An execution $\exec$ is \emph{well-formed} if the target state of every transition is the source state of the following one, i.e.,
$\forall i\in\{2..|\exec|\}.\,\trg(\exec(i-1))=\src(\exec(i))$.
For clarity, we sometimes write an execution $\exec$ as $\exec = \sstate_1 \Rightarrow \sstate_2 \Rightarrow \cdots \sstate_n$.
We use $\ocode \inctx \exec$ to denote that an execution $\exec$ takes objects' code from context $\ocode$. We omit the context when no confusion is likely.
We say that a transition $\transition$ \emph{appears in} a $\exec$, denoted by $\transition \in \exec$, if $\exec = \_ \cdot \transition \cdot \_ $.
We say that a state $\sstate$ \emph{appears in} a $\exec$, denoted by $\sstate \in \exec$, if there is a transition $\transition \in \exec$ such that
$\sstate \in \{\src(\transition),\trg(\transition)\}$.
We denote the sets of transitions and states that appear in an execution $\exec$ by $\execstates(\exec)$ and $\exectrs(\exec)$, respectively.
An \emph{execution} $\exec'$ is a \emph{subexecution} of an execution $\exec$, denoted by $\exec'\subexec\exec$, if it is a subsequence of $\exec$.
We denote the \emph{first} and \emph{last} states of a non-empty execution $\exec$ by $\src(\exec)=\src(\exec(1))$ and $\trg(\exec)=\trg(\exec(|\exec|))$.
We say that $\exec=\tr\exec'\tr'$ is a \emph{complete execution} if $\src(\exec)$ and $\trg(\exec)$ are quiescent states and $\exec'$ contains only active states.
A \emph{run} is a concatenation of complete executions executed in the same code context $\ocode$,
By abuse of notation, we use $\exec$ to denote runs as well as executions.
In case we want to make the code context $\ocode$ of the run explicit, we write $\ocode \inctx \exec$.
The \emph{minimal} and \emph{maximal depths} of a non-empty execution $\exec$, denoted by
$\minexecdepth(\exec) = \min \{ \execdepth(\sstate) \mid \sstate \in \execstates(\exec)\}$
and
$\maxexecdepth(\exec) = \max \{ \execdepth(\sstate) \mid \sstate \in \execstates(\exec)\}$
are the minimal, respectively, maximal depths of any of the states it contains.
A well formed execution $\exec'$ is an \emph{invocation} in an execution $\exec$ if there exist transition $\tr$ and $\tr'$ such that
$\exec=\_ \cdot \tr \cdot \exec' \cdot \tr' \cdot \_$, where $\statedepth(\src(\tr))=\statedepth(\trg(\tr'))$ and $\minexecdepth(\exec')=\statedepth(\src(\tr))+1$.
We refer to $\statedepth(\trg(\tr))$ as the \emph{depth} of the invocation $\exec'$ and denote it by $\execdepth(\exec')$.
Note that according to this definition, the depth of an invocation that results from calling a contract method on a quiescent state is one.
\paragraph{Traces}
We define an \emph{event} as a pair $\event=\ST{\objectid, \action}$,
consisting of
an \emph{object} $\objectid$,
and a \emph{primitive command} $\action$.
Each transition $\transition$ can be transformed to an event by $\event(\transition)=\ST{\objectid(\transition), \cmd(\transition)}$.
The \emph{object} and \emph{primitive command} of an event are can be retrieved with $\objectid(\event)$ and $\action(\event)$, respectively.
A \emph{trace} is a sequence of events, denoted by $\trace$.
The trace matching an execution $\exec$ is received by point-wise application of $\event(\cdot)$ on all the transitions in $\exec$, denoted $\trace(\exec)$.
We denote by $\trace|_\objectid$ the maximal subsequence of $\trace$ comprised of events whose object is $\objectid$.
\section{Related Work}\label{Sec:RelatedWork}
\subsection{Modular Reasoning}
Modular reasoning is a topic which has been studied extensively
with the seminal works of~\citet{DBLP:journals/acta/Hoare72} and~\citet{DBLP:books/ph/Dijkstra76}.
For more recent studies on modularity we refer the readers to~\citet{DBLP:conf/fm/LeinoM05,DBLP:journals/jacm/BanerjeeN05}.
\emph{Averroes}~\cite{RW:AliLhotakECOOP13} is
a tool for generalizing call-graphs of applications by leaning on a
\emph{separate compilation assumption} to generate a general stub
library for applications. This allows analysis tools to be modular,
as generating full call-graphs
is both expensive and imprecise.
They show how encapsulation assumptions and proofs can be leveraged to improve
the feasibility and the precision of analyses.
In our work, we give
a sufficient condition, ECF, for the ability to soundly reason about
a single object in isolation from any other object.
The work of~\citet{RW:Leino2002ValidStateBit} presents an idiom for
verifying if an object behaves as expected in the presence of
callbacks, called \emph{Valid/state specification idiom}. Every
object $o$ maintains a `valid' bit that indicates if its state is
valid, i.e., satisfies its object invariants. The bit should be true
in every first invocation of $o$ in an execution. When
$o$ calls a method of another object $o'$, $o$ turns off its `valid'
bit. This way, if the execution of $o'$ leads to another method call
of $o$, before the original call to $o$ completed, the code of $o$
can take into account the fact that its object invariants do not
necessarily hold. The existence of such a
`valid' bit is helpful to achieve modular soundness, that is the
ability to reason about an object in isolation.
This paper achieves modular soundness by relying on the encapsulation of the object's
state.
Essentially, an ECF object is an object for which the
`valid' bit is always turned on, as it is guaranteed that the object
state changes only from within the object's methods, and that those
methods too are only executed where originally the `valid' bit would
be turned on. Thus, with the assumption on all executions being ECF,
there is no need to define a separate behavior of the code for when
the `valid' bit is turned on or off. To enable sound modular
reasoning, we simply ignore external calls and assume any return
value returned from any such external call.
We note that the absence of shared state drastically simplify our life.
\citet{LogozzoCLSS09}~presents a method for \emph{modular inference of class invariants}.
Specifically, it is shown that the trace semantics of an isolated class are sound and complete with respect to the trace semantics of a whole program.
The goal is to find the strongest state-based sound class invariant, that holds in both the isolated and non isolated cases.
Abstraction is used in order to compute such an invariant.
If it the class invariant matches the specification of the class,
then it is ensured that the class itself matches the specification
even in the context of a whole program.
The mentioned work enables modular reasoning by using abstraction.
Our work does not attempt to find such a sound class invariant,
but rather to satisfy the necessary conditions for being able to statically verify
any specification of an object in isolation of other objects.
The benefit here is that we do not depend on the precision of an abstraction,
which may output an invariant that overapproximates the specification, and thus does not meet it.
\subsection{Verification of Smart Contracts}
Even before the events surrounding the bug in the DAO,
there were discussions in the Ethereum community
about formal verification of smart contracts.
Following the extreme measures taken to
avert the effects of the attack on the DAO by hard-forking the blockchain and effectively
rewrite its history of executions, the discussion became more wide-spread.
\citet{RW:LuuCCS16}~characterized a class of security bugs in smart contracts called \emph{Transaction-Ordering Dependence (TOD)}. A contract inflicted with TOD bugs may behave unexpectedly when there is more than one client using the system and the effect of the execution of one client depends on whether the other client already executed or not.
In both TOD and ECF the bugs arise from the fact that the execution is performed in an unexpected state of the contract.
However, TOD bugs arise when there is more than one execution (since smart contracts are executed in a distributed environment), whilst non-ECF arises even in a single execution which contains callbacks.
One of the solutions suggested for TOD bugs is \emph{guarded transactions}.
The idea is to allow contract writers to define guard conditions which are verified by the virtual machine executing the contract code.
The execution must satisfy the guard condition, otherwise it fails without any effect.
However, by enabling modular reasoning on contracts by proving or asserting at runtime that the executions are ECF, we can verify similar conditions statically.
The only addition that may be required for the virtual machine is the online ECF check, which we found to be inexpensive in practice.
Checking arbitrary conditions at runtime may either be inefficient,
or not expressible enough to specify fully correct contract behavior.
In addition, by verifying at runtime that executions are ECF,
we are already able to detect and prevent executions which are,
with high probability, unwanted or unexpected.
\citet{RW:LuuCCS16} presents a tool called \emph{Oyente}~\cite{Eth:OyenteOnline}, based on symbolic execution of contracts.
The tool's web version reports on the existence of `reentrancy bugs', which is how the family of bugs such as the bug in the DAO were dubbed by the Ethereum community.
We attempted to verify both ECF and non-ECF contract variations based on the DAO object presented in \Cref{Fi:DaoContract}.
We received a report on a `reentrancy bug', even on ECF contracts.
We reported the false positives to the web Oyente team, and will submit the issue request by the camera-ready deadline.
The Why3~\cite{RW:Why3filliatre13esop} tool was also applied to verify smart contracts written in Solidity.
This requires whole code analysis and user supplied loop invariants.
\emph{}~\citet{BhargavanPLAS16}~translate a subset of the high-level Solidity language for Smart Contract development to F*, enabling using F*'s verification framework on Smart Contracts.
They also presented a decompiler for EVM bytecode to F*.
Similarly to the Why3 approach, the authors faced the issue of translating peculiar syntactic features
of the smart contract language Solidity to F*.
It should be noted that both F* and Oyente are successful in detecting other bugs, such as mishandled exceptions.
For technical clarity, we omit discussion of the semantics of exceptions and rollbacks in Ethereum.
Primarily, to arrive at general results that can be applied in domains other than Ethereum,
and secondly, to not overbear the reader with technical details
on the myriad ways Ethereum contracts may be invoked, and how exceptions may be handled in each of these ways.
\citet{MillerIACR2015}~discuss their insights from an educational smart contracts lab they held, and
published example contracts used in the lab.
We manually analyzed one such contract, implementing a \emph{rock, paper, scissors} game~\cite{Eth:EthereumLab}.
We identified several control paths in which a non-ECF execution might manifest.
Specifically, there are two control paths in registration to the game (in which players provide a sum as bounty), and three additional paths in the collection of the prize.
However, the authors put a constraint on the ability to execute callbacks by limiting to a minimum the amount of \emph{gas} available to the execution.
\emph{gas} is a novel concept in Ethereum that effectively bounds the runtime by associating with each low-level opcode a cost.
If an execution is not provided with enough \emph{gas} when called, it throws a special \emph{out-of-gas} exception.
\citet{RW:SergeyArxiv17} offer an analogy between the nomenclature of Smart Contracts and that of concurrent objects.
Specifically, the scenario of a contract calling another contract is compared to cooperative multitasking,
in which contract invocation is analogous to the case where the caller yields control.
One of the main challenges mentioned is that of being able to verify contracts in isolation of other contracts.
The ECF property brings us closer to that goal,
by allowing to check properties that can be specified as `contract invariants' in a modular way.
\section{Decidability}\label{Sec:Decidability}
This section discusses the decidability of verifying \ECF.
Using Rice Theorem (see, e.g.,~\citet{ullman}),
it is easy to show that
verifying \SECF, namely, statically verifying whether all executions of an object are
\FSECF{} or $\CECF$, is an undecidable problem.
Interestingly,
checking $\FSECF$ for a single execution ($\DECFo[\subFS]{\objectid}$) is also undecidable.
\changed{
\begin{theorem}
Given an execution $\ocode \inctx \exec$, checking if it is $\DECFo[\subFS]{\objectid}$ is undecidable.
\end{theorem}
\begin{proof}
We show a reduction from the halting problem.
$\PLname$ is Turing-complete, thus we encode the operation of a Turing machine $M$ as a command $c\in\Cmd$.
In \Cref{fig:decfo-decidability-reduction}, we present the code of a contract \texttt{A}.
The store of \texttt{A} has a single field $X$ initialized as $0$, and a single argument denoted $arg$.
The field $X$ is unchanged by $M$.
We also write the code of a contract \texttt{B}.
The method of \texttt{A} is separated to 3 branches. If $X\neq 0,1$ then the method returns without any effect.
If $X=1$, then $X$ is updated to $2$ and the method returns.
If $X=0$, then $X$ is updated to $1$ and if the argument is equal $0$, we execute the TM $M$ and update $X$ to 2 if and when $M$ finished running.
If the argument is not equal $0$, we call contract \texttt{B}.
If right after \texttt{B}'s execution $X$ is not equal $2$, then$X$ is updated to $3$.
The code of \texttt{B} is calling to \texttt{A}. Therefore, when \texttt{A} calls \texttt{B}, \texttt{B} always creates a callback to \texttt{A}.
\begin{figure}
\centering
\begin{subfigure}{2.5in}
\centering
\begin{footnotesize}
\begin{alltt}
\begin{tabbing}
XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=\kill
Object A \+\\
int X \\
M's fields \\
enter (arg) \\
if X == 0 then \+\\
X := 1 \\
if arg != 0 then \+\\
B() \\
if X != 2 then X := 3 \- \\
else \+\\
run M() \\
X := 2 \-\-\\
else if X == 1 then \+\\
X := 2 \-\\
return \-\\
\end{tabbing}
\end{alltt}
\end{footnotesize}
\end{subfigure}
\quad
\begin{subfigure}{2.5in}
\centering
\begin{footnotesize}
\begin{alltt}
\begin{tabbing}
XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=\kill
Object B \+\\
enter \+\\
A(0) \-\\
return \\
\end{tabbing}
\end{alltt}
\end{footnotesize}
\usetikzlibrary{arrows}
\begin{tikzpicture}[line cap=round,line join=round,>=latex,x=0.5cm,y=0.5cm]
\begin{scriptsize}
\draw [->,line width=0.5pt] (5.2,2.1) -- (6.7,2.1);
\draw [->,line width=0.5pt] (6.7,2.1) -- (8.5,2.1);
\draw [->,line width=0.5pt] (8.5,2.1) -- (9.6,3.02);
\draw [->,line width=0.5pt] (8.5,2.1) -- (9.6,1.4);
\draw [->,line width=0.5pt] (9.6,1.4) -- (10.2,0.85);
\draw [->,line width=0.5pt] (10.2,0.85) -- (11.5,0.85);
\draw [->,line width=0.5pt] (11.5,0.85) -- (12.8,0.85);
\draw [->,line width=0.5pt] (12.8,0.85) -- (13.5,1.5);
\draw [->,line width=0.5pt] (13.5,1.5) -- (13.5,0.75);
\draw [->,line width=0.5pt] (13.5,1.5) -- (14.6,1.5);
\draw [->,line width=0.5pt] (14.6,1.5) -- (16.3,1.5);
\draw [->,line width=0.5pt] (16.3,1.5) -- (17,1.5);
\draw [->,line width=0.5pt] (9.6,3.) -- (10.6,3.);
\draw [->,line width=0.5pt] (10.6,3.) -- (12.,3.);
\draw [->,line width=0.5pt] (12.,3.) -- (12.8,3.);
\draw (3.6,2.2) node[anchor=north west] {$q_{X=0}$};
\draw (6.0,2.0) node[anchor=north west] {$e_{A_{X=0}}$};
\draw (6.6,2.8) node[anchor=north west] {$X\!:=\!1$};
\draw (5.1,2.8) node[anchor=north west] {$A(a)$};
\draw (8.,2.6) node[anchor=north west, rotate=43] {$a\!=\!0$};
\draw (8.5,2.2) node[anchor=north west, rotate=-41 ] {$a\!\neq\! 0$};
\draw (9.3,3.8) node[anchor=north west] {$M()$};
\draw (10.5,3.8) node[anchor=north west] {$X\!:=\!2$};
\draw (12.5,2.9) node[anchor=north west] {$r_A$};
\draw (9.6,2.) node[anchor=north west] {$B()$};
\draw (9.8,0.85) node[anchor=north west] {$e_B$};
\draw (10.2,1.65) node[anchor=north west] {$A(0)$};
\draw (10.85,0.9) node[anchor=north west] {$e^{cb}_{A_{X=1}}$};
\draw (11.3,1.6) node[anchor=north west] {$X\!:=\!2$};
\draw (12.3,0.85) node[anchor=north west] {$r_B$};
\draw (13.5,0.85) node[anchor=north west] {$r_A$};
\draw (13.3,2.3) node[anchor=north west] {$X\!\neq\! 2$};
\draw [->,line width=0.5pt] (12.8,3) to[out=50,in=60] (5.2,3.2);
\draw (14.5,2.3) node[anchor=north west] {$X:=3$};
\draw (16.2,1.3) node[anchor=north west] {$q_{X=3}$};
\draw [->,line width=0.5pt] (13.5,0.75) to[out=-85,in=-110] (5.2,3.2);
\draw (3.6,3.3) node[anchor=north west] {$q_{X=2}$};
\draw [fill=black] (5.2,2.1) circle (1.0pt);
\draw [fill=black] (6.7,2.1) circle (1.0pt);
\draw [fill=black] (8.5,2.1 ) circle (1.0pt);
\draw [fill=black] (9.62,3.02) circle (1.0pt);
\draw [fill=black] (9.62,1.4) circle (1.0pt);
\draw [fill=black] (10.5,3.) circle (1.0pt);
\draw [fill=black] (10.2,0.85) circle (1.0pt);
\draw [fill=black] (11.5,0.85) circle (1.0pt);
\draw [fill=black] (12.8,0.85) circle (1.0pt);
\draw [fill=black] (13.5,1.5) circle (1.0pt);
\draw [fill=black] (13.5,0.75) circle (1.0pt);
\draw [fill=black] (12.,3.) circle (1.0pt);
\draw [fill=black] (12.8,3.) circle (1.0pt);
\draw [fill=black] (14.6,1.5) circle (1.0pt);
\draw [fill=black] (16.3,1.5) circle (1.0pt);
\draw [fill=black] (17,1.5) circle (1.0pt);
\draw [fill=black] (5.2,3.2) circle (1.0pt);
\end{scriptsize}
\end{tikzpicture}
\end{subfigure}
\caption{The codes of two objects \texttt{A} and \texttt{B} showing that $\DECFo[\subFS]{\objectid}$ is undecidable, and a diagram showing the possible flows of the system.
Nodes marked $q$ are quiescent states, with the current value of \texttt{A}'s $X$ variable in subscript.
Nodes starting with $e_o$ indicate entry to an object $o$, and $r_o$ a return from an object $o$.
The notation $e^{cb}_o$ indicates the call is a callback.
Entry nodes are marked with the value of $X$ for better readability.
Syntactic references to the code (assignments, calls, conditions) appear on the edges.
Missing call edges from quiescent states indicate that the resulting quiescent state is the same, hence, $q_{X=2}$ and $q_{X=3}$ are sinks.
}\label{fig:decfo-decidability-reduction}
\end{figure}
We consider the execution $\exec$ starting from the initial state in which \texttt{A}'s $X$ field is equal $0$, with $arg\neq 0$.
This execution calls the object \texttt{B}, which calls back to \texttt{A}, and in the callback the value of $X$ is set to $2$.
We show that $\exec$ is $\DECFo[\subFS]{\objectid}$ if and only if $M()$ halts.
For the `if' direction, we note that if $M()$ halts, then the execution of $A(0)$ from the state where $X=0$ leads to $X$ being set to $2$ right after $M$'s run finished, and that execution has no callbacks, as required.
For the `only if' direction, we note that the only callback-free execution that starts from $X=0$ and ends with $X=2$ is the call \texttt{A}(0), which is the execution that executes $M()$, and it is a legal execution only if $M()$ halts.
The reason that this must be the only execution, is that for any choice of input argument $arg\neq 0$ and context $\ocode$ that maps a different code for \texttt{B}, does not allow reaching the required final state $X=2$ unless callbacks are used.
If $\ocode(\texttt{B})$ still calls back to \texttt{A} then clearly the resulting execution is not an eligible ECF witness.
If $\ocode(\texttt{B})$ does not call back to \texttt{A}, then $X$ is updated to $3$.
Therefore, when $X=3$, any subsequent call to \texttt{A} cannot modify $X$, and in particular $X=2$ is not reachable.
\end{proof}
}
\changed{
\begin{figure}
\begin{subfigure}[t]{2.5in}
\begin{footnotesize}
$A_\objectid$
\begin{alltt}
\begin{tabbing}
XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=\kill
arg = *; \\
... // Code of o \\
x := o'(y) --> E(); x := *; \\
... \\
\end{tabbing}
\end{alltt}
\begin{alltt}
\begin{tabbing}
XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=\kill
E = while (*) \+\\
arg = *; \\
o(arg); \-\\
\end{tabbing}
\end{alltt}
\end{footnotesize}
\end{subfigure}
\begin{subfigure}[t]{2.5in}
\begin{footnotesize}
$A^0_\objectid$
\begin{alltt}
\begin{tabbing}
XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=\kill
arg = *; \\
... // Code of o \\
x := o'(y) --> x := *; \\
...
\end{tabbing}
\end{alltt}
\end{footnotesize}
\end{subfigure}
\caption{The construction of automatons $A_\objectid$ and $A^0_\objectid$.
The notation \texttt{-->} indicates that a command in the code of $\objectid$ is replaced with another.}\label{Fig:PDAConstruction}
\end{figure}
}
In contrast, checking $\CECF$ for a single execution (\DECF[\subC]) is obviously decidable, as we can enumerate all of the permutations of a particular input trace.
Thus, we focus on verifying \SECF, namely, statically verifying whether all executions of an object are
\FSECF{} or $\CECF$,
where the domains of the object variables are restricted to finite sets.
Hence, such objects can be modeled with a pushdown-automaton ($\PDA$).
\changed{
Such a PDA for an object $\objectid$ is able to simulate any modular well-formed execution $\ocode \in \exec$ where the active object of all states in $\exec$ is $\objectid$.
We denote this construction $A_\objectid$. Its code is shown in the left side of \Cref{Fig:PDAConstruction}.
It executes the code of $\objectid$ with a non-deterministically chosen argument, and replaces every call to an external object with a sequence of arbitrary calls to $\objectid$ (the code in \texttt{E}), and a non-deterministic choice of the return value of the call.
}
We begin with a rather simple lemma that shows $\FSECF$ of objects is indeed decidable in this model.
We assume that
variables may take values coming from a finite domain.
\begin{lemma
\label{prop:FSECFDecidable}
Let $\objectid$ be an object, assuming a finite domain for variables.
Then there is an algorithm that decides if $\objectid$ is $\SECF[\subFS]$.
\end{lemma}
\begin{proof}
\changed{
We consider the pushdown automaton $A_\objectid$ and an automaton that allows only callback-free behaviors for $\objectid$, $A^0_\objectid$.
Both are shown in \Cref{Fig:PDAConstruction}.
Each execution of $A_\objectid$ consists of first choosing non-deterministically an argument $arg$ with which the object $\objectid$ is called.
The automaton simulates the steps of $\objectid$ with the only difference being in invocations of other objects, e.g. $\objectid'$.
This command is replaced with the code of the ``environment'' object $E$, that performs a sequence of calls to $\objectid$ with an arbitrary argument chosen in each iteration.
The sequence length is arbitrary and may be also $0$, i.e., no callbacks.
After running the callbacks, the automaton non-deterministically chooses a return value, stored in the original variable intended to store the return value of $\objectid'$, which in the figure is the variable $x$.
The automaton that allows only callback-free behaviors for $\objectid$ is simpler:
the only difference is that it does not run the code of $E$.
In fact, $A^0_\objectid$ is a finite state machine.
We now utilize the result by~\cite{bouajjani1997reachability} for reachability of a regular set of configurations of a pushdown automaton.
Here, the set of configurations is $\State$, and the subset of configurations which we are interested in is the one with with an empty stack, i.e. $\{\ST{}\} \times \Sigma_\objectid$ where $\Sigma_\objectid$ is the set of $\objectid$'s possible states, which is finite and thus regular.
We consider an arbitrary pair of states $\sigma_1,\sigma_2\in\Sigma_\objectid$.
We first check if there is an execution of $A_\objectid$ starting from $\sigma_1$ and ending in $\sigma_2$.
This can be done by checking if $\sigma_1$ is reachable from the initial state and as well as if $\sigma_2$ is reachable from $\sigma_1$.
As the path that shows reachability in $A_\objectid$ may not be callback-free, we check for reachability also in the finite state machine $A^0_\objectid$, in the same manner.
Reachability in finite-state machines is known to be decidable.
If there is no execution in $A^0_\objectid$ from $\sigma_1$ to $\sigma_2$, then the object $\objectid$ is not $\SECF[\subFS]$.
After checking for all pairs in $\Sigma_\objectid \times \Sigma_\objectid$
that for each pair of states $(\sigma_1,\sigma_2)$ there is an execution of $A_\objectid$ starting in $\sigma_1$ and ending in $\sigma_2$ if and only if there is such an execution in $A^0_\objectid$, we verified that $\objectid$ is $\SECF[\subFS]$.
The set of all pairs of states is finite, thus it describes a decision procedure for verifying $\SECF[\subFS]$ for an object $\objectid$.
}
\end{proof}
Showing the decidability of $\SECF[\subC]$ of objects is not that easy, because it requires
reasoning on permutations of events, which is not a regular property, even in the case of finite-state machines.
\changed{
Our strategy for proving the decidability of $\SECF[\subC]$ will be as follows.
As before, we consider executions of $A_\objectid$, which identify with the set of possible projected executions on $\objectid$.
We will show that it is enough to check subexecutions of $A_\objectid$ which involve at most two elements of the stack.
Namely, it is the same as checking the set of executions of $A_\objectid$ with a limit of two to the depth of the stack,
where the initial state of the execution may include both the initial states $\Sigma_\objectid$ of an execution of $A_\objectid$, and in addition any state $\sstate$ where a callback may be called.
Explicitly, $\sstate$ is a state whose primitive command is a call.
We describe such states using the set $\mathcal{I}_0=\{\sstate \mid \exists x,\oid',e. \SelCmd(\topstk(\sstate))=\SCallC{x}{\oid'}{e} \}$.
It is regular as we only consider the top element of the stack.
Finding for the set of states $\mathcal{I}_0$
its subset of reachable states in $A_\objectid$
can therefore be done using~\cite{bouajjani1997reachability}.
We denote by $A^2_\objectid$ the automaton that is received by limiting the stack depth of $A_\objectid$ to two, and having initial states $\mathcal{I}_0\cup\Sigma_\objectid$.
For $A^2_\objectid$, deciding if all its executions are $\DECFo[\subC]{\objectid}$ is decidable.
We do so by constructing a monitor that receives as input execution traces of $A^2_\objectid$ and outputs an error if the executions are not $\DECFo[\subC]{\objectid}$.
Then we will show that $M$ is a finite state machine and thus checking reachability of its error states is decidable.
Such a construction is possible because even though conflict-equivalence demands finding a permutation of a trace, which is not a regular property, the actual choice of permutations that have to be checked is much more limited.
\begin{lemma}
Let $\Pi$ denote the set of all executions of $A^2_\objectid$.
There is an automaton $M$ that for each $\exec\in\Pi$, ends in an accepting state if $\exec \vDash \DECFo[\subC]{\objectid}$, and in a rejecting state otherwise.
\end{lemma}
\begin{figure}
\begin{footnotesize}
{\begin{tt}
\begin{tabbing}
XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=XX\=\kill
$M(\exec)$ = \+\\
d = 0, prefix = $(\emptyset,\emptyset)$, delayedCbs = $(\emptyset,\emptyset)$ \\
for ($\event \in \trace(\exec)$) \+\\
if ($\cmd(\event)=\SAssign{x}{F}$) \+ \\
R := R$\cup \{F\}$ \- \\
if ($\cmd(\event)=\SAssign{F}{x}$) \+ \\
W := W$\cup \{F\}$ \- \\
if ($\cmd(\event)=\SEnterV$ \&\& d=0) // Execution starts \+ \\
R, W := $(\emptyset, \emptyset)$ \\
d := d+1 \- \\
if ($\cmd(\event)=\SReturnV$ \&\& d=1) // Execution ends \+ \\
assert((R,W) commutes with delayedCbs) \\
R, W := $(\emptyset, \emptyset)$ \\
d := d-1 \\
return \- \\
if ($\cmd(\event)=\SEnterV$ \&\& d=1) // Callback starts \+ \\
assert((R,W) commutes with delayedCbs) \\
prefix := (R(prefix) $\cup$ R, W(prefix) $\cup$ W) \\
R, W := $(\emptyset, \emptyset)$ \\
d := d+1 \-\\
if ($\cmd(\event)=\SReturnV$ \&\& d=2) // Callback ends \+ \\
if (!((R,W) commutes with prefix) || !((R,W) commutes with delayedCbs)) \+ \\
delayedCbs := (R(delayedCbs) $\cup$ R, W(delayedCbs) $\cup$ W) \- \\
R, W := $(\emptyset, \emptyset)$ \\
d := d-1
\end{tabbing}
\end{tt}
}
\end{footnotesize}
\caption{The code of $M$ which accepts an execution $\exec$ of $A^2_\objectid$ and verifies if it is $\DECFo[\subC]{\objectid}$}\label{fig:DECFForLevel2DecisionProcedure}
\end{figure}
\begin{proof}
We write the code of the automaton $M$ in \Cref{fig:DECFForLevel2DecisionProcedure}.
The automaton loops on each event in the trace of the execution.
The automaton states consist of a depth variable {\tt d}, and two pairs of read and write sets: {\tt prefix} and {\tt delayedCbs}.
The sets {\tt R} and {\tt W} are updated in each command that reads from or writes to the object store.
They are reset when a callback starts or ends.
The {\tt prefix} pair retains the accumulated reads and writes in the invocations at depth $1$,
i.e. the first invocation of $\objectid$ and which we refer to sometimes as the \emph{main} invocation.
The {\tt delayedCbs} pair retains the accumulated reads and writes by callbacks that we choose to execute after the main invocation ends.
Intuitively, the monitor checks each time a pair of {\tt (R,W)} is finalized and before it is reset, if it satisfies conditions that will allow to find a conflict-equivalent execution.
For a callback execution, we check if the pair commutes with the prefix, i.e. all portions of the main invocation already executed.
If it does not commute with the prefix, we mark it as a delayed callback, and update the {\tt delayedCbs} read and write sets.
As we do not know whether delayed callbacks actually commute with the rest of the invocation,
and with future callbacks that may be executed before the main invocation,
we retain the conflict information in {\tt delayedCbs} to be checked later against any finalized {\tt (R,W)} sets that is belonging either to the main invocation, or to a callback that is chosen to be executed before the main invocation.
Therefore, even in the case when a callback commutes with the prefix, but in which it does not commute with the delayed callbacks (which are `jumping over' the callback under consideration in order of execution),
we have to try to execute it as a delayed callback as well.
Otherwise, the execution is surely not $\DECFo[\subC]{\objectid}$, and we move to an error state.
For a portion of the main invocation (the first or last one, or between callbacks), we check if it does not conflict with any of the relevant delayed callbacks.
The {\tt delayedCbs} pair is always updated correctly with the currently delayed callbacks' read and write locations.
It can be seen from the construction that if $M$ accepts, then we can build from $M$'s execution a witness for $\exec$ being $\DECFo[\subC]{\objectid}$ by ordering all non-delayed callbacks before the main invocation and the delayed callbacks after the main invocation, both of those in the order in which they appear in the original execution.
E.g., if callbacks $i_1$ and $i_2$ are both delayed callbacks, then if $i_1$ is executed before $i_2$ in $\exec$ then the ordering between them in the witness is not changed. The same argument applies for non-delayed callbacks.
In addition, if $\objectid$ is $\DECFo[\subC]{\objectid}$ then $M$ must accept.
This is because any witness for $\exec$ being $\DECFo[\subC]{\objectid}$ is conflict-equivalent to the witness implicitly produced by $M$.
The reason for that is that in the witness, we can also identify a set of callbacks executed before the main invocation and a set of callbacks executed after it.
The internal ordering of callbacks that are executed before the main invocation does not matter as long as it does not reorder conflicts, and thus may in fact be the same as in the original execution, ditto for the other set of callbacks.
\end{proof}
We note that $M$ has a finite state: {\tt d} in $\{0,1,2\}$, and {\tt prefix} and {\tt delayedCbs} in $2^F \times 2^F$ where $F$ is the set of $\objectid$'s fields.
Therefore, $M$ is a finite state machine, and thus reachability is decidable. In particular, we can check if the error states are reachable.
However, $M$ is a machine that works on any execution of any object.
It is not difficult, though, to build $M$ as a monitor of $A^2_\objectid$. We denote this construction $M_{A^2_\objectid}$.
The result is still a finite state machine with decidable reachability, since $A^2_\objectid$, which has a bounded stack depth of 2, is also a finite state machine.
\begin{corollary}\label{corol:ConflictECFA2Decidable}
It is decidable to check if all executions of $A^2_\objectid$ are $\DECFo[\subC]{\objectid}$
\end{corollary}
We now show that the previous corollary implies that if all executions of $A^2_\objectid$ are $\DECFo[\subC]{\objectid}$, then then so are all executions of $A_\objectid$.
In particular, it shows that $\objectid$ is $\SECF[\subC]$.
Importantly, the converse is also true: if an object $\objectid$ is $\SECF[\subC]$, then all executions of $A^2_\objectid$ are $\DECFo[\subC]{\objectid}$.
This is trivial to show:
for an execution $\exec$ of $A^2_\objectid$ such that $\exec \neg{\vDash} \DECFo[\subC]{\objectid}$
it is easy to find a complete execution $\exec'$ in $A_\objectid$ of the form $\exec' = \exec_0 \hat{\exec} \exec'_0$ which is also not $\DECFo[\subC]{\objectid}$.
Specifically, $\exec_0$ will be a subexecution that reaches the initial state of $\exec$,
$\hat{\exec}$ will be identical to $\exec$, only changing the stacks in $\exec$'s states to account for $\exec_0$'s transitions,
and $\exec'_0$ will complete the execution.
Any permutation of $\trace(\exec')$ will induce a conflict equivalence breaking permutation on $\hat{\exec}$,
whose conflicts are the same as those of $\exec$,
therefore $\exec'$ is not $\DECFo[\subC]{\objectid}$.
\begin{lemma}\label{lem:ConflictECFA2ToGeneral}
An object $\objectid$ is $\SECF[\subC]$ if all executions of $A^2_\objectid$ are $\DECFo[\subC]{\objectid}$.
\end{lemma}
\begin{proof}
We show that for every execution $\exec$ of $A_\objectid$ there is a witness $\exec'$ for it being $\DECFo[\subC]{\objectid}$.
The witness is built recursively.
We initialize $\exec'=\exec$.
We denote $D=\maxexecdepth(\exec)$.
We start with the longest subexecution in $\exec_0 \subexec \exec$ which has $\maxexecdepth(\exec_0)=D$ and $\minexecdepth(\exec_0)=D-1$.
By ignoring the first $D-2$ stack frames,
$\exec_0$ can be seen as an execution of $A^2_\objectid$,
thus it is $\DECFo[\subC]{\objectid}$.
Hence, by taking the witness in $A^2_\objectid$ and returning the first $D-2$ stack frames,
we have a subexecution $\exec'_0 \ceq^\objectid \exec_0$ which is callback-free: $\minexecdepth(\exec'_0)=\maxexecdepth(\exec'_0)=D-1$.
We update $\exec'$ by replacing the transitions of $\exec_0$ with $\exec'_0$.
The process continues by taking in each step the longest and deepest subexecution which has a callback and replacing it with a callback-free subexecution which is conflict-equivalent to it.
In each step, the resulting $\exec'$ is conflict-equivalent to $\exec$.
The premise of the lemma ensures that this process will not stop until $\exec'$ is callback-free.
\end{proof}
From \Cref{corol:ConflictECFA2Decidable} and \Cref{lem:ConflictECFA2ToGeneral} we immediately get the following result:
\begin{theorem}
Let $\objectid$ be an object, assuming a finite domain for variables.
Then there is an algorithm that decides if $\objectid$ is $\SECF[\subC]$.
\end{theorem}
}
|
\section{Introduction}
While the discovery of the Higgs boson in 2012 marked the completion
of the Standard Model (SM) as we know it today, important questions such as
the nature of dark matter, the origin of the matter-antimatter
imbalance, the stability of the electroweak vacuum, and the relative
lightness of the Higgs boson remain open.
In answering these questions, many theories beyond the SM
predict new massive particles to be produced at the LHC. Their
subsequent decays may contain, among other signatures, top quarks with
$p_T/m > 1$. If these \textit{boosted} top quarks decay hadronically, their decay products
become collimated and can be experimentally reconstructed as a jet
with large distance parameter (\textit{large-R jet}).
Distinguishing such top decays from jets initiated by light quarks or
gluons is the goal of top tagging algorithms. In the following we
review three broad classes of such algorithms: generic
QCD inspired jet substructure variables, inclusive top tagging
algorithms, and machine learning approaches\footnote{The reconstruction of b quarks inside a candidate jet is another
handle to identify the decays of top quarks. The current experimental
approach is to largely treat this as an independent problem and we
will not discuss it further. However, the recent advent of
maximum-information machine learning techniques will likely blur this
distinction.}.
While this document can serve as a starting point, the following
reviews as well as proceedings of the annual BOOST conference provide a more
thorough survey of the
field~\cite{BoostReport2016Theory,BoostReport2012,TilmanTagsTops}.
\section{QCD Inspired Jet Substructure Variables}
Over the last decade, a number of techniques
--- known as \textit{grooming algorithms} ---
have been developed to remove radiation due to the underlying event and pile-up from jets while extracting
information about the hard substructure, especially its mass.
Trimming~\cite{Trimming} first reclusters the constituents of a given jet
with a smaller distance parameter $R_{\textrm{sub}}$ resulting in a
set of so-called subjets. In a second step all subjets with a
transverse momentum of less than a fraction $f_\textrm{cut}$ of the
initial jet transverse momentum are discarded. Finally, the trimmed
jet is obtained by summing up the remaining subjets. Trimming
is the default
grooming algorithm employed by the ATLAS collaboration~\cite{CONF-17-063}.
Conversely, the soft drop algorithm~\cite{SoftDrop} attempts to find the hard
substructure using a step-wise unclustering procedure. The initial jet is recursively unclustered until
no constituents remain or until
the soft drop condition
\begin{equation}
\frac{\min(p_{T1},p_{T2})}{p_{T1}+p_{T2}} > z_{cut} \left( \frac{\Delta R_{12}}{R_0}\right)^\beta
\end{equation}
is fulfilled.
Here $p_{T1}$ ($p_{T2}$) denotes the transverse momentum of the subjet with higher (lower) transverse momentum,
$\Delta R_{12}$ their angular distance and $z_{cut}$, $R_0$, and $\beta$
are free parameters of the algorithm.
The recursion proceeds at each step by discarding the lower momentum subjet and unclustering the higher momentum one. For $\beta=0$ the algorithm corresponds to the modified mass drop algorithm~\cite{MMDT}. The soft drop algorithm is widely used by the CMS collaboration~\cite{JME-16-003}. Recent attempts have been made to analytically calculate the properties of soft dropped jets to high precision~\cite{SoftDropFact,SoftDropMassCalcI,SoftDropMassCalcII} and to perform corresponding measurements with LHC collision data~\cite{PAS-SMP-16-010,ATLASSDMeas}.
Beyond the mass, the number of distinct centers of radiation is well
suited to identify top quarks (with three centers) from other objets
(which are expected to have one or two). One way to quantify this intuition is given by the n-subjettiness~\cite{NSubjettiness} $\tau_N$.
The n-subjettiness is small
when the radiation inside a jet is compatible with $N$ hard centers of
radiation. Commonly the ratio $\tau_3/\tau_2$ is used for top
tagging. Further developments include the dichroic n-subjettiness~\cite{DichroicNSubjettiness} ---
the ratio of n-subjettiness values after different grooming techniques
are applied --- and its use for jet clustering~\cite{XCone}.
More recently, energy correlation functions have been proposed which
generalize the n-subjettiness by calculating n-particle correlators
and therefor do not rely on the explicit finding of subjet axes~\cite{ECF,ECF2}.
\section{Inclusive Top Tagging Algorithms}
In addition to the generic variables discussed in the previous
section, a number of inclusive top tagging algorithms have been
developed.
The HEPTopTagger~V2~\cite{HTT,HTTv2} algorithm identifies top quarks by a mass-drop
filtering procedure followed by selection criteria on the top mass and
W to top mass ratio. This procedure is repeated for a number jet of
distance parameters, starting from a seed size of $R=1.5$ until the
optimal jet size is found.
The Heavy Object Tagger with Variable R~\cite{HOTVR} (HOTVR) also implements jet
clustering with variable distance parameter. Soft radiation is already removed during
the clustering procedure, using a mass drop criterion leading to a
reconstructed jet mass stable over a large transverse momentum range.
Shower deconstruction~\cite{ShowersI,ShowersII} is a full information approach similar to the
matrix element method that assigns signal and background
probabilities to individual jets instead of events. First, microjets
are built by clustering the constituents of the large-R jet into
smaller jets.
The classifying variable
$\chi$ is then calculated as the probability quotient that a set of
microjets was created by the decay of a top quark, divided by the
probability that it was created by light quarks or gluons.
The Template Overlap Method~\cite{TemplateI,TemplateII} uses a library of templates which encode
the distribution of parton momenta inside a jet. The classifier is
then calculated as the maximum agreement between the jet under study
and all templates for a given phase space.
Finally, the Boosted Event Shape Tagger~\cite{BEST} (BEST) starts by boosting jet
constituents individually into reference frames corresponding to a
particle origin hypothesis (t, W, Z, H). Subsequently, angular
distributions such as
Fox-Wolfram moments or sphericity are calculated in each
frame. A fully connected neural network then yields the compatibility
with each particle hypothesis for simultaneous classification.
\section{(Deep) Machine Learning Approaches}
Deep learning is a novel development in computer science, largely
based on creating artificial neural networks that feature multiple
layers and correspondingly increasingly higher levels of
abstraction. Neural networks connect a number of inputs to an output
decision via intermediate nodes. The output of a given node is
determined by its inputs and a set of weights that are adjusted
(trained) for a given decision task. How many internal nodes to use
and how to connect them is referred to as the
network architecture. In addition to using large numbers of internal
layers, deep network approaches typically operate directly on
minimally preprocessed input data (as opposed to variables designed by
experts discussed in the previous sections).
An important decision when designing a deep neural network for a specific
task is the choice of the architecture. In the following we outline
several network architectures used for top tagging.
In a fully connected network (FCN), each node in a given layer is connected
to all nodes of the previous and subsequent layers. This architecture
is very generic, but does not make explicit use of symmetries inherent
to the problem at hand. An application of FCNs to
top tagging is presented in~\cite{FCN}. The individual constituents of a jet
are given to a FCN with four hidden layers. Several steps of pre-processing
including trimming and scaling, translation, rotation, and flipping
steps increase the achieveable performance.
Image recognition tasks are commonly solved using
convolutional neural networks (CNNs). In these networks the
weights to be learned are arranged in a so-called convolutional kernel
(or several thereof) that is applied to different parts of the
image. This approach can thus detect features useful for
discrimination no matter where on the image they occur.
Top tagging can be viewed as an image
recognition task by equating the energy deposits in the
calorimeters with the pixels of a grayscale image~\cite{DeepTop}.
An alternative approach, inspired by the processing of human language
are recursive neural networks (RNNs). A sentence (sequence)
consists of an arbitrary number of words (inputs). RNNs process the
elements of an input sequence one after the other. Each input in the
sequence is handled by the same processing layer (thus
recursive). Importantly, RNNs also possess internal memory.
Recently, a combination of RNNs and jet
clustering algorithms was used to develop a competitive identification
technique for boosted top quarks~\cite{LSTM}.
Finally, the physics inspired LoLa algorithm~\cite{LoLa} starts with a list of input
four vectors which can either be calorimeter towers, particle flow
candidates or generated partons. In a first step a combination layer
(CoLa) yields a set of trainable linear combinations of
the inputs. In a second step, the eponymous Lorentz layer (LoLa) uses
the resulting matrix to extract physics features based on convolutions
with the Minkowski metric. Its output is then passed to a FCN to identify top quarks.
\section{Conclusions}
A varied menu of top tagging techniques is available for use in
physics analyses. While developments previously focused on searches for new
physics, recent precision calculations might make substructure
increasingly interesting for SM measurements. At the same time, deep learning based tagging algorithms
already show promising gains in performance.
\bigskip \bigskip \begin{center} \begin{large
The author would like to thank the organisers of TOP 2017 for the invitation to this wonderfully organised and stimulating conference.
|
\section{Introduction}
5-brane webs in type IIB string theory have been used to study five-dimensional (5d) superconformal field theories (SCFTs) that are ultraviolet (UV) completions of a certain class of 5d $\mathcal{N}=1$ supersymmetric gauge theories \cite{Seiberg:1996bd,Morrison:1996xf,Douglas:1996xp,Intriligator:1997pq, Aharony:1997ju, Aharony:1997bh}. A 5-brane web configuration provides us with a tool to compute the instanton partition function that captures the BPS spectrum of a 5d theory realized on a 5-brane web as well as with a perspective of qualitative understandings of the SCFTs such as
global symmetry enhancements and various dualities.
By introducing an orientifold plane like an O7-plane or an O5-plane,
5-brane webs can be enriched so that one can describe 5d theories with some other classical gauge group, such as $SO(N), USp(2N)$ \cite{Brunner:1997gk, Bergman:2015dpa, Hayashi:2015vhy, Zafrir:2015ftn}, in addition to the standard classical gauge group $SU(N)$. In recent years there has been some progress on brane configurations with the orientifold planes. For example, in \cite{Hayashi:2015zka,Hayashi:2016abm,Yun:2016yzw}, whether we resolve an O7$^-$-plane into two $[p ,q]$ 7-branes or not in a certain 5-brane web configuration gives an explanation for equivalence proposed in \cite{Gaiotto:2015una} of two theories at the UV fixed point, an $SU(N+1)$ theory with $N_f \le 2N+6$ hypermultiplets in the fundamental representation (flavors) and the Chern-Simons (CS) level $\kappa=\pm\left(N+3-N_f/2\right)$, and an $USp(2N)$ theory with the same number of flavors. In particular, it has been discussed in \cite{Zafrir:2015ftn, Zafrir:2016jpu} that 5-brane configurations with an O5-plane can realize an $SO(N)$ gauge theory even with hypermultiplets in the spinor/conjugate spinor representation. It is also noticeable that one can differentiate the discrete theta angle ($\theta=0,\pi$) of the 5d pure $USp(2N)$ gauge theory \cite{Morrison:1996xf,Douglas:1996xp,Intriligator:1997pq} from 5-brane webs with an O5-plane. The two different theta angles turn out to imply two distinct phase structures for their 5-brane webs, that are characterized by two distinct ``generalized" flop transitions which may be applied to 5-branes intersecting at the same point with an O5-plane \cite{Hayashi:2017btw}.
There has been also progress along a quantitative side on 5-brane webs with an O5-plane. The conventional topological vertex formalism \cite{Aganagic:2003db, Iqbal:2007ii,Awata:2008ed} enables one to systematically compute the Nekrasov instanton partition function of a 5d theory on a 5-brane web via the correspondence~\cite{Leung:1997tw} between a toric diagram and a certain $(p, q)$ 5-brane web diagram. Although 5-brane webs for $SU(N)$ gauge theories with a large number of flavors or a large CS level often lead to non-toric Calabi-Yau geometries \cite{Benini:2009gi}, the topological vertex formulation is still applicable to reproduce the correct partition function \cite{Hayashi:2013qwa,Hayashi:2014wfa, Hayashi:2015xla, Kim:2015jba, Hayashi:2016abm, Hayashi:2016jak}.
Quite recently,
the topological vertex formulation has been further extended to 5-brane webs with an O5-plane \cite{Kim:2017jqn}. Together with a generalized flop transition,
the new method
utilizes a configuration where one-half of the original brane configuration is glued to the other half from the mirror image due to an O5-plane in a specific manner.
The purpose of this paper is to further extend the study of 5-brane webs with an O5-plane and propose 5-brane web diagrams for 5d $\mathcal{N}=1$ gauge theories of an exceptional gauge group $G_2$, using these recent developments on 5-brane webs. We then compute their Nekrasov partition functions based on the topological vertex formalism for 5-brane webs with an O5-plane.
Our strategy is as follows: A 5-brane web diagram for the $SO(7)$ gauge theory with a hypermultiplet in the spinor representation has been constructed in \cite{Zafrir:2015ftn}. We then consider the Higgs branch of the $SO(7)$ gauge theory with one spinor in terms of the web diagram, which should yield a 5-brane web configuration for the pure $G_2$ gauge theory\footnote{We thank Gabi Zafrir for illuminating discussion about this strategy.}.
We note that there are two ways to obtain the web diagram for the $SO(7)$ gauge theory with one spinor. One way is to Higgs the $SO(8)$ gauge theory with a hypermultiplet in the vector representation and a hypermultiplet in the spinor or conjugate spinor representation. The other way is to Higgs the $SO(8)$ gauge theory with a hypermultiplet in the spinor representation and a hypermultiplet in the conjugate spinor representation. These two $SO(8)$ gauge theories should be equivalent to each other due to the triality of $SO(8)$, and both Higgsings hence give rises to the $SO(7)$ gauge theory with one spinor, while the resulting brane configurations look different. Further Higgsing of the two types of the diagrams leads to two different 5-brane webs for the pure $G_2$ gauge theory, which therefore gives two different configuration for the same $G_2$ gauge theory. It is possible to add flavors to the pure $G_2$ gauge theory by Higgsing the $SO(7)$ gauge theory with hypermultiplets either in the vector representation or the spinor representation in addition to one spinor. We test our proposal by comparing the area of compact faces that a D3-brane wraps on the 5-brane webs with the tension of a monopole string which can be calculated from the effective prepotential of the theory in question.
In fact the analysis implies an interesting feature like which faces of a 5-brane web a D3-brane wraps in the presence of an O5-plane.
We then go on to compute the Nekrasov partition function for 5d $\mathcal{N}=1$ $G_2$ gauge theories by applying the recently proposed topological vertex method for 5-brane webs with an O5-plane \cite{Kim:2017jqn}. We check that our partition function for the pure $G_2$ gauge theory reproduces the one-instanton result \cite{Benvenuti:2010pq, Keller:2011ek} and also the two-instanton result \cite{Hanany:2012dm, Keller:2012da, Cremonesi:2014xha}. We also show that the partition function of the $G_2$ gauge theory with one flavor is consistent with flavor decoupling.
The paper is organized as follows: In section \ref{sec:G2fromO5tilde}, we first discuss a 5-brane web for the $SO(7)$ gauge theory with a hypermultiplet in the spinor representation.
By Higgsing the $SO(7)$ theory with one spinor, we propose a 5-brane web for the pure $G_2$ gauge theory. We check that the proposed diagram is consistent with the effective prepotential of the pure $G_2$ gauge theory. We also present two ways to introduce flavors.
In section \ref{sec:G2fromO5}, using the triality of $SO(8)$ gauge theory among hypermultiplets in the vector, spinor, and conjugate spinor representations, we propose another 5-brane web for the pure $G_2$ theory through successive Higgsings of the $SO(8)$ gauge theory theory with one spinor and one conjugate spinor.
In section \ref{sec:Nekrasov}, we first review a recent proposal for the topological vertex formulation with an O5-plane and extend it to the cases with an $\widetilde{\text{O5}}$-plane. We then use it to compute the partition functions of 5d $\mathcal{N}=1$ $G_2$ gauge theories with no flavor and with one flavor.
In section \ref{sec:conclusion}, we summarize the results and comment on further directions. \\
{\bf Note added:} We are informed that the authors of \cite{Kim:2018gjo} computed the partition function for 5d $\mathcal{N}=1$ $G_2$ gauge theories using the ADHM-like method, which will appear in arXiv.
\section{$G_2$ gauge theories from an $\widetilde{\text{O5}}$-plane}\label{sec:G2fromO5tilde}
In string theory,
a wide class of
5d theories with eight supercharges can be constructed by $(p, q)$ 5-brane webs in type IIB string theory \cite{Aharony:1997ju, Aharony:1997bh, Leung:1997tw} or M-theory on Calabi-Yau threefolds \cite{Morrison:1996xf, Douglas:1996xp, Intriligator:1997pq}. We will make use of the 5-brane web description for constructing 5d $\mathcal{N}=1$ gauge theories in this paper.
In this section, we
present 5-brane web diagrams which realize 5d $G_2$ gauge theories by using an $\widetilde{\text{O5}}$-plane. Although D5-branes on top of an O5-plane or an $\widetilde{\text{O5}}$-plane usually generate an $SO/USp$ gauge theory, we will argue that some simple 5-brane web diagram with an $\widetilde{\text{O5}}$-plane can yield a 5d $G_2$ gauge theory in an intriguing way.
\subsection{$SO(7)$ gauge theory with spinor matter}
\label{sec:SO7Spinor}
Before constructing 5-brane webs for $G_2$ gauge theories,
we first discuss 5-brane web realization of $SO(7)$ gauge theories using an $\widetilde{\text{O5}}$-plane
in this subsection.
By using this construction,
we will see in section \ref{sec:HiggstoG2}
that a Higgsing of the 5-brane web diagrams of the $SO(7)$ gauge theories with
a hypermultiplet in spinor representation
can generate 5-brane webs for the pure $G_2$ gauge theory.
A 5-brane web diagram for the pure $SO(7)$ gauge theory can be realized using an $\widetilde{\text{O5}}$-plane.
Naively, a 5-brane web with an $\widetilde{\text{O5}}$-plane
may look problematic since the difference of
the RR charge between an $\widetilde{\text{O5}}^+$-plane and an $\widetilde{\text{O5}}^-$-plane is fractional,
which implies the appearance of $(p,q)$ 5-branes with non-integer $p$.
A way out is that an $\widetilde{\text{O5}}$-plane may be thought of as an O5-plane with a half monodromy branch cut associated to a half D7-brane \cite{Zafrir:2015ftn}.
Namely, an effective description of an $\widetilde{\text{O5}}^-$-plane is an O5$^-$-plane and a half D5-brane plus the half monodromy cut. An $\widetilde{\text{O5}}^+$-plane is also effectively described by an O5$^+$-plane plus the half monodromy cut.
Since the monodromy created by this cut is half of the original monodromy associated to one full D7-brane,
it changes the potential fractional charge to integer charge.
The web diagram for the pure $SO(7)$ gauge theory can be derived by
a Higgsing of the $SO(8)$ gauge theory with a hypermultiplet in the vector representation.
Through this process, we will see that the $\widetilde{\text{O5}}$-plane is accompanied by the half monodromy cut.
A 5-brane web diagram of the $SO(8)$ gauge theory with one flavor is
constructed with O5$^-$-plane as given in Figure \ref{fig:SO8w1flvrHiggsa}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=7cm]{SO8w1flvrHiggs1.pdf}\label{fig:SO8w1flvrHiggsa}}
\subfigure[]{
\includegraphics[width=7cm]{SO8w1flvrHiggs2.pdf} \label{fig:SO8w1flvrHiggsb}}
\subfigure[]{
\includegraphics[width=7cm]{SO8w1flvrHiggs3.pdf} \label{fig:SO8w1flvrHiggsc}}
\subfigure[]{
\includegraphics[width=7cm]{SO8w1flvrHiggs4.pdf} \label{fig:SO8w1flvrHiggsd}}
\caption{A Higgsing procedure of the 5-brane web of the $SO(8)$ gauge theory with one flavor to the 5-brane web of the pure $SO(7)$ gauge theory. (a): Moving the flavor D7-brane to the middle of the diagram. The branch cut is denoted by the dashed line. (b): Lowering the flavor D7-brane as well as the bottom color D5-brane to the O5$^-$-plane. (c): Splitting the D7-branes into two half D7-branes. We have an $\widetilde{\text{O5}}^-$-plane between the half D7-branes and there are effectively three fractional D5-branes between the half D7-branes. We also have monodromy branch cuts for the half D7-branes represented by the red and orange dashed lines. (d): Removing the two fractional D5-branes. We have a half D5-brane denoted by the blue line stretched between the $(2, 1)$ 5-brane and the $(1, -1)$ 5-brane. The diagram gives the pure $SO(7)$ gauge theory.}
\label{fig:SO8w1flvrHiggs}
\end{figure}
We put a floating D7-brane to realize one flavor
and we put the branch cut associated with this D7-brane in the left direction.
Note that $(p,q)$ charges of the 5-branes change when they go across this cut.
In order to perform the Higgsing, we lower this D7-brane
as well as the bottom color D5-brane to the position of the O5$^-$-plane. Since the left part of the web diagram crosses the branch cut, the 5-brane charges change accordingly as in Figure \ref{fig:SO8w1flvrHiggsb}.
On the O5$^-$-plane, the D7-brane can be split into two half D7-branes,
generating an $\widetilde{\text{O5}}^-$-plane in between
\cite{Evans:1997hk, Giveon:1998sr, Feng:2000eq, Bertoldi:2002nn}.
Counting the half D5-brane associated to the $\widetilde{\text{O5}}^-$-plane together with the bottom color D5-brane, we have effectively three half D5-branes between these two half D7-branes.
Note also that half monodromy cut appears between the two half D7-branes.
The 5-brane web after the splitting is depicted in Figure \ref{fig:SO8w1flvrHiggsc}.
At this stage, it is possible to move the two half D5-branes between the half D7-branes off the plane of the 5-brane web,
which degree of freedom corresponds to the one-dimensional Higgs branch.
Thus, removing the two half D5-branes infinitely far away should correspond to the Higgsing of the $SO(8)$ gauge theory with one flavor
down to the pure $SO(7)$ gauge theory.
After removing the two half D5-branes,
we see that one half D5-brane
stretch between the $(2, 1)$ 5-brane and the $(1, -1)$ 5-brane
including the one coming from the $\widetilde{\text{O5}}^-$-plane.
In addition, we have one half D5-brane connecting the $(2, 1)$ 5-brane to the left half D7-brane and also another half D5-brane connecting the $(1, -1)$ 5-brane to the right half D7-brane as in Figure \ref{fig:SO8w1flvrHiggsd},
which configuration preserves the s-rule.
Although the 5-brane web diagram in Figure \ref{fig:SO8w1flvrHiggsd} still has an remaining O5-plane, one can change it into an $\widetilde{\text{O5}}^-$-plane by moving the left/right half D7-brane to infinitely left/right respectively.
No half D5-branes are attached to the half D7-branes
after they go across the $(2, 1)$ 5-brane or $(1, -1)$ 5-brane
due to Hanany-Witten effect \cite{Hanany:1996ie}.
Note that the half monodromy cut between the two half D7-brane still remains even after moving them to infinity.
That is, we see that the $\widetilde{\text{O5}}$-plane is accompanied by the half monodromy cut.
The 5-brane web diagram after moving the half D7-branes in the opposite directions is given in Figure \ref{fig:pureSO7}.
This is exactly the web diagram for the pure $SO(7)$ gauge theory,
which have three color D5-branes with the $\widetilde{\text{O5}}^-$-plane.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO8w1flvrHiggs5.pdf}
\caption{Another 5-brane web diagram for the 5d pure $SO(7)$ gauge theory by removing the half D7-branes in the opposite directions. The diagram is constructed with an $\widetilde{\text{O5}}$-plane. The $\widetilde{\text{O5}}$-plane is realized by the O5-plane with the half monodromy cut which is denoted by the orange dashed line. }
\label{fig:pureSO7}
\end{figure}
It is straightforward to include a hypermultiplet in the vector representation of $SO(7)$. The vector matter can be introduced by adding a flavor D5-brane to the 5-brane web diagram of the pure $SO(7)$ gauge theory. In fact, one can also introduce spinor matter to the 5-brane web of the $SO(7)$ gauge theory. From the 5-brane web viewpoint, the spinor matter can be realized non-perturbatively \cite{Zafrir:2015ftn}. To see that, let us consider a 5-brane web diagram in Figure \ref{fig:SO9USp2quiver}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO9USp2quiver.pdf}
\caption{A 5-brane web diagram for the $[1] - SO(9) - USp(2) - \left[\frac{3}{2}\right]$ quiver theory.}
\label{fig:SO9USp2quiver}
\end{figure}
The web diagram gives a $[1] - SO(9) - USp(2) - \left[\frac{3}{2}\right]$ quiver theory. Here $[n] - G$ stands for $n$ flavors attached to the $G$ gauge theory. The quiver theory has a Higgs branch associated to moving a D5-brane off the plane of the 5-brane web. The Higgsing yields a 5-brane web diagram given in Figure \ref{fig:SO7wspinor}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO7USp0quiver.pdf}
\caption{A 5-brane web diagram for a $SO(7)$ gauge theory with spinor matter obtained by removing the D5-brane from the diagram for the $[1] - SO(9) - USp(2) - \left[\frac{3}{2}\right]$ quiver theory.}
\label{fig:SO7wspinor}
\end{figure}
The resulting theory might naively look like the pure $SO(7)$ gauge theory. However, the gauge coupling of the ``$USp(0)$'' gauge group is not still turned off. We may expect some additional degrees of freedom whose mass is the inverse of the gauge coupling. In other words, the 5-brane web appears to have ``$USp(0)$'' instantons. Before the Higgsing, we indeed have $USp(2)$ instantons and the instantons carry charges in the spinor representation of
$SO(9)$.
Hence, after the Higgsing, a natural candidate for the ``$USp(0)$'' instantons is a hypermultiplet in the spinor representation of the $SO(7)$ gauge theory. Namely, the 5-brane web in Figure \ref{fig:SO7wspinor} gives rise to an $SO(7)$ gauge theory with a hypermultiplet in the spinor representation.
\subsection{5-brane web for pure $G_2$ gauge theory}
\label{sec:HiggstoG2}
We then move on to construct a 5-brane web diagram for a pure $G_2$ gauge theory.
By Higgsing the $SO(7)$ gauge theory with one hypermultiplet in the spinor representation yields the pure $G_2$ gauge theory.
Thus, the corresponding process in the 5-brane web diagram for the $SO(7)$ gauge theory with one spinor in Figure \ref{fig:SO7wspinor} should lead to the web diagram for the pure $G_2$ gauge theory.
The web diagram in Figure \ref{fig:SO7wspinor} implies that the theory possesses an $SU(2)$ flavor symmetry
associated to parallel external $(2, 1)$ 5-branes.
This global symmetry is expected to act on the hypermultiplet in the spinor representation.
Thus, the distance of the two parallel external $(2, 1)$ 5-branes should be associated to the mass of this spinor
and the Higgs branch should open up in the massless limit
at certain subspace in the Coulomb moduli.
The question is how one can take the massless limit. It seems to be difficult take the massless limit from the diagram in Figure \ref{fig:SO7wspinor} since we need to
``flop'' the $\widetilde{\text{O5}}^+$-plane between the $(2, 1)$ 5-brane and the $(1, -1)$ 5-brane.
In order to resolve the issue, it turns out to be useful to consider an equivalent but different 5-brane web diagram for the $SO(7)$ gauge theory with one spinor.
Analogous to Figure \ref{fig:pureSO7} being obtained from Figure \ref{fig:SO8w1flvrHiggsd},
the diagram in Figure \ref{fig:SO7wspinor} may be obtained by moving the two half D7-branes in the opposite directions from the diagram in Figure \ref{fig:SO7wspinor2a}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=7cm]{SO7USp0quiver1.pdf} \label{fig:SO7wspinor2a}}
\subfigure[]{
\includegraphics[width=7cm]{SO7USp0quiver2.pdf} \label{fig:SO7wspinor2b}}
\caption{5-brane web diagrams for the $SO(7)$ gauge theory with one spinor. (a): The 5-brane web diagram which is obtained by Higgsing the diagram of the $SO(8)$ gauge theory with one vector and one spinor. Compared to the diagram in Figure \ref{fig:SO8w1flvrHiggsd}, the $(2, 1)$ 5-brane is attached on the right, yielding the spinor matter. Removing the half D7-branes in the opposite direction gives the diagram in Figure \ref{fig:SO7wspinor}. (b): An equivalent diagram to the one in Figure \ref{fig:SO7wspinor2a} (and also to the one in Figure \ref{fig:SO7wspinor}). We move the half D7-branes in the left direction compared to the diagram in Figure \ref{fig:SO7wspinor2a}. Then the monodromy cut disappears and the diagram is constructed with an O5-plane.}
\label{fig:SO7wspinor2}
\end{figure}
Instead, we can consider another deformation from the diagram in Figure \ref{fig:SO7wspinor2a} by moving both the half D7-branes to infinitely left. The resulting 5-brane web is depicted in Figure \ref{fig:SO7wspinor2b}. After this deformation, the monodromy cuts completely disappears from the diagram and hence we have only an O5-plane. Therefore, a 5-brane web diagram with an $\widetilde{\text{O5}}$-plane has an equivalent 5-brane web diagram without an $\widetilde{\text{O5}}$-plane but with only an O5-plane. The transition is done by moving a half D7-brane from the infinitely right to the infinitely left in the diagram with an $\widetilde{\text{O5}}$-plane. We will make use of this transition in this paper.
By using this diagram, we can take the massless limit of the spinor by
using ``generalized flop transition''
\cite{Hayashi:2017btw}.
Let us focus on the local part of the diagram describing the ``$USp(0)$'' gauge theory. On the left side a $(1, -1)$ 5-brane intersects with an O5-plane together with a full D5-brane. On the other hand, a $(2, 1)$ 5-brane intersects with the O5-plane on the right side. The local structure exactly appears in the 5-brane diagram of the $E_2$ theory,
and hence we can perform a generalized flop transition for this local part by using the results in \cite{Hayashi:2017btw}.
Through the process of studying the phase diagram of the $E_2$ theory,
it is proposed that the diagram in Figure \ref{fig:flopO5a}
is flopped either to the one in Figure \ref{fig:flopO5b} or Figure \ref{fig:flopO5c}.
Whether the diagram is flopped to the one in Figure \ref{fig:flopO5b} or Figure \ref{fig:flopO5c} is translated to the sign of the mass parameter of the $USp(0)$ gauge theory, which is the sign of the Coulomb
branch parameter in our case.
Although this transition is obtained for the $E_2$ theory, it is natural to assume that this transition is always available regardless of the detail of the remaining diagram to which this subdiagram is attached.
On one hand, by considering a limit where the D5-brane comes down to the position of an O5-plane
as we did in the process of the Higgsing,
Figure \ref{fig:flopO5a} reduces to Figure \ref{fig:flopO5tildea}, which corresponds to the one appearing in the local part of the diagram in Figure \ref{fig:SO7wspinor2b}.
On the other hand, by the same limit, Figure \ref{fig:flopO5b} and Figure \ref{fig:flopO5c} both reduce to Figure \ref{fig:flopO5tildeb}.
Therefore, we propose that the $USp(0)$ part in Figure \ref{fig:SO7wspinor2b} can be flopped to the form in Figure \ref{fig:flopO5tildeb}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=4cm]{O5flop1.pdf} \label{fig:flopO5a}}
\subfigure[]{
\includegraphics[width=4cm]{O5flop2.pdf} \label{fig:flopO5b}}
\subfigure[]{
\includegraphics[width=4cm]{O5flop3.pdf} \label{fig:flopO5c}}
\caption{A generalized flop transition for a 5-brane web with an O5-plane. Whether the diagram in (a) is flopped to (b) or (c) depends on the sign of the mass parameter associated to the D5-brane. }
\label{fig:flopO5}
\end{figure}
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=6cm]{O5tildeflop1.pdf} \label{fig:flopO5tildea}}
\subfigure[]{
\includegraphics[width=6cm]{O5tildeflop2.pdf} \label{fig:flopO5tildeb}}
\caption{The generalized flop transition in Figure \ref{fig:flopO5} in the case when the height of the D5-brane in Figure \ref{fig:flopO5} is set to zero.}
\label{fig:flopO5tilde}
\end{figure}
The equivalent flop transition in the presence of an $\widetilde{\text{O5}}$-plane is given in Figure \ref{fig:flopO5tilde2}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=6cm]{O5tildeflop3.pdf} \label{fig:flopO5tilde2a}}
\subfigure[]{
\includegraphics[width=6cm]{O5tildeflop4.pdf} \label{fig:flopO5tilde2b}}
\caption{A generalized flop transition from (a) to (b) which is equivalent to the one in Figure \ref{fig:flopO5tilde}.}
\label{fig:flopO5tilde2}
\end{figure}
After considering the transition in Figure \ref{fig:flopO5tilde} for the local $USp(0)$ part in Figure \ref{fig:SO7wspinor2b}, the 5-brane web diagram becomes the one in Figure \ref{fig:SO7wspinorfloppeda}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=7cm]{SO7USp0quivertransformed.pdf} \label{fig:SO7wspinorfloppeda}}
\subfigure[]{
\includegraphics[width=7cm]{SO7USp0quivertransformed3.pdf} \label{fig:SO7wspinorfloppedb}}
\subfigure[]{
\includegraphics[width=7cm]{SO7USp0quivertransformedHiggs.pdf} \label{fig:SO7wspinorfloppedc}}
\subfigure[]{
\includegraphics[width=7cm]{SO7USp0quivertransformedHiggsed.pdf} \label{fig:SO7wspinorfloppedd}}
\subfigure[]{
\includegraphics[width=7cm]{SO7USp0quivertransformedHiggs2.pdf} \label{fig:SO7wspinorfloppede}}
\caption{A Higgsing procedure which gives rise to a 5-brane web diagram for the pure $G_2$ gauge theory. (a): We performed the generalized flop transition compared to the diagram in Figure \ref{fig:SO7wspinor2b}. (b): We performed two standard flop transitions to the diagram in Figure \ref{fig:SO7wspinorfloppeda}. (c): Putting the two parallel external $(2, 1)$ 5-branes on top of each other. We also introduced the $(2, 1)$ 7-branes at each end of the external $(2, 1)$ 7-branes. (d): Removing one $(2, 1)$ 5-brane between the $(2, 1)$ 7-branes. (e): Sending the $(2, 1)$ 5-brane to infinity. The diagram yields the pure $G_2$ gauge theory. }
\label{fig:SO7wspinorflopped}
\end{figure}
Hence, we obtain another 5-brane web diagram given in Figure \ref{fig:SO7wspinorfloppeda} which also realizes the $SO(7)$ gauge theory with one spinor.
From the 5-brane web diagram in Figure \ref{fig:SO7wspinorfloppeda} it is now straightforward to perform the Higgsing:
After two flop transitions, we obtain the diagram in Figure \ref{fig:SO7wspinorfloppedb},
from which we can take the massless limit by putting the parallel external $(2, 1)$ 5-branes on top of each other.
We also tune the Coulomb branch moduli of the $SO(7)$ gauge theory, leaving two Coulomb branch moduli
as in Figure \ref{fig:SO7wspinorfloppedc}.
By putting a $(2, 1)$ 7-brane at each end of the two parallel external $(2, 1)$ 5-branes,
we can move a segment of a $(2, 1)$ 5-brane between the external $(2, 1)$ 7-branes as in Figure \ref{fig:SO7wspinorfloppedd},
which degrees of freedom corresponds to the Higgs branch of the $SO(7)$ gauge theory with one spinor.
Removing the D5-brane implies that we take a far infrared limit of the $SO(7)$ gauge theory with one spinor at the Higgs branch. Then the resulting 5-brane web diagram should describe the pure $G_2$ gauge theory. Therefore, we conclude that the 5-brane web diagram in Figure \ref{fig:SO7wspinorfloppede} realizes the pure $G_2$ gauge theory.
We can write an equivalent diagram with an $\widetilde{\text{O5}}$-plane as in Figure \ref{fig:pureG2}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO7USp0quivertransformedHiggs3.pdf}
\caption{Another 5-brane web with an $\widetilde{\text{O5}}$-plane which gives the pure $G_2$ gauge theory. The diagram can be obtained by applying the same Higgsing procedure in Figure \ref{fig:SO7wspinorflopped} to the diagram in Figure \ref{fig:SO7wspinor}.}
\label{fig:pureG2}
\end{figure}
The diagram in Figure \ref{fig:pureG2} is obtained by moving one of the half D7-brane which was sent to the infinitely left to the infinitely right. Then the D5-brane on the O5-plane disappears and we have an $\widetilde{\text{O5}}$-plane realized with the half monodromy cut.
\subsection{Check from effective prepotential}
\label{sec:G2prep1}
In this subsection,
we
see
evidence that the 5-brane web diagram
in Figure \ref{fig:pureG2}
yields the 5d pure $G_2$ gauge theory from the analysis effective prepotentials.
In general, the effective prepotential of a 5d gauge theory with a gauge group $G$ is given by \cite{Seiberg:1996bd, Morrison:1996xf, Intriligator:1997pq}
\begin{equation}
\mathcal{F}(\phi) = \frac{1}{2}m_0h_{ij}\phi_i\phi_j + \frac{\kappa}{6}d_{ijk}\phi_i\phi_j\phi_k + \frac{1}{12}\left(\sum_{r \in \text{roots}}\left|r\cdot \phi\right|^3 - \sum_f\sum_{w \in {\bf R}_f}\left|w\cdot\phi + m_f\right|^3\right), \label{prepotential}
\end{equation}
where $m_0$ is the inverse of the squared gauge coupling, $\phi_i$ are the Coulomb branch moduli, $\kappa$ is the classical Chern-Simons level and $m_f$ is the mass of a hypermultiplet in the representation ${\bf R}_f$ of $G$. $r$ are the roots of $G$ and $w$ are the weights of the representation ${\bf R}_f$. We also used $h_{ij} = \text{Tr}(T_iT_j)$ and $d_{ijk} = \frac{1}{2}\text{Tr}\left(T_i\{T_j, T_k\}\right)$ where $T_i$ are the Cartan generators of $G$.
In the following, we consider the prepotential of the pure $SO(7)$ gauge theory, the $SO(7)$ gauge theory with a spinor and then the pure $G_2$ gauge theory step by step.
\paragraph{Pure $SO(7)$.}
The first example is the 5d pure $SO(7)$ gauge theory. When we parameterize the Coulomb branch moduli $\phi_i$ in the Dynkin basis, the prepotential of the pure $SO(7)$ gauge theory becomes
\begin{align}
\mathcal{F}_{SO(7)}(\phi) = &m_0(\phi_1^2-\phi_1\phi_2+\phi_2^2-2\phi_2\phi_3+2\phi_3^2)\cr
&+\frac{4}{3}\phi_1^3 - \frac{1}{2}\phi_1^2\phi_2 - \frac{1}{2}\phi_1\phi_2^2 + \frac{4}{3}\phi_2^3 - 3\phi_2^2\phi_3 + 2\phi_2\phi_3^2 + \frac{4}{3}\phi_3^3, \label{pureSO7prep}
\end{align}
where we chose $[2, -1, 0], [-1, 2, -2], [0, -1, 2]$ for the simple roots for defining the Weyl chamber. The tension of a monopole string may be computed by taking a derivative with respect to a Coulomb branch modulus. The monopole string tension from the prepotential \eqref{pureSO7prep} is
\begin{align}
\frac{\partial \mathcal{F}_{SO(7)}}{\partial \phi_1} =& \frac{1}{2}(2\phi_1 - \phi_2)(2m_0 + 4\phi_1 + \phi_2),\label{pureSO7monopole1}\\
\frac{\partial \mathcal{F}_{SO(7)}}{\partial \phi_2} =& \frac{1}{2}(-\phi_1 + 2\phi_2 - 2\phi_3)(2m_0 + \phi_1 + 4\phi_2 - 2\phi_3),\label{pureSO7monopole2}\\
\frac{\partial \mathcal{F}_{SO(7)}}{\partial \phi_3} =& (-\phi_2 + 2\phi_3)(2m_0 + 3\phi_2 + 2\phi_3). \label{pureSO7monopole3}
\end{align}
It is also possible to compute the tension of a monopole string from 5-brane web diagrams.
Monopole strings in a 5d gauge theory are realized by D3-branes which stretch along some face bounded by 5-branes. Therefore, the tension of a monopole string corresponds to the area of a face where a D3-brane can extend. Hence we need to compute the area of faces in the diagram in Figure \ref{fig:pureSO7} for the pure $SO(7)$ gauge theory.
For that, we first need to identify the gauge theory parameters with the length of 5-branes in the diagram in Figure \ref{fig:pureSO7}. The height of the color D5-branes is the Coulomb branch modulus and we denote the height of the bottom color D5-brane, the middle color D5-brane and the top color D5-brane by $a_1, a_2, a_3$ respectively. In order to compute the inverse of the squared gauge coupling, we first turn off all the Coulomb branch moduli. Then the external $(2, 1)$ 5-brane and the external $(1, -1)$ directly intersect with the O5-plane and the length of the D5-branes between the $(2, 1)$ 5-brane and the $(1, -1)$ 5-brane on the O5-plane gives $m_0$. Alternatively, one can extrapolate the external $(2, 1)$ 5-brane and the external $(1, -1)$ 5-brane in Figure \ref{fig:pureSO7} to the position of the O5-plane and measure the distance between the external $(2, 1)$ 5-brane and the external $(1, -1)$ 5-brane on the O5-plane. The gauge theory parameterization for the pure $SO(7)$ gauge theory is summarized in Figure \ref{fig:pureSO7parameter}.
\begin{figure}
\centering
\includegraphics[width=8cm]{pureSO7parameter.pdf}
\caption{A gauge theory parameterization for the pure $SO(7)$ gauge theory. $a_1, a_2, a_3$ are the Coulomb branch moduli and $m_0$ is the inverse of the squared gauge coupling.}
\label{fig:pureSO7parameter}
\end{figure}
By using the parameterization depicted in Figure \ref{fig:pureSO7parameter}, we compute the area of faces in the pure $SO(7)$ diagram in Figure \ref{fig:pureSO7}. We label the faces as in Figure \ref{fig:pureSO7monopole}.
\begin{figure}
\centering
\includegraphics[width=8cm]{pureSO7monopole.pdf}
\caption{Labeling for the three faces in the pure $SO(7)$ diagram.}
\label{fig:pureSO7monopole}
\end{figure}
The area of the three faces becomes
\begin{align}
\textcircled{\scriptsize 1} =&\frac{1}{2}(a_3-a_2)(2m_0+a_2+5a_3),\label{pureSO7monopole4}\\
\textcircled{\scriptsize 2} =&\frac{1}{2}(a_2 - a_1)(2m_0 - a_1 + 3a_2 + 4a_3), \label{pureSO7monopole5}\\
\textcircled{\scriptsize 3} =&\frac{1}{2}a_1(2m_0 + a_1 + 4a_2 + 4a_3). \label{pureSO7monopole6}
\end{align}
We can then compare the are \eqref{pureSO7monopole4}, \eqref{pureSO7monopole5} and \eqref{pureSO7monopole6} with the tension \eqref{pureSO7monopole1}, \eqref{pureSO7monopole2} and \eqref{pureSO7monopole3}. In the computation of the area from the pure $SO(7)$ diagram we parameterized the Coulomb branch moduli $a_1, a_2, a_3$ in the orthonormal basis of $\mathbb{R}^3$. The relation between $a_1, a_2, a_3$ and $\phi_1, \phi_2, \phi_3$ is
\begin{align}
\phi_1 = a_3, \quad \phi_2 = a_2 + a_3, \quad \phi_3 = \frac{1}{2}(a_1 + a_2 + a_3). \label{SO7relation}
\end{align}
Then the comparison of \eqref{pureSO7monopole4}, \eqref{pureSO7monopole5} and \eqref{pureSO7monopole6} with \eqref{pureSO7monopole1}, \eqref{pureSO7monopole2} and \eqref{pureSO7monopole3} yields
\begin{align}
\textcircled{\scriptsize 1} =& \frac{\partial \mathcal{F}_{SO(7)}}{\partial \phi_1},\\
\textcircled{\scriptsize 2} =& \frac{\partial \mathcal{F}_{SO(7)}}{\partial \phi_2},\\
2\times\textcircled{\scriptsize 3} =& \frac{\partial \mathcal{F}_{SO(7)}}{\partial \phi_3}.
\end{align}
Therefore, a D3-brane can stretch along the region $\textcircled{\scriptsize 1}$ or the region $\textcircled{\scriptsize 2}$. On the other hand, one needs to double the area of the region $\textcircled{\scriptsize 3}$, which implies that a D3-brane will not end on the $\widetilde{\text{O5}}^-$-plane and will end on a mirror D5-brane.
\paragraph{$SO(7)$ with a massless spinor.}
The next example is the $SO(7)$ gauge theory with a hypermultiplet in the spinor representation. For simplicity we consider a case where the spinor matter is massless. In this case, there are several phases where the effective mass of the hypermultiplets vanishes. We here choose a phase where $[0, 0, 1], [0, 1, -1], [1, -1, 1], [-1, 0, 1]$ among the weights of the spinor representation are positive. Then the prepotential of the $SO(7)$ gauge theory with a massless spinor becomes
\begin{align}
\mathcal{F}_{SO(7)_s} =&m_0(\phi_1^2 - \phi_1\phi_2 + \phi_2^2 - 2\phi_2\phi_3 + 2\phi_3^2)\cr
&+\frac{4}{3}\phi_1^3 - \phi_1\phi_2^2 + \frac{4}{3}\phi_2^3 - \phi_1^2\phi_3 + \phi_1\phi_2\phi_3 - 3\phi_2^2\phi_3 + 2\phi_2\phi_3^2 + \phi_3^3. \label{SO7wsprep}
\end{align}
The monopole string tension is then
\begin{align}
\frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_1} =& (2\phi_1 - \phi_2)(m_0 + 2\phi_1 + \phi_2 - \phi_3), \label{SO7wsmonopole1}\\
\frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_2} =& (-\phi_1 + 2\phi_2 - 2\phi_3)(m_0 + 2\phi_2 - \phi_3),\label{SO7wsmonopole2}\\
\frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_3} =& m_0(-2\phi_2 + 4\phi_3) - \phi_1^2 + \phi_1\phi_2 -3\phi_2^2 + 4\phi_2\phi_3 + 3\phi_3^2. \label{SO7wsmonopole3}
\end{align}
Let us then compute the area of faces in a diagram for the $SO(7)$ gauge theory with a massless spinor. We need to use a particular diagram which corresponds to the phase we chose to compute the prepotential \eqref{SO7wsprep}. Such a diagram is depicted in Figure \ref{fig:SO7wmlspinor} and we label the four faces in the diagram.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO7wspinormonopole.pdf}
\caption{A 5-brane web diagram of the $SO(7)$ gauge theory with a massless spinor. The mass is related to the length between the parallel $(2, 1)$ 5-branes and they are on top of each other in the massless limit. We also label the four faces in the diagram.}
\label{fig:SO7wmlspinor}
\end{figure}
The Coulomb branch moduli $a_1, a_2, a_3$ are again the height of the bottom color D5-brane, the middle color D5-brane and the top color D5-brane respectively. We also extrapolate the external $(2, 1)$ 5-brane and the external $(1, -1)$ 5-brane on top of the $\widetilde{\text{O5}}$-plane and the distance between the extrapolated external 5-branes is the inverse of the squared gauge coupling $m_0$. The gauge theory parameterization is summarized in Figure \ref{fig:SO7wmlspinorparameter}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO7wspinorparameter.pdf}
\caption{A gauge theory parameterization for the $SO(7)$ gauge theory with a massless spinor. $a_1, a_2, a_3$ are the Coulomb branch moduli and $m_0$ is the inverse of the squared gauge coupling.}
\label{fig:SO7wmlspinorparameter}
\end{figure}
We can now compute the area of the four faces in Figure \ref{fig:SO7wmlspinor} by the parameterization in Figure \ref{fig:SO7wmlspinorparameter}. The result is
\begin{align}
\textcircled{\scriptsize 1} =&\frac{1}{2}(a_3-a_2)(2m_0-a_1+a_2+5a_3),\label{SO7wsmonopole4}\\
\textcircled{\scriptsize 2} =&\frac{1}{2}(a_2 - a_1)(2m_0 - a_1 + 3a_2 + 3a_3), \label{SO7wsmonopole5}\\
\textcircled{\scriptsize 3} =&\frac{1}{2}a_1(2m_0 + a_1 + 3a_2 + 3a_3), \label{SO7wsmonopole6}\\
\textcircled{\scriptsize 4} =&\frac{1}{4}(-a_1^2 + 2a_1a_2 - a_2^2 + 2a_1a_3 + 2a_2a_3 - a_3^2). \label{SO7wsmonopole7}
\end{align}
The comparision of \eqref{SO7wsmonopole4}, \eqref{SO7wsmonopole5}, \eqref{SO7wsmonopole6} and \eqref{SO7wsmonopole7} with \eqref{SO7wsmonopole1}, \eqref{SO7wsmonopole2} and \eqref{SO7wsmonopole3} yields relations
\begin{align}
\textcircled{\scriptsize 1} =& \frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_1},\\
\textcircled{\scriptsize 2} =& \frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_2},\\
2\times\textcircled{\scriptsize 3} + \textcircled{\scriptsize 4} =& \frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_3},
\end{align}
using the relation \eqref{SO7relation}. As in the case of the pure $SO(7)$ gauge theory, we need to double the area of the region $\textcircled{\scriptsize 3}$. In fact, we further need to add the area of the region $\textcircled{\scriptsize 4}$ for the monopole string tension corresponding to $\frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_3}$. This fact becomes important also for the comparison of the monopole tension in the case of the pure $G_2$ gauge theory.
\paragraph{Pure $G_2$.}
Finally we consider the prepotential of the pure $G_2$ gauge theory. When we use the Dynkin basis for parametrizing the Coulomb branch moduli $\phi_i$, the prepotential of the 5d pure $G_2$ gauge theory becomes
\begin{equation}
\mathcal{F}_{G_2}(\phi) = m_0(\phi_1^2 - 3\phi_1\phi_2 + 3\phi_2^2) + \frac{4}{3}\phi_1^3 - 4\phi_1^2\phi_2 + 3\phi_1\phi_2^2 + \frac{4}{3}\phi_2^3, \label{G2prep}
\end{equation}
where we chose $[2, -3]$ and $[-1, 2]$ for the simple roots for defining the Weyl chamber.
Hence the expected monopole tension from the prepotential \eqref{G2prep} is
\begin{eqnarray}
\frac{\partial \mathcal{F}_{G_2}}{\partial \phi_1} &=& (m_0 + 2\phi_1 - \phi_2)(2\phi_1 - 3\phi_2),\label{G2monopole1}\\
\frac{\partial \mathcal{F}_{G_2}}{\partial \phi_2} &=&(- \phi_1 + 2\phi_2)(3m_0 + 4\phi_1 + 2\phi_2). \label{G2monopole2}
\end{eqnarray}
The gauge theory parameterization for the pure $G_2$ gauge theory realized in Figure \ref{fig:pureG2} can be understood in a similar way to the case of the $SO(7)$ gauge theories. We denote the height of the lowest color D5-brane by $a_1$ and the height of the second lowest color D5-brane by $a_2$ as in Figure \ref{fig:pureG2parameter1}.
\begin{figure}
\centering
\includegraphics[width=8cm]{pureG2parameter1.pdf}
\caption{A gauge theory parameterization for the pure $G_2$ gauge theory. $a_1, a_2$ are the Coulomb branch moduli and $m_0$ is related to the inverse of the gauge coupling.}
\label{fig:pureG2parameter1}
\end{figure}
The inverse of the square gauge coupling can be calculated from the distance between the extrapolated external $(2, 1)$ 5-brane and the external $(1, -1)$ 5-brane and it is denoted by $m_0$ in Figure \ref{fig:pureG2parameter1}.
We are now able to compute the area corresponding to the tension of monopoles strings by using the parameters in Figure \ref{fig:pureG2parameter1}. We label four faces in the pure $G_2$ diagram as in Figure \ref{fig:pureG2monopole1}.
\begin{figure}
\centering
\includegraphics[width=8cm]{pureG2monopole1.pdf}
\caption{Labeling for the four faces in the pure $G_2$ diagram. Note that the regions $\textcircled{\scriptsize 1}$ and $\textcircled{\scriptsize 4}$ are not separated by a 5-brane and are connected to each other.}
\label{fig:pureG2monopole1}
\end{figure}
The area of the four regions is
\begin{eqnarray}
\textcircled{\scriptsize 1} &=& a_1(m_0 + 2a_1 + 3a_2),\\
\textcircled{\scriptsize 2} &=& (a_2 - a_1)(m_0 + a_1 + 3a_2),\\
\textcircled{\scriptsize 3} &=& a_1(m_0 + 2a_1 + 3a_2),\\
\textcircled{\scriptsize 4} &=& a_1a_2.
\end{eqnarray}
We can deduce which area we should compare with the monopole string tension \eqref{G2monopole1} and \eqref{G2monopole2} from the analysis of the $SO(7)$ gauge theory with a massless spinor.
For the $SO(7)$ gauge theory with a spinor, the area corresponding to the monopole string tension is $\textcircled{\scriptsize 1}$, $\textcircled{\scriptsize 2}$, $2\times\textcircled{\scriptsize 3} + \textcircled{\scriptsize 4}$. After the Higgsing,
the region $\textcircled{\scriptsize 1}$ is combined with $\textcircled{\scriptsize 4}$, hence the area corresponding to the monopole string tension for the pure $G_2$ gauge theory should be $\textcircled{\scriptsize 2}$ and $\textcircled{\scriptsize 1} + 2\times \textcircled{\scriptsize 3} + \textcircled{\scriptsize 4}$. Such an area is given by
\begin{eqnarray}
\textcircled{\scriptsize 2} &=& (a_2 - a_1)(m_0 + a_1 + 3a_2), \label{G2monopole3}\\
\textcircled{\scriptsize 1} +2\times\textcircled{\scriptsize 3} + \textcircled{\scriptsize 4} &=&a_1(3m_0 + 6a_1 + 10a_2). \label{G2monopole4}
\end{eqnarray}
We compare the tension \eqref{G2monopole1} and \eqref{G2monopole2} computed from the prepotential \eqref{G2prep} with the tension \eqref{G2monopole3} and \eqref{G2monopole4} from the area of the 5-brane web in Figure \ref{fig:pureG2monopole1}. Note that the Coulomb branch moduli $a_1, a_2$ are related to $\phi_1, \phi_2$ by
\begin{equation}
2\phi_1 - 3\phi_2 = a_2 - a_1, \quad -\phi_1 + 2\phi_2 = a_1. \label{G2relation}
\end{equation}
Using the relation \eqref{G2relation}, we can see that
\begin{align}
\textcircled{\scriptsize 2} =&\frac{\partial \mathcal{F}_{G_2}}{\partial \phi_1},\label{Eq:G2monotension1}\\
\textcircled{\scriptsize 1} +2\times\textcircled{\scriptsize 3} + \textcircled{\scriptsize 4} = &\frac{\partial \mathcal{F}_{G_2}}{\partial \phi_2}.\label{Eq:G2monotension2}
\end{align}
We note that the Higgsing of the $SO(7)$ gauge theory with a spinor enforces the parameters for the $SO(7)$ gauge theory to be $a_3=a_1+a_2$ in Figure \ref{fig:SO7wmlspinorparameter}. This means that the $SO(7)$ Coulomb branch moduli satisfy $\phi_1=\phi_3$. With the proper map between the $SO(7)$ Coulomb branch moduli $\phi_i^{SO(7)}$ and the $G_2$ Coulomb moduli $\phi^{G_2}_j$ given by $\phi_1^{SO(7)}\to \phi^{G_2}_2$ and $\phi_2^{SO(7)}\to\phi_1^{G_2}$, one can easily see that the prepotential for 5d $SO(7)$ gauge theory with a spinor \eqref{SO7wsprep} becomes the $G_2$ prepotential \eqref{G2prep}. It follows that the tensions of the monopole strings are also consistent with the Higgsing
\begin{align}\label{Eq:so7toG2Higgsing}
\frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_2} {}\Big|_{\rm Higgsing} &=~ \frac{\partial \mathcal{F}_{G_2}}{\partial \phi_1},
\qquad
\Big(
\frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_1}+\frac{\partial \mathcal{F}_{SO(7)_s}}{\partial \phi_3}\Big) {}\Big|_{\rm Higgsing} = \frac{\partial \mathcal{F}_{G_2}}{\partial \phi_2},
\end{align}
or in other words, through the Higgsing, \eqref{SO7wsmonopole3} $\to$ \eqref{Eq:G2monotension1} and agrees with \eqref{SO7wsmonopole1} $+$\eqref{SO7wsmonopole3}, $\to$ \eqref{Eq:G2monotension2}.
The analysis of the prepotential therefore presents further support for the claim that the diagram in Figure \ref{fig:pureG2} realizes the 5d pure $G_2$ gauge theory.
\subsection{Adding flavors to $G_2$}
\label{sec:addingF}
5d $G_2$ gauge theories may have hypermultiplets in the fundamental representation and the maximal number flavors for a $G_2$ gauge theory to have a 5d UV fixed point is five \cite{Zafrir:2015uaa, Jefferson:2017ahm}.
From the viewpoint of 5-brane webs, one can also add hypermultiplets in the fundamental representation of $G_2$ to the 5-brane web diagram in Figure \ref{fig:pureG2}. There are two ways to introduce flavors for the $G_2$ theory. One way uses the vector matter of the $SO(7)$ gauge theory and the other way utilizes the spinor matter of the $SO(7)$ gauge theory. After the Higgsing from $SO(7)$ to $G_2$, the former becomes the fundamental matter of the $G_2$ gauge theory and the latter becomes the fundamental matter plus a singlet hypermultiplet of the $G_2$ gauge theory. The singlet appears since the spinor representation of $SO(7)$ is the eight-dimensional representation and the fundamental representation of $G_2$ is the seven-dimensional representation. Hence, the Higgsing of the $SO(7)$ gauge theory with a hypermultiplet in the vector representation and a hypermultiplet in the spinor representation gives the $G_2$ gauge theory with one flavor, and similarly the Higgsing of the $SO(7)$ gauge theory with two hypermultiplets in the spinor representation gives the $G_2$ gauge theory with one flavor and a singlet. The two 5-brane diagrams giving the $G_2$ gauge theories with one flavor are depicted in Figure \ref{fig:G2w1flvra} and \ref{fig:G2w1flvrb}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=7cm]{G2w1flvr1.pdf} \label{fig:G2w1flvra}}
\subfigure[]{
\includegraphics[width=7cm]{G2w1flvr2.pdf} \label{fig:G2w1flvrb}}
\caption{5-brane web diagrams of the $G_2$ gauge theory with one flavor. (a): The diagram is obtained by Higgsing the $SO(7)$ gauge theory with one vector and one spinor. The resulting theory is $G_2$ gauge theory with one flavor. (b): The diagram is obtained by Higgsing the $SO(7)$ gauge theory with two spinors. The resulting theory is the $G_2$ gauge theory with one flavor and a singlet.}
\label{fig:G2w1flvr}
\end{figure}
It is straightforward to add more flavors to $G_2$ gauge theories by Higgsing the 5d $SO(7)$ gauge theory with one spinor and more than one hypermultiplets either in the vector representation or the spinor representation. An example of a 5-brane web diagram for the $G_2$ gauge theory with two flavors is depicted in Figure \ref{fig:G2w2flvr}.
\begin{figure}
\centering
\includegraphics[width=8cm]{G2w2flvr.pdf}
\caption{A 5-brane diagram for the $G_2$ gauge theory with two flavors.}
\label{fig:G2w2flvr}
\end{figure}
As for the 5-brane web obtained by Higgsing the $SO(7)$ gauge theories with two spinors, one can also perform the generalized flop transition in Figure \ref{fig:flopO5tilde2} and then the 5-brane diagram becomes the one in Figure \ref{fig:G2w1flvrfloppeda}. It is also possible to obtain an equivalent diagram by moving the half D7-brane associated the monodromy cut from the infinitely right to the infinitely left in Figure \ref{fig:G2w1flvrfloppeda}, and the resulting diagram after the transition is given in Figure \ref{fig:G2w1flvrfloppedb} without any branch cut.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=7cm]{G2w1flvr3.pdf} \label{fig:G2w1flvrfloppeda}}
\subfigure[]{
\includegraphics[width=7cm]{G2w1flvr4.pdf} \label{fig:G2w1flvrfloppedb}}
\caption{5-brane diagrams for the $G_2$ gauge theory with one flavor and a singlet after performing the generalized flop transitions. (a): A 5-brane web for a $G_2$ gauge theory with one flavor which is obtained after performing the generalized flop transition to the diagram in Figure \ref{fig:G2w1flvrb}. The diagram contains only an $\widetilde{\text{O5}}^-$-plane. (b): An equivalent 5-brane web diagram to the one in Figure \ref{fig:G2w1flvrfloppeda}. We move the half D7-brane in the infinitely right to the infinitely left. Then the $\widetilde{\text{O5}}$-plane disappears and we only have an O5-plane. The diagram still yields the $G_2$ gauge theory with one flavor and a singlet. }
\label{fig:G2w1flvrflopped}
\end{figure}
\section{Another 5-brane web for pure $G_2$ gauge theory}
\label{sec:G2fromO5}
In this section, we present another 5-brane web diagram for the 5d pure $G_2$ gauge theory without using an $\widetilde{\text{O5}}$-plane different to the one in section \ref{sec:G2fromO5tilde}. This diagram turns out to be useful for the topological vertex computation in section \ref{sec:Nekrasov}.
\subsection{Higgsing 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor}
In section \ref{sec:G2fromO5tilde}, we used the 5-brane web diagram in Figure \ref{fig:SO7wspinor2} for the 5d $SO(7)$ gauge theory with one spinor and the Higgsing of the diagram yielded the web diagram for the pure $G_2$ gauge theory in Figure \ref{fig:SO7wspinorfloppede}. The 5-brane diagram of the 5d $SO(7)$ gauge theory has been originally obtained by the Higgsing associated to vector matter of the 5d $SO(8)$ gauge theory. In other words, the 5-brane web diagram of the pure $G_2$ gauge theory in Figure \ref{fig:SO7wspinorfloppede} was obtained by the two successive Higgsings from the 5d $SO(8)$ gauge theory with one vector and one spinor. Due to the triality of the 5d $SO(8)$ gauge theory, the 5d $SO(8)$ gauge theory with one vector and one spinor is equivalent to the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor. Therefore, we should again obtain a 5-brane web diagram for the pure $G_2$ gauge theory by Higgsing a 5-brane web for the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor.
To introduce a hypermultiplet in the spinor representation to the 5-brane web diagram for the 5d $SO(8)$ gauge theory, we consider a ``quiver theory'' of $SO(8)- USp(0)$ and the $USp(0)$ instanton plays a role of the spinor matter \cite{Zafrir:2015ftn}. For introducing two spinors, we consider a quiver $USp(0) - SO(8) - USp(0)$. However, we need two spinors of opposite chirality. The difference between a spinor and a conjugate spinor can be realized by considering different discrete theta angles for the two $USp(0)$ gauge groups \cite{Zafrir:2015ftn}. Namely, we consider a 5-brane diagram of the quiver $USp(0) - SO(8)- USp(0)$ but the two $USp(0)$ gauge groups have different discrete theta angles\footnote{There is another 5-brane web diagram for the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor and it is given by a diagram for a quiver $SO(8) - USp(0) - [1]$. However, in order to perform a Higgsing, it is useful to consider a 5-brane web for the $USp(0) - SO(8) - USp(0)$ quiver theory.}.
A 5-brane web diagram of the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor is given in Figure \ref{fig:SO8wspinorAcspinor}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO8wsacs.pdf}
\caption{A 5-brane web diagram for the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor. The discrete theta angle of the two $USp(0)$ theories is different from each other although the difference is not explicitly expressed in the diagram. }
\label{fig:SO8wspinorAcspinor}
\end{figure}
Note that there are two parallel external $(2, 1)$ 5-branes and two parallel external $(2, -1)$ 5-branes. Each two parallel external 5-branes implies an $SU(2)$ flavor symmetry and hence the theory shows an $SU(2) \times SU(2)$ perturbative flavor symmetry from one spinor and one conjugate spinor. At the level of the diagram in Figure \ref{fig:SO8wspinorAcspinor}, the difference of the discrete theta angle is invisible. However, the difference appears after the generalized flop transition \cite{Hayashi:2017btw}. The two different types of the flop transitions depending on the discrete theta angle are depicted in Figure \ref{fig:gflopO5}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=4cm]{USp0flop1.pdf} \label{fig:gflopO5a}}
\subfigure[]{
\includegraphics[width=4cm]{USp0flop2.pdf} \label{fig:gflopO5b}}
\subfigure[]{
\includegraphics[width=4cm]{USp0flop3.pdf} \label{fig:gflopO5c}}
\caption{The generalized flop transition for the 5-brane web for $USp(0)$. Depending on the discrete theta angle, the transition changes the figure (a) into either figure (b) or figure (c).}
\label{fig:gflopO5}
\end{figure}
We can then apply the generalized flop transition in Figure \ref{fig:gflopO5} to the diagram in Figure \ref{fig:SO8wspinorAcspinor} and it yields another 5-brane web diagram for the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor. The resulting diagram is depicted in Figure \ref{fig:SO8wspinorAcspinor2}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO8wsacs2.pdf}
\caption{A 5-brane web diagram for the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor after performing the generalized flop transitions. }
\label{fig:SO8wspinorAcspinor2}
\end{figure}
In order to perform the Higgsing to the 5d pure $G_2$ gauge theory, we consider a further transition given in Figure \ref{fig:flopO5}, yielding a 5-brane web in Figure \ref{fig:SO8wspinorAcspinor3}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO8wsacs3.pdf}
\caption{Another 5-brane web diagram for the 5d $SO(8)$ gauge theory with one spinor and one conjugate spinor. }
\label{fig:SO8wspinorAcspinor3}
\end{figure}
We can use the 5-brane web in Figure \ref{fig:SO8wspinorAcspinor3} to obtain another 5-brane web diagram for the 5d pure $G_2$ gauge theory by two Higgsings. Let us first perform a Higgsing associated to the parallel external $(2, 1)$ 5-branes on the right part in Figure \ref{fig:SO8wspinorAcspinor3}. The procedure is essentially the same as what has been done in Figure \ref{fig:SO7wspinorflopped} and the resulting 5-brane web diagram is given in Figure \ref{fig:SO7wspinor3}.
\begin{figure}
\centering
\includegraphics[width=8cm]{SO7wspinor.pdf}
\caption{Another 5-brane web diagram for the 5d $SO(7)$ gauge theory with one spinor.}
\label{fig:SO7wspinor3}
\end{figure}
Due to the triality, the 5-brane web diagram in Figure \ref{fig:SO7wspinor3} should give rise to the 5d $SO(7)$ gauge theory with one spinor. Hence the diagram in Figure \ref{fig:SO7wspinor3} realizes the $SO(7)$ gauge group without introducing an $\widetilde{\text{O5}}$-plane different from the diagram in Figure \ref{fig:SO7wspinor}.
Since the diagram in Figure \ref{fig:SO7wspinor3} still contains parallel external $(2, -1)$ 5-branes, we can perform a Higgsing associated to them. Note that after the Higgsing, the consistency of the diagram restricts the position of the lowest color D5-brane to the location of the O5-plane. The resulting diagram is depicted in Figure \ref{fig:G2pure2}.
\begin{figure}
\centering
\includegraphics[width=8cm]{pureG2.pdf}
\caption{Another 5-brane web diagram for the 5d pure $G_2$ gauge theory.}
\label{fig:G2pure2}
\end{figure}
The 5-brane web diagram of the pure $G_2$ gauge theory is given without an $\widetilde{\text{O5}}$-plane in this case. In fact, it turns out that this diagram is more useful to apply the topological vertex technique to compute the partition function than the diagram in Figure \ref{fig:pureG2}.
\subsection{Check from effective prepotential}
As we have done in section \ref{sec:G2prep1}, we can give
evidence that the diagram in Figure \ref{fig:G2pure2} yields the pure $G_2$ gauge theory by computing the tension of a monopole string from the diagram. For that, we first associate the gauge theory parameters, the Coulomb branch moduli $a_1, a_2$ and the inverse of the squared gauge coupling $m_0$ to some lengths of 5-branes. $a_1$ is the height of the lowest color D5-brane and $a_2$ is the height of the second lowest D5-brane. $m_0$ is determined by extrapolating the external $(2, 1)$ 5-brane and the external $(2, -1)$ 5-brane to the location of the O5-plane. The relations between the gauge theory parameters and the lengths of 5-branes are depicted in Figure \ref{fig:pureG2parameter2}.
\begin{figure}
\centering
\includegraphics[width=8cm]{pureG2parameter2.pdf}
\caption{A gauge theory parameterization for the pure $G_2$ gauge theory for the diagram in Figure \ref{fig:G2pure2}. $a_1, a_2$ are the Coulomb branch moduli and $m_0$ is related to the inverse of the squared gauge coupling.}
\label{fig:pureG2parameter2}
\end{figure}
The tension of a monopole string is given by the area of a face where a D3-brane stretch. There are five faces in the 5-brane web in Figure \ref{fig:G2pure2} and we denote the five faces by $\textcircled{\scriptsize 1}$ - $\textcircled{\scriptsize 5}$ as in Figure \ref{fig:pureG2monopole2}.
\begin{figure}
\centering
\includegraphics[width=8cm]{pureG2monopole2.pdf}
\caption{Labeling for the five faces in the pure $G_2$ diagram. Note that the regions $\textcircled{\scriptsize 1}$, $\textcircled{\scriptsize 4}$ and $\textcircled{\scriptsize 5}$ are not separated by a 5-brane and are connected to each other.}
\label{fig:pureG2monopole2}
\end{figure}
The areas of the five faces are respectively
\begin{eqnarray}
\textcircled{\scriptsize 1} &=& a_1(m_0 + 2a_1 + 4a_2),\\
\textcircled{\scriptsize 2} &=& (a_2 - a_1)(m_0 + a_1 + 3a_2),\\
\textcircled{\scriptsize 3} &=& a_1(m_0 + 2a_1 + 2a_2),\\
\textcircled{\scriptsize 4} &=& a_1a_2,\\
\textcircled{\scriptsize 5} &=& a_1a_2.
\end{eqnarray}
One face where a D3-brane stretch is $\textcircled{\scriptsize 2}$. As for the other case, note that we needed to double the area of the region $ \textcircled{\scriptsize 3}$ and further add the area $ \textcircled{\scriptsize 4}$ as in Figure \ref{fig:pureG2monopole1}.
Two sequences of Higgsing connect the region $ \textcircled{\scriptsize 1}$, $ \textcircled{\scriptsize 4}$ and $ \textcircled{\scriptsize 5}$.
Hence, in this case the other face should be $\textcircled{\scriptsize 1} + 2\times \textcircled{\scriptsize 3} + \textcircled{\scriptsize 4} + \textcircled{\scriptsize 5}$. Then, the area of the faces corresponding to the monopole string tension is
\begin{eqnarray}
\textcircled{\scriptsize 2} &=& (a_2 - a_1)(m_0 + a_1 + 3a_2), \label{G2monopole5}\\
\textcircled{\scriptsize 1} + 2\times\textcircled{\scriptsize 3} + \textcircled{\scriptsize 4} + \textcircled{\scriptsize 5} &=& a_1(3m_0 + 6a_1 + 10a_2), \label{G2monopole6}
\end{eqnarray}
which exactly reproduce the area \eqref{G2monopole3} and \eqref{G2monopole4} calculated from the diagram in Figure \ref{fig:pureG2}. Therefore the monopole tension computation gives another evidence that the diagram in Figure \ref{fig:G2pure2} yields the 5d pure $G_2$ gauge theory.
\section{5d Nekrasov partition functions of $G_2$ gauge theories}
\label{sec:Nekrasov}
In section \ref{sec:G2fromO5tilde} and section \ref{sec:G2fromO5}, we have constructed 5-brane web diagrams which realize 5d $G_2$ gauge theories. In section \ref{sec:G2fromO5tilde} we presented 5-brane webs for $G_2$ gauge theories using an $\widetilde{\text{O5}}$-plane and in section \ref{sec:G2fromO5} we presented the 5-brane web for the pure $G_2$ gauge theory with an O5-plane.
One of the applications of these 5-brane webs is to compute the BPS partition functions of 5d theories realized by the webs. Since 5-brane webs can be reinterpreted as toric diagrams \cite{Leung:1997tw}, we can apply the topological vertex formalism \cite{Aganagic:2003db, Iqbal:2007ii,Awata:2008ed}.
In \cite{Kim:2017jqn}, the topological vertex formalism for webs with an O5-plane has been developed and hence we can utilize the technique to compute the partition function for the pure $G_2$ gauge theory realized by the 5-brane web with an O5-plane. Although the 5-brane web diagrams in section \ref{sec:G2fromO5tilde} are realized with an $\widetilde{\text{O5}}$-plane, we extend the formalism so that it can apply to webs with an $\widetilde{\text{O5}}$-plane in section \ref{sec:vertexO5tilde}. Then we apply the method to compute the partition functions of the 5d pure $G_2$ gauge theory and the 5d $G_2$ gauge theory with one flavor.
\subsection{Vertex formalism with an O5- and $\widetilde{\text{O5}}$-plane}
\label{sec:vertexO5tilde}
Here, we first briefly review the topological vertex formalism
for 5-brane webs with an O5-plane proposed in \cite{Kim:2017jqn}.
As for the 5-branes which do not touch the O5-plane,
the rule is exactly the same as the conventional topological vertex formalism \cite{Aganagic:2003db}:
We first assign different Young diagrams $Y_i$ to different $(p,q)$ 5-branes.
Then, we introduce the edge factor $(-Q)^{|Y|} f_Y{}^n$ to each edge,
where $Q$ is given by the exponential of the length of the corresponding $(p,q)$ 5-brane.
Here, $f_Y$ is the framing factor defined as
\begin{align}\label{eq:framing}
f_{Y} = (-1)^{|Y|} g^{\frac{1}{2} (||Y^t||^2 - ||Y||^2)},
\end{align}
with $|Y|=\sum_i Y_i$ and $||Y||^2=\sum_i Y_i{}^2$. $n$ is a certain integer which is associated to the relative difference of the framing when we glue two topological vertices.
The parameter $g$ is related to the Omega deformation parameters by
$g=e^{-\epsilon_1} = e^{+\epsilon_2}$. We also introduce the topological vertex $C_{Y_1 Y_2 Y_3}$ to each vertex of a diagram, where
the Young diagrams $Y_1, Y_2, Y_3$ are ordered clockwise.
On top of that, the additional rule is given for $(p,-1)$ and $(-p,-1)$ 5-branes which intersect with each other on the O5-plane
as depicted in Figure \ref{fig:NewRule} (a).
The point is to assign the {\it identical} Young diagram to these two 5-branes and to assign an edge factor
\begin{align}
( + Q_1 Q_2)^{|Y|} f_Y{}^\mathfrak{n},
\qquad
( \mathfrak{n} = p_1 q_2 + p_2 q_1 + 1 ),
\end{align}
corresponding to this part,
where $Q_1$ and $Q_2$ are the exponential of the (rescaled) length of the two 5-branes respectively.
By multiplying all these factors and by summing over all the possible Young diagrams,
we obtain the topological string partition function.
\begin{figure}
\centering
\includegraphics[width=11cm]{new_rule.pdf}
\caption{
(a): The $(p,-1)$ and $(-p,-1)$ 5-branes which intersect with each other on the O5-plane.
(b): An equivalent diagram to (a) but we use the mirror image for a part of the diagram.
}
\label{fig:NewRule}
\end{figure}
Although we can compute the topological string partition function directly using the rules above with a configuration in Figure \ref{fig:NewRule} (a),
it turns out to be more convenient to cut the D5-branes with a finite length
and to use the mirror image for a part of the diagram
so that the two 5-branes which originally intersected with each other on the O5-plane become a single
edge as in Figure \ref{fig:NewRule} (b) \cite{Kim:2017jqn}. This operation corresponds to using the identity
\begin{align}\label{eq:identityC}
C_{Y_1 Y_2 Y_3} = (-1)^{|Y_1|+|Y_2|+|Y_3|} f_{Y_1}^{-1} f_{Y_2}^{-1} f_{Y_3}^{-1}
C_{Y_3^t Y_2^t Y_1^t}.
\end{align}
to all the vertices reflected along the O5-plane
because a clockwise direction is mapped to a counter-clockwise direction under this reflection.
Note that the Young diagrams assigned to the edges reflected along the O5-plane
should be transposed.
After the proper reflection, each sub-diagram can be seen as a strip diagram.
Such sub-amplitudes are already computed in \cite{Iqbal:2004ne}.
The strip amplitudes are written in terms of the product of
the factor defined as
\begin{align}
Z_{\nu}(g) = \prod_{(i,j) \in \nu} (1-g^{\nu_i+\nu_j^t-i-j+1}),
\end{align}
and the factor defined as
\begin{align}
\mathcal{R}_{XY}(Q) =\mathcal{M}(Q)^{-1} \,\mathcal{N}_{X^t Y}(Q),
\end{align}
with
\begin{align}
\mathcal{M}(Q) = {\rm PE} \left[ \frac{g}{(1-g)^2} Q \right],
\end{align}
where $\rm PE$ is the Plethystic exponential defined as
\begin{align}
{\rm PE}[f(\cdot)] = \exp \bigg[ \sum_{i=1}^{\infty} \frac1n f(\cdot^n)\bigg],
\end{align}
and
\begin{align}
\mathcal{N}_{\lambda \mu}(Q) =
\prod_{(i,j) \in \lambda} (1-Q g^{\lambda_i+\mu_j^t-i-j+1} )
\prod_{(i,j) \in \mu} (1-Q g^{-\lambda_j^t-\mu_i+i+j-1} ).
\end{align}
What remains is to glue each strip at the edges where we cut before the reflection
by multiplying corresponding edge factors
and by summing over Young diagram assigned to these edges.
In this process, we also need to take into account the additional factors of the form $(-1)^{|Y|} f_{Y}^{-1}$ in \eqref{eq:identityC}.
If two vertices connected to the same edge are both reflected along the O5-plane,
such contribution cancels with each other as
$(-1)^{|Y|} f_{Y}^{-1} \times (-1)^{|Y^t|} f_{Y^t}^{-1} = 1$.
Only when one vertex is reflected while the other vertex is not, we need to multiply such a factor .
\begin{figure}
\centering
\includegraphics[width=12cm]{O5.pdf}
\caption{Configurations with an O5-plane}
\label{fig:O5config}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=12cm]{O5Mirror.pdf}
\caption{Configurations with an O5-plane with mirror image.}
\label{fig:O5mirror}
\end{figure}
This method can be generalized to webs with an $\widetilde{\text{O5}}$-plane.
As we did in section \ref{sec:SO7Spinor},
we can interpret that an $\widetilde{\text{O5}}$-plane is realized between two fractional D7-branes
on top of the O5-plane placed at infinitely left and infinitely right, respectively \cite{Zafrir:2015ftn, Feng:2000eq}.
The point is to convert an $\widetilde{\text{O5}}$-plane to an O5-plane by moving
one of these fractional D7-branes to the other side by using Hanany-Witten transition
as discussed in Figure \ref{fig:SO7wspinor2}.
After that, it can be seen as a special case of a 5-brane web with an O5-plane.
For example,
as is also mentioned in section \ref{sec:G2fromO5tilde},
the diagram in Figure \ref{fig:O5config} (a) can be understood
as a special case of the ones in Figure \ref{fig:O5config} (b) or (c).
As discussed above,
by considering the mirror image as in Figure \ref{fig:O5mirror} (b) or (c),
we can interpret the configurations as a part of trip diagrams.
Therefore, a configuration with Figure \ref{fig:O5config} (a) should be
also considered as a special case of Figure \ref{fig:O5mirror} (b) or (c),
where we tune the K$\ddot{\text{a}}$hler parameter $\Delta\to1$ so that the position of the D5-brane
comes to exactly the place where the O5-plane exists.
\subsection{Pure $G_2$ gauge theory}
\label{sec:NekpureG2}
We apply the technique of the topological vertex for webs involving an O5-plane or an $\widetilde{\text{O5}}$-plane described in section \ref{sec:vertexO5tilde} to the 5-brane web of 5d $G_2$ gauge theories. We first consider a 5-brane web for the 5d pure $G_2$ gauge theory. So far we have presented two types of the 5-brane webs for the pure $G_2$ gauge theory. One is depicted in Figure \ref{fig:pureG2} with an $\widetilde{\text{O5}}$-plane and the other is depicted in Figure \ref{fig:G2pure2} with an O5-plane. For applying the topological vertex formalism it is appropriate to use the one in Figure \ref{fig:G2pure2} since the diagram in Figure \ref{fig:pureG2} has a configuration where a single 5-brane directly intersects with an O5-plane and we have not yet known the vertex rule for such a configuration.
We use the 5-brane diagram in Figure \ref{fig:G2pure2} to compute the partition function of the 5d pure $G_2$ gauge theory. First we introduce the gauge theory parameters for the diagram as in Figure \ref{fig:pureG2parameter2}. The Coulomb branch moduli $a_1, a_2$ of the pure $G_2$ gauge theory are given by the height of the bottom and the second bottom color D5-brane respectively. We then define $A_1, A_2$ by
\begin{align}
A_1 = e^{- a_1}, \qquad
A_2 = e^{- a_2}, \qquad
\end{align}
which are the parameters directly appearing in the partition function. For the convenience of the later computation, we also introduce
\begin{align}\label{eq:A3=1}
A_0 = 1, \qquad
A_{-1} = A_1^{-1} A_2^{-1},
\end{align}
and define
\begin{align}
Q_{ij} = A_i A_j{}^{-1},
\qquad
(-1 \le j < i \le 2).
\end{align}
The inverse of the squared gauge coupling is given by $m_0$ in Figure \ref{fig:pureG2parameter2} and we define the instanton fugacity by
\begin{align}
q = e^{-m_0}.
\end{align}
Then the K$\ddot{\text{a}}$hler parameters associated to the length of the horizontal lines in Figure \ref{fig:fig01left} are\footnote{The length of the 5-branes in the diagram is given by a linear combination of $a_1, a_2$ and $m_0$. On the other hand, the instanton partition function is written by the exponentiated parameters $A_1, A_2, q$. Hence when we put for example $A_1$ to some length in a diagram, it means that the length is $a_1$. We will make use of this notation for the topological string partition function computation.}
\begin{align}
Q_{B_2} = q A_2{}^{4}, \qquad
Q_{B_1} = Q_{B_0} = q A_1{}^{2} A_2{}^{2}, \qquad
Q_{B_{-1}} = q A_1{}^{4} A_2{}^{4}.
\end{align}
\begin{figure}[t]
\centering
\subfigure[]{\includegraphics[width=6cm]{pureG2top1.pdf}
\label{fig:fig01left} }
\hspace{20mm}
\subfigure[]{\includegraphics[width=6cm]{pureG2top2.pdf}
\label{fig:fig01right} }
\caption{(a): The assignment of the Young diagrams $Y_i, i=-1, 0, 1, 2$ and the K$\ddot{\text{a}}$hler parameter $Q_{B_i}, i=-1, 0, 1, 2$ for the diagram in Figure \ref{fig:G2pure2} of the pure $G_2$ gauge theory. (b): An equivalent diagram to the one in Figure \ref{fig:fig01left}. We use the mirror image for the outmost 5-brane in Figure \ref{fig:fig01left}. As for the bottom 5-brane in Figure \ref{fig:fig01left}, we use the mirror image for the left part and use the original one for the right part.}
\label{fig:pureG2top}
\end{figure}
With the parameterization, we compute the topological string partition function by applying the topological vertex to the 5-brane web in Figure \ref{fig:G2pure2}.
Since the diagram involves an O5-plane, we utilize the method in section \ref{sec:vertexO5tilde}.
First, we cut the color D5-branes in Figure \ref{fig:fig01left} in the middle and
use the mirror image for the outmost 5-brane for describing the diagram as in Figure \ref{fig:fig01right}.
The diagram now consists of two strip diagrams. Note that the diagram in Figure \ref{fig:fig01right} is almost the same as the diagram of the pure $SU(4)$ gauge theory of CS level zero with $A_0$ set to $1$. A difference comes from the assignment of the Young diagram for $Y_0$. In order to see it, recall the diagram of the 5d $SO(7)$ gauge theory with one spinor in Figure \ref{fig:SO7wspinor3}, which is the diagram before Higgsing to the one in Figure \ref{fig:G2pure2}. Let us cut then the diagram in Figure \ref{fig:SO7wspinor3} in the middle into the left part and the right part and use the mirror image for a part of the diagram to apply the topological vertex computation in section \ref{sec:vertexO5tilde}. Then we use the mirror image for the bottom color D5-brane in the left part while we use the original bottom color D5-brane for the right part in the upper half plane. After the Higgsing, the bottom color D5-brane in Figure \ref{fig:SO7wspinor3} becomes the color D5-brane on top of an O5$^-$-plane in Figure \ref{fig:G2pure2}. Therefore, when we consider the diagram in Figure \ref{fig:fig01right}, we should transpose the Young diagram $Y_0$ in the left strip while we keep the Young diagram $Y_0$ in the right strip. This is essentially the only difference from the 5-brane web of the pure $SU(4)$ gauge theory. The Young diagram assignment is summarized in Figure \ref{fig:fig01right}.
It is now straightforward to apply the topological vertex formalism to the diagram in Figure \ref{fig:fig01right}. It is useful to compute the partition function for the left strip and the right strip separately first and glue them together later. The partition function for the left strip is given by
\begin{align}\label{strip}
Z^{\rm strip1} (\{ Q_{ij} \}; Y_{-1}, Y_{0},Y_{1},Y_{2}) =
\prod_{i=-1}^2 g^{\frac{1}{2} ||Y_i||^2}
\prod_{i=-1}^2 Z_{Y_i}(g)
\prod_{-1 \le j < i \le 2} \mathcal{R}_{Y_i Y_j^t} (Q_{ij})^{-1}.
\end{align}
The partition function of the right strip can be expressed as
$Z^{\rm strip2} (\{ Q_{ij} \}; Y_{-1}, Y_{0}^t ,Y_{1},Y_{2})$ with
\begin{align}
Z^{\rm strip2} (\{ Q_{ij} \}; Y_{-1}, Y_{0},Y_{1},Y_{2}) =
\prod_{i=-1}^2 g^{\frac{1}{2} ||Y_i^t||^2}
\prod_{i=-1}^2 Z_{Y_i}(g)
\prod_{-1 \le j < i \le 2} \mathcal{R}_{Y_i Y_j^t} (Q_{ij})^{-1}.
\end{align}
The rest is gluing the contribution of the left strip and the right strip to each other. When we glue these two sub-diagram, we should also take into account the effect of flipping the Young diagram, which is given by the extra factor of the form $(-1)^{|Y_0|} f_{Y_0}$ coming from the identity \eqref{eq:identityC}.
Note that such factor for the upper most D5-brane cancels out as discussed in section \ref{sec:vertexO5tilde}
since the whole edge is reflected.
Then, the topological string partition function for the diagram in Figure \ref{fig:fig01right} is
\begin{align}\label{part.pureG2}
Z_{G_2} = & \sum_{\{Y_i\}}
(-1)^{|Y_0|} f_{Y_2}^{3} f_{Y_1} f_{Y_0} f_{Y_{-1}}^{-3} \prod_{i=-1}^2 (-Q_{B_i}){}^{| Y_i |}
\cr
& Z^{\rm strip1} (\{ Q_{ij} \}; Y_{-1}, Y_{0},Y_{1},Y_{2})
Z^{\rm strip2} (\{ Q_{ij} \}; Y_{-1}, Y_{0}^t ,Y_{1},Y_{2}).
\end{align}
We claim that this is the partition function of the 5d pure $G_2$ gauge theory up to the perturbative contribution of the Cartan subalgebra of the pure $G_2$ which should be added by hand in the topological vertex computation. Notice that the only difference between this $G_2$ partition function and the pure $SU(4)$ partition function with $A_0=1$ in \eqref{eq:A3=1} is as follows: The Young diagram $Y_0$ in $Z^{\rm strip2}$ is transposed, the power of framing factor $f_{Y_0}$ is 1 instead of $-1$,
and also the associated sign factor $(-)^{|Y_0|}$ is present.
\paragraph{Perturbative part.}
Let us look at the partition function \eqref{part.pureG2} in more detail. The perturbative part of the partition function is obtained by considering a limit $q \to 0$. The limit corresponds to the restriction of the Young diagrams $Y_i = \emptyset$ for $i=-1, 0, 1, 2$. Then the partition function \eqref{part.pureG2} is simplifies and is given by
\begin{align}\label{pert.pureG2}
Z_{\text{pert}} =&\mathcal{M}(A_1)^2\mathcal{M}(A_1^{-1}A_2)^2\mathcal{M}(A_2)^2\mathcal{M}(A_1A_2)^2\mathcal{M}(A_1^2A_2)^2\mathcal{M}(A_1A_2^2)^2\cr
=&{\rm PE} \left(
\frac{2g}{(1-g)^2}
(A_1 + A_2 + A_1 A_2 + A_1{}^{-1} A_2
+ A_1 A_2{}^{2} + A_1{}^{2} A_2)
\right).\
\end{align}
We can compare \eqref{pert.pureG2} with the field theory result of the perturbative part of the partition function of the pure $G_2$ gauge theory, which is given by
\begin{align}\label{locpert.pureG2}
Z'_{\text{pert}} = {\rm PE} \left(
\frac{g}{(1-g)^2} \chi_{14}
\right),
\end{align}
where $\chi_{14}$ is the character of the adjoint representation,
\begin{align}
\chi_{14}{} =&
A_1 + A_1{}^{-1} + A_2 + A_2{}^{-1}
+ A_1 A_2 + A_1{}^{-1} A_2{}^{-1}
\cr
&
+ A_1 A_2{}^{-1} + A_1{}^{-1} A_2
+ A_1 A_2{}^2 + A_1{}^{-1} A_2{}^{-2}
+ A_1{}^2 A_2 + A_1{}^{-2} A_2{}^{-1}
+ 2.
\end{align}
The partition function \eqref{pert.pureG2} obtained from the topological vertex is indeed consistent with \eqref{locpert.pureG2} up to the terms independent of the Coulomb moduli, namely the Cartan part,
and up to the procedure corresponding to the ``flop transition''
\begin{align}
{\rm PE} \left( \frac{g}{(1-g)^2} Q \right)
\to
{\rm PE} \left( \frac{g}{(1-g)^2} Q^{-1} \right).
\end{align}
\paragraph{Instanton part.}
Let us then look at the instanton part of the partition function \eqref{part.pureG2}. The instanton part is given by removing the perturbative part
\begin{align}\label{inst.pureG2}
Z_{G_2,\text{inst}} = \frac{Z_{G_2}}{Z_{\text{pert}}} = \sum_{k}Z_kq^k.
\end{align}
The $k$-instanton contribution is given by the function $Z_k$ for the $q^k$ order. The explicit form of the one-instanton contribution is
\begin{align}\label{1inst.pureG2}
Z_1 =
\frac{
2 g A_1{}^3 A_2{}^3 (1 + A_1 + A_1 A_2) (1 + A_2 + A_1 A_2)
}{
(1 - g)^2
(A_1 - A_2)^2
(1 - A_1{}^2 A_2)^2
(1 - A_1 A_2{}^2)^2
}.
\end{align}
The explicit form of the two-instanton contribution is
\begin{align}\label{2inst.pureG2}
Z_2 = g^5 A_1{}^{10} A_2{}^{10} \frac{({\rm Numerator})}{({\rm Denominator})},
\end{align}
with
\begin{align}
({\rm Denominator}) =
&(1 - g)^4
(1 + g)^2
(A_1 - A_2)^2
(1 - A_1{}^2 A_2)^2
(1 - A_1 A_2{}^2)^2
\cr
&\times
(1 - g A_1{}^2 A_2)^2
(1 - g A_1 A_2{}^2)^2
(1 - g^{-1} A_1{}^2 A_2 )^2
(1 - g^{-1} A_1 A_2{}^2 )^2
\cr
& \times
(1 - g A_1 A_2{}^{-1})^2
(1 - g A_1{}^{-1} A_2)^2,
\end{align}
and
\begin{align}
& \!\!\!\!\!\!\!\!\!\! ({\rm Numerator})
\cr
= &
(-2 g_1 + 6) \chi_{14}{}^3
+ (3 g_1 - 1) \chi_{7}{}^2 \chi_{14}{}^2
+ (10 g_1 + 2) \chi_{7}{} \chi_{14}{}^2
+ (g_3 + 5 g_2 - g_1 + 3) \chi_{14}{}^2 \cr
&
+ (2 g_2 + 16 g_1 + 22) \chi_{7}{}^3 \chi_{14}
+ (38 g_2 + 66 g_1 + 12) \chi_{7}{}^2 \chi_{14}\cr
&+ (26 g_3 + 66 g_2 + 38 g_1 + 14) \chi_{7} \chi_{14}
+ (2 g_4 + 24 g_3 + 18 g_2 + 12 g_1 + 8) \chi_{14}\cr
&
+ (-8 g_1 - 8) \chi_{7}{}^5
+ (-18 g_2 - 11 g_1 + 3) \chi_{7}{}^4
+ (-12 g_3 + 4) \chi_{7}{}^3 \cr
&+ (13 g_3 + 7 g_2 - 16 g_1 - 38) \chi_{7}{}^2
+ (6 g_4 + 12 g_3 - 22 g_2 - 60 g_1 - 8) \chi_{7}
\cr
&
+ (2 g_4 - 9 g_3 - 19 g_2 - 21 g_1 - 21),
\end{align}
where
\begin{align}
g_n =& \sum_{k=-n}^n g^k,
\\
\chi_{7}{} =&
A_1 + A_1{}^{-1}
+ A_2 + A_2{}^{-1}
+ A_1 A_2 + A_1{}^{-1} A_2{}^{-1}
+ 1,
\\
\chi_{14}{} =&
A_1 + A_1{}^{-1} + A_2 + A_2{}^{-1}
+ A_1 A_2 + A_1{}^{-1} A_2{}^{-1}
\cr
&
+ A_1 A_2{}^{-1} + A_1{}^{-1} A_2
+ A_1 A_2{}^2 + A_1{}^{-1} A_2{}^{-2}
+ A_1{}^2 A_2 + A_1{}^{-2} A_2{}^{-1}
+ 2.
\end{align}
We can compare the result \eqref{1inst.pureG2} and \eqref{2inst.pureG2} with the field theory result. The explicit expression of the one-instanton partition function of the pure $G_2$ gauge theory is obtained in \cite{Benvenuti:2010pq, Keller:2011ek}, and it is generalized to higher instantons in \cite{Hanany:2012dm, Keller:2012da, Cremonesi:2014xha}. We checked that the one-instanton partition function \eqref{1inst.pureG2} and the two-instanton partition function \eqref{2inst.pureG2} perfectly agree with the known results.
\subsection{$G_2$ gauge theory with one flavor}
In section \ref{sec:NekpureG2}, we have computed the partition function of the pure $G_2$ gauge theory from the topological vertex. In this section, we apply the method to a diagram for the 5d $G_2$ gauge theory with one flavor. We have two types of the diagram for the $G_2$ gauge theory with one flavor in Figure \ref{fig:G2w1flvr}. In order to apply the topological vertex, we need to avoid a configuration where a single 5-brane intersects with an O5-plane. Hence we will use the diagram in Figure \ref{fig:G2w1flvrb} or equivalently the one in Figure \ref{fig:G2w1flvrfloppedb}, which yields the 5d $G_2$ gauge theory with one flavor and a singlet, for the application of the topological vertex
In order to apply the topological vertex to the 5-brane web diagram in Figure \ref{fig:G2w1flvrfloppedb}, we first divide the diagram into the left strip and the right strip and assign the K$\ddot{\text{a}}$hler parameters as in Figure \ref{fig:G2w1flvrTop}.
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=6cm]{G2strip1.pdf} \label{fig:G2w1flvrTopa}}
\subfigure[]{
\includegraphics[width=6cm]{G2strip2.pdf} \label{fig:G2w1flvrTopb}}
\caption{(a): The left strip. (b): The right strip. }
\label{fig:G2w1flvrTop}
\end{figure}
The diagram in Figure \ref{fig:G2w1flvrTopa} gives the left strip and the one in Figure \ref{fig:G2w1flvrTopb} gives the right strip. Instead of using the upper half plane with an O5-plane, we write the diagram as a strip diagram by using the mirror image as discussed in section \ref{sec:vertexO5tilde}. Then, the application of the topological vertex to the diagram in Figure \ref{fig:G2w1flvrTop} yields the result
\begin{align}\label{leftstrip}
&Z^{\rm left strip} (\{ Q_i \}; \mu,\nu,\rho,\sigma) =
g^{\frac{1}{2} ( || \mu ||^2 + || \nu ||^2 + || \rho ||^2) + || \sigma^t ||^2}
Z_{\mu} (g) Z_{\nu} (g) Z_{\rho} (g) Z_{\sigma} (g){}^2
\cr
&\frac{
\mathcal{R}_{\rho \sigma^t} (Q_1)
\mathcal{R}_{\nu \sigma^t} (Q_{1,2})
\mathcal{R}_{\rho \emptyset} (Q_{0,1})
\mathcal{R}_{\sigma \sigma} (Q_{0,0})
\mathcal{R}_{\mu \sigma^t} (Q_{1,2,3})
\mathcal{R}_{\nu \emptyset} (Q_{0,1,2} )
\mathcal{R}_{\mu \emptyset} (Q_{0,1,2,3})
}{
\mathcal{R}_{\mu \nu^t}(Q_3)
\mathcal{R}_{\nu \rho^t} (Q_2)
\mathcal{R}_{\mu \rho^t} (Q_{2,3})
\mathcal{R}_{\rho \sigma} (Q_{0,0,1} )
\mathcal{R}_{\nu \sigma} (Q_{0,0,1,2})
\mathcal{R}_{\mu \sigma} (Q_{0,0,1,2,3})
},
\end{align}
where we denote
\begin{align}
Q_{i,j}=Q_i Q_j, \qquad Q_{i,j,k}=Q_i Q_j Q_k, \qquad Q_{i,j,k,\ell}=Q_i Q_j Q_k Q_{\ell}, \quad \cdots
\end{align}
for the left strip.
As for the right strip, we obtain
\begin{align}\label{rightstrip}
&Z^{\rm right strip} (\{ \tilde{Q}_i \}; \mu,\nu,\rho) =
g^{\frac{1}{2} ( || \mu ||^2 + || \nu^t ||^2 + || \rho^t ||^2) }
Z_{\mu} (g) Z_{\nu} (g) Z_{\rho} (g)
\cr
&\frac{1}{
\mathcal{R}_{\nu \rho^t}(\tilde{Q}_3)
\mathcal{R}_{\rho \emptyset}(\tilde{Q}_{2})
\mathcal{R}_{\emptyset \mu}(\tilde{Q}_{1})
\mathcal{R}_{\nu \emptyset} (\tilde{Q}_{2,3})
\mathcal{R}_{\rho \mu} (\tilde{Q}_{1,2} )
\mathcal{R}_{\nu \mu} (\tilde{Q}_{1,2,3})
},
\end{align}
with
\begin{align}
\tilde{Q}_1 = Q_0 Q_1 Q_2 Q_3, \qquad
\tilde{Q}_2 = Q_0 Q_1, \qquad
\tilde{Q}_3 = Q_2.
\end{align}
Note also that the web diagram gives a constraint
\begin{equation}
Q_3 = Q_0Q_1. \label{Q3}
\end{equation}
The full partition function is given by gluing the left strip \eqref{leftstrip} to the right strip \eqref{rightstrip} with framing factors and the final result is given by
\begin{align}\label{G2part1}
\tilde{Z}_{G_2, N_f=1} = &\sum_{\mu,\nu,\rho,\sigma}
(+Q_B Q_3{}^2)^{|\mu|} (-Q_B)^{|\nu|+|\rho|} (+Q_0)^{|\sigma|}
f_{\mu}{}^{3} f_{\nu}{}^{} f_{\rho}{}^{-1} f_{\sigma}
\cr
&Z^{\rm left strip} (\{ Q_i \}; \mu,\nu,\rho,\sigma)
Z^{\rm right strip} (\{ \tilde{Q}_i \}; \mu,\nu,\rho),
\end{align}
where $Q_B$ is the K$\ddot{\text{a}}$hler parameter for the bottom color D5-brane in Figure \ref{fig:G2w1flvrflopped}. Since the 5-brane web in Figure \ref{fig:G2w1flvrfloppedb} has parallel external legs. We need to remove an extra factor \cite{Bergman:2013ala, Bao:2013pwa, Hayashi:2013qwa, Bergman:2013aca, Hwang:2014uwa}. Hence the partition function after removing the extra factor is
\begin{equation}
Z_{G_2, N_f=1} = \frac{\tilde{Z}_{G_2, N_f=1}}{\mathcal{M}(Q_0Q_1^3Q_2^2Q_3)}. \label{G2part}
\end{equation}
We claim that the partition function \eqref{G2part} is the partition function of the $G_2$ gauge theory with one flavor and a singlet.
Since the partition function \eqref{G2part} is still written by the K$\ddot{\text{a}}$hler parameter of the 5-brane web in Figure \ref{fig:G2w1flvrTop}, we will rewrite it by the gauge theory parameters. First the $G_2$ gauge theory should have two Coulomb branch moduli $A_1, A_2$
and one instanton fugacity $q$. There is also a mass parameter $M$ for the flavor. Since the one flavor and the singlet both come from the same spinor matter of the $SO(7)$ gauge theory, the mass parameter for the singlet is the same as the mass parameter for the one flavor.
The Coulomb branch parameterization is essentially the same as the parameterization in the case of the pure $G_2$ gauge theory. Namely, the Coulomb branch moduli are the height of the color D5-branes. Hence we impose
\begin{equation}
Q_0Q_1 = A_1, \quad Q_0Q_1Q_2 = A_2. \label{CB}
\end{equation}
Note that the web diagram in Figure \ref{fig:G2w1flvrfloppedb} has parallel $(1, -1)$ 5-branes. The parallel $(1, -1)$ 5-branes imply an $SU(2)$ flavor symmetry for the one flavor of the $G_2$ gauge theory. Therefore, the distance between the two parallel $(1, -1)$ 5-branes is the mass parameter. This leads to a condition
\begin{equation}
Q_0Q_1^3Q_2^2Q_3 = M^2. \label{Mass}
\end{equation}
The instanton fugacity can by computed similarly to the way when we calculated the instanton fugacity for the 5d pure $G_2$ gauge theory. We extrapolate the leftmost external $(1,-1)$ 5-brane and the external $(2, 1)$ 5-brane on the right to the position of the O5-plane in Figure \ref{fig:G2w1flvrfloppedb} and compute the length between the two extrapolated 5-branes on the O5-plane. In fact it turns out that we need to extrapolate the leftmost external $(1, -1)$ 5-brane instead of the other external $(1, -1)$ 5-brane. This can be justified by the comparison with the instanton fugacity obtained by another 5-brane web digram for the $G_2$ gauge theory with one flavor in Figure \ref{fig:G2w1flvra}. Then the instanton fugacity is given by
\begin{equation}
Q_BQ_0^{-1}Q_1Q_2^{-1}Q_3^{-1} = q. \label{instfugacity}
\end{equation}
Hence the relations \eqref{CB}, \eqref{Mass}, \eqref{instfugacity} with the condition \eqref{Q3} yields
\begin{align}\label{gaugeparameters}
Q_0 = M^{-1} A_1 A_2, \quad
Q_1 = M A_2{}^{-1}, \quad
Q_2 = A_1{}^{-1} A_2, \quad
Q_B = q M{}^{-2} A_1 A_2{}^3.
\end{align}
Namely, we can compare the partition function \eqref{G2part} with the parameterization \eqref{gaugeparameters} with the Nekrasov partition function of the $G_2$ gauge theory with one flavor and a singlet.
\paragraph{Perturbative part.}
Since $Q_B$ is proportional to the instanton fugacity $q$, the perturbative part of the partition function can be obtained by setting $Q_B \to 0$. The limit corresponds to the restriction $\mu = \emptyset, \nu = \emptyset, \rho=\emptyset$. In this case, the sum of Young diagrams for the right strip partition function disappears and it reduces to
\begin{eqnarray}
Z^{\text{right}}_{\text{pert}} &=&\mathcal{M}(\tilde{Q}_3)
\mathcal{M}(\tilde{Q}_{2})
\mathcal{M}(\tilde{Q}_{1})
\mathcal{M}(\tilde{Q}_{2,3})
\mathcal{M}(\tilde{Q}_{1,2} )
\mathcal{M}(\tilde{Q}_{1,2,3})\nonumber\\
&=&\mathcal{M}(A_1^{-1}A_2)\mathcal{M}(A_1)\mathcal{M}(A_1A_2)\mathcal{M}(A_2)\mathcal{M}(A_1^2A_2)\mathcal{M}(A_1A_2^2).
\end{eqnarray}
On the other hand, one Young diagram summation remains in the left strip partition function and it becomes
\begin{eqnarray}
Z^{\text{left}}_{\text{pert}} &=& Z^{\text{left}}_{\text{pert 1}}Z^{\text{left}}_{\text{pert 2}},\nonumber\\
Z^{\text{left}}_{\text{pert 1}}&=&\frac{
\mathcal{M}(Q_3)
\mathcal{M}(Q_2)
\mathcal{M}(Q_{2,3})
\mathcal{M}(Q_{0,0,1} )
\mathcal{M}(Q_{0,0,1,2})
\mathcal{M}(Q_{0,0,1,2,3})}{
\mathcal{M}(Q_1)
\mathcal{M}(Q_{1,2})
\mathcal{M}(Q_{0,1})
\mathcal{M}(Q_{0,0})
\mathcal{M}(Q_{1,2,3})
\mathcal{M}(Q_{0,1,2} )
\mathcal{M}(Q_{0,1,2,3})},\label{leftpertproduct}\\
Z^{\text{left}}_{\text{pert 2}}&=&\sum_{\sigma}Q_0^{|\sigma|}f_{\sigma}g^{||\sigma^t||^2}Z_{\sigma}^2(g)\frac{
\mathcal{N}_{\emptyset \sigma^t} (Q_1)
\mathcal{N}_{\emptyset \sigma^t} (Q_{1,2})
\mathcal{N}_{\sigma^t \sigma} (Q_{0,0})
\mathcal{N}_{\emptyset \sigma^t} (Q_{1,2,3})
}{
\mathcal{N}_{\emptyset \sigma} (Q_{0,0,1} )
\mathcal{N}_{\emptyset \sigma} (Q_{0,0,1,2})
\mathcal{N}_{\emptyset \sigma} (Q_{0,0,1,2,3})
}.\label{leftpertsum}
\end{eqnarray}
By the explicit computation of the Young diagram summation of \eqref{leftpertsum}, we argue that the partition function \eqref{leftpertsum} becomes
\begin{eqnarray}
Z^{\text{left}}_{\text{pert 2}}&=&\frac{
\mathcal{M}(Q_{0,0})\mathcal{M}(Q_{0,1})\mathcal{M}(Q_{0,1,2})\mathcal{M}(Q_{0,0,1,1,2})\mathcal{M}(Q_{0,1,2,3})\mathcal{M}(Q_{0,0,1,1,2,3})}{\mathcal{M}(Q_0)\mathcal{M}(Q_{0,0,1})\mathcal{M}(Q_{0,0,1,2})\mathcal{M}(Q_{0,1,1,2})\mathcal{M}(Q_{0,0,1,2,3})\mathcal{M}(Q_{0,1,1,2,3})}\nonumber\\
&&\times\frac{
\mathcal{M}(Q_{0,0,1,1,2,2,3})\mathcal{M}(Q_{0,1,1,1,2,2,3})
}{
\mathcal{M}(Q_{0,1,1,2,2,3})\mathcal{M}(Q_{0,0,1,1,1,2,3})
}.\label{leftpertsumup}
\end{eqnarray}
We checked the equality \eqref{leftpertsumup} until the order $Q_0^6$. Then combining \eqref{leftpertsumup} with \eqref{leftpertproduct} yields
\begin{eqnarray}
Z^{\text{left}}_{\text{pert}}&=&\frac{
\mathcal{M}(Q_3)\mathcal{M}(Q_2)\mathcal{M}(Q_{2,3})\mathcal{M}(Q_{0,0,1,1,2})\mathcal{M}(Q_{0,0,1,1,2,3})\mathcal{M}(Q_{0,0,1,1,2,2,3})
\mathcal{M}(Q_{0,1,1,1,2,2,3})
}{
\mathcal{M}(Q_{1})\mathcal{M}(Q_{1,2})\mathcal{M}(Q_{1,2,3})\mathcal{M}(Q_0)\mathcal{M}(Q_{0,1,1,2})\mathcal{M}(Q_{0,1,1,2,3})\mathcal{M}(Q_{0,1,1,2,2,3})\mathcal{M}(Q_{0,0,1,1,1,2,2,3})
}\nonumber\\
&=&\frac{
\mathcal{M}(A_1)\mathcal{M}(A_1^{-1}A_2)\mathcal{M}(A_2)\mathcal{M}(A_1A_2)\mathcal{M}(A_1^2A_2)\mathcal{M}(A_1A_2^2)\mathcal{M}(M^2)
}{\mathcal{M}(MA_2^{-1})\mathcal{M}(MA_1^{-1})\mathcal{M}(M)\mathcal{M}(M^{-1}A_1A_2)\mathcal{M}(M)\mathcal{M}(MA_1)\mathcal{M}(MA_2)\mathcal{M}(MA_1A_2)
}.\nonumber\\
\end{eqnarray}
Therefore, the perturbative part of the partition function of \eqref{G2part} is
\begin{eqnarray}
Z_{\text{pert}} &=& \frac{Z^{\text{left}}_{\text{pert}}Z^{\text{right}}_{\text{pert}}}{\mathcal{M}(Q_0Q_1^3Q_2^2Q_3)}\nonumber\\
&=&\frac{
\mathcal{M}(A_1)^2\mathcal{M}(A_1^{-1}A_2)^2\mathcal{M}(A_2)^2\mathcal{M}(A_1A_2)^2\mathcal{M}(A_1^2A_2)^2\mathcal{M}(A_1A_2^2)^2
}{\mathcal{M}(MA_2^{-1})\mathcal{M}(MA_1^{-1})\mathcal{M}(M)\mathcal{M}(M^{-1}A_1A_2)\mathcal{M}(M)\mathcal{M}(MA_1)\mathcal{M}(MA_2)\mathcal{M}(MA_1A_2)
}.\nonumber\\\label{pertpart}
\end{eqnarray}
The perturbative partition function then consists of three factors as follows:
\begin{eqnarray}
Z_{\text{pert}} = Z_{\text{pert}}^{\text{vector}} \cdot Z_{\text{pert}}^{\text{fund}} \cdot Z_{\text{pert}}^{\text{singlet}},
\end{eqnarray}
where
\begin{eqnarray}
Z_{\text{pert}}^{\text{vector}} &=&\mathcal{M}(A_1)^2\mathcal{M}(A_1^{-1}A_2)^2\mathcal{M}(A_2)^2\mathcal{M}(A_1A_2)^2\mathcal{M}(A_1^2A_2)^2\mathcal{M}(A_1A_2^2)^2,\label{pertvector}\\
Z_{\text{pert}}^{\text{fund}} &=&\frac{1}{\mathcal{M}(MA_2^{-1})\mathcal{M}(MA_1^{-1})\mathcal{M}(M^{-1}A_1A_2)\mathcal{M}(M)\mathcal{M}(MA_1)\mathcal{M}(MA_2)\mathcal{M}(MA_1A_2)},\nonumber\\\label{pertflavor}\\
Z_{\text{pert}}^{\text{singlet}} &=&\frac{1}{\mathcal{M}(M)}.\label{pertsinglet}
\end{eqnarray}
\eqref{pertvector}, \eqref{pertflavor} and \eqref{pertsinglet} are exactly equal to the perturbative part of the partition functions of the $G_2$ vector multiplets, hypermultiplets in the fundamental representation of $G_2$ and a singlet hypermultiplet up to the Cartan parts and flop transitions as in the case of the pure $G_2$ gauge theory.
\paragraph{Instanton part.}
Let us move on to the instanton part of the $G_2$ partition function \eqref{G2part}. The instanton part is obtained after removing the perturbative part obtained in \eqref{pertpart}. Namely the instanton partition function of the $G_2$ gauge theory with one flavor is
\begin{equation}
Z_{G_2, N_f=1, \text{inst}} = \frac{Z_{G_2, N_f=1}}{Z_{\text{pert}}}, \label{instpart}
\end{equation}
where $Z_{G_2, N_f=1}$ is \eqref{G2part} and $Z_{\text{pert}}$ is \eqref{pertpart}. The order $q^k$ of \eqref{instpart} gives the $k$-instanton contribution of the $G_2$ gauge theory with one flavor.
Since an explicit form of the instanton partition function of the $G_2$ gauge theory with one flavor is not known, we compare the expression \eqref{instpart} with the instanton partition function of the pure $G_2$ gauge theory after decoupling the flavor. After sending $M \to 0$, then the 5-brane web diagram will become the one for the pure $G_2$ gauge theory in Figure \ref{fig:SO7wspinorfloppede}. Hence, $Q_0Q_1, Q_2, Q_B$ are still finite in the limit and the instanton fugacity for the pure $G_2$ gauge theory is given by $q' = qM^{-2}$. Since the limit $M \to 0$ should be compatible with the expansion of $Q_0$ in \eqref{instpart} for the comparison, we will take the following steps. At each order of $q'$, we first rewrite \eqref{instpart} by $M, A_1$ and $Q_2$ and then expand it by $M$ and $A_1$. In this case, the term $Q_B^kQ_0^aQ_1^b$ becomes
\begin{equation}
Q_B^kQ_0^aQ_1^b = q'^k M^{-a+b}A_1^{2a-b+4k}Q_2^{a-b+3k} \label{order.comparison}
\end{equation}
The term which survives after the limit $M \to 0$ satisfies $a=b$.\footnote
It turns out that an explicit evaluation of \eqref{instpart} shows that there are no terms in $M$ of the negative powers until $a=6, k=2$.} In order to obtain an reliable expression until the order $q'^kA_1^{a+4k}$, we need to sum up the Young diagrams in \eqref{instpart} until the order $Q_0^aQ_B^k$. Then we can compare the result which is expanded by $A_1$ with the gauge theory computation for the pure $G_2$ gauge theory. We checked that the partition function \eqref{instpart} after applying the limit $M \to 0$ indeed agrees with the gauge theory result until $a=6$ for the one-instanton and the two-instanton parts of the Nekrasov partition function of the pure $G_2$ gauge theory obtained in \cite{Benvenuti:2010pq, Keller:2011ek, Hanany:2012dm, Keller:2012da, Cremonesi:2014xha}. This gives an evidence that we obtain the correct the partition function \eqref{instpart} for the $G_2$ gauge theory with one flavor.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we studied 5d $\mathcal{N}=1$ $G_2$ gauge theories from 5-brane web diagrams with an O5-plane in type IIB string theory.
The result that we obtained is summarized as follows:
\begin{itemize}
\item Two equivalent types of 5-brane webs for 5d pure $G_2$ gauge theory are presented, in Figures \ref{fig:SO7wspinorfloppede}, \ref{fig:pureG2}, and in Figure \ref{fig:G2pure2}. Webs for 5d $G_2$ gauge theory with matter are also discussed in subsection \ref{sec:addingF}.
\item Based on the 5-brane webs that we obtained, we computed the partition functions of the BPS spectrum. In particular, the partition function for the pure $G_2$ theory is given in \eqref{part.pureG2}, and the $G_2$ theory with one flavor is given in \eqref{G2part}.
\end{itemize}
For the first type of pure $G_2$ diagram, corresponding to Figures \ref{fig:SO7wspinorfloppede} and \ref{fig:pureG2}, we started from a conventional 5-brane web diagram for the 5d $SO(7)$ gauge theory with a hypermultiplet in the spinor representation. Applying a generalized flop transition such as in Figure \ref{fig:flopO5tilde2}, we perform a Higgsing associated to the spinor matter to obtain a 5-brane web diagram for the 5d pure $G_2$ gauge theory.
The detail discussion regarding two diagrams is presented in section \ref{sec:G2fromO5tilde}.
For the second type of pure $G_2$ diagram, corresponding to Figure \ref{fig:G2pure2}, we started with a 5-brane configuration for 5d $SO(8)$ gauge theory with a hypermultiplet in the spinor representation and a hypermultiplet in the conjugate spinor representation. We made use of triality of $SO(8)$ gauge theory so that one can interpret this configuration as 5d $SO(8)$ gauge theory with a vector and a spinor rather than a spinor and a conjugate spinor. With this, we performed a successive Higgsing associated with the vector and the spinor
to obtain another 5-brane web for pure $G_2$ theory,
which is of an ${\text{O5}}$-plane only.
Our results are tested in two different ways:
\begin{itemize}
\item Based on the webs, we computed the tension of monopole strings and compared it with the tension of monopole strings obtained from derivatives of the prepotential of 5d $G_2$ gauge theories. They completely agree as shown in \eqref{Eq:G2monotension1} and \eqref{Eq:G2monotension2}.
\item For pure $G_2$ case, we explicitly compared both the perturbative and instanton parts of the obtained pure $G_2$ partition function with the literature. For one flavor case, however, as an explicit expression for the partition function for $G_2$ theory with one flavor is not known, we instead checked an important consistency. Namely, the flavor decoupling limit of the instanton part for the $G_2$ gauge theory with one flavor reproduces that for the pure $G_2$ gauge theory.
\end{itemize}
Regarding the 5-brane web for pure $G_2$ gauge theory, it is worth noting that the 5-brane diagram of the second type
has an interesting feature. The diagram is almost identical to the diagram for the pure $SU(4)$ gauge theory of Chern-Simons level zero with a restriction of the Coulomb branch moduli. From the point of view of the topological string computation, the only difference between this $G_2$ gauge theory and the pure $SU(4)$ gauge theory is whether a Young diagram is transposed or not. This means that the (unrefined) partition function computations only differ by whether the power of the framing factor is $+1$ or $-1$, and also by whether an associated sign factor is included or not.
It would be interesting to see how this feature is modified when one fully refines the $G_2$ partition functions \cite{HKLY}.
As the topological vertex formalism is now applicable to a large class of 5-brane webs with an O5-plane or an $\widetilde{\text{O5}}$-plane, it would be interesting to see whether one can use this method to compute partition functions of 5d $SO(N)$ gauge theories with spinors \cite{HKLY}. Another direction to pursue is to obtain 5d Seiberg-Witten curves dictating M5-brane configurations based on their dual diagrams \cite{HKLY}.
|
\section*{Introduction and statement of results}
The classical Whitney extension theorem \cite{whitney} states that for a closed subset of euclidean space, every Whitney jet (i.e.\ a family of continuous functions which yields a (formal) Taylor expansion) can be extended to a smooth mapping on the whole euclidean space. This leads immediately to the question, when the extension can be realised as a continuous linear operator between the Fr\'{e}chet spaces of Whitney jets and smooth functions. In the one-dimensional case the problem is solved by Whitney \cite{MR1501749} (see \cite[Section 2.4]{MR2882877} for an exposition), but it is known that not every closed set in the multidimensional case admits a linear extension operator (cf.\ \cite[p.\ 123]{MR2300454}).
This situation stimulated research to identify conditions on the closed set which would guarantee the existence of a linear extension operator. We refer the reader to \cite{MR1501749} and the monograph \cite{MR2882877,MR2868143} as well as the references given there for more information on these results and related questions for other classes of functions.
In particular, for Banach spaces of $C^m$ functions the results of Fefferman \cite{MR2373373} give a much stronger result, since the set to which functions are restricted can be \emph{any} subset of Euclidean space.
In \cite[Remark 22.18]{MR1471480}, Kriegl and Michor note the problem of constructing, in the so-called convenient calculus\footnote{Convenient calculus is one generalisation of calculus for infinite-dimensional spaces beyond the realm of Banach spaces.}, a continuous extension operator for even smooth one-variable functions on general subsets.
Kriegl does prove, in \cite[Theorem 1.8]{K97}, that subsets of a convenient vector space with \emph{smooth} boundary admit a continuous extension operator for functions valued in a convenient vector space.
\smallskip
In the present article we wish to establish a suitable version of these results for manifolds. However, since spaces of Whitney jets are already quite technical in the vector space setting, we would rather like to replace them by a space of smooth functions. It is known that this can be done under suitable assumptions on the closed set. Namely, for H\"older domains with dense interior one gets an extension operator for vector spaces of smooth functions instead of Whitney jet (see e.g.\ \cite[Theorem~2.1]{Frerick_07b}). Replacing vector spaces of smooth mappings by manifolds of mappings, our aim is to prove that the restriction of smooth functions to a "suitably nice" subset is a submersion of the infinite dimensional manifolds (i.e.\ it locally is a projection onto a splitting subspace \cite{1502.05795v4}).
Since the Whitney extension theorem generalises to paracompact manifolds by the usual partition of unity argument, the problem addressed here is twofold:
Firstly, one needs to formulate conditions on the closed set (in the manifold) such that a linear extension operator exists.
Secondly, we would like to establish differentiability and the submersion property on manifolds of mappings, which are in general not even Fr\'{e}chet spaces.
To this end we base our investigation on the so called Bastiani calculus \cite{bastiani} to make sense of (differentiable) manifolds in this setting (a short overview is included in Appendix \ref{app: calculus} for the reader's convenience).
Results of the type established in the present paper have been used in \cite{1610.05904} to prove that mapping groupoids between Lie groupoids are (infinite-dimensional) Lie groupoids.
We describe our results now in more detail. Fix a pair of finite-dimensional manifolds $M,N$ with $M$ being a $\sigma$-compact and equipped with a Riemannian metric.
Let $C \subset M$ be a closed set satisfying a cusp condition, defined below in Definition \ref{def:cusp_condition}.
This condition allows general Lipschitz domains, but also much rougher boundary conditions, for instance Koch snowflake-like sets.
Since the interior $C^\circ$ is dense in $C$, it makes sense to define differentiable mappings as those which are differentiable on the interior and whose differentials extend continuously (this is the setup considered in \cite{Frerick_07b}).
Note that at this stage we do not assume that $C$ carries any submanifold structure of its own, whence smoothness is only a meaningful concept because we can test in charts of the manifolds $M$ and $N$ which do not have a boundary.
In this sense, one can talk about smooth (compactly supported) sections on $C$ with values in vector bundles over $M$. We then discuss a natural locally convex structure on the space $\Gamma_c (C,E)$ of all such sections such that the Whitney extension theorem yields our first result.
\paragraph{Theorem A} \emph{
Let $E\to M$ be a (finite-rank) vector bundle and $C\subset M$
a closed set satisfying the cusp condition.
The restriction map $\res_C\colon \Gamma_c(M,E) \to \Gamma_c(C,E), \sigma \mapsto \sigma|_C$ on compactly-supported smooth sections has a continuous linear splitting.
}\smallskip
We wish to give a global version of Theorem A in terms of manifolds of mappings, i.e.\ for the map $\res^M_C \colon C^\infty (M,N) \rightarrow C^\infty (C,N)$.
Recall that for a smooth manifold $M$ (possibly with corners), the space $C^\infty_{\text{fS}} (M,N)$ of smooth mappings with the fine very strong topology (see \cite{HS17} and \cite{michor}, where the fS-topology is called $\mathcal{FD}$-topology) can be turned into an infinite dimensional manifold. If $M$ is compact the fine very strong topology coincides with the well known compact open $C^\infty$-topology.
Hence if the subset $C$ is a manifold, smoothness of $\res_C^M$ is a well defined concept. We prove in Section \ref{sect: MFDMAP} that Theorem A yields local sections of $\res_C^M$ if $C$ is a submanifold with corners of $M$.
If $C$ is compact, we can even relax the condition and allow \emph{submanifolds with rough boundary}, a definition introduced by Karl-Hermann Neeb \cite{GloecknerNeebBuch}.
We establish in Proposition~\ref{prop: rsub:nnf} that closed submanifolds with rough boundary satisfy the necessary cusp condition, so as to make Theorem~A applicable.
Note however, that for this generalised boundary one has first to establish the manifold structure of the manifold of mappings.
We give the construction in Section \ref{sect: mfdmap:rb} and remark that it might be of independent interest as it is to the best of our knowledge the only source currently available for the construction.
Thus our next main result can be formulated as follows:
\paragraph{Theorem B}\emph{
For $C\subset M$ a submanifold with corners, or compact and a submanifold with rough boundary, then the restriction map $\res_C^M \colon C^\infty_{\text{fS}} (M,N) \rightarrow C^\infty_{\text{fS}} (C,N)$ is a submersion of locally convex manifolds.
}\smallskip
Recall that for infinite-dimensional manifolds whose model spaces are more general than Banach spaces, a submersion is a map that locally, in submersion charts, looks like a projection out of a product.
This is a stronger condition that the map on tangent spaces being a split surjection (cf.\ \cite{1502.05795v4} for a detailed study). Looking at charts, the map in Theorem B splits locally on the spaces of compactly-supported sections of certain vector bundles by Theorem A.
We remark here that Theorem B does not imply that $\res_C^M$ is surjective as not necessarily all smooth functions on closed submanifolds with (rough) boundary will admit extensions to the ambient manifold (compare \cite[Corollary 6.27]{MR2954043}).
A simple example is the case where $M=S^2$, $C\subset S^2$ is a closed equatorial `belt' and $N=S^1$.
A map $C \to S^1$ cannot extend to $S^2$ if has non-zero winding number.
Finally, we look at nested closed subsets which satisfy the assumptions of Theorem B. The following corollary allows a more general result.
\paragraph{Corollary C} \emph{
With the manifolds $M,N$ as above and closed sets $C\subset D\subset M$ which both satisfy the assumptions of Theorem B, the restriction map
$$\res_C^{D}\colon C^\infty_{\text{fS}} (D,N) \to C^\infty_{\text{fS}} (C,N),\quad f \mapsto f|_{C},$$
is a submersion of locally convex manifolds.
}\smallskip
The proof follows again by looking at charts (of the manifold of mappings), and by composing the extension operator $\Gamma_c(C,E) \to \Gamma_c(M,E)$ with the restriction $\res^M_{C'}$.
It is elementary that this composite is locally a splitting of the restriction map $\res_C^{C'}$.
A more specific corollary applies the above collection of results to
closed sets that are geodesically strongly convex\footnote{Recall that a subset $S$ is strongly convex if for every pair $p,q \in S$ there is a unique shortest geodesic segement $\overline{pq}$ connecting $p$ and $q$ such that $\overline{pq} \subseteq S$.}, for example closures $\overline{U_{i\ldots j}}$ of iterated finite intersections $U_{i\ldots j} = U_i \cap \ldots \cap U_j$ of geodesically strongly convex charts $U_i$.
Such closed sets satisfy the required cusp condition and we prove in Lemma \ref{lem: sconv:rbsmfd} that they are submanifolds with rough boundary.
Thus Corollary C immediately yields.
\paragraph{Corollary D}\emph{
Let $M$ be a smooth Riemannian manifold with geodesically
strongly convex compact sets $C \subset D \subset M$ and $N$ another smooth manifold.
Then the restriction map}
\[
\res_C^{D}\colon C^\infty (D,N)
\to
C^\infty(C,N)
\]
\emph{ is a submersion of Fr\'{e}chet manifolds.}\smallskip
A close analogue of this corollary was stated as \cite[Proposition 3]{1610.05904}, with only a rough sketch of a proof, ignoring the function space topologies, and also allowing $M$ to be a manifold with corners.
This hypothesis was only included so as to recover Corollary~D, where a closed set $\overline{U_{i\ldots j}}$ played the r\^ole of the ambient ``manifold with corners'' $M$.
However, \emph{it is not necessarily the case} that these closed sets are manifolds with corners, as defined in \cite[\S~2.4]{michor} (see also Definition \ref{defn: RBM} below).
Thus Corollary~D should be taken to replace \cite[Proposition 3]{1610.05904}.
One can ask the obvious questions as to how much further the results here can be pushed, especially in light of the results of Frerick on general sets satisfying the cusp condition.
The biggest obstacle in pursuing this, is to define the relevant locally convex topologies or manifold structures in the linear and non-linear cases respectively. In light of this, an extension of the results in the present paper might be possible but there seems to be no straightforward way to do this.
\smallskip
A brief outline of the paper is as follows.
In Section~\ref{sect:prelim} we give basic notions that are needed for the paper, relegating most technical results for infinite-dimensional calculus and manifolds to Appendix~\ref{app: calculus}.
Section~\ref{sect: WElin} gives the necessary ingredients to build up to Theorem~A, namely various bits of extension theory and patching results in the linear setting, and these are assembled in Section~\ref{sect: proof of thm A}.
We then recall (from the forthcoming \cite[Chapter 1.4]{GloecknerNeebBuch}) the fundamentals of the theory of manifolds with rough boundary in Section~\ref{sect: mfdmap:rb} and construct the smooth manifolds of maps in that case.
In Section~\ref{sect: MFDMAP} we then finally prove Theorem~B.
Appendix~\ref{app: Whitney} is a summary of the theory of Whitney jets, for ease of reference.
\section{Preliminaries and Notation}\label{sect:prelim}
We wish to study an extension operator between spaces of smooth functions on manifolds. In the end, we will see that, as for the vector space case, an extension operator for functions defined on a ``suitably nice'' subset $C$ of a manifold $M$ to smooth functions on the whole manifold exists. Further, we want to establish that the \emph{restriction of $N$-valued functions} is a submersion in the sense of \cite{1502.05795v4}.
\begin{setup}[Notation]
We write $\ensuremath{\mathbb{N}} \colonequals \{1,2,\ldots\}$ and $\ensuremath{\mathbb{N}}_0 \colonequals \ensuremath{\mathbb{N}} \cup \{0\}$.
Frequently we will use standard multiindex notation to denote (iterated) partial derivatives of a (smooth) function $f \colon \ensuremath{\mathbb{R}}^d \supseteq U \rightarrow \ensuremath{\mathbb{R}}^m$ as $\partial^\alpha f$ for $\alpha \in \ensuremath{\mathbb{N}}_0^d$ (see \ref{setup: multiindex}).
\end{setup}
\begin{setup}[Conventions]
For a subset $S$ of a topological space we denote by $S^\circ$ its interior.
We say that a subset $C$ of a topological space is \emph{regular}, if $C^\circ$ is dense in $C$.
We note that closed subsets satisfying the cusp condition to be defined below are always regular.
Further, every finite-dimensional manifold considered in the following will always be assumed to be Hausdorff and $\sigma$-compact.
We say $M$ is a \emph{Banach} (or \emph{Fr\'echet}) manifold if all its modelling spaces are Banach (or Fr\'echet) spaces. In general, infinite-dimensional manifolds will not required to be $\sigma$-compact or paracompact.
\end{setup}
We consider functions on non-open sets following \cite{Wockel06} (where these mappings are used to define manifolds with boundary). Further we frequently have need for smooth functions on possibly infinite-dimensional manifolds (think manifold of mappings).
To this end we base our investigation on the so called Bastiani calculus \cite{bastiani} which readily generalises beyond the realm of Banach spaces (cf.\ Appendix \ref{app: calculus} for a short introduction).
\begin{definition}
Let $E,F $ be locally convex spaces and $C \subseteq E$ be a set with dense interior. A continuous mapping $f \colon C \rightarrow F$ is called $C^1$-map if
$f\big|_{C^\circ}$ is $C^1$ in the sense of Bastiani calculus and the derivative $d(f\big|_{C^\circ})$ extends (necessarily uniquely) to a continuous mapping $df \colon C \times E \rightarrow F$.
Similarly we say $f$ is $C^k$ for $k \in \ensuremath{\mathbb{N}} \cup \{ \infty \}$ if $f\big|_{C^\circ}$ is $C^k$ and the iterated differentials extend (uniquely) to all of $C$.
We say $f$ is \emph{smooth (or $C^\infty$)} if $f$ is $C^k$ for every $k \in \ensuremath{\mathbb{N}}$ and write $C^\infty (C,F)$ for the set of all smooth maps on $C$.
\end{definition}
We have the following version of the chain rule:
\begin{lemma}[{\cite[Remark 5]{Wockel06} and \cite[Lemma 3.17]{alas2012}}]\label{lem: reschain}
Let $C \subseteq E$ and $D \subseteq F$ be regular subsets of locally convex spaces $E,F$ and $H$ be another locally convex space.
Consider $C^k$ mappings $f \colon C \rightarrow D \subseteq F$ and $g \colon D \rightarrow H$ then $g \circ f$ is a $C^k$ mapping if one of the following conditions is satisfied
\begin{enumerate}
\item $f(C^\circ) \subseteq D^\circ$ (no condition on $C$ and $D$),
\item $C,D$ are locally convex sets, i.e.\ every point has a neighborhood in the set which is convex (no condition on $f$ and $g$).
\end{enumerate}
\end{lemma}
Note that for an open set $C \cap U$ is a regular set if $C$ is regular.
Thus the chain rule allows us to make sense of $C^k$-mappings on regular subsets of smooth manifolds \textbf{without boundary}.
\begin{definition}
Let $C \subseteq M$ be a regular subset of a manifold without boundary. A continuous map $f \colon C \rightarrow N$ into a manifold $N$ without boundary is a $C^k$-mapping if for every $x\in C$ there is a pair of charts $(\varphi, U), (\psi,V)$ such that $\psi \circ f \circ \varphi^{-1}|_{\varphi (U \cap C)}$ makes sense and is a $C^k$-mapping.
\end{definition}
Clearly by Lemma \ref{lem: reschain} condition 1.\ this definition is independent of the choice of charts.
However, we note that many of the familiar rules of calculus are no longer valid for $C^k$-mappings on sets with dense interior which are not locally convex.
In any case, these results are not needed to treat spaces of sections as locally convex spaces in Section \ref{sect: WElin} below and to prove Theorem A.
To retain the ``usual behaviour'' of differentiable functions (most importantly, the Mean Value Theorem, and hence the chain rule) one needs to require in addition that the set is locally convex.
This will be important to establish the global setting required in Theorem B.
Namely, the usual rules of calculus enable the construction of manifolds of mappings as outlined in Section \ref{sect: mfdmap:rb}.
\section{Whitney's extension theorem for linear spaces of functions}\label{sect: WElin}
The aim of this section is to recall the Whitney extension theorem in the vector space case.
Further, we discuss conditions under which the space of Whitney jets can be identified with spaces of smooth functions on a closed regular set.
In this section we let $C \subseteq \ensuremath{\mathbb{R}}^d, d \in \ensuremath{\mathbb{N}}$ be a regular closed set.
\begin{setup}[Ideals of functions vanishing on closed sets]
Let $m \in \ensuremath{\mathbb{N}}$ and $W \subseteq \ensuremath{\mathbb{R}}^d$ be an open neighborhood of the regular closed set $C$. We consider
\[
\mathcal{I}_C (W,\ensuremath{\mathbb{R}}^m) \colonequals \{g \in C^\infty (W,\ensuremath{\mathbb{R}}^m) \mid \partial^\alpha g|_C \equiv 0 \quad \forall \alpha \in \ensuremath{\mathbb{N}}_0^d \}.
\]
Since the operators $\partial^\alpha \colon C^\infty_{\text{co}} (U,\ensuremath{\mathbb{R}}^m) \rightarrow C_{\text{co}} (U,\ensuremath{\mathbb{R}}^m), f \mapsto \partial^\alpha f$ and $\ev_x \colon C_{\text{co}} (W,\ensuremath{\mathbb{R}}^m) \rightarrow \ensuremath{\mathbb{R}}^m, f \mapsto f(x)$ are continuous linear (cf.\ \cite[Definition 2.5 and Proposition 3.20]{alas2012} with respect to the compact open $C^\infty$-topology (cf.\ Appendix \ref{app: calculus}),
\[
\mathcal{I}_C (W,\ensuremath{\mathbb{R}}^m) = \bigcap_{\alpha \in \ensuremath{\mathbb{N}}_0^d} \bigcap_{x \in C} (\ev_x \circ \partial^\alpha)^{-1} (0)
\]
is a closed vector subspace of the Fr\'{e}chet space $C^\infty_{co} (U,\ensuremath{\mathbb{R}}^m)$.
Indeed, if we denote by $\mathcal{E} (C,\ensuremath{\mathbb{R}}^m)$ the $\ensuremath{\mathbb{R}}^m$-valued Whitney jets on $C$ (see Appendix \ref{app: Whitney}), we can view $\mathcal{I}_C (W,\ensuremath{\mathbb{R}}^m)$ as the kernel of the linear restriction map $r_W \colon C^\infty (W,\ensuremath{\mathbb{R}}^m) \rightarrow \mathcal{E}(C,\ensuremath{\mathbb{R}}^m), g \mapsto (\partial^\alpha g)_\alpha$.
Recall from \cite[p.\ 126]{MR2300454} that $r_W$ is continuous if $m=1$.\footnote{Indeed the article claims this only for $W=\ensuremath{\mathbb{R}}^d$ but continuity follows directly from the remarks above Definition 2.1 in loc.cit. as explained in \ref{Whitney:fun}.}
Identifying $C^\infty_{co} (W,\ensuremath{\mathbb{R}}^m) \cong C^\infty_{co} (W,\ensuremath{\mathbb{R}})^m$ (cf.\ \cite[Lemma 3.4]{MR1934608}) we obtain continuity of $r_W$ for arbitrary neighborhoods $W$ and $m \in \ensuremath{\mathbb{N}}$.
\end{setup}
\begin{setup}[{Whitney extension theorem \cite[Theorem 1]{whitney}, or \cite[Theorem 2.2]{MR2300454} for a modern introduction}]\label{setup: whitney:ext}
The following sequence of Fr\'{e}chet spaces is exact:
\begin{equation}\label{eq: seq0}
\begin{xy}
\xymatrix{
0 \ar[r] & \mathcal{I}_C (W, \ensuremath{\mathbb{R}}^m) \ar[r] & C^\infty_{\text{co}} (W, \ensuremath{\mathbb{R}}^m) \ar[r] & \mathcal{E} (C, \ensuremath{\mathbb{R}}^m)\ar[r] & 0
}
\end{xy}
\end{equation}
\end{setup}
\begin{remark}
Recall that in the category of locally convex spaces, a sequence
\begin{displaymath}
\begin{xy}
\xymatrix{
0 \ar[r] & A \ar[r]^i & B \ar[r]^q & C \ar[r] & 0
}
\end{xy}
\end{displaymath}
of continuous linear maps is \emph{exact} if it satisfies both of the following conditions
\begin{enumerate}
\item \emph{algebraically exact}, i.e.\ images of maps coincide with kernels of the next map,
\item \emph{topologically exact}, i.e.\ $i$ and $q$ are open mappings onto their images.
\end{enumerate}
If $A$, $B$ and $C$ are Fr\'{e}chet spaces topological exactness follows from algebraic exactness by virtue of the open mapping theorem; for general locally convex spaces this is not the case (cf.\ e.g.\ \cite{MR1977923}).
\end{remark}
Note that the Whitney extension theorem in general requires only a closed set $C$ and not (as we required) a closed and regular set.
However, in our approach we will replace the space of Whitney jets by a space of smooth functions on a closed set.
Here the regularity assumption comes into play (cf.\ Appendix \ref{app: calculus}) and we will now construct a mapping which deals with the identification:
\begin{setup}\label{Dmap}
Consider the mapping
$$D \colon C^\infty (C,\ensuremath{\mathbb{R}}^m) \rightarrow \prod_{\alpha \in \ensuremath{\mathbb{N}}_0^d} C_{\text{co}}(C,\ensuremath{\mathbb{R}}^m), \quad f \mapsto (\partial^\alpha f)_{\alpha}.$$
Then $D$ makes sense by our definition of $C^\infty (C,\ensuremath{\mathbb{R}}^m)$ and is injective and linear.
Arguing as in \cite[Section 2]{MR2300454} the image of $D$ is a closed subspace of the Fr\'{e}chet space $\prod_{\alpha \in \ensuremath{\mathbb{N}}_0^d} C_{\text{co}}(C,\ensuremath{\mathbb{R}}^m)$ (note that we have compact convergence of functions and all derivatives on the dense interior of $C$!).
\end{setup}
As the mapping $D$ takes a smooth function on $C$ to a jet expansion (i.e.\ its family of derivatives), one is tempted to think that $D$ takes its image in the space $\mathcal{E}(C,\ensuremath{\mathbb{R}}^m)$ of Whitney jets.
However, this is wrong in general as the following example from \cite[Example 2.18]{MR671464} shows:
\begin{example}
Let $C$ be the complement of the open subset $\{(x,y) \in \ensuremath{\mathbb{R}}^2 \mid 0 < y < \exp (-1/x^2), x > 0\}$.
Then $C$ is a closed regular set and we define a function $f \in C^\infty (C,\ensuremath{\mathbb{R}}^2)$ as follows
\[
f(x,y) = \begin{cases} \exp (-1/x^2) & \text{ if } x>0 ,\ y \geq \exp (-1/x^2) \\
0 & \text{ otherwise }
\end{cases}
\]
A computation of $\left(f\!\left(x,\exp(-1/x^2)\right)-f(x,0)\right)/(\exp(-1/x^2)-0) =1$ shows that $f$ cannot be extended to a smooth function on $\ensuremath{\mathbb{R}}^2$.
Thus in particular, the image of $f$ under the mapping $D$ from \ref{Dmap} is not a Whitney jet.
\end{example}
As a consequence $D$ can take its image in the space of Whitney jets only if every smooth function on $C$ extends to a smooth function on an open neighborhood of $C$.
It turns out that the non existence of extensions is tied to the exponential type cusps of the set $X$ in the example.
Prohibiting such inward cusps, which we shall call \emph{narrow fjords}, ensures that every smooth function can indeed be extended.
\begin{definition}[{\cite[2.16.1]{MR671464}}]\label{defn: incusp}
Let $A$ be a regular closed subset of $\ensuremath{\mathbb{R}}^d$. We say $A$ has \emph{no narrow fjords} if for all $a\in A$ exists an integer $p$, a compact neighborhood $K$ of $a$ in $A$ and a constant $C>0$ such that any $x,y\in K$ can be joined by a rectifiable path $\gamma$ lying inside $A^\circ$, except perhaps for finitely many points, and the length $\ell (\gamma)$ of $\gamma$ satisfies
\[
\lVert x-y\rVert \geq C\,\ell (\gamma)^p.
\]
\end{definition}
Note that this definition gives control over how fast the width of fjords can shrink as one moves inwards along them, see Figure~\ref{fig:fjords_condition}. Further, the no narrow fjords condition is closely related to the conditions called $C$-quasiconvexity and the $(C,\omega)$-convexity from \cite[Definition 2.63]{MR2882877}.
\begin{figure}
\begin{center}
\usetikzlibrary{decorations.markings}
\begin{tikzpicture}[decoration={
markings
mark=at position 0.08 with {\arrow{stealth}};,
mark=at position 0.3 with {\arrow{stealth}};,
mark=at position 0.8 with {\arrow{stealth}}
}]
\path [fill=lightgray]
(-1,0)
to (-1,5.5)
to (6,5.5)
to (5.5,4.5)
to (4,4.5)
to [out=-50] (4.5,3.5)
to [out=-165,in=0] (0.5,3)
to [out=0,in=170] (3.5,2.5)
to (4,2.75)
to [out=-80,in=135] (5,1.5)
to [out=-75] (6,0)
to (-1,0);
\draw (6,5.5) to (5.5,4.5)
to (4,4.5)
to [out=-50] (4.5,3.5)
to [out=-165,in=0] (0.5,3)
to [out=0,in=170] (3.5,2.5)
to (4,2.75)
to [out=-80,in=135] (5,1.5)
to [out=-75] (6,0);
\node [below] at (0,0.5) {$A$};
\draw [fill] (0.5,3) circle [radius=0.03];
\node [left] at (0.5,3) {$a$};
\draw (5.19,1) -- (-0.5,1) -- (-0.5,4.63) -- (5.56,4.63);
\node [above right] at (-0.5,4.15) {$K$};
\draw [fill] (5.5,4.5) circle [radius=0.03];
\node [below left] at (5.5,4.5) {$x$};
\draw [fill] (5,1.5) circle [radius=0.03];
\node [right] at (5,1.5) {$y$};
\draw [dotted] (5.5,4.5) to (5,1.5);
\draw[postaction={decorate}, red] (5.5,4.5) to (5,4.75)
to (4,4.5)
to (0.5,3)
to (5,1.5);
\node [left] at (3,4.2) {$\gamma$};
\draw [fill] (3.5,3.25) circle [radius=0.03];
\draw [fill] (3.5,2.5) circle [radius=0.03];
\draw [dotted] (3.5,3.25) to (3.5,2.5);
\draw [fill] (2.3,3.07) circle [radius=0.03];
\draw [fill] (2.3,2.76) circle [radius=0.03];
\draw [dotted] (2.3,3.07) to (2.3,2.76);
\end{tikzpicture}
\caption{No narrow fjords condition}\label{fig:fjords_condition}
\end{center}
\end{figure}
\begin{example}
Let $A$ be a regular closed set.
Recall that the open set $A^\circ$ satsifes the \emph{bounded turning condition} if there is a constant $C>0$ such that for all $x,y\in A$, there is a rectifiable path $\gamma$ from $x$ to $y$ such that $\lVert x - y\rVert \geq C\, \ell(\gamma)$.
If $A^\circ$ satisfies the bounded turning condition then it has no narrow fjords.
Any uniform domain \cite{Martio-Sarvas} (see, for example, \cite[Definition 2.2]{Mourgoglou} for an updated formulation) satisfies the bounded turning condition, which includes all H\"older domains and NTA domains ("non-tangentially accessible domains" as introduced by \cite{JK82}), and so the closures of all these sets all have no narrow fjords.
\end{example}
\begin{setup}\label{setup: WJ}
Let now $C$ be a regular closed set with no narrow fjords.
Then \cite[Proposition 2.16]{MR671464} asserts that $D \colon C^\infty (C,\ensuremath{\mathbb{R}}^m) \rightarrow \bigoplus_{\alpha \in \ensuremath{\mathbb{N}}_0^d} C_{\text{co}} (C,\ensuremath{\mathbb{R}}^m)$ takes its image in $\mathcal{E}(C,\ensuremath{\mathbb{R}}^m)$.
As a consequence of the Whitney extension theorem \ref{setup: whitney:ext}, every element in $C^{\infty} (C,\ensuremath{\mathbb{R}}^m)$ extends to a smooth map on $\ensuremath{\mathbb{R}}^d$, whence the image of $D$ coincides with the space of Whitney jets $\mathcal{E} (C, \ensuremath{\mathbb{R}}^m)$ (see Definition \ref{defn: WJm}).
Thus we topologize $C^\infty (C,\ensuremath{\mathbb{R}}^m)$ with the identification topology induced by $D$, turning it into a Fr\'{e}chet space isomorphic to the space of Whitney jets on $C$.
In particular, the exact sequence \eqref{eq: seq0} yields an exact sequence of Fr\'{e}chet spaces
\begin{equation}\label{eq: WSQ1}
\begin{xy}
\xymatrix{
0 \ar[r] & \mathcal{I}_C (W, \ensuremath{\mathbb{R}}^m) \ar[r] & C^\infty_{\text{co}} (W, \ensuremath{\mathbb{R}}^m) \ar[r] & C^\infty (C, \ensuremath{\mathbb{R}}^m)\ar[r] & 0 .
}
\end{xy}
\end{equation}
\end{setup}
In the next section we are going to investigate \emph{outward} cusp conditions on the boundary of closed subsets and show how they can be transferred to Riemannian manifolds.
\section{The cusp condition
In the last section we have already encountered a cusp condition preventing the occurrence of certain (inward) cusps on the boundary of the closed set on which we are working.
The key functional-analytic result we use to extend sections is due to Frerick in \cite{MR2300454}.
It uses a metric condition on a closed domain $F$ in $\ensuremath{\mathbb{R}}^n$ to ensure there is a continuous extension operator for Whitney jets on $F$ to smooth functions on $\ensuremath{\mathbb{R}}^n$.
The following definition abstracts the hypothesis from \cite[Theorem 3.16]{MR2300454} and from Definition \ref{defn: incusp} so as to apply to closed sets in a metric space more general than $\ensuremath{\mathbb{R}}^n$.
\begin{definition}\label{def:cusp_condition}
Let $(M,d)$ be a locally compact metric space $(M,d)$. A closed set $F\subset M$ has
\begin{enumerate}
\item \emph{no narrow fjords} if for each $x \in F$ there exists $p\in \ensuremath{\mathbb{N}}$, $K \subseteq F$ a compact neighborhood of $x$ and $D>0$ such that all $y,z \in K$ can be joined by a rectifiable curve $\gamma$ lying inside $F^\circ$, except perhaps for finitely many points, such that its length $\ell_d(\gamma)$ satisfies $d(y,z) \geq D\,\ell_d (\gamma)^p$.
\item \emph{at worst polynomial outward cusps} if for all compact $K\subset M$ there exist $\varepsilon_0,\rho >0$ and $r\geq 1$ such that for all $z\in K\cap \partial F$ and $0<\varepsilon<\varepsilon_0$ there is an $x\in F$ with $d(x,z)<\varepsilon$ such that if $d(x,y)<\rho\varepsilon^r$ then $y\in F$ and $d(z,y)<\varepsilon$.
\end{enumerate}
If $F$ has at worst polynomial outward cusps and no narrow fjords we simply say that $F$ \emph{satisfies the cusp condition}.
\end{definition}
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node at (10,6) {$M$};
\draw [fill=lightgray]
(1.5,5.5) to [out=-45,in=-90] (3,6.5)
to [out=-80,in=135] (4,5.5)
to [out=45,in=64] (6.5,3.5)
to [out=66,in=45] (3,2.6)
to [out=100,in=90] (1.5,2);
\node at (2,4) {$F$};
\node [left] at (1.6,4) {$\ldots$};
\draw (5.5,3) -- (6,6) -- (7,5.5) -- (7,2.5) -- (5.5,3);
\node [right] at (7,5) {$K$};
\draw [fill] (6.5,3.5) circle [radius=0.03];
\node [right] at (6.5,3.5) {$z$};
\draw [thick,dotted] (6.5,3.5) circle [radius=2.8];
\draw [dashed] (6.5,3.5) -- (6.5,0.7);
\node [right] at (6.5,1.5) {$\varepsilon_0$};
\draw [red, dotted, thick] (6.5,3.5) circle [radius=2.3];
\draw [red,dashed] (6.5,3.5) -- (4.2,3.5);
\node [below] at (4.9,3.5) {$\varepsilon$};
\draw [fill=red,red] (5.5,4.7) circle [radius=0.05];
\node [below] at (5.5,4.7) {$x$};
\draw [red, dotted,thick] (5.5,4.7) circle [radius=0.55];
\draw [red,dashed] (5.5,4.7) -- (4.95,4.7);
\draw [->,dashed] (5.3,6.3) -- (5.3,4.8);
\node [above] at (5.3,6.3) {$\rho\varepsilon^r$};
\end{tikzpicture}
\caption{Polynomial outward cusps}\label{fig:cusp_condition}
\end{center}
\end{figure}
In the case that $r=1$, the condition on outward polynomial cusps is sometimes called the (interior) corkscrew condition \cite[p.~123]{MR2305115}, and so our polynomial cusps can be seen as corkscrews with nonlinear growth.
\begin{remark}\label{wlog_constants_nice}
In Definition~\ref{def:cusp_condition}.2, if the constants $\varepsilon_0,\rho,r$ work for the compact set $K$, then so do smaller such constants, and if $\varepsilon_0 \leq 1$ then we can also increase $r$.
Putting this together, we can assume that $\varepsilon_0 = \rho < 1$ and increase $r$ as needed, and as a result can replace $\rho \varepsilon^r$ by $\varepsilon^{r+1}$.
Hence we can, without loss of generality, assume that $\rho=1$ and $r \geq 2$.
\end{remark}
\begin{example}\label{ex: cusp}
Every Lipschitz domain satisfies the cusp condition, as do H\"older domains and NTA domains.
The compact subset of $\ensuremath{\mathbb{R}}^2$ whose boundary is the Koch curve satisfies the cusp condition.
\end{example}
Recall that $\mathcal{E}(F)$ denotes the space of Whitney jets on the closed set $F$.
In the following Theorem, $\ensuremath{\mathbb{R}}^n$ is taken with the Euclidean metric.
\begin{theorem}[{\cite[Theorem 3.16]{MR2300454}}]\label{thm:Frerick}
Let $F\subset \ensuremath{\mathbb{R}}^n$ be closed and have at worst polynomial outward cusps.
Then the surjective map $C^\infty(\ensuremath{\mathbb{R}}^n) \to \mathcal{E}(F)$ of Fr\'echet spaces has a continuous splitting.
\end{theorem}
Moreover, Theorem \ref{thm:Frerick} combined with \ref{setup: WJ} yields the following Corollary which generalises \cite[Theorem~2.1]{Frerick_07b}.
\begin{corollary}\label{cor:cusps!}
Let $F\subset \ensuremath{\mathbb{R}}^n$ be closed and satisfy the cusp condition, then the surjective map $C^\infty(\ensuremath{\mathbb{R}}^n) \to C^\infty (F,\ensuremath{\mathbb{R}})$ of Fr\'echet spaces has a continuous splitting.
\end{corollary}
We want to be able to sensibly transfer both Frerick's Theorem and Corollary \ref{cor:cusps!} in Euclidean space to a Riemannian manifold, so we will need a result that allows change of metric.
The following result is stated in more generality than we need, since it should be of independent interest.
\begin{lemma}\label{lem:transfer_cusp_condition}
Let $(M,d_1)$ be a locally compact, complete metric space, $F\subset M$ be closed and let $F$ have at worst polynomial outward cusps using the metric $d_1$.
If $d_2$ is another metric on $M$ that is locally bi-H\"older equivalent to $d_1$, then $F$ has at worst polynomial outward cusps using the metric $d_2$.
\end{lemma}
\begin{proof}
Let $K\subset M$ be any compact set and $\varepsilon_{0,1}$, $\rho_1$ and $r_1$ be the constants guaranteed to exist for $K$ by virtue of $F$ satisfying Definition~\ref{def:cusp_condition}.2 for $d_1$.
By Remark~\ref{wlog_constants_nice} we will assume $\rho_1 = 1$, $r_1\geq 2$ and $\varepsilon_{0,1} < 1$.
Define the compact set
\[
N \colonequals \overline{\{ x\in M \mid
d_1(F\cap K,x) < 2
\text{ and } d_2(F\cap K,x) < 2\}}.
\]
Now as $d_1$ and $d_2$ are locally bi-H\"older equivalent there are constants $C \geq 1$ and $0<\alpha\leq 1$ such that
\[
\frac{1}{C}\, d_1(a,b)^{\frac1\alpha} \leq d_2(a,b) \leq C\, d_1(a,b)^\alpha
\]
for all $a,b\in N$.
Take
\begin{align*}
\varepsilon_{0,2} &\colonequals \min\{C\,\varepsilon_{0,1}^\alpha,{\textstyle \frac12} \},\\
\rho &\colonequals \frac{1}{C^{1+r_1/\alpha^2}},\text{ and}\\
r_2 &\geq r_1/\alpha^2\quad \text{such that } \rho\,\varepsilon_{0,2}^{r_2} \leq \varepsilon_{0,1}
\end{align*}
to be the putative uniform constants required so that $F$ satisfies Definition~\ref{def:cusp_condition}.2 for the metric $d_2$.
Note that since $\varepsilon_{0,2} < 1$ it does makes sense to enlarge $r_2$ until the upper bound on $\rho\,\varepsilon_{0,2}^{r_2}$ is satisfied.
Let $z\in \partial F\cap K$ be arbitrary, and take any $\varepsilon_2$ such that $0<\varepsilon_2 < \varepsilon_{0,2}$.
Define $\varepsilon_1 = (\varepsilon_2/C)^{\frac1\alpha}$.
Since $\varepsilon_1 = (\varepsilon_2/C)^{\frac1\alpha} < (\varepsilon_{0,2}/C)^{\frac1\alpha} = \varepsilon_{0,1}$ then there is an $x\in F$ satisfying $d_1(x,z) < \varepsilon_1$ such that
\[
d_1(x,y) < \varepsilon_1^r \quad \Rightarrow \quad d_1(z,y) < \varepsilon_1\,\text{ and }\, y\in F.
\]
Note that as $z\in K$ and $d_1(x,z) < \varepsilon_1 < \varepsilon_{0,1} < 2$, we have $x\in N$.
Hence $d_2(x,z) \leq C\, d_1(x,z)^\alpha < \, \varepsilon_1^\alpha = \varepsilon_2$, as required.
Now take $y\in M$ such that $d_2(x,y) < \rho\, \varepsilon_2^{r_2}$.
Then $d_2(y,z) \leq d_2(y,x) + d_2(x,z) < \rho\,\varepsilon_2^{r_2} + \varepsilon_2 < \varepsilon_{0,1} + 1 < 2$, and so $y\in N$. So we can calculate that
\begin{align*}
d_1(x,y) & \leq \big( C\, d_2(x,y)\big)^\alpha\\
& < \left(C\, \rho_2\right)^\alpha \varepsilon_2^{r_2 \alpha} \\
& = \left(\frac{\varepsilon_2^{\alpha^2r_2/r_1}}{C} \right)^{\frac{r_1}{\alpha}}\\
& \leq \left(\frac{\varepsilon_2}{C}\right)^{\frac{r_1}{\alpha}} = \varepsilon_1^{r_1}
\end{align*}
where we have used that $\alpha^2r_2 \geq r_1$ and $\varepsilon_2<1$.
Using the cusp condition for $K$ in $d_1$,
\begin{align*}
d_1(z,y) & < \varepsilon_1 \quad (\text{and } y\in F) \\
\Rightarrow
d_2(z,y) & \leq C\, d_1(z,y)^\alpha \\
& < C\,\varepsilon_1^\alpha = \varepsilon_2.
\end{align*}
Hence $F$ has at worst polynomial cusps for $d_2$.
\end{proof}
Note that if we have \emph{uniformly} bi-H\"older equivalent metrics then we can dispense with the assumption of completeness; the proof goes through the same without the need to define the compact set $N$.
We also have the following simple result for transferring the other half of the cusp condition.
\begin{lemma}\label{lem: trans:nnf}
Let $(M,d_1)$ be a locally compact, complete metric space, $F\subset M$ be closed and let $F$ have no narrow fjords using the metric $d_1$. Then if $d_2$ is another metric on $M$ that is locally bi-Lipschitz to $d_1$, then $F$ has no narrow fjords using the metric $d_2$.
\end{lemma}
This follows once recalling that rectifiable paths can be taken to be Lipschitz functions $I\to M$.
\begin{corollary}
Take a manifold $M$ with a continous Riemannian metric $g$, and a locally bi-Lipschitz chart, $\phi \colon U \xrightarrow{\sim} \ensuremath{\mathbb{R}}^n$ on $M$.
Here $\ensuremath{\mathbb{R}}^n$ is given the Euclidean metric, and $U$ the restriction of the geodesic metric $d^g$ on $(M,g)$.
Then if $C \subset M$ is closed and satisfies the cusp condition for the metric $d^g$, then $F = \phi(C\cap U)\subset \ensuremath{\mathbb{R}}^n$ satisfies the cusp condition in the Euclidean metric.
\end{corollary}
\begin{remark}\label{rem: standard:arguments}
Note that by standard arguments\footnote{see eg. the answer by Beno\^it Kloeckner at \url{https://mathoverflow.net/a/236851/}}, every $C^1$ manifold with a continuous Riemannian metric $g$ has an atlas of charts that are locally bi-Lipschitz to Euclidean space, hence \emph{a fortiori} locally bi-H\"older.
\end{remark}
We can apply this (perhaps overly general) result to our setup, namely where we take a relatively compact smooth chart $U$ on the smooth manifold $M$
Thus we have a commutative diagram of Fr\'{e}chet spaces (cf.\ \ref{setup: WJ} and Appendix \ref{app: Whitney} for a description of the topologies)
\[
\xymatrix{
C^\infty_{\text{co}}(U) \ar[d] & \ar[l]_-{\simeq} C^\infty_{\text{co}}(\ensuremath{\mathbb{R}}^n) \ar[d] \\
C^\infty(C\cap U) & \ar[l]_-{\simeq} C^\infty(F) \ar@/_1pc/[u]_{\text{Theorem \ref{thm:Frerick}}}
}
\]
where the vertical arrows are surjective, and a continuous section of the restriction map $C^\infty(\ensuremath{\mathbb{R}}^n) \to C^\infty(F)$. Thus:
\begin{lemma}\label{lemma:splitting restriction in charts}
Let $C \subset M$ be a
closed set satisfying the cusp condition and $U \xrightarrow{\simeq} \ensuremath{\mathbb{R}}^n$ be a smooth chart on $M$.
Then the restriction map $C^\infty_{\text{co}}(U) \to C^\infty(C\cap U)$ of Fr\'{e}chet spaces has a continuous section.
\end{lemma}
\section{Proof of Theorem A}\label{sect: proof of thm A}
In this section we provide the necessary details for the proof of Theorem A from the introduction.
As a first step, we consider spaces of sections on certain closed regular subsets of a Riemannian manifold. After these sections have been discussed, it will turn out that we only need to collect the bits and pieces from the previous sections to obtain the result.
Throughout this section we fix the following data:
\begin{setup}\label{setup: prep:ThmA}
In the following $M$ will be a $d$-dimensional $\sigma$-compact manifold with a fixed choice of Riemannian metric $g$, $E\rightarrow M$ a rank $m$-vector bundle and $C \subseteq M$ a
closed subset which satisfies the cusp condition with respect to the geodesic metric $d^g$.
Further, we choose and fix auxiliary data as outlined in \ref{setup: PData}. In particular, denote the locally finite atlas by $(U_i, \varphi_i)_{i \in I}$, $C_i = U_i \cap C$ and the relatively compact charts by $V_i \subseteq U_i$ which satisfy $C \subseteq \bigcup_{i\in I} V_i$ .
\end{setup}
The main idea of the proof of Theorem A is as follows: We take a section and use local triviality of the bundle to cut it into pieces which can be extended due to the cusp condition. Then we reassemble the pieces into a section by using a classical local to global approach with a partition of unity.
In the next subsections we provide the necessary tools: First we define the spaces of sections, then we prepare the local to global result.
\subsection*{Smooth bundle sections on a closed set without narrow fjords}
\addcontentsline{toc}{subsection}{Smooth bundle sections on a regular closed set without narrow fjords}
Our first task is to construct a suitable topology for the vector space of sections into $E$ on $C$.
\begin{definition}\label{defn: sectsp:cl}
For a closed regular set $C$ which has no narrow fjords we define \begin{align*}
\Gamma_c (C,E) \colonequals \{\sigma \in C^\infty (C,E) \mid \pi_E \circ \sigma = \id_C \text{ and } \supp \sigma \text{ is compact}\}
\end{align*}
the \emph{compactly supported smooth sections on $C$}.
Further, define
\begin{align*}
\mathcal{I}_{c} (C,E) \colonequals \{\sigma \in \Gamma_c (M,E) \mid T^k_x \sigma = 0\quad \forall x \in C, k \in \ensuremath{\mathbb{N}}\}
\end{align*}
the subspace of all compactly supported sections vanishing (with all their derivatives) on $C$.\footnote{Here we use the notation $T^k = T \circ T \cdots \circ T (k$ times) to denote the $k$-fold iterated tangent functor $T$. Note that unpacking the definition of the iterated tangent functors (see e.g.\ \cite[Lemma 1.14]{MR1911979} for a local version) the vanishing of all iterated tangent functors at a point is equivalent to the vanishing of all iterated partial derivatives in any chart containing the point.}
\end{definition}
\begin{remark}\label{rem: closedideal}
Clearly the pointwise operations turn $ \Gamma_c (C,E)$ and $\mathcal{I}_{c} (C,E)$ into vector spaces. One can argue as in the vector space case to see that $\mathcal{I}_{c} (C,E)$ is a closed subspace of $\Gamma_c (M,E)$ (with the fine very strong topology). Indeed using an atlas of $M$ we can use Lemma \ref{lem: top:sect} and Remark \ref{rem: loc:sect} to rewrite the problem in charts, where closedness follows from the argument in \ref{setup: whitney:ext}. (Avoiding localisation in charts, one can alternatively use Lemma \ref{lem: top:sect} together with \cite[Lemma 3.8]{MR1934608}.)
\end{remark}
\begin{setup}\label{setup: PData}
Let $\ensuremath{\mathcal{U}} = (U_i, \varphi_i)_{i \in I}$ be a locally finite atlas of relatively compact charts of $M$ such that $\varphi_i (U_i) = \ensuremath{\mathbb{R}}^{d}$ and there is a collection $\ensuremath{\mathcal{V}}$ of open sets $V_i \subseteq \overline{V}_i \subseteq U_i, \ i \in I$ with
\begin{itemize}
\item $C \subseteq \bigcup_{i \in I} V_i$
\item $(\chi_i)_{i\in I}$ is a smooth partition of unity with $\supp \chi_i \subseteq V_i$
\end{itemize}
We set $C_i \colonequals C \cap U_i$ for $i \in I$ and note that $\varphi_i (C_i) \subseteq \varphi_i (U_i) = \ensuremath{\mathbb{R}}^{d}$ is closed.
\end{setup}
The following proposition is (apart from the topological assertions and the fact that we are working with smooth functions and not jets) is a folklore fact which easily follows from the Whitney extension theorem \ref{setup: whitney:ext} in charts and a gluing argument. Since this argument will be the basis of our construction we give full details.
\begin{proposition}[Whitney extension theorem for sections on a manifold]\label{prop: Whit:mfd}
The linear restriction map $\res_C \colon \Gamma_c (M,E) \rightarrow \Gamma_c (C,E)$ is surjective and endows $\Gamma_c (C,E)$ with a quotient topology such that
\begin{equation}\label{eq: ex:vssect} \begin{aligned}
\begin{xy}\xymatrix{
0 \ar[r] & \mathcal{I}_c (C,E) \ar[r] \ar[d] & \Gamma_c (M,E) \ar[r]^{\res_C} \ar[d]^{\rho_\ensuremath{\mathcal{U}}} & \Gamma_c (C,E) \ar[r] \ar[d]^r & 0 \\
0 \ar[r] & \bigoplus_i \mathcal{I}_{C_i} (U_i, \ensuremath{\mathbb{R}}^m) \ar[r] & \bigoplus_{i} C^\infty_{\text{co}} (U_i, \ensuremath{\mathbb{R}}^m) \ar[r]^q & \bigoplus_{i} C^\infty (C_i, \ensuremath{\mathbb{R}}^m) \ar[r] & 0}
\end{xy}\end{aligned}
\end{equation}
is commutative with exact rows in the category of locally convex spaces.\\
Here $r\colon \Gamma_c (C,E) \rightarrow \bigoplus_{i \in I} C^\infty (C_i,\ensuremath{\mathbb{R}}^m)$ sends $f \mapsto (\pr_2 \circ T\varphi_i \circ f|_{C_i})_{i\in I}$.
\end{proposition}
\begin{proof}
Let us first deal with the lower row: Since $\varphi_i$ is a diffeomorphism, we can use precomposition by $\varphi_i$ to identify $C^\infty_{\text{co}} (U_i,\ensuremath{\mathbb{R}}^m) \cong C^\infty_{\text{co}} (\varphi_i(U_i),\ensuremath{\mathbb{R}}^m)$ and $C^\infty (C_i,\ensuremath{\mathbb{R}}^m) \cong C^\infty (\varphi_i(C_i),\ensuremath{\mathbb{R}}^m)$. Now $F_i :=\varphi_i (C_i)$ is a closed subset of the ambient space and $\varphi_i (U_i)$ is an open neighborhood of $F_i$. Moreover, since $C$ has no narrow fjords, Lemma \ref{lem: trans:nnf} implies that $F_i$ has no narrow fjords, whence \ref{setup: WJ} yields for every $i\in I$ an exact sequence
$$\begin{xy}\xymatrix{
0 \ar[r] & \mathcal{I}_{C_i} (U_i, \ensuremath{\mathbb{R}}^m) \ar[r] & C^\infty_{\text{co}} (U_i, \ensuremath{\mathbb{R}}^m) \ar[r]^q & C^\infty (C_i, \ensuremath{\mathbb{R}}^m) \ar[r] & 0}
\end{xy},$$
where we set $\mathcal{I}_{C_i} (U_i,\ensuremath{\mathbb{R}}^m) \cong \mathcal{I}_{F_i} (\varphi_i (U_i),\ensuremath{\mathbb{R}}^m)$, $C^\infty (C_i,\ensuremath{\mathbb{R}}^m) \cong C^\infty (F_i,\ensuremath{\mathbb{R}}^m)$ and suppress the identifications in the notation. Using that taking countable direct sums in the category of locally convex spaces is exact, we see that the lower row of \eqref{eq: ex:vssect} is exact.
By Lemma \ref{lem: top:sect} we have canonical embeddings $\rho_\ensuremath{\mathcal{U}}$ of $\Gamma_c (M,E)$ into $\bigoplus_i \Gamma(E|_{U_i})$ and $\rho_\ensuremath{\mathcal{V}}$ of $\Gamma_c (M,E)$ into $\bigoplus_i \Gamma (E|_{V_i})$. We identify $\Gamma (E|_{U_i}) \cong C^\infty_{\text{co}} (U_i ,\ensuremath{\mathbb{R}}^m)$ as in Remark \ref{rem: loc:sect} and suppress this in the notation.
Since $\mathcal{I}_c (C, E)$ is clearly contained in the kernel of $\res_C$, we obtain a commutative diagram of vector spaces:
\begin{align}\label{diag: res} \begin{aligned}
\begin{xy}
\xymatrix{
0 \ar[r] & \mathcal{I}_{c} (C, E) \ar@{-->}[d]^{\rho_\ensuremath{\mathcal{U}}|_C} \ar[r] & \Gamma_c (M,E) \ar[r]^{\res_C} \ar[d]^{\rho_\ensuremath{\mathcal{U}}} &\Gamma_c (C,E) \ar[d]^{r= (r_{C_i})_{i\in I}} & \\
0 \ar[r] & \bigoplus_i \mathcal{I}_{C_i} (U_i, \ensuremath{\mathbb{R}}^m) \ar[r] & \bigoplus_{i} C^\infty_{\text{co}} (U_i, \ensuremath{\mathbb{R}}^m) \ar[r]^q & \bigoplus_{i} C^\infty (C_i, \ensuremath{\mathbb{R}}^m) \ar[r] & 0
}
\end{xy}\end{aligned}
\end{align}
Here $r_{C_i} (f) \colonequals f|_{C_i},\ i\in I$ and $\rho_\ensuremath{\mathcal{U}}|_C$ is induced from $\rho_\ensuremath{\mathcal{U}}$ and realises $\mathcal{I}_{c} (C,E)$ as the closed subspace $\{(c_i)_{i\in I} \in \bigoplus_i \mathcal{I}_{C_i} (U_i, \ensuremath{\mathbb{R}}^m) \mid c_i|_{U_i \cap U_j} = \Phi_{ij}(\id_M,c_j)|_{U_i\cap U_j}\}$.\footnote{Recall that point evaluations and postcomposition with fixed smooth functions are continuous in the compact open $C^\infty$-topology (see e.g.\ \cite{alas2012}). An easy adaption of the argument in \cite[proof of Lemma 3.21 (b)]{MR3328452} shows that the subspace indeed is a closed subspace of the direct sum.} Note that apart from the space $\Gamma_c (C,E)$ which is not yet topologised, \eqref{diag: res} is a commutative diagram in the category of locally convex spaces
\paragraph{$\res_C$ is surjective.} Consider $f\in \Gamma_c (C,E)$ and choose $(g_i)_{i\in I} \in \bigoplus_{i\in I} C^\infty (U_i,\ensuremath{\mathbb{R}}^m)$ with $q((g_i)_i) = r (f)$. In general $(g_i)_{i\in I}$ will not be contained in the image of $\rho_\ensuremath{\mathcal{U}}$, but we see that
\begin{equation}\label{eq: cuts:equiv}
g_i|_{C_i\cap C_j} = f|_{C_i \cap C_j} = g_j|_{C_i \cap C_j} \quad \text{for every }i,j\in I.
\end{equation}
Using the partition of unity from \ref{setup: PData}, we construct smooth functions on $M$ by continuing $\chi_j \cdot g_j|_{M\setminus U_j} \equiv 0$. Hence
$$h_i \colonequals \sum_{j \in I} (\chi_j \cdot g_j)|_{V_i} \in C^\infty (V_i, \ensuremath{\mathbb{R}}^m),\quad i\in I,$$
By construction $(h_i)_{i\in I} \in \bigoplus_{i \in I} C^\infty (V_i,\ensuremath{\mathbb{R}}^m)$ and we have $h_i|_{V_i \cap V_j} = h_j|_{V_i\cap V_j}$ for every pair $(i,j)\in I^2$.
Thus $(h_i)_{i\in I}$ is contained in the image of $\rho_\mathcal{V}$ and we can choose $h \in \Gamma_c (M,E)$ with $\rho_V (h) = (h_i)_{i\in I}$.
Now \eqref{eq: cuts:equiv} implies that $h_i|_{V_i \cap C_j} =f|_{V_i \cap C_j}$. As the $V_i$ cover $C$ (see \ref{setup: PData}), we see that $\res_C (h)=f$.
Thus $\res_C$ is surjective and we can endow $\Gamma_c (C,E)$ with the quotient topology, thus turning it into a locally convex space.
\paragraph{$r$ is continuous with respect to the quotient topology}
Follows directly from the commutativity of \eqref{diag: res} and the definition of the quotient topology. Note that $r$ is linear, whence \eqref{eq: ex:vssect} indeed is a commutative diagram in the category of locally convex spaces.
\paragraph{The upper row of diagram \eqref{eq: ex:vssect} is exact.}
In Remark \ref{rem: closedideal} we have seen that $\mathcal{I}_c (C,E)$ is a closed subspace and we know that $\res_C$ is surjective, open and continuous. Hence we only need to prove that its kernel coincides with $\mathcal{I}_c (C,E)$.
Consider $g \in \ker(\res_C)$.
Since $\rho_\ensuremath{\mathcal{U}}$ is injective, the commutativity of \eqref{diag: res} implies that $\rho_\ensuremath{\mathcal{U}} (g)$ is contained in the kernel of $q$ and by exactness of the bottom row and the definition of $\rho_\ensuremath{\mathcal{U}}|_C$ we must have $g \in \mathcal{I}_c (C,E)$. The converse inclusion is trivial and in conclusion \eqref{eq: ex:vssect} is exact in the category of locally convex spaces. Finally we remark that this implies that $\Gamma_c (C,E)$ is a Hausdorff space (as the quotient of a Hausdorff space modulo a closed linear subspace).
\end{proof}
An important ingredient in the proof of the last lemma was the local to global argument using a partition of unity. We will see in Lemma \ref{lem: sm:mix} that this construction is continuous with respect to the function space topologies.
\begin{lemma}\label{lemma: r_is_injective}
The map $r \colon \Gamma_c (C,E) \rightarrow \bigoplus_{i} C^\infty (C_i,\ensuremath{\mathbb{R}}^m)$ is injective and its image is the closed subspace
$$\mathcal{A} \colonequals \left\{(h_i) \in \bigoplus_{i \in I} C^\infty (C_i,\ensuremath{\mathbb{R}}^m) \middle| h_i|_{C_i \cap C_j} = \Phi_{ij} (\id_{C_j} , h_j)|_{C_i\cap C_j} \forall i,j \in \ensuremath{\mathbb{N}}\right\}.$$
If $C$ is compact, $r$ induces an isomorphism of locally convex spaces $\Gamma_c (C,E) \cong \mathcal{A}$.
\end{lemma}
\begin{proof}
We already know that $r$ is continuous and it is clearly injective and takes its image in $\mathcal{A}$. Now every $C_i \subseteq C$ is
contained in the compact set $\overline{U}_i$. Hence for a family $(f_i) \in \mathcal{A}$, the obvious mapping
$$f \colon C \rightarrow E, \quad f(x) \colonequals T\varphi_i^{-1} f_i (x) , \text{ for } x \in C_i$$
makes sense and is a compactly supported smooth section over $C$, i.e.\ it is contained in $\Gamma_c(C,E)$. Hence $\mathcal{A}$ is the image of $r$. Again since point evaluation and postcomposition by fixed smooth mappings are continuous in the compact open $C^\infty$-topology, \eqref{eq: WSQ1}
shows that this is also the case for the Fr\'{e}chet topology on $C^\infty (C,\ensuremath{\mathbb{R}}^m)$. An easy adaptation of the argument in \cite[proof of Lemma 3.21 (b)]{MR3328452} establishes closedness of $\mathcal{A}$.
Let us now assume that $C$ is compact. Then there are only finitely many $i\in I$ such that $C_i \neq \emptyset$ we conclude that $\mathcal{A}$ is a Fr\'{e}chet space as a closed subspace of a finite product of such spaces. Furthermore, $\Gamma_c (M,E)$ is isomorphic to a closed subspace (cf.\ Lema \ref{lem: top:sect}) of the webbed space $\bigoplus_{i} C^\infty_{\text{co}} (U_i,\ensuremath{\mathbb{R}}^m)$, whence $\Gamma_c (C,E)$ is webbed as a qotient of a webbed space (\cite[Lemma 24.28]{MR1483073}). Now the open mapping theorem \cite[24.30]{MR1483073} shows that $r$ is open as a mapping into $\mathcal{A}$, whence $r$ induces the claimed isomorphism of locally convex spaces.
\end{proof}
\begin{remark}
Note that the topology on $\Gamma_c (C,E)$ does not automatically coincide with the quotient topology of $\res_C$ if $C$ is not compact.
Studying the above proof, the open mapping theorem is not applicable since $\mathcal{A}$ is not necessarily ultrabornological (as it is not clear that it would be a limit subspace of the direct sum).
In fact the authors do not know whether the quotient topology may be properly finer in the non-compact case.
\end{remark}
However, the problem mentioned in the last remark is not relevant for us, since we will only consider sets which allow continuous extension operators. In the presence of such a section, the two topologies coincide:
\begin{lemma}\label{lem: op:ext}
Assume that there exists a continuous section $s \colon \mathcal{A} \rightarrow \Gamma_c (M,E)$ of the map $r\circ \res_C$, then the quotient topology turns $r$ into an isomorphism $\Gamma_c (C,E)\cong \mathcal{A}$.
\end{lemma}
\begin{proof}
Since $r\circ \res_C$ is continuous surjective and admits a (global) continuous section, it is a quotient map between locally convex spaces. As $r^{-1} \circ (r \circ \res_C) = \res_C$ we deduce that $r^{-1} \colon \mathcal{A} \rightarrow \Gamma_c (C,E)$ is continuous, whence $r$ induces and an isomorphism of locally convex spaces onto its image.
\end{proof}
Thus we in the situation of Theorem A (to be proved in the end of the section), the topologies coincide.
\subsection*{Interlude: Patching by partition of unity
\addcontentsline{toc}{subsection}{Interlude: Patching by partition of unity}
In this interlude, we discuss continuity properties for the map which patches mappings on a locally finite-covering by means of a partition of unity.
\begin{setup}\label{setup: settingup}
Recall that for a given compact subset $K$ of $M$ only finitely many members of
the locally finite open cover $\ensuremath{\mathcal{U}}$ have a non-trivial
intersection with $K$. Thus for each $i\in I$ we obtain a finite subset of
$I$ by setting
\begin{displaymath}
J_i \colonequals \{ j \in I \mid U_j \cap U_i \neq \emptyset\}
\end{displaymath}
\end{setup}
\begin{setup}\label{abuse_notation}
Fix $n\in \ensuremath{\mathbb{N}}$ and consider for $i\in I$ maps $f_j \in C^\infty (U_j,\ensuremath{\mathbb{R}}^n)$ for $j \in J_i$. Multiplying with the
partition of unity \ref{setup: prep:ThmA}, we obtain for every such pair a smooth mapping
$f_{ji} \colonequals \chi_j|_{V_j \cap U_i} \cdot f_j|_{V_j \cap U_i}$ defined on the (possibly empty) set $V_j \cap U_i$. Note that since $\supp \chi_j \subseteq V_j$, the mapping vanishes in a neighborhood of the boundary of $V_j \cap U_i$ in $U_i$. Thus we can extend $f_{ji}$ by $0$ to a smooth map on all of $U_i$ (or by a similar argument to all of $V_i$). In the following we will extend
these mappings to all of $U_i$ (or similarly to $V_i$) and suppress the extension in the notation.
\end{setup}
\begin{lemma}\label{lem: sm:mix}
Using the notation from \ref{abuse_notation}, the mixing map
\begin{align*}
\mu \colon \bigoplus_{i\in I} C^\infty_{co} (U_i,\ensuremath{\mathbb{R}}^n) & \longrightarrow \bigoplus_{i \in I} C^\infty_{co} (V_i,\ensuremath{\mathbb{R}}^n) \\
(f_i)_{i \in I} & \mapsto \left(\sum_{j\in J_i}
(\chi_j|_{V_i\cap V_j}) \cdot f_j|_{V_{i}\cap V_j} \right)_{i\in I}
\end{align*}
is continuous linear. Its image is contained in the closed subspace $A \colonequals
\{(g_i)_{i\in I} \mid g_j|_{V_i \cap V_j} = g_i|_{V_i\cap V_j} \forall
i,j \in I \}$.
\end{lemma}
\begin{proof}
\textbf{The mapping $\mu$ makes sense.}
As argued in \ref{abuse_notation}, every component of $\mu (f_i)_{i \in I}$ is a smooth function as a finite sum of such functions.
Note that every $i \in I$ appears only in finitely many of the
sets $J_k, k \in I$. Thus every $f_i$ appears at most in finitely many of the sums of the definition of $\mu$, whence $\mu$ makes sense as a mapping
between direct sums. Clearly $\mu$ is linear.
\paragraph{$\mu$ takes its image in $A$}
Observe that by
construction we have $\supp \chi_k \subseteq V_k$. Hence if $\chi_k$
does not vanish on $V_i \cap V_j$ we must have $k \in J_i \cap J_j$.
Thus the sum in $\sum_{k\in J_i} (\chi_k|_{V_k \cap V_i}) \cdot f_k|_{V_i \cap V_k}$
coincides on $V_i \cap V_j$ (up to vanishing
summands) with the one summing over $J_j$. In conclusion,
$\mu$ takes its image in $A$.
\paragraph{Continuity of the auxiliary mappings $m_i$}
Let us first
fix $i\in I$ and consider the linear map
$$m_i \colon \bigoplus_{j \in
J_i} C^\infty_{co} (U_j,\ensuremath{\mathbb{R}}^n) \rightarrow C^\infty (V_i,\ensuremath{\mathbb{R}}^n),\quad (f_j)_j
\mapsto \sum_{j\in J_i} (\chi_j|_{V_i\cap V_j}) \cdot f_j|_{V_i \cap V_j} .
$$
As $J_i$ is finite and $C^\infty_{co} (V_i , \ensuremath{\mathbb{R}}^n)$ is a
topological vector space it clearly suffices to establish smoothness for all of the mappings
\begin{align*}
c_j\colon C^\infty_{co} (U_j,\ensuremath{\mathbb{R}}^n) &\rightarrow C^\infty_{co} (V_i, \ensuremath{\mathbb{R}}^n),\quad j \in J_i\\
f &\mapsto \chi_j|_{V_i\cap V_j} \cdot f|_{V_i \cap V_j}.
\end{align*}
Recall that the space $C^\infty_{co} (U_j,\ensuremath{\mathbb{R}}^n)$ is a topological $C^\infty_{co} (U_j,\ensuremath{\mathbb{R}})$-module (see e.g.\ \cite[Corollary F.13]{math/0408008v1}). Thus the map $\kappa_j (f) \colonequals \chi_j|_{U_j} \cdot f$ is continuous, takes its image in the linear subspace $C^\infty_{\overline{V}_j} (U_j,\ensuremath{\mathbb{R}}^n) \subseteq C^\infty_{\text{co}} (U_j,\ensuremath{\mathbb{R}}^n)$ of smooth functions supported in $\overline{V}_j$.
Now \cite[Lemma 4.24 and Lemma 4.6]{math/0408008v1}
extending functions in $C^\infty_{\overline{V}_j} (U_j, \ensuremath{\mathbb{R}}^n)$ by $0$ to all
of $M$ and restricting then to $V_i\cap V_j$ yields a continuous linear map which, composed with $\kappa_j$, coincides with $c_j$. We conclude that $c_j$ and thus $m_i$ is continuous linear.
\paragraph{Continuity of $m$}
We define the mapping
\[
\tilde{\mu} \colonequals \oplus_{i\in I} m_i \colon \bigoplus_{i \in I} \bigoplus_{j \in J_i} C^\infty_{co}
(U_i ,\ensuremath{\mathbb{R}}^n) \rightarrow \bigoplus_{i \in I} C^\infty_{co} (V_i ,\ensuremath{\mathbb{R}}^n).
\]
This mapping is continuous linear, since the
mappings $m_i$ are so by the previous step.
It follows from the universal property of the locally convex direct sum that
\begin{align*}
B_{diag} \colon \bigoplus_{i \in I} C^\infty (U_i, \ensuremath{\mathbb{R}}^n) &\rightarrow
\bigoplus_{i\in I} \bigoplus_{\# \{k \in I \mid i \in J_k\}} C^\infty (U_i,\ensuremath{\mathbb{R}}^n) \cong
\bigoplus_{i\in I} \bigoplus_{j \in J_i} C^\infty (U_j,\ensuremath{\mathbb{R}}^n)\\
(f_i)_{i\in I} &\mapsto\quad ( \oplus_{\# \{k \in I \mid i \in J_k\}} f_i)_{i\in I} \quad \mapsto \quad ((f_j)_{j\in J_i})_{i\in I}
\end{align*}
is continuous linear (where due to the construction, there is a
bijection between the index sets of both sums).
Now, we have $\mu = \tilde{\mu} \circ B_{diag}$ and thus $m$ is continuous linear as a composition of such mappings.
\end{proof}
\subsection*{Global extensions of bundle sections on a closed set}
\addcontentsline{toc}{subsection}{Global extensions of bundle sections on a closed set}
We will now prove Theorem A from the introduction, whose statement we repeat here for convenience.
\begin{theorem}[Theorem A]\label{thmA}
Let $C\subseteq M$ be a closed set satisfying the cusp condition. Then the restriction map $\res_C \colon \Gamma_c (M,E) \rightarrow \Gamma_c (C,E)$ admits a continuous linear section $\mathcal{E}_C^M$.
\end{theorem}
\begin{proof}
We use the notation and data introduced in \ref{setup: prep:ThmA}.
For the proof we consider a commutative diagram of locally convex spaces (where the numbers indicate where the (continuous) linear map was constructed):
\begin{equation} \begin{aligned}
\begin{xy}
\xymatrix{
\Gamma_c (C,E) \ar[r]^-{r}_-{\eqref{eq: ex:vssect}} \ar@/^1pc/@{-->}[dd]^{\mathcal{E}_C^M} & \bigoplus_{i \in I} C^\infty (C_i)^m\ar[rr]^-{\text{extension}}_-{\text{Lemma \ref{lem: multiFrerick}}} & & \bigoplus_{i \in I} C^{\infty}_{\text{co}} (U_i)^m \ar[ldd]|-{\stackrel{\mu|^{\im(\rho_\ensuremath{\mathcal{V}})}}{\text{Lemma }\ref{lem: sm:mix}}} \ar[dd]^{\mu} \\
& & & \\
\Gamma_c (M,E) \ar[uu]^{\res_C} \ar[rr]^{\rho_\ensuremath{\mathcal{V}}} && \im(\rho_\ensuremath{\mathcal{V}}) \ar[r]^-{\text{inc}} & \bigoplus_{i \in I} C^\infty_{\text{co}} (V_i,\ensuremath{\mathbb{R}}^m)
}
\end{xy}\end{aligned}\label{diag: l:con}
\end{equation}
We postpone the proof of the commutativity of \eqref{diag: l:con} to Lemma \ref{lem: multiFrerick} below, where also the extension map needed in the computation is defined.
Now $\mathcal{E}_C^M$ is defined via the right half of the diagram (using that $\rho_\ensuremath{\mathcal{V}}$ is a topological embedding onto its image by Lemma \ref{lem: top:sect}). Since all the mappings in the definition are continuous and linear, $\mathcal{E}_C^M$ is a continuous linear section of $\res_C$.
\end{proof}
\begin{lemma}\label{lem: multiFrerick}
There exists a continuous linear extension map $\varepsilon \colon \bigoplus_{i \in I}C^\infty(C_i)^m \rightarrow \bigoplus_{i\in I} C^\infty_{\text{co}} (\ensuremath{\mathbb{R}}^d)^m$ which makes \eqref{diag: l:con} commutative.
\end{lemma}
\begin{proof}
We construct continuous linear mappings $\varepsilon_i \colon C^\infty (C_i)^m \rightarrow C^\infty_{\text{co}} (\ensuremath{\mathbb{R}}^d)^m$ and define $\varepsilon \colonequals \oplus_{i \in I} \varepsilon_i$. Thus $\varepsilon$ will be continuous linear by the properties of the direct sum. For the construction we distinguish two cases depending on $i\in I$:
\paragraph{Case 1: $C_i = \emptyset$.} Since the chart does not intersect the domain of our map, only have to extend the empty function, whence $E_i$ is simply defined as the constant $0$-map in this case.
\paragraph{Case 2: $C_i\neq \emptyset$.} Due to our setup, the sets $C_i$ satisfy the assumptions made in the statement of Lemma \ref{lemma:splitting restriction in charts}. Hence in this case there is a continuous linear extension operator $C^\infty (C_i) \rightarrow C^\infty (U_i)$. We define $\varepsilon_i$ as the $m$-fold product of this extension operator.
This completes the construction of the extension map and all there is left is to prove that the diagram \eqref{diag: l:con} commutes.
However, this is obvious from a trivial calculation if one recalls the following facts:
\begin{itemize}
\item For each pair $i,j \in I$ and $(f_k)_k \in \im r$ we have $f_i|_{C_i \cap C_j} = f_j|_{C_i \cap C_j}$,
\item the extension operators $\varepsilon_i$ do not change the map on the $C_i$,
\item composition with $\rho_V^{-1}\mu$ is just mixing and restricting with a partition of unity and then reconstruction via the sheaf property of smooth maps.
\end{itemize}
Composing again with the restriction map, the outer square of \eqref{diag: l:con} commutes.
\end{proof}
As a direct consequence of the above theorem, we obtain:
\begin{corollary}\label{splitting consequences}
If $C\subset M$ is a
closed subset which satisfies the cusp condition in the geodesic metric on $M$, the exact sequence \eqref{eq: ex:vssect} splits, i.e.\ we have the isomorphim of topological spaces $\Gamma_c (M,E) \cong \mathcal{I}_c (C,N) \oplus \Gamma_c (C,N)$.
\end{corollary}
\begin{remark}
As a consequence of Theorem \ref{thmA} and Lemma \ref{lem: op:ext} the map $r$ from Proposition \ref{prop: Whit:mfd} and \eqref{diag: l:con} is a topological embedding onto a closed subspace for every
closed subset which satisfies the cusp condition.
\end{remark}
\section{Manifolds of mappings for manifolds with rough boundary}
\label{sect: mfdmap:rb}
In this section we recall some essentials on manifolds with rough boundary from \cite[Chapter 1.4]{GloecknerNeebBuch}.
Then we recall the classical construction of manifolds of mappings and how to apply them to the rough boundary case.
\begin{definition}\label{defn: RBM}
We recall from \cite{GloecknerNeebBuch} (cf.\ \cite[Section 4]{alas2012}) that a \emph{manifold with rough boundary} modelled on a locally convex space~$E$ is a Hausdorff topological space $M$ with an atlas of smoothly compatible
homeomorphisms $\phi\colon U_\phi\to V_\phi$ from open subsets $U_\phi$ of $M$ onto locally convex subsets $V_\phi\subseteq E$ with dense interior (to distinguish from ordinary manifold charts, they are also called \emph{rough} $E$-charts).
If $x\in M$ we call $x$ a \emph{formal boundary point} if there is a rough $E$-chart $(U_\varphi, \varphi)$ around $x$ such that $\varphi(x) \in \partial \varphi(U_\varphi)$. Denote by $\partial M$ the \emph{(formal) boundary} of $M$, i.e.\ the set of formal boundary points of $M$.
If each $V_\phi$ is open, $M$ is an ordinary manifold (without boundary). If each $V_\phi$ is relatively open in a closed hyperplane $\lambda^{-1}\!\left([0,\infty[\,\right)$, where $\lambda\in E'$ (the space of continuous linear functional on $E$), then $M$ is a \emph{manifold with smooth boundary}. In the case of a \emph{manifold with corners}, each $V_\phi$ is a relatively open
subset of $\lambda^{-1}_1\!\left([0,\infty[\,\right)\cap\cdots\cap \lambda^{-1}_n\!\left([0,\infty[\,\right)$, for suitable $n\in \ensuremath{\mathbb{N}}$ (which may depend on $\phi$) and linearly independent $\lambda_1,\ldots,\lambda_n\in E'$.
\end{definition}
The boundary of manifolds with rough boundary is characterised by the following.
\begin{lemma}\label{lem: rb:boundary}
Let $M$ be a manifold with rough boundary and $(U_\varphi, \varphi)$, $(U_\psi,\psi)$ be rough $E$-charts around $x\in M$. Then $\varphi (x) \in \partial \varphi(U_\varphi)$ if and only if $\psi (x) \in \partial \psi (U_\psi)$.\footnote{A full proof is contained in the forthcoming \cite[Section 3]{GloecknerNeebBuch}. However here is a rough sketch: Argue by contradiction. In the chart where the image is in the boundary choose a convex neighborhood $W$. Now apply the Hahn-Banach theorem to separate the image of $x$ from the interior of $W$ by a functional $\lambda$. Taking the derivative $\kappa$ of the change of charts, one derives a contradiction by considering $\lambda \circ \kappa^{-1}$ on $\kappa (W)$.}
\end{lemma}
In essence Lemma \ref{lem: rb:boundary} shows that the formal boundary of $M$ arises from the topological boundary of the images of charts in the model space.
\begin{setup}
By virtue of the chain rule Lemma \ref{lem: reschain} 2.\ we can define smooth mappings on manifolds with rough boundary in the usual way.
Direct products of manifolds with rough boundary, tangent spaces and tangent bundles\footnote{The definition of tangent vectors at $x$ as equivalence classes of vectors in the model space, i.e.\ $(\psi,x,v) \sim (\varphi,x,w)$ if and only if $d(\psi\circ \varphi^{-1}) (\varphi (x),v)=w$ where $\varphi,\psi$ are manifold charts, makes sense for a manifold with rough boundary and yields the usual results and identifications for iterated tangent bundles, cf.\ also \cite[Section 2]{michor}.} as well as vector bundles may be defined as usual.
\end{setup}
We recall the compact open $C^\infty$-topology on the space $C^\infty (M,N)$ of smooth mappings from a manifold with rough boundary into a manifold without boundary.
\begin{definition}\label{defn: topo:init}
Let $M,N$ be manifolds with rough boundary.
We define compact-open $C^\infty$-topology on $C^\infty (M,N)$ as the intial topology induced by the mappings
\begin{displaymath}
T^k \colon C^\infty (M,N) \rightarrow C_{\text{co}}(T^k M,T^k N),\quad f \mapsto T^k f, \quad k \in \ensuremath{\mathbb{N}}_0
\end{displaymath}
where the right hand side carries the compact open topology. We let $C^\infty_{\text{co}} (M,N)$ denote the space with the compact open $C^\infty$-topology.
\end{definition}
\begin{remark} \label{rem: coinfty}
If $N=\ensuremath{\mathbb{R}}^m$ the compact open $C^\infty$ topology coincides with topology described in \cite[Definition 3.21 and Section 4]{alas2012} by adapting the argument of \cite[Lemma 1.14]{MR1911979} to manifolds with rough boundary.
In addition, if $M$ is a locally convex regular closed subset of $\ensuremath{\mathbb{R}}^m$ and $N=\ensuremath{\mathbb{R}}^m$, the compact open $C^\infty$-topology coincides with the identification topology from \ref{setup: WJ} (by an argument analogous to \ref{Whitney:fun}).
\end{remark}
\begin{proposition}\label{prop: fcomp:cont}
Let $f \colon N\rightarrow B$ and $g \colon A \rightarrow M$ be smooth mappings between finite-dimensional manifolds with rough boundary. Then
\begin{displaymath}
f^* g_* \colon C^\infty_{\text{co}} (M,N) \rightarrow C^\infty_{\text{co}} (A,B), \quad h \mapsto f\circ h \circ g
\end{displaymath}
is continuous.
\end{proposition}
\begin{proof}
The usual proof for manifolds without boundary (see e.g.\ \cite[Lemma 5.5.]{Wockel13}) carries over without any changes.
\end{proof}
\begin{lemma}\label{top: embedding}
The initial topology turns the map
$$\mathcal{T} \colon C^\infty_{\text{co}} (M,N) \rightarrow \prod_{k \in\ensuremath{\mathbb{N}}_0} C_{\text{co}}(T^k M,T^k N),\quad f \mapsto (T^k f)_{k \in \ensuremath{\mathbb{N}}_0}$$
into a topological embedding with closed image.
\end{lemma}
\begin{proof}
By definition of an initial topology, the map $\mathcal{T}$ is a topological embedding. Let now $(T^kf^\alpha)_{k\in \ensuremath{\mathbb{N}}_0, \alpha \in I}$ be a net in the image of $\mathcal{T}$ which converges to $(f_k)_{k \in \ensuremath{\mathbb{N}}_0}$.
If we can prove that $f_k = T^k f_0$ holds for every $k\in \ensuremath{\mathbb{N}}$ then the image of $\mathcal{T}$ is closed. Clearly we can verify the formula $f_k = T^k f_0$ locally in charts. As the sequence converges with respect to the compact open $C^\infty$-topology, the usual inductive proof using the fundamental theorem of calculus (\cite[Theorem 1.5 and Lemma 1.7]{MR1911979} which is valid on locally convex regular subsets of $\ensuremath{\mathbb{R}}^d$!) carries over without any changes, see e.g.\ \cite[Lemma 5.13 and Theorem 5.14]{Wockel13}.
\end{proof}
As a consequence of the above identification, we obtain the following completeness and metrisation results (which are well known in the case of a manifold without boundary).
\begin{lemma}\label{lem: met:top}
Let $M,N$ be manifolds with rough boundary, such that $M$ is locally compact\footnote{In contrast to manifolds without boundary, manifolds with rough boundary need not be locally compact. For example, recall that regular locally convex subsets of $\ensuremath{\mathbb{R}}^d$ are in general not locally compact, e.g.\ $\{(0,0)\} \cup \{(x,y) \in \ensuremath{\mathbb{R}}^2 \mid x >0\}$ in $\ensuremath{\mathbb{R}}^2$.} and $\sigma$-compact and $N$ metrisable and modelled on a metrisable space. Then $C^\infty_{\text{co}} (M,N)$ is metrisable.
\end{lemma}
\begin{proof}
For convenience let $d$ be the dimension of $M$. Since $M$ is locally compact, for every manifold chart $(\varphi, U)$ the domain $U$ is locally compact, whence $\varphi (U)$ is locally compact subset of $\ensuremath{\mathbb{R}}^d$. As $\ensuremath{\mathbb{R}}^d$ is second countable, $\varphi (U)$ is locally compact and second countable, whence $\sigma$-compact.
Using the canonical atlas for the iterated tangent bundle, we see that locally over $U$ we get a bundle trivialisation $T^k M \supseteq T^k U \cong \varphi (U) \times (\ensuremath{\mathbb{R}}^d)^{2^k-1}$. Thus $T^kM$ is $\sigma$-compact for every $k \in \ensuremath{\mathbb{N}}_0$.
Since $N$ is metrisable and modelled on a metrisable space, $T^kN$ is metrisable by \cite{EG54}. Thus the spaces $C_{\text{co}}(T^kM,T^kN)$ are metrisable by \cite[Proposition 5.10 e)]{Wockel13} whence the embedding from Lemma \ref{top: embedding} indentifies $C^\infty_{\text{co}} (M,N)$ as a subspace of a metrisable space.
\end{proof}
\begin{corollary}\label{cor: complete}
Let $M$ be a locally compact manifold with rough boundary and $F$ a Fr\'{e}chet space. Then $C^\infty_{\text{co}} (M,F)$ is a Fr\'{e}chet space.
\end{corollary}
\begin{proof}
It is well known that the pointwise operations turn $C^\infty_{\text{co}} (M,F)$ into a topological vector space (cf.\ \cite[Section 2]{alas2012}).
By Lemma \ref{lem: met:top} the space $C^\infty_{\text{co}} (M,F)$ is metrisable. In addition, \cite{EG54} implies that every $T^kM$ is metrisable, whence a $k$-space. Now $T^kF\cong F^{2^k}$ is complete and we infer from \cite[Remark 3.2 (a)]{MR1934608} that $C_{\text{co}} (T^kM,T^kF)$ is complete. Now Lemma \ref{top: embedding} identifies $C^\infty_{\text{co}} (M,F)$ with a closed subspace of the complete space $\prod_{k \in \ensuremath{\mathbb{N}}_0} C_{\text{co}} (T^kM,T^kF)$, whence $C^\infty_{\text{co}} (M,F)$ is a Fr\'{e}chet space.
\end{proof}
Finally we turn to smooth sections of bundles over a manifold with rough boundary.
\begin{setup}\label{setup: sect:loc}
Let $p \colon E \rightarrow M$ be a vector bundle with typical fibre $F$.
Assume that $F$ is a Fr\'{e}chet space and $M$ is a compact manifold with rough boundary. Then we define
$$\Gamma (M,E) \colonequals \{\sigma \in C^\infty (M,E) \mid p \circ \sigma = \\id_M\},$$
and endow it with the subspace topology induced by $C^\infty_{\text{co}} (M,E)$.
Note that since $M$ is compact, we can choose an open cover $(W_i,\kappa_i)_{1\leq i \leq n}$ of domains of bundle trivialisations for $E$ and denote by $E|_{U_i}$ the restricted bundle over $U_i$. Then define the map
$$\rho \colon \Gamma (M,E) \rightarrow \prod_{1 \leq i \leq n} \Gamma (W_i,E|_{W_i)}),\quad \sigma \mapsto (\sigma|_{W_i})_i.$$
By Lemma \ref{prop: fcomp:cont} the map $\rho$ is continuous (as each of its components are given by mappings $(\iota_i)^*$, where $\iota_i \colon W_i \rightarrow M$ is the inclusion). Clearly $\rho$ is injective, linear and identifies $\Gamma (M,E)$ with the subspace $\mathcal{C}\colonequals \{(\gamma_i)_i \mid \gamma_i|_{W_i \cap W_j} = \gamma_j|_{W_i\cap W_j}\}$. Since the the evaluation map is continuous by \cite[Proposition 3.20]{alas2012}, $\mathcal{C}$ is closed. Working with subbasic neighborhoods, one can also prove that $\rho$ is open onto its image. We refer to \cite[Lemma 6.4]{Wockel13} for details.
Now we obtain isomorphisms $\Gamma (W_i , E|_{W_i}) \rightarrow C^\infty_{\text{co}} (W_i,F), \sigma \mapsto \pr_2 \circ \psi_i \circ \sigma$ of topological vector spaces.
Thus Corollary \ref{cor: complete} implies that $\Gamma (M,E)$ is a Fr\'{e}chet space as a closed subspace of a direct product of Fr\'{e}chet spaces.
\end{setup}
Having discussed the topology on the manifold of mappings, we will now construct an infinite-dimensional manifold structure on $C^\infty_{\text{co}} (M,N)$ for $M$ a compact manifold with rough boundary and $N$ a manifold without boundary.
If $M$ is a manifold with corners, such a construction (even for $M$ non-compact) can be found in \cite{michor}. Our proof follows the general idea of loc.cit.\ but we avoid using an instance of the so called $\Omega$-Lemma.
For the rest of this section $M$ will be a compact manifold with rough boundary and $N$ will be a Fr\'{e}chet manifold without boundary which admits a local addition.
\begin{setup}
A manifold $N$ admits a local addition, if there is a \emph{local addition} $\Sigma_N$ on $N$, i.e.\ a smooth mapping $\Sigma_N \colon TN \supseteq \Omega \rightarrow N$ on an open $0$-neighborhood $\Omega$, such that
\begin{enumerate}
\item $\Sigma_N \circ 0_N = \id_N$, where $0_N \colon N \rightarrow TN$ is the $0$-section.
\item $(\pi_{N}|_{\Omega}, \Sigma_N) \colon \Omega\rightarrow N \times N$ induces a diffeomorphism onto an open neighborhood of the diagonal in $N \times N$.
\end{enumerate}
\end{setup}
Using the local addition, we obtain canonical charts for the mapping space:
\begin{setup}[Canonical charts]\label{setup: can:charts}
For $f \in C^\infty (M,N)$ we let $f^*TN$ be the pullback bundle of the tangent bundle of $N$ with associated bundle map $F\colon f^*TN \rightarrow TN$.
Via the bundle map $F$, we can identify the Fr\'{e}chet space of sections $\Gamma (M,f^*TN)$ (cf.\ \ref{setup: sect:loc}) with
$$C_f^\infty (M,TN) \colonequals \{g \in C^\infty (M,N) \mid \pi_N \circ g = f\} \subseteq C^\infty_{\text{co}} (M,TN)$$
in the subspace topology. In the following we will suppress this (harmless) identification without further notice.
Then define the sets
\begin{align*}
V_f \colonequals \{g \in C^{\infty} (M,E) \mid g(M) \subseteq \Omega\} \cap C^\infty_f (M,TN), \\
U_f \colonequals \{g \in C^\infty (M,N) \mid (f,g)(M) \subseteq (\pi,\Sigma_N) (\Omega)\} \subseteq C^\infty_{\text{co}} (M,N)
\end{align*}
and note that both are open in the compact open $C^\infty$-topology.
Now Proposition \ref{prop: fcomp:cont} implies that $\varphi_f \colon U_f \rightarrow W_f, g \mapsto (\pi_N, \Sigma_N)^{-1} \circ (f,g)$ is a homeomorphism with inverse $\varphi_f^{-1} (\tau) = (\Sigma_N)_* (\tau)$.
\end{setup}
\begin{theorem}\label{thm: mm:rb}
Let $M$ be a compact manifold with rough boundary and $N$ a metrisable manifold modelled on a Fr\'{e}chet space which admits a local addition.
The atlas of canonical charts $(U_f,\varphi_f)_{f\in C^\infty (M,N)}$ turns $C^\infty_{\text{co}} (M,N)$ into a Fr\'{e}chet manifold modelled on spaces of sections $\Gamma (M, f^*TN)$.
The manifold structure does not depend on the choice of local addition.
\end{theorem}
\begin{proof}
The compact open $C^\infty$-topology on $C^\infty (M,N)$ is clearly finer than the compact open topology, whence $C^\infty_{\text{co}} (M,N)$ is Hausdorff.
In \ref{setup: can:charts} we have already seen that the canonical charts form an atlas of homeomorphisms. Hence we only have to prove that the change of charts are smooth.
\paragraph{Change of charts formula.} To this end, one observes that $h \colonequals \varphi_f \circ \varphi_g^{-1}$ has an open domain, $O_{f,g} \subseteq \Gamma (M,g^*TN)$ for each pair $f,g \in C^\infty (M,N)$.
Let now $\tau$ be in $O_{f,g}$ and $x\in M$, then we obtain the formula
\begin{equation}\label{eq: chch:form}
h \colon O_{f,g} \rightarrow \Gamma (M,f^*TN),\quad h(\tau)(x) = (\pi_N, \Sigma_N)^{-1} (f(x) , \Sigma_N \circ \tau (x))
\end{equation}
\paragraph{Localisation in charts}
Choose an atlas of local trivialisations $(W_i,\kappa_i^f)_{1\leq i\leq n}$, for the bundle $f^*TN$. Adjusting our choices if necessary, we may assume that for each $1\leq i\leq n$ $W_i$ is \begin{enumerate}
\item a domain of a bundle trivialisations $(W_i,\kappa_i^g)$ of $g^*TN$
\item the domain of a manifold chart $(W_i,\psi_i)$ of $M$.
\end{enumerate}
Now construct the topological embedding $\rho \colon \Gamma (M,f^*TN) \rightarrow \prod_{1\leq i\leq n} C^\infty_{\text{co}} (W_i,F)$ as in \ref{setup: sect:loc}, where $F$ is the model space of $N$.
Since precomposition with a smooth is continuous in the compact open $C^\infty$-topology by Proposition \ref{prop: fcomp:cont}, we see that $C^\infty_{\text{co}} (W_i,F) \cong C^\infty_{\text{co}} (\psi_i (W_i),F)$ as locally convex spaces (similarly for $g^*TN$).
We remark for later use that since $M$ is compact, $W_i$ and also $\psi_i (W_i)$ are locally compact.
\paragraph{Smoothness via the exponential law.}
Now $\rho$ is a topological embedding with closed image, whence $h$ will be smooth if and only if $\rho \circ h$ is smooth and this is the case if and only if for each $1\leq i \leq n$ the mapping
$$h_i \colon O_{f,g} \rightarrow C^\infty (W_i,F), \tau \mapsto \pr_2 \circ \kappa_i \circ h(\tau)|_{W_i}$$
is smooth.
Now we recall that $O_{f,g} \subseteq \Gamma (M,g^*TN)$ is open and the spaces $\Gamma (M,g^*TN)$ and $C^\infty (W_i,F)$ are Fr\'{e}chet spaces by \ref{setup: sect:loc}. Since $W_i$ is a manifold with rough boundary (being an open subset of $M$) the exponential law \cite[Theorem B]{alas2012} for smooth mappings on manifolds with rough boundary yields: The $h_i$ (and thus $h$) are smooth if and only if the mapping
$$h_i^\vee \colon O_{f,g} \times W_i \rightarrow F,\quad h_i^\vee (\tau,x) \colonequals h_i (\tau)(x)$$
is smooth. However, \eqref{eq: chch:form} allows us to write
\begin{equation}\label{eq: chch:exp}
h_i^\vee (\tau) (x) = \pr_2 \circ (\pi_N, \Sigma_N)^{-1} \circ (f, \Sigma_N \circ (\ev(\tau|_{W_i})) (x),
\end{equation}
where $\ev \colon \Gamma (W_i, g^*TN|_{W_i}) \times W_i \rightarrow g^*TN|_{W_i}, (f,x)\mapsto f(x)$ is the evaluation map.
Using that $g^*TN|_{W_i}$ is trivial we identify $\Gamma (W_i, g^*TN|_{W_i}) \cong C^\infty_{\text{co}} (\psi_i (W_i), F)$ and deduce from \cite[Proposition 3.20]{alas2012} that $\ev$ is smooth. As the restriction of $\tau$ is smooth by \ref{setup: sect:loc}, we deduce from \eqref{eq: chch:exp} that $h_i^\vee$ is smooth as a composition of smooth functions.
Summing up the change of charts are smooth and the canonical charts form indeed a smooth atlas turning $C^\infty (M,N)$ into a Fr\'{e}chet manifold
\paragraph{The construction is independent of the choice of local addition.} Replacing the local addition $\Sigma_N$ by $\tilde{\Sigma}_N$, the change of charts formula \eqref{eq: chch:form} shows that the change of charts between a chart with respect to $\Sigma_N$ and with respect to $\tilde{\Sigma}_N$ will be smooth.
Hence the manifold structure does not depend on the choice of local addition.
\end{proof}
\begin{remark}
A crucial ingredient in the proof of Theorem \ref{thm: mm:rb} was compactness of $M$ to endow the function space $C^\infty (M,N)$ with the compact open $C^\infty$ topology and to use the exponential law instead of the so called $\Omega$-Lemma \cite[8.7]{michor}.
Though the authors believe that for non-compact $M$, the space $C^\infty (M,N)$ can be endowed with a manifold structure along the lines described in \cite{michor}, this would involve two significant steps: One has to define a version of the fine very strong topology for mapping spaces on non-compact manifolds of mappings and reprove the results outlined in \cite{HS17} (mostly trivial with the notable exception of continuity of the composition).
Then one needs an analogue of the $\Omega$-Lemma for manifolds with rough boundary (which will be contained in \cite{GloecknerNeebBuch}, due to H.\ Gl\"{o}ckner, private communications).
\end{remark}
\section{Submanifolds with rough boundary and the proof of Theorem B}\label{sect: MFDMAP}
In this section we establish the global version of the splitting of spaces of compactly supported sections.
Our aim is to construct submersions between the infinite-dimensional manifolds of mappings.
To this end we need to clarify first our concept of a submanifold with boundary sitting inside of manifolds without boundary. If we require the submanifold with rough boundary to be closed, then it will automatically have no narrow fjords. The authors believe that they will also automatically satisfy the cusp condition but were not able to prove the latter statement.
\begin{definition}[Submanifold with rough boundary]
Let $M$ be a finite-dimensional manifold (possibly with rough boundary). A subset $S \subseteq M$ is called \emph{(embedded) submanifold with rough boundary} of $M$ if for every $p \in S$ there is a chart $(U_p,\varphi_p)$ of $M$ with $p\in U_p$ and $\varphi_p (p)=0$ and a regular locally convex subset $R_p$ such that $\varphi_p (S \cap U_p) = \varphi_p(U_p) \cap R_p$.
If for every $p \in S$ the regular locally convex set is a relatively open set in a quadrant (cf.\ Definition \ref{defn: RBM}), then we say that $S$ is an (embedded) \emph{submanifold with corners}.
If in addition $S$ is a closed subset which satisfies the above conditions, we say that $S$ is a \emph{closed submanifold with rough boundary} (or \emph{with corners}, respectively).
\end{definition}
\begin{remark}
\begin{enumerate}
\item A submanifold with rough boundary inherits the structure of a manifold with rough boundary from the ambient manifold and this structure turns the inclusion $\iota_S \colon S \rightarrow M$ into a smooth embedding.
Thus submanifolds with rough boundary as defined here are initial submanifolds, i.e.\ a mapping $f \colon N \rightarrow S \subseteq M$ between manifolds with rough boundary is smooth as a map into $S$ if and only if it is smooth as a map into $M$.
Note that closed submanifolds of $\sigma$-compact manifolds are again $\sigma$-compact.
\item We remark that our definition of an embedded submanifold with corners is a special case of a submanifold with corners as in \cite[2.5]{michor}.
Since we are only interested in a very specialised case, we do not need the more general definition.
In particular, we refrain from defining submanifolds of lower dimension (which could be done as usual but is not needed here).
\item Due to our definition an embedded submanifold with rough boundary $S\subseteq M$ is regular: Consider $x\in \partial S$ and let $(U_\varphi,\varphi)$ be a submanifold chart.
Then $\varphi(x) \in \varphi (U_\varphi \cap S) = \varphi(U_\varphi) \cap C$ for a regular locally convex set $C$.
Hence $W \colonequals \varphi^{-1}(\varphi (U_\varphi \cap C^\circ)) \subseteq S$ is an open set in $M$, whence contained in the interior of $S$.
Choosing a sequence in W we can approximate $\varphi (x)$, whence $x \in \overline{S^\circ}$.
Thus $S$ is regular and we see in addition that $x \in \partial S$ entails $\varphi (x) \in \partial \varphi(U_\varphi)$.
\end{enumerate} \label{rem: submfd:rb}
\end{remark}
Before we continue, let us construct a class of examples for submanifolds with rough boundary of a Riemannian manifold which will be used to prove Corollary D from the introduction.
\begin{lemma}\label{lem: sconv:rbsmfd}
Let $M$ be a Riemannian manifold and $C$ be a regular closed subset which is strongly convex, i.e.\ for every $p,q \in C$ exists a unique minimal geodesic segment $\overline{pq}$ connecting $p$ and $q$ such that $\overline{pq} \subseteq C$.
Then $C$ is a submanifold of $M$ with rough boundary.
\end{lemma}
\begin{proof}
By standard Riemannian geometry, we can choose for every $p\in C$ an open $0$-neighborhood $W_p \subseteq T_pM$ such that the restriction of the Riemannian exponential map $\exp_p \colonequals \exp|_{W_p} \colon W_p \rightarrow M$ induces a diffeomorphism onto an (open) $p$-neighborhood in $M$.
We will show that the manifold charts $\exp_p$ induce suitable submanifold charts with rough boundary.
Clearly if $p \in C^\circ$ we can just shrink $W_p$ to obtain such a submanifold chart. For $q \in \partial C$ (the boundary of $C$) we have to work harder.
Define for $q \in \partial C$ the set
$$V_q \colonequals \{\lambda w \in T_q M \mid 0< \lambda < \infty, w\in W_q \text{ and } \exp_q (w) \in C^\circ\}.$$
We observe that $V_q = \bigcup_{0 < \lambda < \infty }\lambda \cdot \exp_q^{-1} (C^\circ)$ is an open subset of $T_q M$.
Now we exploit the geometric properties of strongly convex sets following \cite{MR0226542}, where these sets are called "konvex".\footnote{Loc.cit.\ assumes that $M$ is a complete Riemannian manifold.
We do not assume completeness as the parts of \cite{MR0226542} needed here do not rely on the completeness of $M$. In fact, the geodesic segments needed in the proofs exist since we are working in the strongly convex set $C$.}
As $C$ is strongly convex, also the interior $C^\circ$ is strongly convex \cite[Korollar 4.5.1]{MR0226542}, whence it is also "schwach konvex" (weakly convex) in the terminology of loc.cit..
Now regularity of $C$ implies that $q \in \partial C = \partial C^\circ$.
We can thus copy the argument in the proof of \cite[4.9.2]{MR0226542} verbatim (note that the geodesics occuring there are only needed locally in a small neighborhood around $q$!) to establish that $V_q$ is a convex cone in $T_qM$ whose tip is $0_q$.
In particular $\overline{V}_q$ is a convex cone, i.e.\ a closed subset with dense interior that is (locally) convex.
\paragraph{Claim:} $\exp_q (W_q \cap \overline{V}_q) = C \cap \exp_q (W_q)$.
If this is true then $\exp_q$ restricts to a submanifold chart (with rough boundary) for $C$ around $q$ as $W_q \cap \overline{V}_q$ has dense interior (namely $W_q \cap V_q$) and is locally convex as an intersection of two (locally) convex sets in $T_pM$.
We conclude that $C$ is a closed embedded submanifold with rough boundary of the Riemannian manifold $M$.
\paragraph{Proof of the claim:}
Let us first observe that $\exp_q (V_q \cap W_q) = \exp_q (W_q) \cap C^\circ$ by construction of $V_q$ (and the diffeomorphism property of $\exp_q$).
Now let $x \in W_q \cap \overline{V}_q$. Since the interior of this subset is the dense set $W_q \cap V_q$, we can choose and fix a sequence $(x_n)_{n\in \ensuremath{\mathbb{N}}} \subseteq V_q \cap W_q$ with $\lim_{n} x_n = x$. By continuity of $\exp_q$ and since $\exp_q (x_n) \in C^\circ \cap \exp_q (W_q)$ we have
$\exp_q (x) \in \exp_q (W_q) \cap \overline{C^\circ} = \exp_q (W_q) \cap C$. Conversely, if $p_n \in C^\circ \cap \exp_q (W_q)$ is a sequence converging to $p \in C \cap \exp_q (W_q)$ we use continuity of $\exp_q^{-1}$ to see that $p \in \exp_q (W_q \cap \overline{V}_q)$.
Summing up, the claim follows.
\end{proof}
\begin{remark}\label{rem: boundary: strong:conv}
Actually, strongly convex subsets have very nice boundary behaviour. For example, it is known that they have Lipschitz boundary (cf.\ e.g.\ \cite{MR638797}).
However, we are not aware of another source in the literature where submanifold charts of the above kind are explicitely constructed.
In light of Example \ref{ex: cusp} this implies that a strongly convex regular closed subset satisfies the cusp condition.
\end{remark}
Encouraged by these results, we shall now prove that every closed submanifold with rough boundary satsifies the cusp condition, Definition~\ref{def:cusp_condition}.
\begin{proposition}\label{prop: rsub:nnf}
Let $M$ be a Riemannian manifold and $C$ be a closed submanifold with rough boundary. Then $C$ satisfies the cusp condition.
\end{proposition}
\begin{proof}
Following Remark \ref{rem: submfd:rb} we already know that $C$ is a regular closed set. We first have to check the no narrow fjord condition. To this end, fix $x\in C$ together with a manifold chart $\varphi \colon U \rightarrow \varphi (U) \subseteq \ensuremath{\mathbb{R}}^n$ around $x$ and $\varphi (U\cap C) = \varphi(U)\cap R$ for a suitable regular locally convex set $R$.
Shrinking the chart if necessary, we can assume that it is bi-Lipschitz with respect to the geodesic length metric and the Euclidean metric (cf.\ Remark~\ref{rem: standard:arguments}). Then arguing as in Lemma~\ref{lem: trans:nnf}, it suffices to check the no narrow fjords condition for $\varphi (x)$ as an element of the locally convex subset $\varphi(U) \cap R \subseteq \ensuremath{\mathbb{R}}^n$ (note that this set need not be closed whence it does not make sense to say that it has no narrow fjords!)
Now since $U \cap C$ is the intersection of an open and a closed subset of a locally compact space, it is locally compact, whence $\varphi(U\cap C) = \varphi (U) \cap R$ is locally compact.
We can thus choose a compact neighborhood $K$ of $\varphi(x)$ in $\varphi (U) \cap R$.
Then local convexity of $\varphi (U) \cap R$ allows us to choose a neighborhood $\varphi(x) \in W \subseteq K$ which is convex.
Denote now by $K_{\varphi(x)} = \overline{W} \subseteq K$ the closure of $W$.
It is again a convex set and compact by construction.
Now an easy but tedious computation involving metric estimates, convexity of $K_{\varphi(x)}$ and the boundary behavior observation $\partial (\varphi (U) \cap R) \cap K_{\varphi(x)} \subseteq \partial K_{\varphi(x)}$ yields the constants needed to verify the no narrow fjords condition for $K_{\varphi (x)}$.
Alternatively, observe that convex sets have Lipschitz boundary (cf.\ again \cite{MR638797}) which implies that $K_{\varphi (x)}$ has no narrow fjords.
To check the outward polynomial cusp condition, we use again the compact, convex neighborhood $K_{\varphi(x)}$ of $\varphi(x)$ in $\varphi(C\cap U)$, and a diffeomorphism $\varphi(U) \simeq \ensuremath{\mathbb{R}}^n$ (which is locally bi-Lipschitz). Since the image of the compact convex neighborhood with Lipschitz boundary yields the required estimates for the polynomial outward cusp condition, the diffeomorphism transfers them from $\varphi(U)$ to $\ensuremath{\mathbb{R}}^n$ (and so satisfying Frerick's version of the definition). Invoking Lemma \ref{lem:transfer_cusp_condition}, we see that $ C\cap U $ satisfies the condition.
Hence $C$ has at worst polynomial outward cusps, completing the proof.
\end{proof}
For regular closed subsets which are at the same time submanifolds with rough boundary we prove now that for every vector bundle the space of sections from Definition~\ref{defn: sectsp:cl} can canonically be identified with the sections of the corresponding pullback bundle over the submanifold with rough boundary.
\begin{proposition}\label{prop: quot:secsmfd}
Let $\pi_E \colon E \rightarrow M$ be a rank $k$ vector bundle over a manifold without boundary.
Let $C \subseteq M$ be a closed submanifold with rough boundary. Assume in addition that
\begin{enumerate}
\item $C$ is a submanifold with corners, or
\item $C$ is compact.
\end{enumerate}
Then the pullback $p_C \colon \Gamma_c (M,E) \rightarrow \Gamma_c (C, \iota_C^* E),\ \sigma \mapsto \sigma \circ \iota_C$ by the inclusion $\iota_C \colon C \rightarrow M$ is a linear quotient map.
\end{proposition}
\begin{proof}
Observe that since $C \subseteq M$ is closed, the inclusion $\iota_C$ is a proper mapping (i.e.\ preimages of compact sets are compact).
Thus by our definition of submanifold with rough boundary (or with corners), every section $\sigma \in \Gamma_c (M,E)$ induces a smooth pullback section $\sigma \circ \iota_C \in \Gamma_c (C,\iota_C^* E)$ and the pullback map $p_C$ makes sense and is clearly linear.
Further, taking canonical identifications, a section $\tau \in \Gamma_c (C,\iota_C^* E)$ clearly coincides with a smooth mapping $\tilde{\tau}\colon C \rightarrow E$ with compact support such that $\pi_E \circ \tilde{\tau} = C$ (here we use $\tilde{\tau}$ to mark the difference in the codomain).
Now Proposition \ref{prop: rsub:nnf} implies that $C$ has no narrow fjords, whence by Proposition \ref{prop: Whit:mfd} there is $\hat{\tau} \in \Gamma_c (M,E)$ which restricts to $\tilde{\tau}$ on $C$. We deduce that $p_C (\hat{\tau}) = \tau$, whence $p_C$ is surjective.
To establish continuity we have to distinguish the two cases, due to the difference in the function space topologies.
\begin{enumerate}
\item ($C$ is a submanifold with corners) The map $\iota_C^* \colon C^\infty_{\text{fS}} (M,E) \rightarrow C^\infty_{\text{fS}} (C,E), f \mapsto f\circ \iota_C$ is continuous by \cite[Theorem 7.3]{michor} (and even smooth) as $\iota_C$ is proper.
Consider the linear subspace
$\mathcal{D}_{\iota_C} (C,E) \colonequals \{g \in C^\infty (C,E) \mid \pi_E \circ g = \iota_C, g \equiv 0 \text{ off some compact set in } C\} $ of $C^\infty_{fS} (C,E)$.
It is easy to see that $\iota_C^*$ restricts to a continuous mapping
$I \colon \Gamma_c (M,E) \rightarrow \mathcal{D}_{\iota_C} (C,E).$
However, due to the definition of the pullback bundle (see \cite[1.18 and 1.19]{michor}), the space $\mathcal{D}_{\iota_C} (C,E)$ is isomorphic as a linear and topological space to $\Gamma_c (C,\iota_C^* (E))$, composing $I$ with this isomorphism we obtain $p_C$ which is thus continuous.
\item ($C$ is compact) Since $M$ is $\sigma$-compact, we can choose and fix a locally finite (countable) atlas of bundle trivialisations $(W_i,\kappa_i)_{i\in \ensuremath{\mathbb{N}}}$ for $E$ such that every $W_i$ is relatively compact. Since $C$ is compact, only finitely many $W_i$ intersect $C$. After reordering, we may assume that $W_i \cap C \neq \emptyset$ iff $1\leq i \leq n$ for some $n\in \ensuremath{\mathbb{N}}$.
Observe that since $W_i$ is an open subset of $M$ and $C$ is an embedded submanifold (with rough boundary), for $i\leq n$, the set $\iota_C^{-1} (W_i) = W_i \cap C$ is a submanifold with rough boundary of $W_i$ (as submanifold charts $\varphi \colon U \rightarrow \varphi (U)$ are bijections, we find $\varphi(W_i \cap U \cap C) = \varphi (W_i \cap U) \cap R$).
Hence Proposition \ref{prop: fcomp:cont} implies that $p_i \colon C^\infty_{\text{co}} (W_i,\ensuremath{\mathbb{R}}^k) \rightarrow C^\infty_{\text{co}} (W_i\cap C, \ensuremath{\mathbb{R}}^k), f \mapsto f|_{C \cap W_i}$ is continuous linear for $i \leq n$.
Thus we obtain a continuous (linear) map $q \colon \bigoplus_{i \in \ensuremath{\mathbb{N}}} C^\infty_{\text{co}} (W_i, \ensuremath{\mathbb{R}}^d) \rightarrow \bigoplus_{i \leq n} C^\infty_{\text{co}} (W_i \cap C, \ensuremath{\mathbb{R}}^d), (f_i)_i \mapsto (p_i(f_i))_{1\leq i \leq n}$.
Following Lemma \ref{lem: top:sect} and \ref{setup: sect:loc} we obtain a commutative diagram
\begin{displaymath}
\begin{xy}
\xymatrix{
\Gamma_c (M,E) \ar[r]^-\rho \ar[d]^{p_C} & \displaystyle\bigoplus_{i\in \ensuremath{\mathbb{N}}} \Gamma (W_i, E|_{W_i}) \ar[r]^\cong & \displaystyle \bigoplus_{i\in \ensuremath{\mathbb{N}}} C^\infty_{\text{co}} (W_i,\ensuremath{\mathbb{R}}^d) \ar[d]^{q} \\
\Gamma (C,E) \ar[r]^-\rho & \displaystyle\bigoplus_{1\leq i\leq n} \Gamma (W_i \cap C, (\iota_C^* E)|_{W_i \cap C}) \ar[r]^\cong & \displaystyle \bigoplus_{1\leq i\leq n} C^\infty (W_i\cap C, \ensuremath{\mathbb{R}}^d)
}
\end{xy}
\end{displaymath}
where the the image the $\rho$ are topological embeddings with closed images. Hence $p_C$ is continuous in this case.
\end{enumerate}
Finally, let us establish that $p_C$ is a quotient map, i.e.\ $p_C$ is open.
To this end, recall from Appendix \ref{App: topo:funcsp} that $\Gamma_c (M,E)$ is an (LF)-space, i.e. webbed and ultrabornological.
Further, if $C$ is a manifold with corners, also $\Gamma_c (C,\iota_C^* E)$ is an (LF) space.
If $C$ is compact and a manifold with rough boundary, then $\Gamma (C,\iota_C^*E)$ is even a Fr\'{e}chet space by \ref{setup: sect:loc}.
In both cases, $p_C$ is open by the open mapping theorem \cite[24.30]{MR1483073}.
\end{proof}
\begin{proposition}\label{prop: ident:sectsp}
Let $M$ be a manifold and $E \rightarrow M$ be a rank $k$ vector bundle.
If $C$ is closed submanifold with rough boundary which satisfies the assumptions of Proposition \ref{prop: quot:secsmfd}, then $\Gamma_c (C,E) = \Gamma_c (C,\iota_C^*E)$ as locally convex vector spaces.
\end{proposition}
\begin{proof}
Since $C$ is an embedded submanifold, a section in the pullback bundle is smooth if and only if it is a smooth as a mapping $C \rightarrow E$. Thus as sets we canonically identify $\Gamma_c (C,E) = \Gamma_c (C,\iota_C^*E)$.
Now Proposition \ref{prop: quot:secsmfd} and Proposition \ref{prop: Whit:mfd} yield a commutative diagram
\begin{displaymath}
\begin{xy}
\xymatrix{
& \Gamma_c (M,E)\ar[ld]_{\res_C} \ar[rd]^{p_C} &\\
\Gamma_c (C,E) \ar[rr]^\id & &\ar@<1ex>[ll]^\id \Gamma_c (C,\iota_C^* E)}
\end{xy}
\end{displaymath}
where the diagonal arrows are quotient mappings.
\end{proof}
\begin{setup}[Short reminder: Manifolds of mappings for non-compact source manifolds]
Assume that $C$ is a (sub-)manifold with corners which is possibly non-compact.
Then the function space $C^\infty (C,N)$ can be endowed with an infinite-dimensional manifold structure which constructed similarly to the construction outlined in Section \ref{sect: mfdmap:rb}: One endows $C^\infty (C,N)$ withe the $\mathcal{FD}$-topology described in \cite{michor} (a Whitney type topology).
In the boundaryless case \cite{HS17}, this topology is also called the fine very strong topology and therefore we denote by $C^\infty_{\text{fS}} (C,N)$ the function space with the fine very strong (=$\mathcal{FD}$-) topology.
Choosing a local addition on $N$, the construction of manifold charts is completely analogous to the construction outlined in \ref{setup: can:charts} with the notable exception that one has to restrict to $\Gamma_c (C,f^*TN)$ and one has to intersect $U_f$ with
$$\{g \in C^\infty (C,N) \mid \exists K \text{ compact, such that } \forall x \in C\setminus K\ f(x)=g(x)\}.$$
The rest of the construction is completely analogous to the one outlined in \ref{setup: can:charts} and yields the same structure as in Theorem \ref{thm: mm:rb} if $C$ is compact (note that we will thus also write $C^\infty_{fS} (C,N) = C^\infty_{\text{co}} (C,N)$ if $C$ is compact with rough boundary).
\end{setup}
We are now ready to prove Theorem B which we restate here for the reader's convenience. Recall that $M$ is equipped with a Riemannian metric.
\begin{theorem}\label{thm: Bproof}
For $C\subset M$ a submanifold with corners, or compact and a submanifold with rough boundary, then the restriction map $\res_C^M \colon C^\infty_{\text{fS}} (M,N) \rightarrow C^\infty_{\text{fS}} (C,N)$ is a submersion of locally convex manifolds.
\end{theorem}
\begin{proof}
Let $\iota_C \colon C\rightarrow M$ be the canonical inclusion, which is smooth as $C$ is an embedded submanifold. Hence $\res_C^M=\iota_C^*$ is smooth by \cite[Theorem 7.3]{michor} (if $C$ is a submanifold with corners).
Since the compact-open $C^\infty$ topology is coarser than the fine very strong topology (cf.\ \cite{HS17}), Proposition \ref{prop: fcomp:cont} implies that $\res_C^M$ is continuous if $C$ is compact and a submanifold with rough boundary (note that $M$ might be non-compact).
Hence to establish smoothness and the submersion property, it suffices to construct submersion charts for $\res_C^M$.
Let now $F\in C^\infty (M,N)$ and $f \colonequals \res_C^M (F)$. Then we use that $C$ satisfies the cusp condition and consider the canonical charts $(U_F,\varphi_F)$ and $(U_f,\varphi_f)$ (cf.\ \ref{setup: can:charts}) to obtain a commutative diagram
\begin{displaymath}
\begin{xy}
\xymatrix{
C^\infty_{\text{fS}} (M,N) \supseteq U_F \ar[d]^{\res_C^M}\ar[r]^{\varphi_F} & \Gamma_c (M,F^*E) \ar[rr]^-{\cong}_-{\text{Corollary }\ref{splitting consequences}} \ar[d]^{\res_C}& & \mathcal{I}_c (C,M) \oplus \Gamma_c (C,F^*TN) \ar[d]^{\pr_2} \\
C^\infty_{\text{fS}} (C,N) \supseteq U_f \ar[r]^{\varphi_f} & \Gamma_c (C,f^*TN) \ar[rr]^-\cong_-{\text{Proposition } \ref{prop: ident:sectsp}} & & \Gamma_c (C,F^*TN)
}
\end{xy}
\end{displaymath}
Observe that $\res_C^M$ is a smooth submersion as the canonical charts conjugate it to a projection onto a complemented closed subspace, which is continuous linear.
\end{proof}
Note that Corollary D from the introduction follows from the results in this section and Corollary C in the wash as Lemma \ref{lem: sconv:rbsmfd} asserts that strongly convex regular closed subsets of Riemannian manifolds are submanifolds with rough boundary.
\ifdefined\MathZ
\else
\section*{Acknowledgements}
\thankyous
\fi
|
\section{Introduction}
The interactions of open-shell atoms, molecules, and ions are of great significance to many areas of molecular physics and chemistry.
As a matter of fact, all chemical reactions at some stage have to experience bond breaking and rearrangement and to describe such processes one has to deal with interacting fragments of reactants or products.
Many electronically excited states of molecules have open electronic shells, so in studies of the interactions of such molecules with background gas or solvent similar problems arise.
Interacting radicals are of key importance in atmosphere, where they undergo chain reactions, form metastable states and stable complexes \cite{Aloisio:00,Galano:10}.
Collisions of radicals (such as OH, CH$_2$, CN, and many more) with hydrogen (H and H$_2$) are a subject of study for astrochemistry \cite{Roueff:13} as they provide information about the conditions which molecules experience in the interstellar clouds.
Stabilized organic radicals are relevant in organic and materials chemistry~\cite{Ratera:12}, for example, as building blocks of molecular magnets.
One should also mention the high relevance of interactions of radicals in cold chemistry: open-shell molecules can be manipulated with magnetic fields to control their collisions at low temperatures, giving unique information on the interaction potential \cite{Kirste:12,Lavert-Ofir:14,Chefdeville:13}.
In modelling the dynamics of weakly interacting systems, accurate potential energy surfaces are crucial. In case of open-shell systems, despite large progress in recent years, electronic structure calculations are still far from routine or robust.
Although the coupled cluster (CC) method, most widely used in studying molecular interactions (in particular in its variant with single, double, and perturbative triple excitations, CCSD(T)), has been extended to the high-spin open-shell case a while ago by Knowles et al.~\cite{Knowles:93} and Watts et al.~\cite{Watts:93}, there are many examples where the coupled cluster method fails and a multireference approach is necessary.
In particular, one of the most important cases for which the single-reference CC method does not work is the dissociation of a low-spin system into high-spin subsystems: for example, a breakdown of a singlet, stable molecule into two doublet molecules.
However, despite many attempts and progress in theory since the introduction of the multireference coupled-cluster (MRCC) ansatz by Jeziorski and Monkhorst~\cite{Jeziorski:81}, the proposed variations of MRCC are far from being black-box methods (see Refs.~\cite{Bartlett:07} and \cite{Jeziorski:10} for a review).
At present, to account for both static and dynamical correlation effects in molecular interactions for systems exhibiting a non-unique electronic configuration, multireference configuration interaction~ \cite{Werner:88}~ (MRCI) or multireference perturbation theories \cite{Andersson:90,Finley:98,Angeli:02} are often applied.
In the case of the MRCI method, it is possible to reproduce the short-range part of the potential well when size-consistency corrections are applied.
Such {\em a posteriori} corrections are not exact and non-unique, and the remaining size-consistency error might be on the order of the interaction energy.
In case of the perturbation methods, the same problem also arises in general; moreover, some of these methods suffer from the problem of intruder states \cite{Camacho:10}.
One should also mention that all multireference methods rely on a proper selection of the active space, which is not necessarily straightforward, and that the complexity of calculations grows exponentially with the size of the active space. This is a particularly severe problem for interacting open-shell molecules as the active space required for the complex, typically the union of the monomers' active spaces, might be intractable.
An interesting alternative for open-shell systems is the family of the so-called spin-flip methods in which it is possible to reach arbitrary, multireference low-spin states of a given system starting from a single-reference high-spin state~\cite{Krylov:01}.
The spin-flip equation of motion approach has been implemented within a variety of different theories, in particular, Hartree-Fock~\cite{Krylov:01}, density functional theory~\cite{Shao:03}, configuration interaction doubles~\cite{Krylov:01} and finally, using the CC framework~\cite{Levchenko:04,Krylov:06}. The theory has been generalized to multiple spin flips \cite{Zimmerman:12}, however, the resulting calculations can be quite demanding, and it is often advantageous to describe the exchange splittings between all multiplets by a single exchange coupling constant $J_{AB}$ within the Heisenberg spin Hamiltonian. In order to determine $J_{AB}$, it is sufficient to perform a single spin-flip calculation starting from the high-spin, single-determinant reference state \cite{Mayhall:15}.
Recent progress notwithstanding, the current arsenal of methods which can be used for multireference open-shell complexes is limited, and it is desirable to explore alternative approaches which provide good insight and reliable interaction potentials.
The closed-shell symmetry-adapted perturbation theory (SAPT)~\cite{Jeziorski:94,Hohenstein:12} has been widely recognized as a highly useful tool for calculations of interaction energies with an accuracy comparable to CCSD(T), but also as a robust analysis tool which provides insightful physical interpretation of the nature of the interaction in terms of its electrostatic, induction, dispersion, and exchange components.
For general single-reference open-shell systems, \.Zuchowski et al. \cite{Zuchowski:08} and Hapka et al. \cite{Hapka:12} introduced the symmetry-adapted perturbation theory based on, respectively, restricted and unrestricted Slater determinants (with Hartree-Fock or Kohn-Sham orbitals).
Present implementations of SAPT, however, are valid only for the single-reference high-spin case of the dimer in which the spin quantum number of the complex is equal to the sum of spin quantum numbers of monomers ($S = S_{\rm A} + S_{\rm B}$).
The interaction of two open-shell species with the total spin of the dimer smaller than the sum of spin quantum numbers of the monomers still poses a challenge.
This is the case since low-spin configurations are typical examples of multireference systems. The knowledge of the so-called exchange splitting, which is defined as
the difference between the highest and lowest possible spin state of a given dimer, is very important in interactions of small open-shell molecules (eg. in the O$_2$~$^3\Sigma^-_g$ dimer~\cite{Aquilanti:99,Bartolomei:08,Bartolomei:10})
as well as for interactions involving stable organic radicals such as phenalenyl~\cite{Cui:14} or unsaturated metal sites within metal organic frameworks in the O$_2$ adsorption process \cite{Parkes:15}.
The new approaches of calculating low-spin potential energy surfaces could be applied to model fragments of interaction potentials in reaction dynamics, for instance near the channels which correspond to dissociation into two radicals, or near the bond-breaking geometry of the reactive complex.
Spin splitting also plays an important role in cold physics, since it is the term that strongly couples hyperfine states of alkali-metal atoms and drives Feshbach resonances \cite{Bartenstein:05,Chin:04}.
It should be stressed that in the SAPT approach exchange splitting does indeed come exclusively from exchange corrections: the electrostatic, induction, and dispersion SAPT terms are identical for all asymptotically degenerate states of different spin multiplicity.
This paper is the first step toward the development of SAPT for low-spin dimer states. Here, we derive the first-order exchange energy in terms of spin-restricted orbitals which preserve the correct value of the squared-spin operator $\widehat{S}^2$.
The theory introduced in this paper is beneficial for several reasons: (i) as usually in SAPT, the interaction energy components are calculated {\em directly}, which contrasts with supermolecular calculations that always involve subtraction; (ii) there is no need for the active space selection for the complex: in fact, no multireference calculation is necessary (similar to the spin-flip electronic structure approaches); (iii) it is possible to calculate the energy in the monomer-centered basis set, and the overall scaling of the method (once transformed molecular-orbital (MO) integrals are available) is just $o^4$ where $o$ is the number of occupied orbitals;
and (iv) the low-spin exchange energy component, together with other SAPT corrections, can provide an insight into the nature of the interaction in a radical-radical system.
The formalism presented here is valid for any spin-restricted Slater determinants. At this level, similarly to the high-spin open-shell theory \cite{Zuchowski:08}, intramonomer correlation effects are not included except for the possible use of Kohn-Sham orbitals \cite{Williams:01}. As we will show in Sec.~\ref{theory}, the
first-order SAPT term responsible for the multiplet splittings involves matrix elements between the
zeroth-order wavefunction and a function where one of the unpaired spins on one monomer has been lowered and
one of the unpaired spins on the other monomer has been raised, in other words, a function which was
obtained from the zeroth-order SAPT wavefunction via an intermolecular spin flip. Therefore, we will
refer to our new approach as spin-flip SAPT (SF-SAPT), borrowing the nomenclature from spin-flip
electronic structure theories \cite{Krylov:01}. However, while in the latter theories the $z$ component of the total
spin of the system (the $M_S$ quantum number) is changed upon the spin flip, in SF-SAPT only the $M_{S_A}$ and $M_{S_B}$ numbers for individual molecules
are changed: the overall $M_S$ value is conserved.
To date, there are only a few applications of perturbation theory to low-spin complexes.
In 1984, Wormer and van der Avoird used the Heitler-London formula with appropriately spin-adapted linear combinations of Slater determinants (in the specific case of 4 active electrons) to predict the splittings between different multiplets of the O$_2\cdots$O$_2$ system \cite{Wormer:84}.
Extensive studies of SAPT performance up to very high order of perturbation theory exist for few-electron systems in which monomers can be represented by exact (in a given Gaussian basis) wavefunctions.
\'Cwiok et al. \cite{Cwiok:92} studied the singlet and triplet states of the H$\cdots$H interaction, while Patkowski et al. reported the convergence of various SAPT variants for the lowest singlet and triplet states of the Li$\cdots$H complex \cite{Patkowski:01, Patkowski:02}.
In the past few years, the exchange splitting of interaction energy has been studied for the H$_2^+$ system by Gniewek and Jeziorski~\cite{Gniewek:14,Gniewek:15,Gniewek:16}. Although this is a single-electron system, its two lowest states are asymptotically degenerate and correspond to symmetric ($g$) and antisymmetric ($u$) combinations of atomic orbitals.
The difference between $u$ and $g$ states results from the resonance tunneling of the electron
between the nuclei, in a similar way
as the difference between triplet and singlet states in the interaction of two doublets.
Gniewek and Jeziorski used perturbation theory to study the asymptotic expansion of exchange energy in H$_2^+$ and discussed a new approach to exchange energy via a variational volume-integral formula.
The plan of this paper is as follows: in the next section, we derive the arbitrary-spin first-order exchange energy formula for two monomers described by their spin-restricted determinants and show the connection with the Heisenberg model of the scalar interaction of two spins.
In Section III we give the most important details regarding the implementation of the theory, and in Section IV we report the test calculations for several important open-shell$\cdots$open-shell systems. In the last section of the paper, we summarize our results and discuss prospects for the future.
\section{Theory}
\label{theory}
Similar to the existing closed-shell and high-spin open-shell SAPT
approaches, the permutational symmetry adaptation, leading to the
exchange corrections, is performed in the simplest manner, within the
symmetrized Rayleigh-Schr\"{o}dinger (SRS)
formalism \cite{Jeziorski:78,Jeziorski:94}. Specifically, the
SRS wavefunction corrections are computed from the standard
Rayleigh-Schr\"{o}dinger (RS) perturbation expansion without any symmetry adaptation:
only later, the energy corrections are computed from a formula that
involves a symmetry projector. In the spin-free formalism used, for example,
in Ref.~\cite{Patkowski:01}, this symmetry operator projects onto a
particular irreducible representation of the permutation group
$S_{N_A+N_B}$ ($N_X$ is the number of electrons of molecule $X$): the choice
of this representation is determined by the spin multiplicity. In the more
conventional, spin formalism employed here (and used in the high-spin case
so far), this operator has to be replaced by the $(N_A+N_B)$-electron antisymmetrizer
and an appropriate spin projection.
In the already existing high-spin open-shell
SAPT theories the monomer wavefunctions $\Psi_A$ and $\Psi_B$
as well as the zeroth-order dimer
wavefunction $\Psi_0=\Psi_A\Psi_B$ are pure spin functions
so the spin projection is not needed. This is only the case if all unpaired
electrons on both monomers have the same spin, that is, when
$M_{S}=\pm (S_A+S_B)$ so that no contamination by lower-spin states is
possible. Indeed, the restricted open-shell Hartree-Fock and Kohn-Sham
(ROHF/ROKS) SAPT exchange corrections
have been developed \cite{Zuchowski:08} under the assumption that all unpaired
spins point the same way.
For the low-spin case, it is known \cite{Patkowski:02,Jeziorski:10} that
the antisymmetrizer ${\cal A}$ in the SRS energy expression has to be accompanied
by an operator $P_{SM_S}$ that projects a dimer function onto
a subspace corresponding to a particular value of the total spin $S$ (with an
appropriate $M_S$). In this way,
the exchange corrections, different for different dimer spin
states, are obtained by expanding the energy expression
\begin{equation}\label{estart}
E_{\rm int}=\frac{\langle \Psi_A\Psi_B|V{\cal A}P_{SM_S}|\Phi_{AB}\rangle}{%
\langle \Psi_A\Psi_B|{\cal A}P_{SM_S}|\Phi_{AB}\rangle},
\end{equation}
in powers of the intermolecular interaction operator $V$ and separating the
polarization and exchange effects in each order. In Eq.~(\ref{estart}),
${\cal A}$ is the $(N_A+N_B)$-electron antisymmetrizer and
$\Phi_{AB}$ is the polarization (RS) expansion of the wave function
for the complex \cite{Jeziorski:94}. We will
assume that $\Psi_A$ and $\Psi_B$ are ground-state high-spin open-shell determinants.
The occupied orbitals in $\Psi_A$ are $\chi_{i_1}$, $\chi_{i_2}$, \ldots,
$\chi_{i_{k_A}}$ (inactive, doubly occupied) and $\chi_{a_1}$, $\chi_{a_2}$,
\ldots, $\chi_{a_{2S_A}}$ (active, occupied by $\alpha$ electrons only).
The occupied orbitals in $\Psi_B$ are $\chi_{j_1}$, $\chi_{j_2}$, \ldots,
$\chi_{j_{k_B}}$ (inactive, doubly occupied) and $\chi_{b_1}$, $\chi_{b_2}$,
\ldots, $\chi_{b_{2S_B}}$ (active, occupied by $\beta$ electrons only).
The spin projector $P_{SM_S}$ commutes with $V$ and the zeroth-order Hamiltonian
$H_0=H_A+H_B$ as the latter operators are spin-independent. Moreover, the operator
$\hat{S}^2$, and thus also $P_{SM_S}$, is independent of the ordering of electrons and
hence commutes with any electron permutations or their linear combinations
including ${\cal A}$. As a result, we can rewrite Eq.~(\ref{estart}) in an
alternative form where the spin projection acts in the bra instead of the ket:
\begin{equation}\label{estart_alt}
E_{\rm int}=\frac{\langle P_{SM_S}\Psi_A\Psi_B|V{\cal A}|\Phi'_{AB}\rangle}{%
\langle P_{SM_S}\Psi_A\Psi_B|{\cal A}|\Phi'_{AB}\rangle}.
\end{equation}
While Eqs.~(\ref{estart}) and (\ref{estart_alt}) are fully equivalent, they have
quite different interpretations.
Eq.~(\ref{estart}) corresponds to a standard RS treatment where the wavefunction
corrections in the expansion of $\Phi_{AB}$ are the same for all asymptotically
degenerate multiplets: the spin
projection enters later, at the calculation of SRS energy corrections.
Eq.~(\ref{estart_alt}) corresponds to a perturbation expansion initiated from a
spin-adapted zeroth-order function $P_{SM_S}\Psi_A\Psi_B$, with all resulting
perturbation corrections (in the expansion of $\Phi'_{AB}$) also being of pure spin.
Importantly, both approaches
allow the use of standard nondegenerate perturbation theory
(unless degeneracy arises from the orbital part of the monomer wavefunctions,
which is outside the scope of this work), albeit for quite different reasons.
For Eq.~(\ref{estart_alt}), the projector $P_{SM_S}$ reduces the zeroth-order
space of $(2S_A+1)(2S_B+1)$ asymptotically degenerate
product functions corresponding to a pair of
interacting multiplets to a single spin-adapted combination with the requested
$(S,M_S)$. For Eq.~(\ref{estart}), any of the $(2S_A+1)(2S_B+1)$ product
functions, including $\Psi_A\Psi_B$, is a valid zeroth-order state for nondegenerate RS perturbation
theory as neither $V$ nor $H_0$ mix states with different $M_{S_A}$ or $M_{S_B}$.
We choose Eq.~(\ref{estart}) as the starting point for the derivation below.
We now introduce the single exchange approximation (often called the $S^2$ approximation
as it neglects terms higher than quadratic in the intermolecular overlap integrals):
\begin{equation}\label{singex}
{\cal A}\sim (1+{\cal P}){\cal A}_A {\cal A}_B,
\end{equation}
where the single-exchange operator ${\cal P}$ is given by
\begin{equation}
{\cal P}=-\sum_{i\in A}\sum_{j\in B} {\cal P}_{ij}
\end{equation}
(for non-approximated alternatives in the closed-shell
case, see Refs.~\cite{Jeziorski:76,Schaffer:12,Schaffer:13}).
We now multiply both sides of Eq. (\ref{estart}) by the denominator,
keeping only the terms of first order in $V$ (thus, replacing
$\Phi_{AB}$ by its zeroth-order term $\Psi_A\Psi_B$), and make
the approximation (\ref{singex}). Introducing the shorthand
notation $\langle X\rangle\equiv\langle \Psi_A\Psi_B|X|\Psi_A\Psi_B\rangle$, we obtain (note that ${\cal A}$,
${\cal A}_A$, and ${\cal A}_B$ all commute with $P_{SM_S}$)
\begin{equation}\label{est2}
\left(\langle V\rangle+E^{(10)}_{\rm exch}\right)
\left(\langle P_{SM_S}\rangle
+\langle {\cal P}P_{SM_S}\rangle\right)
=\langle VP_{SM_S}\rangle
+\langle V{\cal P}P_{SM_S}\rangle
\end{equation}
Note that we have separated the electrostatic contribution
$E^{(10)}_{\rm elst}=\langle V\rangle$ from the exchange one, and
we introduced the zero into the $E^{(10)}$ superscript to remind the
reader that no intramolecular correlation effects are included.
Before we proceed, we need to understand a bit better the action of the
operator $P_{SM_S}$ on the product $\Psi_A\Psi_B$. The RHF determinants $\Psi_A$
and $\Psi_B$ are pure spin functions corresponding to the quantum numbers
$(S_A,M_{S_A}=S_A)$ and $(S_B,M_{S_B}=-S_B)$, respectively.
In the zeroth-order space spanned by $(2S_A+1)(2S_B+1)$ product functions that correspond
to total spin $S_A$ for molecule A and total spin $S_B$ for molecule B (and any
combination of $M_{S_A},M_{S_B}$), there exists exactly one function $\Psi_{S,M_S}$
with a total spin $S\in\{|S_A-S_B|,\ldots,S_A+S_B\}$ and its projection
$M_S\in\{-S,\ldots,S\}$. This function is a linear combination of only those
product functions that correspond to $M_{S_A}+M_{S_B}=M_S$. Therefore, the
projection $P_{SM_S}\Psi_A\Psi_B$ picks, up to normalization, this very
function $\Psi_{S,M_S}$ out of the zeroth-order space.
In other words, $P_{SM_S}\Psi_A\Psi_B$
produces a linear combination of products of functions of A and B with the
same values $S_A,S_B$ but different $M_{S_A},M_{S_B}$ (however, the latter two
numbers add up to the same total $M_S$). The coefficients in this linear
combination are the Clebsch-Gordan coefficients
$\langle SM_S | S_A M_{S_A} S_B M_{S_B}\rangle$ --- we will denote the
coefficients
$\langle S\: (S_A-S_B) | S_A\: S_A\: S_B\: -S_B\rangle$,
$\langle S\: (S_A-S_B) | S_A\: (S_A-1)\: S_B\: (-S_B+1)\rangle$,
and $\langle S\: (S_A-S_B) | S_A\: (S_A-2)\: S_B\: (-S_B+2)\rangle$ by
$c_0$, $c_1$, and $c_2$, respectively.
Specifically, the spin
projection can be expressed as follows:
\begin{equation}\label{pspsi0}
P_{SM_S}\Psi_A\Psi_B=c_0 \Psi_A\Psi_B
+c_1
\Psi_A^{\downarrow}\Psi_B^{\uparrow}
+c_2
\Psi_A^{\downarrow\downarrow}\Psi_B^{\uparrow\uparrow}+\ldots
\end{equation}
where, for example, $\Psi_A^{\downarrow}$ is a normalized wavefunction
(linear combination of determinants) corresponding to the spin quantum
numbers $(S_A,S_A-1)$. Up to a constant, this function can be obtained from
$\Psi_A$ by the action of the spin-lowering operator $\hat{S}_{-}$.
Eq. (\ref{pspsi0}) implies that
\begin{equation}
\langle P_{SM_S}\rangle=c_0
\end{equation}
\begin{equation}
\langle VP_{SM_S}\rangle=c_0 \langle V\rangle
\end{equation}
Therefore, Eq. (\ref{est2}) becomes
\begin{equation}\label{est3}
\langle V\rangle\langle {\cal P}P_{SM_S}\rangle
+c_0 E^{(10)}_{\rm exch}
+E^{(10)}_{\rm exch}
\langle {\cal P}P_{SM_S}\rangle
=
\langle V{\cal P}P_{SM_S}\rangle
\end{equation}
Now, the last term on the l.h.s. is neglected as it requires at least two
electron exchanges (one in $E^{(10)}_{\rm exch}$ and one in $\langle {\cal P}P_{SM_S}\rangle$),
and we arrive at the final formula for the SRS-like
first-order exchange energy:
\begin{widetext}
\begin{equation}\label{e1exch}
E^{(10)}_{\rm exch}
=c_0^{-1}\left(\langle V{\cal P}P_{SM_S}\rangle
-\langle V\rangle
\langle {\cal P}P_{SM_S}\rangle\right)
\end{equation}
\end{widetext}
Let us now go back to Eq. (\ref{pspsi0}) and examine the spin-flipped
monomer wavefunctions such as $\Psi_A^{\downarrow}$. Up to normalization, this
function is equal to $\hat{S}_{-}\Psi_A$, where $\hat{S}_{-}=\sum_{k=1}^{N_A} \hat{S}_{-}(k)$
applies the conventional spin-lowering operator
$\hat{S}_{-}(k)=\hat{S}_{x}(k)-i\hat{S}_{y}(k)$ to
each of the $N_A$ spins. It can be shown that when $\Psi_A$
is a restricted high-spin determinant like in this work, the lowering
operator acts on it as follows,
\begin{widetext}
\begin{equation}\label{sminuspsi}
\hat{S}_{-}\Psi_A=\sum_{n=1}^{2S_A} \left|\chi_{i_1}\alpha \chi_{i_1}\beta \ldots
\chi_{i_{k_A}}\alpha \chi_{i_{k_A}}\beta \chi_{a_1}\alpha \ldots
\chi_{a_{n-1}}\alpha
\chi_{a_n}\beta \chi_{a_{n+1}}\alpha \ldots \chi_{a_{2S_A}}\alpha \right|
\end{equation}
\end{widetext}
where the spin orbitals present in each (normalized) determinant are
explicitly listed. Note that only the active spin orbitals end up being
spin-flipped. This observation is true for the ROHF determinants only:
lowering of the spin for an inactive spin orbital results in a duplicate spin
orbital and the determinant vanishes. If UHF determinants were considered,
this simplification would not take place. Analogously, the repeated application
of the spin-lowering operator, $\hat{S}_{-}\hat{S}_{-}\Psi_A$, gives a linear combination
of all possible determinants where two active spinorbitals have been flipped
from $\alpha$ to $\beta$. The function $\hat{S}_{-}\Psi_A$, Eq. (\ref{sminuspsi}), is
not normalized, but all $2S_A$ determinants entering the linear combination are
clearly orthonormal. Therefore, $\Psi_A^{\downarrow}=(1/\sqrt{2S_A}) \hat{S}_{-}\Psi_A$
is normalized.
Inserting Eq. (\ref{pspsi0}) into Eq. (\ref{e1exch}), we obtain
\begin{widetext}
\begin{eqnarray}\label{e1excha}
E^{(10)}_{\rm exch}
&=&
\langle V{\cal P}\rangle
+c_1/c_0\langle \Psi_A\Psi_B|V{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle
\nonumber \\ &&
+c_2/c_0\langle \Psi_A\Psi_B|V{\cal P}|
\Psi_A^{\downarrow\downarrow}\Psi_B^{\uparrow\uparrow}\rangle
+\ldots+
\nonumber \\ &&
-\langle V\rangle \left(
\langle {\cal P}\rangle
+c_1/c_0\langle \Psi_A\Psi_B|{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle
\right. \nonumber \\ && \left.
+c_2/c_0\langle \Psi_A\Psi_B|{\cal P}|
\Psi_A^{\downarrow\downarrow}\Psi_B^{\uparrow\uparrow}\rangle
+\ldots\right)
\end{eqnarray}
\end{widetext}
The first interesting observation from Eq. (\ref{e1excha}) is that the
standard, closed-shell like contribution (the ``spin-diagonal'' term)
$\langle V{\cal P}\rangle-\langle V\rangle\langle {\cal P}\rangle$
is present in the first-order exchange energy of any dimer spin state
regardless of the values of the Clebsch-Gordan coefficients. It is the
off-diagonal ``spin-flip'' terms that are responsible for the splittings (and, as we will
see below, only the single spin-flip term survives). Starting with the
diagonal term and using the density-matrix approach to exchange energy that
is valid in monomer- as well as dimer-centered basis sets, we can use
the general spinorbital
expression given by Eq. (39) of Moszy\'nski et al. \cite{Moszynski:94a}:
\begin{widetext}
\begin{eqnarray}\label{e1exmosz}
\langle V{\cal P}\rangle -\langle V\rangle\langle {\cal P}\rangle &=&
-\left[ \tilde{v}^{\beta\alpha}_{\alpha\beta}
+S^{\beta}_{\alpha'}(\tilde{v}^{\alpha\alpha'}_{\alpha\beta}
-\tilde{v}^{\alpha'\alpha}_{\alpha\beta})
+S^{\alpha}_{\beta'}(\tilde{v}^{\beta'\beta}_{\alpha\beta}
-\tilde{v}^{\beta\beta'}_{\alpha\beta})
\right. \nonumber \\ && \left.
-S^{\beta}_{\alpha'}S^{\alpha'}_{\beta'}\tilde{v}^{\alpha\beta'}_{\alpha\beta}
-S^{\beta'}_{\alpha'}S^{\alpha}_{\beta'}\tilde{v}^{\alpha'\beta}_{\alpha\beta}
+S^{\beta}_{\alpha'}S^{\alpha}_{\beta'}\tilde{v}^{\alpha'\beta'}_{\alpha\beta}
\right]
\end{eqnarray}
\end{widetext}
In that reference, $\alpha,\alpha'$ ($\beta,\beta'$) run over all occupied
spin orbitals of A (B). Equation (\ref{e1exmosz}) applies in the open-shell
case as well (for the diagonal term) but we have to break up the summation
over e.g. the occupied spin orbitals of A into the inactive orbitals
with spins $\alpha$ and $\beta$ and active orbitals with spin $\alpha$.
When we do that and perform all spin integrations, the resulting expression
for the diagonal term is
\begin{widetext}
\begin{eqnarray}\label{e1exdiag}
E^{(10)}_{\rm exch,diag}=\langle V{\cal P}\rangle -\langle V\rangle\langle {\cal P}\rangle &=&
- \tilde{v}^{ja}_{aj}
- \tilde{v}^{bi}_{ib}
-2 \tilde{v}^{ji}_{ij}
\nonumber \\ &&
+ \tilde{v}^{a'a}_{aj} S^{j}_{a'}
- \tilde{v}^{aa'}_{aj} S^{j}_{a'}
+ \tilde{v}^{jj'}_{aj} S^{a}_{j'}
-2 \tilde{v}^{j'j}_{aj} S^{a}_{j'}
+ \tilde{v}^{ai}_{ij} S^{j}_{a}
-2 \tilde{v}^{ia}_{ij} S^{j}_{a}
- \tilde{v}^{jb}_{ab} S^{a}_{j}
\nonumber \\ &&
+ \tilde{v}^{bb'}_{ib} S^{i}_{b'}
- \tilde{v}^{b'b}_{ib} S^{i}_{b'}
+ \tilde{v}^{i'i}_{ib} S^{b}_{i'}
-2 \tilde{v}^{ii'}_{ib} S^{b}_{i'}
+ \tilde{v}^{jb}_{ij} S^{i}_{b}
-2 \tilde{v}^{bj}_{ij} S^{i}_{b}
- \tilde{v}^{ai}_{ab} S^{b}_{i}
\nonumber \\ &&
+ \tilde{v}^{ia}_{aj} S^{j}_{i}
+ \tilde{v}^{bj}_{ib} S^{i}_{j}
-2 \tilde{v}^{ai}_{aj} S^{j}_{i}
-2 \tilde{v}^{jb}_{ib} S^{i}_{j}
\nonumber \\ &&
+2 \tilde{v}^{i'i}_{ij} S^{j}_{i'}
+2 \tilde{v}^{jj'}_{ij} S^{i}_{j'}
-4 \tilde{v}^{ii'}_{ij} S^{j}_{i'}
-4 \tilde{v}^{j'j}_{ij} S^{i}_{j'}
\nonumber \\ &&
+ \tilde{v}^{a'b}_{ab} S^{j}_{a'} S^{a}_{j}
+ \tilde{v}^{ab'}_{ab} S^{b}_{i} S^{i}_{b'}
+2 \tilde{v}^{ab}_{ib} S^{j}_{a} S^{i}_{j}
+2 \tilde{v}^{ab}_{aj} S^{j}_{i} S^{i}_{b}
\nonumber \\ &&
+2 \tilde{v}^{a'j}_{aj} S^{j'}_{a'} S^{a}_{j'}
+2 \tilde{v}^{aj'}_{aj} S^{j}_{i} S^{i}_{j'}
+ \tilde{v}^{aj'}_{aj} S^{j}_{a'} S^{a'}_{j'}
- \tilde{v}^{a'j'}_{aj} S^{j}_{a'} S^{a}_{j'}
\nonumber \\ &&
+2 \tilde{v}^{ib'}_{ib} S^{b}_{i'} S^{i'}_{b'}
+2 \tilde{v}^{i'b}_{ib} S^{j}_{i'} S^{i}_{j}
+ \tilde{v}^{i'b}_{ib} S^{b'}_{i'} S^{i}_{b'}
- \tilde{v}^{i'b'}_{ib} S^{b}_{i'} S^{i}_{b'}
\nonumber \\ &&
+4 \tilde{v}^{aj}_{ij} S^{j'}_{a} S^{i}_{j'}
-2 \tilde{v}^{aj'}_{ij} S^{j}_{a} S^{i}_{j'}
+4 \tilde{v}^{ib}_{ij} S^{j}_{i'} S^{i'}_{b}
-2 \tilde{v}^{i'b}_{ij} S^{j}_{i'} S^{i}_{b}
\nonumber \\ &&
+2 \tilde{v}^{i'j}_{ij} S^{b}_{i'} S^{i}_{b}
+4 \tilde{v}^{i'j}_{ij} S^{j'}_{i'} S^{i}_{j'}
+2 \tilde{v}^{ij'}_{ij} S^{j}_{a} S^{a}_{j'}
+4 \tilde{v}^{ij'}_{ij} S^{j}_{i'} S^{i'}_{j'}
\nonumber \\ &&
-2 \tilde{v}^{i'j'}_{ij} S^{j}_{i'} S^{i}_{j'}
\end{eqnarray}
\end{widetext}
where the implicit summations now run over orbitals: $i$--(inactive A),
$j$--(inactive B), $a$--(active A), $b$--(active B). It is worth noting
that an active-active exchange term $\tilde{v}^{ba}_{ab}$ is absent from
Eq. (\ref{e1exdiag}) due to the fact that all active orbitals on A are
paired with $\alpha$ spins and all active orbitals on B are paired with
$\beta$ spins.
For the first off-diagonal term, corresponding to a single spin flip between
the interacting monomers, Eq. (\ref{e1excha}) involves two matrix
elements, $\langle \Psi_A\Psi_B|V{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle$
and $\langle \Psi_A\Psi_B|{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle$.
Analogously to the conventional closed-shell formalism, these elements
can be expressed through the ``spin-flip interaction density matrix'':
\begin{equation}
\langle \Psi_A\Psi_B|V{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle=
\int \tilde{v}(12) \rho_{int}^{\downarrow\uparrow}(12)\,{\rm d}\tau_{12}
\end{equation}
\begin{equation}\label{ren1sf}
\langle \Psi_A\Psi_B|{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle=
\frac{1}{N_A N_B} \int \rho_{int}^{\downarrow\uparrow}(12)\,{\rm d}\tau_{12}
\end{equation}
where
\begin{equation}
\rho_{int}^{\downarrow\uparrow}(12)=N_A N_B \int \Psi_A\Psi_B {\cal P}
\Psi_A^{\downarrow}\Psi_B^{\uparrow}\,{\rm d}\tau'_{12}
\end{equation}
and, as always, ${\rm d}\tau'_{12}$ means the integration over coordinates of
all electrons except 1 and 2. The spin-flip interaction density matrix can
be expressed by the (also spin-flip) one- and two-electron reduced density
matrices in full analogy to Eq. (99) of Ref.~\cite{Patkowski:06}:
\begin{widetext}
\begin{eqnarray}\label{rhointsf}
\rho_{int}^{\downarrow\uparrow}(12)&=&
-\rho_A^{\downarrow}(1|2)\rho_B^{\uparrow}(2|1)
\nonumber \\ &&
-\int \rho_A^{\downarrow}(1|4)\Gamma_B^{\uparrow}(24|21)\,{\rm d}\tau_4
\nonumber \\ &&
-\int \Gamma_A^{\downarrow}(13|12)\rho_B^{\uparrow}(2|3)\,{\rm d}\tau_3
\nonumber \\ &&
-\int \Gamma_A^{\downarrow}(13|14)\Gamma_B^{\uparrow}(24|23)\,{\rm d}\tau_3
{\rm d}\tau_4
\end{eqnarray}
\end{widetext}
The spin-flip reduced density matrices on the r.h.s. of Eq.~(\ref{rhointsf}) are
in fact no different from ordinary transition density matrices: for example, for
monomer A
\begin{equation}\label{sf1rdm}
\rho_A^{\downarrow}(1|1')=N_A\int \Psi_A^{\ast}(1,2,\ldots,N_A)\Psi_A^{\downarrow}(1',2,\ldots,N_A)
{\rm d}\tau'_1
\end{equation}
\begin{equation}\label{sf2rdm}
\Gamma_A^{\downarrow}(12|1'2')=N_A(N_A-1)
\int \Psi_A^{\ast}(1,2,3,\ldots,N_A)\Psi_A^{\downarrow}(1',2',3,\ldots,N_A) {\rm d}\tau'_{12}
\end{equation}
Therefore, in order to obtain spinorbital expressions for these matrices, we
can use Eqs. (92) and (94) of Ref.~\cite{Patkowski:06},
except that a general $\alpha\to\rho$ excitation is replaced by a linear
combination of spin-flip
excitations. Therefore, we can straightforwardly use Eq. (92)
of Ref.~\cite{Patkowski:06} to write
\begin{equation}
\rho_A^{\downarrow}(1|1')= \frac{1}{\sqrt{2S_A}} \sum_{n=1}^{2S_A}
\chi_{a_n}\alpha(1)\chi_{a_n}\beta(1')
\end{equation}
The application of Eq. (94) from Ref.~\cite{Patkowski:06}
to obtain a formula for $\Gamma_A^{\downarrow}$
is a little more complicated because there is an additional summation over
an occupied orbital in this equation. This summation needs to be split
into three: over inactive orbitals with spin $\alpha$, over inactive orbitals
with spin $\beta$, and over active orbitals (which have spin $\alpha$). The
resulting expression is
\begin{widetext}
\begin{eqnarray}
\Gamma_A^{\downarrow}(12|1'2')&=&\frac{1}{\sqrt{2S_A}}\sum_{n=1}^{2S_A}\left[
\sum_{m=1}^{k_A} \left(
\chi_{a_n}\alpha(1)\chi_{i_m}\alpha(2)\chi_{a_n}\beta(1')\chi_{i_m}\alpha(2')
\right. \right. \nonumber\\ && \left.\left.
-\chi_{i_m}\alpha(1)\chi_{a_n}\alpha(2)\chi_{a_n}\beta(1')\chi_{i_m}\alpha(2')
\right. \right. \nonumber\\ && \left.\left.
-\chi_{a_n}\alpha(1)\chi_{i_m}\alpha(2)\chi_{i_m}\alpha(1')\chi_{a_n}\beta(2')
\right. \right. \nonumber\\ && \left.\left.
+\chi_{i_m}\alpha(1)\chi_{a_n}\alpha(2)\chi_{i_m}\alpha(1')\chi_{a_n}\beta(2')
\right)
\right. \nonumber\\ && \left.
+\sum_{m=1}^{k_A} \left(
\chi_{a_n}\alpha(1)\chi_{i_m}\beta(2)\chi_{a_n}\beta(1')\chi_{i_m}\beta(2')
\right. \right. \nonumber\\ && \left.\left.
-\chi_{i_m}\beta(1)\chi_{a_n}\alpha(2)\chi_{a_n}\beta(1')\chi_{i_m}\beta(2')
\right. \right. \nonumber\\ && \left.\left.
-\chi_{a_n}\alpha(1)\chi_{i_m}\beta(2)\chi_{i_m}\beta(1')\chi_{a_n}\beta(2')
\right. \right. \nonumber\\ && \left.\left.
+\chi_{i_m}\beta(1)\chi_{a_n}\alpha(2)\chi_{i_m}\beta(1')\chi_{a_n}\beta(2')
\right)
\right. \nonumber\\ && \left.
+\sum_{m=1}^{2S_A} \left(
\chi_{a_n}\alpha(1)\chi_{a_m}\alpha(2)\chi_{a_n}\beta(1')\chi_{a_m}\alpha(2')
\right. \right. \nonumber\\ && \left.\left.
-\chi_{a_m}\alpha(1)\chi_{a_n}\alpha(2)\chi_{a_n}\beta(1')\chi_{a_m}\alpha(2')
\right. \right. \nonumber\\ && \left.\left.
-\chi_{a_n}\alpha(1)\chi_{a_m}\alpha(2)\chi_{a_m}\alpha(1')\chi_{a_n}\beta(2')
\right. \right. \nonumber\\ && \left.\left.
+\chi_{a_m}\alpha(1)\chi_{a_n}\alpha(2)\chi_{a_m}\alpha(1')\chi_{a_n}\beta(2')
\right)\raisebox{0pt}[10pt][10pt]{}
\right]
\end{eqnarray}
\end{widetext}
Developing analogous formulas for the spin-flip
reduced density matrices of monomer
B, employing Eq.~(\ref{rhointsf}), and performing spin integration, one
arrives at the following formula:
\begin{widetext}
\begin{eqnarray}\label{offdiag1}
\langle \Psi_A\Psi_B|V{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle&=&
\frac{1}{2\sqrt{S_A S_B}}\left[
- \tilde{v}_{ab}^{ba} \right.
\nonumber \\ &&
+ \tilde{v}_{ab}^{bj} S_{j}^{a}
+ \tilde{v}_{aj}^{jb} S_{b}^{a}
-2 \tilde{v}_{aj}^{bj} S_{b}^{a}
+ \tilde{v}_{ab}^{bb'} S_{b'}^{a}
- \tilde{v}_{ab'}^{bb'} S_{b}^{a}
\nonumber \\ &&
+ \tilde{v}_{ab}^{ia} S_{i}^{b}
+ \tilde{v}_{ib}^{ai} S_{a}^{b}
-2 \tilde{v}_{ia}^{ib} S_{b}^{a}
+ \tilde{v}_{ab}^{a'a} S_{a'}^{b}
- \tilde{v}_{a'a}^{a'b} S_{b}^{a}
\nonumber \\ &&
-2 \tilde{v}_{ab}^{ij} S_{i}^{b} S_{j}^{a}
-2 \tilde{v}_{aj}^{ib} S_{i}^{j} S_{b}^{a}
+4 \tilde{v}_{aj}^{ij} S_{i}^{b} S_{b}^{a}
+4 \tilde{v}_{ib}^{ij} S_{a}^{b} S_{j}^{a}
-4 \tilde{v}_{ij}^{ij} S_{a}^{b} S_{b}^{a}
\nonumber \\ &&
+2 \tilde{v}_{ib}^{ib'} S_{a}^{b} S_{b'}^{a}
-2 \tilde{v}_{ab}^{ib'} S_{i}^{b} S_{b'}^{a}
-2 \tilde{v}_{ib'}^{ib'} S_{a}^{b} S_{b}^{a}
+2 \tilde{v}_{ib'}^{ab'} S_{a}^{b} S_{b}^{i}
\nonumber \\ &&
+2 \tilde{v}_{aj}^{a'j} S_{a'}^{b} S_{b}^{a}
-2 \tilde{v}_{ab}^{a'j} S_{a'}^{b} S_{j}^{a}
-2 \tilde{v}_{a'j}^{a'j} S_{a}^{b} S_{b}^{a}
+2 \tilde{v}_{a'b}^{a'j} S_{a}^{b} S_{j}^{a}
\nonumber \\ && \left.
+ \tilde{v}_{a'b}^{a'b'} S_{a}^{b} S_{b'}^{a}
- \tilde{v}_{a'b'}^{a'b'} S_{a}^{b} S_{b}^{a}
- \tilde{v}_{ab}^{a'b'} S_{a'}^{b} S_{b'}^{a}
+ \tilde{v}_{ab'}^{a'b'} S_{a'}^{b} S_{b}^{a} \right]
\end{eqnarray}
\end{widetext}
For the renormalization term, Eq.~(\ref{ren1sf}), we find (for example,
by reanalyzing Eq.~(\ref{offdiag1}), replacing each $\tilde{v}$ by
$\frac{1}{N_A N_B}$ times the appropriate product of overlap integrals) that
\begin{equation}\label{ren1a}
\langle \Psi_A\Psi_B|{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle=
-\frac{1}{2\sqrt{S_A S_B}}S_{a}^{b} S_{b}^{a}
\end{equation}
The ROHF electrostatic energy is equal to
\begin{equation}\label{ren1b}
E^{(10)}_{\rm elst}=\langle V\rangle=
4\tilde{v}^{ij}_{ij}+2\tilde{v}^{aj}_{aj}+2\tilde{v}^{ib}_{ib}
+\tilde{v}^{ab}_{ab}
\end{equation}
Combining Eqs.~(\ref{offdiag1})--(\ref{ren1b}), we obtain the final formula
for the single spin-flip contribution to Eq.~(\ref{e1excha}):
\begin{widetext}
\begin{eqnarray}\label{offdiag1final}
\lefteqn{\frac{1}{2\sqrt{S_A S_B}} E^{(10)}_{\rm exch,1flip}=
\langle \Psi_A\Psi_B|V{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle
-\langle V\rangle
\langle \Psi_A\Psi_B|{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle=}
\\ &&
\frac{1}{2\sqrt{S_A S_B}}\left[ - \tilde{v}_{ab}^{ba} \right.
+ \tilde{v}_{ab}^{bj} S_{j}^{a}
+ \tilde{v}_{aj}^{jb} S_{b}^{a}
-2 \tilde{v}_{aj}^{bj} S_{b}^{a}
+ \tilde{v}_{ab}^{bb'} S_{b'}^{a}
- \tilde{v}_{ab'}^{bb'} S_{b}^{a}
\nonumber \\ &&
+ \tilde{v}_{ab}^{ia} S_{i}^{b}
+ \tilde{v}_{ib}^{ai} S_{a}^{b}
-2 \tilde{v}_{ia}^{ib} S_{b}^{a}
+ \tilde{v}_{ab}^{a'a} S_{a'}^{b}
- \tilde{v}_{a'a}^{a'b} S_{b}^{a}
\nonumber \\ &&
-2 \tilde{v}_{ab}^{ij} S_{i}^{b} S_{j}^{a}
-2 \tilde{v}_{aj}^{ib} S_{i}^{j} S_{b}^{a}
+4 \tilde{v}_{aj}^{ij} S_{i}^{b} S_{b}^{a}
+4 \tilde{v}_{ib}^{ij} S_{a}^{b} S_{j}^{a}
\nonumber
+2 \tilde{v}_{ib}^{ib'} S_{a}^{b} S_{b'}^{a}
-2 \tilde{v}_{ab}^{ib'} S_{i}^{b} S_{b'}^{a}
+2 \tilde{v}_{ib'}^{ab'} S_{a}^{b} S_{b}^{i}
\nonumber \\ &&
+2 \tilde{v}_{aj}^{a'j} S_{a'}^{b} S_{b}^{a}
-2 \tilde{v}_{ab}^{a'j} S_{a'}^{b} S_{j}^{a}
+2 \tilde{v}_{a'b}^{a'j} S_{a}^{b} S_{j}^{a}
\nonumber \left.
+ \tilde{v}_{a'b}^{a'b'} S_{a}^{b} S_{b'}^{a}
- \tilde{v}_{ab}^{a'b'} S_{a'}^{b} S_{b'}^{a}
+ \tilde{v}_{ab'}^{a'b'} S_{a'}^{b} S_{b}^{a} \right]
\end{eqnarray}
\end{widetext}
Finally, we will show that the double spin-flip term in Eq. (\ref{e1excha}),
and all subsequent terms, vanish identically. Specifically, we will
prove that
$\langle \Psi_A\Psi_B|V{\cal P}|
\Psi_A^{\downarrow\downarrow}\Psi_B^{\uparrow\uparrow}\rangle=0$, where
$\Psi_A^{\downarrow\downarrow}$ and $\Psi_B^{\uparrow\uparrow}$ are
proportional to $\hat{S}_{-}\hat{S}_{-}\Psi_A$ and $\hat{S}_{+}\hat{S}_{+}\Psi_B$, respectively.
Similar to the previous term,
\begin{equation}\label{offdiag2sf}
\langle \Psi_A\Psi_B|V{\cal P}|\Psi_A^{\downarrow\downarrow}
\Psi_B^{\uparrow\uparrow}\rangle=
\int \tilde{v}(12) \rho_{int}^{\downarrow\downarrow\uparrow\uparrow}(12)
\,{\rm d}\tau_{12}
\end{equation}
where the double spin-flip interaction density matrix
$\rho_{int}^{\downarrow\downarrow\uparrow\uparrow}(12)$ is given by a formula
analogous to Eq.~(\ref{rhointsf}). However, as a double spin flip can be
treated as a special case of a double excitation, the one- and two-electron
reduced spin-flip density matrices in the modified Eq.~(\ref{rhointsf}) are
now given by Eqs. (93) and (95), respectively, of Ref.~\cite{Patkowski:06}.
Consequently,
\begin{equation}\label{rho12sf}
\rho_A^{\downarrow\downarrow}(1|1')=0
\end{equation}
and the two-electron density matrix is proportional to
\begin{eqnarray}\label{gamma12sf}
\Gamma_A^{\downarrow\downarrow}(12|1'2')&\sim&\sum_{m,n=1}^{2S_A}
\left[
\chi_{a_n}\alpha(1)\chi_{a_m}\alpha(2)\chi_{a_n}\beta(1')\chi_{a_m}\beta(2')
\right. \nonumber\\ && \left.
-\chi_{a_m}\alpha(1)\chi_{a_n}\alpha(2)\chi_{a_n}\beta(1')\chi_{a_m}\beta(2')
\right. \nonumber\\ && \left.
-\chi_{a_n}\alpha(1)\chi_{a_m}\alpha(2)\chi_{a_m}\beta(1')\chi_{a_n}\beta(2')
\right. \nonumber\\ && \left.
+\chi_{a_m}\alpha(1)\chi_{a_n}\alpha(2)\chi_{a_m}\beta(1')\chi_{a_n}\beta(2')
\right]
\end{eqnarray}
According to Eq.~(\ref{rho12sf}), only the last term in the expression for
$\rho_{int}^{\downarrow\downarrow\uparrow\uparrow}$ survives:
\begin{equation}
\rho_{int}^{\downarrow\downarrow\uparrow\uparrow}(12)=
-\int \Gamma_A^{\downarrow\downarrow}(13|14)
\Gamma_B^{\uparrow\uparrow}(24|23)\,{\rm d}\tau_3
{\rm d}\tau_4
\end{equation}
Such an expression vanishes upon the integration over ${\rm d}\tau_1$
and ${\rm d}\tau_2$, either alone or with a spin-independent operator such
as $\tilde{v}(12)$ (like in Eq. (\ref{offdiag2sf})). In particular, in the
integration over ${\rm d}\tau_1$, the two spinorbitals in
$\Gamma_A^{\downarrow\downarrow}(13|14)$ that depend on the coordinates of
electron 1 always occur with opposite spins (cf. Eq. (\ref{gamma12sf})),
leading to a zero spin integral.
This concludes the proof that the double spin-flip contribution to
Eq. (\ref{e1excha}) vanishes.
To conclude, the first-order SAPT exchange correction for an interaction
of two high-spin open-shell molecules with spins $S_A$ and $S_B$, in an
arbitrary dimer spin state $|S_A-S_B|\le S\le S_A+S_B$, is given by the formula
which combines the orbital expressions for $E^{(10)}_{\rm exch,diag}$ and
$E^{(10)}_{\rm exch,1flip}$ defined in Eqs. (\ref{e1exdiag}) and
(\ref{offdiag1final}), respectively. By simplifying the $c_1/\left(2c_0\sqrt{S_A S_B}\right)$ coefficient,
as outlined in the Appendix,
we obtain
\begin{eqnarray}\label{e1exchf}
E^{(10)}_{\rm exch}
&=&
\langle V{\cal P}\rangle
-\langle V\rangle
\langle {\cal P}\rangle
\nonumber \\ &&
+c_1/c_0\left[
\langle \Psi_A\Psi_B|V{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle
-\langle V\rangle
\langle \Psi_A\Psi_B|{\cal P}|\Psi_A^{\downarrow}\Psi_B^{\uparrow}\rangle\right]
\nonumber \\ & =& E^{(10)}_{\rm exch,diag} + Z(S_A,S_B,S) E^{(10)}_{\rm exch,1flip}
\end{eqnarray}
where
\begin{equation}\label{clebsch}
Z(S_A,S_B,S) = \frac{S(S+1)+2S_A S_B-S_A(S_A+1) -S_B(S_B+1)}{ 4 S_A S_B}.
\end{equation}
Note that $Z(S_A,S_B,S_A+S_B)=1$, while $Z(S_A,S_B,|S_A-S_B|)=-\frac{1}{2\max{(S_A,S_B)}}$.
All of the resulting matrix elements are the same
for the full set of asymptotically degenerate multiplets of the dimer: the
only factors in Eq.~(\ref{e1exchf}) that are different for different spin
states are $Z(S_A,S_B,S)$. Thus, as expected from
the Heisenberg Hamiltonian model, the ratios of splittings between different
multiplets are simple, system-independent numbers (expressible through
the Clebsch-Gordan coefficients) as long as the single exchange approximation
is applied. The observation that the single exchange approximation implies the
Heisenberg picture is not new: in fact, it dates back to the work of
Matsen {\em et al.} \cite{Matsen:71} on spin-free quantum chemistry.
For practical applications to large systems, it is advantageous to recast
the newly derived $E^{(10)}_{\rm exch}$ expressions from the molecular-orbital
(MO) basis into the atomic orbital (AO) one, similar to the AO first-order
exchange expressions without the single-exchange approximation for
closed-shell SAPT \cite{Hesselmann:05} and UHF-based high-spin open-shell
SAPT \cite{Hapka:12}. The AO approach is advantageous as it avoids the
AO-MO transformation of many different types of two-electron integrals
present in Eqs. (\ref{e1exdiag}) and (\ref{offdiag1final}). Moreover, as
we will see below, the AO expressions make heavy use of the generalized
Coulomb and exchange operators, whose evaluation in the {\sc psi4}
code \cite{Parrish:17} is highly optimized both with and without
density fitting (DF).
In the following, we will denote AO indices by capital letters assuming, for
simplicity, that the same AO basis set has been used to expand molecular
orbitals of A and B (according to the so-called dimer-centered basis set
formalism \cite{Williams:95}). The SCF coefficients of the molecular
spinorbital $\lambda$ will be denoted by $C^K_{\lambda}$ --- obviously,
in the ROHF formalism, the SCF
coefficients are the same for spin $\alpha$ and spin $\beta$, for example, $C^K_{i\alpha}=C^K_{i\beta}$.
We will use boldface letters for matrices, and denote by
${\mathbf A}\cdot {\mathbf B}=\sum_{KL} {\mathbf A}_{KL}{\mathbf B}_{KL}$ the scalar product of two matrices.
The inactive and active parts of density matrices for monomer A are ${\mathbf P}^{iA}_{KL}=C^K_{i\alpha}C^L_{i\alpha}=
C^K_{i\beta}C^L_{i\beta}$, ${\mathbf P}^{aA}_{KL}=C^K_{a\alpha}C^L_{a\alpha}$, and similarly for monomer B (note that the active electrons on A all have spin $\alpha$, the active electrons on B all have spin $\beta$).
Therefore, the
total density matrix is ${\mathbf P}^{A}=2{\mathbf P}^{iA}+{\mathbf P}^{aA}$.
Now, the
generalized Coulomb and exchange matrices are defined as
\begin{equation}
{\mathbf J}[{\mathbf X}]_{KL} = (KL|MN) {\mathbf X}_{MN}
\end{equation}
\begin{equation}
{\mathbf K}[{\mathbf X}]_{KL} = (KM|NL) {\mathbf X}_{MN}
\end{equation}
and reduce to standard Coulomb and exchange matrices, or their inactive/active contributions, in special cases, for example,
${\mathbf J}^{iA}={\mathbf J}[{\mathbf P}^{iA}]$, ${\mathbf K}^{aA}={\mathbf K}[{\mathbf P}^{aA}]$, \ldots The electrostatic
potential matrix for monomer A is
\begin{equation}
\bm{\omega}_A={\mathbf v}_A+2{\mathbf J}^{iA}+{\mathbf J}^{aA}
\end{equation}
where ${\mathbf v}_A$ is the matrix of the nuclear attraction operator.
By adding the exchange matrices, we obtain the following spin-dependent Fock matrices ${\mathbf h}$:
\begin{equation}
{\mathbf h}_A^{\alpha}={\mathbf v}_A+2{\mathbf J}^{iA}+{\mathbf J}^{aA}-{\mathbf K}^{iA}-{\mathbf K}^{aA}
\end{equation}
\begin{equation}
{\mathbf h}_A^{\beta}={\mathbf v}_A+2{\mathbf J}^{iA}+{\mathbf J}^{aA}-{\mathbf K}^{iA}
\end{equation}
\begin{equation}
{\mathbf h}_B^{\alpha}={\mathbf v}_B+2{\mathbf J}^{iB}+{\mathbf J}^{aB}-{\mathbf K}^{iB}
\end{equation}
\begin{equation}
{\mathbf h}_B^{\beta}={\mathbf v}_B+2{\mathbf J}^{iB}+{\mathbf J}^{aB}-{\mathbf K}^{iB}-{\mathbf K}^{aB}
\end{equation}
With this notation, the diagonal and single spin-flip terms of the ROHF-based $E^{(10)}_{\rm exch}$ formula,
Eqs.~(\ref{e1exdiag}) and (\ref{offdiag1final}), respectively, can be rewritten as follows:
\begin{align}
E^{(10)}_{\rm exch,diag}=&
- {\mathbf P}^{iB} \cdot (2{\mathbf K}^{iA}+{\mathbf K}^{aA})
- {\mathbf P}^{aB} \cdot {\mathbf K}^{iA}
\nonumber \\ &
- ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot ({\mathbf h}_A^{\alpha}+{\mathbf h}_A^{\beta}+{\mathbf h}_B^{\alpha}+{\mathbf h}_B^{\beta})
\nonumber \\ &
- ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot ({\mathbf h}_A^{\alpha}+{\mathbf h}_B^{\alpha})
- ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot ({\mathbf h}_A^{\beta}+{\mathbf h}_B^{\beta})
\nonumber \\ &
+2 ({\mathbf P}^{iB} {\mathbf S}^{AO} {\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot \bm{\omega}_A
+2 ({\mathbf P}^{iB} {\mathbf S}^{AO} {\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot \bm{\omega}_A
\nonumber \\ &
+ ({\mathbf P}^{aB} {\mathbf S}^{AO} {\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot \bm{\omega}_A
+ ({\mathbf P}^{iB} {\mathbf S}^{AO} {\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot \bm{\omega}_A
\nonumber \\ &
+2 ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB} {\mathbf S}^{AO} {\mathbf P}^{iA}) \cdot \bm{\omega}_B
+2 ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB} {\mathbf S}^{AO} {\mathbf P}^{aA}) \cdot \bm{\omega}_B
\nonumber \\ &
+ ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB} {\mathbf S}^{AO} {\mathbf P}^{iA}) \cdot \bm{\omega}_B
+ ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB} {\mathbf S}^{AO} {\mathbf P}^{aA}) \cdot \bm{\omega}_B
\nonumber \\ &
-2 ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot {\mathbf K}[{\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB}]
-2 ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot {\mathbf K}[{\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB}]
\nonumber \\ &
-2 ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot {\mathbf K}[{\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB}]
- ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot {\mathbf K}[{\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}]
\nonumber \\ &
- ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot {\mathbf K}[{\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}]
\label{aodiag}
\end{align}
\begin{align}
E^{(10)}_{\rm exch,1flip}=&
- {\mathbf P}^{aB} \cdot {\mathbf K}^{aA}
- ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot ({\mathbf h}_A^{\alpha}+{\mathbf h}_B^{\beta})
\nonumber \\ &
+ ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot {\mathbf K}^{aB}
+ ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot {\mathbf K}^{aA}
\nonumber \\ &
+2 ({\mathbf P}^{iB} {\mathbf S}^{AO} {\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot \bm{\omega}_A
+ ({\mathbf P}^{aB} {\mathbf S}^{AO} {\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot \bm{\omega}_A
\nonumber \\ &
+2 ({\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB} {\mathbf S}^{AO} {\mathbf P}^{aA}) \cdot \bm{\omega}_B
+ ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB} {\mathbf S}^{AO} {\mathbf P}^{aA}) \cdot \bm{\omega_B}
\nonumber \\ &
-2 ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot {\mathbf K}[{\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{iB}]
-2 ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot {\mathbf K}[{\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}]
\nonumber \\ &
-2 ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot {\mathbf K}[{\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}]
-2 ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{iB}) \cdot {\mathbf K}[{\mathbf P}^{iA} {\mathbf S}^{AO} {\mathbf P}^{aB}]
\nonumber \\ &
- ({\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}) \cdot {\mathbf K}[{\mathbf P}^{aA} {\mathbf S}^{AO} {\mathbf P}^{aB}]
\label{aoflip}
\end{align}
where ${\mathbf S}^{AO}$ is the overlap matrix in the AO basis.
\section{Computational Details}
The MO-based formulas for the arbitrary-spin first-order exchange energy, Eqs.~(\ref{e1exdiag}) and (\ref{offdiag1final}),
have been implemented in two versions to verify the numerical correctness of the codes: into the ROHF-based SAPT code of
Ref.~\cite{Zuchowski:08} forming a part of the {\sc SAPT2012} package \cite{SAPT2012}, and into the development version of
the {\sc psi4} package \cite{Parrish:17} using the straightforward {\sc psi4numpy} framework \cite{Smith:18} in which
each term corresponds to a single line of Python code.
The exchange energy for the high-spin state of the dimer could also be computed with the code of Ref.~\cite{Zuchowski:08}
and we have verified that our new code gives identical results in the high-spin case. In fact, for two interacting
doublets, the formulas for $E^{(10)}_{\rm exch,diag}$ and $E^{(10)}_{\rm exch,1flip}$ (Eqs.~(\ref{e1exdiag}) and (\ref{offdiag1final}),
respectively) add up exactly to the formula for $E^{(10)}_{\rm exch}$ in the high-spin (triplet) state in accordance
with Eq.~(\ref{clebsch}) for $S_A=S_B=1/2$, $S=1$.
The AO-based first-order exchange expressions,
Eqs.~(\ref{aodiag}) and (\ref{aoflip}), have been implemented into the development version of {\sc psi4}, making use of its
efficient evaluation of generalized Coulomb and exchange matrices, both with and without the DF
approximation. Whenever the latter approximation was applied, the standard aug-cc-pV$X$Z/JKFIT set \cite{Weigend:02}
was used as the auxiliary basis accompanying the orbital set aug-cc-pV$X$Z \cite{Kendall:92}. As shown in Sec.~\ref{sec:phenalenyl},
the DF approximation to $E^{(10)}_{\rm exch}$ performs very well, and the density fitted AO-based code
can be applied to much larger systems than the ones studied here.
Unless stated otherwise, we used the aug-cc-pVTZ basis set.
\section{Numerical Results}
In this section, we study the applications of theory developed in Sec. \ref{theory} to several test systems.
For the first one, the Li$\cdots$H complex, it is possible to compare the spin-flip theory with exact SAPT calculations
of Patkowski et al.~\cite{Patkowski:01}. Then, we focus on other atom-atom complexes including the lithium dimer,
Li$\cdots$N, and N$\cdots$N systems. We also examine the Mn$\cdots$Mn system,
very extensively studied in the past, for which the spin-exchange splitting is surprisingly low. Finally, we also show a test of
exchange splittings for molecular cases: the O$_2$($^3\Sigma^-_g$) dimer and the phenalenyl dimer.
For the atomic Li$\cdots$Li, Li$\cdots$N, and N$\cdots$N systems we can compare the energy differences between low-spin and high-spin dimers
with results obtained from multireference {\em ab initio} methods: complete active space self-consistent field (CASSCF) and internally contracted
multireference configuration interaction (MRCI) with the Davidson correction implemented in {\sc molpro} 2012 \cite{Werner:12}.
The exchange splitting $\Delta E$ between the interaction energies for the highest ($E_{\rm int} (S_A+S_B) $) and lowest ($E_{\rm int} (|S_A-S_B|) $ ) multiplicities from {\em ab initio} calculations can be compared with the spin-flip SAPT term, Eq.~(\ref{offdiag1final}),
by multiplying the latter by a following factor:
\begin{equation}
\Delta E^{(10)} = (1+ \frac{1}{ 2 \max{(S_A,S_B)} } ) E^{(10)}_{\rm exch,1flip},
\end{equation}
where from now on we use $\Delta E^{(10)}$ to denote the difference between the highest- and lowest multiplicity of a complex at the
SAPT level introduced in this work.
\subsection{Li$\cdots$H}
The first system we investigate is the Li$\cdots$H complex, where the interaction between two doublets gives rise to
a triplet state and a singlet state. For both of them, full configuration interaction (FCI) interaction energies as
well as SAPT corrections to high order (in several variants including SRS) have been obtained before \cite{Patkowski:01}.
Therefore, to facilitate our comparisons with the data of Ref.~\cite{Patkowski:01}, we use the same (4s4p1d/5s2p) basis
set in our first-order exchange calculations. The exchange energies of both spin states as functions
of the interatomic separation $R$ are presented in Fig.~\ref{fig:lih}. The excellent agreement between the results of
Ref.~\cite{Patkowski:01} (where the electron correlation within the lithium atom was described at the FCI level) and
our new values (where the Li atom was described by the ROHF theory) not only validates our implementation, but also
indicates that the intramonomer correlation effect on $E^{(1)}_{\rm exch}$ is negligible for this system.
At the
range of distances presented in Fig.~\ref{fig:lih}, the diagonal exchange energy term, Eq.~(\ref{aodiag}), is much
smaller than the spin-flip term, Eq.~(\ref{aoflip}). The $E^{(10)}_{\rm exch,diag}$ contribution stems from
the difference between the ``apparent Coulomb energy'', the arithmetic mean of the singlet
and triplet energies, and the actual mathematical Coulomb energy, defined in SAPT as a weighted average of the energies
of all asymptotically degenerate states contributing to the zeroth-order wavefunction, including the Pauli-forbidden
ones \cite{Kutzelnigg:80,Patkowski:01}. It can be shown that, in the first order, the mathematical Coulomb energy is
identical to the electrostatic correction $E^{(10)}_{\rm elst}$ (Eq.~(\ref{ren1b})) up to terms that vanish as the
fourth or higher powers of intermolecular overlap integrals \cite{Jeziorski:priv}.
The apparent Coulomb
energy does not have this property and the difference between the two (that is, $E^{(10)}_{\rm exch,diag}$)
does not vanish in the $S^2$ approximation. The only exception are the interactions involving
one- and two-electron systems, such as H$\cdots$H \cite{Cwiok:92}, He ($^3S$)$\cdots$H \cite{Przybytek:04}, and
He ($^3S$)$\cdots$He ($^3S$) \cite{Przybytek:05}, when the mathematical Coulomb energy coincides with the mean energy
of physical states. In these cases,
$E^{(10)}_{\rm exch,diag}$ (Eq.~(\ref{e1exdiag})) is identically zero as there are no inactive orbitals
(the range of summation over $i$ and $j$ is
empty). For the Li$\cdots$H complex, as already observed in Ref.~\cite{Patkowski:01},
the two Coulomb energies are particularly close and the $E^{(10)}_{\rm exch,diag}$ term is
much smaller than $E^{(10)}_{\rm exch,1flip}$.
As a result, the singlet and triplet exchange energies are almost exactly the negatives of each other.
Assuming that this is precisely the case, one could define the exchange energy at the
FCI level as one half of the difference between the FCI interaction energies of triplet and singlet.
The FCI exchange energies defined in this way are also displayed in
Fig.~\ref{fig:lih}. Interestingly, $E^{(10)}_{\rm exch}$ represents about two thirds of the total exchange energy,
showing that the first-order description of exchange is qualitatively correct. This result should be contrasted with
the first-order electrostatic correction $E^{(10)}_{\rm elst}$ that is nearly negligible for this dispersion-bound complex.
As a result, the sum $E^{(10)}_{\rm elst}+E^{(10)}_{\rm exch,diag}$ recovers only a small fraction
(8\%\ at the triplet van der Waals minimum distance of 11.5$a_0$) of the FCI apparent Coulomb
energy while $E^{(10)}_{\rm exch,1flip}$ recovers 71\%\ of the FCI exchange energy.
\subsection{Li$\cdots$Li, Li$\cdots$N, and N$\cdots$N}
The total spin for the Li$\cdots$Li complex can be either singlet or triplet, for Li$\cdots$N - triplet or quintet, and for two interacting nitrogen atoms it can take multiplicities from singlet to septet.
For these systems, unlike for the Li$\cdots$H interaction, the benchmark SAPT corrections with exact monomer wavefunctions are not available. In Fig.~\ref{fig:li2_lin_n2:split} we show the comparison of exchange splittings from SF-SAPT, CASSCF, and MRCI (Davidson corrected).
As expected, the first-order result fails to recover the exchange splitting around the chemical minima, which are at 5$a_0$ for Li$\cdots$Li, 3.5$a_0$ for Li$\cdots$N, and 2.1$a_0$ for the nitrogen dimer.
Nonetheless, in all these cases SF-SAPT exchange splitting is very close to the CASSCF one for the separations corresponding to the
van der Waals minima of the highest spin states: 7.9$a_0$ for the lithium dimer~\cite{Semczuk:13}, 10.2$a_0$ for Li$\cdots$N (this value was obtained in present work using the spin-restricted CCSD(T) method with the aug-cc-pVQZ basis set), and 7.2$a_0$ for the nitrogen dimer~\cite{Tscherbul:10}.
Clearly, the exchange splitting from first-order SAPT exhibits correct asymptotic behavior. It is also worth noting that for Li$\cdots$Li,
unlike Li$\cdots$H, the diagonal and spin-flip exchange terms are of the same order as the apparent and mathematical Coulomb energies
are no longer close.
Moreover, the splittings, as well as
$E^{(10)}_{\rm elst}$ and $E^{(10)}_{\rm exch,diag}$, show very little basis set dependence: our Li$\cdots$Li tests at $R=7.9a_0$
show that each of these three first-order contributions agrees between the (dimer-centered) cc-pVDZ and aug-cc-pVQZ basis sets to below 1\%.
In case of the lithium dimer, we could also compare the exchange splittings to the experimentally derived values given by C{\^{o}}t{\'{e}} et al. \cite{Cote:94} which confirm the correct behavior of the MRCI method. Since the $1s$ shells were frozen in MRCI
calculations, we experienced no problems with a size-inconsistent behavior of this method. Quite large differences between splittings obtained from MRCI and CASSCF/SAPT can possibly be attributed to strong influence of intramonomer dynamical correlation on the interaction effects, in particular exchange induction.
For the Li$\cdots$N and N$\cdots$N systems, the differences between MRCI, CASSCF, and first-order SAPT are smaller. On the other hand, it is worth noting that the MRCI method fails to reproduce the exponential decay of the exchange splitting for large interatomic separations for Li$\cdots$N and N$\cdots$N, due to its lack of size-consistency.
It might be quite useful to compare the quality of the $S^2$ approximation by inspecting the ratio of $E^{(10)}_{\rm exch}(S^2)$ and $E^{(10)}_{\rm exch}$ for the high-spin state (the nonapproximated values were computed using the high-spin code of Ref.~\cite{Zuchowski:08}).
From Fig. \ref{fig:li2_lin_n2:s2} it is clear that while for the lithium dimer the single exchange approximation is quite drastic, even in the van der Waals minimum region, it works very well for the nitrogen dimer and the Li$\cdots$N system. Such behavior can be attributed to a very small ionization potential $I_p$ of the Li atom (4.2 eV) compared to the N atom (14.5 eV) which directly affects the decay rates of the wavefunctions (at long range, this decay rate is proportional to $\exp{(-\sqrt{2 I_p}r)} $ \cite{Ahlrichs:81}).
\subsection{ O$_2\cdots$O$_2$}
Oxygen dimer is one of the best studied van der Waals complexes with relevance to the chemistry of atmosphere \cite{Aquilanti:99}. Hence, this was one of very first systems for which the exchange interaction was studied.
Wormer and van der Avoird calculated the exchange splitting within the Heisenberg model and the $J_{AB}$ parameter was obtained with variation-perturbation theory \cite{Wormer:84}.
Later, the exchange splitting was studied by several {\em ab initio} multireference methods. In particular, global CASPT2 surfaces for all multiplicities were obtained by Bartolomei and coworkers \cite{Bartolomei:08}.
To facilitate comparisons to Ref.~\cite{Bartolomei:08}, our SAPT calculations for this system use the same ANO-VTZ basis set and the same bond length in the oxygen molecules (2.28$a_0$).
In Fig.~\ref{fig:O2O2fig1} we compare the splitting between the highest and lowest spin states of the O$_2\cdots$O$_2$ complex for four basic angular geometries: H-shape, linear, T-shape, and X-shape with previous studies of Wormer and van der Avoird~\cite{Wormer:84} and Bartolomei and coworkers~\cite{Bartolomei:08}, the latter including CASSCF, MRCI, and CASPT2 calculations.
The overall agreement of the spin-flip SAPT exchange energy with these references is very satisfactory.
The exchange splittings for the studied geometries are in a very good agreement with the Bartolomei {\em et al.} CASSCF and Wormer and van der Avoird perturbation calculations for the H-, T-, and L-shape complexes.
For the X-shape geometry, the first-order exchange splittings perform similarly to MRCI and are significantly smaller compared to Heitler-London calculations of Ref.~\cite{Wormer:84} and to CASSCF.
In the CASPT2 calculations, the exchange splittings exhibit a node for the X-shape orientation,
which is also the case for the SAPT exchange at about 7.5$a_0$ but it is not clearly seen in Fig.~\ref{fig:O2O2fig1}.
This node can, however, be seen on the logarithmic plot of the diagonal and spin-flip parts of the exchange splitting, Fig.~\ref{fig:O2O2fig2}, as a dip in the absolute value of the splitting.
Note that the diagonal part of the exchange energy is very large in the oxygen dimer for all configurations including X-shape, almost two orders of magnitude larger than the spin-flip part.
Such a big domination of the diagonal part is responsible for the fact that the oxygen molecules strongly repel each other at short range and form only weakly bound complexes for all spin states of the dimer. Finally, let us remark that the single exchange approximation works very well for the oxygen dimer, as shown by the comparison
of full and $S^2$-approximated $E^{(10)}_{\rm exch}$ values for the high-spin (quintet) complex displayed in
Fig.~\ref{fig:O2O2fig_S2}. This behavior is consistent with the observation made by Wormer and van der Avoird~\cite{Wormer:84} that the single-parameter Heisenberg model recovers the quintet-triplet-singlet splittings accurately.
\subsection{Mn$\cdots$Mn}
The manganese dimer has attracted many studies of its exceptional exchange splitting. Its outermost electronic shell $4s$ is doubly filled and its zero orbital angular momentum ($S$ state) originates from a cancellation of half-filled $d$-shell momenta of electrons.
For this reason, the open shells in the Mn atom are to a large extent screened by the $4s^2$ shell and the resulting spin exchange splitting is very small, on the order of 10 cm$^{-1}$ in the minimum of the potential energy curve, which is 2 orders of magnitude less than the binding energy.
The first {\em ab initio} study of exchange interaction in the Mn$_2$ system was initiated by Nesbet~\cite{Nesbet:64} who used the Heisenberg model and estimated the $J_{AB}$ parameter to be small (up to 62 K at an interatomic separation of 4.5$a_0$).
More advanced methods were introduced to this system after significant progress in the multireference methods was made~\cite{Wang:04,Yamamoto:06,Negodaev:08,Tzeli:08,Camacho:08,Buchachenko:10}.
This system is however, very challenging for multireference methods: there are 5 electrons in the submerged $d$-shell plus two electrons on the outermost $4s$ shell per atom, which makes the active space needed for dimer calculations quite large.
Moreover, the perturbation theory approaches like CASPT2 exhibit problems for Mn$\cdots$Mn related to a presence of intruder states~\cite{Camacho:08}.
Clearly, a development of new methods for systems similar to the manganese dimer is warranted.
The ROHF method produces correct orbitals for the Mn atom (with a correct degeneracy of $d$ orbitals) and can be straightforwardly applied to first-order SF-SAPT.
In Fig.~\ref{fig:mn2}, the spin-flip SAPT exchange term is very small, nearly two orders of magnitude smaller than the diagonal exchange energy.
Similarly to the oxygen dimer, this causes a very small exchange splitting and very small differences between the potential well depths and equilibrium distances for different multiplets.
It is also worthwhile to inspect the quality of the $S^2$ approximation by comparing the high-spin exchange energies. Near the van der Waals minimum of undecaplet Mn$\cdots$Mn, the single exchange approximation reproduces about 90\% of the full exchange.
In order to assess the performance of first-order SF-SAPT, we have computed the $J_{AB}$ parameter and compared it with previous literature data in panel (c) of Fig.~\ref{fig:mn2}.
In the literature, the $J_{AB}$ parameter is usually given at the minimum separation for the high-spin complex (in case of the Buchachenko et al.~\cite{Buchachenko:10} work, the full curve was provided).
Since the spin-flip splitting is obtained here at the ROHF level, it works remarkably well.
In particular, our results agree very well with the experimental result of Cheeseman et al. \cite{Cheeseman:90} (although we realize this might be somewhat fortuitous).
Except for the value by Negodaev et al. \cite{Negodaev:08}, all presented values of $J_{AB}$ are smaller than the result derived from SF-SAPT.
\subsection{Phenalenyl dimer} \label{sec:phenalenyl}
The dimer of the doublet phenalenyl radical is an example of ``pancake bonding'', a strong interaction between $\pi$-stacked radicals that has gathered significant interest in recent years \cite{Preuss:14}.
The ground state of the phenalenyl dimer is a multireference singlet that exhibits pancake bonding with a binding energy of 11.5 kcal/mol, while its asymptotically degenerate triplet state exhibits only a van der Waals minimum with a depth
of 3.6 kcal/mol \cite{Cui:14}. At the interplanar separation of 3.104 \AA\ corresponding to the pancake bonded minimum depicted in Fig.~\ref{fig:phen}, the reference singlet-triplet splitting, computed using the high-level multireference
averaged quadratic coupled cluster theory (MR-AQCC) approach \cite{Szalay:93} in Ref.~\cite{Cui:14} (with a (2,2) active space and the
6-31G(d) basis set), amounts to 17.2 kcal/mol.
It should be noted that an accurate description of pancake bonding remains a challenge to many quantum-chemical approaches, most notably those based on density functional theory \cite{Mou:17}.
In this section, we will examine how well the singlet-triplet splitting in the phenalenyl dimer is recovered by the first-order SF-SAPT approach.
In this way, we expect to find out whether (1) the simple first-order treatment of the splitting remains valid for such a strong intermolecular attraction, (2) our AO implementation in {\sc psi4} is efficient enough to enable $E^{(10)}_{\rm exch}$ calculations for large complexes, and (3) the DF approximation is just as accurate as for conventional high-spin SAPT. For this purpose, we select the pancake-bonded minimum geometry of the singlet state as established in Ref.~\cite{Cui:14},
vary the intermolecular separation $R$, and perform MO- and AO-based SF-SAPT calculations in a number of basis sets. We did verify that, in the absence of the DF approximation, the SF-SAPT exchange energies from the MO and AO formalisms are identical.
The singlet-triplet splittings obtained in the aug-cc-pVDZ basis set are presented in Fig.~\ref{fig:phen}. In addition to the first-order SF-SAPT calculations,
we perform CASSCF(2,2)/aug-cc-pVDZ computations across the same potential energy curve. At the pancake-bonded minimum,
the SAPT/aug-cc-pVDZ corrections $E^{(10)}_{\rm elst}$, $E^{(10)}_{\rm exch,diag}$, and
$E^{(10)}_{\rm exch,1flip}$ amount to $-19.9$, 50.6, and 4.3 kcal/mol, respectively. A basis set increase to aug-cc-pVTZ changes these values to
$-19.7$, 50.2, and 4.3 kcal/mol, respectively, confirming that the $E^{(10)}$ corrections converge quickly with the basis set size.
Thus, the singlet-triplet splitting from first-order SF-SAPT amounts to 8.6 kcal/mol or about half of the benchmark value.
On the other hand, the CASSCF method overestimates the benchmark splitting, giving a value of 23.4 kcal/mol. Fig.~\ref{fig:phen} shows that at larger
intermonomer separations the SF-SAPT and CASSCF splittings get closer to each other and both quantities exhibit the same long-range decay.
Thus, the underestimated splitting from SF-SAPT results likely from the single-exchange approximation applied in our calculations breaking down
for the very short intermolecular distances that are characteristic of pancake bonding.
The largest error from the DF approximation, as computed in a smaller, cc-pVDZ basis set, appears for $E^{(10)}_{\rm exch,diag}$ and amounts to 0.065 kcal/mol
(using the cc-pVDZ/JKFIT auxiliary basis), thus, the DF approximation is fully adequate for the first-order
SF-SAPT approach.
The cost of AO-based SF-SAPT expressions is dominated by the evaluation of generalized JK matrices, of which a total of 8 are required. As such, SF-SAPT
takes typically about as long as 4 ROHF iterations, each of them requiring two JK builds. For the phenalenyl dimer on a six-core i7-5930K Intel CPU,
the density-fitted evaluation of Eqs.~(\ref{aodiag})--(\ref{aoflip}) takes 26 seconds for the cc-pVDZ basis set (674 functions) and 187 seconds for the
cc-pVTZ basis set (1556 functions), making this method tractable for very large molecules. As the algorithm uses the native JK builders of the {\sc psi4}
code \cite{Parrish:17}, any improvements made to these core {\sc psi4} objects will also be extended to the SF-SAPT calculations.
\section{Summary}
We derived and implemented the molecular-orbital and atomic-orbital formulas for the first-order SAPT
exchange energy for two high-spin open-shell species (described by their ROHF determinants) combined to give
an arbitrary spin state of the complex (the previously existing
open-shell SAPT approaches \cite{Zuchowski:08,Hapka:12} were restricted to the high-spin state of the complex
except for a few system-specific studies). Within the single exchange
approximation, the resulting exchange energies for different asymptotically degenerate multiplets are
linear combinations of two common terms: the diagonal exchange (common to all spin states) and the
spin-flip term (responsible for the multiplet splittings). Thus, the single exchange approximation (which
makes double- and higher-spin-flip terms vanish identically) is equivalent to the Heisenberg Hamiltonian model
where all splittings within an asymptotically degenerate set of multiplets are expressible by a single
parameter $J_{AB}$. This equivalence was proven long ago by Matsen {\em et al.} \cite{Matsen:71,Wormer:84}, however,
this work is the first one to give an explicit, general SAPT expression for the splitting parameter.
We investigated the behavior of the diagonal and spin-flip components of the first-order SAPT exchange
correction on a number of interatomic and intermolecular complexes.
For the Li$\cdots$H system, we compared the exchange energies with existing FCI-based SAPT
calculations~\cite{Patkowski:01} and found a very good agreement. In particular, ROHF-based first-order SF-SAPT
reproduces 71\%\ of the FCI exchange splitting in the van der Waals minimum. For several other diatomic complexes:
Li$\cdots$Li, Li$\cdots$N, and N$\cdots$N we compared the difference between $E^{(10)}_{\rm exch}$ for the highest
and lowest spin state to the splittings obtained with the CASSCF and MRCI methods. The $E^{(10)}_{\rm exch}$
results agree very well with CASSCF for a wide range of distances, from an infinite separation to roughly half
the distance between the high- and low-spin minima.
The power of the perturbative approach is particularly
impressive in the asymptotic region: the $E^{(10)}_{\rm exch}$ splitting very accurately reproduces
the CASSCF values and, unlike MRCI, ensures proper exponential decay.
For the oxygen dimer in four characteristic configurations, we compared the exchange splittings with
literature data. Again, the first-order SAPT predictions agree very well with the MRCI and CASSCF results
of Bartolomei et al. \cite{Bartolomei:08}. A particularly interesting test of our new theory was the manganese
dimer, which poses a great challenge for supermolecular calculations with a variety of multireference methods.
The Heisenberg exchange coupling constant $J_{AB}$ derived from
$E^{(10)}_{\rm exch}$ compares qualitatively
well with literature data and predicts small spin splitting due to screening from the outermost doubly
occupied ($4s^2$) shell. However, it should be noted that for such a challenging system the overall agreement
between existing supermolecular data is far from satisfactory. Interestingly, first-order spin-flip SAPT
predicts the value of $J_{AB}$ within the experimental bounds. Finally, a highly efficient AO-based implementation
(with density fitting) of $E^{(10)}_{\rm exch}$ allowed us to apply SF-SAPT to a larger system, the
phenalenyl dimer. In that case, we obtained a qualitative agreement with reference CASSCF calculations.
The agreement improves at larger separations which might be attributed to the importance of terms beyond the
$S^2$ approximation.
The new first-order exchange
corrections have been implemented into the development version of the {\sc psi4} code \cite{Parrish:17}
as well as the {\sc SAPT2012} package \cite{SAPT2012}.
The AO-based, density fitted SF-SAPT calculation is particularly efficient, leading to very fast
(and entirely single-reference) qualitatively correct estimates of the strength of multiplet splittings.
Our new development is merely the first step towards extending SAPT to arbitrary spin states of the
interacting complex.
However, even at the present level of theory, SF-SAPT could be highly useful as a complementary method for
transition metal complexes or potential energy surfaces near dissociation. The method introduced here can
be also used with Kohn-Sham orbitals, provided they were obtained from a density functional which yields
asymptotically correct exchange-correlation potentials \cite{Misquitta:05}.
The ideas presented here can be generalized to the second-order exchange-induction
and exchange-dispersion corrections as well as to a multireference, CASSCF-based description of the
noninteracting wavefunctions. Moreover, while the single exchange approximation implies that double-
and higher-spin-flip terms vanish, it is a different approximation than merely neglecting multiple
spin flips. In fact, going beyond the $S^2$ approach might be useful even for two interacting doublets
at short range (such as in the phenalenyl dimer) where there is only a single active spin on each monomer
to be flipped. The work in all of these directions is in progress in our groups.
\section*{Acknowledgments}
We are indebted to Bogumi{\l} Jeziorski for our stimulating discussions on the topic of this work.
We thank Massimiliano Bartolomei for sharing the oxygen dimer data and Alexiei Buchachenko for his comments
regarding the Mn$\cdots$Mn dimer. K.P. and D.G.A.S. were supported by the U.S. National Science Foundation
CAREER award CHE-1351978. P.S.Z. is grateful to the Polish
National Science Center (NCN), grant number 015/19/B/ST4/02707.
\section*{Appendix: Proof of Eq.~(\ref{clebsch})}
In this Appendix, we calculate the coefficient $Z(S_A,S_B,S)=\frac{c_1}{2c_0\sqrt{S_A S_B}}$ that appears
when inserting Eq.~(\ref{offdiag1final}) into Eq.~(\ref{e1excha}), proving Eq.~(\ref{clebsch}).
Recall that
\begin{equation}
c_0=\langle S\: (S_A-S_B) | S_A\: S_A\: S_B\: -S_B \rangle,
\end{equation}
\begin{equation}
c_1=\langle S\: (S_A-S_B) | S_A\: (S_A-1)\: S_B\: (-S_B+1) \rangle.
\end{equation}
In the following, we will extend our previous notation of spin raising and lowering operators
$\hat{S}_{+}$ and $\hat{S}_{-}$, stating explicitly the molecule (A or B) that the operator is acting on.
We start by noting that the action of the spin-flip operator $\hat{S}_{A-}\hat{S}_{B+}$ on the
initial product state $|S_A\: S_A\: S_B\: -S_B\rangle$ produces
\begin{eqnarray}
\lefteqn{\hat{S}_{A-}\hat{S}_{B+} |S_A \: S_A\: S_B\: -S_B \rangle} \nonumber \\ &&
= \sqrt{S_A(S_A+1) -S_A(S_A-1) }\sqrt{S_B(S_B+1) +S_B(1-S_B) } |S_A\: (S_A -1)\: S_B\: (- S_B+1) \rangle
\nonumber \\ && = 2\sqrt{S_AS_B}|S_A\: (S_A-1)\: S_B \: (-S_B+1) \rangle \label{eq:clebschproof1}
\end{eqnarray}
Since $\hat{S}^2=\hat{S}_A^2 + \hat{S}_B^2 + 2(\hat{S}_A)_z(\hat{S}_B)_z + \hat{S}_{A-}\hat{S}_{B+} + \hat{S}_{A+} \hat{S}_{B-}$,
one might rewrite the action of the operator $\hat{S}_{A-}\hat{S}_{B+}$ as
\begin{eqnarray}
\lefteqn{\hat{S}_{A-}\hat{S}_{B+} |S_A \: S_A\: S_B\: -S_B \rangle} \nonumber \\ && =
\left[\hat{S}^2-(\hat{S}_A^2 + \hat{S}_B^2 + 2(\hat{S}_A)_z(\hat{S}_B)_z + \hat{S}_{A+} \hat{S}_{B-})\right]
|S_A\: S_A\: S_B\: -S_B \rangle \nonumber \\ &&
= \left[\hat{S}^2-(S_A(S_A+1) + S_B(S_B+1) - 2S_A S_B + 0)\right] |S_A\: S_A\: S_B\: -S_B \rangle \label{eq:clebschproof2}
\end{eqnarray}
The last term in the above equation yields zero since the reverse spin-flip operator $\hat{S}_{A+}\hat{S}_{B-}$ cannot raise any more spins on A
or lower any more spins on B.
Now, by projecting Eqs.~(\ref{eq:clebschproof1}) and (\ref{eq:clebschproof2}) onto $\langle S\: (S_A-S_B) |$ and using the fact that
$\langle S\: (S_A-S_B) | \hat{S}^2 = S(S+1)\langle S\: (S_A-S_B) |$, one finds that
\begin{equation}
2c_1 \sqrt{S_AS_B} = \left[S(S+1) -S_A(S_A+1) - S_B(S_B+1) + 2S_A S_B\right] c_0
\end{equation}
which, after a simple rearrangement, gives the formula (\ref{clebsch}) for $Z(S_A,S_B,S)$.
|
\section{Error Analysis of the Schalkwijk-Kailath Protocol}
\label{appx:SK}
We reproduce for completeness the explicit error analysis of the Schalkwijk-Kailath protocol~\cite{schalkwijk1966coding} as given in~\cite{el2011network} but with our notation. Consider the vector $\mathbf{N} \equiv (N_0, \dots, N_n)$. This is clearly jointly Gaussian since $N_i$ are independent and Gaussian. We claim that $(N_0 , Y_1, \dots, Y_n)$ is also jointly Gaussian where $Y^n$ are mutually independent and $Y_i \sim \mathcal{N}(0,N_S+ \sigma^2)$.
\begin{proof}
We proceed by induction. For the base case, we first note that $Y_1 = \gamma_1 N_0 + N_1$, so we obtain $(N_0, Y_1 , N_2 , \dots N_n)$ by a linear transformation, so it's jointly Gaussian with $\mathbb{E}[Y_1] = 0$. Furthermore, $\mathbb{E}[Y_1^2] = \gamma_1^2 +\sigma^2 = N_S+\sigma^2$ since $N_0, N_1$ are independent.
Now we consider the inductive case. Assume $(N_0, Y_1, \dots , Y_k , N_{k+1}, \dots, N_n)$ is jointly Gaussian and $Y^k$ are mutually independent and $Y_i \sim N(0,N_S+\sigma^2)$. Then, $(N_0, Y_1, \dots , Y_k)$ is jointly Gaussian, so $\mathbb{E}[N_0 \vert Y^k]$, which is also the minimum mean squared error (MMSE) estimate, is affine in $Y^k$. Since $Y^k$ and $N_0$ both have mean zero, it is actually linear. Thus, $X_{k+1} = \gamma_{k+1} (N_0 - \mathbb{E}[N_0 \vert Y^k])$ is linear in $N_0, Y^k$. Hence, $(Y_1, \dots, Y_{k+1} = X_{k+1} +N_{k+1})$ is jointly Gaussian. By the orthogonality principle for linear MMSE estimates, $X_{k+1}$ is independent of $Y_i$ for $i \in [1:k]$. Furthermore, $N_{k+1}$ is also independent of $Y_i$. Since $Y^k$ are mutually independent, we conclude that $(Y_1, \dots , Y_{k+1})$ is jointly Gaussian and uncorrelated. Thus, they are mutually independent. It is also clear that $\mathbb{E}[Y_{k+1}] = 0 $. Furthermore, $\mathbb{E}[Y_{k+1}^2] = N_S+\sigma^2$ since $\mathbb{E}[X_{k+1}^2] = N_S$ by construction and $X_{k+1},N_{k+1}$ are independent. Finally, $(N_0, Y_1, \dots, Y_{k+1} , N_{k+2}, \dots, N_n)$ is jointly Gaussian, so we're done.
\end{proof}
We next expand $I(N_0 ; Y^n)$ in two different ways. First,
\begin{align*}
I(N_0 ; Y^n) &= \sum_{i=1}^{n} I(N_0 ; Y_i \vert Y^{i-1})\\
&= \sum_{i=1}^{n} h(Y_i \vert Y^{i-1}) - h(Y_i \vert N_0, Y^{i-1})\\
&= \sum_{i=1}^{n} h(Y_i) - h(N_i \vert N_0, Y^{i-1}) \\
&= \sum_{i=1}^{n} h(Y_i) - h(N_i) \\
& = \frac{n}{2} \log\left( 1+ \frac{N_S}{\sigma^2}\right) \\
& = n P_H
\end{align*}
where the third equality holds because $Y^n$ are mutually independent and $Y_i$ is a function of $Y^{i-1}$, $N_0$, and $N_i$. $P_H$ is the capacity of the AWGN channel with power $N_S$ and noise $\sigma^2$ given in~\eqref{eq:shannonHom}. The second way to calculate this gives
\begin{align*}
I(N_0 ; Y^n) & = h(N_0) - h(N_0 \vert Y^n) \\
& = \frac{1}{2} \log(2\pi e \Var[N_0]) - \frac{1}{2} \log(2\pi e \Var[N_0 \vert Y^n])\\
& = -\frac{1}{2} \log\frac{\Var[N_0 \vert Y^n]}{\sigma^2},
\end{align*}
where the second equality follows from the fact that $N_0 \vert Y^n = y^n$ is Gaussian with variance independent of $y^n$ by the joint Gaussianity of $(N_0, Y_1, \dots, Y^n)$. Note that $\Var[N_0 \vert Y^n]$ is a random variable, but since it does not depend on $y^n$, we can identify it with the value it takes on almost surely.
\iffalse
The third equality follows since $N_0 | Y^n$ is Gaussian by joint Gaussianity of $(N_0, Y_1, \dots, Y_n)$. For the last equality, we used the fact that $\Var[N_0 \vert Y^n =y^n]$ does not depend on $y^n$, which allows us to identify $\Var[N_0 \vert Y^n]$, a random variable, by the value that it takes almost surely. We can see this fact by considering the random variable
\[V \equiv N_0 -\sum_{j=1}^{n} (P+f)^{-1} \mathbb{E}[N_0 Y_j] Y_j.\]
We show that $V$ is independent of $Y_i$ for all $i$:
\begin{align*}
\mathbb{E}[V Y_i] = \mathbb{E}[N_0 Y_i] - (P+f)^{-1} \mathbb{E}[N_0 Y_i] \mathbb{E}[Y_i^2] = 0.
\label{}
\end{align*}
Mutual independence of $V Y^n$ then follows from joint Gaussianity. Now,
\begin{equation*}
\Var[N_0 \vert Y^n = y^n] = \Var[V \vert Y^n =y^n],
\label{}
\end{equation*}
so the fact follows.
\fi
We conclude from the two ways to write the mutual information that $\Var[N_0 \vert Y^n] = \sigma^2 2^{-2n P_H}$. Now, $\Theta_n = Y_0 - \mathbb{E}(N_0 \vert Y^n)$ is Gaussian since it's linear in $Y_0^n$. Furthermore,
\begin{align*}
\Var{\Theta_n} & = \mathbb{E}[(N_0 - \mathbb{E}[N_0 \vert Y^n])^2] \\
& = \mathbb{E}[\mathbb{E}[(N_0 - \mathbb{E}[N_0 \vert Y^n])^2 \vert Y^n]] \\
& = \Var[N_0 \vert Y^n],
\label{}
\end{align*}
where the second equality follows by the law of iterated expectation. Thus, $\Theta_n \sim \mathcal{N}( \theta(m), \sigma^2 2^{-2n P_H})$.
We make a decoding error only if $\abs{\Theta_n-\theta(m)} > \sqrt{N_S} 2^{-nR}$. Hence, $p_e \le 2 Q(2^{n(P_H- R)} \sqrt{N_S/\sigma^2})$, where
\begin{equation*}
Q(x)\equiv \int_x^\infty \frac{1}{\sqrt{2\pi}} e^{-t^2/2} dt.
\label{}
\end{equation*}
Now, we know that for $x \ge 1$, $Q(x) \le \frac{1}{\sqrt{2\pi}} e^{-x^2/2}$~\cite{durrett2010probability}, so if $R< P_H$ and $n$ is large enough,
\begin{equation*}
p_e \le \sqrt{\frac{2}{\pi}} \exp\left( -\frac{2^{2n(P_H-R)}N_S}{2\sigma^2} \right)
\label{}
\end{equation*}
Note that we used the channel $n+1$ times, so the rate we achieve is actually $\frac{n}{n+1}R$.
\section{Discussion}
We have proven that with respect to the communication model of asymmetric feedback where Eve obtains a slightly noisy copy of the feedback on the initial round of the protocol, it is possible to achieve a much higher rate of private communication over a lossy thermal-noise bosonic wiretap channel, even one that is loss unlimited~\eqref{eq:shannonHom}.
It is worthwhile to make some observations about this result, as well as discuss its implications, limitations, and possible extensions.
\subsection{Extensions}
Eq.\ \eqref{eq:shannonHom} is achievable with a probability of decoding error that decays doubly exponentially with the number of channel uses. That is, the rate is achievable with an infinite error exponent. Indeed, this was the main thrust of the result in~\cite{schalkwijk1966coding} for the setting of feedback-assisted classical communication. It was then extended by~\cite{gallager2010variations}, where it was shown that we can actually achieve a probability of error that goes as an exponential tower with an order that increases linearly with the blocklength. This was shown to be optimal by a corresponding lower bound on the error probability. This lower bound would also apply in our case since it only concerns the part where Alice reliably sends a message to Bob. The only essential difference between the protocol in~\cite{gallager2010variations} and that of Schalkwijk and Kailath is that the former preserves the discrete structure of the message. In particular, in both, only the initial transmission contains information about the message. It is then straightforward to make the same extension to our setting.
That is, we can achieve~\eqref{eq:shannonHom} with a block probability of error
\begin{equation*}
p_e \le \frac{1}{^{f(n)} e}
\end{equation*}
for sufficiently large $n$, where $^n a$ denotes the tetration operation,
\begin{equation*}
f(n) \equiv \left\lfloor n[1-\varphi^{-1}(R)] - \frac{5(1-\varphi^{-1}(R))}{P_H - R} \right \rfloor,
\end{equation*}
and
\begin{equation*}
\varphi(\nu) \equiv \frac{\nu}{2} \log \left(1 + \frac{N_S}{\sigma^2 \nu} \right).
\end{equation*}
Note that this expression has minor differences compared to the bound in~\cite{gallager2010variations} because we use bits instead of nats. Some simplifications were also made for presentation.
Another possible extension follows from observing that the achieved rate~\eqref{eq:shannonHom} is not optimal. This is because it is known that squeezed state encoding can achieve higher rates over the thermal noise lossy bosonic channel with a homodyne receiver~\cite{guha2004classical}. However, the induced classical channel with squeezed states encoding is AWGN, so our argument trivially applies to this case. With optimal squeezed-state encoding, we can achieve a rate~\cite{guha2004classical}:
\begin{equation}
\label{eq:shannonsqueezedHom}
P_{\text{sq}} = \frac{1}{2}\log \left(1+ \frac{4N_S + 2 - f(\eta, N_S) + f(\eta, N_S)^{-1}}{((1-\eta)/\eta) + f(\eta, N_S)^{-1}}\right),
\end{equation}
where
\begin{equation*}
f(\eta, N_S) \equiv \frac{\eta\left[\left(1 + \frac{2(1-\eta)}{\eta} \left(\frac{1+\eta}{2\eta}+2N_S\right)\right)^{1/2}-1\right]}{1-\eta}.
\end{equation*}
When Bob uses a homodyne detection receiver, a squeezed state encoding is conjectured to be the optimal encoding for any Gaussian bosonic channel. If this conjecture is proven true, Eq.\ \eqref{eq:shannonsqueezedHom} would be optimal for asymmetric feedback-assisted private communication assuming that Bob makes a homodyne detection since the private capacity is trivially upper bounded by the classical capacity.
\subsection{Non-Gaussian Channels}
An apparent limitation of our result is that it strongly relies on the induced classical channel from Alice to Bob being an AWGN channel. This fact was heavily used in for instance the error analysis of the Schalkwijk-Kailath protocol. However, as noted in Remark 17.3 of~\cite{el2011network}, this exact protocol, that is, the protocol with the conditional expectations $\mathbb{E}[N_0 | Y^{i-1}]$ realized in the AWGN case, can be applied for channels with non-Gaussian additive noise. In fact, it would work for any affine channel whose output is given by
\begin{equation}
\label{eq:linearChannel}
a (X + N),
\end{equation}
where $X$ is the input, $N$ is some general additive noise independent of $X$, and $a \neq 0$ is some scaling. The error analysis follows by observing that the estimate $\Theta_n$ is linear in $X_0, N_0^n$ as shown in~\cite{el2011network}. Hence, in the affine channel case the variance of $\Theta_n/a$ has the exact same expression in terms of $X_0, N_0^n$ as in the AWGN case, and so $\Var[\Theta_n] = a^2 \Var[N] 2^{-2n C(\Var[N], N_S)} $, where
\begin{align}
\label{eq:nonGaussian}
C(\Var[N], N_S) & \equiv \frac{1}{2} \log \left( 1+ \frac{N_S}{\Var[N]} \right)
\end{align}
is the capacity of the AWGN channel with the corresponding noise statistics. Then, by Chebyshev's inequality,
\begin{align*}
p_e & \le \Prob\left\{ \abs{\Theta_n - \theta(m)} > \sqrt{N_S} 2^{-nR} \right\} \\
& \le a^2 2^{-2n(C(\Var[N], N_S) - R)} \frac{\Var[N]}{N_S}.
\end{align*}
This works for general affine channels, albeit at the cost of losing the doubly exponential decay. The privacy analysis can also be generalized since the argument did not make use of Gaussianity.
\iffalse
However, for general channels we might not have a bound on $S(E_0)$ so we instead argue in~\eqref{eq:privacyBound2}
\begin{align*}
I(M Y_0; E_0 | W_0 ) & \le I(M Y_0 W_0; E_0)\\
& \le I(M Y_0 W_0; B_0 E_0)\\
& \le I(M; B_0 E_0)\\
& \le I(M; A_0)\\
& \le C(\mathbbm{1}, N_S)\\
& = g(N_S).
\end{align*}
where $\mathbbm{1}$ is the identity channel.
\fi
Note that~\eqref{eq:nonGaussian} is in general less than the actual capacity of the channel, but it is still loss unlimited. In particular, our result applies to a general lossy bosonic wiretap channel where the additive noise (state of the ${\hat f}$ mode) can be any state, that is, it does not even need to be Gaussian. This is shown in~\cref{fig:nonGauss}.
\begin{figure}[h]
\centering
\includegraphics[width=0.3\textwidth]{nonGauss.pdf}
\caption{A non-Gaussian channel corresponding to a bosonic wiretap channel where the additive noise $\rho_N$ is not Gaussian.}
\label{fig:nonGauss}
\end{figure}
It is simple to see that the induced classical channel in this case is an affine channel. This subsumes a large class of physical optical channels.
However, it would be interesting to extend our result to a general quantum wiretap channel.
Unfortunately, there are channels for which it is not possible to induce a classical affine channel via a particular encoding and decoding. A trivial example is the constant channel which outputs a state regardless of the input. Admittedly, such a channel is useless for communication, so it is not surprising that our protocol fails there. For the specific case of coherent encoding and homodyne detection, we can find explicit nontrivial channels for which we do not induce an affine channel. For instance, consider the channel, a unitary self-Kerr interaction, which maps a coherent state to a cat state: $\ket{\alpha} \mapsto \frac{1}{\sqrt{2}} (\ket{\alpha} + \ket{-\alpha})$. It is not difficult to see that a homodyne measurement of a cat state will always have an expected value of zero. If the channel is affine: $X \mapsto a(X +N)$, then given $\mathbb{E}[ a(X+N)]=0$, either $a= 0$ or $\mathbb{E}[N] = - \mathbb{E}[X]$. If $a =0$, then the measurement result is always zero, which is clearly wrong. The latter possibility is also impossible since additive noise must be independent of the message. The conclusion follows. Thus, it is not clear whether our protocol can be extended to general quantum channels.
\subsection{Communication Model}
We assumed that Eve can eavesdrop on both the (noisy, quantum) forward and (classical) backward channels, but that her copy of the initial round of classical feedback has some noise. From a physical standpoint, this is a natural assumption since there is always \textit{some} noise in the physical layer. It might appear unnatural that the feedback error happens on the initial round, but this could be realized by Alice and Bob somehow knowing the time window in which the feedback noise might occur and starting the protocol during that window. Our constraint on feedback noise is reasonable as well, that any physical Eve-to-Bob channel should not have infinite capacity.
One might say that in order to make the model more physically relevant, there should be noise in the feedback channel to Alice as well. Indeed, it is this very noiselessness that is key to the result that we obtain. The channel to Alice is a noiseless CV classical channel, which has an~\textit{infinite} capacity, while we constrained the channel to Eve to have \textit{finite} capacity. It is this infinite difference that allows us to get a loss unlimited rate~\footnote{Note that this is an infinite difference in capacity between the channels from Bob to Alice and Eve, and we are trying to achieve forward private communication from Alice to Bob. Thus it is not clear a priori that this should be possible.}. Hence, it is unlikely that we can achieve loss unlimited rates if we allow noise in the channel to Alice as well. Indeed, it is known that linear schemes such as Schalkwijk-Kailath cannot achieve any positive rates in this regime~\cite{kim2007gaussian}.
However, this does not undermine our thesis that~\eqref{eq:shannonHom} indicates issues with the usual assumption of noiselessness of the public side channel. The fact that a classical noiseless CV channel has infinite capacity introduces significant unphysical features to the model and makes achievable rates strongly dependent on the model. For by just adding an infinitesimal amount of noise to the eavesdropper's copy of the initial round, one single step towards physical relevance, we obtain a loss unlimited rate instead of a loss limited one.
Now, it is somewhat queer that we can achieve a markedly different rate even if we allow Eve to have noiseless access to the feedback after the initial round. For an intuition for why this works, we observe that the Schalkwijk-Kailath protocol can be summarized as follows: Bob on the initial round receives the message with additive noise $N_0$ and subsequently learns the value of $N_0$ to exquisite precision. Since Eve has \emph{additional noise} to what Bob received on the initial round, even if Eve can also learn $N_0$ to exquisite precision, the accuracy with which she can recover the message is still limited by the additional noise. Indeed, we can even \textit{further alter} our communication model so as to give Eve $N_0$ exactly as done by~\eqref{eq:privacyBound}, thereby allowing her to completely simulate the protocol for rounds $i>0$, and our result would still hold. We can also give her a copy $X_0'$ of the quantum system~\footnote{Note that this is \textit{not} the classical random variable that is encoded into the quantum system, but the encoded quantum system itself.} $X_0$ that Alice sends on round $i=0$. This would simply replace $E_0$ in the privacy analysis by $X_0' E_0$, whose entropy we can again bound via subadditivity and then by bounds on photon number. Thus, we can alter our communication model so that Eve is \textit{strengthened}, and our result still holds. This reflects again that our protocol essentially depends on that initial round of feedback where there is an infinite difference in capacities.
At first glance the fact that we only need noise on the initial round appears to have implausibly strong implications. One might think our result holds in the usual feedback-assisted private communication model but where Alice and Bob have a pre-shared secret key which they can use to simulate the noise. For instance, by performing bitwise addition with the pre-shared key, they can simulate an additive noise channel with noise $S_0$ on the initial round of feedback. However, since any reasonable pre-shared key has finite Shannon entropy, $S_0$ is discrete. Then, the capacity of the feedback channel would be infinite, which precludes us from bounding the information leaked to Eve.
\iffalse
term $I(M Y_0 ; W_0)$ in~\eqref{eq:privacyBound2} scales with $n$. Indeed, as $M$ is also discrete,
\begin{align*}
& I(M Y_0; W_0) = I(M Y_0 ; Y_0 + S_0) = I(X_0 N_0 ; X_0 + N_0 + S_0) \\
& = I(X_0; X_0 + S_0) = H(X_0) + H(S_0) = nR - H(S_0)
\end{align*}
This is exactly the difference between the private information communicated and the preshared secret key consumed, which implies as much information is leaked as is communicated.
\fi
Hence it does not seem sufficient to have pre-shared secret key to obtain our result.
Lastly, it is worth looking at the closely related communication models of two way private communication, secret key agreement, and quantum key distribution. If we make the same alteration, that is, introduce noise in the classical feedback in the initial round, Bob can simply send an arbitrarily long locally generated random bit string to Alice during that round. Alice will then receive this key noiselessly, while Eve will only receive a noisy copy whose mutual information with the key is finite since the capacity of the noisy feedback channel is finite. Hence, in these models the infinite difference in capacities trivially leads to an infinite rate. This again shows the dependence of the rates to the model considered and suggests that in general we should more seriously address blatant unphysical features in the communication model.
\section{Introduction}
We consider the task of continuous variable (CV) private communication over a lossy bosonic channel with additive noise, with a quantum-limited adversary.
We consider a forward wiretap channel from Alice to Bob with an eavesdropper Eve. There is also a backward
public side channel from Bob to Alice. All parties, including Eve, have perfectly noiseless access to communication on this side channel. This immediately presents a problem. All communication channels in reality have a non-zero level of noise. With error-correction the noise can be reduced, but never completely removed. Therefore this model is clearly unphysical. Indeed, most of the information theoretic capacity analyses of private communication and secret-key generation have traditionally assumed such a zero-error feedback. Admittedly, the process of abstraction that allows a problem to be mathematically analyzed inevitably causes such unphysical features to emerge. However, we will find in this paper that this particular feature can be problematic in that ameliorating it can create a significant difference in achievable rates.
\iffalse
It is known~\cite{pirandola2017fundamental,takeoka2015fundamental} that for the pure-loss bosonic channel $\mathcal{N}$ with transmissivity $\eta$,
the feedback-assisted private capacity is upper bounded as: $P_B(\mathcal{N}) \le - \log(1-\eta)$~\footnote{Throughout this paper, logarithms are base 2.}. What is worthwhile to note is that this is a bound on the capacity in the limit of infinite input power, which in this case is photon number per mode. In other words, no matter how many photons per mode we are allowed for transmission, the capacity for private communication is some finite function of $\eta$, that is, it is a loss limited rate. For instance, for $\eta \ll 1$, $-\log(1-\eta) \approx 1.44 \eta$. If the pure-loss channel is an optical fiber, $\eta = e^{-\alpha L}$, where $L$ is the length of the fiber and $\alpha$ is some positive constant, and so the capacity reduces at least exponentially with the length.
Despite the promise of unconditional security, the capacity ultimately being proportional to $\eta$ is a highly pessimistic result and casts doubt on the practicality of quantum cryptographic protocols over optical fiber with feedback and whether they can compete with conventional protocols~\footnote{It would be a different matter if we had high-quality quantum repeaters, but this is still quite far from being realized.}. However, again recall that this is a result with respect to a particular choice of communication model, and we will argue that the result can change greatly by slightly ameliorating an unphysical feature in the model. We therefore look more closely at the model we are considering and try to understand what exactly is limiting the rate. As mentioned above, there are unphysical features surrounding the perfect noiselessness of the side channel. Hence, one plausible limiting factor is the assumption of Eve's perfectly noiseless access to the side channel.
Indeed, by slightly weakening only this assumption, ceteris paribus, we can achieve a much higher rate, even one that diverges with the input power: a loss unlimited rate.
\fi
More explicitly, we propose a protocol over a lossy bosonic channel and consider an alteration of the feedback-assisted private communication model in which Eve's copy of the public communication sent by Bob to Alice on just the initial round of the protocol is corrupted by a small amount of noise, a small step towards physical relevance. Eve obtains the remainder of the public communication noiselessly, and is otherwise quantum equipped, i.e., has perfect quantum memories and the ability to make arbitrary collective quantum measurements. We show that with such a seemingly inconsequential weakening of the eavesdropper, Alice and Bob can achieve a rate of $\frac{1}{2}\log(1+4\eta N_S)$ secure bits per mode~\footnote{Throughout this paper, logarithms are base 2.} using a simple laser-light modulation and homodyne detection, $N_S$ being the mean input photon number per mode. This is a loss unlimited rate, in sharp contrast to the loss limited upper bound known for the original model: $-\log(1-\eta)$~\cite{takeoka2015fundamental, pirandola2017fundamental}.
The exact same protocol also works if the channel has noise in addition to loss. We will argue that the noise can even be non-Gaussian.
A closely related classical result~\footnote{Note that in the classical case the feedback-assisted private capacity of the Gaussian wiretap channel also has a loss limited capacity. This can be seen for instance by upper bounding it by the secret key capacity and upper bounding that via Theorem 4 of~\cite{maurer1993secret}.} was proven in~\cite{gunduz2008secret}. However, their altered communication model assumes Eve's copy of the feedback is corrupted by \textit{Gaussian additive noise} and on \textit{every} round. Thus this is a more specialized and significant departure from the usual model. They do consider more general correlated noise between the different channels involved, but this can be integrated into our model as well.
In addition to the difference in model alteration, in the discussion section we will see that we can slightly \textit{strengthen} Eve from what is discussed above, by giving her copies of the initial transmission and noise, and still obtain our main result with the same protocol. Note that our results trivially apply to a classical additive white Gaussian noise (AWGN) channel and since adding more noise to Eve's feedback can only increase the achievable rate, implies the result in~\cite{gunduz2008secret}.
\section{Main Result}
Alice is to send a message to Bob over a bosonic (quantum) wiretap channel ${\cal N}$, which is to be kept private from Eve, by transmitting bosonic states on the forward channel $\cal N$ and using a backward noiseless classical feedback channel. Eve obtains the quantum output of the complement of the isometric extension of ${\cal N}$, and can eavesdrop on the feedback channel, but she obtains a noisy version of the classical feedback for the initial round. Denoting the classical feedback system as $Y$, for this round Eve obtains the output of a classical noisy channel $W(Y)$.
Furthermore, we assume that the capacity of the noisy channel $W(Y)$ from Bob to Eve, is finite for a finite input power, that is,
\begin{equation*}
C(W, P) \equiv \max_{Y: \mathbb{E}[Y^2] \le P} I(Y; W)
\end{equation*}
is finite for finite~\footnote{Note that the capacity is finite at all finite cost constraints iff it is finite at some finite cost constraint. The forward direction is trivial. For the reverse direction, assume for contradiction that there is some finite cost constraint at which the capacity is infinite. Then, by time-sharing we can conclude that is infinite at all finite cost constraints, a contradiction.} $P$. Note that this condition is very mild. For instance, if $W(Y) = Y + S$, with $S \sim {\rm Gaussian}(0,N_0)$, is AWGN, $N_0$ can be arbitrarily small as long as it is nonzero. Hence, we only need an iota of noise.
We shall refer to this setting as \emph{asymmetric feedback-assisted private communication.} We say asymmetric because the feedback is noiseless to Alice but noisy to Eve on the initial round. A diagram of the communication model for the initial round is shown in~\cref{fig:model}. $B'$ is some leftover system that Bob keeps. For subsequent rounds no noise is applied to Eve's copy of $Y$.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{model.pdf}
\caption{A diagram of the asymmetric feedback-assisted private communication model for the initial round of communication.}
\label{fig:model}
\end{figure}
The privacy criterion we use is the same as the one used in definitions of private communication in various classic papers~\cite{wyner1975wire,csiszar1978broadcast,maurer1993secret} and in the classical result~\cite{gunduz2008secret}; i.e., we assume the message $M$ is uniformly distributed and require:
\begin{equation*}
\lim_{k \to \infty} \frac{1}{k} I(M; \mathcal{E}) = 0,
\end{equation*}
where $k$ is the number of rounds and $\mathcal{E}$ is the aggregation of Eve's systems after all $k$ rounds. Note that this is in general weaker than the privacy criterion used in quantum Shannon theory analyses, for instance~\cite{wilde2013quantum}, which requires the trace distance between Eve's states and a constant state independent of the message to vanish. We leave open whether we can obtain our result with this stronger privacy criterion.
We will consider the case where the forward channel is a single mode thermal bosonic wiretap channel: denoting Alice's, Bob's and Eve's optical modes as $\hat a, \hat b,$ and $\hat e$, respectively, the channel is described by the Heisenberg evolutions:
\begin{equation}
\hat b = \sqrt{\eta} \hat a + \sqrt{1 - \eta} \hat f; \qquad \hat e = \sqrt{1-\eta} \hat a - \sqrt{\eta} \hat f,
\end{equation}
where $\hat f$ is the optical mode of the environment, which is in a thermal state ${\rho_{\rm th}}(0, n_{\rm th})$ with thermal number $n_\mathrm{th}$, and $\eta$ is the transmissivity. A thermal state is given by $\rho_\text{th}(\alpha, \bar n) \equiv \int_{\mathbb{C}}^{} \frac{d^2 \beta}{\pi \bar n} e^{-| \beta |^2 / \bar n} \state{\alpha + \beta}.$ We enforce an average photon number constraint on Alice's input over the $i \in [0:n]$ rounds:
$\frac{1}{n+1} \sum_{i=0}^{n} \tr[\hat a^\dagger \hat a \rho_{A_i}] \le N_S$.
\iffalse
We also assume that Bob performs a homodyne measurement on his output on every round. Eve, however, has no such restrictions and can do anything with the quantum states she obtains. The setting is given by
\begin{center}
\includegraphics[width=0.5\textwidth]{setting.jpg}
\end{center}
\fi
We now state our main theorem.
\begin{thm}
\label{thm:main}
The rate
\begin{equation}
\label{eq:shannonHom}
P_H = \frac{1}{2}\log \left (1+ \frac{N_S}{\sigma^2}\right),
\end{equation}
where $\sigma^2 \equiv \frac{1}{4\eta} + \frac{1-\eta}{2\eta} n_\mathrm{th}$, is achievable with mean photon per mode $N_S$ for asymmetric feedback-assisted private communication over the thermal bosonic wiretap channel, under the aforesaid model.
\iffalse satisfies the technical conditions
\begin{equation}
\label{eq:techCond}
\Var[S] < \infty, \quad h(S) > -\infty,
\end{equation}
where $h$ denotes the differential entropy.\fi
\end{thm}
|
\section{Introduction}\label{sec:introduction}
It is well known that a $\beta$-reduction sequence of
a simply typed $\lambda$-term can be extremely long.
Beckmann~\cite{Beckmann01} showed that, for any $k \geq 0$,
\[
\max\{ \beta(t) \mid t \text{ is a simply typed } \lambda\text{-term of order } k \text{ and size }
n \} = \Exp{k}{\Theta(n)}
\]
where $\beta(t)$ is the maximum length of the $\beta$-reduction sequences
of the term $t$, and $\Exp{k}{x}$ is defined by: $\Exp{0}{x} \triangleq x$ and $\Exp{k+1}{x}
\triangleq 2^{\Exp{k}{x}}$.
Indeed, the following order-$k$ term~\cite{Beckmann01}:
\newcommand\twice{\mathit{Twice}}
\newcommand\Twice[1]{\ctwok{#1}}
\[
(\Twice{k})^n \Twice{k-1} \cdots \Twice{2} (\lambda x^{\mathtt{o}}.a\,x\,x)
{((\lambda x^{\mathtt{o}}. x)c)},
\]
where $\Twice{j}$ is the twice function $\lambda f^{\Ttwice{j-1}}.\lambda x^{\Ttwice{j-2}}.f(f\,x)$
(with $\Ttwice{j}$ being the order-$j$ type defined by: $\Ttwice0=\mathtt{o}$ and $\Ttwice{j}=\Ttwice{j-1}\to\Ttwice{j-1}$),
has a $\beta$-reduction sequence of length $\Exp{k}{\Omega(n)}$.
Although the worst-case length of the longest $\beta$-reduction sequence is well known as above,
much is not known about the \emph{average-case} length
of the longest $\beta$-reduction sequence:
\emph{how often} does one encounter a term having a very long $\beta$-reduction sequence?
In other words, suppose we pick a simply-typed $\lambda$-term $t$ of order $k$ and
size $n$ \emph{randomly}; then
what is the probability that $t$ has a $\beta$-reduction sequence longer than a certain bound, like
$\Exp{k}{cn}$ (where $c$ is some constant)? One may expect that, although there exists
a term (such as the one above) whose reduction sequence is as long as $\Exp{k}{\Omega(n)}$,
such a term is rarely encountered.
In the present paper, we provide a partial answer to the above question, by showing that
almost every simply typed $\lambda$-term of order $k$ has a \(\beta\)-reduction
sequence as long as $(k-1)$-fold exponential in the term size, under
a certain assumption. More precisely, we shall
show:
\[
\lim_{n \rightarrow \infty} \frac{\card{\{ [t]_\alpha \in
\termsn{k}{\I}{\K}
\mid \beta(t) \geq \Exp{k-1}{n^p}
\}}}{\card{\termsn{k}{\I}{\K}}} = 1
\]
for some constant \(p>0\), where
$\termsn{k}{\I}{\K}$ is the set of ($\alpha$-equivalence
classes $[-]_\alpha$ of) simply-typed $\lambda$-terms
such that the term size is $n$, the order is up to $k$, the (internal)
arity is up to $\I \geq k$ and the number of variable names is up to
$\K$ (see the next section for the precise definition).
Related problems have been studied in the context of
the quantitative analysis of
\emph{untyped} $\lambda$-terms~\cite{SN,grygiel2013counting,Bendkowski}.
For example, David et al.~\cite{SN} have shown that
almost all untyped $\lambda$-terms are strongly normalizing, whereas the result
is opposite in the corresponding combinatory logic.
A more sophisticated analysis is, however, required in our case,
for considering only well-typed terms, and also for reasoning about
the \emph{length} of a reduction sequence instead of a qualitative property like strong normalization.
To prove our main result above, we have extended the infinite monkey theorem
(a.k.a. ``Borges's theorem''~\cite[p.61, Note I.35]{Flajolet}) to a parameterized version
for regular tree languages. The infinite monkey theorem states that for any word \(w\),
a sufficiently long word almost surely contains \(w\) as a subword (see Section~\ref{sec:mainresults} for
a more precise statement). Our extended theorem, roughly speaking, states that,
for any regular tree grammar \(\mathcal{G}\)
that satisfies a certain condition and any \emph{family} \((U_m)_{m \in \mathbb{N}}\) of \emph{trees}
(or tree contexts) generated by \(\mathcal{G}\) \new{such that $|U_m| = \mathrm{O}(m)$},
a sufficiently large tree \(T\) generated by \(\mathcal{G}\) almost
surely contains \(U_{\ceil{p\log{|T|}}}\) as a subtree
(where \(p\) is some positive constant).
Our main result is then obtained by preparing a regular tree grammar for simply-typed \(\lambda\)-terms,
and using as \(U_m\) a term having a very long \(\beta\)-reduction sequence, like
\((\Twice{k})^m \Twice{k-1} \cdots \Twice{2} (\lambda x^{\mathtt{o}}.a\,x\,x)
\new{((\lambda x^{\mathtt{o}}. x) c)}\) given above.
The extended infinite monkey theorem mentioned above may be of independent interest and applicable to
other problems.
Our work is a part of our long-term project on the quantitative
analysis of the complexity of higher-order model checking~\cite{Knapik02FOSSACS,Ong06LICS}. The
higher-order model checking asks
whether the (possibly infinite) tree generated by a ground-type term
of the $\lambda$Y-calculus (or, a higher-order recursion scheme)
satisfies a given regular property, and it is known that the problem is
$k$-EXPTIME complete for order-$k$ terms~\cite{Ong06LICS}. Despite the huge worst-case
complexity, practical model checkers~\cite{Kobayashi13JACM,Kobayashi13horsat,Ramsay14POPL}
have been built, which run fast for many typical inputs, and have successfully been
applied to automated verification of functional programs~\cite{Kobayashi09POPL,KSU11PLDI,Ong11POPL,SUK13PEPM}.
The project aims to provide a theoretical justification for it, by studying \emph{how many}
inputs actually suffer from the worst-case complexity.
Since the problem appears to be hard due to recursion,
as an intermediate step towards the goal, we aimed to analyze the variant of the problem
considered by Terui~\cite{Terui12RTA}: given a term of the simply-typed $\lambda$-calculus (without recursion)
of type Bool, decide whether it evaluates to true or false (where Booleans are Church-encoded; see \cite{Terui12RTA}
for the precise definition). Terui has shown that even for the problem, the complexity is $k$-EXPTIME complete
for order-$(2k+2)$ terms. If, contrary to the result of the present paper,
the upper-bound of the lengths of $\beta$-reduction sequences \emph{were} small for almost every term,
then we could have concluded that the decision problem above is easily solvable for most of the inputs.
The result in the present paper does not necessarily provide a negative answer to the question above,
because one need not necessarily apply $\beta$-reductions to solve Terui's decision problem.
The present work may also shed some light on other problems on typed $\lambda$-calculi
with exponential or higher worst-case complexity. For example,
despite DEXPTIME-completeness of ML typability~\cite{Mairson90POPL,Kfoury90},
it is often said that the exponential behavior is rarely seen \emph{in practice}. That is,
however, based on only empirical studies.
A variation of our technique may be used to provide a theoretical justification (or possibly \emph{un}justification).
A preliminary version of this article appeared in Proceedings of FoSSaCS 2017~\cite{SAKT17FOSSACS}.
Compared with the conference version, we have strengthened the main result (from ``almost every \(\lambda\)-term
of size \(n\) has \(\beta\)-reduction sequence as long as \(\exp_{k-2}(n)\)'' to
``... as long as \(\exp_{k-1}(n^p)\)''), and added proofs. We have also generalized the argument in
the conference version to formalize the parameterized infinite monkey theorem for regular tree languages.
The rest of this paper is organized as follows.
\Sec{mainresults} states our main result formally.
\Sec{proofimf} proves the extended infinite monkey theorem, and
\Sec{proofbeta}
proves the main result.
\Sec{relatedwork} discusses related work, and
\Sec{conclusion} concludes this article.
\section{Main Results}\label{sec:mainresults}
This section states the main results of this article: the result on the quantitative analysis of
the length of \(\beta\)-reduction sequences of simply-typed \(\lambda\)-terms (Section~\ref{sec:mainresult})
and the parameterized infinite monkey theorem for
regular tree grammars (Section~\ref{sec:RegTreeGram}).
The latter is used in the proof of the former in Section~\ref{sec:proofbeta}.
In the course of giving the main results, we also introduce various notations used
in later sections; the notations are summarized in Appendix~\ref{sec:symbols}.
We assume some familiarity with the simply-typed \(\lambda\)-calculus and
regular tree grammars and omit to define some standard concepts (such as \(\beta\)-reduction);
readers who are not familiar with them
may wish to consult \cite{TypedLambda} about the simply-typed \(\lambda\)-calculus
and \cite{Saoudi92,tata2007} about regular tree grammars.
For a set $A$, we denote by $\card{A}$ the cardinality of $A$; by abuse of notation, we write \( \card{A} = \infty \) to mean that \( A \) is infinite.
For a sequence \(s\), we also denote by $\len{s}$ the length of $s$.
For a sequence \(s=a_1 a_2 \cdots a_{\len{s}}\) and \(i\le\len{s}\),
we write \(\prj{s}{i}\) for the \(i\)-th element \(a_i\). We write \(s_1\cdot s_2\) or just \(s_1s_2\)
for the concatenation of two sequences \(s_1\) and \(s_2\).
We write \(\NIL\) for the empty sequence.
We use \(\biguplus\) to denote the union of disjoint sets.
For a set $I$ and a family of sets $(A_i)_{i \in I}$, we define
\(\coprod_{i \in I} A_i \triangleq \biguplus_{i \in I} (\{i\} \times A_i) = \set{(i,a) \mid i\in I, a\in A_i}\).
For a map \(f\), we denote the domain and image of \(f\) by \(\mathop{\mathrm{Dom}}(f)\) and \(\image{f}\), respectively.
We denote by \(\log{x}\) the binary logarithm \(\log_{2}{x}\).
\nk{Omitted ``...and by \(\ln x\) the natural logarithm \(\log_{e}{x}\).'' since
it was used only in the proof of convergence, which has now been omitted.}
\subsection{The Main Result on Quantitative Analysis of the Length of \(\beta\)-Reductions}
\label{sec:mainresult}
The set of (\emph{simple}) \emph{types}, ranged over by \(\tau,\tau_1,\tau_2,\ldots\), is defined by the grammar:
\[\tau ::= \mathtt{o} \mid \tau_1 \rightarrow \tau_2.\]
Note that the set of types can also be generated by the following grammar:
\[\tau ::= \tau_1 \to\cdots\to \tau_k\to \mathtt{o}\]
where \(k \geq 0\); we sometimes use the latter grammar for inductive definitions of functions on types.
We also use \(\sigma\) as a metavariable for types.
\begin{rem}
We have only a single base type \(\mathtt{o}\) above. The main result
(Theorem~\ref{mainthm}) would not change even if there are
multiple base types.
\end{rem}
Let $V$ be a countably infinite set, which is ranged over by \(x, y, z
\, (\text{and } x', x_1, \text{etc.})\).
The set of \emph{$\lambda$-terms} (or \emph{terms}), ranged over by $t,t_1,t_2,\ldots$, is defined by:
\begin{eqnarray*}
t ::= x \mid \lambda \overline{x}^\tau\!. t \mid t_1 \, t_2\qquad\qquad
\overline{x} ::= x \mid {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,].
\end{eqnarray*}
We call elements of \(V \cup \set{{\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]}\) \emph{variables}, and use meta-variables
\(\overline{x}, \overline{y}, \overline{z} \, (\text{and } \overline{x}', \overline{x}_1, \text{etc.})\) for them.
We sometimes omit type annotations and just write $\lambda \overline{x}.t$ for $\lambda
\overline{x}^\tau\!. t$.
We call
the special variable ${\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]$ an \emph{unused variable}, which may be
bound by \(\lambda\), but must not occur in the body.
In our quantitative analysis below, we will count
the number of variable names occurring in a term, except \({\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]\).
For example, the term \(\lambda x.\lambda y.\lambda z.x\) in the standard syntax can be represented by \(\lambda x.\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,].\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,].x\),
and the number of used variables in the latter is counted as \(1\).
Terms of our syntax can be translated to usual $\lambda$-terms by regarding
elements in \(V\cup\set{{\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]}\) as usual variables.
Through this identification we define the notions of free variables, closed terms, and
\(\alpha\)-equivalence \(\aeq\).
The $\alpha$-equivalence class of a term $t$ is written as $[t]_\alpha$.
In this article, we distinguish between a term and its $\alpha$-equivalence class, and
we always use $[-]_\alpha$ explicitly.
For a term \(t\), we write \(\mathbf{FV}(t)\) for the set of all the free variables of \(t\).
For a term $t$,
we define the set $\V(t)$ of variables (except ${\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]$) in $t$ by:
\[
\V(x) \triangleq \{x\} \quad\V(\lambda x^\tau\!.t) \triangleq \{x\} \cup \V(t)
\quad
\V(\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^\tau\!.t) \triangleq \V(t) \quad \V(t_1 t_2) \triangleq \V(t_1) \cup \V(t_2).
\]
Note that neither \(\V(t)\) nor even \(\card{\V(t)}\) is preserved by \(\alpha\)-equivalence.
For example,
\(t= \lambda x. (\lambda y. y)(\lambda z. x)\)
and
\(t'=\lambda x. (\lambda x. x)(\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]. x)\)
are \(\alpha\)-equivalent, but \(\card{\V(t)}=3\) and \(\card{\V(t')}=1\).
We write \(\numofvars{(t)}\) for
\(\min_{t' \in [t]_\alpha}\!\! \card{\V(t')}\), i.e.,
the minimum number of variables required
to represent an \(\alpha\)-equivalent term.
For example, for \(t=\lambda x. (\lambda y. y)(\lambda z. x)\) above,
\(\numofvars{(t)}=1\),
because \(t\sim_\alpha t'\) and
\(\card{\V(t')}=1\).
A \emph{type environment} $\Gamma$ is a finite set of type bindings of the form
$x: \tau$ such that if $(x:\tau) ,(x:\tau') \in \Gamma$ then \(\tau = \tau'\);
sometimes we regard an environment also as a function.
When we write \(\Gamma_1\cup\Gamma_2\), we implicitly require that
$(x:\tau) \in\Gamma_1$ and $(x:\tau') \in \Gamma_2$ imply \(\tau = \tau'\),
so that \(\Gamma_1\cup\Gamma_2\) is well formed.
Note that \(({\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]:\tau)\) cannot belong to a type environment;
we do not need any type assumption for \({\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]\) since it does not
occur in terms.
The type judgment relation \(\Gamma \p t : \tau\) is inductively defined by the following typing rules.
\\[1pt]
\[
\begin{gathered}
\frac{}{x:\tau \p x : \tau}
\qquad
\frac{
\Gamma_1 \p t_1 : \sigma \ft \tau
\qquad
\Gamma_2 \p t_2 : \sigma
}{\Gamma_1 \cup \Gamma_2 \p t_1 t_2 : \tau}
\\[10pt]
\frac{
\Gamma' \p t:\tau
\qquad \Gamma'=\Gamma\text{ or }\Gamma' = \Gamma\cup\set{\overline{x}:\sigma}
\qquad \overline{x} \notin \mathop{\mathrm{Dom}}(\Gamma)
}{\Gamma \p \lambda \overline{x}^{\sigma}\!.t:\sigma \ft \tau}
\end{gathered}
\]
\\[5pt]
The type judgment relation is equivalent to the usual one for the simply-typed \(\lambda\)-calculus, except that a
type environment for a term may contain only variables that occur free in the term
(i.e., if \(\Gamma \p t :\tau\) then \(\mathop{\mathrm{Dom}}(\Gamma)=\mathbf{FV}(t)\)).
Note that if \(\Gamma \p t :\tau\) is derivable, then the derivation is unique.
\nk{I have omitted the terminology ``a \emph{\(\typing{\Gamma}{\tau}\)-term}''.}
Below we consider only well-typed $\lambda$-terms, i.e., those that are inhabitants of some typings.
\begin{defi}[size, order and internal arity of a term]
The \emph{size} of a term $t$, written $|t|$, is defined by:
\begin{eqnarray*}
|x| \triangleq 1 \;\;\;\;\;\; |\lambda \overline{x}^\tau \!. t| \triangleq |t| + 1 \;\;\;\;\;\;
|t_1 t_2| \triangleq |t_1| + |t_2| + 1.
\end{eqnarray*}
The \emph{order} and \emph{internal arity} of a type $\tau$, written
$\mathtt{ord}(\tau)$ and $\mathtt{iar}(\tau)$ respectively, are defined by:
\[
\begin{aligned}
&\mathtt{ord}(\tau_1 \rightarrow \cdots \rightarrow \tau_n \rightarrow \mathtt{o})
\triangleq \max(\{0\} \cup \{ \mathtt{ord}(\tau_i) + 1 \mid 1 \le i \le n \})
\\
&\mathtt{iar}(\tau_1 \rightarrow \cdots \rightarrow \tau_n \rightarrow \mathtt{o})
\triangleq \max(\{n\} \cup \{ \mathtt{iar}(\tau_i) \mid 1 \le i \le n \}) \end{aligned}
\]
where \(n \ge 0\). We denote by $\types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}$ the set of types $\{ \tau \mid \mathtt{ord}(\tau) \leq \delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,
\mathtt{iar}(\tau) \leq \I \}$.
For a judgment \(\Gamma \p t: \tau\),
we define
the \emph{order} and \emph{internal arity} of \(\Gamma \p t: \tau\), written
$\tord{\Gamma}{t}{\tau}$ and $\tar{\Gamma}{t}{\tau}$ respectively, by:
\[
\begin{aligned} &
\tord{\Gamma}{t}{\tau} \triangleq \max\set{ \mathtt{ord}(\tau') \mid (\Gamma' \p t' : \tau') \text{ occurs in } \td }
\\&
\tar{\Gamma}{t}{\tau} \triangleq \max\set{ \mathtt{iar}(\tau') \mid (\Gamma' \p t' : \tau') \text{ occurs in } \td }
\end{aligned}
\]
where $\td$ is the (unique) derivation tree for $\Gamma \p t: \tau$.
\end{defi}
Note that the notions of size, order, internal arity, and \(\beta(t)\)
(the maximum length of \(\beta\)-reduction sequences of \(t\), as defined in Section~\ref{sec:introduction})
are well-defined with respect to
\(\alpha\)-equivalence.
\begin{exa}
Recall the term \(\Twice{j} \triangleq \lambda f^{\Ttwice{j-1}}.\lambda x^{\Ttwice{j-2}}.f(f\,x) \) (with $\Ttwice{j}$ being the order-$j$ type defined by: $\Ttwice0=\mathtt{o}$ and $\Ttwice{j}=\Ttwice{j-1}\to\Ttwice{j-1}$) in Section~\ref{sec:introduction}.
For \(t=\Twice{3}\,\Twice{2}\,(\lambda x^{\mathtt{o}}.x)\,y\),
we have \(\tord{y\COL\mathtt{o}}{t}{\mathtt{o}} = \tar{y\COL\mathtt{o}}{t}{\mathtt{o}}=3\).
Note that the derivation for \(y\COL\mathtt{o}\p t:\mathtt{o}\) contains the type judgment \(\emptyset\p \Twice{3}:
\Ttwice{3}\), where \(\Ttwice{3} = ((\mathtt{o}\to\mathtt{o})\to(\mathtt{o}\to\mathtt{o}))\to(\mathtt{o}\to\mathtt{o})\to\mathtt{o}\to\mathtt{o}\), and
\(\mathtt{ord}(\Ttwice{3})=\mathtt{iar}(\Ttwice{3})=3\).
\end{exa}
We now define the sets of ($\alpha$-equivalence classes of) terms
with bounds on the order, the internal arity, and the number of variables.
\begin{defi}[terms with bounds on types and variables]
\label{df:set-of-terms}
Let $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,\I,\K \geq 0$ and $n \geq 1$ be integers.
For each $\Gamma$ and $\tau$,
\(\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) and
\(\termsng{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) are defined by:
\[
\begin{aligned}
\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} &\triangleq \{ [t]_\alpha \mid
\Gamma \p t: \tau,
\tord{\Gamma}{t}{\tau} \leq \delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#, \tar{\Gamma}{t}{\tau} \leq \I,
\numofvars{(t)}\le \K
\}
\\
\termsng{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} &\triangleq
\{ [t]_\alpha \in \termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \mid |t| = n \}.
\end{aligned}
\]
We also define:
\begin{equation*}
\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \triangleq \, \biguplus_{\mathclap{\tau \in
\types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}}}\, \termsg{\emptyset}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \qquad
\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \triangleq \{ [t]_\alpha \in \terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \mid
|t| = n \}.
\end{equation*}
\end{defi}
\noindent
Intuitively, \(\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\)
is the set of (the equivalence classes of)
terms of type \(\tau\) under \(\Gamma\), within
given bounds \(\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#\), \(\I\), and \(\K\) on the order, (internal) arity, and the number of variables respectively;
and \(\termsng{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) is the subset of
\(\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) consisting of terms of size \(n\).
The set \( \terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) consists of (the equivalence classes of)
all the closed (well-typed) terms, and
\(\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) consists of those of size \(n\).
For example,
\(t=[\lambda x^\mathtt{o}.\lambda y^\mathtt{o}.\lambda z^\mathtt{o}.x]_\alpha\) belongs to
\( \termsN{4}{1}{3}{1}\); note that
\(|t|=4\) and
\(\numofvars{\new{(}t\new{)}}
= \numofvars{(\lambda x^\mathtt{o}.\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,].\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,].x)}=1\).
We are now ready to state our main result, which is proved in Section~\ref{sec:proofbeta}.
\begin{thm}\label{mainthm}
For $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#, \I, \K \geq 2$ and $k = \min\{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,
\I\}$, there exists a real number $p > 0$ such that
\[
\lim_{n \rightarrow \infty} \frac{\card{\{ [t]_\alpha \in
\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}
\mid \beta(t) \geq \Exp{k-1}{n^p}
\}}}{\card{\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}} = 1.
\]
\end{thm}
\noindent
As we will see later (in Section~\ref{sec:proofbeta}), the
denominator \(\card{\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}\) is nonzero if \(n\) is sufficiently
large.
The theorem above says that if the order, internal arity, and the number of used variables are
bounded independently of term size, most of the simply-typed \(\lambda\)-terms of size \(n\) have
a very long \(\beta\)-reduction sequence, which is as long as $(k-1)$-fold exponential in \(n\).
\begin{rem}
Recall that, when \(k=\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#\), the worst-case length of \(\beta\)-reduction sequence is \(k\)-fold
exponential~\cite{Beckmann01}.
We do not know whether \(k-1\) can be replaced with \(k\) in the theorem above.
Note that in the above theorem, the order \(\delta\), the internal arity \(\I\) and the number \(\K\) of
variables are bounded above by a constant, \emph{independently of the term size} \(n\). Our proof of the theorem
(given in Section~\ref{sec:proofbeta}) makes use of this assumption to model the set of simply-typed \(\lambda\)-terms
as a regular tree language.
It is debatable whether our assumption is reasonable. A slight change of the assumption
may change the result, as is the case for strong normalization of untyped \(\lambda\)-terms~\cite{SN,Bendkowski}.
When \(\lambda\)-terms are viewed as models of functional programs,
our rationale behind the assumption is as follows.
The assumption that the size of types (hence also the order and the internal arity) is fixed
is sometimes assumed in the context of type-based program analysis~\cite{Heintze97}. The assumption on the number
of variables comes from the observation that a large program usually consists of a large number of
\emph{small} functions, and that the number of variables is bounded by the size of each function.
\end{rem}
\subsection{Regular Tree Grammars and Parameterized Infinite Monkey Theorem}
\label{sec:RegTreeGram}
To prove Theorem~\ref{mainthm} above,
we extend the well-known \emph{infinite monkey theorem} (a.k.a. ``Borges's
theorem''~\cite[p.61, Note I.35]{Flajolet}) to a parameterized version for regular tree grammars,
and apply it to the regular tree grammar that generates
the set of ($\alpha$-equivalence classes of) simply-typed
\(\lambda\)-terms. Since the extended infinite monkey theorem may be of independent interest,
we state it (as Theorem~\ref{thm:imt}) in this section as one of the main results. The theorem is proved in Section~\ref{sec:proofimf}.
We first recall some basic definitions for regular tree grammars
in Sections~\ref{sec:context} and \ref{sec:gram}, and then state
the parameterized infinite monkey theorem in Section~\ref{sec:pimf}.
\subsubsection{Trees and Tree Contexts}
\label{sec:context}
A \emph{ranked alphabet} $\Sigma$ is a map from a finite set of \emph{terminal} symbols
to the set of natural numbers. We use the metavariable \(a\) for a terminal,
and often write \(\Ta,\Tb,\Tc,\ldots\) for concrete terminal symbols.
For a terminal $a \in \mathop{\mathrm{Dom}}(\Sigma)$, we call $\Sigma(a)$ the \emph{rank}
of $a$.
A \emph{$\Sigma$-tree} is a tree constructed from terminals in $\Sigma$
according to their ranks:
$a(T_1, \ldots, T_{\Sigma(a)})$ is a $\Sigma$-tree if
$T_i$ is a $\Sigma$-tree for each $i \in \set{1,\ldots, \Sigma(a)}$.
Note that \(\Sigma(a)\) may be \(0\): $a(\,)$ is a $\Sigma$-tree if \(\Sigma(a)=0\).
We often write just \(a\) for \(a(\,)\).
For example, if \(\Sigma=\set{\Ta\mapsto 2, \Tb\mapsto 1, \Tc\mapsto 0}\),
then \(\Ta(\Tb(\Tc), \Tc)\) is a tree; see (the lefthand side of) Figure~\ref{fig:tree} for
a graphical illustration of the tree.
We use the meta-variable $T$ for trees.
The \emph{size} of $T$,
written $|T|$, is the number of occurrences of terminals in $T$.
We denote the set of all $\Sigma$-trees by $\TR{\Sigma}$,
and the set of all $\Sigma$-trees of size \(n\) by $\TRn{n}{\Sigma}$.
\begin{figure}
\begin{minipage}[t]{.4\textwidth}
\(
\Tree[ [ [].{$\Tc$} ].{$\Tb$} [ ].{$\Tc$} ].{$\Ta$}
\)
\end{minipage}
\begin{minipage}[t]{.4\textwidth}
\(
\Tree[ [ [].{$\hole$} [].{$\Tc$} ].{$\Ta$} [ [].{$\hole$} ].{$\Tb$} ].{$\Ta$}
\)
\end{minipage}
\caption{A tree \(\Ta(\Tb(\Tc), \Tc)\) (left) and a context
\(\Ta(\Ta(\hole, \Tc), \Tb(\hole))\) (right).}
\label{fig:tree}
\end{figure}
Before introducing grammars, we define tree contexts,
which
play an important role in our formalization and proof of the (parameterized) infinite
monkey theorem for regular tree languages.
The set of \emph{contexts} over a ranked alphabet $\Sigma$, ranged over by
$C$, is defined by:
\[
C ::= \hole \mid a(C_1, \ldots, C_{\Sigma(a)}).
\]
In other words, a context is a tree where the alphabet is extended
with the special nullary symbol \(\hole\). For example, \(\Ta(\Ta(\hole, \Tc), \Tb(\hole))\)
is a context over \(\set{\Ta\mapsto 2, \Tb\mapsto 1, \Tc\mapsto 0}\);
see (the \new{righthand} side of) Figure~\ref{fig:tree} for
a graphical illustration.
We write \(\cs{\Sigma}\) for the set of contexts over $\Sigma$.
For a context \(C\),
we write \(\hn{C}\) for the number of occurrences of \(\hole\) in $C$.
We call \(C\) a \emph{\(k\)-context}
if \(\hn{C}=k\),
and an \emph{\affine{context}} if \(\hn{C} \le 1\).
A 1-context is also called a \emph{\linear{context}}.
We use the metavariable \(S\) for \linear{contexts} and \(U\) for \affine{contexts}.
We write \(\hole_i\) for the
\(i\)-th hole occurrence (in the left-to-right order) of \(C\).
For contexts \(C\), \(C_1, \dots, C_{\hn{C}}\),
we write \( C[C_1, \dots, C_{\hn{C}}] \) for the context obtained by
replacing each \(\hole_i\) in $C$ with \(C_i\). For example, if
\(C=\Ta(\hole, \Ta(\hole,\hole))\),
\(C_1 = \Tb(\hole)\), \(C_2 = \Tc\), and \(C_3 = \hole\), then
\(C[C_1,C_2,C_3] = \Ta(\Tb(\hole), \Ta(\Tc,\hole))\);
see Figure~\ref{fig:cfill} for a graphical illustration.
Also, for contexts \(C\), \(C'\) and \(i \in \set{1,\dots,\hn{C}}\),
we write \(C[C']_i\) for the context obtained by
replacing \(\hole_i\) in $C$ with \(C'\). For example, for \(C\) and \(C_1\) above,
\(C[C_1]_2 = \Ta(\hole, \Ta(\Tb(\hole), \hole))\).
For contexts \(\ctr\) and \(\ctr'\),
we say that \(\ctr\) is a \emph{subcontext of} \(\ctr'\) and write $\ctr \preceq \ctr'$ if
there exist \(\ctr_0\), \(\ctr_1,\ldots,\ctr_{\hn{\ctr}}\), and \(1 \le i \le \hn{\ctr_0}\) such that \(\ctr' =
\ctr_0[\ctr[\ctr_1, \dots ,\ctr_{\hn{\ctr}}]]_i\). For example,
\(\ctr=\Ta(\Tb(\hole),\hole)\)
is a subcontext of \(\ctr'=\Ta(\hole, \Ta(\Tb(\hole), \Tc))\), because
\(\ctr'=\ctr_0[\ctr[\hole,\Tc]]_2\) for \(\ctr_0=\Ta(\hole,\hole)\); see also
Figure~\ref{fig:subcontext} for a graphical illustration.
The subcontext relation \(\preceq\) may be regarded as a generalization
of the subword relation.
In fact, a word \(w = a_1\cdots a_k\) can be viewed as a
\linear{context} \(w^\sharp = a_1(\cdots (a_k \hole)\cdots)\); and
if \(w\) is a subword of \(w'\), i.e., if
\(w'=w_1ww_2\), then \(w'^\sharp = w_1^\sharp[w^\sharp[w_2^\sharp]]\).
\begin{figure}
\begin{minipage}[t]{.2\textwidth}
\(C:\)\\
\(
\Tree[ [].{$\hole$} [ [].{$\hole$} [].{$\hole$} ].{$\Ta$} ].{$\Ta$}
\)
\end{minipage}
\begin{minipage}[t]{.2\textwidth}
\(C_1:\)\\
\(
\Tree[ [].{$\hole$} ].{$\Tb$}
\)
\end{minipage}
\begin{minipage}[t]{.2\textwidth}
\(C[C_1,C_2,C_3]:\)\\
\(
\Tree[ [ [].{$\hole$} ].{$\Tb$} [ [].{$\Tc$} [].{$\hole$} ].{$\Ta$} ].{$\Ta$}
\)
\end{minipage}
\caption{Context filling (where \(C_2=c\) and \(C_3=\hole\)).}
\label{fig:cfill}
\end{figure}
\begin{figure}
\includegraphics[scale=0.5]{contexts}
\caption{A graphical illustration of a subcontext. The part surrounded by
the rectangle is the subcontext \(\Ta(\Tb(\hole),\hole)\).}
\label{fig:subcontext}
\end{figure}
The \emph{size} of a context \(U\), written \(|U|\), is defined by:
\(|\hole| \triangleq 0\) and \(|a(\ctr_1,\dots,\ctr_{\Sigma(a)})| \triangleq 1 + |\ctr_1|+\dots+|\ctr_{\Sigma(a)}|\).
Note that \(\hole\) and 0-ary terminal \(a\) have different sizes: \(|\hole|=0\) but \(|a|=1\).
For a \(0\)-context \(\ctr\), \(|\ctr|\) coincides with the size of \(\ctr\) as a tree.
\subsubsection{Tree Grammars}
\label{sec:gram}
A \emph{regular tree grammar}~\cite{Saoudi92,tata2007} (\emph{grammar} for short) is a triple ${\mathcal G} = (\Sigma,
{\mathcal N}, {\mathcal R})$ where
(i) $\Sigma$ is a ranked alphabet;
(ii) ${\mathcal N}$ is a finite set of symbols called \emph{nonterminals};
(iii) ${\mathcal R}$ is a finite set of \emph{rewriting rules} of the form
$N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} C[N_1, \dots, N_{\hn{C}}]$ where
$C \in \cs{\Sigma}$ and $N, N_1, \dots, N_{\hn{C}} \in {\mathcal N}$.
We use the metavariable \(N\) for nonterminals.
We write \(\Sigma\cup{\mathcal N}\) for
the ranked alphabet \(\Sigma\cup\set{N \mapsto 0 \mid N\in {\mathcal N}}\) and often regard the right-hand-side of a rule as a \( (\Sigma\cup{\mathcal N}) \)-tree.
The rewriting relation \(\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{\mathcal G}\) on \(\TR{\Sigma \cup {\mathcal N}}\) is inductively defined by the following rules:
\\
\begin{align*}
\frac{(N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} C[N_1, \dots, N_{\hn{C}}]) \in {\mathcal R}}
{N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{\mathcal G} C[N_1, \dots, N_{\hn{C}}]}
\qquad
\frac{T_i \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{\mathcal G} T'_i}{a(T_1,\dots,T_k)
\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{\mathcal G} a(T_1,\dots,T_{i-1},T'_i,T_{i+1},\dots,T_k)}
\end{align*}
\\[4pt]
We write \( \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*_{{\mathcal G}} \) for the reflexive and transitive closure of \( \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{{\mathcal G}} \).
For a tree grammar ${\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R})$ and a nonterminal
$N \in {\mathcal N}$, \emph{the language $\inhav{{\mathcal G},N}$ of} $N$ is defined by
$\inhav{{\mathcal G},N} \triangleq \{ T \in \TR{\Sigma} \mid N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*_{\mathcal G} T \}$.
We also define
$\inhavn{{\mathcal G},N} \triangleq \{ T \in \inhav{{\mathcal G},N} \mid |T| = n\}$.
If \({\mathcal G}\) is clear from the context, we often omit \({\mathcal G}\) and
just write \(T \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T'\), \(\inhav{N}\), and \(\inhavn{N}\) for
\(T \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*_{\mathcal G} T'\), \(\inhav{{\mathcal G},N}\), and \(\inhavn{{\mathcal G},N}\), respectively.
We are interested in, not only trees, but also contexts generated by a grammar.
Let \({\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R})\) be a grammar.
The set of \emph{\ctyping}s of \({\mathcal G}\), ranged over by $\f$,
is defined by:
\[
\f ::= N_1\cdots N_k\Ar N
\]
where \(k \in \mathbb{N}\) and $N, N_1, \ldots, N_k \in {\mathcal N}$. Intuitively, \(N_1\cdots N_k\Ar N\)
denotes the type of \(k\)-contexts \(C\) such that \(N\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* C[N_1,\ldots,N_k]\).
Based on this intuition, we define
the sets \(\pinhav{{\mathcal G},\f}\)
and \(\pinhavn{{\mathcal G},\f}\) of contexts by:
\begin{align*}
\pinhav{{\mathcal G},N_1 \cdots N_k \Ar N}
&\triangleq \set{ C \in \cs{\Sigma} \mid \hn{C} = k, N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^{*}_{{\mathcal G}}
C[N_1,\ldots,N_k]
}
\\
\pinhavn{{\mathcal G},\f} &\triangleq \set{C \in \pinhav{{\mathcal G},\f} \mid |C| = n}.
\end{align*}
We also write \(C\COL\f\) when \(C\in\pinhav{{\mathcal G},\f}\) (assuming that \({\mathcal G}\) is clear from the context).
Note that \(\pinhav{{\mathcal G},\Ar N} = \inhav{{\mathcal G},N}\), and
we identify a \ctyping{} \(\Ar N\) with the nonterminal \(N\).
We call an element in \(\pinhav{{\mathcal G},\f}\) a \emph{\(\f\)-context},
and also a \emph{\(\f\)-tree} if it is a tree.
It is clear that,
if \(\ctr \in \pinhav{{\mathcal G},N_1 \cdots N_k \Ar N}\)
and \(\ctr_i \in \pinhav{{\mathcal G},N^i_1 \cdots N^i_{\ell_i} \Ar N_i}\) (\(i=1,\dots,k\)), then
\( \ctr[\ctr_1,\dots,\ctr_{k}]
\in \pinhav{{\mathcal G}, N^1_1 \cdots N^1_{\ell_1} \cdots N^k_1 \cdots N^k_{\ell_k} \Ar N}\).
We also define:
\begin{align*}
\inhav{{\mathcal G}} &\triangleq \, \bigcup_{\mathclap{N \in {\mathcal N}}}\, \pinhav{{\mathcal G},N}
&
\inhavn{{\mathcal G}} &\triangleq \{ T \in \inhav{{\mathcal G}} \mid |T| = n\}
\\
\ctx{{\mathcal G}} &\triangleq \, \bigcup_{\mathclap{N,N' \in {\mathcal N}}}\, \pinhav{{\mathcal G},N\Ar N'}
&
\ctxn{{\mathcal G}} &\triangleq \{ S \in \ctx{{\mathcal G}} \mid |S| = n\}
\\
\octx{{\mathcal G}} &\triangleq \inhav{{\mathcal G}} \cup \ctx{{\mathcal G}}
&
\octxn{{\mathcal G}} &\triangleq \set{U \in \octx{{\mathcal G}} \mid |U| = n}.
\end{align*}
\begin{exa}\label{ex:grammar}
Consider the grammar \({\mathcal G}_0 = (\set{\Ta\mapsto 2,\Tb\mapsto 1,\Tc\mapsto 0},\set{A_0,B_0},{\mathcal R}_0)\), where \({\mathcal R}_0\)
consists of:
\[
A_0 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Ta(B_0, B_0) \qquad A_0 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tc \qquad B_0 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tb(A_0).
\]
Then, \(\Ta(\hole,\hole)\in \pinhavnn{1}{{\mathcal G}_0,B_0\,B_0 \Ar A_0}\) and
\(\Ta(\Tb(\hole),\Tb(\Tc))\in \pinhavnn{4}{{\mathcal G}_0, A_0 \Ar A_0}\).
\end{exa}
This article focuses on grammars that additionally satisfy two properties called
\emph{strong connectivity} and \emph{unambiguity}.
We first define strong connectivity.
\begin{defi}[Strong Connectivity~\cite{Flajolet}\footnote{\label{foot:StrConApr}
In~\cite{Flajolet}, strong connectivity is defined for \emph{context-free specifications}. Our
definition is a straightforward adaptation of the definition for regular
tree grammars.
}]
Let ${\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R})$ be a regular tree grammar.
We say that \(N'\) is \emph{reachable} from \(N\) if
\(\pinhav{{\mathcal G}, N'\Ar N}\) is non-empty, i.e., if
there exists
a \linear{context} \( \mvlinctr\in \cs{\Sigma} \) such that \( N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* \mvlinctr[N'] \).
We say that ${\mathcal G}$ is \emph{strongly connected} if for any pair $N, N' \in {\mathcal N}$ of nonterminals,
\(N'\) is reachable from \(N\).
\end{defi}
\begin{rem}
There is another reasonable, but slightly weaker condition of reachability:
\( N' \) is \emph{reachable} from \( N \) if there exists
a \((\Sigma \cup {\mathcal N})\)-tree \(T\) such that \(N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T\) and $N'$ occurs in \(T\).
The two notions coincide if \( \inhav{{\mathcal G}, N} \neq \emptyset \) for every \( N \in {\mathcal N} \).
Furthermore this condition can be easily fulfilled by simply removing all the nonterminals \( N \)
with \( \inhav{{\mathcal G}, N} = \emptyset \) and the rules containing those nonterminals.
\end{rem}
\begin{exa}\label{ex:sc}
The grammar ${\mathcal G}_0$ in Example~\ref{ex:grammar} is strongly connected,
since $\Tb(\hole) \in \inhav{{\mathcal G}_0, A_0 \Ar B_0}$ and $\Ta(\hole, \Tb(\Tc)) \in
\inhav{{\mathcal G}_0, B_0 \Ar A_0}$.
Next, consider the grammar ${\mathcal G}_1 = (\set{ \Ta \mapsto 2, \Tb \mapsto 1, \Tc
\mapsto 0}, \set{A_1, B_1}, {\mathcal R}_1)$, where ${\mathcal R}_1$ consists of:
\[
A_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Ta(\Tc, A_1) \qquad A_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tb(B_1) \qquad B_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tb(B_1) \qquad B_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tc.
\]
${\mathcal G}_1$ is not strongly connected, since $\inhav{{\mathcal G}_1, A_1 \Ar B_1}
= \emptyset$.
\end{exa}
To define the unambiguity of a grammar, we define (the standard notion of) leftmost rewriting.
The \emph{leftmost rewriting relation} \( \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{{\mathcal G},\ell} \) is the restriction of the rewriting relation that allows only the leftmost occurrence of nonterminals
to be rewritten.
Given \( T \in \TR{\Sigma} \) and \( N \in {\mathcal N} \),
a \emph{leftmost rewriting sequence of \( T \) from \( N \)} is a sequence \( T_1, \dots, T_n \) of \( (\Sigma \cup {\mathcal N}) \)-trees such that
\[
N = T_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{{\mathcal G},\ell} T_2 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{{\mathcal G},\ell} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{{\mathcal G},\ell} T_n = T.
\]
It is easy to see that \( T \in \inhav{{\mathcal G}, N} \) if and only if there exists a leftmost rewriting sequence from \( N \) to \( T \).
\begin{defi}[Unambiguity]\label{def:unambiguity}
A grammar \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) is \emph{unambiguous} if, for each \( N \in {\mathcal N} \) and \( T \in \TR{\Sigma} \),
there exists
at most one leftmost rewriting sequence from \( N \) to \( T \).
\end{defi}
\subsubsection{Parameterized Infinite Monkey Theorem for Tree Grammars}
\label{sec:pimf}
We call a partial function \(f\) from \(\mathbb{N}\) to \(\mathbb{R}\) with infinite definitional domain
(i.e., \(\mathop{\mathrm{Dom}}(f)=\set{n\in\mathbb{N} \mid f(n)\mbox{ is defined}}\) is infinite)
a \emph{partial sequence}, and write \(f_n\) for \(f(n)\).
For a partial sequence \(f = (f_n)_{n\in \mathbb{N}}\), we define
\(\ds{f}{(-)}\) as the bijective monotonic function from \(\mathbb{N}\) to \(\mathop{\mathrm{Dom}}(f)\), and then we define:
\[
\limd_{n \rightarrow \infty} f_n \triangleq \lim_{n \rightarrow \infty} f_{\ds{f}{n}}.
\]
All the uses of \(\limd\) in this article are of the form:
\(\limd_{n\rightarrow \infty}\fr{b_n}{a_n}\)
where \((a_n)_{n \in \mathbb{N}}\) and \((b_n)_{n \in \mathbb{N}}\) are (non-partial) sequences (hence, \(\fr{b_n}{a_n}\) is
undefined only when \(a_n=0\)).
The following property on \(\limd\)
is straightforward, which we leave for the reader to check.
\begin{lem}
\label{lem:convSum}
If \(\limd\limits_{n \rightarrow \infty} \fr{b^{(i)}_n}{a^{(i)}_n} = 1\) for \(i=1,\dots,k\),
and if \(0 \le b^{(i)}_n \le a^{(i)}_n\) for \(i=1,\dots,k\) and for \(n \in \mathbb{N}\),
then \(\limd\limits_{n \rightarrow \infty} \fr{\sum_i b^{(i)}_n}{\sum_i a^{(i)}_n} = 1\).
\end{lem}
Now we are ready to state the second main result of this article.
\newcommand\thmstatementIMTcap{Parameterized Infinite Monkey Theorem for Regular Tree Languages}
\newcommand\thmstatementIMT{Let ${\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R})$ be an unambiguous and
\SC{}
regular tree grammar
such that \(\card{\inhav{{\mathcal G}}}=\infty\),
and \((S_n)_{n \in \mathbb{N}}\) be a family of \linear{contexts} in \(\ctx{{\mathcal G}}\)
such that $|S_n| = \mathrm{O}(n)$.
Then there exists a real constant \(\cn > 0\) such that
for any \(N \in {\mathcal N}\) the following equation holds:
\[
\limd_{n \rightarrow \infty} \frac{\card{\{ T \in \inhavn{{\mathcal G},N} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\inhavn{{\mathcal G},N}}} = 1.
\]
}
\begin{thm}[\thmstatementIMTcap]
\label{thm:imt}
\thmstatementIMT
\end{thm}
Intuitively, the equation above says that if \(n\) is large enough,
almost all the trees of size \(n\) generated from \(N\) contain
\(S_{\ceil{\cn \log n}}\) as a subcontext.
\begin{rem}
\label{rem:imf}
The standard infinite monkey theorem says that for any finite word \(w=a_1\cdots a_k\in A^*\), the probability that
a randomly chosen word of length \(n\) contains \(w\) as a subword converges to \(1\), i.e.,
\[\lim_{n\to\infty}\frac{\card{\set{w'\in A^n \mid \exists w_0,w_1\in
A^*.w'=w_0 w w_1 }}}{\card{A^n}} = 1.\]
This may be viewed as a special case of our parameterized infinite monkey theorem above,
where (i) the components of the grammar are given by
\(\Sigma = \set{a\mapsto 1 \mid a\in A}\cup \set{e {\, \mapsto 0}}\), \({\mathcal N} = \set{N}\), and
\({\mathcal R} = \set{N\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N)\mid a\in A}\cup \set{N\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} e}\), and (ii)
\(S_n= a_1(\cdots a_k(\hole)\cdots)\) independently of \(n\).
\end{rem}
Note that in the theorem above, we have used \(\limd\) rather than \(\lim\).
To avoid the use of \(\limd\), we need an additional condition on \({\mathcal G}\), called
aperiodicity.
\begin{defi}[Aperiodicity~\cite{Flajolet}]\label{def:aperiodicity}
Let \({\mathcal G}\) be a grammar.
For a nonterminal \(N\), \({\mathcal G}\) is called \(N\)-\emph{aperiodic} if
there exists \(n_0\) such that
\(\card{\inhavn{N}} > 0\) for any \(n\ge n_0\).
Further, \({\mathcal G}\) is called \emph{aperiodic} if
\({\mathcal G}\) is \(N\)-aperiodic for any nonterminal \(N\).
\end{defi}
\noindent
In Theorem~\ref{thm:imt},
if we further assume that \({\mathcal G}\) is \(N\)-aperiodic,
then \(\limd\limits_{n \rightarrow \infty}\) in the statement can be replaced with
\(\lim\limits_{n \rightarrow \infty}\).
In the rest of this section, we reformulate the theorem above in the form more convenient for
proving Theorem~\ref{mainthm}.
In Definition~\ref{df:set-of-terms},
\(\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) was defined
as the disjoint union of \((\termsng{\emptyset}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K})_{\tau \in \types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}}\).
To prove Theorem~\ref{mainthm}, we will construct a grammar \({\mathcal G}\) so that, for each \(\tau\),
\(\termsng{\emptyset}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \cong \inhav{{\mathcal G},N}\) holds for some nonterminal \(N\).
Thus, the following style of statement is more convenient, which is obtained as an immediate corollary of
Theorem~\ref{thm:imt}
and Lemma~\ref{lem:convSum}.
\begin{cor}
\label{cor:imt}
Let ${\mathcal G}$ be an unambiguous and
\SC{} regular tree grammar such that \(\card{\inhav{{\mathcal G}}}=\infty\),
and \((S_n)_{n \in \mathbb{N}}\) be a family of \linear{contexts} in \(\ctx{{\mathcal G}}\)
such that $|S_n| = \mathrm{O}(n)$.
Then there exists a real constant \(\cn > 0\) such that
for any non-empty set \(\nset\) of nonterminals of \({\mathcal G}\) the following equation holds:
\[
\limd_{n \rightarrow \infty} \frac{\card{\{ (N,T) \in \coprod_{N \in \nset} \inhavn{{\mathcal G},N} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\coprod_{N \in \nset} \inhavn{{\mathcal G},N}}} = 1.
\]
\end{cor}
We can also replace
a family of \linear{contexts} \((S_n)_n\) above
with that of trees \((T_n)_n\). We need only this corollary in the proof of Theorem~\ref{mainthm}.
\begin{cor}
\label{cor:imtTree}
Let ${\mathcal G}$ be an unambiguous and \SC{} regular tree grammar
such that \(\card{\inhav{{\mathcal G}}}=\infty\) and
there exists \(C \in \cup_{\f} \inhav{{\mathcal G},\f}\) with \(\hn{C}\ge2\).
Also, let \((T_n)_{n \in \mathbb{N}}\) be a family of trees in \(\inhav{{\mathcal G}}\)
such that $|T_n| = \mathrm{O}(n)$.
Then there exists a real constant \(\cn > 0\) such that
for any non-empty set \(\nset\) of nonterminals of \({\mathcal G}\) the following equation holds:
\[
\limd_{n \rightarrow \infty} \frac{\card{\{ (N,T) \in \coprod_{N \in \nset} \inhavn{{\mathcal G},N} \mid
T_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\coprod_{N \in \nset} \inhavn{{\mathcal G},N}}} = 1.
\]
\end{cor}
\begin{proof}
Let \({\mathcal G}=(\Sigma,{\mathcal N},{\mathcal R})\).
Let \(C \in \inhav{{\mathcal G},\f}\) be the context in the assumption and let \(\f\) be of the form \(N'_1\cdots
N'_k\Ar N'\), i.e., \(N'\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* C[N'_1,\ldots,N'_k]\), where \(k = \hn{C} \ge 2\).
By Corollary~\ref{cor:imt}, it suffices to construct a family of
\linear{contexts} \((S_n)_n\) such that (i) \(S_n\in\ctx{{\mathcal G}}\),
(ii) \(T_n\preceq S_n\), and (iii) $|S_n| = \mathrm{O}(n)$.
For each \(n\), there exists \(N_n\) such that
\(N_n\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T_n\).
Since \(\card{\inhav{{\mathcal G}}}=\infty\) and by the strong connectivity,
for each \(i \le k\) there exists \(T'_i\) such that
\(N'_i \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T'_i\).
For each \(n\), by the strong connectivity, there exists
\(S'_n \) such that \(N_1' \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* S'_n[N_n]\),
and since \({\mathcal N}\) is finite, we can choose \(S'_n\) so that the size is bounded above by a constant (that is independent of \(n\)). Let \(S_n\) be \(C[S'_n[T_n],\hole,T'_3,\dots,T'_k]\).
Then (i) \(S_n\in \inhav{{\mathcal G},N'_2\Ar \new{N'}}\subseteq \ctx{{\mathcal G}}\) because \(N'\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*
C[N'_1,\ldots,N'_k] \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* C[S'_n[T_n],N'_2,T'_3,\dots,T'_k]\),
(ii) \(T_n\preceq S_n\), and (iii) \(|S_n|=\mathrm{O}(n)\), as required.
\end{proof}
\begin{rem}
\label{rem:avoidLimDef}
In each of Corollaries~\ref{cor:imt} and~\ref{cor:imtTree},
if we further assume that \({\mathcal G}\) is \(N\)-aperiodic for any \(N \in \nset\),
then \(\limd\limits_{n \rightarrow \infty}\) in the statement can be replaced with
\(\lim\limits_{n \rightarrow \infty}\).
\end{rem}
\begin{rem}
On Theorem~\ref{mainthm} (and similarly on Corollaries~\ref{cor:imt} and~\ref{cor:imtTree}),
one might be interested in the following form of probability:
\[
\limd_{n \rightarrow \infty} \frac{\card{\uplus_{m \le n}\{ T \in \inhavnn{m}{{\mathcal G},N} \mid
S_{\ceil{\cn \log m}} \preceq T
\}}}{\card{\uplus_{m \le n}\inhavnn{m}{{\mathcal G},N}}} = 1,
\]
which discusses trees of size \emph{at most} \(n\) rather than \emph{exactly} \(n\).
Under the assumption of aperiodicity,
the above equation follows from Theorem~\ref{mainthm},
by the following \emph{Stolz-Ces\`aro theorem}~\cite[Theorem~1.22]{Marian}:
Let $(a_n)_{n \in \mathbb{N}}$ and $(b_n)_{n \in \mathbb{N}}$ be a sequence of real numbers and
assume $b_n > 0$ and \(\sum_{n=0}^{\infty} b_n = \infty\). Then
for any \(c \in [-\infty, +\infty]\),
\[
\lim_{n \rightarrow \infty} \frac{a_n}{b_n} = c
\qquad\text{implies}\qquad
\lim_{n \rightarrow \infty} \fr{\sum_{m\le n}a_m}{\sum_{m\le n}b_m} = c.
\]
(We also remark that the inverse implication also holds if there exists \(\gamma \in \mathbb{R} \setminus \{1\}\) such that
\(\lim_{n \rightarrow \infty} \fr{\sum_{m\le n}b_m}{\sum_{m\le n+1}b_m} = \gamma\)~\cite[Theorem~1.23]{Marian}.)
\end{rem}
\section{Proof of the Parameterized Infinite Monkey Theorem for Regular Tree Languages (Theorem~\ref{thm:imt})}\label{sec:proofimf}
Here we prove Theorem~\ref{thm:imt}.
In Section~\ref{sec:IMTforWords}, we first prove
a parameterized version\footnote{Although the parameterization is a simple extension, we are not aware of
literature that explicitly states this parameterized version.} of infinite monkey
theorem for \emph{words}, and explain how the proof for the word case can be extended to deal with
regular tree languages. The structure of the rest of the section is explained at the end of Section~\ref{sec:IMTforWords}.
\subsection{Proof for Word Case and Outline of this Section}\label{sec:IMTforWords}
Let $A$ be an alphabet, i.e.,\; a finite non-empty set of symbols.
For a word $w = a_1\cdots a_n$, we write $|w|$ for \(n\) and call it the \emph{size} of $w$.
As usual, we denote by $A^n$ the set of all words of length $n$ over $A$,
and by $A^*$ the set of all finite words over $A$: $A^* = \bigcup_{n
\geq 0} A^n$. For two words $w, w' \in A^*$, we say $w'$ is a
\emph{subword of $w$} and write $w' \sqsubseteq w$ if $w = w_1 w' w_2$ for
some words $w_1, w_2 \in A^*$.
The infinite monkey theorem
states that, for any word $w \in A^*$,
the probability that a randomly chosen word of size $n$ contains $w$ as
a subword tends to one if $n$ tends to infinity (recall Remark~\ref{rem:imf}).
The following theorem is
a parameterized version, where \(w\) may depend on \(n\).
It may also be viewed as a special case of Theorem~\ref{thm:imt}, where
\(\Sigma = \set{a\mapsto 1 \mid a\in A}\cup \set{e}\), \({\mathcal N} = \set{N}\),
\({\mathcal R} = \set{N\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N)\mid a\in A}\cup \set{N\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} e}\), and
\(S_n= a_1(\cdots a_k(\hole)\cdots)\) for each \(w_n = a_1\cdots a_k\).
\begin{prop}[Parameterized Infinite Monkey Theorem for Words]\label{prop:monkey}
Let $A$ be an alphabet and $(w_n)_n$ be a family of words over $A$ such
that $|w_n|= \mathrm{O}(n)$.
Then, there exists a real constant \(p>0\) such that we have:
\[
\lim_{n \rightarrow \infty} \fr{\card{\{ w \in A^n \mid w_{\ceil{p \log n}} \sqsubseteq w\}}}{\card{A^n}} = 1.
\]
\end{prop}
\begin{proof}\upshape
Let $\zps{n}$ be $1 - \card{\{ w \in A^n \mid w_{\ceil{p \log n}} \sqsubseteq w \}}/\card{A^n}$,
i.e., the probability that a word of size $n$ does \emph{not} contain $w_{\ceil{p \log n}}$.
By the assumption \(|w_n|=O(n)\),
there exists \(b>0\) such that \(|w_n| \le bn\) for sufficiently large \(n\).
Let \(0<q<1\) be an arbitrary real number, and we define
\(p \triangleq \frac{q}{2b\log{\card{A}}}\).
We write $s(n)$ for \(|w_{\ceil{p \log n}}|\) and $c(n)$ for
$\floor{n/s(n)}$.
Let \(n \in \mathbb{N}\).
Given a word $w = a_1\cdots a_n\in A^n$, let us decompose it to subwords of length $s(n)$ as follows.
\[
w = \underbrace{a_1 \cdots a_{s(n)}}_{\text{first subword}}
\cdots
\underbrace{a_{(c(n)-1) s(n)+1} \cdots a_{c(n)s(n)}}_{c(n)\text{-th subword}}
\underbrace{a_{c(n) s(n)+1} \cdots a_{n}}_{\text{remainder}}.
\]
Then,
\begin{align*}
0 \le \zps{n} &\leq \mbox{``the probability that none of the $i$-th subword is $w_{\ceil{p \log n}}$''}
\\ &=
\fr{\card{(A^{s(n)} \setminus \{w_{\ceil{p \log n}}\})^{c(n)}A^{n-c(n)s(n)}}}{\card{A^{n}}}
\\ &=
\left(\fr{\card{A^{s(n)} \setminus \{w_{\ceil{p \log n}}\}}}{\card{A^{s(n)}}}
\right)^{c(n)} =
\left(\fr{\card{A^{s(n)}} - 1}{\card{A^{s(n)}}}
\right)^{c(n)} =
\left(1-\fr{1}{\card{A}^{s(n)}}\right)^{c(n)}
\end{align*}
and we can show \(\lim_{n \rightarrow \infty} \left(1-\fr{1}{\card{A}^{s(n)}}\right)^{c(n)} = 0\)
as follows.
For sufficiently large \(n\),
we have \(s(n) \le b\ceil{p \log n} \le
2bp\log n = q\fr{\log n}{\log {\card{A}}} = q\log_{\card{A}}n\),
and hence
\[
\left(1-\fr{1}{\card{A}^{s(n)}}\right)^{c(n)}
\le \left(1-\fr{1}{n^q}\right)^{c(n)}
\le \left(1-\fr{1}{n^q}\right)^{\fr{n}{s(n)}-1}
\le \left(1-\fr{1}{n^q}\right)^{\fr{\log \card{A}}{q}\fr{n}{\log n}-1}.
\]
Also we have \(\lim_{n \rightarrow \infty} \left(1-\fr{1}{n^{q}}\right)^{\fr{n}{\log n}} = 0\) (which we leave for the reader to check);
therefore \(\lim_{n \rightarrow \infty} \zps{n} = 0\).
\end{proof}
The key observations in the above proof were:
\begin{itemize}
\item[(W1)] Each word \(w=a_1\cdots a_n\) can be decomposed to
\[(w_{\mathtt{rem}}', w'_1,\ldots,w'_{\new{c}(n)})\in
A^{n-c(n)s(n)}\times \Big(\prod_{i=1}^{c(n)} A^{s(n)}\Big) \]
where \(w_{\mathtt{rem}}'=a_{c(n) s(n)+1} \cdots a_{n}\) and \(w'_i = a_{(i-1)s(n)+1}\cdots a_{is(n)}\).
\item[(W2)]
The word decomposition above induces the following
decomposition of the set \(A^n\) of words of length \(n\):
\begin{align*}
A^n \ &\cong\ A^{n-c(n)s(n)}\times \Big(\prod_{i=1}^{c(n)} A^{s(n)}\Big)
\\
\ &\cong\
\coprod_{w \in A^{n-c(n)s(n)}} \prod_{i=1}^{c(n)} A^{s(n)}
\end{align*}
Here, \(\cong\) denotes the existence of a bijection.
\item[(W3)] Choose \(s(n)\) and \(c(n)\) so that
(i) \(A^{s(n)}\) contains at least one element that contains \(w_n\) as a subword
and (ii) \(c(n)\) is sufficiently large. By condition (i), the probability that
an element of \(A^n\) does not contain \(w\) as a subword can be bounded above
the probability that none of \(w'_i\, (i\in\set{1,\ldots,c(n)})\) contains \(w_n\) as a subword,
i.e., \((1-\frac{1}{\card{A^{s(n)}}})^{c(n)}\), which converges to \(0\) by condition (ii).
\end{itemize}
The proof of Theorem~\ref{thm:imt} in the rest of this section is based on similar observations:
\begin{itemize}
\item[(T1)] Each tree \(T\) of size \(n\) can be decomposed to
\[(E, U_1,\ldots,U_{c_E}),\]
where \(U_1,\ldots,U_{c_E}\) are (affine) subcontexts,
\(E\), called a \emph{second-order context} (which will be formally defined later),
is the ``remainder'' of \(T\) obtained by extracting \(U_1,\ldots,U_{c_E}\), and
\(c_E\) is a number that depends on \(E\).
For example, the tree on the lefthand side of Figure~\ref{fig:partitioning}
can be decomposed to the second-order context and affine contexts
shown on the righthand side. By substituting each affine context for
\(\Hhole{}{}\) in the preorder traversal order, we recover the tree on the lefthand side.
This decomposition of a tree may be regarded as a generalization of the word decomposition
above, (by viewing a word as an affine context), where the part \(E\)
corresponds to the remainder
\(a_{c(n) s(n)+1} \cdots a_{n}\) of the word decomposition.
\item[(T2)] The tree decomposition above induces
the decomposition of
the set \(\inhavn{{\mathcal G},N}\) in the following form:
\begin{align}
\inhavn{{\mathcal G},N} \cong \coprod_{E\in \mathcal{E}} \prod_{j=1}^{c_E} \mathcal{U}_{E,j}
\label{eq:bijection}
\end{align}
where
\(\mathcal{E}\) is a set of second-order contexts and
\(\mathcal{U}_{E,j}\) is a set of \affine{contexts}.
(At this point, the reader need not be concerned about
the exact definitions of \(\mathcal{E}\) and
\(\mathcal{U}_{E,j}\), which will be given later.)
\item[(T3)] \new{Design} the above decomposition so that
(i) each \(\mathcal{U}_{E,j}\) contains at least one element that
contains \(S_{\ceil{\cn \log n}}\) as a subcontext,
and (ii) \(c_E\) is sufficiently large.
By condition (i),
the probability that an element of \(\inhavn{{\mathcal G},N}\) does not contain \(S_{\ceil{\cn \log n}}\)
as a subcontext is bounded above by
the probability that none of \(U_i\,(i\in\set{1,\ldots,c_E})\)
contain \(S_{\ceil{\cn \log n}}\) as a subcontext,
which is further bounded above by
\(\displaystyle{\max_{E\in \mathcal{E}}\Big(\prod_{j=1}^{c_E}(1-\frac{1}{\card{\mathcal{U}_{E,j}}})\Big)}\).
The bound can be proved to converge to \(0\) by using condition (ii).
\end{itemize}
\begin{figure}[tbp]
\centering\includegraphics[scale=0.4,bb=0 0 842 391]{img/partitioning}
\caption{An example of tree decomposition. The parts
surrounded by rectangles
on the lefthand side show the extracted affine contexts, and the remaining part of
the tree is
the second-order tree context.}
\label{fig:partitioning}
\end{figure}
The rest of the section is organized as follows.
Before we jump to the decomposition of \(\inhavn{{\mathcal G},N}\), we
first present
a decomposition of \(\TRn{n}{\Sigma}\), i.e., the set of \(\Sigma\)-trees of size \(n\)
in Section~\ref{sec:decomposition}. The decomposition of
\(\TRn{n}{\Sigma}\) may be a special case of that of \(\inhavn{{\mathcal G},N}\), where
\({\mathcal G}\) generates all the trees in \(\TRn{n}{\Sigma}\).
For a technical convenience in extending the decomposition of
\(\TRn{n}{\Sigma}\) to that of \(\inhavn{{\mathcal G},N}\),
we normalize grammars to \emph{canonical form}
in Section~\ref{sec:canonical-grammar}. We then give the decomposition of
\(\inhavn{{\mathcal G},N}\) given in Equation (\ref{eq:bijection}) above,
and prove that it satisfies the required properties
in Sections~\ref{sec:rtl-decomposition} and
\ref{sec:adjusting}.
Finally, we prove Theorem~\ref{thm:imt} in Section~\ref{sec:proofimfmain}.
\subsection{Grammar-Independent Tree Decomposition}
\label{sec:decomposition}
In this subsection, we will define a decomposition
function $\closeddecomp$ (where $m$ is a parameter) that decomposes a tree $T$
into (i) a (sufficiently long) sequence $P=U_1\cdots U_k$ consisting of \affine{subcontexts} of size no less than \(m\), and (ii) a ``second-order'' context \(E\) (defined
shortly in Section~\ref{sec:second-order-context}), which is the remainder of extracting \(P\) from \(T\).
Recall Figure~\ref{fig:partitioning}, which illustrates how a tree is decomposed by
$\closeddecompm{3}$. Here, the symbols $\hhole$ in the second-order context on the
right-hand side represents the original position of each subcontext.
By filling the \(i\)-th occurrence (counted in the depth-first, left-to-right pre-order)
of \(\hhole\) with
the \(i\)-th \affine{context}, we can recover the original tree on the left hand side.
As formally stated later (in Corollary~\ref{cor:gindependent-tree-decomposition}),
the decomposition function \(\closeddecomp\) provides a witness for a bijection of the form
\[
\TRn{n}{\Sigma} \cong \coprod_{E\in \mathcal{E}} \prod_{j=1}^{c_E} \mathcal{U}_{E,j},
\]
a special case of Equation \eqref{eq:bijection} mentioned in Section~\ref{sec:IMTforWords}.
\subsubsection{Second-Order Contexts}
\label{sec:second-order-context}
We first define the notion of second-order contexts and operations on them.
The set of \emph{second-order contexts} over $\Sigma$,
ranged over by $E$, is defined by:
\[
E ::= \Hhole{k}{n}[E_1, \ldots, E_k] \;\mid\; a(E_1, \ldots,
E_{\Sigma(a)}) \ \ (a \in \mathop{\mathrm{Dom}}(\Sigma)).
\]
Intuitively, the second-order context is an expression having holes of
the form \(\Hhole{k}{n}\) (called \emph{second-order holes}), which should be filled with a \(k\)-context of size \( n \).
By filling all the second-order holes, we obtain a \(\Sigma\)-tree.
Note that
\(k\) may be \(0\). In the technical development below,
we only consider second-order holes \(\Hhole{k}{n}\) such that
\(k\) is \(0\) or \(1\).
We write \(\shn(E)\) for the number of the second-order holes in \(E\).
Note that $\Sigma$-trees can be regarded as second-order contexts \(E\) such that $\shn(E) = 0$, and vice versa.
For \(i \le \shn(E)\), we write \(\sh{E}{i}\) for the \(i\)-th second-order hole
(counted in the depth-first, left-to-right pre-order).
We define the \emph{size} \(|E|\) by:
\(|\hhole^n_{k}[E_1,\dots,E_k] | \triangleq n + |E_1|+\cdots+|E_k|\)
and
\(|a(E_1,\dots,E_{\Sigma(a)})| \triangleq |E_1|+\cdots+|E_k| + 1\).
Note that \(|E|\) includes the size of contexts to fill the second-order holes in \(E\).
\begin{exa}
\label{ex:decomp}
The second-order context on the right hand side of Figure~\ref{fig:partitioning}
is expressed as \(E = \Tb(\Hhole{1}{3}[\Ta(\Hhole{0}{3}, \Tb(\Hhole{0}{5}))])\), where \(\shn(E)=3\),
\(|E|=14\), \(\sh{E}{1}=\Hhole{1}{3}\), \(\sh{E}{2}=\Hhole{0}{3}\), and \(\sh{E}{3}=\Hhole{0}{5}\).
\end{exa}
Next we define the substitution operation on second-order contexts.
For a context \(C\) and a second-order hole \(\hhole^n_k\),
we write \(\cbf{C}{\hhole^n_k}\) if \(C\) is a \(k\)-context of size \( n \).
Given \(E\) and \(C\) such that \(\shn(E) \ge 1\) and
\(\cbf{C}{\sh{E}{1}}\), we write $E\Hfill{C}$ for the second-order context obtained by replacing the leftmost second-order hole of $E$ (i.e.,\; $\sh{E}{1}$)
with $C$ (and by interpreting the syntactical bracket \([-]\) as the substitution operation).
Formally, it is defined by induction on $E$ as follows:
\[
\begin{array}{lcl}
\left(\hhole^n_k[E_1, \ldots, E_k]\right)
\Hfill{C} &\triangleq&
C[E_1, \ldots, E_k]
\\
(a(E_1, \ldots, E_{\Sigma(a)})) \Hfill{C} &\triangleq&
a(E_1, \ldots, E_i \Hfill{C}, \ldots, E_{\Sigma(a)})\\&&\hfill
\text{ where } i = \min\{ j \mid \shn(E_j) \geq 1, 1 \leq j \leq \Sigma(a) \}.
\end{array}
\]
In the first clause, $C[E_1,\ldots,E_k]$ is the second-order context obtained by
replacing $\hole_i$ in $C$ with $E_i$ for each $i \leq k$.
Note that
we have \(|E\Hfill{C}| = |E|\) whenever \(E\Hfill{C}\) is well-defined,
i.e.,~if \(\cbf{C}{\sh{E}{1}}\) (cf. Lemma~\ref{lem:second-substitution-preserves-size} below).
We extend the substitution operation for a sequence of contexts.
We use metavariable \(P\) for sequences of contexts.
For $E$ and $P = C_1 C_2 \cdots C_{\shn(E)}$, we
write $\cbf{P}{E}$ if $\cbf{C_i}{\sh{E}{i}}$ for each $i \leq \shn(E)$.
Given \(E\) and a sequence of contexts \(P=C_1 C_2\cdots
C_\ell\) such that
\(\ell \le \shn(E)\) and \(\cbf{C_i}{\sh{E}{i}}\) for each \(i \le \ell\),
we define \(E\Hfill{P}\) by induction on \(P\):\tk{Is the partial application of length \( \ge 2 \) used anywhere? I think it's not.}
\begin{align*}
&E\Hfill{\epsilon} \triangleq E
&
&E\Hfill{C \cdot P} \triangleq (E\Hfill{C})\Hfill{P}
\end{align*}
Note that \(\shn(E\Hfill{C}) = \shn(E) - 1\), so if \(\cbf{P}{E}\)
then \(E\Hfill{P}\) is a tree.
\begin{exa}
\label{ex:decomp-cont}
Recall the second-order context \(E = \Tb(\Hhole{1}{3}[\Ta(\Hhole{0}{3}, \Tb(\Hhole{0}{5}))])\) in
Figure~\ref{fig:partitioning} and Example~\ref{ex:decomp}.
Let \(P\) be the sequence of \affine{contexts} given in Figure~\ref{fig:partitioning}:
\[
\Ta(\Tb(\hole),\Tc)\cdot \Tb(\Tb(\Tc)) \cdot \Ta(\Tb(\Tc), \Tb(\Tc)).
\]
Then
\[E\Hfill{\prj{P}{1}} = \Tb(\prj{P}{1}[\Ta(\Hhole03, \Tb(\Hhole05))]) = \Tb(\Ta(\Tb(\Ta(\Hhole{0}{3}, \Tb(\Hhole{0}{5}))), \Tc)),\]
and
\begin{align*}
E\Hfill{P} &= ((E\Hfill{\prj{P}{1}}])\Hfill{\prj{P}{2}})\Hfill{\prj{P}{3}}\\
&= \Tb(\Ta(\Tb(\Ta(\prj{P}{2}, \Tb(\prj{P}{3}))), \Tc)) =
\Tb(\Ta(\Tb(\Ta(\Tb(\Tb(\Tc)), \Tb(\Ta(\Tb(\Tc), \Tb(\Tc))))), \Tc)),
\end{align*}
which is the tree shown on
the left hand side of Figure~\ref{fig:partitioning}.
\end{exa}
Thanks to the size annotation, the substitution operation preserves the size of a second-order context.
\begin{lem}\label{lem:second-substitution-preserves-size}
Let \( E \) be a second-order context and \( P \) be a sequence of contexts such that \(P\COL E\).
Then \( |E| = |E\Hfill{P}| \).
\end{lem}
\begin{proof}
The proof is given by a straightforward induction on \( E \). In the case \(E=\Hhole{k}{n}[E_1, \ldots, E_k]\),
we use the fact \(|C[E_1,\ldots,E_k]| = |C|+|E_1|+\cdots +|E_k|\) (which can be shown by induction on \(C\)).
\end{proof}
\subsubsection{Grammar-Independent Decomposition Function}
\label{sec:gindependent-decomposition}
Now we define the decomposition function \(\closeddecomp\).
Let \(\Sigma\) be a ranked alphabet and \(m \ge 1\).
We define \(\mr_{\Sigma} \triangleq \max(\image{\Sigma})\), which is also written as \(\mr\) for short.
We always assume that \(\mr \ge 1\),
since it holds whenever \(\card{\inhav{{\mathcal G}}}=\infty\),
which is an assumption of Theorem~\ref{thm:imt}. We shall define the decomposition function \(\closeddecompp{\Sigma}\)
(we omit $\Sigma$ and just write \(\closeddecomp\) for \(\closeddecompp{\Sigma}\)below)
so that $\closeddecomp(T) = (E,P)$ where
(i) \(E\) is the second-order context, and (ii)
\(P\) is a sequence of \affine{contexts}, (iii) \(E\Hfill{P} = T\),
and (iv) \(m \le |\prj{P}{i}| \le \mr(m-1)+1\) for each \(i\in \set{1,\ldots,\card{P}}\).
The function \(\closeddecomp\) is defined as follows, using an auxiliary decomposition function
\(\decomp\) given below.
\[
\closeddecomp(T) \triangleq (U[E], P)\mbox{ where }(U,E,P)=\decomp(T).
\]
\noindent
The auxiliary decomposition function (just called ``decomposition function'' below) \(\decomp\)
traverses a given tree \(T\) in a bottom-up manner, extracts a sequence \(P\) of subcontexts,
and returns it along with a \linear{context} \(U\) and a second-order context \(E\);
\(U\) and \(E\) together represent
the ``remainder'' of extracting \(P\) from \(T\).
During the bottom-up traversal, the \(U\)-component for each subtree \(T'\) represents a context
containing the root of \(T'\); whether it is extracted as (a part of) a subcontext or becomes a part of the remainder will be decided later based on the surrounding context.
The \(E\)-component will stay as a part of the remainder during
the decomposition of the whole tree (unless the current subtree \(T'\) is too small, i.e., \(|T'|<m\)).
We define \(\decomp\) by:
\begin{itemize}
\item If \(|T|<m\), then \(\decomp(T) \triangleq(\hole, T, \NIL)\).
\item If \(|T|\geq m\), $T = a(T_1, \ldots, T_{\Sigma(a)})$, and
\(\decomp(T_i) = (U_i, E_i, P_i) \ (\mbox{for each }i \le \Sigma(a))\), then:
\begin{equation}
\label{eq:defOfDecomp}
\decomp(T) \triangleq
\left\{
\begin{aligned}
& (\hole,\; a\bigl(U_1[E_1], \dots, U_{\Sigma(a)}[E_{\Sigma(a)}] \bigr),\; P_1 \cdots P_{\Sigma(a)})\\
& \quad \mbox{if there exist $i,j$ such that \(1\leq i<j \le \Sigma(a)\) and \(|T_i|,|T_j|\geq m\)}\\
& (\hole,\; \hhole^n_1 [E_i],\; a(T_1, \ldots, U_i, \ldots, T_{\Sigma(a)}) \cdot P_i)\\
& \quad \mbox{if $|T_j|<m$ for every \(j\neq i\), $|T_i| \geq m$, and}\\
& \quad \mbox{\phantom{if} $ n \triangleq |a(T_1, \ldots, U_i, \ldots,
T_{\Sigma(a)})| \geq m$}\\ & (a(T_1, \ldots, U_i, \ldots, T_{\Sigma(a)}),\; E_i,\; P_i)\\
& \quad \mbox{if $|T_j|<m$ for every \(j\neq i\), $|T_i| \geq m$, and}\\
& \quad \mbox{\phantom{if} $|a(T_1, \ldots, U_i, \ldots,
T_{\Sigma(a)})| < m$} \\%& (\hole,\; \hhole^{n}_0,\; T)\\
& \quad \mbox{if $|T_i| < m$ for every $i \le \Sigma(a)$, and \( n \triangleq |T|\) }
\end{aligned}
\right.
\end{equation}
\end{itemize}
As defined above, the decomposition is carried out by case analysis on the size of a given tree.
If \(T\) is not large enough (i.e., \(|T|<m\)), then \(\decomp(T)\) returns
an empty sequence of contexts, while keeping \(T\) in the \(E\)-component.
If \(|T|\geq m\), then \(\decomp(T)\) returns a non-empty sequence of contexts, by case analysis on the
sizes of \(T\)'s subtrees. If there are more than one subtree whose size is no less than \(m\) (the first case above),
then \(\decomp(T)\) concatenates the sequences of contexts extracted from the subtrees, and returns the remainder
as the second-order context. If only one of the subtrees, say \(T_i\), is large enough (the second and third cases), then
it basically returns the sequence \(P_i\) extracted from \(T_i\); however, if the remaining part
\(a(T_1, \ldots, U_i, \ldots,
T_{\Sigma(a)})\) is also large enough, then it is added to the sequence (the second case).
If none of the subtrees is large enough (but \(T\) is large enough), then \(T\) is returned as the \(P\)-component
(the last case).
\begin{exa}
\label{ex:decomposition-function}
Recall Figure~\ref{fig:partitioning}. Let \(T_0\) be the tree on the left hand side.
For some of the subtrees of \(T_0\),
\(\decompm{3}\) can be calculated as follows.
\[
\begin{array}{l}
\decompm{3}(\Tb(\Tc))=(\hole, \Tb(\Tc), \epsilon)\\
\decompm{3}(\Tb(\Tb(\Tc))) = (\hole, \Hhole{0}{3}, \Tb(\Tb(\Tc)))\hfill \mbox{ (by the last case of~\eqref{eq:defOfDecomp})}\\
\decompm{3}(\Ta(\Tb(\Tc),\Tb(\Tc))) =
(\hole, \Hhole{0}{5}, \Ta(\Tb(\Tc), \Tb(\Tc))) \qquad \hfill\mbox{ (by the last case of~\eqref{eq:defOfDecomp})}\\
\decompm{3}(\Ta(\Tb(\Tb(\Tc)), \Tb(\cdots))) =
(\hole, \Ta(\Hhole{0}{3}, \Tb(\Hhole{0}{5})), \Tb(\Tb(\Tc))\cdot \Ta(\Tb(\Tc), \Tb(\Tc)))\\
\hfill \mbox{ (by the first case of~\eqref{eq:defOfDecomp})}\\
\decompm{3}(T_0) = (\Tb(\hole), \Hhole{1}{3}[\Ta(\Hhole{0}{3}, \Tb(\Hhole05))],
\Ta(\Tb(\hole),\Tc)\cdot \Tb(\Tb(\Tc)) \cdot \Ta(\Tb(\Tc), \Tb(\Tc)))\\
\hfill \mbox{ (by the third case of~\eqref{eq:defOfDecomp})}\\
\end{array}
\]
From \(\decompm{3}(T_0)\) above, we obtain:
\[\closeddecompm{3}(T_0) = \left(\Tb(\Hhole13[\Ta(\Hhole03, \Tb(\Hhole05))]),\ \Ta(\Tb(\hole),\Tc)\cdot \Tb(\Tb(\Tc)) \cdot \Ta(\Tb(\Tc), \Tb(\Tc))\right).\]
\end{exa}
\subsubsection{Properties of the Decomposition Function}
\label{sec:gindependent-decomposition-properties}
We summarize important properties of \(\decomp\) in this subsection.
We say that \anaffine{context} $U$ is \emph{good for $m$}
if
\(|U| \ge m\) and
$U$ is of the form \(a(U_1,\dots,U_{\Sigma(a)})\)
where \(|U_i| < m\) for each \(i \le \Sigma(a)\).
In other words, \(U\) is good if \(U\) is of an appropriate size:
it is large enough (i.e. \(|U|\ge m\)), and not too large (i.e. the size
of any proper subterm is less than \(m\)).
For example, \(\Ta(\Tb(\hole),\Tb(\Tc))\) is good for \(3\), but
neither \(\Tb(\Tb(\hole))\) nor \(\Ta(\Tb(\hole),\Tb(\Tb(\Tc)))\) is.
The following are basic properties of the auxiliary decomposition function
\(\decomp\).
The property (\ref{charphi:split}) says that the original tree can be
recovered by composing the elements obtained by the decomposition,
and the property (\ref{charphi:componentsize})
ensures that \(\closeddecomp(T)\) extracts only good contexts from \(T\).
\begin{lem}\label{lemma:charphi}
\tk{to do: check which property is used}
Let T be a tree. If $\decomp(T) = (U,E,P)$, then:
\begin{enumerate}
\item \label{charphi:split}
\(\cbf{P}{E}\),\, \(\hn{U}=1\), and \((U[E])\Hfill{P}=T\)
\item \label{charphi:garbagesize}
\(|U|<m\)
\item \label{charphi:componentsize}
For each \(i \in \set{1,\dots,\len{P}}\),
\(\prj{P}{i}\) is good for \(m\).
\item \label{charphi:size-of-e}\(|T| = |U[E]|\).
\end{enumerate}
\end{lem}
\begin{proof}
(\ref{charphi:split})--(\ref{charphi:componentsize}) follow by straightforward simultaneous induction on $|T|$.
(\ref{charphi:size-of-e}) follows from (\ref{charphi:split}) and
Lemma~\ref{lem:second-substitution-preserves-size}.
\end{proof}
The following lemma ensures that \(\shn(E)\) is sufficiently large whenever \(\closeddecomp(T)=(E,P)\).
Recall condition (ii) of (T3) in Section~\ref{sec:IMTforWords}; \(\shn(E)\) corresponds to \(c_E\).
\begin{lem}[Decomposition produces sufficiently many affine contexts]\label{lemma:partition}
For any tree $T$ and $m$ such that $1 \leq m \leq |T|$,
if $\decomp(T) = (U,E,P)$, then
\[
\len{P} \geq \frac{|T|}{2\mr m}.
\]
\end{lem}
\begin{proof}
Recall that \(\mr \triangleq \max(\image{\Sigma})\) and we assume \(\mr \ge 1\).
We show below that
\[
\begin{aligned}
|U| + 2\mr m\len{P}- \mr m \ge |T|
\end{aligned}
\]
by induction on \(T\).
Then it follows that
\[
2\mr m\len{P} \ge |T|+ \mr m - |U|>|T|+ \mr m-m \geq |T|
\]
(where the second inequality uses Lemma~\ref{lemma:charphi}(\ref{charphi:garbagesize})),
which implies \(\len{P} > \fr{|T|}{2\mr m}\) as required.
Since \(|T| \ge m\), $\decomp(T)$ is computed by~Equation~\eqref{eq:defOfDecomp},
on which we perform a case analysis.
Let $T = a(T_1, \dots, T_{\Sigma(a)})$ and \(\decomp(T_i) = (U_i, E_i, P_i) \ (i \le \Sigma(a))\).
\begin{itemize}
\item The first case: In this case, we have
\[
U = \hole \qquad P = P_{i_1} P_{i_2} \cdots P_{i_s}
\]
where $\{i_1, \dots, i_s\} = \{i \le\Sigma(a) \mid |T_i| \geq m \}$ and $s \geq 2$,
since if \(|T_i|<m\) then \(P_i = \NIL\).
Note that we have \(\mr \ge 2\) in this case.
Then,
\begin{align*}&
|U|+2\mr m\len{P}-\mr m
\\ =\ &
2\mr m \left(\textstyle\sum_{j \le s} \len{P_{i_j}}\right) -\mr m
\\ \ge\ &
\left(\textstyle\sum_{j \le s}
(|T_{i_j}| + \mr m - |U_{i_j}|)\right) - \mr m
\qquad(\because \text{by induction hypothesis for \(T_{i_j}\)})
\\ \ge\ &
\left(\textstyle\sum_{j \le s} (|T_{i_j}| + \mr m - (m-1))\right) -\mr m
\qquad(\because |U_{i_j}|\le m-1 \mbox{ by Lemma~\ref{lemma:charphi}(\ref{charphi:garbagesize})})
\\ =\ &
\left(\textstyle\sum_{j \le s} |T_{i_j}|\right) + s \mr m - s m + s -\mr m
\\ \ge\ &
\left(\textstyle\sum_{j \le s} |T_{i_j}|\right) + 2 \mr m - s m + s -\mr m -r +1
\qquad(\because \text{\(s \ge 2\) and \(\mr \ge 1\)})
\\ =\ &
\left(\textstyle\sum_{j \le s} |T_{i_j}|\right) + (r-s)(m-1)+1
\\ \ge\ &
\left(\textstyle\sum_{j \le s} |T_{i_j}|\right) + (\Sigma(a)-s)(m-1)+1
\qquad(\because r \ge \Sigma(a))
\\ \ge\ &
\left(\textstyle\sum_{j \le s} |T_{i_j}|\right) + (\textstyle\sum_{i\in \set{1,\ldots,\Sigma(a)}\setminus\set{i_1,\ldots,i_s}}|T_i|)+1\\ &
\qquad(\because m-1 \ge |T_i| \mbox{ for }i\in \set{1,\ldots,\Sigma(a)}\setminus\set{i_1,\ldots,i_s})
\\ =\ &
|T|
\end{align*}
as required.
\item The second case: In this case, we have
\[
U = \hole
\qquad
P = a(T_1, \ldots, U_i, \ldots, T_{\Sigma(a)}) P_i
\]
and \(|T_{j}| \ge m\) if and only if \(j=i\).
Also we have \(r \ge 1\).
Then,
\begin{align*}&
|U|+2\mr m\len{P}-\mr m
\\ =\ &
2\mr m(1+\len{P_i}) - \mr m
\\ =\ &
\mr m + 2\mr m \len{P_i}
\\ \ge\ &
\mr m + (|T_i| +\mr m -|U_i|)
\qquad(\because \text{by induction hypothesis for \(T_{i}\)})
\\ \ge\ &
|T_i| +\mr m -(m-1)
\qquad(\because \mr m \ge 0, |U_i| \le m-1\mbox{ by Lemma~\ref{lemma:charphi}(\ref{charphi:garbagesize})})
\\ \ge\ &
|T_i|+\mr m -\mr -m + 2
\qquad(\because \mr \ge 1)
\\ =\ &
|T_i|+(\mr-1)(m-1)+1
\\ \ge\ &
|T_i|+(\Sigma(a)-1)(m-1)+1 \qquad(\because r \ge \Sigma(a))
\\ \ge\ &
|T_i|+(\textstyle\sum_{j\in \set{1,\ldots,\Sigma(a)}\setminus\set{i}}|T_j|)+1
\\ =\ &
|T|
\end{align*}
as required.
\item The third case of~\eqref{eq:defOfDecomp}:
In this case, we have
\[
U = a(T_1, \ldots, U_i, \ldots, T_{\Sigma(a)})
\qquad
P = P_i
\]
and \(|T_{j}| \ge m\) if and only if \(j=i\).
Then,
\begin{align*}&
|U|+2\mr m\len{P}-\mr m
\\ =\ &
1+(\textstyle\sum_{j\neq i}|T_j|)+|U_i|+2\mr m \len{P_i} -\mr m
\\ \ge\ &
1+(\textstyle\sum_{j\neq i}|T_j|)+|T_i|
\qquad(\because \text{by induction hypothesis for \(T_{i}\)})
\\ =\ &
|T|
\end{align*}
as required.
\item The forth case of~\eqref{eq:defOfDecomp}:
In this case, we have
\[
U = \hole
\qquad
P = T.
\]
Then
\begin{align*}&
|U|+2\mr m\len{P}-\mr m
\\ =\ &
\mr m
\\ \ge\ &
\mr(m-1)+1
\qquad(\because r \ge 1) \\ \ge\ &
\Sigma(a)(m-1)+1
\qquad(\because r \ge \Sigma(a))
\\ \ge\ &
(\textstyle\sum_{j\in \set{1,\ldots,\Sigma(a)}}|T_j|)+1
\\ =\ &
|T|
\end{align*}
as required.
\qedhere
\end{itemize}
\end{proof}
\subsubsection{Decomposition of \(\TRn{n}{\Sigma}\)}
\label{sec:decomp-trees}
This subsection shows that
the decomposition function \(\closeddecomp\) above provides a witness for a bijection of the form
\[
\TRn{n}{\Sigma} \cong \coprod_{E\in \mathcal{E}} \prod_{j=1}^{c_{E}} \mathcal{U}_{E,\,j}.
\]
\asd{Here I removed a garbage: ``where \(\ell_i\) is sufficiently large for each \(i\).''}
We prepare some definitions to precisely state the bijection.
We define the set \(\mathcal{E}_n^m\) of second-order contexts and the set \(\ucomp\) of affine contexts
by:
\begin{align*}
\mathcal{E}_n^m &\triangleq \{ E \mid (E, P) = \closeddecomp(T) \text{ for some } T \in \TRn{n}{\Sigma} \text{ and } P \}\\
\ucomp &\triangleq \set{ U \mid U : \hhole_k^n, U \text{ is good for } m }.
\end{align*}
Intuitively, \(\mathcal{E}_n^m\) is the set of second-order contexts obtained by decomposing a tree of
size \(n\), and \(\ucomp\) is the set of good contexts that match the second-order context
\(\hhole_k^n\).
The bijection is then stated as the following lemma.
\begin{lem}
\label{cor:gindependent-tree-decomposition}
\begin{align}
\TRn{n}{\Sigma}\;\cong
\coprod_{E \in \mathcal{E}_n^m} \prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}].
\label{eq:bijection-for-trees}
\end{align}
\end{lem}
The rest of this subsection is devoted to a proof of the lemma above;
readers may wish to skip the rest of this subsection upon the first reading.
\begin{lem}
\label{lem:size-of-E-in-uframe}
If \(E\in \mathcal{E}_n^m\), then \(|E|=n\).
\end{lem}
\begin{proof}
Suppose \(E\in \mathcal{E}_n^m\). Then
\( (E, P) = \closeddecomp(T) \) { for some } \(T \in \TRn{n}{\Sigma}\) and \(P\).
By Lemma~\ref{lemma:charphi}\eqref{charphi:size-of-e},
\(|E|=|T|=n\).
\end{proof}
For a second-order context \(E\), and \(m \ge 1\), we define the set \(\mathcal{P}_E^m\) of sequences of affine contexts:
\begin{align*}
\mathcal{P}_E^m &\triangleq \set{ P \mid (E, P) = \closeddecomp(T) \text{ for some } T \in \TR{\Sigma} }.
\end{align*}
The set \(\mathcal{P}_E^m\) consists of sequences \(P\) of affine contexts that
match \(E\) and are obtained by the decomposition function \(\closeddecomp\).
In the rest of this subsection,
we prove the bijection in
Lemma~\ref{cor:gindependent-tree-decomposition}
in two steps.
We first show \(\TRn{n}{\Sigma}\cong \coprod_{E\in \mathcal{E}_n^m} \mathcal{P}_E^m\) (Lemma~\ref{lemma:union-sig},
called ``coproduct lemma''), and then show
\(\mathcal{P}_E^m = \prod_{j=1}^{\shn(E)} \ucomph{\sh{E}{j}}\) (Lemma~\ref{lemma:union-sig}, called ``product lemma'').
\begin{lem}[Coproduct Lemma (for Grammar-Independent Decomposition)]
\label{lemma:union-sig}
For any
$n \geq 1$ and \(m \ge 1\),
there exists a bijection
\begin{flalign*}
\TRn{n}{\Sigma}\cong \coprod_{E\in \mathcal{E}_n^m} \mathcal{P}_E^m
\end{flalign*}
that maps each element \((E,P)\) of the set
\(\coprod_{E\in \mathcal{E}_n^m} \mathcal{P}_E^m\) to \(E\Hfill{P}\in \TRn{n}{\Sigma}\).
\end{lem}
\begin{proof}
We define a function
\[
f:
\coprod_{E \in \mathcal{E}_n^m} \mathcal{P}_E^m
\longrightarrow
\TRn{n}{\Sigma}
\]
by \(f(E,P) \triangleq E\Hfill{P}\),
and a function
\[
g:
\TRn{n}{\Sigma}
\longrightarrow
\coprod_{E\in\mathcal{E}_n^m} \mathcal{P}_E^m
\]
by \(g(T) \triangleq \closeddecomp(T) \).
Let us check that these are functions into the codomains:
\begin{itemize}
\item \( f(E,P) \in \TRn{n}{\Sigma} \):
Since \(P \in \mathcal{P}_E^m\), there exists \( T \in \TR{\Sigma} \) such that
\( (E,P) = \closeddecomp(T) \).
By Lemma~\ref{lemma:charphi}(\ref{charphi:split}),
we have \(f(E,P) = E\Hfill{P}=T \in \TR{\Sigma}\).
By the condition \(E\in\mathcal{E}_n^m\) and by
Lemmas~\ref{lem:second-substitution-preserves-size} and \ref{lem:size-of-E-in-uframe},
\(|E\Hfill{P}|=|E|=n\). Thus, \(f(E,P)\in\TRn{n}{\Sigma}\) as required.
\item \( g(T) \in \coprod_{E \in \mathcal{E}_n^m} \mathcal{P}_E^m \):
Obvious from the definitions of \( \mathcal{E}_n^m \) and \( \mathcal{P}_E^m \).
\end{itemize}
We have \(f(g(T)) = T\) by Lemmas~\ref{lemma:charphi}(\ref{charphi:split}). Let \( (E,P) \in \coprod_{E \in \mathcal{E}_n^m} \mathcal{P}_E^m \).
By definition, there exists \( T \in \TR{\Sigma} \) such that \( (E,P) = \closeddecomp(T)\).
By using Lemmas~\ref{lem:second-substitution-preserves-size} and \ref{lem:size-of-E-in-uframe} again,
we have \(|T|=|E\Hfill{P}|=|E| = n\). Thus, \((E,P)=g(T)\).
Then
\[
g(f(E,P))
=g(f(g(T))
=g(T)
=(E,P).
\]
\end{proof}
It remains to show the product lemma:
\(\mathcal{P}_E^m = \prod_{j=1}^{\shn(E)} \ucomph{\sh{E}{j}}\).
To this end, we prove a few more properties about the auxiliary decomposition function \(\decomp\).
\begin{lem}\label{lem:size-of-E}
If \(|T|\geq m\) and \(\decomp(T)=(U,E,P)\), then \(|E|\geq m\).
\end{lem}
\begin{proof}
Straightforward induction on \(T\).
\end{proof}
The following lemma states that, given $\decomp(U[T]) = (U,E,P)$, $P$ is
determined only by $T$ ($U$ does not matter); this is because the decomposition is
performed in a bottom-up manner.
\begin{lem}
\label{lemma:ehole}
For $m \geq 1$, \(E\), \(P\), \(T\), and a \linear{context} \(U\) with {\(|T|\ge m\)} and \( |U|<m \),
we have \( \decomp(T) = (\hole, E, P) \) if and only if \( \decomp(U[T]) = (U, E, P) \).
\end{lem}
\begin{proof}
The proof proceeds by induction on \(|U|\).
If \( U = \hole \) the claim trivially holds.
If \(U \neq \hole\), \(U\) is of the form \(a(T_1,\dots,U_i,\dots,T_{\Sigma(a)})\).
Since \(|U| < m\), we have \( |U_i| < m \) and \(|T_j|<m\) for every \(j\neq i\).
Assume that \( \decomp(T) = (\hole, E, P) \).
By the induction hypothesis, we have \( \decomp(U_i[T]) = (U_i, E, P) \).
Since \( |U| = |a(T_1, \dots, T_{i-1}, U_i, T_{i+1}, \dots, T_{\Sigma(a)})| < m \), we should apply the third case of Equation~\eqref{eq:defOfDecomp} to compute \( \decomp(U[T]) \).
Hence, we have \[ \decomp(U[T]) = (a(T_1, \dots, T_{i-1}, U_i, T_{i+1}, \dots, T_{\Sigma(a)}),E,P) = (U, E, P). \]
Conversely, assume that
\[
\decomp(U[T]) = (U, E, P) = (a(T_1, \dots, T_{i-1}, U_i, T_{i+1}, \dots, T_{\Sigma(a)}), E, P) .
\]
Let \( (U'_i, E'_i, P'_i) = \decomp(U_i[T]) \) and \( (U'_j, E'_j, P'_j) = \decomp(T_j) \) for each \( j \neq i \).
Then the final step in the computation of \( \decomp(U[T]) \) must be the third case; otherwise \( U = \hole \), a contradiction.
By the position of the unique hole in \( U \), it must be the case that \( U = a(T_1, \dots, T_{i-1}, U'_i, T_{i+1}, \dots, T_{\Sigma(a)}) \), \( E = E'_i \) and \( P = P'_i \).
So \( (U_i, E, P) = \decomp(U_i[T]) \).
By the induction hypothesis, \( \decomp(T) = (\hole, E, P) \).
\end{proof}
The following is the key lemma for the product lemma, which says that
if \((U,E,P)=\decomp(T)\), the decomposition is actually independent of the \(P\)-part.
\begin{lem}\label{lem:decomposition-replace}
If \( (U, E, P) = \decomp(T) \), then
\( \decomp(U[E]\Hfill{P'}) = (U, E, P') \) for any \(P'\in \prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}]\).
\end{lem}
\begin{proof}
The proof proceeds by induction on \(|T| (= |U[E]|)\). If \(|T|< m\), then \(E=T\), hence \(\shn(E)=0\). Thus,
\(P=P'=\epsilon\), which implies
\( \decomp(U[E]\Hfill{P'}) = \decomp(T)=(U,E,P)=(U,E,P')\).
For the case \(|T|\ge m\),
we proceed by the case analysis on which rule of
Equation~\eqref{eq:defOfDecomp} was
used to compute \( (U, E, P) = \decomp(T) \).
Assume that \( T = a(T_1, \dots, T_{\Sigma(a)}) \) and \( (U_i, E_i, P_{i}) = \decomp(T_i) \) for each \( i = 1, \dots, \Sigma(a) \).
By Lemma~\ref{lemma:charphi}, we have \( \cbf{P_{i}}{E_i} \) for each \( i \).
\begin{itemize}
\item
The first case of Equation~\eqref{eq:defOfDecomp}:
We have \( |T_j|, |T_{j'}|\ge m\)
for some \( 1 \le j < j' \le \Sigma(a) \), and:
\[
U = \hole
\qquad
E = a(U_1[E_1], \dots, U_{\Sigma(a)}[E_{\Sigma(a)}])
\qquad
P = P_{1} \cdots P_{\Sigma(a)}.
\]
Since \(P, P':E\), we can split
\( P'=P'_{1} \cdots P'_{\Sigma(a)} \) so that \( \cbf{P'_{i}}{E_i} \) for each \( i \).
By the induction hypothesis,
\[
\decomp(U_i[E_i]\Hfill{P'_{i}}) = (U_i, E_i, P'_{i})
\qquad\mbox{(for each \( i = 1, \dots, \Sigma(a) \)).}
\]
Since \(|U_j[E_j]\Hfill{P'_{j}}|= |U_{j}[E_{j}]| =|T_j|\ge m \) and
\(|U_{j'}[E_{j'}]\Hfill{P'_{j'}}|= |U_{j'}[E_{j'}]|=|T_{j'}| \ge m \) for some \( 1 \le j < j' \le \Sigma(a) \), we have
\begin{align*}
\decomp(U[E\Hfill{P'}])
&= \decomp(a(U_1[E_1\Hfill{P'_1}], \dots, U_{\Sigma(a)}[E_{\Sigma(a)}\Hfill{P'_{\Sigma(a)}}])) \\
&= (\hole, a(U_1[E_1], \dots, U_{\Sigma(a)}[E_{\Sigma(a)}]), P'_{1} \cdots P'_{\Sigma(a)})\\
&= (U, E, P')
\end{align*}
as required.
\item
The second case of Equation~\eqref{eq:defOfDecomp}:
We have \( |T_j| \ge m \) for a unique \( j \in \{ 1, \dots, \Sigma(a) \} \) and:
\[
n \triangleq |U_0| \ge m \qquad
U = \hole
\qquad
E = \hhole^n_1[E_{j}]
\qquad
P = U_0 \cdot P_j
\]
for \( U_0 \triangleq a(T_1, \dots, T_{j-1}, U_{j}, T_{j+1}, \dots, T_{\Sigma(a)}) \).
Because \(|T_j|\geq m\), we have \(|E_j|\ge m\) by Lemma~\ref{lem:size-of-E}.
Now we have \(|E_j\Hfill{P_j}|=|E_j|\ge m\), \( |U_j|<m \), and \( \decomp(U_j[E_j\Hfill{P_j}]) =
\decomp(T_j) = (U_j, E_j, P_j) \); hence
by Lemma~\ref{lemma:ehole}, we have \( \decomp(E_j\Hfill{P_j}) = (\hole, E_j, P_j) \).
Let \( P' = U_0' \cdot P'' \).
By the assumption, \( \cbf{U_0'}{\hhole^n_1} \) and thus \( U_0' \) is a \( 1 \)-context.
Also, \( U_0' \) is good for \( m \); thus \( |U_0'| \ge m \)
and \(U'_0\) is of the form \( a'(T_1', \dots, T'_{j'-1}, U'_{j'}, T'_{j+1}, \dots, T'_{\Sigma(a')}) \)
where \( |U'_{j'}| < m \) and \( |T'_i| < m \) for every \( i \neq j' \).
Now we have
\begin{enumerate}
\item \( \decomp(U'_{j'}[E_j]\Hfill{P_j}) = (U'_{j'}, E_j, P_j) \) by Lemma~\ref{lemma:ehole} and \( \decomp(E_j\Hfill{P_j}) = (\hole, E_j, P_j) \);
\item \(P''\in \prod_{i=1}^{\shn(E_j)} \ucomp[\sh{E_j}{i}]\) since
\(P'=U_0'\cdot P''\in \prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}] \) where \(E=\hhole^n_1[E_{j}]\);
\item \( |U'_{j'}[E_j]| \;<\; |U_0'[E_j]| \;=\; n + |E_j| \;=\; |E| \;=\; |U[E]| \).
\end{enumerate}
Therefore, we can apply the induction hypothesis to \( U'_{j'}[E_j] \), resulting in
\[
\decomp(U'_{j'}[E_j]\Hfill{P''}) = (U'_{j'}, E_j, P'').
\]
We have \( |U'_{j'}[E_j]|\ge |E_j| \ge m \) and \( |T'_i| < m \) for every \( i \neq j' \).
Furthermore
\[
|a'(T'_1, \dots, T'_{j'-1}, U'_{j'}, T'_{j'+1}, \dots, T'_{\Sigma(a')})| \;=\; |U_0'| \;=\; n \;\ge\; m.
\]
Hence we have
\begin{align*}
\decomp(U[E]\Hfill{P'})
&=
\decomp(U_0'[E_j]\Hfill{P''}) \\
&= \decomp(a'(T'_1, \dots, T'_{j'-1}, U'_{j'}, T'_{j'+1}, \dots, T'_{\Sigma(a')})[E_j]\Hfill{P''}) \\
&= \decomp(a'(T'_1, \dots, T'_{j'-1}, U'_{j'}[E_j]\Hfill{P''}, T'_{j'+1}, \dots, T'_{\Sigma(a')})) \\
&= (\hole, \hhole^n_1[E_j], U'_0 \cdot P'')\\
&= (U,E,P')\\
\end{align*}
\item
The third case of Equation~\eqref{eq:defOfDecomp}:
We have \( |T_j| = |U_j[E_{j}]| \ge m \) for a unique \( j \in \{ 1, \dots, \Sigma(a) \} \) (and thus \( |T_i| < m \) for every \( i \neq j \)) and: \[
U = a(T_1, \dots, T_{j-1}, U_{j}, T_{j+1}, \dots, T_{\Sigma(a)})
\qquad
|U|<m
\qquad
E = E_j
\qquad
P = P_j.
\]
By the induction hypothesis,
\[
\decomp(U_j[E_j]\Hfill{P'}) = (U_j, E_j, P').
\]
Since \( |T_i| < m \) for every \( i \neq j \) and \( |a(T_1, \dots, T_{j-1}, U_{j}, T_{j+1}, \dots, T_{\Sigma(a)})| < m \), we have
\begin{align*}
\decomp(U[E]\Hfill{P'})
&= \decomp(a(T_1, \dots, T_{j-1}, U_j[E_j]\Hfill{P'}, T_{j+1}, \dots, T_{\Sigma(a)})) \\
&= (a(T_1, \dots, T_{j-1}, U_{j}, T_{j+1}, \dots, T_{\Sigma(a)}), E_j, P')\\
&= (U, E, P')\\
\end{align*}
\item
The fourth case of Equation~\eqref{eq:defOfDecomp}:
Let \( n \) be \( |T| \); we have \( n \ge m \) and:
\[
U = \hole
\qquad
E = \hhole^n_0
\qquad
P = T.
\]
Since \( P' : E \), \(P'\) must be a singleton sequence consisting of a tree,
say \( T'=a'(T_1', \dots, T_{\Sigma(a')}') \).
By the assumption, \( T' \) is good for \( m \).
Hence \( |T'| \ge m \) and \( |T_i'| < m \) for every \( i = 1, \dots, \Sigma(a') \).
So
\[
\decomp(U[E]\Hfill{P'})
\;=\; \decomp(T')
\;=\; (\hole, \hhole^n_0, T')
\;=\; (U,E,P').
\]
\end{itemize}
\end{proof}
\begin{cor}\label{cor:closed-decomposition-replace}
If \( (E, P) = \closeddecomp(T) \), then
\( \closeddecomp(E\Hfill{P'}) = (E, P') \) for any \(P'\in \prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}]\).
\end{cor}
\begin{proof}
If \( (E, P) = \closeddecomp(T) \), then \((U,E',P)=\decomp(T)\) and \(U[E']=E\) for some \(U, E'\).
Since \(\prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}] = \prod_{i=1}^{\shn(E')} \ucomp[\sh{E'}{i}]\), Lemma~\ref{lem:decomposition-replace}
implies \(\decomp(E\Hfill{P'}) = \decomp(U[E']\Hfill{P'}) = (U,E',P')\). Thus, we have
\(\closeddecomp(E\Hfill{P'}) = (U[E'],P')=(E,P')\) as required.
\end{proof}
We are now ready to prove the product lemma.
\begin{lem}[Product Lemma (for Grammar-Independent Decomposition)]\label{lem:product}
For \(E\) and \(m \ge 1\), if \(\mathcal{P}_E^m\) is non-empty, then
\[
\mathcal{P}_E^m = \prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}].
\]
\end{lem}
\begin{proof}
The direction \(\subseteq\) follows from
the fact that \( (E,P) = \closeddecomp(T) \) implies \( \cbf{P}{E} \) and \( \prj{P}{i} \) is good for \( m \)
(Lemma~\ref{lemma:charphi}).
We show the other direction.
Since \( \mathcal{P}_E^m \neq \emptyset \), there exist \(P\) and \( T \) such that \( \closeddecomp(T) = (E, P) \).
Let \( P' \in \prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}] \).
By Corollary~\ref{cor:closed-decomposition-replace}, we have \( \closeddecomp(E\Hfill{P'}) = (E, P') \).
This means that \( P' \in \mathcal{P}_E^m \).
\end{proof}
Lemma~\ref{cor:gindependent-tree-decomposition} follows as
an immediate corollary of Lemmas~\ref{lemma:union-sig} and \ref{lem:product}.
\subsection{Grammars in Canonical Form}
\label{sec:canonical-grammar}
As a preparation for generalizing the decomposition of \(\TRn{n}{\Sigma}\) (Lemma~\ref{cor:gindependent-tree-decomposition})
to that of \(\inhavn{{\mathcal G},N}\),
we first transform a given regular tree grammar into \emph{canonical form}, which will be defined
shortly (in Definition~\ref{def:canonical-gram}).
We prove that the transformation preserves unambiguity and (a weaker version of) strong connectivity.
\begin{defi}[Canonical Grammar]
\label{def:canonical-gram}
A rewriting rule of a regular tree grammar \( {\mathcal G} \) is in \emph{canonical form} if it is of the form
\[
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{\mathcal G} a(N_1, \dots, N_{\Sigma(a)}).
\]
A grammar \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) is \emph{canonical} if every rewriting rule is in canonical form.
\end{defi}
We transform a given regular tree grammar \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) to
an equivalent one in canonical form.
The idea of the transformation is fairly simple: we replace a rewriting rule
\[
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(T_1, \dots, T_n) \qquad\mbox{(\( T_1, \dots, T_n \) are \( (\Sigma \cup {\mathcal N}) \)-trees)}
\]
such that \( T_i \notin {\mathcal N} \) with rules
\[
\{\;\; N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(T_1, \dots, T_{i-1}, N', T_{i+1}, \dots, T_n),
\qquad
N' \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T_i \;\;\}
\]
where \( N' \) is a fresh nonterminal that does not appear in \( {\mathcal N} \).
After iteratively applying the above transformation, we next replace a rewriting rule of the form \( N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} N' \) with rules
\[
\{ N_0 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(\dots, N', \dots) \mid (N_0 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(\dots, N, \dots)) \in {\mathcal R} \}.
\]
Again by iteratively applying this transformation, we finally obtain a grammar in canonical form.
The transformation, however, does not preserve strong connectivity.
For example, consider the grammar \( {\mathcal G} = (\{ \Ta \mapsto 2, \Tb \mapsto 1, \Tc \mapsto 0 \}, \{ N \}, {\mathcal R}) \) where
\[
{\mathcal R} = \{\;\;
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tb(\Tc), \qquad
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Ta(N, N)
\;\;\}.
\]
Then the above transformation introduces a nonterminal \( N' \) as well as rules
\[
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tb(N')
\qquad\mbox{and}\qquad
N' \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tc.
\]
Then \( N \) is not reachable from \( N' \).
Observe that the problem above was
caused by a newly introduced nonterminal that generates a single finite tree.
To overcome the problem, we introduce a weaker version of strong connectivity
called \emph{essential strong-connectivity}. It
requires strong connectivity only for nonterminals generating infinite languages;
hence, it is preserved by the above transformation.
\begin{defi}[Essential Strong-connectivity]
Let \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) be a regular tree grammar.
We say that \( {\mathcal G} \) is \emph{essentially strongly-connected} if for any nonterminals \( N_1, N_2 \in {\mathcal N} \) with \( \card{\inhav{{\mathcal G},N_1}} = \card{\inhav{{\mathcal G},N_2}} = \infty \), \( N_2 \) is reachable from \( N_1 \).
\end{defi}
Note that by the definition, every strongly-connected grammar is
also essentially-strongly connected.
In the definition above, as well as in the arguments below, nonterminals \( N \) with \( \card{\inhav{{\mathcal G},N}} = \infty \) play an important role.
We write \( \nontinfp{\nont} \) for the subset of \( {\mathcal N} \) consisting of such nonterminals.
\begin{rem}
A regular tree grammar that is essentially strongly-connected can be easily transformed into a strongly-connected grammar, hence the terminology.
Let \( {\mathcal G} \) be an essentially strongly-connected grammar and \( N \in \nontinfp{\nont} \).
We say that a nonterminal is \emph{inessential} if it generates a finite language.
Let \( N_0 \) be an inessential nonterminal of a grammar \( {\mathcal G} \) such that \( \inhav{{\mathcal G}, N_0} = \{ T_1, \dots, T_n \} \).
Then by replacing each rule
\[
N_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} C[N_0,\ldots,N_0]
\]
(where \( C \) is a \( k \)-context possibly having nonterminals other than \(N_0\)) with rules
\[
\{ N_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} C[T_{i_1},\ldots,T_{i_k}] \mid i_1,\ldots,i_k\in\set{1, \dots, n} \},
\]
one can remove the inessential nonterminal \( N_0 \) from the grammar.
A grammar \( {\mathcal G} \) is essentially strongly-connected if and only if the grammar \( {\mathcal G}' \) obtained by removing all inessential nonterminals is strongly-connected.
This transformation preserves the language in the following sense:
writing \( {\mathcal G}' = (\Sigma, \nontinfp{\nont}, {\mathcal R}') \) for the resulting grammar,
we have \( \inhav{{\mathcal G}, N} = \inhav{{\mathcal G}', N} \) for each \(N \in \nontinfp{\nont}\).
Note that the process of erasing inessential nonterminals breaks canonicity; in fact,
{the class of languages generated by strongly-connected canonical regular tree grammars
is a \emph{proper} subset of that of essentially strongly-connected canonical regular tree grammars.}
\end{rem}
Recall that the second main theorem (Theorem~\ref{thm:imt}) takes a family \( (S_n)_{n \in \mathbb{N}} \)
from \( \ctx{{\mathcal G}} = \bigcup_{N, N' \in {\mathcal N}} \inhav{{\mathcal G}, N \Ar N'} \).
In order to restate the theorem for essentially strongly connected grammars,
we need to replace
\(\ctx{{\mathcal G}}\) with the ``essential'' version, namely,
\[
\ctxinf{{\mathcal G}} \triangleq \bigcup_{N, N' \in \nontinfp{\nont}} \inhav{{\mathcal G}, N \Ar N'}.
\]
\begin{lem}[Canonical Form]
\label{lem:canonical-grammar}
Let \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) be a regular tree grammar that is unambiguous and strongly connected.
Then one can (effectively) construct a grammar \( {\mathcal G}' = (\Sigma, {\mathcal N}', {\mathcal R}') \) and
a family \( ( \nset_N )_{N \in {\mathcal N}} \) of subsets \( \nset_N \subseteq {\mathcal N}' \) that satisfy the following conditions:
\begin{itemize}
\item \( {\mathcal G}' \) is canonical, unambiguous and essentially strongly-connected.
\item \( \inhav{{\mathcal G},N} = \biguplus_{N' \in \nset_N} \inhav{{\mathcal G}',N'} \) for every \(N \in {\mathcal N}\).
\item If \( \card{\inhav{{\mathcal G}}} = \infty \), then \( \ctx{{\mathcal G}} \subseteq \ctxinf{{\mathcal G}'} \).
\end{itemize}
\end{lem}
\begin{proof}
See Appendix~\ref{appx:canonical}.
\end{proof}
\subsection{Decomposition of Regular Tree Languages}
\label{sec:rtl-decomposition}
This subsection generalizes the decomposition of \(\TRn{n}{\Sigma}\)
in Section~\ref{sec:decomposition}:
\[ \TRn{n}{\Sigma}\;\cong
\coprod_{E \in \mathcal{E}_n^m} \prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}]
\]
to
that of \(\inhavn{{\mathcal G},N}\), and proves a bijection of the following form:
\[
\inhavn{{\mathcal G},N}\;\cong
\coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G},N)} \prod_{i=1}^{\shn(\widetilde{E})} \tcomp[\sh{\widetilde{E}}{i}]. \]
Here, \(\widetilde{E}\) denotes a \emph{typed} second-order context
(which
will be defined shortly in Section~\ref{sec:typed-context}),
each of whose second-order holes
\(\sh{\widetilde{E}}{i}=\Hhole{\f}{n}\) carries not only the size \(n\) but the \emph{context type} \(\f\) of a context to be substituted for the hole.
Accordingly, we have replaced
\(\ucomp[\sh{E}{i}]\) with \(\tcomp[\sh{\widetilde{E}}{i}]\), which denotes the set of contexts that respect
the context type specified by \(\sh{\widetilde{E}}{i}\).
In the rest of this subsection, we first define the notion of typed second-order contexts in
Section~\ref{sec:typed-context}, extend the decomposition function accordingly in Section~\ref{sec:grespecting-decomposition}, and use it to prove the above bijection in Section~\ref{sec:grespecting-bijection}.
Throughout this subsection (i.e.,~Section~\ref{sec:rtl-decomposition}),
we assume that \({\mathcal G}\) is a canonical and unambiguous grammar.
We emphasize here that the discussion in this subsection
essentially relies on both unambiguity and canonicity of the grammar.
The essential strong connectivity is not required for the results in this
subsection; it will be used in Section~\ref{sec:adjusting}, to show
that each component \(\tcomp[\sh{\widetilde{E}}{i}]\) contains an affine
context that has \(S_{\ceil{\cn \log n}}\) as a subcontext
(recall condition (i) of (T3) in Section~\ref{sec:IMTforWords}).
\begin{rem}
Note that any deterministic bottom-up tree automaton
(without any \(\epsilon\)-rules) ~\cite{tata2007}
can be considered an unambiguous canonical tree grammar, by regarding
each transition rule \(a(q_1,\ldots,q_k) \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} q\) as a rewriting rule
\(N_q \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_{q_1},\ldots, N_{q_k})\). Thus, by the equivalence between the
class of tree languages generated by regular tree grammars and the class of those
accepted by deterministic bottom-up tree automata, any regular tree grammar can be
converted to an unambiguous canonical tree grammar.
\end{rem}
\subsubsection{Typed Second-Order Contexts}
\label{sec:typed-context}
The set of \emph{\( {\mathcal G} \)-typed second-order contexts}, ranged over by \(\widetilde{E}\), is defined by:
\[
\widetilde{E} ::= \hhole_{N_1 \cdots N_k \Ar N}^n[\widetilde{E}_1, \ldots, \widetilde{E}_k] \;\mid\; a(\widetilde{E}_1, \ldots,
\widetilde{E}_{\Sigma(a)}) \ \ (a \in \mathop{\mathrm{Dom}}(\Sigma)).
\]
The subscript \( \kappa = (N_1 \cdots N_k \Ar N) \) describes the type of
first-order contexts that can be substituted for this second-order hole;
the superscript \( n \) describes the size as before.
Hence a (first-order) context \( C \) is suitable for
filling a second-order hole \( \hhole_{\kappa}^n \) if \( C : \kappa \) and \( |C| = n \).
We write \( \cbf{C}{\hhole_{\kappa}^n} \) if \( C : \kappa \) and \( |C| = n \).
The operations such as \( \sh{\widetilde{E}}{i} \), \( \widetilde{E}\Hfill{C} \) and \( \widetilde{E}\Hfill{P} \) are defined
analogously.
For a sequence of contexts \(P=C_1C_2\cdots C_{\ell}\),
we write \( \cbf{P}{\widetilde{E}} \)
if \( \len{P} (= \ell)=\shn(\widetilde{E}) \) and \( \cbf{C_i}{\sh{\widetilde{E}}{i}} \) for each \( i \le \shn(\widetilde{E}) \).
We define the second-order context typing relation \(\p \widetilde{E}:N\) inductively
by the rules in Figure~\ref{fig:TypeRuleE}. Intuitively,
\(\p \widetilde{E}:N\) means that \(\widetilde{E}\Hfill{P}\in
\inhav{{\mathcal G},N}\) holds for any \(P\) such that \(P\COL \widetilde{E}\) (as confirmed in Lemma~\ref{lem:second-order-substitution-lemma} below).
As in the case of untyped second-order contexts,
we actually use only typed second-order contexts with holes
of the form \(\hhole_{N_1 \cdots N_k \Ar N}^n\) where \(k\) is \(0\) or \(1\).
\begin{figure}[t]
\[
\begin{array}{cr}
\dfrac{
\p \widetilde{E}_i: N_i
\quad
\mbox{(for each \( i = 1, \dots, k \))}
}{
\p \hhole_{N_1 \cdots N_k \Ar N}^{n}[\widetilde{E}_1,\dots,\widetilde{E}_k] : N
}
&
(\text{SC-Hole})
\\[25pt]
\dfrac{
(N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1,\dots,N_{\Sigma(a)})) \in {\mathcal R}
\qquad
\p \widetilde{E}_i : N_i \quad\mbox{(for each $i = 1,\dots,\Sigma(a)$)}
}{
\p a(\widetilde{E}_1,\dots,\widetilde{E}_{\Sigma(a)}) : N
}
&
(\text{SC-Term})
\end{array}
\]
\caption{Typing rules for \(\widetilde{E}\)}\label{fig:TypeRuleE}
\end{figure}
\begin{exa}
\label{ex:gctx}
Recall the second-order context
\(E = \Tb(\Hhole{1}{3}[\Ta(\Hhole{0}{3}, \Tb(\Hhole{0}{5}))])\) in
Figure~\ref{fig:partitioning} and Example~\ref{ex:decomp}.
Given the grammar consisting of the rules:
\[A\to \Ta(B,B) \qquad B\to \Tb(A)\qquad B\to \Tb(B)\qquad B\to \Tc,\]
the corresponding \emph{typed} second-order context \(\widetilde{E}\) is:
\[\Tb(\Hhole{A\Ar A}{3}[\Ta(\Hhole{\Ar B}{3}, \Tb(\Hhole{\Ar A}{5}))])\]
and we have \(\p \widetilde{E} : B\).
For \(P = \Ta(\Tb(\hole),\Tc)\cdot \Tb(\Tb(\Tc)) \cdot \Ta(\Tb(\Tc), \Tb(\Tc))\), we have \(P\COL\widetilde{E}\),
and:
\[\widetilde{E}\Hfill{P} = \Tb(\Ta(\Tb(\Ta(\Tb(\Tb(\Tc)), \Tb(\Ta(\Tb(\Tc), \Tb(\Tc))))), \Tc)) \in \inhav{{\mathcal G},B}.\]
\end{exa}
\begin{lem}
The following rule is derivable:
\[
\begin{array}{cr}
\dfrac{
C \in \inhav{{\mathcal G}, N_1 \dots N_k \Ar N}
\qquad
\p \widetilde{E}_i : N_i
\quad
\mbox{\rm (for each \( i = 1, \dots, k \))}
}{
\p C[\widetilde{E}_1, \dots, \widetilde{E}_k] : N
}
&
(\mbox{\rm SC-Ctx})
\end{array}
\]
\end{lem}
\begin{proof}
The proof proceeds by induction on \(C\).
Assume that the premises of the rule hold.
If \(C=\hole\), then \(C[\widetilde{E}_1, \dots, \widetilde{E}_k] =\widetilde{E}_1\)
and \(N=N_1\); thus the result follows immediately from the assumption.
If \(C=a(C_1,\ldots,C_\ell)\), then by the assumption that the grammar is canonical
and \(C \in \inhav{{\mathcal G}, N_1 \dots N_k \Ar N}\),
we have \(N\to a(N'_1,\ldots,N'_\ell)\) with \(C_i\in
\inhav{{\mathcal G},N_{k_{i-1}+1}\dots N_{k_i}\Ar N'_i}\)
for \(i=1,\ldots,\ell\) where \(k_0=0\) and \(k_\ell=k\).
By the induction hypothesis,
we have \(\p C_i[\widetilde{E}_{k_{i-1}+1},\ldots,\widetilde{E}_{k_{i}}]:N'_i\).
Thus, by using rule (\rm{SC-Term}), we have
\(\p C[\widetilde{E}_1, \dots, \widetilde{E}_k] : N\) as required.
\end{proof}
\begin{lem}
\label{lem:second-order-substitution-lemma}
Assume that \( \p \widetilde{E} : N \).
\begin{enumerate}
\item\label{item:second-order-substitution-lemma-tree}
If \( \shn(\widetilde{E}) = 0 \), then \( \widetilde{E} \) is a tree and \( \widetilde{E} \in \inhav{{\mathcal G}, N} \).
\item\label{item:second-order-substitution-lemma-sbst}
If \( \shn(\widetilde{E}) \ge 1 \) and \( \cbf{C}{\sh{\widetilde{E}}{1}} \), then \( \p \widetilde{E}\Hfill{C} : N \).
\item\label{item:second-order-substitution-lemma-sbstall}
If \( \cbf{P}{\widetilde{E}} \), then \( \widetilde{E}\Hfill{P} \in \inhav{{\mathcal G}, N} \).
\end{enumerate}
\end{lem}
\begin{proof}
\((1)\) By induction on the structure of \( \p \widetilde{E} : N \).
\((2)\) By induction on the structure of \( \p \widetilde{E} : N \).
The base case is when \( \widetilde{E} = \hhole_{N_1 \dots N_k \Ar N}^n[\widetilde{E}_1, \dots, \widetilde{E}_k] \).
Since \( \p \widetilde{E} : N \), we have \( \p \widetilde{E}_i : N_i \) for each \( i \).
Then by the derived rule (\textsc{SC-Ctx}), we have \( \p C[\widetilde{E}_1, \dots, \widetilde{E}_k] : N \).
\((3)\) By induction on \( \shn(\widetilde{E}) \) (using \((1)\) and \((2)\)).
\end{proof}
\subsubsection{Grammar-Respecting Decomposition Function \(\widetilde{\Phi}_{m}^{\bullet,\grm}\)}
\label{sec:grespecting-decomposition}
We have defined in Section~\ref{sec:gindependent-decomposition}
the function \(\closeddecomp\) that decomposes a tree \(T\) and
returns a pair \((E,P)\) of a second-order context \( E \) and
a sufficiently long sequence \(P\) of (first-order) contexts.
The aim here is to extend \(\closeddecomp\) to a grammar-respecting one
\(\widetilde{\Phi}_{m}^{\bullet,\grm}\) that takes a pair \((T,N)\) such that
\(T\in \inhav{{\mathcal G},N}\) as an input, and returns a pair \((\widetilde{E},P)\),
which is the same as \(\closeddecomp(T)=(E,P)\), except that
\(\widetilde{E}\) is a ``type-annotated'' version of \(E\).
For example, for the tree
in Figure~\ref{fig:partitioning} and the grammar in Example~\ref{ex:gctx},
we expect that \(\widetilde{\Phi}_{m}^{\bullet,\grm}(T) = (\widetilde{E},P)\) where:
\[
\begin{array}{l}
T = \Tb(\Ta(\Tb(\Ta(\Tb(\Tb(\Tc)), \Tb(\Ta(\Tb(\Tc), \Tb(\Tc))))), \Tc))\\
\widetilde{E} = \Tb(\Hhole{A\Ar A}{3}[\Ta(\Hhole{\Ar B}{3}, \Tb(\Hhole{\Ar A}{5}))])\qquad
P = \Ta(\Tb(\hole),\Tc)\cdot \Tb(\Tb(\Tc)) \cdot \Ta(\Tb(\Tc), \Tb(\Tc)).
\end{array}
\]
We say that \emph{\( \widetilde{E} \) refines \( E \)}, written \( \widetilde{E} \lhd E \), if \( E \) is obtained by simply forgetting type annotations, i.e.,~replacing \( \hhole_{N_1 \cdots N_k \Ar N}^n \) in \( \widetilde{E} \) with \( \hhole_k^n \).
This relation is formally defined by induction on the structures of \( E \) and \( \widetilde{E} \) by the rules in Figure~\ref{fig:refinement-of-second-order-contexts}.
\begin{figure}[t]
\[
\begin{array}{cr}
\dfrac{
\widetilde{E}_i \lhd E_i
\quad
\mbox{(for each \( i = 1, \dots, k \))}
}{
\hhole_{N_1 \cdots N_k \Ar N}^{n}[\widetilde{E}_1,\dots,\widetilde{E}_k] \lhd \hhole_k^n[E_1,\dots,E_n]
}
&
(\text{Ref-Hole})
\\[25pt]
\dfrac{
\widetilde{E}_i \lhd E_i \quad\mbox{(for each $i = 1,\dots,\Sigma(a)$)}
}{
a(\widetilde{E}_1,\dots,\widetilde{E}_{\Sigma(a)}) \lhd a(E_1, \dots, E_n)
}
&
(\text{Ref-Term})
\end{array}
\]
\caption{Refinement relation}\label{fig:refinement-of-second-order-contexts}
\end{figure}
The following lemma is obtained by straightforward induction on \(\widetilde{E}\lhd E\).
\begin{lem}
\label{lem:ref-basic}
Assume \( \widetilde{E} \lhd E \).
\begin{enumerate}
\item\label{item:ref-basic-size}
\( |\widetilde{E}| = |E| \).
\item\label{item:ref-basic-type}
If \( \cbf{P}{\widetilde{E}} \), then \( \cbf{P}{E} \).
\item\label{item:ref-basic-sbst}
If \( \cbf{P}{\widetilde{E}} \), then \( \widetilde{E}\Hfill{P} = E\Hfill{P} \).
\end{enumerate}
\end{lem}
Given \( T \in \inhav{{\mathcal G}, N} \) and \( (E, P) = \closeddecomp(T) \),
the value of \( \widetilde{\Phi}_{m}^{\bullet,\grm}(T, N) \) should be \( (\widetilde{E}, P) \)
where \( \widetilde{E} \) is a \( {\mathcal G} \)-typed second-order context that satisfies the following conditions:
\[
\widetilde{E} \lhd E,
\qquad
\p \widetilde{E} : N
\quad\mbox{and}\quad
\cbf{P}{\widetilde{E}}.
\]
We prove that, for every \(T \in \inhav{{\mathcal G},N}\), there exists exactly one typed second-order context \( \widetilde{E} \) that satisfies the above condition.
Hence the above constraints defines the function \( \widetilde{\Phi}_{m}^{\bullet,\grm}(T, N) \).
We first state and prove a similar result for first-order contexts (in Lemma~\ref{lem:first-unique-decomposition}),
and then prove the second-order version (in Lemma~\ref{lem:second-unique-decomposition}),
using the former.
\begin{lem}
\label{lem:first-unique-decomposition}
Let \( C \) be a (first-order) \(k\)-context and \( T_1, \dots, T_k \) be trees. Assume that \( C[T_1, \dots, T_k] \in \inhav{{\mathcal G}, N} \).
Then there exists a unique family \( (N_i)_{i \in \set{1, \dots, k}} \) such that
\( C \in \inhav{{\mathcal G}, N_1 \dots N_k \Ar N} \) and \( T_i \in \inhav{{\mathcal G}, N_i} \) for every \( i = 1, \dots, k \).
\end{lem}
\begin{proof}
By induction on \( C \).
\begin{itemize}
\item Case \( C = \hole \):
Then \( k = 1 \).
The existence follows from \( \hole \in \inhav{{\mathcal G}, N \Ar N} \) (as \( N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* N \)) and \( T_1 = C[T_1] \in \inhav{{\mathcal G}, N} \).
The uniqueness follows from the fact that \( \hole \in \inhav{{\mathcal G}, N_1 \Ar N} \) implies \( N_1 = N \).
\item Case \( C = a(C_1, \dots, C_{\Sigma(a)}) \):
Let \( \ell = \Sigma(a) \), \( k_i = \hn{C_i} \) for each \( i = 1, \dots, \ell \) and
\[
(T_{1,1}, \dots, T_{1,k_1}, T_{2,1}, \dots, T_{2,k_2}, \dots, T_{\ell,1}, \dots, T_{\ell,k_\ell}) = (T_1, \dots, T_k).
\]
Then \( C[T_1, \dots, T_k] = a(C_1[T_{1,1}, \dots, T_{1,k}], \dots, C_{\ell}[T_{\ell,1}, \dots T_{\ell,k_{\ell}}]) \).
Since \( C[T_1, \dots, T_k] \in \inhav{{\mathcal G}, N} \), there exists a rewriting sequence
\[
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1, \dots, N_\ell) \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* a(C_1[T_{1,1}, \dots, T_{1,k_1}], \dots, C_{\ell}[T_{\ell,1}, \dots T_{\ell,k_{\ell}}]).
\]
Thus \( N_i \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* C_i[T_{i,1}, \dots, T_{i,k_i}] \), i.e.,~\( C_i[T_{i,1}, \dots, T_{i,k_i}] \in \inhav{{\mathcal G}, N_i} \), for each \( i = 1, \dots, \ell \).
By the induction hypothesis, there exist \( N_{i,1}, \dots, N_{i,k_i} \) such that \( C_i \in \inhav{{\mathcal G}, N_{i,1} \dots N_{i,k_i} \Ar N_i} \) and \( T_{i,j} \in \inhav{{\mathcal G}, N_{i,j}} \) for each \( j = 1, \dots, k_i \).
Then (the rearrangement of) the family \( (N_{i,j})_{i \le \ell, j \le k_i} \) satisfies the requirement.
We prove the uniqueness.
Assume that both \( (N^1_{i,j})_{i,j} \) and \( (N^2_{i,j})_{i,j} \) satisfy the requirement.
Then, for each \( m = 1, 2 \), there exist \(N^m_i\) (\(i=1,\dots,\ell\)) such that
\begin{align*}
N
\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N^m_1, \dots, N^m_\ell)
& \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* a(C_1[N^m_{1,1}, \dots, N^m_{1,k_1}], \dots, C_\ell[N^m_{\ell,1}, \dots, N^m_{\ell,k_\ell}]) \\
& \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* a(C_1[T_{1,1}, \dots, T_{1,k_1}], \dots, C_{\ell}[T_{\ell,1}, \dots, T_{\ell,k_{\ell}}]).
\end{align*}
Since \( {\mathcal G} \) is unambiguous, \( N^1_i = N^2_i \) for each \( i = 1, \dots, \ell \).
By the induction hypothesis, we have \( N^1_{i,j} = N^2_{i,j} \) for each \( i \) and \( j \).
\qedhere
\end{itemize}
\end{proof}
\begin{lem}
\label{lem:second-unique-decomposition}
Let \( {\mathcal G} \) be a canonical unambiguous regular tree grammar and \( T \in \inhav{{\mathcal G}, N} \).
Given \(E\) and \(P\), assume that \( \cbf{P}{E} \) and \( E\Hfill{P} = T \).
Then \( E \) has a unique refinement \( \widetilde{E} \lhd E \) such that \( \p \widetilde{E} : N \) and \( \cbf{P}{\widetilde{E}} \).
\end{lem}
\begin{proof}
We prove by induction on \( E \).
\begin{itemize}
\item
Case \( E = \hhole_k^n[E_1, \dots, E_k] \):
The sequence \( P \) can be decomposed as \( P = C \cdot P_1 \cdots P_k \) so that \( \cbf{P_i}{E_i} \) for each \( i = 1, \dots, k \).
Furthermore \( \hn{C} = k \) and \( |C| = n \).
We have \( E\Hfill{P} = C[E_1\Hfill{P_1}, \dots, E_k\Hfill{P_k}] \in \inhav{{\mathcal G}, N} \).
We prove the existence.
By Lemma~\ref{lem:first-unique-decomposition}, there exists a family \( (N_i)_{i = 1, \dots, k} \) such that \( C \in \inhav{{\mathcal G}, N_1 \dots N_k \Ar N} \) (i.e.,~\( C : \hhole_{N_1 \dots N_k \Ar N}^n \)) and \( E_i\Hfill{P_i} \in \inhav{{\mathcal G}, N_i} \) for each \( i = 1, \dots, k \).
By the induction hypothesis, for each \( i = 1, \dots, k \), there exists \( \widetilde{E}_i \lhd E_i \) such that \( \p \widetilde{E}_i : N_i \) and \( \cbf{P_i}{\widetilde{E}_i} \).
Let \( \widetilde{E} \triangleq \hhole_{N_1\dots N_k \Ar N}^n[\widetilde{E}_1, \dots, \widetilde{E}_k] \).
Then \( \widetilde{E} \lhd E \), \( \p \widetilde{E} : N \), and \( \cbf{P}{\widetilde{E}} \).
We prove the uniqueness.
Assume \( \widetilde{E}^1 \) and \( \widetilde{E}^2 \) satisfy that \( \widetilde{E}^j \lhd E \), \( \p \widetilde{E}^j : N \) and \( \cbf{P}{\widetilde{E}^j} \) for \( j = 1, 2 \).
Since \( \widetilde{E}^j \lhd E \) and \( \p \widetilde{E}^j : N \), \(\widetilde{E}^j\) must be of the form: \( \widetilde{E}^j =
\hhole_{N_1^j \dots N_k^j \Ar N}^n[\widetilde{E}^j_1, \dots, \widetilde{E}^j_k] \) with \( \widetilde{E}^j_i \lhd E_i \) and \( \p \widetilde{E}^j_i : N^j_i \).
Since \( \cbf{P}{\widetilde{E}^j} \), we have \( C \in \inhav{{\mathcal G}, N_1^j \dots N_k^j \Ar N} \)
and \(P_i : \widetilde{E}^j_i\) for \(i=1,\dots,k\).
By Lemma~\ref{lem:second-order-substitution-lemma}, \( \widetilde{E}^j_i\Hfill{P_i} \in \inhav{{\mathcal G}, N_i^j} \) for each \( i = 1, \dots, k \) and \( j = 1, 2 \).
By Lemma~\ref{lem:ref-basic}\eqref{item:ref-basic-sbst},
\(E\Hfill{P} = \widetilde{E}^j\Hfill{P} = C[\widetilde{E}^j_1\Hfill{P_1}, \dots, \widetilde{E}^j_k\Hfill{P_k}] \in \inhav{{\mathcal G},N}\).
Hence by Lemma~\ref{lem:first-unique-decomposition}, \( N_i^1 = N_i^2 \) for each \( i = 1, \dots, k \).
By the induction hypothesis, \( \widetilde{E}_i^1 = \widetilde{E}_i^2 \) for each \( i \).
Hence \( \widetilde{E}^1 = \widetilde{E}^2 \).
\item
Case \( E = a(E_1, \dots, E_{\Sigma(a)}) \):
The sequence \( P \) can be decomposed as \( P = P_1 \cdots P_\Sigma(a) \) so that \( \cbf{P_i}{E_i} \) for each \( i = 1, \dots, \Sigma(a) \).
Then \( E\Hfill{P} = a(E_1\Hfill{P_1}, \dots, E_{\Sigma(a)}\Hfill{P_{\Sigma(a)}}) \).
We prove the existence.
Since \( N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* a(E_1\Hfill{P_1}, \dots, E_{\Sigma(a)}\Hfill{P_{\Sigma(a)}}) \),
there exists a rule \( N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1, \dots, N_{\Sigma(a)}) \) such that
\[
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1, \dots, N_{\Sigma(a)}) \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* a(E_1\Hfill{P_1}, \dots, E_{\Sigma(a)}\Hfill{P_{\Sigma(a)}}).
\]
So \( N_i \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* E_i\Hfill{P_i} \) for each \( i = 1, \dots, \Sigma(a) \).
By the induction hypothesis, there exists \( \widetilde{E}_i \lhd E_i \) such that \( \p \widetilde{E}_i : N_i \) and \( \cbf{P_i}{\widetilde{E}_i} \).
Let \( \widetilde{E} \triangleq a(\widetilde{E}_1, \dots, \widetilde{E}_{\Sigma(a)}) \).
Then \( \widetilde{E} \lhd E \), \( \p \widetilde{E} : N \), and \( \cbf{P}{\widetilde{E}} \).
We prove the uniqueness.
Assume \( \widetilde{E}^1 \) and \( \widetilde{E}^2 \) satisfy that \( \widetilde{E}^j \lhd E \), \( \p \widetilde{E}^j : N \) and \( \cbf{P}{\widetilde{E}^j} \) for \( j = 1, 2 \).
Since \( \widetilde{E}^j \lhd E \), \(\widetilde{E}^j\) must be of the form:
\( \widetilde{E}^j = a(\widetilde{E}^j_1, \dots, \widetilde{E}^j_{\Sigma(a)}) \) with \( \widetilde{E}^j_i \lhd E_i \).
By \( \p \widetilde{E}^j : N \), there exists a rule \( N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1^j, \dots, N_{\Sigma(a)}^j) \)
such that \( \p \widetilde{E}^j_i : N_i^j \) for each \( i \).
Since \(P : \widetilde{E}^j\), we have \(P_i : \widetilde{E}^j_i\) for each \(i\).
By Lemmas~\ref{lem:second-order-substitution-lemma} and~\ref{lem:ref-basic}\eqref{item:ref-basic-sbst},
we have \( \widetilde{E}_i^j\Hfill{P_i} = E_i\Hfill{P_i} \in \inhav{{\mathcal G}, N_i^j} \).
Now we have
\[
N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1^j, \dots, N_{\Sigma(a)}^j) \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* a(E_1\Hfill{P_1}, \dots, E_{\Sigma(a)}\Hfill{P_{\Sigma(a)}})
\]
for \( j = 1, 2 \).
Since \( {\mathcal G} \) is unambiguous, \( N_i^1 = N_i^2 \) for each \( i = 1, \dots, \Sigma(a) \).
By the induction hypothesis, \( \widetilde{E}_i^1 = \widetilde{E}_i^2 \) for each \( i \).
Hence \( \widetilde{E}^1 = \widetilde{E}^2 \).
\qedhere
\end{itemize}
\end{proof}
\subsubsection{Decomposition of \(\inhavn{{\mathcal G},N}\)}
\label{sec:grespecting-bijection}
To formally state the decomposition lemma, we prepare some definitions.
For a canonical unambiguous grammar ${\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R})$, $N \in {\mathcal N}$, $n \ge 1$, and $m \ge 1$,
we define \(\widetilde{\mathcal{E}}_n^m({\mathcal G}, N)\),
\(\widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N)\), and \(\tcomp\) by:
\begin{align*}
\widetilde{\mathcal{E}}_n^m({\mathcal G}, N) &\triangleq \{ \widetilde{E} \mid (\widetilde{E}, P) = \widetilde{\Phi}_{m}^{\bullet,\grm}(T,N) \text{ for some } T \in \inhavn{{\mathcal G}, N} \text{ and } P \}.
\end{align*}
\begin{align*}
\widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) &\triangleq \set{ P \mid (\widetilde{E}, P) = \widetilde{\Phi}_{m}^{\bullet,\grm}(T, N) \text{ for some } T \in \inhav{{\mathcal G}, N} }.
\end{align*}
\begin{align*}
\tcomp &\triangleq \set{ U \mid U : \hhole_{\kappa}^n, U \text{ is good for } m }.
\end{align*}
The set \(\widetilde{\mathcal{E}}_n^m({\mathcal G}, N)\) consists of second-order contexts that are obtained by
decomposing trees of size \(n\), and \(\widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N)\) consists of
affine context sequences that match \(\widetilde{E}\). The set \(\widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N)\) is the set of contexts
that match the hole \(\hhole_{\kappa}^n\).
The following is the main result of this subsection.
\begin{lem}
\label{lem:grm-bijection}
\begin{flalign*}
\inhavnn{n}{{\mathcal G}, N} \quad \cong &
\coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)} \prod_{i=1}^{\shn(\widetilde{E})} \tcomp[\sh{\widetilde{E}}{i}].
\end{flalign*}
\end{lem}
The lemma above is a direct consequence of
typed versions of coproduct and product lemmas (Lemmas~\ref{lemma:union-grm} and
\ref{lemma:product} below).
The following coproduct lemma can be shown in a manner similar to Lemma~\ref{lemma:union-sig}:
\begin{lem}[Coproduct Lemma (for Grammar-Respecting Decomposition)]
\label{lemma:union-grm}
For any
$n \geq 1$ and \(m \ge 1\),
there exists a bijection
\begin{flalign*}
\inhavnn{n}{{\mathcal G}, N} \quad \cong &
\coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)} \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N).
\end{flalign*}
such that \((\widetilde{E},P)\) in the right hand side is mapped to \(\widetilde{E}\Hfill{P}\).
\end{lem}
\begin{proof}
We define a function
\[
f:
\coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)} \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N)
\longrightarrow
\inhavnn{n}{{\mathcal G}, N}
\]
by \(f(\widetilde{E},P) \triangleq \widetilde{E}\Hfill{P}\),
and a function
\[
g:
\inhavnn{n}{{\mathcal G}, N}
\longrightarrow
\coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)} \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N)
\]
by \(g(T) \triangleq \widetilde{\Phi}_{m}^{\bullet,\grm}(T, N) \).
Let us check that these are functions into the codomains:
\begin{itemize}
\item \( f(\widetilde{E},P) \in \inhavn{{\mathcal G},N} \):
Since \(P \in \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N)\), there exists \( T \in \inhavn{{\mathcal G},N} \) such that \( (\widetilde{E},P) = \widetilde{\Phi}_{m}^{\bullet,\grm}(T,N) \).
By Lemmas~\ref{lemma:charphi} and \ref{lem:ref-basic}, we have \( f(\widetilde{E},P) = \widetilde{E}\Hfill{P} = T \in \inhavn{{\mathcal G},N} \).
\item \( g(T) \in \coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)} \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) \):
Obvious from the definitions of \( \widetilde{\mathcal{E}}_n^m({\mathcal G},N) \) and \( \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) \).
\end{itemize}
We have \(f(g(T)) = T\) by Lemmas~\ref{lemma:charphi}(\ref{charphi:split}) and \ref{lem:ref-basic}.
Let \( (\widetilde{E},P) \in \coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)} \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) \).
By definition, there exists \( T \in \inhavn{{\mathcal G},N} \) such that \( (\widetilde{E},P) = \widetilde{\Phi}_{m}^{\bullet,\grm}{(T,N)} = g(T) \).
Then
\[
g(f(\widetilde{E},P))
=g(f(g(T))
=g(T)
=(\widetilde{E},P).
\]
\end{proof}
The following is a key lemma used for proving a typed version of the product lemma.
\begin{lem}\label{lem:ppind}
For a nonterminal \(N\),
$n \geq 1$, \(m \ge 1\) and $ \widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)$, let \( E \) be the unique second-order context such that \( \widetilde{E} \lhd E \).
Then we have
\begin{align*}
\widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) = \mathcal{P}_E^m \cap \{ P \mid \cbf{P}{\widetilde{E}} \}.
\end{align*}
\end{lem}
\begin{proof}
The direction \(\subseteq\) is clear.
We prove the converse.
Let \(P\) be in the right hand side.
Since \( \widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)\), there exist \( T' \) and \( P' \) such that
\[
\widetilde{\Phi}_{m}^{\bullet,\grm}(T',N) = (\widetilde{E}, P').
\]
By the definition of \( \widetilde{\Phi}_{m}^{\bullet,\grm}(T',N) \), we have \( \p \widetilde{E} : N \).
Since \(P \in \mathcal{P}_E^m\), there exists \( T \) such that \( (E,P) = \closeddecomp(T) \) and thus, by Lemma~\ref{lemma:charphi},
\[
\closeddecomp(E\Hfill{P}) = (E,P).
\]
Since \( \closeddecomp(E\Hfill{P}) = (E,P) \), \( \widetilde{E} \lhd E \), \( \p \widetilde{E} : N \),
and \( \cbf{P}{\widetilde{E}} \), by the definition of \( \widetilde{\Phi}_{m}^{\bullet,\grm} \), we have \( \widetilde{\Phi}_{m}^{\bullet,\grm}(E\Hfill{P}, N) = (\widetilde{E},P) \).
By Lemmas~\ref{lem:second-order-substitution-lemma}\eqref{item:second-order-substitution-lemma-sbstall}
and~\ref{lem:ref-basic}\eqref{item:ref-basic-sbst}, \(E\Hfill{P} = \widetilde{E}\Hfill{P} \in \inhav{{\mathcal G},N}\).
Hence \( P \in \widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) \).
\end{proof}
The following is the typed version of the product lemma, which follows from
Lemmas~\ref{lem:product} and~\ref{lem:ppind}.
\begin{lem}[Product Lemma (for Grammar-Respecting Decomposition)]
\label{lemma:product}
For any nonterminal \(N\),
$n \geq 1$,
\(m \ge 1\) and
\( \widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G},N) \), we have
\[
\widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) = \prod_{i=1}^{\shn(\widetilde{E})} \tcomp[\sh{\widetilde{E}}{i}].
\]
\end{lem}
\begin{proof}
Let \( E \) be the unique second-order context such that \( \widetilde{E} \lhd E \).
By Lemmas~\ref{lem:product} and~\ref{lem:ppind}, we have:
\begin{align*}
\widetilde{\mathcal{P}}_{\gctx}^m({\mathcal G},N) &= \mathcal{P}_E^m \cap \{ P \mid \cbf{P}{\widetilde{E}} \}\\
&= (\textstyle\prod_{i=1}^{\shn(E)} \ucomp[\sh{E}{i}])\cap (\textstyle\prod_{i=1}^{\shn(\widetilde{E})} \set{U \mid U\COL \sh{\widetilde{E}}{i}})\\
&= \textstyle\prod_{i=1}^{\shn(\widetilde{E})} (\ucomp[\sh{E}{i}]\cap \set{U \mid U\COL \sh{\widetilde{E}}{i}})\\
&= \textstyle\prod_{i=1}^{\shn(\widetilde{E})} \tcomp[\sh{\widetilde{E}}{i}].
\end{align*}
\end{proof}
Lemma~\ref{lem:grm-bijection} is an immediate
corollary of Lemmas~\ref{lemma:union-grm} and \ref{lemma:product}.
\subsection{Each Component Contains the Subcontext of Interest}
\label{sec:adjusting}
In Section~\ref{sec:rtl-decomposition},
we have shown that the set \( \inhavn{{\mathcal G}, N} \) of trees can be decomposed as:
\[
\inhavnn{n}{{\mathcal G}, N} \quad \cong
\coprod_{\widetilde{E} \in \widetilde{\mathcal{E}}_n^m({\mathcal G}, N)} \prod_{i=1}^{\shn(\widetilde{E})} \tcomp[\sh{\widetilde{E}}{i}],
\]
assuming that \({\mathcal G}\) is canonical and unambiguous.
In this subsection, we further assume that \({\mathcal G}\) is essentially strongly connected,
and prove that, for each tree context \( S \in \ctxinf{{\mathcal G}} \), every component ``contains'' \( S \), i.e., there exists \(U\in\; \tcomp[\sh{\widetilde{E}}{i}]\) such that
\(S\preceq U\)
if \( m \) is sufficiently large, say \( m \ge m_0 \) (where \( m_0 \) depends on \( |S| \)).
More precisely, the goal of this subsection is to prove the following lemma.
\begin{lem}
\label{lemma:adjust}
Let \( {\mathcal G} = (\Sigma,{\mathcal N},{\mathcal R}) \) be an unambiguous, essentially strongly-connected grammar in canonical form and \((S_n)_{n \in \mathbb{N}} \) be a family of \linear{contexts} in \(\ctxinf{{\mathcal G}}\) such that \( |S_{n}| = \mathrm{O}(n) \).
Then there exist integers \(b,c \ge 1\) that satisfy the following:
For any \(N \in \nontinfp{\nont}\), \( n\ge 1 \), \( m \ge b \), \( \widetilde{E} \in \tframeNM{n}{cm}({\mathcal G},N) \)
and \(i \in \set{1,\dots,\shn(\widetilde{E})}\),
there exists \( U \in \tcompCM{\sh{\widetilde{E}}{i}}{cm}\) such that $S_{m} \preceq U$.
\end{lem}
The rest of this \new{subsection} is devoted to a proof of the lemma above.
The idea of the proof is as follows.
Assume \( S \in \inhav{{\mathcal G}, N_1 \Ar N_2} \) (\( N_1, N_2 \in \nontinfp{\nont} \)) and \( \hhole^n_{\f} = \sh{\widetilde{E}}{i} \).
Recall that\tk{where is the first occurrence of the claim below?}
\[
\tcomp[\hhole^n_{\f}] = \{ U \mid U \in \inhav{{\mathcal G}, \f}, \; |U| = n, \mbox{ and \( U \) is good for \( m \)} \}.
\]
It is not difficult to find a context \( U \) that satisfies both \( U \in \inhav{{\mathcal G}, \f} \) and \( S \preceq U \).
For example, assume that \( \f = N_0 \Ar N_3 \) (\( N_0, N_3 \in \nontinfp{\nont} \)).
Then, since the grammar is assumed to be essentially strongly-connected,
there exist \( S_{0,1} \in \inhav{{\mathcal G}, N_0 \Ar N_1} \) and \( S_{2,3} \in \inhav{{\mathcal G}, N_2 \Ar N_3} \)
and then \( U \triangleq S_{2,3}[S[S_{0,1}]] \) satisfies \( U \in \inhav{{\mathcal G}, \f} \) and \( S \preceq U \).
What is relatively difficult is to show that \( S_{2,3} \) and \( S_{0,1} \) can be chosen so that they meet the required size constraints (i.e.,~\( |U| = n \) and \( U \) is good for \( m \)).
The following is a key lemma, which states
that any essentially strongly connected grammar \({\mathcal G}\) is
\emph{periodic} in the sense that there is a constant \(c\) (that depends on \({\mathcal G})\)
and a family of constants \(d_{N,N'}\) such that,
for each \(N,N'\in \nontinfp{\nont}\),
and sufficiently large \(n\),
\( \inhavn{{\mathcal G}, N \Ar N'} \neq \emptyset \) if and only if
\(n\equiv d_{N,N'} \mod c\).
\begin{lem}
\label{lem:period-grammar2}
Let \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) be a regular tree grammar.
Assume that \( {\mathcal G} \) is essentially strongly-connected and \( \card{\inhav{{\mathcal G}}} = \infty \).
Then there exist constants \( n_0, c > 0 \) and a family \( ( d_{N,N'} )_{N, N' \in \nontinfp{\nont}} \) of natural numbers \( 0 \le d_{N,N'} < c \) that satisfy the following conditions:
\begin{enumerate}
\item\label{item:period-grammar2-every}
For every \( N, N' \in \nontinfp{\nont} \), if \( \inhavn{{\mathcal G}, N \Ar N'} \neq \emptyset \), then \( n \equiv d_{N,N'} \mod c \).
\item\label{item:period-grammar2-suff}
The converse of \eqref{item:period-grammar2-every} holds for sufficiently large \( n \):
for any \( N, N' \in \nontinfp{\nont} \) and \( n \ge n_0 \),
if \( n \equiv d_{N,N'} \mod c \) then \( \inhavn{{\mathcal G}, N \Ar N'} \neq \emptyset \).
\item\label{item:period-grammar2-id}
\( d_{N,N} = 0 \) for every \( N \in \nontinfp{\nont} \).
\item\label{item:period-grammar2-comp}
\( d_{N,N'} + d_{N', N''} \equiv d_{N,N''} \mod c \) for every \( N, N', N'' \in \nontinfp{\nont} \).
\end{enumerate}
\end{lem}
The proof of the above lemma is rather involved;
we defer it to Appendix~\ref{sec:periodicity}.
We give some examples below, to clarify what the lemma means.
\begin{exa}
Consider the grammar \({\mathcal G}_1\) consisting of the following rewriting rules:
\[
A \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Ta(B)\qquad B \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tb(A)\qquad A \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tc(C)\qquad C\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Tc
\]
Then \(\nontinfp{\nont}=\set{A,B}\) and the conditions of the lemma above hold for:
\[n_0=0\qquad c=2\qquad d_{A,A}=d_{B,B}=0\qquad d_{A,B}=d_{B,A}=1.\]
In fact, \(\inhav{{\mathcal G}_1,A\Ar A}=\set{(\Ta\Tb)^k\hole \mid k\geq 0}\),
\(\inhav{{\mathcal G}_1,B\Ar B}=\set{(\Tb\Ta)^k\hole \mid k\geq 0}\),
\(\inhav{{\mathcal G}_1,B\Ar A}=\set{(\Ta\Tb)^k\Ta\hole \mid k\geq 0}\),
and
\(\inhav{{\mathcal G}_1,A\Ar B}=\set{(\Tb\Ta)^k\Tb\hole \mid k\geq 0}\).
Here, since the arities of \(\Ta\) and \(\Tb\) are \(1\), we have
used regular expressions to denote \linear{contexts}.
Consider the grammar \({\mathcal G}_2\),
obtained by adding the following rules to the grammar above.
\[ A\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Ta(A_1) \qquad A_1\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Ta(A_2) \qquad A_2\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \Ta(A).\]
Then, the conditions of the lemma above holds for
\(n=6\), \(c=1\), \(d_{N,N'}=0\) for \(N,N'\in \set{A,A_1,A_2,B}\). Note that
\(\inhav{{\mathcal G}_2,A_2\Ar A_1} = \set{\Ta^2(\Ta^3|\Ta\Tb)^k\Ta^2\hole\mid k\geq 0}\).
\qed
\end{exa}
\begin{rem}
With some additional assumptions on a grammar,
Lemma~\ref{lem:period-grammar2} above can be easily proved.
For example, consider a canonical, unambiguous and essentially strongly-connected grammar \( {\mathcal G} \)
and assume that
(i) \({\mathcal G}\) is \(N\)-aperiodic for some \(N\)
and (ii)
there exist a \( 2 \)-context \( C \) and nonterminals \( N, N_1, N_2 \in \nontinfp{\nont} \) such that \( N \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}_{{\mathcal G}}^* C[N_1,N_2] \).
Then Lemma~\ref{lem:period-grammar2} for the grammar \( {\mathcal G} \) trivially holds with \( c = 1 \).
In fact, this simpler approach is essentially what we adopted in the conference version~\cite{SAKT17FOSSACS} of this article.
\end{rem}
Using the lemma above, we prove that
for every \( S \in \ctxinf{{\mathcal G}} \) and any sufficiently large
\(n\) such that \(\inhavnn{n}{{\mathcal G}, \f}\neq \emptyset\),
we can find a context \(U\in \inhavnn{n}{{\mathcal G}, \f}\) such that \( S \preceq U \)
(Lemma~\ref{lem:period-sub} below).
\begin{lem}
\label{lem:period-sub}
Let \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) be a regular tree grammar in canonical form.
Assume that \( {\mathcal G} \) is unambiguous and essentially strongly-connected and \( \card{\inhav{{\mathcal G}}} = \infty \).
Then there exists a constant \( n_0 \in \mathbb{N} \) that satisfies the following condition:
For every
\begin{itemize}
\item \( S \in \ctxinf{{\mathcal G}} \),
\item \( n \ge n_0 + |S| \), and
\item \( \f = ({}\Ar N') \) or \( (N \Ar N') \) where \( N, N' \in \nontinfp{\nont} \),
\end{itemize}
if \( \inhavnn{n}{{\mathcal G}, \f} \neq \emptyset \), then there exists \( U \in \inhavnn{n}{{\mathcal G}, \f} \) with \( S \preceq U \).
\end{lem}
\begin{proof}
First, let us choose the following constants:
\begin{itemize}
\item \( m_0 \in \mathbb{N} \) such that, for every \( N, N' \in {\mathcal N}^{\mathrm{inf}} \),
there exists \( S_{N,N'} \in \inhav{{\mathcal G}, N \Ar N'} \) with \( |S_{N, N'}| \le m_0 \).
The existence of \(m_0\) is a consequence of essential strong-connectivity and of finiteness of \( {\mathcal N} \).
\item \( m_1 \) which is the constant \( n_0 \) of Lemma~\ref{lem:period-grammar2}.
\item \( m_2 \in \mathbb{N} \) such that, for every \( N \in {\mathcal N} \backslash {\mathcal N}^{\mathrm{inf}} \)
and every \( T \in \inhav{{\mathcal G}, N} \), we have \( |T| < m_2 \).
The existence of \(m_2\) follows from the fact that
\( \bigcup_{N \in {\mathcal N} \backslash {\mathcal N}^{\mathrm{inf}}} \inhav{{\mathcal G}, N} \) is a finite set.
\end{itemize}
Let \( c \)
and \( (d_{N, N'})_{N, N' \in {\mathcal N}^{\mathrm{inf}}} \) be the constant and the family obtained by Lemma~\ref{lem:period-grammar2}.
We define \( n_1 \triangleq m_0 + m_1 \) and \( n_2 \triangleq m_0 + m_1 + \mr m_2 + 1 \), where \(\mr \triangleq \max(\image{\Sigma})\). Below we shall show:
(i) the current lemma for the case \( \f = (N \Ar N') \), by setting
\(n_0 \triangleq n_1\) and then (ii) the lemma for the case \( \f = (\Ar N') \),
by setting \(n_0 \triangleq n_2\); we use (i) to show (ii).
The whole lemma
then follows immediately from (i) and (ii) with \(n_0 \triangleq \max\set{n_1,n_2} \, (=n_2)\).
\begin{itemize}
\item
Case (i):
We define \( n_0 \triangleq n_1 = m_0 + m_1 \).
Assume that: \( S \in \inhav{{\mathcal G}, N_1 \Ar N_2} \) where \( N_1, N_2 \in {\mathcal N}^{\mathrm{inf}} \);
\( n \ge n_0 + |S| \); \( \f = (N \Ar N') \) where \( N, N' \in {\mathcal N}^{\mathrm{inf}} \);
and \( \inhavnn{n}{{\mathcal G}, \f} \neq \emptyset \).
Let \( S_1 \in \inhav{{\mathcal G}, N_2 \Ar N'} \) with \( |S_1| \le m_0 \).
Then \( S_1[S] \in \inhav{{\mathcal G}, N_1 \Ar N'} \).
It suffices to show that \( \inhavnn{n - |S_1[S]|}{{\mathcal G}, N \Ar N_1} \neq \emptyset \).
By Lemma~\ref{lem:period-grammar2}\eqref{item:period-grammar2-every}\eqref{item:period-grammar2-comp},
\[
\mspace{20mu}
\begin{aligned}
n &\equiv d_{N, N'} \equiv d_{N, N_1} + d_{N_1, N_2} + d_{N_2, N'} && \bmod c
&&(\because \inhavnn{n}{{\mathcal G}, \f} = \inhavnn{n}{{\mathcal G}, N\Ar N'} \neq \emptyset)
\\
|S_1| &\equiv d_{N_2, N'} && \bmod c
&&(\because (S_1 \in)\, \inhavnn{|S_1|}{{\mathcal G}, N_2 \Ar N'} \neq \emptyset)
\\
|S| &\equiv d_{N_1, N_2} && \bmod c
&&(\because (S \in)\, \inhavnn{|S|}{{\mathcal G}, N_1 \Ar N_2} \neq \emptyset)
\end{aligned}
\]
and thus
\[
n - |S_1[S]| \equiv d_{N, N_1} \mod c.
\]
Since \(n \ge m_0 + m_1 + |S|\), we have \( n - |S_1[S]| \ge m_1 \),
and hence by Lemma~\ref{lem:period-grammar2}\eqref{item:period-grammar2-suff}, we have
\( \inhavnn{n - |S_1[S]|}{{\mathcal G}, N \Ar N_1} \neq \emptyset \).
\item
Case (ii):
We define \( n_0 \triangleq n_2 = m_0 + m_1 + \mr m_2 + 1 \).
Assume that: \( S \in \inhav{{\mathcal G}, N_1 \Ar N_2} \) where \( N_1, N_2 \in {\mathcal N}^{\mathrm{inf}} \);
\( n \ge n_0 + |S| \); \( \f = ({} \Ar N') \) where \( N' \in {\mathcal N}^{\mathrm{inf}} \);
and \( \inhavnn{n}{{\mathcal G}, \f} \neq \emptyset \).
Let \( T \in \inhavnn{n}{{\mathcal G}, \f} \).
Since the arity of terminal symbols are bounded by \( \mr \) and \( |T| \ge \mr m_2 + 1 \),
there exists a subtree \( T_0 \preceq T \) such that \( m_2 \le |T_0| \le \mr m_2 + 1 \),
which can be shown by induction on tree \(T\).
Let \( U \) be a \linear{context} such that \( T = U[T_0] \).
Since \( {\mathcal G} \) is canonical and unambiguous, by Lemma~\ref{lem:first-unique-decomposition},
\( U \in \inhav{{\mathcal G}, N \Ar N'} \) and \( T_0 \in \inhav{{\mathcal G}, N} \) for some \( N \in {\mathcal N} \).
Since \( m_2 \le |T_0| \le \mr m_2 + 1 \), we have \( N \in {\mathcal N}^{\mathrm{inf}} \) and \( |U| \ge m_0 + m_1 + |S| \).
By using the case \((1)\), since \((U \in)\, \inhavnn{|U|}{{\mathcal G},N \Ar N'} \neq \emptyset\),
there exists \(U' \in \inhavnn{|U|}{{\mathcal G},N \Ar N'}\) with \(S \preceq U'\).
Let \(T' \triangleq U'[T_0]\); then \(T' \in \inhavn{{\mathcal G},\Ar N'}\) and \(S \preceq T'\).
\qedhere
\end{itemize}
\end{proof}
\newcommand{\widehatu}[3]{\widehat{\mathcal{U}}_{{#1},{#2}}^{#3}}
\obsolete{ We refine the lemma above, to show that
if
\( \tcompCM{\hhole^{n}_{\f}}{c_0 m} \neq \emptyset \)
(instead of \(\inhavnn{n}{{\mathcal G}, \f}\neq \emptyset\);
note that
\(\tcompCM{\hhole^{n}_{\f}}{c_0 m}\) is the subset of
\(\inhavnn{n}{{\mathcal G}, \f}\neq \emptyset\) consisting of only good contexts),
we can find a context
\(U\in \tcompCM{\hhole^{n}_{\f}}{c_0 m} \) such that \( S \preceq U \).
\begin{lem}\label{lem:period-main}
Let \( {\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R}) \) be a regular tree grammar in canonical form.
Assume that \( {\mathcal G} \) is unambiguous and essentially strongly-connected and
\( \card{\inhav{{\mathcal G}}} = \infty \).\tk{Unambiguity is needed to use the previous lemma.}
Then there exist integers \( m_0, c_0 \ge 1 \) that satisfy the following condition:
For every
\begin{itemize}
\item \( n \ge 1 \),
\item \( m \ge m_0 \),
\item \( S \in \ctxinf{{\mathcal G}} \) where \( |S| \le m \),
and
\item \( \f = ({}\Ar N') \) or \( (N \Ar N') \) where \( N, N' \in \nontinfp{\nont} \),
\end{itemize}
if \( \tcompCM{\hhole^{n}_{\f}}{c_0 m} \neq \emptyset \), then there exists \( U \in \tcompCM{\hhole^{n}_{\f}}{c_0 m} \) with \( S \preceq U \).
\end{lem}
\begin{proof}
Let \( n_0 \) be the constant of the previous lemma (Lemma~\ref{lem:period-sub}) and choose natural numbers \( m_0 \) and \( c_0 \) that satisfy the following conditions:
\begin{enumerate}
\item \( m_0 > n_0 \) and \( m_0 > |T| \) for any \( T \in \bigcup_{N \in {\mathcal N}\backslash\nontinfp{\nont}} \inhav{{\mathcal G},N} \).
\item \( c_0 = 2\mr + 1 \).
\end{enumerate}
Let \( n \ge 1 \), \( m \ge m_0 \), \( S \in \ctxinf{{\mathcal G}} \) where \( |S| \le m \),
and \( \f = ({}\Ar N') \) or \( (N \Ar N') \) where \( N, N' \in \nontinfp{\nont} \).
Suppose that \( \tcompCM{\hhole^{n}_{\f}}{c_0 m} \neq \emptyset \).
Let \( U' \in \tcompCM{\hhole^{n}_{\f}}{c_0 m} \).
Assume that \( U' = a(U_1, \dots, U_{\Sigma(a)}) \).
Since \( U' \) is good for \( c_0 m \), we have
\( |U'| \ge c_0 m = (2\mr + 1)m \ge 2\mr m + 1 \).
Since \( \Sigma(a) \le r \), there exists \( i \le \Sigma(a) \) such that \( |U_i| \ge 2m\).
Since \( U' \in \inhav{{\mathcal G}, \f} \), there exist \(N_1, \dots, N_{\Sigma(a)} \in {\mathcal N}\) such that
\[
N' \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1, \dots, N_{\Sigma(a)}) \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T
\]
where \( T = U'[N] \) if \( \f = (N \Ar N') \) and \( T = U' \) if \( \f = (\Rightarrow N') \).
Since \( |U_i| \ge 2m \ge m_0 \) and by the condition \((1)\) on \(m_0\),
we have \( N_i \in \nontinfp{\nont} \).
Now \( U_i \in \inhav{{\mathcal G}, \f'} \) for \( \f' = N \Ar N_i \) or \( {} \Ar N_i \),
depending on \( \hn{U_i} \).
Since \( |U_i| \ge 2m \ge n_0 + |S| \), by Lemma~\ref{lem:period-sub},
there exists \( U_i' \in \inhav{{\mathcal G}, \f'} \) such that \( S \preceq U_i' \) and \( |U_i'| = |U_i| \).
Since \( |U_i| = |U'_i| \) and \( U' \) is good for \( c_0 m \),
\( U \triangleq a(U_1, \dots, U_{i-1}, U'_i, U_{i+1}, \dots, U_{\Sigma(a)}) \) is also good for \( c_0 m \).
Obviously \( |U| = |U'| \) and thus \( U \in \tcompCM{\hhole^{n}_{\f}}{c_0 m} \).
Since \( S \preceq U_i' \) and \( U_i' \preceq U \), we have \( S \preceq U \) as required.
\end{proof}
}
We prepare another lemma.
\begin{lem}
\label{lem:remainsHigh}
If \(\decomp(T) = (U,E,P)\), \(\ell \in \set{1,\dots,\shn(E)}\),
\(\prj{P}{\ell}\) is a \linear{context}, and
the second-order hole \(\sh{E}{\ell}\) occurs in \(E\) in the form \((\sh{E}{\ell})[E']\),
then \(|E'| \ge m\).
\end{lem}
\begin{proof}
By straightforward induction on \(|T|\) and case analysis of
Equation~\eqref{eq:defOfDecomp} in the definition of \(\decomp\).
We can use Lemma~\ref{lem:size-of-E} in the second case of
Equation~\eqref{eq:defOfDecomp} when \(\ell=1\);
all the other cases immediately follow from induction hypothesis.
\end{proof}
We are now ready to prove the main lemma of this subsection.
\nk{I have removed an intermediate lemma and merged it with
the proof of the main lemma below; I do understand the motivation for the intermediate lemma,
but renaming the constants \(m_0,n_0,...\) so many times seemed to
make the proof more difficult to read.}
\begin{proof}[Proof of Lemma~\ref{lemma:adjust}]
Since \(S_0 \in \ctxinf{{\mathcal G}}\), we have
\(\nontinfp{\nont} \neq \emptyset\) and hence \(\card{\inhav{{\mathcal G}}} = \infty\).
Let \( n_0\) be the constant of Lemma~\ref{lem:period-sub}.
Let \( c_1 \) be a positive integer such that \( |S_n| \le c_1 n \) for every \( n \in \mathbb{N} \).
We define \( c \triangleq (2r+1)c_1 \) (recall that
\(r\) is the largest arity of \(\Sigma\)) and choose \(b\) so that \(b \ge n_0\) and
\( b > |T| \) for any \( T \in \bigcup_{N \in {\mathcal N}\backslash\nontinfp{\nont}} \inhav{{\mathcal G},N} \).
Assume that \(N \in \nontinfp{\nont}\), \( n \ge 1 \), \( m \ge b \), \( \widetilde{E} \in \tframeNM{n}{cm}({\mathcal G},N) \)
and \( i \in \set{1, \dots, \shn(\widetilde{E})} \).
Let \( \sh{\widetilde{E}}{i} = \hhole^{n'}_{\f'} \).
We need to show that there exists
\( U \in \tcompCM{\hhole^{n'}_{\f'}}{cm}\) such that $S_{m} \preceq U$.
Since \( \widetilde{E} \in \tframeNM{n}{cm}({\mathcal G},N) \), there exist \(T \in \inhavn{{\mathcal G}, N}\) and \(P\) such that
\((\widetilde{E}, P) = \closedgdecompm{cm}(T,N)\);
hence \(\prj{P}{i} \in \tcompCM{\sh{\widetilde{E}}{i}}{cm} = \tcompCM{\hhole^{n'}_{\f'}}{c m}\).
The affine context \(\prj{P}{i}\) must be of the form
\( a(U_1, \dots, U_{\Sigma(a)}) \). Since \(\prj{P}{i}\) is good for \(cm\),
we have \(|\prj{P}{i}|\geq cm = (2r+1)c_1m\geq 2rc_1m+1\).
Since \(\Sigma(a)\leq r\), there exists \(j\leq \Sigma(a)\) such that
\(|U_j|\geq 2c_1m\).
\new{We have \(\f' = (N' \Ar N'')\) or \((\Ar N'')\) for some \(N', N'' \in {\mathcal N}\).}
Since \( P.i = a(U_1, \dots, U_{\Sigma(a)}) \in \inhav{{\mathcal G}, \f'} \), there exist \(N_1, \dots, N_{\Sigma(a)} \in {\mathcal N}\) such that
\[
N'' \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} a(N_1, \dots, N_{\Sigma(a)}) \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T'
\]
where \( T' = P.i[N'] \) if \( \f' = (N' \Ar N'') \),
and \( T' = P.i \) if \( \f' = (\Ar N'') \).
\new{Let \(\f\) be \(N'\Ar N_j\) if \(U_j\) is a \linear{context}, and \(\Ar N_j\) otherwise.}
Then \(U_j\in \inhavnn{|U_j|}{{\mathcal G},\f}\neq \emptyset\).\asd{
To show this sentence, I changed the definition of \(\f\).
I could not check if \(U_j\) is always a \linear{context} in the case that \(\f' = (N' \Ar N'')\).
If \(U_j\) is always a \linear{context}, the change of the definition is meaningless but also harmless,
and if \(U_j\) is not necessarily a \linear{context}, the change is necessary;
anyway, the change is harmless and the current proof is correct.}
In order to apply Lemma~\ref{lem:period-sub} (for \(S=S_m\) \new{and \(n= |U_j|\)}),
we need to check the conditions: (i) \(|U_j| \geq n_0+|S_m|\)
and (ii) \(\f\) consists of only nonterminals in \(\nontinfp{\nont}\).
Condition (i) follows immediately from \(|U_j|\geq 2c_1m
\geq m+c_1m \geq n_0+|S_m|\). As for (ii), it suffices to check that
\(\inhav{{\mathcal G},N_j}\) and \(\inhav{{\mathcal G},N'}\)
contain a tree whose size is no less than \(b\)
(where the condition on \(\inhav{{\mathcal G},N'}\) is required only if
\new{\(U_j\) is a \linear{context}}).
The condition on \(\inhav{{\mathcal G},N_j}\) follows
from \(|U_j|\geq 2c_1m\geq m\geq b\).
If
\new{\(U_j\) is a \linear{context},}
by Lemma~\ref{lem:remainsHigh},
there exist \(S\) and \(T''\)
such that:
\[
T=S[(\prj{P}{i})[T'']]
\qquad
|T''| \ge cm\ge b
\qquad
T'' \in \inhav{{\mathcal G},N'}
\]
as required.
Thus, we can apply Lemma~\ref{lem:period-sub}
and obtain \(U_j' \in \inhav{{\mathcal G},\f}\) such that \(S_m \preceq U_j'\) and \(|U_j'|=|U_j|\).
Since \(P.i = a(U_1,\ldots,U_{\Sigma(a)})\) is good for \( c m \),
\( U \triangleq a(U_1, \dots, U_{j-1}, U'_j, U_{j+1}, \dots, U_{\Sigma(a)}) \) is also good for \( c m \).
Obviously \( |U| = |P.i| \) and thus \( U \in
\tcompCM{\hhole^{n'}_{\f'}}{cm}\).
Since \( S_m \preceq U_j' \) and \( U_j' \preceq U \), we have \( S_m \preceq U \) as required.
\end{proof}
\subsection{Main Proof}\label{sec:proofimfmain}
Here, we give a proof of Theorem~\ref{thm:imt}.
Before the proof, we prepare a simple lemma.
We write \( \cszon{\Sigma} \) for the set of \affine{contexts} over \( \Sigma \) of size at most \( n \).
Lemma~\ref{lem:upBound} below gives an upper bound of \( \card{\cszon{\Sigma}} \).
A more precise bound can be obtained by using a technique of analytic combinatorics
such as Drmota--Lalley--Woods theorem~({\it cf.\;} \cite{Flajolet}, Theorem~VII.6),
but the rough bound provided by the lemma below is sufficient for our purpose.
\begin{lem}
\label{lem:upBound}
For every ranked alphabet \( \Sigma \), there exists a real constant \( \gamma > 1 \) such that
\[
\card{\cszon{\Sigma}} \le \gamma^n
\]
for every \( n \ge 0 \).
\end{lem}
\begin{proof}
Let \(A\) be the set of symbols:
\(
\{ \grave{a}, \acute{a} \mid a \in \mathop{\mathrm{Dom}}(\Sigma) \} \cup \{\, \square \}
\).
Intuitively \( \grave{a} \) and \( \acute{a} \) are opening and closing tags of an XML-like language.
We can transform an affine context \(U\) to its XML-like string representation \(U^\dagger\in A^*\) by:
\begin{align*}
(a(U_1, \dots, U_{\Sigma(a)}))^\dagger &\triangleq \grave{a}\, U_1^\dagger \cdots U_{\Sigma(a)}^\dagger \,\acute{a} \\
\hole^\dagger &\triangleq \square.
\end{align*}
Obviously, \((\cdot)^\dagger\) is injective. Furthermore, \(|U^\dagger| = 2|U|\) if \(U\) is a
\(0\)-context (i.e., a tree), and \(|U^\dagger| = 2|U|+1\) if \(U\) is a \linear{context}
(note that the size of the hole \(\hole\) is zero, but its word representation is of length \( 1 \)).
Thus, for \(n>0\), we have
\[
\card{\cszon{\Sigma}} \le
\sum_{i = 0}^{2n+1} (\card{A})^i
\le
\sum_{i = 0}^{3n} (\card{A})^i
\le (\card{A} + 1)^{3n} = ((\card{A}+1)^3)^n.
\]
If \(n=0\), then \(\card{\cszon{\Sigma}}=1=((\card{A}+1)^3)^n\),
as \(\cszon{\Sigma}\) is the singleton set \(\set{\hole}\).
Thus, the required result holds for \(\gamma=(\card{A}+1)^3\).
\end{proof}
The following lemma is a variant of
Theorem~\ref{thm:imt}, specialized to a canonical grammar.
\begin{lem}
\label{lem:imt}
Let ${\mathcal G} = (\Sigma, {\mathcal N}, {\mathcal R})$ be a canonical, unambiguous, and essentially strongly-connected regular tree grammar
such that \(\card{\inhav{{\mathcal G}}}=\infty\),
and \((S_n)_{n \in \mathbb{N}}\) be a family of \linear{contexts} in \(\ctxinf{{\mathcal G}}\)
such that $|S_n| = \mathrm{O}(n)$.
Then there exists a real constant \(\cn > 0\) such that for any \(N \in \nontinfp{\nont}\), \[
\limd_{n \rightarrow \infty} \frac{\card{\{ T \in \inhavn{{\mathcal G},N} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\inhavn{{\mathcal G},N}}} = 1.
\]
\end{lem}
\begin{proof}
\nk{The previous proof is found in ``oldimdproof.tex''.}
The overall structure of the proof is the same as that of Proposition~\ref{prop:monkey}.
Let \(Z_n\) be
\[
1- \frac{\card{\{ T \in \inhavn{{\mathcal G},N} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\inhavn{{\mathcal G},N}}}
=
\frac{\card{\{ T \in \inhavn{{\mathcal G},N} \mid
S_{\ceil{\cn \log n}} \not\preceq T
\}}}{\card{\inhavn{{\mathcal G},N}}}.
\]
It suffices to show that \(Z_n\) converges to \(0\).
By Lemma~\ref{lem:grm-bijection},
we have
\[ \inhavn{{\mathcal G},N}\cong
\coprod_{\widetilde{E} \in \tframeNM{n}{m}({\mathcal G}, N)}
\prod_{i=1}^{\shn(\widetilde{E})} \tcompCM{\sh{\widetilde{E}}{i}}{m}
\]
for any \(m>0\).
Thus, we have
\begin{equation}
Z_n \leq
\frac{\sum_{\widetilde{E} \in \tframeNM{n}{m}({\mathcal G}, N)}
\prod_{i=1}^{\shn(\widetilde{E})}
\card{\set{U\in \tcompCM{\sh{\widetilde{E}}{i}}{m}\mid
S_{\ceil{\cn \log n}} \not\preceq U}}
}
{\sum_{\widetilde{E} \in \tframeNM{n}{m(n)}({\mathcal G}, N)} \prod_{i=1}^{\shn(\widetilde{E})} \card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}}. \label{zn}
\end{equation}
Let \(b, c\geq 1\) be the numbers in Lemma~\ref{lemma:adjust}.
Then, by the lemma, for any \(n\) \new{and \(p\)} such that \(\ceil{\cn \log n}\geq b\),
each \(\tcompCM{\sh{\widetilde{E}}{i}}{c\ceil{\cn \log n}}\) contains at least one \(U\)
that has \(S_{\ceil{\cn \log n}}\) as a subcontext.
Thus, for \(m=c\ceil{\cn\log n}\),
\( \card{\set{U\in \tcompCM{\sh{\widetilde{E}}{i}}{m}\mid
S_{\ceil{\cn \log n}} \not\preceq U}}\) is bounded above by:
\[
\begin{array}{ll}
\card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}-1 = (1-\displaystyle\frac{1}{\card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}})\card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}
\leq
(1-\displaystyle\frac{1}{\gamma^{r m}})\card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}.
\end{array}
\]
Here \(\gamma\) is the constant (that only depends on \(\Sigma\))
of Lemma~\ref{lem:upBound}, and \(r\) is the largest arity of \(\Sigma\). In
the last inequality, we have used the fact that
\(\tcompCM{\sh{\widetilde{E}}{i}}{m}\subseteq \cszon[\mr m]{\Sigma}\).
By using the upper-bound above, Equation \eqref{zn}, and \(\shn(\widetilde{E})\geq\frac{n}{2rm}\)
(Lemma~\ref{lemma:partition}),
we have:
\[
\begin{array}{l}
Z_n \leq
\dfrac{\sum_{\widetilde{E} \in \tframeNM{n}{m}({\mathcal G}, N)}
\prod_{i=1}^{\shn(\widetilde{E})}
(1-\displaystyle\frac{1}{\gamma^{r m}})\card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}}
{\sum_{\widetilde{E} \in \tframeNM{n}{m(n)}({\mathcal G}, N)} \prod_{i=1}^{\shn(\widetilde{E})} \card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}} \\
= \dfrac{\sum_{\widetilde{E} \in \tframeNM{n}{m}({\mathcal G}, N)}
(1-\displaystyle\frac{1}{\gamma^{r m}})^{\shn(\widetilde{E})} \prod_{i=1}^{\shn(\widetilde{E})}
\card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}}
{\sum_{\widetilde{E} \in \tframeNM{n}{m(n)}({\mathcal G}, N)} \prod_{i=1}^{\shn(\widetilde{E})} \card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}} \\
\leq \dfrac{\sum_{\widetilde{E} \in \tframeNM{n}{m}({\mathcal G}, N)}
(1-\displaystyle\frac{1}{\gamma^{r m}})^{\frac{n}{2rm}} \prod_{i=1}^{\shn(\widetilde{E})}
\card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}}
{\sum_{\widetilde{E} \in \tframeNM{n}{m(n)}({\mathcal G}, N)} \prod_{i=1}^{\shn(\widetilde{E})} \card{\tcompCM{\sh{\widetilde{E}}{i}}{m}}}
=
(1-\dfrac{1}{\gamma^{r m}})^{\frac{n}{2rm}}
\end{array}
\]
for any \(n\) \new{and \(p\)} such that \(\ceil{\cn \log n}\geq b\) and \(m=c\ceil{\cn\log n}\).
It remains to choose \(\cn\) so that
\((1-\dfrac{1}{\gamma^{r m}})^{\frac{n}{2rm}} =
(1-\dfrac{1}{\gamma^{r c\ceil{\cn\log n}}})^{\frac{n}{2rc\ceil{\cn\log n}}}\)
converges to \(0\).
Let us choose positive real numbers \(a\), \(p\), and \(q\) so that
\(p\) and \(q\) satisfy the following conditions for every \(n \ge a\):
\begin{align}
{\cn \log n} &\ge b
\label{eq:foradjust}
\\
\gamma^{\mr c\ceil{\cn \log n}} &\le n^q
\label{eq:toconv}
\\
q &< 1.
\label{eq:forconv}
\end{align}
For example,
we can choose \(a,\, p,\, q\) as follows:
\begin{align*}&
a = \max\set{\gamma^{b({\mr c+2})}, \gamma^{\mr c({\mr c+2})}}
\qquad
\cn = \fr{1}{(\mr c+2)\log\gamma}
\qquad
q = \fr{\mr c+1}{\mr c+2}
\end{align*}
In fact, condition \eqref{eq:toconv} follows from:
\[
\begin{array}{l}
\mr c\ceil{\cn \log n} - \log_\gamma n^q
\leq
\mr c(\fr{1}{(\mr c+2)\log\gamma} \log n+1) - \fr{\mr c+1}{\mr c+2}\frac{\log n}{\log{\gamma}}\\
= \frac{1}
{(rc+2)\log\gamma}
(rc\log n+ rc(rc+2)\log \gamma - (rc+1)\log n)\\
= \frac{1}
{(rc+2)\log\gamma}
(rc(rc+2)\log \gamma - \log n)\\
\leq \frac{1}
{(rc+2)\log\gamma}
(rc(rc+2)\log \gamma - \log a)\leq 0.\\
\end{array}
\]
Thus, for \(n\geq a\), we have:
\[
(1-\dfrac{1}{\gamma^{r c\ceil{\cn\log n}}})^{\frac{n}{2rc\ceil{\cn\log n}}}
\leq
(1-\dfrac{1}{n^q})^{\frac{n}{2rc\ceil{\cn\log n}}}
\leq
\left(\left(1-\fr{1}{n^{q}}\right)^{\fr{n}{\log n}}\right)^{\fr{1}{3\mr c p}}.
\]
Since \(\lim_{n \rightarrow \infty} \left(1-\fr{1}{n^{q}}\right)^{\fr{n}{\log n}} = 0\),
we have \(\limd\limits_{n\rightarrow\infty} \zps{n} =0\) as required.
\end{proof}
\begin{rem}
\label{rem:convergence}
In the proof above, we used
the fact that if \(0<q<1\) then
\[
\lim_{n\tends \infty}\left(1-\fr{1}{n^q}\right)^{\fr{n}{\log n}} = 0.
\]
We also remark that if \(q\ge 1\) then
\[
\lim_{n\tends \infty}\left(1-\fr{1}{n^q}\right)^{\fr{n}{\log n}} = 1.
\]
Thus, \(p\) should be chosen to be sufficiently small so that Equation (\ref{eq:toconv}) in the proof
holds for some \(q<1\).
\end{rem}
We are now ready to prove Theorem~\ref{thm:imt}.
We restate the theorem.
\renewcommand{\thethmNN}{\ref{thm:imt}}
\begin{thmNN}[\thmstatementIMTcap]
\thmstatementIMT
\end{thmNN}
\begin{proof}
By Lemma~\ref{lem:canonical-grammar}, there exists a canonical, unambiguous and essentially
strongly-connected grammar \({\mathcal G}' = (\Sigma, {\mathcal N}', {\mathcal R}')\) and
a family \( ( \nset_N )_{N \in {\mathcal N}} \) of subsets \( \nset_N \subseteq {\mathcal N}' \)
such that
\( \inhav{{\mathcal G},N} = \biguplus_{N' \in \nset_N} \inhav{{\mathcal G}',N'} \) for every \(N \in {\mathcal N}\)
and \( \ctx{{\mathcal G}} \subseteq \ctxinf{{\mathcal G}'} \).
Let \(\nset'_N \triangleq \nset_N \cap \nontinfp{{\mathcal N}'} = \set{N'\in\nset_N\mid \inhav{{\mathcal G}',N'}=\infty}\).
For any \(N \in {\mathcal N}\),
since \(\inhav{{\mathcal G}}=\infty\) and \({\mathcal G}\) is strongly connected,
we have \(\inhav{{\mathcal G},N}=\infty\), and hence
\(\nset'_N\neq \emptyset\). By Lemma~\ref{lem:imt}, there exists a real constant \(p>0\) such that, for each
\(N'\in \nset'_N\),
\begin{align}
\limd_{n \rightarrow \infty} \frac{\card{\{ T \in \inhavn{{\mathcal G}',N'} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\inhavn{{\mathcal G}',N'}}} = 1.
\label{eq:lim}
\end{align}
Thus, we have
\begin{align*}
&\
\limd\limits_{n \rightarrow \infty} \frac{\card{\{ T \in \inhavn{{\mathcal G},N} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\inhavn{{\mathcal G},N}}}
\\ =&\
\limd\limits_{n \rightarrow \infty} \frac{\sum_{N'\in\nset_N}\card{\{ T \in \inhavn{{\mathcal G}',N'} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\sum_{N'\in\nset_N}\card{\inhavn{{\mathcal G}',N'}}}
\\ =&\
\limd\limits_{n \rightarrow \infty} \frac{\sum_{N'\in\nset'_N}\card{\{ T \in \inhavn{{\mathcal G}',N'} \mid
S_{\ceil{\cn \log n}} \preceq T
\}}}{\sum_{N'\in\nset'_N}\card{\inhavn{{\mathcal G}',N'}}}
\\ =&\
1
\end{align*}
as required; we have used Lemma~\ref{lem:convSum} and
Equation~\eqref{eq:lim} in the last step.
\end{proof}
\section{Proof of the Main Theorem on \(\lambda\)-calculus}\label{sec:proofbeta}
This section proves our main theorem (Theorem~\ref{mainthm}). We first prepare
a regular tree grammar that generates the set of tree representations of
elements of $\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ in Section~\ref{sec:grammar-for-lambda}, and then
apply Corollary~\ref{cor:imtTree} to obtain Theorem~\ref{mainthm},
where \((T_n)_{n \in \mathbb{N}}\) in the corollary are set to (the tree representations of)
the terms in the introduction that have long \(\beta\)-reduction sequences.
\subsection{Regular Tree Grammar $\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ of $\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$}
\label{sec:grammar-for-lambda}
Recall that $\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ is the set of (\(\alpha\)-equivalence classes of)
closed well-typed terms, whose order, internal arity,
and number of variables are bounded above by \(\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#\), \(\I\), and \(\K\)
(consult Definition~\ref{df:set-of-terms} for the precise definition).
The set $\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ can be generated by the following grammar (up to isomorphism).
\begin{defi}[grammar of $\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$]
Let $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,\I,\K \geq 0$ be integers and $X_\K = \{x_1, \ldots, x_\K\}$ be a subset of $V$.
The regular tree grammar \(\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) is defined as
\((\lralph{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}, \lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}, \lrules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K})\) where:
\begin{align*}
\lralph{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \triangleq& \ \{ x \mapsto 0 \mid x \in X_\K \} \cup \{@ \mapsto
2\}\\
\cup & \ \{ \lambda \overline{x}^\tau \mapsto 1 \mid
\overline{x} \in \{{\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]\} \cup X_\K,\ \tau \in \types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#-1}{\I}
\}\\
\lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \triangleq & \ \{\NT{\Gamma}{\tau} \mid
\tau \in \types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}, \mathop{\mathrm{Dom}}(\Gamma) \subseteq X_\K,\ \image{\Gamma} \subseteq \types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#-1}{\I},
\\
& \ \qquad\qquad\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\neq\emptyset\} \\
\lrules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \triangleq & \ \{\NT{\{x_i:\tau\}}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} x_i \}
\cup \{ \NT{\Gamma}{\sigma \ft \tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma}
(\NT{\Gamma}{\tau}) \}\\
\cup & \ \{ \NT{\Gamma}{\sigma \ft \tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_i^{\sigma}
(\NT{\Gamma\cup\set{x_i:\sigma}}{\tau}) \mid i = \min\{ j \mid x_j\in X_\K\setminus \mathop{\mathrm{Dom}}(\Gamma)\}\}
\\ \cup & \
\{\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} @(\NT{\Gamma_1}{\sigma \ft \tau},
\NT{\Gamma_2}{\sigma}) \mid \Gamma = \Gamma_1 \cup \Gamma_2\}
\end{align*}
\end{defi}
The grammar above generates the tree representations of elements of $\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$,
where a variable \(x\), a lambda-abstraction, and an application
are represented respectively as the nullary tree constructor \(x\), unary tree constructor
\(\lambda \overline{x}\), and binary tree constructor \(@\).
The nonterminal \(\NT{\Gamma}{\tau}\) is used to generate (the tree representations of) the elements of
\(\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\); the condition
\(\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\neq\emptyset\) on nonterminal \(\NT{\Gamma}{\tau}\) ensures that
every nonterminal generates at least one tree.
To guarantee that the grammar generates at most one tree for each
\(\alpha\)-equivalence class \([t]_\alpha\), (i) variables are chosen from the fixed set \(X_\K\),
and (ii) in the rule for generating a \(\lambda\)-abstraction, a variable is chosen in a deterministic manner.
Note that \(\lralph{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\), \(\lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) and \(\lrules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) are finite.
The finiteness of \(\lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) follows from that of
\(X_\K\), \(\types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#-1}{\I}\), and
\(\set{\Gamma \mid \mathop{\mathrm{Dom}}(\Gamma) \subseteq X_\K,\ \image{\Gamma} \subseteq
\types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#-1}{\I}}\). The finiteness of \(\lrules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) also follows
immediately from that of \(\lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\).
\begin{exa}\label{exam:lambdaGrammar}
Let us consider the case where $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\# = \I = \K = 1$.
The grammar $\lgrm{1}{1}{1}$ consists of the following components.
\begin{align*}
\lralph{1}{1}{1} &= \{ x_1, @, \lambda x_1^\mathtt{o}, \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^\mathtt{o} \}
\qquad
\lnont{1}{1}{1} = \{ \NT{\emptyset}{\mathtt{o} \ft \mathtt{o}}, \NT{\{x_1: \mathtt{o}\}}{\mathtt{o}},
\NT{\{x_1: \mathtt{o}\}}{\mathtt{o} \ft \mathtt{o}} \}\\
\lrules{1}{1}{1} &=
\begin{cases}
\NT{\emptyset}{\mathtt{o} \ft \mathtt{o}} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_1^\mathtt{o}(\NT{\{x_1:\mathtt{o}\}}{\mathtt{o}})\\
\NT{\{x_1: \mathtt{o} \}}{\mathtt{o}} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}
x_1 \mid @(\NT{\{x_1: \mathtt{o}\}}{{\mathtt{o} \ft \mathtt{o}}},\NT{\{x_1: \mathtt{o}\}}{\mathtt{o}}) \mid
@(\NT{\emptyset}{\mathtt{o} \ft \mathtt{o}},\NT{\{x_1: \mathtt{o}\}}{\mathtt{o}})
\\
\NT{\{x_1: \mathtt{o}\}}{\mathtt{o} \ft \mathtt{o}} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^\mathtt{o}(\NT{\{x_1: \mathtt{o} \}}{\mathtt{o}}).
\end{cases}
\end{align*}
\end{exa}
There is an obvious embedding $\embp{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ ($e$ for short) from trees in
$\TR{\lralph{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}$ into (not necessarily well-typed) $\lambda$-terms.
For $\NT{\Gamma}{\tau} \in \lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ we define
\[
\tecp{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}{\Gamma}{\tau} \triangleq [-]_\alpha \circ e:
\inhav{\NT{\Gamma}{\tau}} \fa \termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}
\]
where $[-]_\alpha$ maps a term to its $\alpha$-equivalence class.
We sometimes omit the superscript and/or the subscript and may write just \(\pi\) for
\(\tecp{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}{\Gamma}{\tau}\).
The following lemma says that
$\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ gives a complete representation system of the \(\alpha\)-equivalence classes.
\begin{lem}\label{prop:cong}
For $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,\I,\K \geq 0$,
$\ \tecp{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}{\Gamma}{\tau}:
\inhav{\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},\NT{\Gamma}{\tau}} \fa \termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ is a size-preserving bijection.
\end{lem}
\begin{proof}
It is trivial that the image of \(\pi_{\typing{\Gamma}{\tau}}\) is contained
in \(\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) and \(\pi\) preserves the size.
The injectivity,
i.e.,\; \(e(T) \aeq e(T')\) implies \(T=T'\) for \(T, T' \in \inhav{\NT{\Gamma}{\tau}}\),
is shown by induction on
the length of the leftmost rewriting
sequence \(\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T\)
and by case analysis of the rewriting rule used in the first step of the reduction sequence.
In the case analysis below, we use the fact that if \(T\in \inhav{\NT{\Gamma}{\tau}}\), then
\(\mathbf{FV}(e(T))=\mathop{\mathrm{Dom}}(\Gamma)\), which can be proved by straightforward induction.
\begin{itemize}
\item Case for the rule $\NT{\{x_i:\tau\}}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} x_i$: In this case, \(T=x_i\). By the assumption
\(e(T)\aeqe(T')\), \(T=T'\) follows immediately.
\item Case for the rule \(\NT{\Gamma}{\sigma \ft \tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}
\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma} (\NT{\Gamma}{\tau}) \): In this case, \(T= \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma} (T_1)\) with
\(\NT{\Gamma}{\tau}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T_1\). By the assumption
\(e(T)\aeqe(T')\), \(T'\) must be of the form \(\lambda \overline{x}^\sigma(T_1')\)
where (i) \(\overline{x}\) does not occur free in \(e(T_1')\), and (ii) \(e(T_1)\aeqe(T_1')\).
Then \(\overline{x} = {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]\) and hence we have \(\NT{\Gamma}{\sigma \ft \tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma} (\NT{\Gamma}{\tau}) \) with \(\NT{\Gamma}{\tau}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T_1'\),
because, if \(\overline{x} \neq {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]\), we have
\(\NT{\Gamma}{\sigma \ft \tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda \overline{x}^{\sigma} (\NT{\Gamma\cup\set{\overline{x}:\sigma}}{\tau}) \) with
\(\NT{\Gamma\cup\set{\overline{x}:\sigma}}{\tau}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T_1'\), and hence
\(\mathbf{FV}(e(T'_1))=\mathop{\mathrm{Dom}}(\Gamma\cup\set{\overline{x}:\sigma})\), which contradicts the condition (i).
Therefore, by the induction hypothesis, we have \(T_1=T_1'\), which implies \(T=T'\) as required.
\item Case for the rule:
\( \NT{\Gamma}{\sigma \ft \tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_i^{\sigma}
(\NT{\Gamma\cup\set{x_i:\sigma}}{\tau})\), where \(i = \min
\set{ j \mid x_j\in X_\K\setminus \mathop{\mathrm{Dom}}(\Gamma)}\).
By the assumption
\(e(T)\aeqe(T')\), \(T'\) must be of the form \(\lambda x^{\sigma}(T_1')\) where \(x\) occurs free in \(e(T_1')\),
and \([x'/x_i]e(T_1)\aeq [x'/x]e(T_1')\) for a fresh variable \(x'\).
Thus, by the definition of \(\lrules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\),
\(T'\) must have also been generated by the same rule, i.e., \(T'=\lambda x_i^{\sigma}(T_1')\)
with \(\NT{\Gamma\cup\set{x_i:\sigma}}{\tau}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T_1'\).
By the induction hypothesis, we have \(T_1=T_1'\), which also implies \(T=T'\) as required.
\item
Case for the rule: \(\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} @(\NT{\Gamma_1}{\sigma \ft \tau}, \NT{\Gamma_2}{\sigma})\)
where \(\Gamma = \Gamma_1 \cup \Gamma_2\).
Then \(T = @(T_1,T_2)\) with \(\NT{\Gamma_1}{\sigma\ft\tau}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* T_1\)
and \(\NT{\Gamma_2}{\sigma}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*T_2\). By the assumption
\(e(T)\aeqe(T')\), \(T'\) must also be of the form \(@(T_1',T_2')\), with
\(e(T_1)\aeq e(T_1')\) and \(e(T_2)\aeq e(T_2')\).
Therefore, \(T'\) must also have been generated by a rule for applications;
hence
\(\NT{\Gamma'_1}{\sigma' \ft \tau}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*T_1'\) and \(\NT{\Gamma'_2}{\sigma'}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*T_2'\) for some
\(\Gamma_1', \Gamma_2', \sigma'\) such that \(\Gamma = \Gamma'_1 \cup \Gamma'_2\).
By the condition \(e(T_i)\aeqe(T_i')\),
\(\mathop{\mathrm{Dom}}(\Gamma_i)=\mathbf{FV}(e(T_i))=\mathbf{FV}(e(T_i'))=\mathop{\mathrm{Dom}}(\Gamma'_i)\) for each \(i\in\set{1,2}\). Thus,
since \(\Gamma_1 \cup \Gamma_2 = \Gamma'_1\cup\Gamma'_2\),
we have \(\Gamma_i=\Gamma_i'\), which also implies \(\sigma=\sigma'\) (since the type of a simply-typed \(\lambda\)-term
is uniquely determined by the term and the type environment).
Therefore, by the induction hypothesis, we have \(T_i=T_i'\) for each \(i\in\set{1,2}\), which implies
\(T=T'\) as required.
\end{itemize}
Next we show the surjectivity, i.e.,\; for any \([t]_\alpha \in \termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) there exists
\(T \in \inhav{\NT{\Gamma}{\tau}}\) such that \(e(T) \aeq t\).
For \(\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#, \I, \K\) with \(X_\K = \{x_1, \ldots, x_\K\}\) and \(\NT{\Gamma}{\tau} \in \lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\),
we define a ``renaming'' function
\(\rnp{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}{\Gamma}{\tau}\) (\(\rn_{\typing{\Gamma}{\tau}}\) or \(\rn\) for short) from \(\set{t\mid [t]_\alpha \in
\termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}\)
to \(\inhav{\NT{\Gamma}{\tau}}\) by induction on the size of \(t\),
so that \(\pi{}(\rn(t)) = [t]_\alpha\) holds.
\[
\newcommand\localspace{\mspace{-50mu}}
\newcommand\localspacee{\mspace{-220mu}}
\begin{array}{ll}
\rn_{\typing{\set{x:\tau}}{\tau}}(x) &\triangleq x
\\
\rn_{\typing{\Gamma}{\sigma\ft\tau}}(\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma}\!.t) &\triangleq \lambda
{\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^\sigma(\rn_{\typing{\Gamma}{\tau}}(t))
\\
\rn_{\typing{\Gamma}{\sigma\ft\tau}}(\lambda x^{\sigma}\!.t) &\triangleq \lambda
{\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^\sigma(\rn_{\typing{\Gamma}{\tau}}(t)) \qquad\qquad\hfill(\text{if
}x\notin\mathbf{FV}(t)) \\
\rn_{\typing{\Gamma}{\sigma\ft\tau}}(\lambda x^{\sigma}\!.t) &\triangleq
\lambda
x_i^\sigma(\rn_{\typing{\Gamma\cup\set{x_i:\sigma}}{\tau}}(\swap{t}{x_i}{x}))
\text{ where \(i \triangleq \min\{ j \mid x_j\in X_\K\setminus
\mathop{\mathrm{Dom}}(\Gamma)\}\)}
\\
&\hfill (\text{if }x\in\mathbf{FV}(t)) \\
\rn_{\typing{\Gamma_1\cup\Gamma_2}{\tau}}(t_1t_2) &\triangleq
@(\rn_{\typing{\Gamma_1}{\sigma\ft\tau}}(t_1),
\rn_{\typing{\Gamma_2}{\sigma}}(t_2))
\hfill(\text{if }
\Gamma_1 \p t_1: \sigma \ft \tau \text{ and } \Gamma_2 \p t_2: \sigma)
\end{array}
\]
Here, \(\swap{t}{x}{y}\) represents the term obtained by swapping every occurrence of \(x\) with that of \(y\);
for example, \(\swap{(\lambda x.xy)}{x}{y} = \lambda y.yx\). Note that in the last clause,
if \(\Gamma\p t_1t_2:\tau\), then there exists a unique triple \((\Gamma_1,\Gamma_2,\sigma)\) such that
\(\Gamma=\Gamma_1\cup\Gamma_2\),
\( \Gamma_1 \p t_1: \sigma \ft \tau\), and \(\Gamma_2 \p t_2: \sigma\).
We can prove that \(\pi{}(\rn(t)) = [t]_\alpha\) holds for every \(\lambda\)-term \(t\) such that
\([t]_\alpha\in \termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\),
by straightforward induction on the size of $t$.
\end{proof}
\begin{lem}\label{lem:unamb}
For $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,\I,\K \geq 0$, \(\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) is unambiguous.
\end{lem}
\begin{proof}
The proof is similar to that of the injectivity of Lemma~\ref{prop:cong}.
We show that, for any \(\NT{\Gamma}{\tau} \in \lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\), \(T \in \TR{\lralph{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}\),
and two leftmost rewriting sequences of the form
\begin{align*}&
\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(1)} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(n)} = T \qquad(n \ge 1)
\\&
\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T'^{(1)} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T'^{(n')} = T \qquad(n' \ge 1),
\end{align*}
\(n=n'\) and \(T^{(k)}=T'^{(k)}\) hold for \(1 \le k < n\).
The proof proceeds by induction on \(n\), with
case analysis on the rule \(\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(1)}\).
As in the proof of Lemma~\ref{prop:cong},
we use the fact that if \(T\in \inhav{\NT{\Gamma}{\tau}}\), then
\(\mathbf{FV}(e(T))=\mathop{\mathrm{Dom}}(\Gamma)\).
\begin{itemize}
\item Case for the rule $\NT{\{x_i:\tau\}}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} x_i$: In this case, \(T=x_i\) and \(n=1\).
Since the root of \(T'^{(1)}\) must be \(x_i\), we have \(T'^{(1)} = T\) and \(n'=1=n\).
\item Case for the rule \(\NT{\Gamma}{\sigma \ft \tau'} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma} (\NT{\Gamma}{\tau'}) \):
In this case, \(n \ge 2\),
\(T^{(k)} = \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma}(T^{(k)}_1)\) (\(k \le n\)),
\(T= \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma} (T_1)\), and we have the following leftmost rewriting sequences:
\[
\NT{\Gamma}{\tau'} = T^{(1)}_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(n)}_1 = T_1.
\]
Since the root of \(T'^{(k)}\) is the same as that of \(T\),
\(T'^{(k)} = \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma}(T'^{(k)}_1)\) for \(k \le n'\),
and we have the following leftmost rewriting sequence:
\[
T'^{(1)}_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T'^{(n')}_1 = T_1.
\]
By the definition of rules, \(T'^{(1)}\) must be of the form \(\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\sigma} (\NT{\Gamma}{\tau})\)
and so \(T'^{(1)}_1 = \NT{\Gamma}{\tau}\).
By the induction hypothesis for \(\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(2)}_1\),
we have \(n=n'\) and \(T^{(k)}_1=T'^{(k)}_1\) for \(2 \le k < n\).
Thus we have also \(T^{(k)}=T'^{(k)}\) for \(1 \le k < n\), as required.
\item Case for the rule:
\( \NT{\Gamma}{\sigma \ft \tau'} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_i^{\sigma}
(\NT{\Gamma\cup\set{x_i:\sigma}}{\tau'})\), where \(i = \min
\set{ j \mid x_j\in X_\K\setminus \mathop{\mathrm{Dom}}(\Gamma)}\).
In this case, \(n \ge 2\),
\(T^{(k)} = \lambda x_i^{\sigma}(T^{(k)}_1)\) (\(k \le n\)),
\(T= \lambda x_i^{\sigma} (T_1)\), and we have the following leftmost rewriting sequences:
\[
\NT{\Gamma\cup\set{x_i:\sigma}}{\tau'} = T^{(1)}_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(n)}_1 = T_1.
\]
Since the root of \(T'^{(k)}\) are the same as that of \(T\),
\(T'^{(k)} = \lambda x_i^{\sigma}(T'^{(k)}_1)\) (\(k \le n'\)),
and we have the leftmost rewriting sequence:
\[
T'^{(1)}_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T'^{(n')}_1 = T_1.
\]
By the definition of rules,
\(\NT{\Gamma}{\sigma \ft \tau'} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_i^{\sigma} (T'^{(1)}_1)\)
implies that \(T'^{(1)}_1 = \NT{\Gamma\cup\set{x_i:\sigma}}{\tau'}\).
Hence by the induction hypothesis for \(\NT{\Gamma\cup\set{x_i:\sigma}}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(2)}_1\),
we have \(n=n'\) and \(T^{(k)}_1=T'^{(k)}_1\) for \(2 \le k < n\).
Thus we also have \(T^{(k)}=T'^{(k)}\) for \(1 \le k < n\).
\item
Case for the rule: \(\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} @(\NT{\Gamma_1}{\sigma \ft \tau}, \NT{\Gamma_2}{\sigma})\)
where \(\Gamma = \Gamma_1 \cup \Gamma_2\).
In this case, \(n=n_1+n_2-1\), \(n_1,n_2 \ge 2\), \(T= @(T_1,T_2)\),
\begin{align*}
T^{(k)} &= @(T^{(k)}_1, \NT{\Gamma_2}{\sigma}) && (1 \le k \le n_1)
\\
T^{(n_1-1+k)} &= @(T_1, T^{(k)}_2) && (1 \le k \le n_2),
\end{align*}
and we have the following leftmost rewriting sequences:
\begin{align*}
\NT{\Gamma_1}{\sigma \ft \tau} &= T^{(1)}_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(n_1)}_1 = T_1
\\
\NT{\Gamma_2}{\sigma} &= T^{(1)}_2 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(n_2)}_2 = T_2.
\end{align*}
Since the root of \(T'^{(1)}\) is \(@\), \(T'^{(1)}\) must be of the form \(@(\NT{\Gamma'_1}{\sigma' \ft \tau}, \NT{\Gamma'_2}{\sigma'})\)
with \(\Gamma'_1\cup\Gamma'_2 = \Gamma\).
Also let us consider \(T'^{(k)}\):
we have \(n=n'_1+n'_2-1\), \(n'_1,n'_2 \ge 2\),
\begin{align*}
T'^{(k)} &= @(T'^{(k)}_1, \NT{\Gamma'_2}{\sigma'}) && (1 \le k \le n'_1)
\\
T'^{(n'_1-1+k)} &= @(T_1, T'^{(k)}_2) && (1 \le k \le n'_2),
\end{align*}
and we have the following leftmost rewriting sequences:
\begin{align*}
\NT{\Gamma'_1}{\sigma' \ft \tau} &= T'^{(1)}_1 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T'^{(n'_1)}_1 = T_1
\\
\NT{\Gamma'_2}{\sigma'} &= T'^{(1)}_2 \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \dots \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T'^{(n'_2)}_2 = T_2.
\end{align*}
Since \(\mathop{\mathrm{Dom}}(\Gamma_i)=\mathbf{FV}(e(T_i))=\mathop{\mathrm{Dom}}(\Gamma'_i)\) for each \(i\in\set{1,2}\)
and \(\Gamma_1 \cup \Gamma_2 = \Gamma'_1\cup\Gamma'_2\),
we have \(\Gamma_i=\Gamma'_i\),
which also implies \(\sigma=\sigma'\).
Hence by the induction hypothesis for \(\NT{\Gamma_1}{\sigma \ft \tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(2)}_1\)
and for \(\NT{\Gamma_2}{\sigma} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T^{(2)}_2\),
we have \(n_i=n'_i\) and \(T^{(k)}_i=T'^{(k)}_i\) for each \(i\in\set{1,2}\) and \(2 \le k < n_i\).
Thus we also have \(T^{(k)}=T'^{(k)}\) for \(1 \le k < n\).
\qedhere
\end{itemize}
\end{proof}
\subsection{Strong Connectivity and Aperiodicity}
In this section, we restrict the grammar \(\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) to
$\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ by removing unnecessary nonterminals, and
show the strong connectivity and aperiodicity of $\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$
for $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,\I, \K \geq 2$ (Lemma~\ref{prop:irr} below).
Recall that the strong connectivity and aperiodicity is required to apply Corollary~\ref{cor:imtTree} and
Remark~\ref{rem:avoidLimDef}, respectively.
We define the restricted grammar \(\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) by:
\[
\begin{aligned}
\enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} &\triangleq \set{\N{\z} \in \lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \mid
\text{\(\N{\z}\) is reachable n \(\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\)
from some \(\NT{\emptyset}{\sigma} \in \lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\)}}
\\
\erules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} &\triangleq \set{\N{\z} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} T \in \lrules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \mid
\N{\z} \in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}
\\
\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} &\triangleq
(\lralph{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}, \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}, \erules{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}).
\end{aligned}
\]
For \(\N{\z} \in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\), clearly
\(\inhav{\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},\N{\z}} = \inhav{\lgrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},\N{\z}}\).
Through the bijection \(\pi\),
we can show that, for any \(\NT{\Gamma}{\tau} \in \lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\), \(\NT{\Gamma}{\tau}\) also belongs to \(\enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\)
if and only if there exists a term in \(\terms{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) whose type
derivation contains a type judgment of the form \(\Gamma \p t:
\tau\).
The strong connectivity of \(\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\)
follows from the following facts: (i) each \(\NT{\Gamma}{\tau}\in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) is reachable from
some \(\NT{\emptyset}{\tau'}\in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) (by the definition of $\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ above),
(ii) each $\NT{\emptyset}{\tau}\in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ is
reachable from $\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}$ (Lemma~\ref{lem:sc1} below),
and (iii)
$\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}$ is reachable from every \(\NT{\Gamma}{\tau}\in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\)
(Lemma~\ref{lem:sc2} below).
\begin{lem}
\label{lem:sc1}
Let $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#, \I \geq 2$ and $\K \geq 1$ be integers.
Then for any nonterminal $\NT{\emptyset}{\tau} \in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$,
$\NT{\emptyset}{\tau}$ is reachable from $\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}$.
\end{lem}
\begin{proof}
Let \(\tau = \tau_1 \ft \dots \ft \tau_n \ft \mathtt{o}\) and
\(\tau_i = \tau^i_1 \ft \dots \ft \tau^i_{\ell_i} \ft \mathtt{o}\)
for \(i=1,\dots,n\).
For \(i=1,\dots,n\), let \(T_{\tau_i} \triangleq \lambda *^{\tau^i_1}(\dots\lambda *^{\tau^i_{\ell_i}}(x_1)\dots)\).
Then, we have:
\[
\begin{aligned}
\NT{x_1:\mathtt{o}}{\tau_i}
\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}
\lambda *^{\tau^i_1}(\NT{x_1:\mathtt{o}}{\tau^i_2 \ft \dots \ft \tau^i_{\ell_i} \ft \mathtt{o}})
&\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*
\lambda *^{\tau^i_1}(\dots\lambda *^{\tau^i_{\ell_i}}(\NT{x_1:\mathtt{o}}{\mathtt{o}})\dots)
\\
&\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}
\lambda *^{\tau^i_1}(\dots\lambda *^{\tau^i_{\ell_i}}(x_1)\dots)
= T_{\tau_i}
\end{aligned}
\]
and hence
\[
\begin{aligned}
\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}
&\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}\phantom{{}^*}
\lambda x_1^{\mathtt{o}}(\NT{x_1:\mathtt{o}}{\mathtt{o}})
\\
&\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}\phantom{{}^*}
\lambda x_1^{\mathtt{o}}(
@(\NT{x_1:\mathtt{o}}{\tau_n\ft\mathtt{o}},\NT{x_1:\mathtt{o}}{\tau_n})
)
\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*
\lambda x_1^{\mathtt{o}}(
@(\NT{x_1:\mathtt{o}}{\tau_n\ft\mathtt{o}},T_{\tau_n})
)
\\
&\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*
\lambda x_1^{\mathtt{o}}(
@({@(
\dots
@(
\NT{x_1:\mathtt{o}}{\tau_2\ft\cdots\ft\tau_n\ft\mathtt{o}}
,T_{\tau_2}) ,
\dots)},T_{\tau_n})
)
\\
&\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}\phantom{{}^*}
\lambda x_1^{\mathtt{o}}(
@({@(
\dots
@(
@(\NT{\emptyset}{\tau_1\ft\tau_2\ft\cdots\ft\tau_n\ft\mathtt{o}},\NT{x_1:\mathtt{o}}{\tau_1})
,T_{\tau_2}) ,
\dots)},T_{\tau_n})
)
\\
&\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*
\lambda x_1^{\mathtt{o}}(
@({@(
\dots
@(
@(\NT{\emptyset}{\tau},T_{\tau_1})
,T_{\tau_2}) ,
\dots)},T_{\tau_n})
).
\end{aligned}
\]
\end{proof}
\begin{lem}
\label{lem:sc2}
Let $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#, \I, \K \geq 2$ be integers.
Then for any nonterminal $\NT{\Gamma}{\tau} \in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$,
$\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}$ is reachable from \(\NT{\Gamma}{\tau}\).
\end{lem}
\begin{proof}
Suppose $\NT{\Gamma}{\tau} \in \enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$.
By the definition of \(\enont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) (and \(\lnont{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\)), there exists \(t\) such that
\([t]_\alpha\in \termsg{\Gamma}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\).
Let \(T \triangleq \pi^{-1}([t]_\alpha) \in \inhav{\NT{\Gamma}{\tau}}\).
Now \(T\) contains at least one (possibly bound) variable, say \(x\), and let \(\sigma\) be the type of \(x\).
Since \(T \in \inhav{\NT{\Gamma}{\tau}}\), there exists a \linear{context} \(\ctr\) such that
\[
\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^{*} \ctr[\NT{\set{x:\sigma}}{\sigma}] \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \ctr[x] = T.
\]
We show, by the induction on the structure of \(\sigma\), that \(\NT{\set{x:\sigma}}{\sigma}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*
\ctr'[\NT{\set{y:\mathtt{o}}}{y}]\) for some \(y\) and \linear{context} \(\ctr'\).
The case for \(\sigma=\mathtt{o}\) is obvious. If \(\sigma=\sigma'\to\tau'\), then since \(\K\geq 2\),
we obtain:
\[
\begin{array}{l}
\NT{\set{x:\sigma}}{\sigma}\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_i^{\sigma'}\NT{\set{x:\sigma,x_i:\sigma'}}{\tau'}
\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_i^{\sigma'}(@(\NT{\set{x:\sigma}}{\sigma}, \NT{\set{x_i:\sigma'}}{\sigma'}))\\
\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}} \lambda x_i^{\sigma'}(@(x, \NT{\set{x_i:\sigma'}}{\sigma'}))
\end{array}
\]
by ``\(\eta\)-expansion''. By the induction hypothesis, we have \(
\NT{\set{x_i:\sigma'}}{\sigma'} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^* \ctr''[\NT{\set{y:\mathtt{o}}}{\mathtt{o}}]\) for some \(y\) and 1-context \(\ctr''\).
Thus, the result holds for \(\ctr'=\lambda x_i^{\sigma'}(@(x,\ctr''))\).
By using the property above, we obtain
\[
\begin{array}{l}
\NT{\Gamma}{\tau} \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^{*} \ctr[\NT{\set{x:\sigma}}{\sigma}] \longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}^*
\ctr[\ctr'[\NT{\set{y:\mathtt{o}}}{\mathtt{o}}]] \\\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}
\ctr[\ctr'[@(\NT{\emptyset}{\mathtt{o}\to\mathtt{o}},\NT{\set{y:\mathtt{o}}}{\mathtt{o}})]]
\longrightarrow} \newcommand{\gf}[1]{\mathbf{F}_{#1}} \newcommand{\thole}[2]{h_{\typing{#1}{#2}}
\ctr[\ctr'[@(\NT{\emptyset}{\mathtt{o}\to\mathtt{o}},y)]].
\end{array}
\]
Thus, \(\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}\) is reachable from \(\NT{\Gamma}{\tau}\).
\end{proof}
\begin{lem}\label{lem:aperiodicity}
Let $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#, \I \geq 2$ and $\K \geq 1$ be integers.
Then for any integer $n \geq 5$, the nonterminal $\NT{\emptyset}{\mathtt{o} \ft \mathtt{o}}$
of $\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}\K$
satisfies \(\inhavn{\NT{\emptyset}{\mathtt{o} \ft \mathtt{o}}} \neq \emptyset\).
\end{lem}
\begin{proof}
For simplicity of the presentation, here we identify trees with terms by
the size-preserving bijection \(\pi\).
The proof proceeds by induction on \(n\).
For \(n=5,6,7\), the following terms
\[
\begin{aligned}&
\lambda x^{\mathtt{o}}\!. (\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\mathtt{o}}\!.x) x
\\&
\lambda x^{\mathtt{o}}\!. (\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\mathtt{o}\ft\mathtt{o}}\!.x) (\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\mathtt{o}}\!.x)
\\&
\lambda x^{\mathtt{o}}\!. ( \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\mathtt{o}\ft\mathtt{o}\ft\mathtt{o}}\!.x) \lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\mathtt{o}}\!.\lambda {\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^{\mathtt{o}}\!.x
\end{aligned}
\]
belong to \(\inhavn{\NT{\emptyset}{\mathtt{o} \ft \mathtt{o}}}\), respectively.
For $n \geq 8$, by the induction hypothesis, there exists
$t \in \inhavnn{n-3}{\NT{\emptyset}{\mathtt{o} \ft \mathtt{o}}}$. Thus we have
$\lambda x. tx \in \inhavn{\NT{\emptyset}{\mathtt{o} \ft \mathtt{o}}}$ as required.
\end{proof}
The following is the main result of this subsection.
\begin{lem}\label{prop:irr}
$\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}$ is strongly connected and aperiodic for any $\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#,\I, \K \geq 2$.
\end{lem}
\begin{proof}
Strong connectivity follows from Lemmas~\ref{lem:sc1} and \ref{lem:sc2},
and the definition of \(\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) (as stated at the beginning of this subsection).
The aperiodicity follows from Lemma~\ref{lem:aperiodicity} and strong connectivity.
(Note that for every nonterminal \(N\), there exists a linear context \(U\in \inhav{{\mathcal G},
\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}\Ar N}\). Thus, for any \(n\ge |U|+5\),
\(\inhavn{{\mathcal G},N} \supseteq \set{U[T]\mid T\in\inhavnn{n-|U|}{{\mathcal G},
\NT{\emptyset}{\mathtt{o}\ft\mathtt{o}}}}\neq \emptyset\).)
\end{proof}
\subsection{Explosive Terms}
\label{sec:explosive}
In this section, we define a family \((\expl{n}{k})_n\)
of \(\lambda\)-terms, which have long \(\beta\)-reduction sequences.
They play the role of \((T_n)_n\) in
Corollary~\ref{cor:imtTree}.
We define a ``duplicating term'' $\dt \triangleq \lambda x^\mathtt{o}. (\lambda x^\mathtt{o}. \lambda
{\ast}} \newcommand{\V}{\mathbf{V}} \newcommand{\hole}{[\,]^\mathtt{o}. x)x\,x$, and \(\idt \triangleq \lambda x^\mathtt{o}. x\).
For two terms $t$, $t'$ and integer $n \geq 1$, we define the ``$n$-fold
application'' operation $\fold{}{n}{}$ by $\fold{t}{0}{t'} \triangleq t'$ and $\fold{t}{n}{t'} \triangleq t (\fold{t}{n-1}{t'})$.
For an integer $k \geq 2$, we define an order-\(k\) term
\[
\ctwok{k} \triangleq \lambda f^{\Ttwice{k-1}}. \lambda x^{\Ttwice{k-2}}.
f (f x)
\]
where $\Ttwice{i}$ is defined by $\Ttwice{0} \triangleq \mathtt{o}$ and $\Ttwice{i+1}
\triangleq \Ttwice{i} \ft \Ttwice{i}$.
\begin{defi}[explosive terms]
Let $m \geq 1$ and $k \geq 2$ be integers.
We define the \emph{explosive term $\expl{m}{k}$} by:
\[
\expl{m}{k} \triangleq
\lambda x^\mathtt{o}.\bigl( (\fold{\ctwok{k}}{m}{\ctwok{k-1}})\ctwok{k-2} \cdots
\ctwok{2} \, \dt (\idt\, x) \bigr).
\]
\end{defi}
We state key properties of $\expl{m}{k}$ below.
\begin{lem}[explosive]
\label{lemma:expl}
\begin{enumerate}
\item\label{item:expltyped} $\emptyset\p\expl{m}{k} : \mathtt{o}\ft\mathtt{o}$ is derivable.
\item\label{item:expldix}
$\mathtt{ord}(\expl{m}{k}) = k$,
$\mathtt{iar}(\expl{m}{k}) = k$ and
$\card{\smash{\V(\expl{m}{k})}}=2$.
\item\label{item:explsize} $|\expl{m}{k}| = 8m + 8k - 2$.
\item\label{item:expluctx} \([\expl{m}{k}]_\alpha \in \termsg{\emptyset}{\mathtt{o}\ft\mathtt{o}}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\) if \(\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#\), \(\I \ge k\) and
\(\K \ge 2\).
\item\label{item:explbeta}
If a term $t$ satisfies $\expl{m}{k} \preceq t$, then $\beta(t) \geq \Exp{k}{m}$ holds.
\end{enumerate}
\end{lem}
\begin{proof}
First, observe that by the definition of $\ctwok{k}$, we have:
\[
|\ctwok{k}| = 7, \quad \p \ctwok{k}: \Ttwice{k}, \quad \mathtt{ord}(\ctwok{k}) =
k, \quad \mathtt{iar}(\ctwok{k}) = k.
\]
(\ref{item:expltyped}) is obvious, and in the derivation of
$\emptyset\p\expl{m}{k} : \mathtt{o}\ft\mathtt{o}$,
\(\ctwok{k}\) is
a subterm that has a type of the largest order and internal arity. Thus,
(\ref{item:expldix}) follows also immediately from
\(\mathtt{ord}(\ctwok{k}) =\mathtt{iar}(\ctwok{k}) = k\), and
\(\smash{\V(\expl{m}{k})}=\set{f,x}\).
(\ref{item:explsize}) follows from the following calculation.
\[
\begin{array}{ll}
|\expl{m}{k}|
&= 1+
|\fold{\ctwok{k}}{m}{\ctwok{k-1}}| + \Sigma_{i=2}^{k-2}(1+|\ctwok{i}|)+(1+|\dt|)+(1+|\idt\,x|)\\
& \hfill\mbox{(by $ |t_0t_1\cdots t_n| = |t_0|+\Sigma_{i=1}^n (1+|t_i|)$)}\\
&= 1+
((|\ctwok{k}|+1)m+|\ctwok{k-1}|) + \Sigma_{i=2}^{k-2}(1+|\ctwok{i}|)+(1+|\dt|)+(1+|\idt\,x|)\\
&= 1+ (8m+7) + 8(k-3)+9+5 =8m + 8k -2.
\end{array}
\]
(\ref{item:expluctx}) follows from (\ref{item:expltyped}) and (\ref{item:expldix}).
The proof for (\ref{item:explbeta}) is the same as the corresponding proof in~\cite{Beckmann01}.
\end{proof}
\subsection{Proof of the Main Theorem}
We are now ready to prove Theorem~\ref{mainthm}.
\begin{proof}[Proof of Theorem~\ref{mainthm}]
We apply Corollary~\ref{cor:imtTree} with Remark~\ref{rem:avoidLimDef} to
the grammar \(\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}\),
which is unambiguous by Lemma~\ref{lem:unamb},
and aperiodic and strongly connected by Lemma~\ref{prop:irr}.
We define \(T_n \triangleq \pi^{-1}([\expl{n}{k}]_\alpha)\);
then by Lemma~\ref{lemma:expl}(\ref{item:explsize}),
we have \(|T_n| = 8n + 8k - 2 = \mathrm{O}(n)\).
Thus, by Corollary~\ref{cor:imtTree} (with Remark~\ref{rem:avoidLimDef},
which allows us to replace \(\limd_{n \rightarrow \infty}\) with
\(\lim_{n \rightarrow \infty}\)),
there exists \(\cn>0\) such that,
for \(\nset \triangleq \set{ \NT{\emptyset}{\tau} \mid \tau \in \types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I} }\),
\begin{equation}
\label{eq:mainCor}
\lim_{n \rightarrow \infty} \frac{\card{\{ (N,T) \in \coprod_{N \in \nset}\inhavn{\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},N} \mid
T_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\coprod_{N \in \nset}\inhavn{\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},N}}} = 1.
\end{equation}
For \(n \in \mathbb{N}\), we have
\begin{equation}
\label{eq:MAINsubset}
\{ [t]_\alpha \in \termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \mid \expl{\ceil{\cn \log n}}{k} \preceq t \}
\subseteq
\{ [t]_\alpha \in \termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K} \mid \beta(t) \geq \Exp{k-1}{n^{\cn}} \}
\end{equation}
by Lemma~\ref{lemma:expl}(\ref{item:explbeta}), and
\begin{equation}
\label{eq:MAINtermGram}
\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}
\ =\
\, \biguplus_{\mathclap{\tau \in
\types{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}}}\, \termsng{\emptyset}{\tau}{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}
\ \cong\
\, \coprod_{\mathclap{N \in \nset}}\,
\inhavnn{n}{\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},N}
\end{equation}
by Lemma~\ref{prop:cong}.
Therefore, we have:
\begin{align*}
1 \ge \ &
\frac{\card{\{ [t]_\alpha \in
\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}
\mid \beta(t) \geq \Exp{k-1}{n^{\cn}}
\}}}{\card{\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}}
\\ \ge \ &
\frac{\card{\{ [t]_\alpha \in
\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}
\mid \expl{\ceil{\cn \log n}}{k} \preceq t
\}}}{\card{\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}}
&& \reason{by~\eqref{eq:MAINsubset}}
\\ =\ &
\frac{\card{\{ (N, T) \in
\coprod_{N \in \nset}\, \inhavnn{n}{\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},N}
\mid
T_{\ceil{\cn \log n}} \preceq T
\}}}{\card{\coprod_{N \in \nset}\, \inhavnn{n}{\egrm{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K},N}}}
&& \reason{by~\eqref{eq:MAINtermGram}}
\end{align*}
Since the right hand side converges to \(1\) by~\eqref{eq:mainCor}, we have
\[ \lim_{n \rightarrow \infty}
\frac{\card{\{ [t]_\alpha \in
\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}
\mid \beta(t) \geq \Exp{k-1}{n^{\cn}}
\}}}{\card{\termsn{\delta} \newcommand{\I}{\iota} \newcommand{\K}{\xi} \newcommand{\cards}{\#}{\I}{\K}}}
=1\]
as required.
\end{proof}
\section{Related Work}\label{sec:relatedwork}
As mentioned in Section~\ref{sec:introduction},
there are several pieces of work on probabilistic properties of
untyped $\lambda$-terms~\cite{SN,grygiel2013counting,Bendkowski}.
David et al.~\cite{SN} have shown that almost all untyped
$\lambda$-terms are strongly normalizing, whereas the result is opposite
for terms expressed in SK combinators (the latter result has later been generalized for arbitrary Turing complete combinator bases~\cite{Bendkowski17}).
Their former result implies that untyped $\lambda$-terms
do \emph{not} satisfy the infinite monkey theorem, i.e.,\; for any term
$t$, the probability that a randomly chosen term of size $n$ contains
$t$ as a subterm tends to \emph{zero}.
Bendkowski et al.~\cite{Bendkowski} proved that almost all terms in de
Brujin representation are not strongly normalizing, by regarding the
size of an index $i$ is $i+1$, instead of the constant $1$.
The discrepancies among those results suggest that this kind of probabilistic property
is quite fragile and depends on the definition of the syntax and the size of terms.
Thus, the setting of our paper, especially the assumption on
the boundedness of internal arities and the number of variables is a matter of debate,
and it would be interesting to study how the result changes for different assumptions.
We are not aware of similar studies on \emph{typed} $\lambda$-terms.
In fact, in their paper
about combinatorial aspects of $\lambda$-terms,
Grygiel and Lescanne~\cite{grygiel2013counting}
pointed out that
the combinatorial study of typed $\lambda$-terms is difficult, due to
the lack of (simple) recursive definition of typed terms.
In the present paper, we have avoided the difficulty
by making the assumption on the boundedness of internal arities and the number
of variables (which is, as mentioned above, subject to a debate though).
Choppy et al.~\cite{Choppy89} proposed a method to evaluate
the average number of reduction steps for
a restricted class of term rewriting systems called \emph{regular} rewriting systems.
In our context, the average number of reduction steps is not of much interest;
note that, as the worst-case number of reduction steps
is \(k\)-fold exponential for order-\(k\) terms,
the average is also \(k\)-fold exponential, even if
it were the case that the number of reduction steps is small for almost all terms.
In a larger context, our work may be viewed as an instance of the studies of
average-case complexity~\cite[Chapter 10]{CCbook}, which discusses ``typical-case
feasibility''. We are not aware of much work on the average-case complexity of
problems with hyper-exponential complexity.
As a result related to our parameterized infinite monkey theorem for trees
(Theorem~\ref{thm:imt}),
Steyaert and Flajolet~\cite{STEYAERT198319} studied the probability that a given pattern
(which is, in our terminology, a tree context of which every leaf is a hole)
occurs at a randomly chosen node of a randomly chosen tree. Their result immediately yields
a \emph{non-parameterized} infinite monkey theorem for trees (which says that
the probability that a given pattern occurs in a sufficiently large tree tends to \(1\)),
but their technique does not seem directly applicable to obtain our \emph{parameterized} version.
\section{Conclusion}\label{sec:conclusion}
We have shown that almost every simply-typed $\lambda$-term of order $k$
has a $\beta$-reduction sequence as long as $(k-1)$-fold exponential in the term size,
under a certain assumption. To our knowledge, this is the first result of this kind for
{typed} $\lambda$-terms. We obtained the result through the parameterized infinite monkey theorem
for regular tree grammars, which may be of independent interest.
A lot of questions are left for future work, such as (i) whether
our assumption (on the boundedness of arities and the number of
variables) is reasonable, and how the result changes for different assumptions, (ii)
whether our result is optimal (e.g., whether almost every term has a $k$-fold exponentially long reduction sequence), and (iii) whether similar results hold for Terui's decision
problems~\cite{Terui12RTA} and/or the higher-order model checking problem~\cite{Ong06LICS}.
To resolve the question (ii) above, it may be useful to conduct
experiments to count the number of reduction steps for randomly generated terms.
\subsubsection*{Acknowledgment}
We would like to thank anonymous referees for useful comments.
This work was supported by JSPS KAKENHI Grant Number JP15H05706.
\bibliographystyle{../splncs.bst}
|
\section{Introduction}
Assume a time-harmonic incoming wave $u^{in}$ is incident onto a bounded penetrable
scatterer $D\subset {\mathbb R}^2$ embedded in a homogeneous isotropic background medium.
We assume that the boundary $\partial D$ is Lipschitz continuous and piecewise analytic, and that the complement $D^e:={\mathbb R}^2\backslash\overline{D}$ of $D$ is connected.
The wave propagation of the total field $u=u^{in}+u^{sc}$ is then modeled by the Helmholtz equation
\begin{equation}\label{eq:Helm}
\Delta u+k^2 q\, u=0\quad\mbox{in}\quad {\mathbb R}^2.
\end{equation}
In this paper the refractive index (potential) function $q$ is supposed to be a piecewise constant function, given by
\begin{eqnarray*}
q(x)=\left\{\begin{array}{lll}
1,\quad\mbox{if}\quad x\in D^e,\\
q_0\neq 1,\quad\mbox{if}\quad x\in \overline{D}.
\end{array}\right.
\end{eqnarray*}
Across the interface $\partial D$, we assume the continuity of the total field and its normal derivative, i.e.,
\begin{eqnarray}\label{TE}
u^+=u^-,\quad \partial_\nu u^+=\partial_\nu u^-\quad\mbox{on}\;\partial D.
\end{eqnarray}
Here the superscripts $(\cdot)^\pm$ stand for the limits taken from outside and inside, respectively, and $\nu\in {\mathbb S}:=\{x\in{\mathbb R}^2: |x|=1\}$ is the unit normal on $\partial D$ pointing into $D^e$.
At the infinity, the perturbed scattered field
$u^{sc}$ is supposed to fulfill the Sommerfeld radiation condition
\begin{equation}\label{eq:radiation}
\lim_{|x|\rightarrow \infty} |x|\,\left\{ \frac{\partial u^{sc}}{\partial |x|}-ik u^{sc} \right\}=0.
\end{equation}
The unique solvability of the scattering problem \eqref{eq:Helm}, \eqref{eq:radiation} and (\ref{TE}) in $H^2_{loc}({\mathbb R}^2)$ is well known (see e.g., \cite[Chapter 8]{CK}). In particular, the Sommerfeld radiation condition (\ref{eq:radiation}) leads to the asymptotic expansion
\begin{equation}\label{eq:farfield}
u^{sc}(x)=\frac{e^{ik |x|}}{\sqrt{|x|}}\; u^\infty(\hat x)+\mathcal{O}\left(\frac{1}{|x|^{3/2}}\right),\quad |x|\rightarrow+\infty,
\end{equation}
uniformly in all directions $\hat x:=x/|x|$, $x\in\mathbb{R}$. The function $u^\infty(\hat x)$ is an analytic function defined on ${\mathbb S}^{2}$ and is referred to as the \emph{far-field pattern} or the \emph{scattering amplitude}. The vector $\hat{x}\in{\mathbb S}$ is called the observation direction of the far field. The classical inverse scattering problem consists of the recovery of the boundary $\partial D$ from the far-field patterns corresponding to one or several incident plane waves. In this paper we are
concerned with the following questions:
\begin{description}
\item[(i)] Does the obstacle $D$ scatter any incident wave trivially (that is, $u^{sc}\equiv 0$) ?
\item[(ii)] Does the far-field pattern of a single incoming wave uniquely determine $\partial D$ ?
\end{description}
A negative answer to the first question means that acoustic cloaking cannot be achieved using isotropic materials, while a positive answer implies that $k^2$ is a non-scattering wavenumber (energy).
The study of non-scattering energies dates back to \cite{KS} in the case of a convex (planar) corner domain, where notion of \emph{scattering support} for an inhomogeneous medium was explored. In one of the authors' previous work \cite{EH2017}, it was shown that variable potential functions with the following corners on $\partial D$:
\begin{itemize}
\item curvilinear polygonal corners in ${\mathbb R}^2$;
\item curvilinear polyhedral corners in ${\mathbb R}^3$;
\item circular conic corners in ${\mathbb R}^3$;
\end{itemize}
scatters every incident wave non-trivially. Earlier publications were devoted to the
absence of non-scattering energies under more restrictive assumptions on the smoothness of the potential or the angle of the corner. Here we mention the following works in the acoustic case:
\begin{itemize}
\item $C^\infty$-potentials with rectangular corners in ${\mathbb R}^n$ ($n\geq 2$) \cite{BLS};
\item H\"older continuous potentials with convex corners in ${\mathbb R}^2$, and with circular conic corners in ${\mathbb R}^3$ whose opening angle is outside of a countable subset of $(0,\pi)$ \cite{PSV};
\item analytical potentials with arbitrary polygonal corners or polyhedral wedge corners \cite{ElHu2015};
\item H\"older continuous potentials with rectangular corners in ${\mathbb R}^3$ \cite{HSV}.
\end{itemize}
The argument of the pioneering work \cite{BLS} was based on the use of complex geometric optics (CGO) solutions, which
was later extended to \cite{PSV} and \cite{HSV} for treating less regular potentials and convex corners.
The approach of \cite{ElHu2015} relies on the expansion of solutions to the Helmholtz equation with real-analytic potentials. For general potentials and corners, the absence of non-scattering energies can be verified via singularity analysis of the inhomogeneous Laplace equation in a cone \cite{EH2017}. We remark that the first question is closely related to the second one, that is, the approach for proving absence of non-scattering wavenumbers implies uniqueness to shape identification problems in inverse medium scattering.
It was first proved in \cite{ElHu2015} that the shape of a convex penetrable obstacle of polygonal or polyhedral type with an unknown analytical potential can be uniquely determined by a single far-field pattern. The CGO-solution methods of \cite{PSV,BLS} also lead to uniqueness in shape identification but are confined so far to convex polygons in ${\mathbb R}^2$ and rectangular boxes in ${\mathbb R}^3$ with H\"older continuous potentials (see \cite{HSV}).
In \cite[Corollay 2.1]{EH2017}, the uniqueness result of \cite{ElHu2015} was extended to
more general potential functions using the data of a single far-field pattern.
\section{Main results}
The main purpose of this paper is to exclude (positive) real non-scattering energies when $\partial D$ contains a weakly singular corner, around which the boundary is allowed to be $C^1$-smooth but piecewise analytic.
The corners mentioned in the previous section are all strongly singular in the following sense.
\begin{definition}\label{def:SS}
A point $O\in \partial D\subset {\mathbb R}^2$ is called strongly singular if the boundary around $O$ can be locally parameterized by a continuous and piecewise analytic function whose derivative is discontinuous at $O$.
\end{definition}
Evidently, every planar corner point of a polygon with flat slides is strongly singular, because the boundary can be locally parameterized by a piecewise linear function, whose
first derivative is piecewise constant. A curvilinear corner of $D$ (see e.g., Definition 2.1 of \cite{EH2017} for a precise description) is also strongly singular by Definition \ref{def:SS}. Below we state the definition of weakly singular corners to be explored within the scope of this paper.
\begin{definition}\label{def:WS}
A point $O=(0,0)\in \partial D$ is called weakly singular if the subboundary $B_\epsilon(O)\cap \partial D$ for some $\epsilon>0 $ can be parameterized by the polynomial function $x_2=f(x_1)$, $x_1\in(-1,1)$, where
\begin{eqnarray}\label{f1}
f(x_1)=\left\{\begin{array}{lll}
c_1\, x_1^{\alpha_1}, &&1>x_1\geq 0,\\
c_2\, x_1^{\alpha_2},&& -1<x_1\leq 0,
\end{array}\right.
\end{eqnarray} and the coefficients $c_j\in {\mathbb R}$ and $\alpha_j\in {\mathbb N}^+$ are assumed to fulfill the relations
\begin{eqnarray*}
(c_1, \alpha_1)\neq (c_2, \alpha_2),\;c_1^2+c_2^2\ne0,\;\;\alpha_j\geq 2.
\end{eqnarray*}
The order of the singularity at $O$ is defined as
\begin{eqnarray*}
\beta:=\left\{\begin{array}{lll}
\min\{\alpha_1, \alpha_2\}&&\mbox{if}\quad c_1\neq 0, c_2\neq 0,\\
\alpha_1 &&\mbox{if}\quad c_2= 0,\\
\alpha_2 && \mbox{if}\quad c_1=0.
\end{array}\right.
\end{eqnarray*}
\end{definition}
The boundary around a weakly singular corner of order $\beta$ is $C^{\beta-1}$-smooth but piecewise $C^{\beta}$-smooth, that is, the $\beta$-th derivative is discontinuous at O.
A singular point of order one must be strongly singular in the sense of Definition \ref{def:SS}.
\begin{definition}\label{def:S}
A point $O\in \partial D$ is called singular if it is either strongly singular in the sense of Definition \ref{def:SS} or weakly singular in the sense of Definition \ref{def:WS}.
\end{definition}
The singular points defined by Definition \ref{def:S} form only a subset of non-analytic points of the boundary. In fact, the polynomial functions described in (\ref{f1}) can be regarded as the leading terms of the Taylor expansion of an analytic function at $x_1=0^\pm$.
If $q$ is a piecewise constant function in ${\mathbb R}^2$, we shall prove that
\begin{theorem}\label{TH}
The obstacle $D\subset {\mathbb R}^2$ scatters every incoming wave, if $\partial D$ contains a singular point.
\end{theorem}
Note that when $\partial D$ possesses a strongly singular corner, Theorem \ref{TH} has been implicitly contained in \cite{EH2017}. The main contribution of this paper is to verify
Theorem \ref{TH} for weakly singular corners in ${\mathbb R}^2$. The above theorem implies that a Lipschitz domain with a singular point on the boundary scatters every incoming wave trivially in two dimensions. Theorem \ref{TH} follows straightforwardly from Lemma \ref{lem:SS} for strongly singular corners and Lemma \ref{lem:WS} for weakly singular corners.
Only local properties of the Helmholtz equation are involved in the proof of Theorem \ref{TH}.
Consequently, we get a local uniqueness result to the inverse scattering for shape identification:
\begin{theorem}\label{TH2} Let $D_j$ ($j=1,2$) be two penetrable obstacles in ${\mathbb R}^2$ with the piecewise constant potential functions $q_j$, respectively. If $\partial D_2$ differs from $\partial D_1$ in the presence of a singular point lying on the boundary of the unbounded component of ${\mathbb R}^2\backslash\overline{(D_1\cup D_2)}$, then the far-field patterns corresponding to $D_j$ and $q_j$ incited by any incoming wave cannot coincide.
\end{theorem}
Theorem \ref{TH2} can be used to distinguish two penetrable scatterers with a piecewise constant potential.
Equivalently, Theorem \ref{TH2} can be reformulated as follows:
\begin{corollary}\label{Coro1}
Let $D_j\subset {\mathbb R}^2$ ($j=1,2$) be two penetrable obstacles in ${\mathbb R}^2$ with the piecewise constant potential functions $q_j$, respectively. Assume that $\partial D_j$ are piecewise analytic and all non-analytical points of the boundary are singular corners defined by Definition
\ref{def:S}. If the far-field patterns corresponding to $(D_j, q_j)$ incited by a single incoming wave are identical, then the boundary of the unbounded component of ${\mathbb R}^2\backslash\overline{(D_1\cup D_2)}$ must be analytic.
\end{corollary}
Theorem \ref{TH2} and Corollary \ref{Coro1} can be verified in the same manner as the proof of Theorem \ref{TH}. We omit the proofs for simplicity. Note that the above shape identification problem is a formally-determined inverse issue. If the far-field data is available for all incident directions but at fixed energy, uniqueness was verified based on the idea of Isakov; see \cite{Isakov90, Kirsch93}. We also refer to \cite{B2008,IY2012} and \cite[Chapter 10]{CK} for unique determination of potential functions from the data of infinitely many plane waves or the Dirichlet-to-Neumann map.
\section{Strongly singular corners always scatter}
This section is devoted to the proof of Theorem \ref{TH} when $\partial D$ contains a strongly singular corner. In the case that $D$ is a polygon, the proof simply follows from \cite{ElHu2015} where variable analytical potential functions were treated.
If $\partial D$ contains a curvilinear corner, the proof was given in \cite{EH2017}. We shall present a proof valid for all strongly singular corners in 2D, under the assumption that $q$ is a piecewise constant potential.
Without loss of generality, we suppose that $O=(0,0)$ is a strongly singular point lying on $\partial D$. Assuming that $u^{sc}$ vanishes in $D^e$, we shall derive a contradiction. Suppose that the boundary $\partial D$ in a neighborhood of $O$ can be expressed as $\Gamma=\{(x_1,f(x_1)): x_1\in(-1,1)\}$, where
$f(x_1)\in C([-1,1])$ is piecewise analytic in $(-1,0]\cup [0,1)$, $f(0)=0$ and $f'(x_1)$ is discontinuous at $x_1=0$.
Since $u^{sc}=0$ in $D^e$, the Cauchy data of $u$ on $\Gamma$ coincide with those of $u^{in}$, which are analytic. Observing that $q$ is a constant on $\overline{D}$ and $\Gamma$ is piecewise analytic, by Cauchy-Kovalevskaya theorem, one may extend $u$ analytically from
$\overline{D}\cap B_1$ to a small neighborhood of $O$ in
the exterior domain
$D^e\cap B_1$. For notational convenience, we suppose the extended domain contains $B_1$. Further, the extended function, which we still denote by $u$, satisfies the Helmholtz equation
\begin{eqnarray*}
\Delta u+k^2 q_0 u=0\quad\mbox{in}\quad B_1.
\end{eqnarray*}
Hence, we deduce the transmission problem for the Helmholtz equations
\begin{eqnarray}\label{transmission}\left\{\begin{array}{lll}
\Delta u_j+q_ju_j=0, \quad j=1,2,&\mbox{ in}\quad B_1,\\
u_1=u_2,\quad\frac{\partial u_1}{\partial \nu}= \frac{\partial u_2}{\partial \nu}&\mbox{ on}\quad \Gamma,
\end{array}\right.\end{eqnarray}
where
\begin{eqnarray*}
u_1=u^{in},\quad u_2=u,\qquad q_1=k^2,\quad q_2=k^2 q_0.
\end{eqnarray*}
To prove Theorem \ref{TH} for strongly singular corners, it is essential to prove that
\begin{lemma} \label{lem:SS}
Suppose that $u_j\in H^2(B_1)$ ($j=1,2$) are solutions to (\ref{transmission}). If $q_1\neq q_2 $, then $u_1=u_2\equiv0$ in $B_1$.
\end{lemma}
By Lemma \ref{lem:SS} and the unique continuation, $u^{in}$ vanishes identically in ${\mathbb R}^2$ which is impossible.
Hence, a piecewise constant potential with a strongly singular point on the boundary of the support always scatter.
The proof of Lemma \ref{lem:SS} follows from an adaption of the arguments in the proof of \cite[Proposition A. 3]{EH2017} to an in homogeneous Helmholtz equation with vanishing Cauchy data.
To make this paper self-contained, we present the proof as follows.
\begin{proof} Setting $u:=u_1-u_2$. Then $u$ is a solution to an inhomogeneous Helmholtz equation with vanishing Cauchy data on $\Gamma$:
\begin{eqnarray}\label{eq:26}\left\{\begin{array}{lll}
\Delta u+q_1u=(q_1-q_2)u_2 &\mbox{in}\quad B_1,\\
u=\frac{\partial u}{\partial \nu}=0&\mbox{on}\quad \Gamma,\\
\Delta u_2+q_2u_2=0 &\mbox{in}\quad B_1.
\end{array}\right.\end{eqnarray}
Note that $u$ and $u_2$ are both real-analytic functions in $B_1$.
Denote by $\tilde{\tau}_j(x_1)$ and $\tilde{\nu}_j(x_1)$ ($j=1,2$) the unit tangential and normal vectors on the curves
\[\Gamma_1:=\{(x_1,f(x_1)): x_1\in[0,1)\},\quad\Gamma_2:=\{(x_1,f(x_1): x_1\in(0,-1]\},
\]
which intersect at the corner $O$. Since $f'(x_1)$ is discontinuous at $x_1=0$, the tangential and normal vectors at the corner point, which we denote by
$\tau_j:=\tilde{\tau}_j(0)$ and $\mu_j=\tilde{\nu}_j(0)$,
are linearly independent. Without loss of generality we suppose that $\nu_1=a_1\tau_1+a_2\tau_2$ with $a_1,a_2\in{\mathbb R}$, $a_2\neq 0$. Hence,
\begin{eqnarray}\label{Derivative}
\partial_{\tau_2}=\frac{1}{a_2}\partial_{\nu_1}-\frac{a_1}{a_2}\partial_{\tau_1}.
\end{eqnarray}
We shall prove by induction that $\nabla^m u(O)=0$ for all $m\in{\mathbb N}_0$, which implies the lemma.
From the Dirichlet and Neumann boundary conditions of $u$ on $\Gamma=\Gamma_1\cup \Gamma_2$ we see that
\begin{eqnarray}\label{D-12}
u=\nabla u=0,\quad
\partial_{\tau_1}^2u=\partial_{\tau_2}^2u=\partial_{\nu_1}\partial_{\tau_1}u=0\quad\mbox{at the corner}\; O.
\end{eqnarray}
Combining (\ref{Derivative}) and (\ref{D-12}) gives the relation
$\partial_{\tau_1}\partial_{\tau_2}u=0$ at $O$.
Since each entry of the vector $\nabla^2$ can be expanded as a linear combination of $\partial_{\tau_1}^2$, $\partial^2_{\tau_2}$ and $\partial_{\tau_1}\partial_{\tau_2}$, we obtain
\begin{eqnarray}\label{eq:27}
\nabla^2u=0\quad \mbox{at}\quad O.
\end{eqnarray}
Consequently, it follows from the equations in (\ref{eq:26}) that
\begin{eqnarray}\label{eq:28}
u_2=\Delta u_2=0\quad \mbox{at}\quad O,
\end{eqnarray}
where we have used the assumption that $q_1\neq q_2$.
To prove that $\nabla^3u(O)=0$, we observe that
\begin{eqnarray*}
\partial_{\tau_1}^3u=\partial_{\tau_2}^3u=
\partial_{\tau_1}^2\partial_{\nu_1}u=
\partial_{\tau_2}^2\partial_{\nu_2}u=0 \quad\mbox{at}\; O.
\end{eqnarray*}
Applying $\partial_{\tau_1}^2$ to both sides of (\ref{D-12}) yields $\partial_{\tau_1}^2\partial_{\tau_2}u(O)=0$. Analogously we can get $\partial_{\tau_2}^2\partial_{\tau_1}u(O)=0$.
Hence, the relation $\nabla^3u(O)=0$ follows from the fact that the differential operators $\partial_{\tau_1}^3, \partial_{\tau_1}^2\partial_{\tau_2}, \partial_{\tau_1}\partial_{\tau_2}^2$ and $\partial_{\tau_2}^3$ span the vector $\nabla^3$. Taking $\nabla$ on the equations in (\ref{eq:26}) gives
\begin{eqnarray}\label{eq:29}
\nabla u_2=\nabla\Delta u_2=0\quad \mbox{at}\quad O.
\end{eqnarray}
Now we want to verify that $\nabla^4u(O)=0$. Arguing as in the previous step we get
\begin{eqnarray}\label{eq1:8}
\partial_{\tau_1}^4u=\partial_{\tau_1}^3\partial_{\tau_2}u=\partial_{\tau_1}\partial_{\tau_2}^3u=
\partial_{\tau_2}^4u=0\quad\mbox{at}\; O.
\end{eqnarray} Hence it suffices to prove $\partial_{\tau_1}^2\partial_{\tau_2}^2u(O)=0$.
Taking $\Delta$ on the first equation in (\ref{eq:26}) and using (\ref{eq:27})-(\ref{eq:28}), we find
\begin{eqnarray*}
\Delta^2 u(O)=(q_1-q_2)\Delta u_2(O)-q_1 \Delta u(O)=0.
\end{eqnarray*}
On the other hand, using (\ref{eq1:8}) and $\partial_{\nu_1}=a_1\partial_{\tau_1}+a_2\partial_{\tau_2}$, we deduce that
\begin{eqnarray*}
\Delta^2u(O)=[\partial_{\nu_1}^2 +\partial_{\tau_1}^2]^2u(O)=[2(1+a_1^2)a_2^2+4a_1^2a_2^2]\,\partial_{\tau_1}^2\partial_{\tau_2}^2u(O),
\end{eqnarray*} from which the relation $\partial_{\tau_1}^2\partial_{\tau_2}^2u(O)=0$ follows. This proves $\nabla^4u(O)=0$.
Now, differentiating the equations in (\ref{eq:26}) yields
\begin{eqnarray*}
\nabla^2 u_2=\nabla^2\Delta u_2=0\quad \nabla\Delta^2 u=0\quad \mbox{at}\quad O.
\end{eqnarray*}
For $m>4$, we make the induction hypothesis that
\begin{eqnarray}\label{induction-hypothesis}
\nabla^j u(O)=\nabla^{j-3}\Delta^2 u(O)=0,\quad \nabla^{j-2}u_2(O)=\nabla^{j-2}\Delta u_2(O)=0\quad\mbox{for all}\;j=0,1,\cdots,m.
\end{eqnarray} We then need to verify that the above relations hold for $j=m+1$, that is,
\begin{eqnarray*}
\nabla^{m+1} u(O)=\nabla^{m-2}\Delta^2 u(O)=0,\quad \nabla^{m-1}u_2(O)=\nabla^{m-1}\Delta u_2(O)=0.
\end{eqnarray*}
We first prove $\nabla^{m+1}u =0$ at $O$.
For $j\in {\mathbb N}_0$, denote by $\nabla^j_\tau$ the vector of all tangential derivatives of order $j$, i.e.,
\begin{eqnarray*}
\nabla^j_\tau u=\left\{\partial_{\tau_1}^{j_1}\partial^{j_1}_{\tau_2}\,u: \quad j_1,j_2\in{\mathbb N}_0,j_1+j_2=j\right\}.
\end{eqnarray*} In view of the vanishing of the Cauchy data on $\Gamma$ and using (\ref{Derivative}) again, we have
\begin{eqnarray*}
\nabla_\tau^{m-3}\Delta^2u=\partial_{\tau_1}^{m+1}u=\partial_{\tau_1}^m\partial_{\tau_2}u
=\partial_{\tau_1}\partial_{\tau_2}^mu=\partial_{\tau_2}^{m+1}u=0\quad\mbox{at}\quad O.
\end{eqnarray*}
It was shown in \cite[Proposition A.3]{EH2017} that the span of the differential operators $\nabla_\tau^{m-3}\Delta^2$,
$\partial_{\tau_1}^{m+1}$, $\partial_{\tau_1}^m\partial_{\tau_2}$,
$\partial_{\tau_1}\partial_{\tau_2}^m$ and $\partial_{\tau_2}^{m+1}$ contains the vector $\nabla^{m+1}_\tau$. Hence, the relation $\nabla^{m+1}u =0$ at $O$ follows.
Taking $\nabla^{m-1}$ on the equations in (\ref{eq:26}) and using $q_1\neq q_2$, we see
\begin{eqnarray*}
\nabla^{m-1} u_2=\nabla^{m-1}\Delta^2 u_2=0\quad \mbox{at}\quad O.
\end{eqnarray*}
Taking $\nabla^{m-1}\Delta^2$ on the first equation in (\ref{eq:26}) gives $\nabla^{m-2}\Delta^2 u(O)=0$.
By induction we obtain $u_1=u_2\equiv 0$ in $B_1$.
\end{proof}
\section{Weakly singular corners always scatter}
To prove Theorem \ref{TH} for weakly singular points lying on $\partial D$, we only need to show that
\begin{lemma} \label{lem:WS}
Let $\Gamma$ be the profile of the function (\ref{f1}) .
Suppose that $u_j$ ($j=1,2$) are solutions to the Helmholtz equation $\Delta u_j+q_ju_j=0, j=1,2 $ in $B_1$ with $q_1\neq q_2 $, subject to the transmission conditions $ u_1=u_2,\frac{\partial u_1}{\partial \nu}= \frac{\partial u_2}{\partial \nu}$ on $\Gamma$. If $O\in \Gamma$ is a weakly singular point and $q_1\neq q_2$, then $u_1=u_2\equiv0$.
\end{lemma}
It seems non-trivial to prove the above lemma by extending the analysis in the proof of Lemma \ref{lem:SS} to the case of weakly singular corners. The analytical approach of using polar coordinates (see \cite{ElHu2015}) also turns out to be complicated. Below we shall present a novel approach by using the expansion of solutions to the Helmholtz equation in the Cartesian coordinate system.
Since $u_j$ satisfies the Helmholtz equation and $q_j$ is constant, the solution $u_j$ is analytic in $B_1$. Hence,
$u_j$ can be expanded into the convergent Taylor expansion
\begin{eqnarray*}
u_j=\sum_{n\ge0} \sum_{m \ge0} a^{(j)}_{n,m}\, x_1^n\, x_2^m \quad\mbox {in} \quad B_1,
\end{eqnarray*}
where the coefficients $a^{(j)}_{n,m}\in {\mathbb C}$ satisfy the relation
\begin{equation}
(n+1)(n+2) a^{(j)}_{n+2,m}+(m+1)(m+2)a^{(j)}_{n,m+2}+q_j a^{(j)}_{n,m}=0.
\label{eq:1}
\end{equation}
\par
Set $u=u_1-u_2$. Then $u$ admits the Taylor expansion
$$u=\sum_{n\ge0} \sum_{m \ge0} a_{n,m} x_1^n x_2^m\quad\mbox{in}\quad B_1,\qquad a_{n,m}:=a_{n,m}^{(1)}-a_{n,m}^{(2)},$$
and satisfies the equation $\Delta u+ q_1 u=(q_2-q_1)u_2$ in $B_1$. The implies that the coefficients $a_{n,m}$ fulfills the recursive relation
\begin{equation}
(n+1)(n+2) a_{n+2,m}+(m+1)(m+2)a_{n,m+2}+q_1 a_{n,m}=(q_2-q_1)a^{(2)}_{n,m}.
\label{eq:2}
\end{equation}
Combining (\ref{eq:1}) and (\ref {eq:2}), we deduce that
\begin {equation}
\begin{split}
0&=(m+4)(m+3)(m+2)(m+1)a_{n,m+4}+(n+4)(n+3)(n+2)(n+1)a_{n+4,m}\\
&\quad+2(n+2)(n+1)(m+2)(m+1)a_{n+2,m+2}\\
&\quad+(q_1+q_2)(n+2)(n+1)a_{n+2,m}
+(q_1+q_2)(m+2)(m+1)a_{n,m+2}\\
&\quad+q_2q_1 a_{n,m}.
\end{split}
\label{eq:3}
\end{equation}
We shall prove $a_{n,m}=0$ for all $n, m\in {\mathbb N}$ through (\ref{eq:3}) and the transmission conditions
\begin{eqnarray}
u=\partial_\nu u=0\quad\mbox{on}\quad \Gamma.
\end{eqnarray*}
This together with (\ref{eq:2}) would give rise to $a^{(2)}_{n,m}=a^{(1)}_{n,m}=0.$
Denote by $\Gamma_1:=\{(x_1, f(x_1): x_1\in[0,1)\}$ and $\Gamma_2=\{(x_1, f(x_1): x_1\in(-1,0])\}$, with the normal directions given by
\[
\nu^{(1)}(x_1)=(\alpha_1 c_1x_1^{\alpha_1-1},-1)^\top, \quad x_1>0;\qquad
\nu^{(2)}(x_1)=(\alpha_2 c_2x_1^{\alpha_2-1},-1)^\top,\quad x_1<0,
\] respectively. Observe that
\[
\frac{\partial u}{\partial x_1}=\sum_{n\ge0}\sum_{m\ge0}a_{n+1,m}(n+1)x_1^n x_2^m,\quad
\frac{\partial u}{\partial x_2}=\sum_{n\ge0}\sum_{m\ge0}a_{n,m+1}(m+1)x_1^n x_2^m.
\]
It follows from $\partial_\nu u=0$ on $\Gamma_j$ ($j=1,2$) that
\begin{eqnarray}\label{eq:4}
\alpha_1\sum_{n\ge0}\sum_{m\ge0}a_{n+1,m}(n+1)x_1^{n+\alpha_1 m+\alpha_1-1} c_1^{m+1}
-\sum_{n\ge0}\sum_{m\ge0}a_{n,m+1}(m+1)x_1^{n+\alpha_1 m}c_1^{m}=0, \\ \label{eq:5}
\alpha_2 \sum_{n\ge0}\sum_{m\ge0}a_{n+1,m}(n+1)x_1^{n+\alpha_2 m+\alpha_2-1} c_2^{m+1}
-\sum_{n\ge0}\sum_{m\ge0}a_{n,m+1}(m+1)x_1^{n+\alpha_2 m}c_2^{m}=0.
\end{eqnarray}
Without loss of generality, we suppose that $\alpha_1\leq \alpha_2$. In order to prove Lemma \ref{lem:WS}, we will consider two cases:
\par
\indent Case 1: $ \alpha_1=\alpha_2\geq 2$;\qquad\qquad\qquad\qquad\qquad
Case 2: $2\leq\alpha_1< \alpha_2$
The proofs in Cases 1 and 2
will be carried out in the subsequent two subsections, separately.
\subsection{Proof of Lemma \ref{lem:WS} when $\alpha_1=\alpha_2\geq 2$.}\label{subsec}
For notational convenience we set $\alpha:=\alpha_1=\alpha_2\geq 2$.
Equating coefficients of ${x_1}^l$ ($l\in {\mathbb N}, l\ge \alpha-1$) in (\ref{eq:4}) and (\ref{eq:5}) and changing properly the summation indices, we obtain
\begin{eqnarray}\label{eq:6}
\sum_{n+\alpha m=l-\alpha+2,n\ge1, m\ge0}\alpha n a_{n,m}c_1^{m+1}-\sum_{n+\alpha m=l+\alpha,n\ge0, m\ge1}ma_{n,m}c_1^{m-1}=0,\\ \label{eq:7}
\sum_{n+\alpha m=l-\alpha +2,n\ge1, m\ge0}\alpha n a_{n,m}c_2^{m+1}-\sum_{n+\alpha m=l+\alpha ,n\ge0, m\ge1}ma_{n,m}c_2^{m-1}=0.
\end{eqnarray} On the other hand, the Dirichlet condition
$u=0$ on $\Gamma$ gives the relations
\begin{equation}
\sum_{n\ge0}\sum_{m\ge0}a_{n,m}x^{n+\alpha m}_1{c_1}^{m}=0,\quad
\sum_{n\ge0}\sum_{m\ge0}a_{n,m}x^{n+\alpha m}_1c_2^m=0,
\label{eq:8}
\end{equation}
Equating coefficients of ${x_1}^l$ ($l\in {\mathbb N}$, $l\geq 0$) in (\ref{eq:8}), we get
\begin{eqnarray}\label{eq:9}
\sum_{n+\alpha m=l,n\ge0,m\ge0}a_{n,m}{c_1}^m=0,\quad
\sum_{n+\alpha m=l,n\ge0,m\ge0}a_{n,m}{c_2}^m=0.
\end{eqnarray}
By (\ref{eq:6}),(\ref{eq:7}) and ({\ref{eq:9}}) we shall prove $a_{n,m}=0$ for all $ n+\alpha m=j$ ($j\in {\mathbb N}$) by an induction argument on the index $j\in {\mathbb N}$. We divide the proof into four steps.
\par
{\bf Step 1:} Prove $a_{j,0}=0$ for all $j=0,1,\ldots \alpha-1$. This follows from
(\ref{eq:9}) with $n+\alpha m=l$ for $l=0,1,\ldots \alpha-1$.
\par
{\bf Step 2:} Prove $a_{j,0}=a_{j-\alpha,1}=0$ for all $j=\alpha, \alpha+1,\cdots, 2\alpha-1$.
Setting $ l=\alpha,\alpha+1,\dots,2\alpha-1$ in (\ref{eq:9}), we obtain
\[
a_{l,0}+c_1a_{l-\alpha,1}=0,\quad
a_{l,0}+c_2a_{l-\alpha,1}=0.
\]
Since $c_1\ne c_2$, we see $a_{l,0}=a_{l-\alpha,1}=0$.
{\bf Step 3:} Prove $a_{j,0}=a_{j-\alpha,1}=a_{j-2\alpha, 2}=0$ for all $j=2\alpha, 2\alpha+1,\cdots, 3\alpha-1$.
As done in previous two steps, setting $ n+\alpha m=2\alpha, 2\alpha+1,\dots,3\alpha-1$ in (\ref{eq:9}), we find
\begin{eqnarray}\label{neq:1}
a_{j,0}+c_1a_{j-\alpha,1}+{c_1}^2\,a_{j-2\alpha,2}=0,\quad
a_{j,0}+c_2a_{j-\alpha,1}+{c_2}^2\,a_{j-2\alpha,2}=0.
\end{eqnarray}
On the other hand, one may conclude from Steps 1 and 2 that
\[
a_{n,m}=0 \quad \mbox{if}\quad \ n+\alpha m < j.
\]
This together with $\alpha\geq 2$ implies that
\begin{eqnarray*}
0=\sum_{n+\alpha m=j-2\alpha+2,n\ge1, m\ge0}\alpha n a_{n,m}c_1^{m+1}=\sum_{n+\alpha m=j,n\ge1, m\ge0}\alpha n a_{n,m}c_2^{m+1}
\end{eqnarray*}
Hence, setting $l=j-\alpha$ in (\ref{eq:6}) and (\ref{eq:7}) gives the relations
\begin{eqnarray}\label{neq:2}
a_{j-\alpha,1}+2{c_1}a_{j-2\alpha,2}=0,\quad
a_{j-\alpha,1}+2{c_2}a_{j-2\alpha,2}=0.
\end{eqnarray}
Therefore, combining (\ref{neq:1}) and (\ref{neq:2}) yields $a_{j,0}=a_{j-\alpha,1}=a_{j-2\alpha,2}=0$. Further, we conclude from Steps 1-3 that
\begin{eqnarray}\label{neq:3}
a_{n,m}=0 \quad \mbox{if}\quad \ n+\alpha m < 3\alpha.
\end{eqnarray}
{\bf Step 4:} Prove $a_{j,0}=a_{j-\alpha,1}=a_{j-2\alpha, 2}=a_{j-3\alpha, 3}=0$ for all $j=3\alpha, 3\alpha+1,\cdots, 4\alpha-1$.
Setting $ n+\alpha m=3\alpha,3\alpha+1,\dots,4\alpha-1$ in (\ref{eq:9}), we get for such $j$ that
\begin{eqnarray*}
a_{j,0}+c_1a_{j-\alpha,1}+{c_1}^2a_{j-2\alpha,2}+{c_1}^3a_{j-3\alpha,3}=0,\\
a_{j,0}+c_2a_{j-\alpha,1}+{c_2}^2a_{j-2\alpha,2}+{c_2}^3a_{j-3\alpha,3}=0,
\end{eqnarray*}
Setting $l=j-\alpha$ in (\ref{eq:6})-(\ref{eq:7}) and making use of (\ref{neq:3}), we obtain
\begin{eqnarray*}
a_{j-\alpha,1}+2c_1a_{j-2\alpha,2}+3{c_1}^2a_{j-3\alpha,3}=0,\\
a_{j-\alpha,1}+2c_2a_{j-2\alpha,2}+3{c_2}^2a_{j-3\alpha,3}=0.
\end{eqnarray*}
For fixed $j\in\{3\alpha, 3\alpha+1,\cdots, 4\alpha-1 \}$, the previous relations can be written as the system
\begin{eqnarray}\label{neq:4}
\begin {pmatrix}
1&c_1&{c_1}^2&{c_1}^3\\
1&c_2&{c_2}^2&{c_2}^3\\
0&1&2c_1&3{c_1}^2\\
0&1&2c_2&3{c_2}^2
\end {pmatrix} \begin{pmatrix}
a_{j,0} \\ a_{j-\alpha, 1} \\ a_{j-2\alpha, 2} \\ a_{j-3\alpha, 3}
\end{pmatrix}=0.
\end{eqnarray}
It is not difficult to check that the determinant of the matrix on the left hand side of (\ref{neq:4}) is $-(c_1-c_2)^4\neq0$, implying that $ a_{j,0}=a_{j-\alpha,1}=a_{j-2\alpha,2}=a_{j-3\alpha,3}=0$.
Hence, it holds that
\begin{eqnarray*}
a_{n,m}=0 \quad \mbox{if}\quad \ n+\alpha m <4\alpha.
\end{eqnarray*}
{\bf Step 5:} Induction arguments. We make the induction hypothesis
$a_{n,m}=0$ for all $n+\alpha m<M$ for some $M\ge4\alpha, M\in {\mathbb N}$.
We need to prove that
\begin{equation}
a_{n,m}=0\quad \mbox{if}\quad n+\alpha m=M.
\label{eq:11}
\end{equation}
We first claim that
\begin{eqnarray}\label{neq:5}
a_{n,m}=0, \quad \mbox{if}\quad n+\alpha m=M, m\ge4.
\end{eqnarray}
Let $n'=n, m'=m-4\ge 0$. Then we see i
\[
n'+\alpha (m'+4)=M, \quad n', m'\geq 0.
\]
One can readily prove that
\[
n'+4+\alpha m'<M,\quad
n'+2+\alpha (m'+2)<M,\quad
n'+2+\alpha m'<M,\quad
n'+\alpha( m'+2)<M.
\]
Therefore, by induction hypothesis,
\[
a_{n'+4,m'}=a_{n'+2,m'+2}=a_{n'+2,m'}=a_{n',m'+2}=0.
\]
Using (\ref{eq:3}), we get the relation
\[
a_{n',m'+4}=0,\ \mbox{if}\quad n'+\alpha(m'+4)=M, \quad n',m' \ge 0,
\] which proves (\ref{neq:5}).
To proceed with the proof we set $ l=M$ in (\ref{eq:9}) to obtain
\begin{eqnarray*}
a_{M,0}+c_1a_{M-\alpha,1}+{c_1}^2a_{M-2\alpha,2}+{c_1}^3a_{M-3\alpha,3}=0,\\
a_{M,0}+c_2a_{M-\alpha,1}+{c_2}^2a_{M-2\alpha,2}+{c_2}^3a_{M-3\alpha,3}=0,
\end{eqnarray*} where the relation (\ref{neq:5}) was again used. On the other hand, setting $l=M-\alpha$ in (\ref{eq:6})-(\ref{eq:7}) and recalling the induction hypothesis, we see
\begin{eqnarray*}
a_{M-\alpha,1}+2c_1a_{M-2\alpha,2}+3{c_1}^2a_{M-3\alpha,3}=0,\\
a_{M-\alpha,1}+2c_2a_{M-2\alpha,2}+3{c_2}^2a_{M-3\alpha,3}=0.
\end{eqnarray*}
Note that the coefficient matrix for the unknowns $a_{M,0}, a_{M-\alpha, 1}, a_{M-2\alpha, 2}$ and $a_{M-3\alpha, 3}$ is the same as the 4-by-4 matrix on the left hand side of (\ref{neq:4}). Since the determinant of this matrix does not vanish, we obtain
$ a_{M,0}=a_{M-\alpha,1}=a_{M-2\alpha,2}=a_{M-3\alpha,3}=0$.
This together with (\ref{neq:5}) proves (\ref{eq:11}).
By induction, it holds that $a_{n,m}=0$ for all $n, m\in {\mathbb N}.$ In view of (\ref{eq:2}) and the condition $q_1\neq q_2$, we obtain $a_{n,m}^{(1)}=a_{n,m}^{(2)}=0$ for all $n, m\in {\mathbb N}$. Finally, we get $u_1=u_2\equiv0$ in $B_1$ by the analyticity of $u_j$ ($j=1,2$).
\subsection{Proof of Lemma \ref{lem:WS} when $2\leq\alpha_1<\alpha_2$.}
We first observe that the powers of $x_1$ in the first summation on the left hand side of
(\ref{eq:4}) and (\ref{eq:5}) are all greater than or equal to $\alpha_j-1$, whereas those in the second summation start from zero. Hence,
equating coefficients of the term $x_1^{l}$ ($l<\alpha_j-1$) in (\ref{eq:4}) and (\ref{eq:5}) yields
\begin{eqnarray}\label{eq:12}
\sum_{n\ge0,m\ge1,n+\alpha_1m=l+\alpha_1}a_{n,m}mc_1^{m-1}=0,\quad l\le\alpha_1-2,\\ \label{eq:14}
\sum_{n\ge0,m\ge1,n+\alpha_2m=l+\alpha_2}a_{n,m}mc_2^{m-1}=0,\quad l\le\alpha_2-2.
\end{eqnarray}
Analogously, equating coefficients of the term $x_1^{l}$ for $l\ge \alpha_j-1$, we obtain
\begin{eqnarray}\label{eq:13}
\sum_{n+\alpha_1m=l-\alpha_1+2,n\ge1, m\ge0}\alpha_1 na_{n,m}c_1^{m+1}-\sum_{n+\alpha_1m=l+\alpha_1,n\ge0, m\ge1}ma_{n,m}c_1^{m-1}=0, l\ge \alpha_1-1,\\ \label{eq:15}
\sum_{n+\alpha_2m=l-\alpha_2+2,n\ge1, m\ge0}\alpha_2 na_{n,m}c_2^{m+1}-\sum_{n+\alpha_2m=l+\alpha_2,n\ge0, m\ge1}ma_{n,m}c_2^{m-1}=0,l\ge \alpha_2-1.
\end{eqnarray}
From the Dirichlet boundary condition $u=0$ on $\Gamma$, we obtain
\[
\sum_{n\ge0}\sum_{m\ge0}a_{n,m}x^{n+\alpha_1 m}_1{c_1}^{m}=0,\quad \sum_{n\ge0}\sum_{m\ge0}a_{n,m}x^{n+\alpha_2 m}_1c_2^m=0,
\]
which implies that
\begin{eqnarray}\label{eq:16}
\sum_{n+\alpha_1m=l,n\ge0,m\ge0}a_{n,m}{c_1}^m=0,\quad l\ge0, \\\label{eq:17}
\sum_{n+\alpha_2m=l,n\ge0,m\ge0}a_{n,m}{c_2}^m=0, \quad l\ge0.
\end{eqnarray}
Since $\alpha_1<\alpha_2$, the proof in this section is more complicated than previous subsection. We shall still apply the induction argument to prove that $a_{n,m}=0$ for all $ n+\alpha m=j$, $j\in {\mathbb N}$. Below we carry out the proof under the assumption that $c_1\neq 0$. If $c_1=0$, we have $c_2\neq 0$ by assumption. Then the interface can be locally parameterized by the function given in (\ref{f1}) with $\alpha_1=\alpha_2$. Hence, the vanishing of $u_j$ in $B_1$ follows from the same arguments used in subsection \ref{subsec}, where the case $c_1=0$ is covered.
\textbf{Step 1:} Prove $a_{j,0}=0$ for all $j=\alpha_1,\alpha_1+1,\ldots, \alpha_2-1 $. This follows from (\ref{eq:17}).
\textbf{Step 2:} Prove $a_{j,0}=a_{j-\alpha_1,1}=0$ when $j=\alpha_1, \alpha_2, \ldots, min ({\alpha_2-1,2\alpha_1-1})$.
Setting $j=\alpha_1,\ldots 2\alpha_1-1$ in (\ref{eq:16}), we obtain
\begin{equation}
a_{j,0}+a_{j-\alpha_1,1}c_1=0.
\label {eq:18}
\end{equation}
This together with the condition $c_1\neq 0$ and the fact that $a_{l,0}=0$ for $l=\alpha_1,\alpha_1+1,\ldots, \alpha_2-1 $ (see Step 1) gives the desired result.
\textbf{Step 3:} Induction arguments. Assuming that
\begin{eqnarray*}
a_{n,m}=0\quad\mbox{for all}\quad n+m\alpha_1<M,\quad M> min({\alpha_2-1,2\alpha_1-1}),
\end{eqnarray*}
we will prove that
\begin{equation}
a_{n,m}=0\quad\mbox{for all}\quad n+m\alpha_1=M.
\label{eq:19}
\end{equation}
Setting $l=M$ in (\ref{eq:17}) gives
\begin{equation}
\sum_{n+m \alpha_2=M,n\ge0,m\ge0}a_{n,m}c_2^m=0.
\label{eq:20}
\end{equation}
Since $\alpha_1<\alpha_2$, the indices $n, m$ appearing in the above summation fulfill
$n+m\alpha_1<M$ if $m\ne0$.
By induction hypothesis, this implies that
\[
a_{n,m}=0, \quad \text{if} \ n+m\alpha_2=M, m\ne0.
\]
When $m=0$, it follows from (\ref{eq:20}) that $a_{M,0}=0$. Now, it remains to prove
\begin{eqnarray}\label{neq:6}
a_{n,m}=0 \quad \text{for all}\ n+\alpha_1m=M, \quad m\ne0.
\end{eqnarray}
in the following cases.
\subsubsection{Case 1: $M\le2\alpha_1-1.$}
Setting $l=M$ in (\ref{eq:16}), we obtain
\[
a_{M,0}+a_{M-\alpha_1,1}c_1=0,
\]
which together with $ a_{M,0}=0$ and $c_1\neq 0$ leads to $a_{M-\alpha_1,1}=0$.
This proves (\ref{neq:6}) when $M\le2\alpha_1-1.$
\subsubsection{Case 2: $2\alpha_1\le M\le3\alpha_1-1$.}
Letting $l=M$ in (\ref{eq:16}) and using again the fact that $a_{M,0}=0$, we obtain
\begin{eqnarray}\label{neq:7}
a_{M-\alpha_1,1}c_1+a_{M-2\alpha_1,2}{c_1}^2=0.
\end{eqnarray}
Setting $l=M-\alpha_1$ in (\ref{eq:13}) and making use of the induction hypothesis
\[
a_{n,m}=0,\quad\text{for all}\; n+\alpha_1m=l-\alpha_1+2<M,
\]
we obtain
\begin{eqnarray}\label{neq:8}
a_{M-\alpha_1,1}+2a_{M-2\alpha_1,2}{c_1}=0.
\end{eqnarray}
Combining (\ref{neq:7}) and (\ref{neq:8}) leads to $a_{M-\alpha_1,1}=a_{M-2\alpha_1,2}=0$,
which proves (\ref{neq:6}).
\subsubsection{Case 3: $3\alpha_1\le M \le 4\alpha_1-1 $.}\label{sub-1}
As done in previous cases, setting $l=M$ in (\ref{eq:16}) and using $a_{M,0}=0$ gives
\begin {equation}
a_{M-\alpha_1,1}c_1+a_{M-2\alpha_1,2}{c_1}^2+a_{M-3\alpha_1,3}{c_1}^3=0.
\label{eq:21}
\end{equation}
Setting $l=M-\alpha_1$ in (\ref{eq:13}). Recalling from the induction hypothesis that
\[
a_{n,m}=0,\quad\text{if}\ n+\alpha_1m=l-\alpha_1+2<M,
\]
we obtain
\begin{equation}
a_{M-\alpha_1,1}+2a_{M-2\alpha_1,2}{c_1}+3 a_{M-3 \alpha_1,3}{c_1}^2=0.
\label{eq:22}
\end{equation}
Next we will show $a_{M-\alpha_1,1}=0$. Write $N=M-\alpha_1+\alpha_2$ for notational simplicity.
If $N\le2\alpha_2-2$, setting $l=N-\alpha_2$ in (\ref{eq:14}) gives
$a_{M-\alpha_1,1}=0.$
If $N\ge2\alpha_2-1$, we have the relation
\[
M-\alpha_1+\alpha_2-2\alpha_2+2=M-\alpha_1-\alpha_2+2 <M.
\]
Letting $l=M-\alpha_1$ in (\ref{eq:15}), we can obtain
\begin{equation}
\sum_{n+\alpha_2m=N,n\ge0, m\ge1}a_{n,m}m{c_2}^{m-1}=0,
\label{eq:23}
\end{equation}
Since $\alpha_2>\alpha_1$, it holds that
\[
n+\alpha_1m <M\quad\mbox{for all}\quad n+\alpha_2m=N, m\ge2,
\]
implying that
\[
a_{n,m}=0\quad\mbox{for all}\quad n+\alpha_2m=N, m\ge2.
\]
due to the induction hypothesis.
Hence, it follows from (\ref{eq:23}) that $a_{M-\alpha_1,1}=0.$
Combining this with (\ref{eq:21}) and (\ref{eq:22}) and the fact that $c_1\ne0$, we obtain $a_{M-2\alpha_1,2}=a_{M-3\alpha_1,3}=0$, which proves (\ref{neq:6}).
\subsubsection{Case 4: $M\ge 4\alpha_1$.}\label{sec:4.2.4}
We first prove that
\begin{equation}
a_{n,m+4}=0, \quad\text{if} \ n+\alpha_1(m+4)=M, n\ge0, m\ge0.
\label{eq:24}
\end{equation}
Supposing that indices $n,m\ge 0$ in (\ref{eq:3}) satisfy the relation $n+\alpha_1 (m+4)=M$.
Then we have
\begin{eqnarray*}
n+4+\alpha_1 m<M,\quad
n+2+\alpha_1 (m+2)<M, \\
n+2+\alpha_1 m<M,\quad
n+\alpha_1( m+2)<M,
\end{eqnarray*}
By induction hypothesis, we see
\[
a_{n+4,m}=a_{n+2,m+2}=a_{n+2,m}=a_{n,m+2}=0.
\]
Hence, the relation (\ref{eq:24}) follows from (\ref{eq:3}).
To prove (\ref{neq:6}) we only need to verify
\begin{equation}
a_{M-\alpha_1,1}=a_{M-2\alpha_1,2}=a_{M-3\alpha_1,3}=0.
\label{eq:25}
\end{equation}
\par
Analogously to the Case 3, we will show that $a_{M-\alpha_1,1}=0$ by setting $N=M-\alpha_1+\alpha_2$.
If $N\le2\alpha_2-2$, letting $l=N$ in (\ref{eq:14}) leads to $a_{M-\alpha_1,1}=0$.
If $N\ge2\alpha_2-1$, letting $l=M-\alpha_1$ in (\ref{eq:15}) and noting that
\[
M-\alpha_1+\alpha_2-2\alpha_2+2=M-\alpha_1-\alpha_2+2 <M,
\]
we obtain
\begin{eqnarray}\label{neq:9}
\sum_{n+\alpha_2m=N,n\ge0, m\ge1}a_{n,m}m{c_2}^{m-1}=0.
\end{eqnarray}
Similar to the arguments in subsection \ref{sub-1}, we can obtain using the induction hypothesis that
\[
a_{n,m}=0\quad \mbox{for all}\quad n+\alpha_2m=N, m\ge2,
\]
because $n+\alpha_1m<M$ for such indices $n$ and $m$.
Therefore, we get $a_{M-\alpha_1,1}=0$ from (\ref{neq:9}).
Now, setting $l=M$ in (\ref{eq:16}), using (\ref{eq:24}) and the fact that $a_{M,0}=a_{M-\alpha_1,1}=0$, we see
\begin{eqnarray}\label{neq:11}
a_{M-2\alpha_1,2}{c_1}^2+a_{M-3\alpha_1,3}{c_1}^3=0.
\end{eqnarray}
On the other hand, setting $l=M-\alpha_1$ in (\ref{eq:13}), using $a_{M-\alpha_1,1}=0$ and the relations
\[
a_{n,m}=0\quad\mbox{for all}\quad n+\alpha_1m=l-\alpha_1+2=M-2\alpha_1-2<M,
\]
we deduce that
\begin{eqnarray}\label{neq:10}
2a_{M-2\alpha_1,2}{c_1}+3 a_{M-3 \alpha_1,3}{c_1}^2=0.
\end{eqnarray}
Since $c_1\ne0$, we obtain $a_{M-2\alpha_1,2}=a_{M-3\alpha_1,3}=0$ by combining (\ref{neq:11}) and (\ref{neq:10}). This finishes the proof of (\ref{neq:6}) when $M\geq 4\alpha_1$.
Finally, the relation (\ref{eq:19}) follows from (\ref{neq:6}) and the fact that
$a_{M,0}=0$. The proof of Lemma \ref{lem:WS} is thus complete under the assumption that $2\leq\alpha_1<\alpha_2$.
\section{Concluding remarks}
We remark that Lemma \ref{lem:WS} does not hold true if the curve $\Gamma$ is analytic. Counterexamples can
be easily constructed when $\Gamma$ is a line segment (see \cite[Remark 3.3]{EH2017} ) or a circle.
If $\Gamma\subset B_1$ is a circle of radius $R<1$ centered at the origin, one may find interior transmission eigenvalues (ITEs) ( or equivalently, $q_1$ and $q_2$) such that the coupling problem
\begin{eqnarray}\label{ITE}
\Delta u_j + q_j u_j =0\quad\mbox{in}\quad \Omega,\quad u_1=u_2,\quad \partial_\nu u_1=\partial_\nu u_2\quad\mbox{on}\quad \Gamma,
\end{eqnarray}
admits non-trivial solutions $u_1$ and $u_2$ in $B_R$ (see e.g.,\cite{CPJ}), which can be analytically extended to $B_1$. Here $\Omega\subset B_1$ denotes the domain enclosed by the closed curve $\Gamma$. Our Lemmas \ref{lem:SS} and \ref{lem:WS} imply that, if $\Gamma$ possesses a singular point, the non-trivial solutions $u_j$ to (\ref{ITE}) can not be analytically extended onto $\overline{B}_1$.
We refer to \cite{CGH2010, CP, CPJ, PJ,S} for the existence of ITEs in inverse scattering theory. Note that all results of this paper carry over to variable potential functions which is a constant in a small neighborhood of the
singular point under question. The singular points considered here form only a subset of non-analytical points of $\Gamma$.
We conjecture that Lemmas \ref{lem:SS} and \ref{lem:WS} remain valid under the weak assumption that $\Gamma$ contains a single non-analytical point. However, the proof requires novel mathematical arguments and the progress along this direction will be reported in our forthcoming publications.
\section{Acknowledgments}
G. Hu would like to thank J. Elschner for private discussions on the proof of Lemma \ref{lem:SS}.
The work of G. Hu is supported by the NSFC grant (No. 11671028), NSAF grant (No. U1530401) and the 1000-Talent Program of Young Scientists in China.
The work of L. Li and J. Yang are partially supported by the National Science Foundation of
China (61421062, 61520106004) and Microsoft Research of Asia.
|
\section{Introduction}
Let $G = (G, \cdot, e, (\,)^{-1})$ be a group (written multiplicatively,
with identity element $e$). A \emph{pseudo-length function} on $G$ is a
map $\normsymb : G \to [0,+\infty)$ that obeys the properties
\begin{itemize}
\item $\norm{e} = 0$,
\item $\norm{ x^{-1} } = \norm{x}$,
\item $\norm{ x y } \leq \norm{x} + \norm{ y }$
\end{itemize}
for all $x,y\in G$. If in addition we have $\norm{x}>0$ for all $x \in G
\setminus \{e\}$, we say that $\normsymb$ is a \emph{length function}. By
setting $d(x,y) \coloneqq \norm{x^{-1}y}$, it is easy to see that
pseudo-length functions (resp. length functions) are in bijection with
left-invariant pseudometrics (resp. left-invariant metrics) on $G$.
From the above properties it is clear that one has the upper bound
\[
\norm{x^n} \leq |n| \, \norm{x}
\]
for all $x \in G$ and $n \in \Z$. Let us say that a pseudo-length
function $\normsymb: G\to [0,+\infty)$ is \emph{homogeneous} if equality
is always attained here, in that one has
\begin{equation}\label{linear-growth}
\norm{x^n} = |n| \, \norm{x}
\end{equation}
for all $x \in G$ and any $n \in \Z$. Using the axioms of a
pseudo-length function, it is not difficult to show that the homogeneity
condition \eqref{linear-growth} is equivalent to the triangle inequality
holding with equality whenever $x=y$ (i.e., that \eqref{linear-growth}
holds for $n=2$); see \cite[Lemma 1]{GK}.
If one has a real or
complex Banach space $\mathbb{B} = (\mathbb{B},\| \, \|)$, and $\phi: G
\to \mathbb{B}$ is any homomorphism from $G$ to $\mathbb{B}$ (viewing the
latter as a group in additive notation), then the function $\normsymb: G
\to [0,+\infty)$ defined by $\norm{x} \coloneqq \| \phi(x) \|$ is easily
verified to be a homogeneous pseudo-length function. Furthermore, if
$\phi$ is injective, then $\normsymb$ is in fact a
homogeneous length function. For instance, the function
$\norm{ (n, m) } \coloneqq |n + \sqrt{2} m|$
is a length function on $\Z^2$, where in this case $\mathbb{B} \coloneqq
\R$ and $\phi((n,m)) \coloneqq n + \sqrt{2} m$. On the other hand, one
can easily locate many length functions that are not homogeneous, for
instance by taking the square root of the length function just
constructed.
The main result of this paper is that such Banach space constructions are
in fact the \emph{only} way to generate homogeneous (pseudo-)length
functions.
\begin{theorem}[Classification of homogeneous length functions]\label{class}
Given a group $G$, let $\normsymb: G \to [0,+\infty)$ be a homogeneous
pseudo-length function. Then there exist a real Banach space $\mathbb{B}
= (\mathbb{B},\| \, \|)$ and a group homomorphism $\phi : G \to
\mathbb{B}$ such that $\norm{x} = \| \phi(x) \|$ for all $x \in G$.
Furthermore, if $\normsymb$ is a length function, one can take $\phi$ to
be injective, i.e., an isometric embedding.
\end{theorem}
We derive Theorem~\ref{class} from a more quantitative result
bounding the pseudo-length of a commutator
\begin{equation}\label{comdef}
[x,y] \coloneqq xyx^{-1}y^{-1};
\end{equation}
see Proposition~\ref{main} below. Our arguments are elementary, relying
on directly applying the axioms of a homogeneous length function to
various carefully chosen words in $x$ and $y$, and repeatedly taking an
asymptotic limit $n \to \infty$ to dispose of error terms that arise in
the estimates obtained in this fashion.
An additional advantage of quantifying Theorem~\ref{class}
in Proposition~\ref{main} is that one can derive from the latter
proposition a ``quasified'' version of Theorem~\ref{class}. See
Theorem~\ref{quasi} below.\footnote{A different variant of
Theorem~\ref{class} involves replacing homogeneity by the assumption that
$\normsymb$ is a pseudo-length function on $G$ whose homogenization is
positive:
\[
\normsymb_{\rm hom}(g) \coloneqq \lim_{n \to \infty} \frac{\norm{g^n}}{n}
> 0, \qquad \forall g \neq e.
\]
(This was studied in \cite[Theorem 2.10(III)]{Niem} in the special case
of abelian $(G,\normsymb)$.) In this case we work with $(G,
\normsymb_{\rm hom})$ instead of $\normsymb$, to conclude that $G$ maps
into a Banach space.}
Finally, as one quick corollary of Theorem \ref{class}, we obtain the
following characterization of the groups that admit homogeneous length
functions.
\begin{corollary}\label{abel}
A group admits a homogeneous length function if and only if it is
abelian and torsion-free.
\end{corollary}
\subsection{Examples and approaches}
We now make some remarks to indicate the nontriviality of
Theorem~\ref{class}. Corollary \ref{abel} implies that there are no
non-abelian groups with homogeneous length functions.
Whether or not such a striking geometric rigidity phenomenon holds was
previously unknown to experts. Moreover, the corollary fails
to hold if one or more of the precise conditions in the theorem are
weakened. For instance, such length functions indeed exist (i)~on
non-abelian monoids, and (ii)~on balls of finite radius in free groups.
We explain these two cases further in Section \ref{remarks-sec}.
Given these cases, one could \textit{a priori} ask if every non-abelian
group admits a homogeneous length function. This is not hard to disprove;
here are two examples.
\begin{example}[Nilpotent groups]
If $G$ is nilpotent of nilpotency class two (e.g., the Heisenberg group),
then $[x,y]^{n^2} = [x^n,y^n]$ for all $x,y \in G$ and integers $n \geq
0$ since the map $(g,h) \mapsto [g,h]$ is now a bihomomorphism
$G \times G \to [G,G] \subset Z(G)$.
If $[x,y]$ is non-trivial, then any homogeneous length function on $G$
would assign a linearly growing quantity to the right-hand side and a
quadratically growing quantity to the left-hand side, which is absurd;
thus such groups cannot admit homogeneous length functions. The claim
then also follows for nilpotent groups of higher nilpotency class, since
they contain subgroups of nilpotency class two.\footnote{One can also
show by relatively simple means that solvable non-abelian groups cannot
admit homogeneous length functions either; see the discussion on
lamplighter groups in the comments to
\href{https://terrytao.wordpress.com/2017/12/16/bi-invariant-metrics-of-linear-growth-on-the-free-group/}{\tt
terrytao.wordpress.com/2017/12/16/}.}
\end{example}
\begin{example}[Connected Lie groups]
As we explain in Remark~\ref{Rinv}, a homogeneous length function
$\normsymb$ induces a \textit{bi-invariant} metric on $G$.
Now if $(G, \normsymb)$ is furthermore a connected Lie group, then by
\cite[Lemma 7.5]{Milnor}, $G \cong K \times \R^n$ for some compact Lie
group $K$ and integer $n \geq 0$. By \eqref{linear-growth}, $K$ cannot
have torsion elements, hence must be trivial. But then $G$ is abelian.
\end{example}
Prior to Corollary \ref{abel}, the above examples left open the question
of whether \textit{any} non-abelian group admits a homogeneous length
function. One may as well consider groups generated by two non-commuting
elements. As a prototypical example, let $\free{2}$ be the free group on
two generators $a,b$. The word length function on $\free{2}$ is a length
function, but it is not homogeneous, since for instance the word length
of $(bab^{-1})^n = ba^n b^{-1}$ is $n+2$, which is not a linear function
of $n$. It is however the case that the word length of $x^n$ has linear
\emph{growth} in $n$ for any non-trivial $x$. Similarly for the
Levenshtein distance (edit distance) on $\free{2}$.
Our initial attempts to construct homogeneous length functions on
$\free{2}$ all failed. Of course, this failure is explained by our main
result. However, many of these methods apply under minor weakening of the
hypotheses, such as working with monoids rather than groups, or weakening
homogeneity. Results in these cases are discussed further in Section
\ref{remarks-sec}.
\subsection{Further motivations}
We next mention some motivations from functional analysis and
probability, or more precisely the study of Banach space embeddings.
If $G$ is an additive subgroup of a Banach space $\mathbb{B}$, then
clearly the norm on $\mathbb{B}$ restricts to a homogeneous length
function on $G$. In~\cite{CSC,GK} one can find several equivalent
conditions for a given length function on a given group to arise in this
way (studied in the broader context of additive mappings and separation
theorems in functional analysis); see also \cite[Theorem 2.10(II)]{Niem}
for an alternative proof. These conditions are summarized in \cite{KR2}.
For instance, given a group $G$ with a length function $\ell$, there
exists an isometric embedding from $G$ to a Banach space $\mathbb{B}$
with $\normsymb$ induced from the metric on $\mathbb{B}$, if and only if
$G$ is amenable and $\norm{x^2} = 2 \norm{x}$ for all $x$.
In view of such equivalences, it is natural to try and characterize the
groups possessing a homogeneous length function. This question is
answered by Corollary~\ref{abel}, which shows these are precisely the
abelian torsion-free groups.
Groups and semigroups with translation-invariant metrics also naturally
arise in probability theory, with the most important `normed' (i.e.,
homogeneous) examples being Banach spaces~\cite{LT}. Notice however that
in certain fundamental stochastic settings, formulating and proving
results does not require the full Banach space structure. In this vein, a
general variant of the Hoffmann-J{\o}rgensen inequality was shown in
\cite{KR1} in arbitrary metric semigroups -- including Banach spaces as
well as (non-abelian) compact Lie groups. Similarly in \cite{KR2}, the
authors transferred the (sharp) Khinchin--Kahane inequality from Banach
spaces to abelian groups $G$ equipped with a homogeneous length function.
To explore extensions of these results to the non-abelian setting (e.g.,
Lie groups with left-invariant metrics), we need to first understand if
such objects exist. As explained above, this question was not answered in
the literature; but it is now settled by our main result.
Finally, there may also be a relation to the Ribe program \cite{Naor},
which aims to reformulate aspects of Banach space theory in purely metric
terms. Indeed, from Corollary~\ref{abel} we see that a metric space $X$
is isometric to an additive subgroup of a Banach space if and only if
there is a group structure on $X$ which makes the metric left-invariant
and the length function $\norm{ x } \coloneqq d(1,x)$ homogeneous.
\section*{Acknowledgements}
This project is an online collaboration that originated from a blog post
at \url{https://terrytao.wordpress.com/2017/12/16}, following the model
of the ``Polymath'' projects \cite{gn}. A full list of participants and
their grant acknowledgments may be found at
{\small
\url{http://michaelnielsen.org/polymath1/index.php?title=linear\_norm\_grant\_acknowledgments}}.
We also thank Michal Doucha for useful references and comments, in
particular in bringing the paper \cite{Niem} to our attention, and the
anonymous referee for helpful suggestions.
\section{Key proposition}
The key proposition used to prove Theorem~\ref{class} is the following
estimate, which can treat a somewhat more general class of functions than
homogeneous pseudo-length functions, in which the symmetry hypothesis is
dropped and one allows for an error in { the homogeneity
property}, which is now { also} only claimed for $n=2$.
\begin{proposition}\label{main}
Let $G = (G,\cdot)$ be a group, let { $c \in \R$}, and let
$\normsymb: G \to \R$ be a function obeying the following axioms:
\begin{itemize}
\item[(i)] For any $x,y \in G$, one has
\begin{equation}\label{ng}
\norm{xy} \leq \norm{x}+\norm{y}.
\end{equation}
\item[(ii)] For any $x \in G$, one has
\begin{equation}\label{double}
\norm{x^2} \geq 2\norm{x} - { c}.
\end{equation}
\end{itemize}
Then for any $x,y \in G$, one has
\begin{equation}\label{xyc}
\norm{ [x,y] } \leq { 5 c},
\end{equation}
where the commutator $[x,y]$ was defined in \eqref{comdef}.
\end{proposition}
Notably, we neither assume symmetry $\norm{ x^{-1} } = \norm{ x }$, not
even up to a constant, nor $\norm{ e } = 0$ { (although $0
\leq \norm{ e } \leq c$ follows from the axioms)}; we also allow
$\normsymb$ to take on negative values. The reader may however wish to
restrict attention to { homogeneous} length functions, and
set { $c=0$ and $\normsymb \geq 0$} for a first reading of
the arguments below. The { factor of $5$ is} probably not
optimal here, but the crucial feature of the bound \eqref{xyc} for our
main application is that the right-hand side vanishes when {
$c=0$} (the right-hand side is also independent of $x$ and $y$, which we
use in other applications).
{ We define a \emph{semi-length function} to be a function
$\normsymb: G \to \R$ such that for all $x, y\in G$, $\norm{xy} \leq
\norm{x} + \norm{y}$, i.e. $\normsymb$ satisfies ~\eqref{ng}. Every
pseudo-length function is a semi-length function. A semi-length function
that satisfies \eqref{double} for some $c\in\R$ is called
\emph{quasi-homogeneous}.
\begin{remark}\label{quasilength}
Suppose $\normsymb: G \to \R$ and there is a constant $k$ such that
$\norm{xy} \leq \norm{x} + \norm{y} + k$ for all $x, y\in G$. Then the
function $\normsymb'(x) := \norm{x} + k$ is a semi-length {
function}. Further,
$\normsymb'$ satisfies~\eqref{double} with $c$ replaced by $c' := k+c$,
{ whenever $\normsymb$ satisfies~\eqref{double} on the nose.}
Thus Proposition~\ref{main} continues to hold if \eqref{ng} is replaced
by the condition $\norm{xy} \leq \norm{x} + \norm{y} + k$ for all $x,
y\in G$, with the bound in the conclusion \eqref{xyc} becoming $5c + 4k$.
\end{remark}
}
We now turn to the proof. For the remainder of this section, let $G$,
{ $c$}, and $\normsymb$ satisfy the hypotheses of the
proposition. Our task is to establish the bound \eqref{xyc}. We shall
now use \eqref{ng} and \eqref{double} repeatedly to establish a number of
further inequalities relating the { semi-}lengths $\norm{x}$ of various
elements $x$ of $G$, culminating in \eqref{xyc}.
Many of our inequalities will involve terms that depend on an auxiliary
parameter $n$, but we will be able to eliminate several of them by the
device of passing to the limit $n \to \infty$. It is because of this
device that we are able to obtain a bound \eqref{xyc} whose right-hand
side is completely uniform in $x$ and $y$.
From \eqref{ng} and induction we have the upper homogeneity
bound
\begin{equation}\label{ng-iter}
\norm{x^n} \leq { n \norm{x}}
\end{equation}
for any natural number $n \geq 1$. Similarly, from \eqref{double} and
induction one has the lower homogeneity bound
\[
\norm{x^n} \geq { n \norm{x} - \log_2(n)\, c \geq n \norm{x} - n c}
\]
whenever $n$ is a power of two. It is convenient to rearrange this
latter inequality as
\begin{equation}\label{rearrange}
\norm{x} \leq \frac{\norm{x^n}}{n} + { c}.
\end{equation}
This inequality, particularly in the asymptotic limit $n \to \infty$,
will be the principal means by which the hypothesis \eqref{double} is
employed.
We remark that by further use of \eqref{ng-iter} one can also obtain a
similar estimate to \eqref{rearrange} for natural numbers $n$ that are
not powers of two, but the powers of two will suffice for the
arguments that follow.
\begin{lemma}[Approximate conjugation invariance]\label{aci}
For any $x,y \in G$, one has
\[
\norm{yxy^{-1} } \leq { \norm{ x } + c}.
\]
\end{lemma}
\begin{remark}\label{Rinv}
Setting { $c=0$}, we conclude that any homogeneous
pseudo-length function is conjugation invariant, and thus determines a
bi-invariant metric on $G$. It should not be surprising that this
observation is used in the proof of Theorem~\ref{class}, since it is a
simple consequence of that theorem.
\end{remark}
\begin{proof}[Proof of Lemma~\ref{aci}]
From \eqref{rearrange} with $x$ replaced by $yxy^{-1}$, one has
\[
\norm{yxy^{-1}} \leq \frac{\norm{y x^n y^{-1}} }{n} + { c}
\]
whenever $n$ is a power of two. On the other hand, from \eqref{ng-iter}
and \eqref{ng} one has
\[
\norm{ y x^n y^{-1} } \leq \norm{y} + n \norm{x} + \norm{y^{-1}}
\]
and thus
\[
\norm{yxy^{-1}} \leq \norm{x} + { c} + \frac{\norm{y} +
\norm{y^{-1}} { - c}}{n}.
\]
Sending $n \to \infty$, we obtain the claim.
\end{proof}
\begin{lemma}[Splitting lemma]\label{split}
Let $x,y,z,w \in G$ be such that $x$ is conjugate to both $wy$ and
$zw^{-1}$. Then one has
\begin{equation}\label{ineq}
\norm{x} \leq \frac{\norm{y} + \norm{ z } }{2} + \frac{3}{2}
{ c}.
\end{equation}
\end{lemma}
\begin{proof}
If we write $x = swys^{-1} = t zw^{-1} t^{-1}$ for some $s,t \in G$, then
from \eqref{rearrange} we have
\begin{align*}
\norm{x} &\leq \frac{\norm{ x^n x^n } }{2n} + { c} \\
&= \frac{\norm{ s(wy)^n s^{-1}t (zw^{-1})^n t^{-1} } }{2n} +
{ c}
\end{align*}
whenever $n$ is a power of two. From Lemma~\ref{aci} and \eqref{ng} one
has
\begin{align*}
\norm{(wy)^{k+1} s^{-1}t (zw^{-1})^{k+1} } &= \norm{ w y (wy)^k s^{-1} t
(zw^{-1})^k z w^{-1} } \\
&\leq \norm{ y (wy)^k s^{-1} t (zw^{-1})^k z } + { c} \\
&\leq \norm{ (wy)^k s^{-1} t (zw^{-1})^k } + \norm{y} + \norm{z} +
{ c}
\end{align*}
for any $k \geq 0$, and hence by induction
\[
\norm{(wy)^n s^{-1}t (zw^{-1})^n } \leq \norm{s^{-1} t} + n ( \norm{y} +
\norm{z} + { c} ).
\]
Inserting this into the previous bound for $\norm{x}$ via two
applications of \eqref{ng}, we conclude that
\[
\norm{x} \leq \frac{\norm{y} + \norm{z}+ { c} }{2} +
\frac{\norm{s} + \norm{s^{-1}t} + \norm{t^{-1}} }{2n} + { c};
\]
sending $n \to \infty$, we obtain the claim.
\end{proof}
\begin{corollary}
If $x,y \in G$, let $f = f_{x,y}: \Z^2 \to \R$ denote the function
\[
f(m,k) \coloneqq \norm{ x^m [x,y]^k }.
\]
Then for any $m,k \in \Z$, we have
\begin{equation}\label{fmk}
f(m,k) \leq \frac{f(m-1,k) + f(m+1,k-1)}{2} + { 2c}.
\end{equation}
\end{corollary}
\begin{proof}
Observe that $x^m [x,y]^k$ is conjugate to both $x (x^{m-1} [x,y]^k)$ and
to $(y^{-1} x^m [x,y]^{k-1} xy) x^{-1}$, hence by \eqref{ineq} one has
\[
\norm{ x^m [x,y]^k } \leq \frac{\norm{ x^{m-1} [x,y]^k } + \norm{ y^{-1}
x^{m} [x,y]^{k-1} xy }}{2} + { \frac{3}{2} c}.
\]
Since $y^{-1} x^{m} [x,y]^{k-1} xy $ is conjugate to $x^{m+1}
[x,y]^{k-1}$, the claim now follows from Lemma~\ref{aci}.
\end{proof}
We now prove Proposition~\ref{main}. Let $x,y \in G$. We can
write the inequality \eqref{fmk} in probabilistic form as
\[
f(m,k) \leq {\mathbf E} f\left( \left(m,k-\frac{1}{2}\right) + Y
\left(1,-\frac{1}{2}\right) \right) + { 2 c}
\]
where $Y = \pm 1$ is a Bernoulli random variable that equals $1$ or $-1$
with equal probability. The key point here is the drift of $\left(0,
-\frac{1}{2} \right)$ in the right-hand side.
Iterating this inequality, we see that
\[
f( 0, n) \leq {\mathbf E} f\left( (Y_1 + \dots + Y_{2n}) \left(1,
-\frac{1}{2}\right) \right) + { 4 c n},
\]
where $n \geq 0$ and $Y_1,\dots,Y_{2n}$ are independent copies of $Y$ (so
in particular $Y_1+\dots+Y_{2n}$ is an even integer).
From \eqref{ng} and \eqref{ng-iter} one has the inequality
\begin{eqnarray*}
f(m,k) & \leq &
|m| \left(\max( \norm{x}, \norm{x^{-1}}) \right)\\
& & + |k| \left(\max(
\norm{[x,y]}, \norm{[x,y]^{-1}}) \right) + \norm{e}
\end{eqnarray*}
for all integers $m,k$, where the final term $\norm{e}$ is used when
$m=k=0$, { but can also be added in the remaining cases since
it is non-negative}. We conclude that
\[
f\left( (Y_1 + \dots + Y_{2n}) \left(1, -\frac{1}{2}\right) \right) \leq
A |Y_1 + \dots + Y_{2n}| + \norm{e}
\]
where $A$ is a quantity independent of $n$; more explicitly, one can take
\[
A \coloneqq \max\left( \norm{x}, \norm{x^{-1}}\right) + \frac{1}{2}
\max\left(\norm{[x,y]}, \norm{[x,y]^{-1}}\right).
\]
Taking expectations, since the random variable $Y_1+\dots+Y_{2n}$ has
mean zero and variance $2n$, we see from the Cauchy--Schwarz inequality
or Jensen's inequality that
\[
{\mathbf E}|Y_1 + \dots + Y_{2n}| \leq \left( {\mathbf E}|Y_1 + \dots +
Y_{2n}|^2 \right)^{1/2} = \sqrt{2n}
\]
and hence
\[
f(0,n) \leq A \sqrt{2n} + \norm{e} + { 4 c n}.
\]
But from \eqref{rearrange}, if $n$ is a power of $2$ then we have
\[
\norm{[x,y]} \leq \frac{f(0,n)}{n} + { c}.
\]
Combining these two bounds and sending $n \to \infty$, we obtain
Proposition~\ref{main}.
\begin{remark}
{
One can deduce a `local' version of Proposition \ref{main} as follows:
notice that the constant $c$ can be described in terms of $\normsymb$
from \eqref{double}, to yield
\begin{equation}\label{9bound}
\norm{ [x,y] } \leq 5 \sup_{z \in G} \left( 2 \norm{z} - \norm{z^2} \right)
\end{equation}
for any group $G$ and function $\normsymb : G \to \R$ for which this supremum exists, and any $x,y \in G$.
(Both sides are zero when $G$ is a Banach space and $\normsymb$ is the
norm, so equality is obtained in that case.)
It is also enough to consider the supremum over the subgroup of $G$
generated by $x$ and $y$ without loss of generality, which may lead to a
better bound on $\norm{ [x,y] }$ than taking the supremum over all of $G$.
Notice also that the constant $c$ must be non-negative, from \eqref{double} and \eqref{ng} with $x=y=e$:
\begin{equation}\label{Enonneg2}
c \geq 2 \norm{e} - \norm{e^2} = \norm{e} \geq \norm{e^2} - \norm{e} = 0.
\end{equation}
In fact, this reasoning and our results imply that the only way to get
$c=0$ on the right-hand side of \eqref{xyc} is when $\normsymb$
arises from pulling back the norm of a Banach space $\mathbb{B}$ along a
group homomorphism $G\to\mathbb{B}$, or equivalently along a group
homomorphism from the torsion-free abelianization of $G$ to $\mathbb{B}$.
}
\end{remark}
\section{Completing the proof of Theorem~\ref{class}}
With Proposition~\ref{main} in hand, it is not difficult to conclude the
proof of Theorem~\ref{class}. Suppose that $G$ is a group with a
homogeneous { semi-}length function $\normsymb: G \to [0,+\infty)$.
Applying Proposition~\ref{main} with { $c=0$}, we conclude that
$\norm{[x,y]}=0$ for all $x,y \in G$, thus by the triangle inequality
$\normsymb$ vanishes on the commutator subgroup $[G,G]$, and therefore
factors through the abelianization $G_{\mathrm{ab}} \coloneqq G/[G,G]$ of
$G$. Observe that this already establishes part of one implication of
Corollary~\ref{abel}. Factoring out by $[G,G]$ like this, we may now
assume without loss of generality that $G$ is abelian. To
reflect this, we now use additive notation for $G$, thus for instance
$\norm{nx} = |n| \norm{x}$ for each $x \in G$ and $n \in \Z$, and one can
also view $G$ as a module over the integers $\Z$.
At this point we repeat the arguments in \cite[Theorem B]{KR2}, which
treated the case when $G$ was separable, though it turns out that this
separability hypothesis is unnecessary.
If $x$ is a torsion element of $G$, i.e.~$nx=0$ for some $n$, then the
homogeneity condition forces $\norm{x}=0$. Thus $\normsymb$ vanishes on
the torsion subgroup of $G$; factoring out by this subgroup, we may thus
assume without loss of generality that $G$ is not only abelian, but is
also torsion-free.
We can view $G$ as a subgroup of the $\Q$-vector space $G \otimes_\Z \Q$,
the elements of which can be formally expressed as $\frac{1}{n} x$ for
natural numbers $n$ and elements $x \in G$ (with two such expressions
$\frac{1}{n} x, \frac{1}{m} y$ identified if and only if $mx = ny$, and
the $\Q$-vector space operations defined in the obvious fashion); the
fact that this is well defined as a $\Q$-vector space follows from the
hypotheses that $G$ is abelian and torsion-free. We can then define
the map $\| \, \|_\Q: G \otimes_\Z \Q \to [0,+\infty)$ by setting
\[
\left\|\frac{1}{n} x\right\|_\Q \coloneqq \frac{1}{n} \norm{x}
\]
for any $x \in G$ and natural number $n$; the linear growth condition
ensures that $\|\,\|_\Q$ is well-defined. It is not difficult to verify
that $\| \, \|_\Q$ is indeed a seminorm over the $\Q$-vector space $G
\otimes_\Z \Q$.
The norm $\| \, \|_\Q$ on $G \otimes_\Z \Q$ gives a metric $d(x, y) = \|
x-y \|_\Q$. Consider the metric completion $\mathbb{B}$ of $G \otimes_\Z
\Q$ with this metric. It is easy to see that the $\Q$-vector space
structure on $G \otimes_\Z \Q$ extends to an $\R$-vector space structure
on $\mathbb{B}$, and the norm $\|\,\|_\Q$ on $G \otimes_\Z \Q$ extends
to a norm $\|\,\|_\R$ on $\mathbb{B}$. As $\mathbb{B}$ is complete by
construction, it is a Banach space. The inclusion of $G$ in $G \otimes_\Z
\Q$ gives a homomorphism $\phi: G \to \mathbb{B}$ as required.
This concludes the proof of Theorem~\ref{class}. Since the homomorphism
$\phi: G \to \mathbb{B}$ can only be injective for abelian torsion-free
$G$, we obtain the ``only if'' portion of Corollary~\ref{abel}.
Conversely, if a group $G$ is abelian and torsion-free, by the above
constructions it embeds into a real vector space $\mathbb{B} \coloneqq G
\otimes_\Z \R$; now by Zorn's lemma $\mathbb{B}$ has a norm (e.g.,
consider the $\ell^1$ norm with respect to a Hamel basis of
$\mathbb{B}$), which restricts to the desired homogeneous length function
on $G$.
We remark that $G \otimes_\Z \R$ is the construction of the smallest,
`enveloping' vector space containing a copy of the abelian, torsion-free
group $G$.
\begin{remark}
The above arguments also show that homogeneous pseudo-length functions on
$G$ are in bijection with seminorms on the real vector space
$G_{\mathrm{ab},0} \otimes_\Z \R$, where $G_{\mathrm{ab},0}$ denotes the
torsion-free abelianization of $G$.
\end{remark}
\section{Further remarks and results}\label{remarks-sec}
If we weaken any of several conditions in Corollary \ref{abel}, then
examples of non-abelian structures with generalized length functions do,
in fact, often exist.
However, the generality of Proposition 2.1 allows us to obtain
non-trivial information in some of these cases. Here we mention several
such cases and discuss other related problems.
\subsection{Monoids and embeddings}
Our first weakening is to replace `groups' by the more primitive
structures `monoids' or `semigroups'. In this case, Robert Young (private
communication) described to us non-abelian monoids with homogeneous,
bi-invariant length functions:
consider the free monoid $\mathrm{FMon}(X)$ on any alphabet $X$ of size
at least $2$, with the edit distance $d(v,w)$ between strings $v,w \in
\mathrm{FMon}(X)$ being the least number of single generator insertions
and deletions to get from $v$ to $w$. The triangle inequality and
positivity are easily verified, while homogeneity of the corresponding
length function $\norm{x} \coloneqq d(e,x)$ is trivial. Moreover, the
metric $d(\cdot,\cdot)$ turns out to be bi-invariant:
\[
d(gxh,gyh) = d(x,y) \ \text{ for all }\ g,h,x,y \in \mathrm{FMon}(X).
\]
This specializes to left- and right-invariance upon taking $g \in X$ and
$h = e$, or $h \in X$ and $g = e$, respectively.
Note moreover that $\mathrm{FMon}(X)$ embeds into the free group
$\mathrm{FGp}(X)$ generated by $X$ and $X^{-1}$, where $X^{-1}$ is the
collection of symbols defined to be inverses of elements of $X$. In
particular, $\mathrm{FMon}(X)$ is cancellative. While this trivially
addresses the embeddability issue, notice that a more refined version of
embeddability fails. Namely, by our main theorem, $\mathrm{FMon}(X)$ does
not embed into any group in the category $\mathcal{C}_{\rm{bi-inv,\
hom}}$ with cancellative semigroups with homogeneous bi-invariant metrics
as objects and isometric semigroup maps as morphisms. Thus, one may
reasonably ask what is a sufficiently small category in which the
embeddability works. The following proposition shows that we just need to
drop homogeneity.
\begin{proposition}\label{Pisometry}
Let $\mathcal{C}_{\rm{bi-inv}}$ denote the category whose objects are
cancellative semigroups with bi-invariant metrics, and morphisms are
isometric semigroup maps. Then $\mathrm{FMon}(X)$ embeds isometrically
into $\mathrm{FGp}(X)$ in $\mathcal{C}_{\rm{bi-inv}}$.
\end{proposition}
\begin{proof}
From above, $\mathrm{FMon}(X)$ is an object of
$\mathcal{C}_{\rm{bi-inv}}$; denote the metric by $d_{FM}$. One can check
that $d_{FM}(w,w')$ equals the difference between $\normsymb(w) +
\normsymb(w')$ and twice the length of the longest common (possibly
non-contiguous) substring in $w,w'$; here, $\normsymb$ denotes the length
of a word in the alphabet $X$.
We next claim $\mathrm{FGp}(X)$ is also an object of
$\mathcal{C}_{\rm{bi-inv}}$. Namely, for a word $w=x_1x_2\cdots x_m$ in
the free group, we consider \emph{non-crossing matchings} in $w$, i.e.,
sets $M$ of pairs of letters in $\{1,2,\dots m\}$ such that the following
hold.
\begin{itemize}
\item If $(i, j)\in M$, then $i< j$ and $x_j=x_i^{-1}$.
\item If $(i, j), (k, l) \in M$, then either $(i, j) = (k, l)$ or
$i,j,k,l$ are distinct.
\item If $i<k<j<l$ and $(i, j)\in M$, then $(k, l)\notin M$.
\end{itemize}
Given a matching $M$ as above, consider the set $U=U(M)$ of indices $k$,
$1\leq k\leq m$ which are not part of a pair in $M$. Define the
\emph{deficiency} of the matching $M$ as the cardinality of the set
$U(M)$, and define the length $\normsymb_{wc}(w)$ of the word $w$ as the
infimum of the deficiency over all non-crossing matchings in $w$ (the
subscript in $\normsymb_{wc}$ stands for Watson--Crick). This length was
previously studied in \cite{gadgil}, including checking that it is
well-defined on all of $\mathrm{FGp}(X)$; moreover, $\normsymb_{wc}(w)$
equals the smallest number of conjugates of elements in $X \sqcup X^{-1}$
whose product is $w$. Now define $d_{FG}(w,w') \coloneqq
\normsymb_{wc}(w^{-1} w')$. It is easy to see that $\normsymb_{wc}$ is a
conjugacy invariant length function.
We claim that $d_{FG} \equiv d_{FM}$ on $\mathrm{FMon}(X)$, which proves
the result. It is easy to show that if two words in $\mathrm{FMon}(X)$
differ by a single insertion or deletion, then their distance in
$\mathrm{FGp}(X)$ is at most one, hence exactly one. In the other
direction, we claim that a non-crossing matching on $w^{-1}w'$, with $w$
and $w'$ containing only positive generators (in $X$), is just a
`rainbow', i.e.~nested arches with one end in $w^{-1}$ and the other in
$w'$. But then $d_{FG}(w,w')$ equals $\normsymb(w) + \normsymb(w')$ minus
twice the length of a common substring, which is maximal by the
minimality of the deficiency. Hence $d_{FG}(w,w') = d_{FM}(w,w')$,
completing the proof.
\end{proof}
{ Note that given weights $\normsymb(a)$ and $\normsymb(b)$,
there is a natural weighted version $\normsymb_{wc;a,b}$ where the
letters of $U$ as above are taken with these weights (symmetrically under
inversion). This corresponds to the weighted edit distance, with
different costs for editing different letters.}
\subsection{ Quasimorphisms and commutator lengths}
{ We now investigate potential applications of Proposition \ref{main} with $c > 0$.}
A \emph{quasi-morphism} on a group $G$ is a map $f : G \to \R$
whose \emph{defect} is bounded,
\[
D(f) \coloneqq \sup_{x,y\in G} | f(xy) - f(x) - f(y) | < +\infty.
\]
{ Every quasi-morphism induces a pseudo-length function {(in particular semi-length function)}
by setting
\begin{equation}\label{quasimortonorm}
\norm{ x } \coloneqq | f(x) | { + D(f),}
\end{equation}
where we can take $c = 2 D(f)$ as a bound on the homogeneity defect. In this case, Proposition~\ref{main} makes a rather
trivial statement: a homogeneous quasi-morphism is bounded on
commutators,
\[
| f([x,y]) | \leq { 10} D(f).
\]
In fact, as observed in \cite[Lemme~1.1]{Br}, for homogeneous quasi-morphisms one can improve the
constant from $10$ to $3$, and a quasi-morphism can always be homogenized by replacing it by $\lim_{n \to \infty} f( x^n) / n$~\cite[p.~135]{Br}, which differs from the original $f$ by at most $D(f)$.}
Nevertheless, quasi-morphisms can be utilized to construct interesting
{ pseudo-length functions}, for example satisfying homogeneity on
specific commutators. The following quasi-morphism is due to
Brooks~\cite[Section~2]{Fu}.
For a given word $w$ in the free group $\free{2}$, written in reduced
form, let $f_w : \free{2} \to \R$ be the function which assigns to every
other $g\in \free{2}$, also written in reduced form, the maximum number
of times such that $w$ occurs in $g$ without overlaps, minus the
analogous number of times that $w^{-1}$ can maximally occur in $g$. Since
$f_w(w^n) = n f_w(w)$, { using~\eqref{quasimortonorm} results in a
pseudo-length function that grows linearly on the powers of $w$}. For example with $w$ being the
commutator of the generators of $\free{2}$, we see that although the
{ pseudo-length} function must be bounded on commutators by
{ Proposition~\ref{main}}, it can nevertheless grow linearly
on the powers of a fixed commutator.
Thus, there exist { examples of quasi-homogeneous semi-length functions on free groups} that are not induced by
norms. Nevertheless, we { will now} see that for a large class of groups, including
amenable groups and $G = SL(n, \Z)$ for $n\geq 3$, { even all quasi-homogeneous { semi-}length} functions are induced by norms on Banach spaces.
Further, the bound from { Proposition~\ref{main}} even in
the case of free groups is sharper than that obtained without using
homogeneity.
Recall that the \emph{commutator length} $\mathrm{cl}(g)$ of a word in
$[G, G]$ is the length $k$ of the shortest expression $g = [a_1,
b_1]\cdot[a_2, b_2]\cdots[a_k, b_k]$ of $g$ as a product of commutators.
The \emph{stable commutator length} is defined as $\lim_{n\to\infty}
\mathrm{cl}(g^n)/n$, where the limit exists by sub-additivity of the
function $n\mapsto \mathrm{cl}(g^n)$.
Then { Proposition~\ref{main}}, together with {$\norm{e} \leq c$ and
\eqref{rearrange} for $n$ a power of two,
\[
\norm{x} \leq \frac{\norm{x^n}}{n} + \log_2(n)\, c,
\]}
easily imply the following estimates:
\begin{proposition}\label{commutatorbound}
{
Let $\normsymb$ and $c$ be as in
Proposition~\ref{main}. Then for
$x\in [G, G]$, $\norm{x}\leq (5\, \mathrm{cl}(x) + 1)c$
and $\norm{x}\leq (5\, \mathrm{scl}(x) + 1)c$.
}
\end{proposition}
We say two { { semi-}length} functions $\normsymb_1,\normsymb_2: G
\to\R$ are \emph{equivalent} if $|\normsymb_1(x) - \normsymb_2(x)|$ is bounded in $x\in G$.
For a group $G$ which is perfect and so that the stable commutator length
vanishes on $G=[G, G]$, for example $SL(n,\Z)$ for $n \geq 3$, it is
immediate that any {homogeneous semi-}length function is bounded,
and hence equivalent to the trivial { semi-}length function
$\normsymb(g)\equiv 0$.
More generally, for groups $G$ for which the stable commutator length
vanishes on $[G,G]$, we can deduce an analogue of Theorem~\ref{class}.
Note that there are several interesting examples of such groups,
including solvable groups, and more generally, amenable groups.
\begin{theorem}\label{quasi}
Let $G$ be a group such that the stable commutator length vanishes on
$[G, G]$ and assume $\normsymb: G \to \R$ satisfies \eqref{ng} and \eqref{double}.
Then there exist a real Banach space $\mathbb{B} = (\mathbb{B},\| \, \|)$
and a group homomorphism $\phi : G \to \mathbb{B}$ such that $\normsymb$
is equivalent to $x\mapsto \| \phi(x) \|$.
\end{theorem}
\begin{remark}
{As in Remark~\ref{quasilength}, we can replace \eqref{ng} by the \emph{a priori}
weaker condition that $\norm{xy} \leq \norm{x} + \norm{y} + k$ for all $x, y\in G$ with $k$ fixed.}
\end{remark}
\begin{proof}
Let $\mathrm{ab}: G \to G_{\mathrm{ab}}= G / [G, G]$ be the abelianization homomorphism.
We first construct a homogeneous { semi-}length function $\bar{\normsymb}$
on { $G_{\mathrm{ab}}$} so that $\normsymb$ is equivalent to $\bar{\normsymb}\circ\mathrm{ab}$.
Let $\eta : G_{\mathrm{ab}} \to G$ be a section of $\mathrm{ab}$ { and let $\bar{\normsymb}_0(x) := \norm{\eta(x)} + c$}. We show that
$\bar{\normsymb}_0$ is a { { semi-}length} function. The required
$\bar{\normsymb}$ will be obtained by homogenizing $\bar{\normsymb}_0$.
By Proposition~\ref{commutatorbound}, as the stable commutator length
vanishes on $[G, G]$, it follows that for $x, y\in G$, if $\mathrm{ab}(x)
= \mathrm{ab}(y)$, then $|\norm{x} - \norm{y}| \leq c$. Now, for
$\alpha, \beta \in G_{\mathrm{ab}}$, $\mathrm{ab}(\eta(\alpha\beta)) =
\mathrm{ab}(\eta(\alpha)\eta(\beta))$, hence
\[
|\norm{\eta(\alpha\beta)} - \norm{\eta(\alpha)\eta(\beta))}| \leq c.
\]
This together with { the triangle inequality~\eqref{ng} gives
\[
\bar{\normsymb}_0(\alpha\beta) \leq \bar{\normsymb}_0(\alpha) + \bar{\normsymb}_0(\beta) + c,
\]
while using \eqref{double} instead} gives the required lower bound for $\bar{\normsymb}_0(\alpha^2)$.
Next, for $x\in G$, as $\mathrm{ab}(\eta(\mathrm{ab}(x))) = \mathrm{ab}(x)$, we have
{ $|\normsymb(x) - (\bar{\normsymb}_0 \circ \mathrm{ab})(x)| \leq c$}. Thus
$\normsymb$ is equivalent to $\bar{\normsymb}_0 \circ \mathrm{ab}$.
Since $(\alpha \beta)^n = \alpha^n \beta^n$ in $G_{\mathrm{ab}}$, we also have
$\bar{\normsymb}_0((\alpha\beta)^n) \leq \bar{\normsymb}_0(\alpha^n) +
\bar{\normsymb}_0(\beta^n) + { c}$. We deduce that the homogenization
$\bar{\normsymb}$ of $\bar{\normsymb}_0$ is a { semi-}length
function on $G_{\mathrm{ab}}$, which is equivalent to $\bar{\normsymb}_0$ { due to the bounds~\eqref{ng-iter} and~\eqref{rearrange}, applied to $\bar{\normsymb}_0$}. Therefore
also $\normsymb$ is equivalent to $\bar{\normsymb}\circ \mathrm{ab}$ on $G$.
The claim now follows upon applying Theorem~\ref{class} to
$(G_{\mathrm{ab}},\bar{\normsymb})$ and taking $\phi$ to be the composition $G
\to G_{\mathrm{ab}} \to \mathbb{B}$.
\end{proof}
The following examples of length functions on the free group show
that some hypotheses are needed to get bounds as strong as those of the
Theorem (naturally the stable commutator length does not vanish in the
free group).
For example, consider the word $[a^k, b^m]$ in the free group $\free{2}$,
generated by $a$ and $b$, for some integers $k$ and $m$.
\begin{itemize}
\item The norm of such an element with respect to the word metric is
$2(|k| + |m|)$.
\item If we have a length function $\normsymb$ which is symmetric and
conjugation-invariant, but not necessarily homogeneous, then we have the
bound $\norm{[a^k, b^m]} \leq 2 \min (|k|\, \norm{a}, |m|\, \norm{b})$.
Furthermore, { the $\normsymb_{wc;a,b}$ from above} are conjugation-invariant length
functions for which these inequalities hold with equality.
Further, $\norm{[a^k, b^m]} \geq 2 \min (|k|\, \norm{a}, |m|\, \norm{b})$
as, for any matching $M$ for $w=[a^k, b^m]$, if some pair $(i, j)$
corresponds to letters $a$ and $a^{-1}$, then no pair corresponds to
letters $b$ and $b^{-1}$ and conversely. Further, it is easy to find a
matching for $w$ for which the deficiency is $\min (|k|\, \norm{a}, |m|\,
\norm{b})$. On the other hand, $\normsymb_{wc;a,b}$ is not homogeneous; for
instance, $\norm{[a,b]} = 2$ and $\norm{[a,b]^3} = 4$. Similarly, we
have $\norm{[a^k,b^k]}=2|k|$ and $\norm{[a^k,b^k]^3} \leq 4|k|$, which
demonstrates that $2\norm{x} - \norm{x^2}$ is unbounded (as must be the
case, according to \eqref{9bound}).
\item On the other hand, the function $\normsymb_{cyc}$ associating to each
word the length of its cyclically reduced form is homogeneous, but not a
{ { semi-}length function}. For this we have $\normsymb_{cyc}([a^k, b^m]) = 2(|k| +
|m|)$.
\end{itemize}
Observe that all of the bounds on $\norm{[a^k, b^m]}$ here become
unbounded as $k,m \to \infty$. This should be compared with
{ Proposition~\ref{main}, which establishes a bound
$\norm{[a^k,b^m]} \leq 5c$} that is uniform in $k$ and $m$ for any
function $\normsymb$ satisfying the hypotheses of that proposition.
\subsection{Finite balls in free groups}
From Proposition~\ref{main} and a standard compactness argument, we can
establish the following local version of the theorem.
\begin{theorem}
For any $\eps>0$ there exists $R \geq 4$ with the following property: if
$a,b$ are two elements of a group $G$, $B_{a,b}(R) \subset G$ is the
collection of all words in $a,b,a^{-1},b^{-1}$ of length at most $R$
\textup{$($}so in particular $B_{a,b}(R)$ contains $[a,b]$\textup{$)$},
and the map $\normsymb: B_{a,b}(R) \to [0,+\infty)$ is a ``local
{ semi}-length function'' which obeys the triangle inequality
\begin{equation}\label{tri}
\norm{ xy } \leq \norm{x} + \norm{y}
\end{equation}
whenever $x,y,xy \in B_{a,b}(R)$, with equality when $x=y$, then one has
\[
\norm{ [a,b] } \leq \eps ( \norm{a} + \norm{b} ).
\]
\end{theorem}
\begin{proof}
By pulling back to the free group $\free{2}$ generated by $a$ and $b$, we
may assume without loss of generality that $G = \free{2}$. Without loss
of generality we may also normalize $\norm{a}+\norm{b}=1$. If the claim
failed, then one could find a sequence $R_n \to \infty$ and local
pseudo-length functions $\normsymb_n: B_{a,b}(R_n) \to [0,+\infty)$ such
that $\normn{a}+\normn{b} = 1$, but that $\normn{[a,b]} \geq \eps$. By
the Arzela--Ascoli theorem, we can pass to a subsequence that converges
pointwise to a homogeneous pseudo-length function $\normsymb: G \to
[0,+\infty)$ such that $\norm{[a,b]} \geq \eps$, which contradicts
Proposition~\ref{main}.
\end{proof}
\begin{remark}
By carefully refining the arguments in the previous section, choosing $n$
to be various small powers of $R$ instead of sending $n$ to infinity, one
can extract an explicit value of $R$ of the form $R = C \eps^{-A}$ for
some absolute constants $C, A > 0$; we leave the details to the
interested reader.
\end{remark}
On the other hand, for any finite $R$ one can construct local length
functions $\normsymb: B(0,R) \to [0,+\infty)$ such that $\norm{x} > 0$
for all $x \in B(0,R) \setminus \{e\}$. One construction is as follows.
Any two matrices $U_a, U_b\in SO(3)$ define a representation $x \mapsto
U_x$ of the free group $\free{2}$ in the obvious fashion. Every $U_x$ is
then a rotation around some axis in $\R^3$ by some angle $0 \leq \theta_x
\leq \pi$ in one of the two directions around that axis; if $U_a$ and
$U_b$ are sufficiently close to the identity, then the angle $\theta_x$
is at most $\pi/2$ for all $x \in B(0,R)$. We set $\norm{x} \coloneqq
\theta_x$ for $x \in B(0,R)$. Also, if $U_a,U_b$ are chosen generically,
the representation is faithful, as follows from the dominance of word
maps on simple Lie groups such as $SO(3)$, see \cite{borel}. Hence
$\norm{x}>0$ for any non-identity $x$. From the triangle inequality for
angles we thus have \eqref{tri} whenever $x,y,xy \in B(0,R)$, with
equality when $x=y$. Note that as one sends $R \to \infty$, the local
length functions constructed here converge to zero pointwise, so in the
limit we do not get any counterexample to the main theorem.
|
\section*{Introduction}
\subsection*{Classical and higher Schur-Weyl duality}
Classical and higher Schur-Weyl dualities are important tools in representation theory. Working over the fixed ground field $\mathbb{C}$, the classical {\it Schur-Weyl duality} for the general linear Lie algebra $\mathfrak{gl}_n$ refers to the double centralizer theorem applied to the commuting actions of $\mathfrak{gl}_n$ and the symmetric group $S_a$
\begin{align}
\mathfrak{gl}_n\;\curvearrowright V^{\otimes a}\;\curvearrowleft S_a \label{SWtriv}
\end{align}
on the tensor product of $a$ copies of the vector representation $V$. By {\it (higher) Schur-Weyl duality} (see \cite{AS}, \cite{BK}) we mean the existence of commuting actions
\begin{align}
\mathfrak{gl}_n\;\curvearrowright M\otimes V^{\otimes a}\;\curvearrowleft H_a \label{SW}
\end{align}
of $\mathfrak{gl}_n$ and the degenerate affine Hecke algebra $H_a$ on the tensor product of an arbitrary $\mathfrak{gl}_n$-representation $M$ with $V^{\otimes a}$. The degenerate affine Hecke algebra $H_a$, introduced by Drinfeld \cite{Drinfeld} and Lusztig \cite{Lusztig}, contains the group algebra $\mathbb{C}[S_a]$ and the polynomial algebra $\mathbb{C}[y_1,\ldots, y_a]$ as subalgebras, and is isomorphic as vector space to $\mathbb{C}[S_a]\otimes \mathbb{C}[y_1,\ldots, y_a]$. In particular it has a basis
\begin{eqnarray*}
\mathcal{B}&=&\{wy_1^{k_1}\cdots y_a^{k_a}\mid w\in S_a, k_i\in \mathbb{N}_0\}.
\end{eqnarray*}
The action of the symmetric group on $M\otimes V^{\otimes a}$ is
given by permuting the tensor factors of $V^{\otimes a}$. To get the action of the polynomial generators $y_i$, one additionally considers the Casimir element
\begin{eqnarray}
\label{Casimir}
\Omega^{\mathfrak{gl}_n}&=&\sum_{1\leq i,j\leq n}E_{ij}\otimes E_{ji}\in \mathfrak{gl}_n\otimes \mathfrak{gl}_n,
\end{eqnarray}
labels the tensor factors of $M\otimes V^{\otimes a}$ as $0,1,\ldots, a$, and then sets
\begin{eqnarray}
\label{yi-action-Hecke}
y_i=\sum_{j=0}^{i-1}\Omega_{ji}^{\mathfrak{gl}_n},
\end{eqnarray}
with $\Omega_{ji}$ denoting the action of $\Omega$ on the $j$-th and $i$-th tensor factors of $M\otimes V^{\otimes a}$.
These operators satisfy $y_{i+1}=s_iy_is_i+ s_i$ for $s_i=(i,i+1)\in S_a$, and define an action of $H_a$. When $M$ is the trivial representation, this action factors through the quotient $H_a\to \mathbb{C}[S_a]$, and (\ref{SW}) reduces to (\ref{SWtriv}). The quotient map $H_a\to \mathbb{C}[S_a]$ sends $y_1,\ldots, y_a$ to the Jucys-Murphy elements of $\mathbb{C}[S_a]$.
The existence of \eqref{SWtriv} and \eqref{SW} allows one to pass knowledge about the representation theory between the two sides of the duality. It is also crucial for the construction and definition of $2$-Kac Moody representations in the sense of Rouquier, \cite{Rouquier}.
\subsection*{Commuting actions for the periplectic Lie superalgebras $\mathfrak{p}(n)$}
We aim to establish a duality analogous to \eqref{SW} in a situation where $\mathfrak{gl}_n$ is replaced by the {\it periplectic Lie superalgebra} $\mathfrak{p}(n)$. The family $\mathfrak{p}(n)$, $n\ge 2$, is the first family of so-called ``strange'' Lie superalgebras in the classification of reductive Lie superalgebras \cite{Kac}. The hope is to use a duality like \eqref{SW} as a tool in understanding the representation theory of $\mathfrak{p}(n)$.
The superalgebra $\mathfrak{p}(n)$ is defined as the subalgebra of the general linear superalgebra $\mathfrak{gl}(n|n)$, consisting of all elements preserving a certain bilinear form $\beta$ on the vector representation $V$ of $\mathfrak{gl}(n|n)$ (see Section \ref{sec:periplectic} for the definition). The duality analogous to \eqref{SWtriv} has been established in \cite{M}, where it was shown that the centralizer algebra $\End_{\mathfrak{p}(n)}(V^{\otimes a})$ is a certain \emph{Brauer superalgebra}, a signed version of the Brauer algebra. One would like to add polynomial generators $y_1,\ldots, y_a$ to the Brauer superalgebra, and define their action on the tensor product $M\otimes V^{\otimes a}$ of an arbitrary $\mathfrak{p}(n)$-representation $M$ with $a$ copies of the vector representation $V$ using an analogue of \eqref{yi-action-Hecke} for some suitably defined element $\Omega\in \mathfrak{p}(n)\otimes \mathfrak{p}(n)$, which centralizes the action of $\mathfrak{p}(n)$ on tensor products. Unfortunately, such an element {$\Omega$ does not exist in $\mathfrak{p}(n)\otimes \mathfrak{p}(n)$.
The main idea is to instead consider a {\it fake} Casimir element (see also \cite{BDEHHILNSS}) $$\Omega=\sum_{b\in \mathcal{X}}{b\otimes b^*}\in \mathfrak{p}(n)\otimes \mathfrak{gl}(n|n).$$
Here $\mathcal{X}$ is a basis of $\mathfrak{p}(n)$, and $\{b^*\mid b\in \mathcal{X}\}$ is the dual basis with respect to the supertrace form on $\mathfrak{gl}(n|n)$. This element does not act on a tensor product $M\otimes N$ of arbitrary $\mathfrak{p}(n)$-representations, but does act on the tensor product $M\otimes V$ of an arbitrary $\mathfrak{p}(n)$-representation $M$ and the vector representation $V$ for $\fgl(n|n)$. A formula analogous to \eqref{yi-action-Hecke} defines the action of commuting elements $y_1,\ldots, y_a$ on $M\otimes V^{\otimes a}$, centralizing the $\mathfrak{p}(n)$ action. We thus obtain, see Proposition~\ref{prop:funct}, commuting actions
\begin{eqnarray}
\label{SWpn}
\mathfrak{p}(n)\;\curvearrowright M\otimes V^{\otimes a}\;\curvearrowleft {\sVW}_a,
\end{eqnarray}
of $\mathfrak{p}(n)$ and a certain {\it affine VW superalgebra} $\sVW_a$. More generally, we establish an action of the affine VW super{\it category} $\sVW$, see Section~\ref{section:sVW}, on the category of modules of the form $M\otimes V^{\otimes a}$ obtained by varying $a$. Our main result (Theorem~\ref{Thm2}) gives an explicit basis of all the morphism spaces in $\sVW$.
The linear independence is proved using the duality \eqref{SWpn} for a specific choice for $M$, namely a Verma module of highest weight $0$. We verify that the PBW filtration on $M$ is compatible with a filtration on the algebras $\sVW_a$, which we build to mimic the filtration by the degree of the polynomials in $\mathbb{C}[y_1,\ldots, y_a]$ in case \eqref{SW}. We explicitly describe the associated graded algebra and deduce the basis theorem from there. As an application we give a description of the centre of all endomorphism algebras involved. The arguments involve the concept of PBW-deformations and (noncommutative) Rees algebras.
\subsection*{Links to other results of this type}
A special feature of the periplectic Lie superalgebras is that $\sVW_a$ are {\it super}algebras, since the involved endomorphism algebra has odd generators. This does not occur in the context of higher Schur-Weyl dualities of the classical Lie superalgebras (see \cite{CW}, \cite{Sergeev} for a general treatment, \cite{BS5}, \cite{ES2}, \cite{LZ} for different cases with $M=\mathbb{C}$, and \cite{BS4}, \cite{DRV}, \cite{ES1}, \cite{RS}, \cite{Sartori} for higher dualities).
The superalgebra $\sVW_a$ is a super (or signed) version of the affine VW algebra, defined in \cite{Nazarov} and studied in \cite{ES1} in the context of higher Schur-Weyl dualities for classical Lie algebras in type $BCD$. In other words, it is a super version of the degenerate BMW algebras, see e.g. \cite{DRV}. This means that, in addition to involving superalgebras, the duality \eqref{SWpn} also has flavours of type $BCD$. In diagrammatic terms, this means working with generalized dotted Brauer diagrams with height moves involving signs.
A basis theorem for the endomorphism algebras of objects in $\sVW$ was obtained independently in \cite{CP} by an algebraic method developed in \cite{Nazarov}, also using the fake Casimir operator. The Brauer superalgebras recently appeared in the literature under the names {\it odd Brauer algebras}, {\it marked Brauer algebras} or {\it periplectic Brauer algebras}, indicating the slightly different points of view on the subject.
Brauer supercategories can be realized as subcategories, as well as quotients, of the VW supercategories. (In terms of representations, this corresponds to taking $M$ to be the trivial representation; they are a super version of the classical Brauer categories as defined e.g.\ in \cite{LZ2}). As a direct consequence of our basis theorem we thus obtain a basis theorem for the Brauer supercategories, hence reprove results from \cite{BE}, \cite{KT} and \cite{M}. Under this quotient, the elements $y_1,\ldots, y_a$ of the superalgebra $\sVW_a$ specialise to Jucys-Murphy elements in the Brauer superalgebras. This allows one to apply the Cherednik \cite{Cherednik} and Okounkov-Vershik \cite{CST}, \cite{OV} approaches in this context. First steps in this direction were already successfully taken in \cite{BDEHHILNSS} and \cite{C} from different perspectives to determine the blocks and decomposition numbers in the category of finite dimensional representations of $\mathfrak{p}(n)$ and of the Brauer superalgebra, and further developed in \cite{EC}. A thorough treatment of the corresponding category $\mathcal{O}$ is so far missing and will be deferred to subsequent work.
\subsection*{The roadmap of the paper}
In Section 2 we define the Brauer supercategory $\sBr$, the VW supercategory $\sVW$, and their endomorphism algebras $\sBr_a$ and $\sVW_a$, and state the main results, Theorems \ref{Thm1} and \ref{Thm2}. In particular, Theorem~\ref{Thm2} gives bases $S_{a,b}^\bullet$ of the endomorphism spaces of $\sVW$. In Section 3 we prove that $S_{a,b}^\bullet$ are spanning sets using a topological argument. In Section 4 we discuss the Lie superalgebra $\mathfrak{p}(n)$ and its representations, the fake Casimir $\Omega$, and prove the existence of the commuting action \eqref{SWpn}. In Section 5 we prove linear independence of the sets $S_{a,b}^\bullet$ by finding large $n$ and large enough $\mathfrak{p}(n)$-representations $M$, so that the set $S_{a,b}^\bullet$ maps into a set of linearly independent operators on $M\otimes V^{\otimes a}$. This proves Theorem~\ref{Thm2}, and Theorem~\ref{Thm1} follows as a corollary. As an application, in Section~\ref{centre} we describe the presentation, the centre, and a certain deformation of the endomorphism algebras $\sVW_a=\End_{\sVW}(a)$.
\textbf{Acknowledgements}. We thank Gwyn Bellamy, Michael Ehrig, Stephen Griffeth, Joanna Meinel, Travis Schedler and Anne Shepler for helpful discussions. This project was started at the WINART workshop in Banff, and was developed and finalised during several visits of some of the authors to the Hausdorff Center of Mathematics (in particular to MPI and HIM) in Bonn. We thank these places for the excellent working conditions.
\addtocontents{toc}{\protect\setcounter{tocdepth}{1}}
\section{Definitions and main results}
\label{sectionone}
In this section we define the Brauer supercategory $\sBr$ and the affine VW supercategory $\sVW$ as monoidal supercategories, and state Theorems \ref{Thm1} and \ref{Thm2}, which give diagrammatic bases for the morphism spaces in these categories.
We fix $\mathbb{C}$ as the ground field for the whole paper.
\subsection{Monoidal supercategories}
We start by recalling some basic facts about monoidal supercategories. For a thorough discussion, see e.g. \cite{BE}.
A \emph{superspace} is a vector space $V$ with a $\mathbb{Z}_2$ grading, $V=V_{\overline{0}}\oplus V_{\overline{1}}$. Homogeneous vectors $v\in V_{\overline{0}}$ are said to be \emph{even} or \emph{of parity $\overline{v}=\overline{0}$}, and $v\in V_{\overline{1}}$ are said to be \emph{odd} or \emph{of parity $\overline{v}=\overline{1}$}. Linear maps between superspaces inherit the grading; homogeneous linear maps are called \emph{even} or \emph{odd}, respectively, depending on whether they preserve or change the parity of homogeneous vectors. Formulas involving parity are usually written for homogeneous elements and extended linearly.
A tensor product of superspaces is again a superspace. For $f,g$ homogeneous linear maps of superspaces, $f\otimes g$ is defined as$$(f\otimes g)(v\otimes w)=(-1)^{\overline{g}\overline{v}} f(v) \otimes g(w)$$
on homogeneous vectors $v \otimes w$.
The following Koszul sign rule holds for compositions
\begin{equation}
\label{signrule}
(f\otimes g)\circ (h\otimes k)= (-1)^{\overline{g}\overline{h}}(f\circ h)\otimes (g\circ k).
\end{equation}
We use the common diagram calculus: the object $a\otimes b$ is depicted by drawing the $b$ to the right of $a$, similar for $f\otimes g$.
A \emph{supercategory} is a category enriched in superspaces; this means all morphism sets are superspaces, and composition preserves parity.
We will be using the usual string calculus for morphisms in strict monoidal supercategories (see e.g. \cite[Definition XI.2.1]{K}). More precisely, we will define strict monoidal supercategories ($\sBr$ and $\sVW$) using generators and relations by
\begin{enumerate} [(i)]
\item specifying a set of generating objects; all objects in the category are obtained as finite tensor products $a_1 \otimes \hdots \otimes a_r$ of generating objects $a_i$ (including the empty tensor product, which is defined to be the unit object $\mathbbm{1}$)
\item specifying a set of generating morphisms; all morphisms in the category are then obtained as linear combinations of finite compositions of \emph{horizontal} (using the tensor product $f\otimes g$) and \emph{vertical} (using the composition $f\circ g$) stackings of compatible generating morphisms and the identity morphisms. Diagrammatically, $f \otimes g$ is presented as placing $f$ to the left of $g$, whereas $f \circ g$ is presented as stacking $f$ on top of $g$; in particular, morphisms are read from bottom to top;
\item specifying a set of generating relations for morphisms; the full set of relations is obtained as the two sided tensor ideal generated by the specified generating relations. Implicitly, we also require the morphisms to respect the sign rule (\ref{signrule}); these are sometimes called the \emph{height moves} in string calculus.
\end{enumerate}
\subsection{The Brauer supercategory $\sBr$}
The \textit{Brauer supercategory} is the $\mathbb{C}$-linear strict monoidal supercategory $\sBr$, generated as a monoidal supercategory by a single object
$\filledstar$ and morphisms
\vspace{-0.5cm}
\begin{eqnarray*}
&s=\TikZ{[scale=.5] \draw (0,0) node{} to (1,1)node{} (1,0) node{} to (0,1)node{} ;}:\;\filledstar\otimes \filledstar\longrightarrow\filledstar\otimes \filledstar,&\nonumber \\
&b=\TikZ{[scale=.5] \draw (1,0) node{} (0,0) node{} (0,0) arc(180:0:0.5) ;}:\;\filledstar\otimes \filledstar\longrightarrow\mathbbm{1},\quad\text{and}\quad
b^*=\TikZ{[scale=.5] \draw (1,1) node{} (0,1) node{} (0,1) arc(-180:0:0.5) ;}:\;\mathbbm{1}\longrightarrow \filledstar\otimes \filledstar ,&
\end{eqnarray*}
with parities $\overline s=\overline 0$, $\overline{b} =\overline{b^*}=\overline 1$, subject to the following defining relations:
\begin{comment}
\begin{eqnarray*}
s^2=1_{\filledstar\otimes\filledstar},&& (\mathbf{1}_{\filledstar}\otimes s )\circ (s\otimes \mathbf{1}_{\filledstar}) \circ (\mathbf{1}_{\filledstar}\otimes s )=
(s\otimes \mathbf{1}_{\filledstar})\circ (\mathbf{1}_{\filledstar}\otimes s ) \circ (s\otimes \mathbf{1}_{\filledstar}),\\
(b\otimes\mathbf{1}_{\filledstar})\circ(\mathbf{1}_{\filledstar}\otimes b^*)=\mathbf{1}_{\filledstar}, &&
(\mathbf{1}_{\filledstar}\otimes b)\circ (b^*\otimes\mathbf{1}_{\filledstar})=-\mathbf{1}_{\filledstar}, \\
b\circ b^*=0, &&(\mathbf{1}_{\filledstar}\otimes s) \circ( b^*\otimes \mathbf{1}_{\filledstar})=(s\otimes \mathbf{1}_{\filledstar})\circ (\mathbf{1}_{\filledstar}\otimes b^*),\\
s\circ b^*=-b^*.
\end{eqnarray*}
\end{comment}
\begin{enumerate}[(R1)]
\item {The {\it braid relations}:} $ \TikZ{[scale=.5] \draw (0,0) node{} to (1,1)node{} to (0,2)node{} (1,0) node{} to (0,1)node{} to (1,2)node{} ;}= \TikZ{[scale=.5] \draw (0,0) node{} to (0,2)node{} (1,0) node{} to (1,2)node{};}$ and $\TikZ{[scale=.5] \draw (0,0) node{} to (1,1)node{} (1,0) node{} to (0,1)node{} (2,0) node{} to (2,1)node{} (0,1) node{} to (0,2)node{} (1,1) node{} to (2,2)node{} (2,1) node{} to (1,2)node{} (0,2) node{} to (1,3)node{} (1,2) node{} to (0,3)node{} (2,2) node{} to (2,3)node{};}=
\TikZ{[scale=.5] \draw (0,0) node{} to (0,1)node{} (1,0) node{} to (2,1)node{} (2,0) node{} to (1,1)node{}
(0,1) node{} to (1,2)node{} (1,1) node{} to (0,2)node{} (2,1) node{} to (2,2)node{}
(0,2) node{} to (0,3)node{} (1,2) node{} to (2,3)node{} (2,2) node{} to (1,3)node{}
;}$,
\item {The {\it snake relations} or {\it adjunctions}:} $ \TikZ{[scale=.5] \draw
(0,0) node{} to (0,1)node{}
(0,1) arc(180:0:0.5)
(1,1) arc(-180:0:0.5)
(2,1)node{} to (2,2)node{};}=
-\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2)node{}
;}$ and $ \TikZ{[scale=.5] \draw
(0,1) node{} to (0,2)node{}
(1,1) arc(180:0:0.5)
(0,1) arc(-180:0:0.5)
(2,0)node{} to (2,1)node{};}=
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2)node{}
;}$,
\item {The {\it untwisting} relations}: $\TikZ{[scale=.5] \draw
(0,1) node{} to (0,2)node{}
(2,0.5) node{} to (2,1)node{}
(1,1) node{} to (2,2)node{}
(2,1) node{} to (1,2)node{}
(0,1) arc(-180:0:0.5)
;}=
\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,1)node{}
(2,1) node{} to (2,2)node{}
(0,1) node{} to (1,2)node{}
(1,1) node{} to (0,2)node{}
(1,1) arc(-180:0:0.5)
;}$ and $ \TikZ{[scale=.5] \draw
(1,1) node{} to (0,2)node{}
(0,1) node{} to (1,2)node{}
(0,1) arc(-180:0:0.5)
;}= -
\TikZ{[scale=.5] \draw
(1,1) node{} to (1,2)node{}
(0,1) node{} to (0,2)node{}
(0,1) arc(-180:0:0.5)
;}$.
\end{enumerate}
The supercategory structure means the {\it height moves} via \eqref{signrule} are also satisfied, e.g.
$\TikZ{[scale=.5] \draw
(0,0) node{} to (0,0.5)node{}
(0,0.5) arc(180:0:0.5)
(1,0) node{} to (1,0.5) node{}
(1.5,0) arc(180:0:0.5);}=b \circ (1 \otimes 1 \otimes b) = b \otimes b =
\TikZ{[scale=.5] \draw
(0,0) arc(180:0:0.5)
(1.5,0) arc(180:0:0.5);} \; , \quad
\TikZ{[scale=.5] \draw
(1.5,0) node{} to (1.5,0.5)node{}
(1.5,0.5) arc(180:0:0.5)
(2.5,0) node{} to (2.5,0.5) node{}
(0,0) arc(180:0:0.5);}=b \circ (b \otimes 1 \otimes 1) =-\; b \otimes b =
-\;\TikZ{[scale=.5] \draw
(0.0,0) arc(180:0:0.5)
(1.5,0) arc(180:0:0.5);}.$
The objects of $\sBr$ are sometimes written as natural numbers $\mathbb{N}_0$, identifying $a\in\mathbb{N}_0$ with $\filledstar^{\otimes a}$, where $\filledstar^{\otimes 0}=\mathbbm{1}$.
A {\it diagram} is a finite composition (horizontally or vertically) of generating morphisms and identity morphisms. It consists of lines, connecting pairs of points among the bottom and top ones, which we call \emph{strings}. Elements of $\Hom_{\sBr}(a,b)$ are linear combinations of diagrams with strings connecting $a$ points at the bottom and $b$ points at the top. We let $1_a\in \Hom_{\sBr}(a,a)$ denote the identity morphism, and let
\begin{eqnarray*}
&b_i=1_{i-1}\otimes b\otimes 1_{a-i+1} \in \Hom_{\sBr}(a+2,a), \quad \quad
b_i^*=1_{i-1}\otimes b^*\otimes 1_{a-i+1} \in \Hom_{\sBr}(a,a+2),\\
&s_i=1_{i-1}\otimes s\otimes 1_{a-i-1} \in \Hom_{\sBr}(a,a)
\end{eqnarray*}
denote the morphisms obtained by applying $b,b^*$ and $s$ on the $i$-th and $(i+1)$-st tensor factors. The supercategory $\sBr$ can alternatively be generated as a supercategory (as opposed to a monoidal supercategory) by vertically stacking compatible $b_i,b_i^*,s_i$.
\subsection{Normal diagrams}
We call a string with both ends at the top of the diagram a \emph{cup}, a string with both ends at the bottom of the diagram a \emph{cap}, a string with one end at the top and one at the bottom a \emph{through string}, and a string with no endpoints a \emph{loop}.
Call a diagram $d\in \Hom_{\sBr}(a,b)$ \emph{normal} if all of the following hold:
\begin{itemize}
\item any two strings intersect at most once;
\item no string intersects itself;
\item no two cups or caps are at the same height;
\item all cups are above all caps;
\item the height of caps decreases when the caps are ordered from left to right with respect to their left ends;
\item the height of cups increases when the caps are ordered from left to right with respect to their left ends.
\end{itemize}
As a consequence, every string in a normal diagram has either one cup, or one cap, or no cups and caps, and there are no closed loops. A diagram with no loops in $\Hom_{\sBr}(a,b)$ has $\frac{a+b}{2}$ strings. In particular, if $a+b$ is odd then this space is zero.
Each normal diagram $d \in \Hom_{\sBr}(a,b)$, where $a,b\in \mathbb{N}_0$, gives rise to a partition $P(d)$ of the set of $a+b$ points into $2$-element subsets given by the endpoints of the strings in $d$. We call such a partition a \emph{connector} and let $\Conn(a,b)$ denote the set of all such connectors; its size is $(a+b-1)!!$. For each connector $c\in \Conn(a,b)$, we pick a normal diagram $d_c\in P^{-1}(c) \subset \Hom_{\sBr}(a,b)$. (Note that different normal diagrams in a single fibre $P^{-1}(c)$ differ only by braid relations, and thus represent the same morphism, see Lemma~\ref{choice of Sab is a choice of sign}.)
\begin{theorem}\label{Thm1}
The set $S_{a,b}=\{ d_c \mid c\in \Conn(a,b)\}$ is a basis of $\Hom_{\sBr}(a,b)$.\end{theorem}
We show that it is a spanning set using topology in Section~\ref{Sect-span}. Linear independence can also be seen directly using topology, since the defining relations of $\sBr$ do not change the underlying connector of a diagram. However, we obtain it using representation theory in Section~\ref{linindep} as a direct consequence of the more general Theorem~\ref{Thm2}. For the special case of $a=b$, this theorem appears as a basis theorem for the algebra $\mathcal{A}_a$ in \cite{M}.
Let us also remark that the above choice of normal diagrams for basis vectors is for convenience only. It is enough to choose one diagram $d'_c$ with no loops in every fibre $P^{-1}(c)$; the set $\{ d'_c \mid c\in \Conn(a,b) \}$ is then also a basis. This choice of basis differs from $S_{a,b}$ by signs only, meaning it is a subset of $\{ \pm d \mid d \in \Conn(a,b)\}$ with exactly one choice of sign for each $d$, see Proposition~\ref{oBr-span}.
\subsection{The affine VW supercategory $\sVW$} \label{section:sVW}
The \emph{affine VW supercategory}, or \emph{affine Nazarov-Wenzl supercategory}, is the $\mathbb{C}$-linear strict monoidal supercategory $\sVW$, generated as a monoidal supercategory by a single object
$\filledstar$, morphisms $s=\TikZ{[scale=.5] \draw (0,0) node{} to (1,1)node{} (1,0) node{} to (0,1)node{} ;}:\;\filledstar\otimes \filledstar\longrightarrow\filledstar\otimes \filledstar$, $b=\TikZ{[scale=.5] \draw (1,0) node{} (0,0) node{} (0,0) arc(180:0:0.5) ;}:\;\filledstar\otimes \filledstar\longrightarrow\mathbbm{1}$ and $b^*=\TikZ{[scale=.5] \draw (1,1) node{} (0,1) node{} (0,1) arc(-180:0:0.5) ;}:\;\mathbbm{1}\longrightarrow \filledstar\otimes \filledstar$ as above, and an additional morphism
\begin{eqnarray*}
&y=\TikZ{[scale=.5] \draw (0,0) node{} to (0,1) node{} (0,0.5) node[fill,circle,inner sep=1.5pt]{};}:\;\filledstar \longrightarrow \filledstar&
\end{eqnarray*}
with $\overline y=0$, subject to relations (R1)-(R3) above, and
\begin{enumerate}
\item[(R4)] {The {\it dot relations}:} $\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2) node{}
(1,0) node{} to (1,2) node{} (1,1) node[fill,circle,inner sep=1.5pt]{}
;} =
\TikZ{[scale=.5] \draw
(0,0) node{} to (1,1) node{} to (0,2) node{}
(1,0) node{} to (0,1) node{} to (1,2) node{} (0,1) node[fill,circle,inner sep=1.5pt]{}
;}+
\TikZ{[scale=.5] \draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{}
;}+
\TikZ{[scale=.5] \draw
(0,0) arc(180:0:0.5) (0,2) arc(-180:0:0.5)
;}$ \textrm{ and }
$\TikZ{[scale=.5] \draw
(0,1) arc(180:0:0.5) (0,0) node{} to (0,1) node{} (1,0) node{} to (1,1) node{} (1,0.5) node[fill,circle,inner sep=1.5pt]{}
;}=
\TikZ{[scale=.5] \draw
(0,1) arc(180:0:0.5) (0,0) node{} to (0,1) node{} (1,0) node{} to (1,1) node{} (0,0.5) node[fill,circle,inner sep=1.5pt]{}
;}+
\TikZ{[scale=.5] \draw
(0,1) arc(180:0:0.5) (0,0) node{} to (0,1) node{} (1,0) node{} to (1,1) node{}
;}.
$
\end{enumerate}
The objects in $\sVW$ can be identified with integers $a\in \mathbb{N}_0$, and the morphisms are linear combinations of dotted diagrams. The category can alternatively be generated by vertically stacking $b_i,b_i^*,s_i$ and $y_i=1_{i-1}\otimes y\otimes 1_{a-i}\in \Hom_{\sVW}(a,a)$. It is a filtered category, in the sense that the spaces $\Hom_{\sVW}(a,b)$ have a filtration with $\Hom_{\sVW}(a,b)^{\le k }$ being the span of all dotted diagrams with at most $k$ dots.
\subsection{Normal dotted diagrams}
Call a dotted diagram $d\in \Hom_{\sVW}(a,b)$ \emph{normal} if:
\begin{itemize}
\item the underlying diagram obtained by erasing the dots is normal;
\item all dots on cups and caps are on the leftmost end, and all dots on the through strings are at the bottom.
\end{itemize}
Let $S^\bullet _{a,b}$ be the set of normal dotted diagrams obtained by taking all diagrams in $S_{a,b}$ and adding dots to them in all possible ways. Let $S^{k} _{a,b}\subset S^\bullet _{a,b}$ and $S^{\le k}_{a,b}=\bigcup _{l=0}^k S^l_{a,b}$ be the sets of such diagrams with exactly $k$ dots, respectively at most $k$ dots. In particular, $S^{0}_{a,b}=S^{\leq 0}_{a,b}=S_{a,b}$. Note that if $a\equiv b \text{ mod } 2$ then the cardinality of $S^k_{a,b}$ is ${\frac{a+b}{2}+k-1 \choose k}\cdot (a+b-1)!!$, and if $ a \not\equiv b \text{ mod } 2$ then the cardinality of $S^k_{a,b}$ is $0$.
The following basis theorem is the main result of this paper.
\begin{theorem}[Basis Theorem]\label{Thm2}
The set $S^{\le k} _{a,b}$ is a basis of $\Hom_{\sVW}(a,b)^{\le k}$, and consequently the set $S^\bullet _{a,b}$ is a basis of $\Hom_{\sVW}(a,b)$.
\end{theorem}
The proof will be given in Sections~\ref{Sect-span} and \ref{linindep}. The identification $S_{a,b}=S^0_{a,b}$ defines an embedding of categories $\sBr \longrightarrow \sVW$ and hence Theorem~\ref{Thm2} directly implies Theorem~\ref{Thm1}.
As an immediate consequence of Theorem~\ref{Thm2} we obtain the following:
\begin{corollary}
The diagrams without dots form a supersubalgebra $\Hom_{\sBr}(a,a)$ of the superalgebra $\Hom_{\sVW}(a,a)$. The dotted diagrams whose underlying undotted diagram is the identity morphism $1_a$ form a polynomial subalgebra $\C[y_1,\ldots,y_a]$, and the subalgebras $\C[y_1,\ldots,y_a]$ and $\Hom_{\sBr}(a,a)$ together generate $\Hom_{\sVW}(a,a)$ as vector superalgebra.
\end{corollary}
\subsection{The affine VW superalgebra $\sVW_a$}
For any $a \in \NN$, the endomorphism space $\sVW_a=\Hom_{\sVW}(a,a)$ has the structure of a superalgebra. It is the signed version of the affine VW algebra (see \cite[Section 2]{ES1} for the setup we use), and the affine version of the Brauer superalgebra $\Hom_{\sBr}(a,a)$. These algebras have an interesting structure, and allow an $\hbar$-deformation. For more details, including a presentation and a description of the centre, see Section \ref{centre}.
One can also define {\it cyclotomic quotients} of the algebras $\sVW_a$ by
mimicking the constructions in \cite{AMR} for affine VW algebras, see also \cite{CP}. We expect Lemma \ref{lem:dot-loop} (stating the vanishing of all loop values) to simplify the necessary admissibility conditions from \cite{AMR} and more explicitly \cite{ES1} drastically, but do not pursue this here.
\section{Spanning sets for $\sBr$ and $\sVW$}\label{Sect-span}
In this section we show that the sets $S_{a,b}$ and $S^{\bullet}_{a,b}$ span the corresponding morphism spaces in the categories $\sBr$ and $\sVW$ (Propositions \ref{oBr-span} and \ref{oVW-span}).
\subsection{Some diagrammatic relations}
First, we establish some additional relations in these categories. Note that these relations are local and hold wherever they are defined within a bigger expression, and we indicate how the local diagram fits into the larger one by specifying the position ($i \in \NN$) of a string (always counted from the left).
The first lemma shows that in $\sBr$ (and consequently in $\sVW$), similar untwisting relations to (R3) hold for caps as they do for cups, and that any isolated loops are zero.
\begin{lemma}[Untwisting relations]
\label{lem:sBRrel} The following relations hold in $\sBr$ and $\sVW$:
\begin{align*}
(a) \quad
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,1)node{}
(2,1) node{} to (2,1.5)node{}
(1,0) node{} to (2,1)node{}
(1,1) node{} to (2,0)node{}
(0,1) arc(180:0:0.5)
;}&=
\TikZ{[scale=.5] \draw
(0,1) node{} to (0,1.5)node{}
(2,0) node{} to (2,1)node{}
(0,0) node{} to (1,1)node{}
(0,1) node{} to (1,0)node{}
(1,1) arc(180:0:0.5)
;} \hspace{0.8cm} &
(b) \quad
\TikZ{[scale=.5] \draw
(1,0) node{} to (0,1)node{}
(0,0) node{} to (1,1)node{}
(0,1) arc(180:0:0.5)
;}&=
\TikZ{[scale=.5] \draw
(1,0) node{} to (1,1)node{}
(0,0) node{} to (0,1)node{}
(0,1) arc(180:0:0.5)
;}\hspace{0.8cm} &
(c) \quad
\TikZ{[scale=.5] \draw
(0,1) arc(0:360:0.5)
;}&=0 \\
b_{i} s_{i+1}&=b_{i+1} s_{i} & b_{i} s_{i}&=b_{i} & b_{i} b_{i}^*&=0
\end{align*}
\end{lemma}
\begin{proof}
\begin{enumerate}[(a)]
\item Using the relations in $\sBr$ and \eqref{signrule}, the morphism $s$ can be rewritten as
$$\TikZ{[scale=.5] \draw
(0,0) node{} to (1,1)node{}
(1,0) node{} to (0,1)node{}
;}\stackrel{(R2)}{=}
-\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,1)node{}
(0,1) arc(180:0:0.5)
(1,1) arc(-180:0:0.5)
(2,1)node{} to (2,1.5)node{}
(2,1.5)node{} to (3,2.5)node{}
(3,1.5)node{} to (2,2.5)node{}
(3,0.5)node{} to (3,1.5)node{}
;}=
-\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,2)node{}
(0,2) arc(180:0:0.5)
(1,2)node{} to (1,1)node{}
(1,1) arc(-180:0:0.5)
(2,2)node{} to (2,2.5)node{}
(2,1)node{} to (3,2)node{}
(3,1)node{} to (2,2)node{}
(3,2)node{} to (3,2.5)node{}
(3,0.5)node{} to (3,1)node{}
;}\stackrel{(R3)}{=}
-\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,2)node{}
(0,2) arc(180:0:0.5)
(2,2.5)node{} to (2,2)node{}
(2,1) arc(-180:0:0.5)
(1,1)node{} to (2,2)node{}
(2,1)node{} to (1,2)node{}
(1,0.5)node{} to (1,1)node{}
(3,1)node{} to (3,2.5)node{}
;}
$$
and therefore
$$
\TikZ{[scale=.5] \draw
(0,0) node{} to (1,1)node{}
(1,0) node{} to (0,1)node{}
(0,1)node{} to (0,1.5)node{}
(1,1) arc(180:0:0.5)
(2,0)node{} to (2,1)node{}
;}=
-\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,2)node{}
(0,2) arc(180:0:0.5)
(2,2)node{} to (2,3)node{}
(2,1) arc(-180:0:0.5)
(1,1)node{} to (2,2)node{}
(2,1)node{} to (1,2)node{}
(1,0.5)node{} to (1,1)node{}
(3,1)node{} to (3,2.5)node{}
(3,2.5) arc(180:0:0.5)
(4,0.5)node{} to (4,2.5)node{}
;}=
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2)node{}
(0,2) arc(180:0:0.5)
(1,2)node{} to (2,1)node{}
(2,1)node{} to (2,0.5)node{}
(2,0.5) arc(-180:0:0.5)
(3,0.5) arc(180:0:0.5)
(4,0.5)node{} to (4,0)node{}
(2,2.5)node{} to (2,2)node{}
(2,2)node{} to (1,1)node{}
(1,1)node{} to (1,0)node{}
;}\stackrel{(R2)}{=}
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,1)node{}
(0,1) arc(180:0:0.5)
(2,1)node{} to (2,1.5)node{}
(1,0)node{} to (2,1)node{}
(2,0)node{} to (1,1)node{}
;}
.$$
\item We use part (a), the relations in $\sBr$ and the Koszul sign rule \eqref{signrule} to show
\begin{align*}\TikZ{[scale=.5] \draw
(0,0) node{} to (1,1)node{}
(1,1) arc(-180:0:-0.5)
(0,1)node{} to (1,0)node{};}&\stackrel{(R2)}{=}
\TikZ{[scale=.5] \draw
(0,-1.5)node{} to (0,-0.5)node{}
(0,-0.5) node{} to (1,0.5)node{}
(1,0.5) arc(-180:0:-0.5)
(0,0.5)node{} to (1,-0.5)node{}
(1,-0.5) node{} to (1,-1)node{}
(1,-1) arc(-180:0:0.5)
(2,-1) arc(180:0:0.5)
(3,-1)node{} to (3,-1.5)node{};}=
-\TikZ{[scale=.5] \draw
(0,-0.5)node{} to (0,0)node{}
(0,0) node{} to (1,1)node{}
(1,1) arc(-180:0:-0.5)
(0,1)node{} to (1,0)node{}
(1,0) arc(-180:0:0.5)
(2,0)node{} to (2,1.5)node{}
(2,1.5) arc(180:0:0.5)
(3,1.5)node{} to (3,-0.5)node{};}\stackrel{(R3)}{=}
-\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,1)node{}
(0,1)node{} to (-1,2)node{}
(-1,2) arc(-180:0:-0.5)
(-2,2)node{} to (-2,1)node{}
(-2,1) arc(-180:0:0.5)
(-1,1)node{} to (0,2)node{}
(0,2)node{} to (0,2.5)node{}
(0,2.5) arc(180:0:0.5)
(1,2.5)node{} to (1,0.5)node{};}=\\
&\stackrel{(a)}{=}
-\TikZ{[scale=.5] \draw
(2,0) node{} to (2,1.5)node{}
(2,1.5) arc(-180:0:-0.5)
(1,1.5)node{} to (0,0.5)node{}
(0,0.5) arc(-180:0:0.5)
(1,0.5)node{} to (0,1.5)node{}
(0,1.5) arc(180:0:1.5)
(3,1.5)node{} to (3,0)node{};}\stackrel{(R3)}{=}
\TikZ{[scale=.5] \draw
(2,0) node{} to (2,1.5)node{}
(2,1.5) arc(-180:0:-0.5)
(1,1.5)node{} to (1,0.5)node{}
(0,0.5) arc(-180:0:0.5)
(0,0.5)node{} to (0,1.5)node{}
(0,1.5) arc(180:0:1.5)
(3,1.5)node{} to (3,0)node{};}\stackrel{(R2)}{=}
\TikZ{[scale=.5] \draw
(0,0)node{} to (0,1)node{}
(0,1) arc(180:0:0.5)
(1,1)node{} to (1,0)node{};}
\end{align*}
\item With $\TikZ{[scale=.5] \draw
(0,0)node{} to (0,1)node{}
(1,0)node{} to (1,1)node{}
(0,1) arc(180:0:0.5)
;}=
\TikZ{[scale=.5] \draw
(0,0)node{} to (1,1)node{} (0,1)node{} to (1,0)node{}
(0,1) arc(180:0:0.5)
;}
$ and $-\TikZ{[scale=.5] \draw
(0,1)node{} to (0,0)node{}
(1,1)node{} to (1,0)node{}
(0,0) arc(-180:0:0.5);}=
\TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5)
(0,1)node{} to (1,2)node{} (1,1)node{} to (0,2)node{}
;}$, we have
$
\TikZ{[scale=.5] \draw
(0,1) arc(360:0:0.5)
;}=
\frac{1}{2} \TikZ{[scale=.5] \draw
(0,1) arc(360:0:0.5)
;}+\frac{1}{2} \TikZ{[scale=.5] \draw
(0,1) arc(360:0:0.5)
;}=
\frac{1}{2} \TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5)
(0,1)node{} to (1,2)node{} (0,2)node{} to (1,1)node{}
(0,2) arc(180:0:0.5)
;}-\frac{1}{2} \TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5)
(0,1)node{} to (1,2)node{} (0,2)node{} to (1,1)node{}
(0,2) arc(180:0:0.5)
;}=0.
$ \qedhere
\end{enumerate}
\end{proof}
The next lemma explains how a dot can be moved within a dotted diagram in $\sVW$. In particular, it can slide through crossings and cups, modulo some diagrams with a smaller number of dots.
\begin{lemma}[Dot sliding relations]\label{lem:dot-slide} The following relations hold in $\sVW$:
\begin{align*}
(a) \;\;\TikZ{[scale=.5] \draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{} (0.75,0.5) node[fill,circle,inner sep=1.5pt]{}
;}&=
\TikZ{[scale=.5] \draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{} (0.25,1.5) node[fill,circle,inner sep=1.5pt]{}
;}+
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2) node{} (1,0) node{} to (1,2) node{} ;}-
\TikZ{[scale=.5] \draw
(0,0) arc(180:0:0.5) (0,2) arc(-180:0:0.5)
;}
&
(b) \;\;\TikZ{[scale=.5] \draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{} (0.25,0.5) node[fill,circle,inner sep=1.5pt]{}
;}&=
\TikZ{[scale=.5] \draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{} (0.75,1.5) node[fill,circle,inner sep=1.5pt]{}
;}-
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2) node{} (1,0) node{} to (1,2) node{} ;}-
\TikZ{[scale=.5] \draw
(0,0) arc(180:0:0.5) (0,2) arc(-180:0:0.5)
;} &
(c) \;\;\TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{} (1,1.5) node[fill,circle,inner sep=1.5pt]{}
;}&=\TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5) (0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{} (0,1.5) node[fill,circle,inner sep=1.5pt]{}
;}-
\TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5) (0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
;}\\
s_iy_{i+1}&=y_is_i+1-b^*_ib_i & s_iy_i&=y_{i+1}s_i-1-b_i^*b_i & y_{i+1}b^*_i&=y_ib^*_i-b^*_i
\end{align*}
\end{lemma}
\begin{proof}
To obtain the relations (a) and (b), we multiply the first relation in (R4) by $s_i$ on the left, respectively on the right, and then use the braid and untwisting relations (R1), (R3) together with Lemma~\ref{lem:sBRrel}{(b)} to simplify.
To prove (c), we compute:
$$\TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5)
(0,1)node{} to (0,2)node{}
(1,1)node{} to (1,2)node{}
(1,1.5) node[fill,circle,inner sep=1.5pt]{}
;}\stackrel{(R2)}{=}
-\TikZ{[scale=.5] \draw
(0,3)node{} to (0,1)node{}
(0,1) arc(-180:0:0.5)
(1,1)node{} to (1,2.5)node{}
(1,1.5) node[fill,circle,inner sep=1.5pt]{}
(1,2.5) arc(180:0:0.5)
(2,2.5) arc(-180:0:0.5)
(3,2.5)node{} to (3,3)node{}
;}=
\TikZ{[scale=.5] \draw
(0,2.5)node{} to (0,1)node{}
(0,1) arc(-180:0:0.5)
(1,1)node{} to (1,2)node{}
(1,1.5) node[fill,circle,inner sep=1.5pt]{}
(1,2) arc(180:0:0.5)
(2,2)node{} to (2,0)node{}
(2,0) arc(-180:0:0.5)
(3,0)node{} to (3,2.5)node{}
;}\stackrel{(R4)}{=}
\TikZ{[scale=.5] \draw
(0,2.5)node{} to (0,1)node{}
(0,1) arc(-180:0:0.5)
(1,1)node{} to (1,2)node{}
(2,1.5) node[fill,circle,inner sep=1.5pt]{}
(1,2) arc(180:0:0.5)
(2,2)node{} to (2,0)node{}
(2,0) arc(-180:0:0.5)
(3,0)node{} to (3,2.5)node{}
;}-
\TikZ{[scale=.5] \draw
(0,2.5)node{} to (0,1)node{}
(0,1) arc(-180:0:0.5)
(1,1)node{} to (1,2)node{}
(1,2) arc(180:0:0.5)
(2,2)node{} to (2,0)node{}
(2,0) arc(-180:0:0.5)
(3,0)node{} to (3,2.5)node{}
;}\stackrel{(R2)}{=}
\TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5)
(0,1) node{} to (0,2) node{}
(1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{}
;}-
\TikZ{[scale=.5] \draw
(0,1) arc(-180:0:0.5)
(0,1) node{} to (0,2) node{}
(1,1) node{} to (1,2) node{}
;}.
\hfill\qedhere$$
\end{proof}
By induction, we obtain formulas for sliding dots along cups or caps:
\begin{lemma}\label{lem:dot-on-cup} The following relations hold in $\sVW$ for any $k\geq 1$.
\begin{align*}
&(a) \quad \TikZ{[scale=.5] \node [right] at (1,0.5) {k};
\draw
(0,1) arc(180:0:0.5) (0,0) node{} to (0,1) node{} (1,0) node{} to (1,1) node{} (1,0.5) node[fill,circle,inner sep=1.5pt]{}
;}=\sum\limits_{j=0}^k{k\choose j}\TikZ{[scale=.5] \node[left] at (0,0.5) {j};
\draw
(0,1) arc(180:0:0.5) (0,0) node{} to (0,1) node{} (1,0) node{} to (1,1) node{} (0,0.5) node[fill,circle,inner sep=1.5pt]{}
;},
&(b) \quad \TikZ{[scale=.5] \node [left] at (0,0.5) {k};
\draw
(0,1) arc(180:0:0.5)
(0,0) node{} to (0,1) node{}
(1,0) node{} to (1,1) node{}
(0,0.5) node[fill,circle,inner sep=1.5pt]{}
;}=\sum\limits_{j=0}^k (-1)^{k+j}{k\choose j}
\TikZ{[scale=.5]
\node[right] at (1,0.5) {j};
\draw
(0,1) arc(180:0:0.5)
(0,0) node{} to (0,1) node{}
(1,0) node{} to (1,1) node{}
(1,0.5) node[fill,circle,inner sep=1.5pt]{}
;}, \\
\vspace{1cm}
&(c) \quad \TikZ{[scale=.5] \node[left] at (0,1.5){k};
\draw
(0,1) arc(-180:0:0.5) (0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{} (0,1.5) node[fill,circle,inner sep=1.5pt]{}
;}=\sum\limits_{j=0}^k{k\choose j}\TikZ{[scale=.5] \node[right] at (1,1.5){j};
\draw
(0,1) arc(-180:0:0.5) (0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{} (1,1.5) node[fill,circle,inner sep=1.5pt]{}
;},
&(d) \quad \TikZ{[scale=.5]
\node[right] at (1,1.5){k};
\draw
(0,1) arc(-180:0:0.5)
(0,1) node{} to (0,2) node{}
(1,1) node{} to (1,2) node{}
(1,1.5) node[fill,circle,inner sep=1.5pt]{}
;}=\sum\limits_{j=0}^k (-1)^{k+j}{k\choose j}
\TikZ{[scale=.5]
\node[left] at (0,1.5){j};
\draw
(0,1) arc(-180:0:0.5)
(0,1) node{} to (0,2) node{}
(1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{}
;},
\end{align*}
where the integers attached to the dots indicate the number of dots on the strand.
\end{lemma}
The following formulas for sliding dots through a crossing can also be verified in a straightforward way using induction, and should be compared with \cite[Lemma 2.3]{AMR}.
\begin{lemma}[Generalized dot sliding]\label{AMR2.3} For any $k\in\mathbb{Z}_{\geq0}$ we have the following relations:
\begin{align*}
(a) \;\;\TikZ{[scale=.5]
\node[right] at (0.75,0.5){k};
\draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{} (0.75,0.5) node[fill,circle,inner sep=1.5pt]{}
;}&=
\TikZ{[scale=.5]
\node[left] at (0.25,1.5){k};
\draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{}
(0.25,1.5) node[fill,circle,inner sep=1.5pt]{}
;}+\sum^{k-1}_{j=0}\left(
\TikZ{[scale=.5]
\node[left] at (0,1.5){k-1-j};
\node[right] at (1,0.5){j};
\draw
(0,0) node{} to (0,2) node{} (1,0) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{}
(1,0.5) node[fill,circle,inner sep=1.5pt]{} ;}-
\TikZ{[scale=.5]
\node[left] at (0.12,1.71){k-1-j};
\node[right] at (0.88,0.3){j};
\draw
(0,0) arc(180:0:0.5) (0,2) arc(-180:0:0.5)
(0.12,1.71) node[fill,circle,inner sep=1.5pt]{}
(0.88,0.3) node[fill,circle,inner sep=1.5pt]{}
;}\right) \\
(b) \;\;\TikZ{[scale=.5]
\node[left] at (0.25,0.5){k};
\draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{}
(0.25,0.5) node[fill,circle,inner sep=1.5pt]{}
;}&=
\TikZ{[scale=.5]
\node[right] at (0.75,1.5){k};
\draw
(0,0) node{} to (1,2) node{} (1,0) node{} to (0,2) node{}
(0.75,1.5) node[fill,circle,inner sep=1.5pt]{}
;}-\sum^{k-1}_{j=0}\left(
\TikZ{[scale=.5]
\node[left] at (0,0.5){k-1-j};
\node[right] at (1,1.5){j};
\draw
(0,0) node{} to (0,2) node{} (1,0) node{} to (1,2) node{}
(0,0.5) node[fill,circle,inner sep=1.5pt]{}
(1,1.5) node[fill,circle,inner sep=1.5pt]{} ;}+
\TikZ{[scale=.5]
\node[left] at (0.12,0.3){k-1-j};
\node[right] at (0.88,1.71){j};
\draw
(0,0) arc(180:0:0.5) (0,2) arc(-180:0:0.5)
(0.12,0.3) node[fill,circle,inner sep=1.5pt]{}
(0.88,1.71) node[fill,circle,inner sep=1.5pt]{}
;}\right)\end{align*}
\end{lemma}
Furthermore, as we show next, as a generalization of Lemma~\ref{lem:sBRrel}{(c)}, isolated loops in $\sVW$ with any number of dots are zero.
\begin{lemma}[Loop values] \label{lem:dot-loop}
For any $k,\ell\in \mathbb{N}_0$, the following relation holds in $\sVW$:
\begin{equation*}
\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\node [right] at (1,1.5) {$\ell$};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{} (1,1.5) node[fill,circle,inner sep=1.5pt]{}
;}=0, \quad \text{that is,} \quad b_i y_i^k y_{i+1}^{\ell}b_i^*=0 \text{ for any } i \geq 1.
\end{equation*}
\end{lemma}
\begin{proof}
Using Relation (R4)
to consecutively slide dots from the right side of the loop to the left, any loop with dots as above can be written as a linear combination of loops with dots on the left only. Hence, without loss of generality, we can assume $\ell=0$.
Applying Relation (R4) and Lemma~\ref{lem:dot-slide}(c), we can rewrite a loop with $k+1$ dots on the left in two different ways (where the integers always indicate the number of dots on the strand):
$$
\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\node [right] at (1,1.5) {};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{} (1,1.5) node[fill,circle,inner sep=1.5pt]{}
;}+\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{}
;}=
\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\node [left] at (0,1) {};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{} (0,1) node[fill,circle,inner sep=1.5pt]{} (0,1.5) node[V]{}
;}
=
\TikZ{[scale=.5]
\node [left] at (0,1.5) {k+1};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{}
;}=
\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\node [left] at (0,2) {};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{} (0,1.5) node[fill,circle,inner sep=1.5pt]{} (0,2) node[fill,circle,inner sep=1.5pt]{}
;}
=\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\node [right] at (1,1.5) {};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{} (1,1.5) node[fill,circle,inner sep=1.5pt]{}
;}-\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{}
;}.$$
Subtracting $
\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\node [right] at (1,1.5) {};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{} (1,1.5) node[fill,circle,inner sep=1.5pt]{}
;}$ from both sides, we get $2\left(\TikZ{[scale=.5]
\node [left] at (0,1.5) {k};
\draw
(0,1) arc(-180:0:0.5) (0,2) arc(180:0:0.5)
(0,1) node{} to (0,2) node{} (1,1) node{} to (1,2) node{}
(0,1.5) node[fill,circle,inner sep=1.5pt]{}
;}\right)=0$.
\end{proof}
\begin{example} Lemma~\ref{lem:dot-loop} shows that all {\it isolated} loops, i.e. those which do not intersect any other strands, with or without dots are equal to zero. This does not mean that all dotted diagrams involving (non-isolated) loops are equal to zero, as the following example shows.
\begin{align*}
d=
\TikZ{[scale=.5]
\draw
(0,1) arc(-180:0:0.5) (0,4) arc(180:0:0.5)
(0,1) node{} to (0,4) node{}
(1,1) node{} to (2,2) node{} to (2,3) node{} to (1,4) node{}
(2,0.5) node{} to (2,1) node{} to (1,2) node{} to (1,3) node{} to (2,4) node{} to (2,4.5) node{}
(1,2.5) node[fill,circle,inner sep=1.5pt]{}
;}&=
\TikZ{[scale=.5]
\draw
(0,1) arc(-180:0:0.5) (0,4) arc(180:0:0.5)
(0,1) node{} to (0,4) node{}
(1,1) node{} to (2,2) node{} to (2,3) node{} to (1,4) node{}
(2,0.5) node{} to (2,1) node{} to (1,2) node{} to (1,3) node{} to (2,4) node{} to (2,4.5) node{}
(1.8,3.8) node[fill,circle,inner sep=1.5pt]{}
;}-
\TikZ{[scale=.5]
\draw
(0,1) arc(-180:0:0.5) (0,4) arc(180:0:0.5)
(0,1) node{} to (0,4) node{}
(1,1) node{} to (2,2) node{} to (2,4.5) node{}
(2,0.5) node{} to (2,1) node{} to (1,2) node{} to (1,4) node{}
;}-
\TikZ{[scale=.5]
\draw
(0,1) arc(-180:0:0.5) (0,4) arc(180:0:0.5) (1,4) arc(-180:0:0.5)
(0,1) node{} to (0,4) node{}
(1,1) node{} to (2,2) node{}
(1,2) arc(180:0:0.5)
(2,0.5) node{} to (2,1) node{} to (1,2) node{}
(2,4) node{} to (2,4.5) node{}
;}=
\TikZ{[scale=.5]
\draw
(0,1) arc(-180:0:0.5) (0,2.5) arc(180:0:0.5)
(0,1) node{} to (0,2.5) node{}
(1,1) node{} to (1,2.5) node{}
(2,0.5) node{} to (2, 3) node{}
(2,1.75) node[fill,circle,inner sep=1.5pt]{}
;}-
\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,1) node{} to (1,2) node{} to (1,2.5) node{}
(0,2.5) arc(180:0:0.5)
(2,1) node{} to (2,3)node{}
(1,1) node{} to (0,2) node{} to (0,2.5)node{}
(1,1) arc(-180:0:0.5)
;}-
\TikZ{[scale=.5]
\draw
(0,1) arc(-180:0:0.5) (0,2.5) arc(180:0:0.5) (1,2.5) arc(-180:0:0.5) (1,1) arc(180:0:0.5)
(0,1) node{} to (0,2.5) node{}
(2,0.5) node{} to (2,1) node{}
(2,2.5) node{} to (2,3) node{}
;} \\
&=
\TikZ{[scale=.5]
\draw
(0,1) arc(-180:0:0.5) (0,2.5) arc(180:0:0.5)
(0,1) node{} to (0,2.5) node{}
(1,1) node{} to (1,2.5) node{}
(2,0.5) node{} to (2, 3) node{}
(2,1.75) node[fill,circle,inner sep=1.5pt]{}
;}-
\TikZ{[scale=.5] \draw
(0,0.5) node{} to (0,2.5) node{}
(0,2.5) arc(180:0:0.5)
(2,1) node{} to (2,3)node{}
(1,1) node{} to (1,2.5) node{}
(1,1) arc(-180:0:0.5)
;}+
\TikZ{[scale=.5]
\draw
(0,0) node{} to (0,2.5) node{}
;}=
0
+
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2.5) node{}
;}+
\TikZ{[scale=.5]
\draw
(0,0) node{} to (0,2.5) node{};}=
2 \TikZ{[scale=.5]
\draw
(0,0) node{} to (0,2.5) node{};}
\end{align*}
Note that although $d$ has one dot, but above calculation shows that it can be rewritten as a diagram with no dots. This is a general phenomenon - resolving loops in a diagram with $k$ dots will produce a linear combination of diagrams without loops which all have $<k$ dots (see the proof of Proposition~\ref{oVW-span}).
\end{example}
\subsection{Spanning set}
We now prove the first part of Theorems \ref{Thm1} and \ref{Thm2} - namely, that the sets $S_{a,b}$ and $S^{\leq k}_{a,b}$ span $\Hom_{\sBr}(a,b)$ and $\Hom_{\sVW}(a,b)^{\leq k}$, respectively.
\begin{lemma}\label{choice of Sab is a choice of sign}
If $d_1,d_2$ in $\Hom_{\sBr}(a,b)$ are any two normal diagrams with the same connector, $P(d_1)=P(d_2)$, then $d_1=d_2 \in \Hom_{sBr}(a,b)$.
\end{lemma}
\begin{proof}
As they are both normal, the diagrams $d_1$ and $d_2$ differ by at most the order of the crossings, so by braid relations (R1), $d_1=d_2$ in $\sBr$.
\end{proof}
\begin{proposition}\label{oBr-span} Any diagram $d$ in $\sBr$ is either equal to zero (if it has loops) or (if it has no loops) to $\pm d_c\in S_{a,b}$, where $c=P(d)$ is the connector corresponding to $d$. In particular, $S_{a,b}$ spans $\Hom_{\sBr}(a,b)$.
\end{proposition}
\begin{proof}
If the diagram $d \in \sBr$ has any loops, we can use relations (R1) -- (R3) together with Lemma~\ref{lem:sBRrel} to isolate the loops to one side, which shows $d=0$.
If the diagram has no loops, we can use relations (R1) -- (R3) and Lemma~\ref{lem:sBRrel} to eliminate any self intersections, double intersections (two strings intersecting twice), and change the height of cups and caps. The resulting normal diagram $d'$ will have the same connector as $d$, $c:=P(d)=P(d')$, and it will differ from $d$ in $\sBr$ by possibly a sign, $d=\pm d'$. It will possibly differ from $d_c\in S_{a,b}$ by the order of the crossings, so by Lemma~\ref{choice of Sab is a choice of sign} it satisfies $d'=d_c$. Thus, $d=\pm d_c$.
\end{proof}
The situation is only slightly more involved for $\sVW$, as transforming a diagram to an element of $S_{a,b}^\bullet$ can produce additional terms with fewer dots, in effect replacing the diagram by
a linear combination of elements of $S_{a,b}^\bullet$. More precisely we have
\begin{proposition}\label{oVW-span} Any dotted diagram $d \in \Hom_{\sVW}(a,b)^{\le k}$ is equal to a linear combination of elements in $S_{a,b}^{\le k}$.
\end{proposition}
\begin{proof}
We argue by induction on $k$, with $k=0$ given by Proposition~\ref{oBr-span}. Assume $k\ge 1$, and let $d$ be a diagram with $k$ dots.
If $d$ contains loops, work with one loop at a time to:
\begin{enumerate}[(i)]
\item slide all the dots on the loop so they are all to the left;
\item slide any dots on other strings away from the loop, so that no dots are in the interior of the loop.
\end{enumerate}
This is accomplished using (R4) and Lemma~\ref{lem:dot-slide}. At each step, we get a linear combination of one diagram with the same number of dots, which are now in a better position, i.e.\ further away from the interior of a loop or more to the left on a loop, and diagrams with fewer dots. Applying the induction assumption to diagrams with fewer dots, it is enough to prove the claim for the diagram with all the dots on loops moved all the way to the left, and no dots in the interior of loops. For such a diagram, any loop can be moved away from the other strings, so by Lemma~\ref{lem:dot-loop} that diagram is equal to zero. This proves the claim for dotted diagrams with loops.
Next, assume that $d$ has no loops. Working with one string at a time,
\begin{enumerate}[(i)]
\item slide all the dots on through strings to the bottom.
\item slide the dots on cups and caps all the way to the left.
\end{enumerate}
Again, this is done using (R4) and Lemma~\ref{lem:dot-slide}. At the end of this process, we have replaced $d$ by a linear combination of a diagram $d'$ with $k$ dots (which are all the way on the bottom of through strings, and on the left of cups and caps), plus diagrams with fewer dots. Apply the induction assumption to diagrams with fewer dots; it remains to prove the claim for $d'$. The position of dots on $d'$ means that it is of the form $\prod_i y^{a_i}_i d'' \prod_j y^{b_j}_j$ for some $a_i,b_j \in \NN_0$ and some undotted diagram $d''\in \sBr$.
Applying Proposition~\ref{oBr-span} to $d''$ completes the proof.
\end{proof}
\subsection{A flipping functor $\iota: \sVW \rightarrow \sVW^{\text{op}}$}
We describe a functor between the supercategory $\sVW$ and its opposite, which on the level of diagrams corresponds to an upside-down flip, with some additional signs.
\begin{proposition}\label{invol} There is an isomorphism of supercategories $\iota: \sVW \rightarrow \sVW^{\text{op}}$, given on objects by the identity and on morphisms by:
$$\iota(s_i)=-s_i, \quad \iota(b_i)=b_i^*, \quad \iota(b_i^*)=-b_i, \quad \iota(y_i)=-y_i.$$
The inverse functor is given by $\iota^3$. It restricts to an anti-isomorphism on each $\operatorname{End}_{\sVW}(a)$, $a\in\mathbb{N}$ (sending $s_i$ , $e_i$, $y_i$ to minus themelves in the notation from Section~\ref{secAbar}).
\end{proposition}
\begin{proof}
To see that $\iota$ respects
the defining relations of $\sVW$, we note that (R1) and the first part of (R4) are invariant under
the diagrams upside-down, the flips of (R3) and the second part of (R4) are a consequence of Lemmas \ref{lem:sBRrel} and \ref{lem:dot-slide}, and the first diagram of (R2) turns into the second after the flip, with the sign changes being consistent as well.
\end{proof}
\section{The periplectic Lie superalgebra $\mathfrak{p}(n)$}\label{sec:periplectic}
We recall some facts from the representation theory of the Lie superalgebra $\mathfrak{p}(n)$. For more details on Lie superalgebras see for instance \cite{Mu}, \cite{Se}, and for $\mathfrak{p}(n)$ see also \cite{BDEHHILNSS}.
\subsection{Definition and bases}
From now on, let $V=\mathbb{C}^{n|n}$ be the superspace of superdimension $n|n$, meaning $V=V_{\overline 0}\oplus V_{\overline 1}$ with $V_{\overline 0}=\C^n$, $V_{\overline 1}=\C^n$. Let $v_1, \ldots , v_n$ be the standard basis of $V_{\overline 0}$ and $v_{1'}, \ldots , v_{n'}$ be the standard basis of $V_{\overline 1}$. We let $[n] := \{1, \dots, n\}$, $[n']:=\{1',\dots,n'\}$ denote the sets of indices.
The \emph{general linear Lie superalgebra} $\fgl(n|n)$ is the Lie superalgebra of endomorphisms of $V$, with $\Z/2\Z-$grading induced by $V$, and the Lie superbracket given by the super commutator $[x,y]=xy-(-1)^{\overline{x}\overline{y}}yx$.
In terms of matrices,
$$\fgl(n|n) = \left\{ \begin{pmatrix} A & B \\ C & D\end{pmatrix} \mid
A, B, C, D \in M_{n,n}(\mathbb{C})\right\},$$
with
$$\fgl(n|n)_{\overline{0}} = \left\{ \begin{pmatrix} A & 0 \\ 0 & D \end{pmatrix} \right\}=\fgl(n) \oplus \fgl(n) \quad \text{and}\quad
\fgl(n|n)_{\overline{1}} = \left\{ \begin{pmatrix} 0 & B \\ C & 0 \end{pmatrix} \right\}.$$
We call $V$ the \emph{vector representation} of $\fgl(n|n)$.
A basis of $\fgl(n|n)$ is given by the matrix units $E_{rs}$ for $r,s \in [n] \cup [n']$, which act on
$V$ as
$E_{rs}v_t = \delta_{st} v_r$ for $t \in [n] \cup [n']$.
Let $\beta: V \otimes V \to \mathbb{C}$ be the bilinear form given by
$$\beta|_{V_{\overline{0}}\otimes V_{\overline{0}}}=\beta|_{V_{\overline{1}}\otimes V_{\overline{1}}}=0 \quad \text{ and } \quad
\beta(v_i,v_{j'})=\beta(v_{j'},v_i)=\delta_{i,j}\hbox{ for all } i,j \in [n].$$
It is symmetric, odd, and non-degenerate on $V$.
Andr\'{e} Weil named such forms \textit{periplectic} by analogy with symplectic forms. The corresponding {\it periplectic Lie superalgebra} $\mathfrak{p}(n)$ is then defined as the Lie supersubalgebra of $\fgl(n|n)$ preserving $\beta$, i.e.\ it is spanned by all homogeneous elements $x$ which satisfy $\beta(x u, v) + (-1)^{\bar{x} \bar{u}} \beta(u, xv) = 0$.
In terms of matrices,
$$\mathfrak{p}(n) = \left\{ \begin{pmatrix} A & B \\ C & -A^t \end{pmatrix} \in \fgl(n|n)
\mid B = B^t, C = -C^t \right\},$$
with
$$\mathfrak{p}(n)_{\overline 0} = \left\{ \begin{pmatrix} A & 0 \\ 0 & -A^t \end{pmatrix} \right\}, \quad \mathfrak{p}(n)_{\overline 1} = \left\{ \begin{pmatrix} 0 & B \\ C & 0 \end{pmatrix} \right\}.$$
\begin{lemma}
The set $\mathcal{X} = \{A_{ij}^- \mid i, j \in [n] \} \cup \{B_{ij}^+ \mid i \leq j \in [n]\} \cup \{C_{ij}^- \mid i < j \in [n]\}$ is a basis for $\mathfrak{p}(n)$, where
$A_{ij}^\pm = E_{ij} \pm E_{j' i'}$,
$B_{ij}^\pm = E_{ij'} \pm E_{j i'}$, $C_{ij}^\pm = E_{i'j} \pm E_{j' i}$, and $\overline{A_{ij}^{\pm}}=0, \; \overline{B_{ij}^{\pm}}=\overline{C_{ij}^{\pm}}=1$.
\end{lemma}
The universal enveloping superalgebra of a Lie superalgebra $\mathfrak{g}$ is the quotient of the tensor algebra $T(\mathfrak{g})$ by the ideal generated by elements of the form $x\otimes y-(-1)^{\overline{x}\overline{y}}y\otimes x - [x,y]$ for all homogeneous $x,y\in \mathfrak{g}$.
Letting
$$\mathfrak{g}=\mathfrak{p}(n), \quad \mathfrak{g}_{-1}=\left\{ \begin{pmatrix} 0 & 0 \\ C & 0 \end{pmatrix} \in \mathfrak{p}(n) \right\},\quad \mathfrak{g}_{0}=\left\{ \begin{pmatrix} A & 0 \\ 0 & -A^t \end{pmatrix} \in \mathfrak{p}(n) \right\}, \quad \mathfrak{g}_{1}=\left\{ \begin{pmatrix} 0 & B \\ 0 & 0 \end{pmatrix} \in \mathfrak{p}(n) \right\},$$
the PBW-Theorem for $\mathfrak{p}(n)$ theorem states that multiplication gives an isomorphism of vector superspaces
$$\Lambda (\mathfrak{g}_{1}) \otimes S(\mathfrak{g}_{0}) \otimes \Lambda (\mathfrak{g}_{-1}) \to \mathcal{U}(\mathfrak{p}(n)).$$
There is a {\it supertrace form} on $\fgl(n|n)$, given by
\begin{equation}\label{eqn:str}\<x,y\> = \str(xy), \quad \text{ with } \quad \str\left(\begin{pmatrix} A&B\\C&D \end{pmatrix}\right) = \tr(A) - \tr(D).\end{equation}
It is bilinear, invariant in the sense $\<[x,y],z\>=\<x,[y,z]\>$ for all $x,y,z \in \fgl(n|n)$, and nongdegenerate. The subalgebra $\mathfrak{p}(n)$ is isotropic with respect to this form; however, one can consider the dual space $\mathfrak{p}(n)^\perp$ of $\mathfrak{p}(n)$ in $\fgl(n|n)$ with respect to this form, which satisfies $\fgl(n|n) = \mathfrak{p}(n) \oplus \mathfrak{p}(n)^\perp$. The basis $\mathcal{X}$ of $\mathfrak{p}(n)$ gives rise to a dual basis $\mathcal{X}^*=\{x^* \mid x \in \mathcal{X}\}$ for $\mathfrak{p}(n)^{\perp}$, in the sense that $\<x^*,y\>=\delta_{xy} \; \; \forall \; y \in \mathcal{X}$. It is explicitly given as $$(A^-_{ij})^* = \frac{1}{2} A_{ji}^+, \quad (B_{ij}^+)^* = -\frac{1}{2} C_{ji}^+, \quad (B_{ii}^+)^* = -\frac{1}{4} C_{ii}^+, \quad \text{ and } \quad (C_{ij}^-)^* = \frac{1}{2} B_{ji}^-.$$
\subsection{The category $\pn$}
We consider the monoidal supercategory $\pn$ of representations of $\mathfrak{p}(n)$ with the set $\Hom_{\mathfrak{p}(n)}(M,N)$ of morphisms from $M$ to $N$ given by linear combinations of homogeneous $\C$-linear maps $f$ from $M$ to $N$ such that $f(x.m)=(-1)^{\overline{x}\overline{f}}x.f(m)$ for homogeneous elements $m\in M$, $x\in \mathfrak{p}(n)$. We in particular allow morphisms to be odd (i.e.\ they change the parity of elements they are applied to).
This supercategory is symmetric, with the braiding given by the superswap
$$\sigma: M\otimes N\to N\otimes M, \qquad \sigma(m\otimes n)=(-1)^{\overline{m}\overline{n}}n\otimes m.$$
We call $V$ the vector representation of $\mathfrak{p}(n)$. The form $\beta$ induces an (odd) identification of $V\to V^*$ as $\mathfrak{p}(n)$-representations, given by
$v\mapsto \beta(v,-).$ Similarly, the bilinear form $(\beta\otimes \beta)\circ (1\otimes \sigma\otimes 1):V^{\otimes 4}\to \C$ induces an identification $(V\otimes V)^*\to V\otimes V$.
With that, the dual map to the form $\beta$ can be thought of as $\beta^*:\mathbb{C}\to V\otimes V$; it is given by
$$\beta^*(1)=\sum_{i}(v_i \otimes v_{i'} - v_{i'}\otimes v_i).$$
\begin{lemma}
\label{lem:homsbb*s}
The following are maps of Lie superalgebra modules of degrees $\overline{1}, \; \overline{1},$ and $\overline{0}$
$$\beta \in \Hom_{\mathfrak{p}(n)}(V\otimes V,\C), \quad
\beta^* \in \Hom_{\mathfrak{p}(n)}(\C,V\otimes V), \quad
\s \in \Hom_{\mathfrak{p}(n)}(V\otimes V,V \otimes V).$$
\end{lemma}
\subsection{A (fake) quadratic Casimir element}\label{Omega}
Because of the absence of the Killing form on $\mathfrak{p}(n)$, there is no Casimir element in $\mathcal{U}(\mathfrak{p}(n))$, nor a quadratic Casimir in $\mathfrak{p}(n)\otimes \mathfrak{p}(n)$. (In fact, the centre of $\mathcal{U}(\mathfrak{p}(n))$ is trivial.) We can however use the supertrace form on $\mathfrak{gl}(n|n) $ to define a \emph{fake Casimir} in $\mathfrak{p}(n)\otimes \mathfrak{gl}(n|n)$ as follows (see also \cite{BDEHHILNSS}). Let
$$\Omega = 2\sum_{x \in \mathcal{X}} x \otimes x^* \in \mathfrak{p}(n) \otimes \fgl(n|n);$$
explicitly,
\begin{equation}
\label{eq:Casimir}
\Omega =
\sum_{i,j} A_{ij}^- \otimes A_{ji}^+
- \frac{1}{2} \sum_{i} B_{ii}^+ \otimes C_{ii}^+
- \sum_{i < j} B_{ij}^+ \otimes C_{ji}^+
+ \sum_{i < j} C_{ij}^- \otimes B_{ji}^-.
\end{equation}
This element does not act on an arbitrary tensor product $M\otimes N$ of $\mathfrak{p}(n)$-representations, but acts on $M \otimes V$, for $M$ any $\mathfrak{p}(n)$-representation, and $V$ the above described vector representation.
Its action gives a morphism in $\pn$ by the following proposition, first observed in \cite[Lemma 4.1.4]{BDEHHILNSS}.
\begin{proposition}\label{prop:Omega-fpn-commute}
The actions of $\Omega$ and $\mathfrak{p}(n)$ on $M \otimes V$ commute, i.e.\
$\Omega\in\operatorname{End}_{\mathfrak{p}(n)}(M\otimes V)$.
\end{proposition}
\begin{proof} The Lie superalgebra $\mathfrak{p}(n)$ acts on $M \otimes V$ via the coproduct $\Delta$ of $\mathcal{U}(\mathfrak{p}(n))$, given by $\Delta(y)=y \otimes 1 + 1 \otimes y$. For any homogeneous element $y \in \mathfrak{p}(n) \subset \fgl(n|n)$, we have
\begin{equation*}[y \otimes 1 + 1 \otimes y,x_i\otimes x^*_i]=[y,x_i]\otimes x^*_i + (-1)^{\bar{y}\bar{x_i}}x_i \otimes [y,x^*_i].\end{equation*}
Furthermore, by expanding in the basis $\{x_i\}_i\cup\{x^*_i\}_i$ of $\fgl(n|n)$, we can see that
$$[y, x_i] = \sum_j \<x_j^*,[y, x_i]\> x_j, \quad \text{ and } \quad
{[y, x_i^*] = \sum_j \<[y, x_i^*], x_j\> x_j^*},$$
Therefore, using the invariance of the supertrace form (\ref{eqn:str}),
\begin{align*}
[\Delta(y), \Omega] &= [ y \otimes 1 + 1 \otimes y, \sum_{i} x_i \otimes x_i^*] = \sum_i [y, x_i] \otimes x_i^* + \sum_i (-1)^{\bar{y} \bar{x}_i} x_i \otimes [y, x_i^*]\\
&= \sum_{i,j} \<x_j^*,[y, x_i] \> (x_j \otimes x_i^*) + \sum_{i,j}(-1)^{\bar{y} \bar{x}_i} \<[y, x_i^*],x_j \> (x_i \otimes x_j^*) \\
&= \sum_{i,j} \<x_j^*,[y, x_i] \> (x_j \otimes x_i^*) - \sum_{i,j} \<[x_i^*,y], x_j\> (x_i \otimes x_j^*) \\
&= \sum_{i,j} \<x_j^*,[y, x_i] \> (x_j \otimes x_i^*) - \sum_{i,j} \<x_i^*, [y,x_j]\> (x_i \otimes x_j^*) = 0.\hfill\qedhere
\end{align*}
\end{proof}
\remark \label{rem:Omegaparity} Note that $\Omega$ is even, $\overline{\Omega}=\overline{0}$, since from (\ref{eq:Casimir}) we see that
$$\Omega \in \left(\fgl(n|n)_{\bar{1}} \otimes \fgl(n|n)_{\bar{1}}\right) \oplus \left(\fgl(n|n)_{\bar{0}} \otimes \fgl(n|n)_{\bar{0}}\right) \subset (\fgl(n|n) \otimes \fgl(n|n))_{\bar{0}}.$$
\vspace{0.3cm}
We consider the special case when $M=V$, and calculate the action of $\Omega$ in that case.
\begin{lemma}\label{lem:OmegaOnVV}
The action of $\Omega$ on $V \otimes V$ is explicitly given by $\s + \beta^* \beta$.
\end{lemma}
\begin{proof}
This is an explicit calculation in the basis $\{v_a \otimes v_b \mid a,b \in [n] \cup [n'] \}$ of $V \otimes V$. We include the computation for the case $a,b\in [n]$. The remaining three cases follow similarly.
Let $a,b\in [n]$. Then
\begin{align*}
(A_{ij}^- \otimes A_{ji}^+)(v_a \otimes v_b)
& = A_{ij}^- v_a \otimes A_{ji}^+ v_b
= \delta_{aj}v_i \otimes \delta_{bi}v_j
= \delta_{aj}\delta_{bi} (v_b \otimes v_a),\\
(B_{ij}^+ \otimes C_{ji}^+) (v_a \otimes v_b)
&= B_{ij}^+ v_a \otimes C_{ji}^+ v_b = 0, \text{ and }\\
(C_{ij}^- \otimes B_{ji}^-) (v_a \otimes v_b)
&= C_{ij}^- v_a \otimes B_{ji}^- v_b = 0,
\end{align*}
and therefore
$\Omega (v_a \otimes v_b)
= \sum_{i, j} \delta_{aj}\delta_{bi} v_b \otimes v_a + 0+ 0 + 0 = v_b \otimes v_a=(\s+\beta^*\beta)(v_a \otimes v_b).$
\end{proof}
\subsection{Jucys-Murphy type elements}\label{JM}
Once we have the above fake Casimir operator, we can define certain commuting elements of $\End_{\mathfrak{p}(n)}(M \otimes V^{\otimes a})$. They are intended to mimic the action of the polynomial generators of the degenerate affine Hecke algebra in case of $\mathfrak{gl}(n)$.
Label the tensor factors of $M\otimes V^{\otimes a}$ by $0, 1,\ldots, a$, and let $\Omega_{ij}$ denote the operator acting as $\Omega$ applied to the $i$th and $j$th factor and the identity everywhere else. For $1\le j \le a$, let
$$Y_j = \sum_{i=0}^{j-1} \Omega_{i j} \; \; \in \; \; \End_{\mathfrak{p}(n)}(M\otimes V^{\otimes a}),$$
(see \cite[Section 4.1]{BDEHHILNSS}) The following result is then standard.
\begin{proposition} The operators $Y_1, Y_2, \dots, Y_a$ pairwise commute.
\end{proposition}
\begin{proof}
Now $\Omega$ commutes with the coproduct $\Delta(y)$, $y \in \mathfrak{p}(n)$, so $\Omega \otimes 1 =Y_1$ commutes with
$$(\Delta \otimes 1)\Omega = \sum_{x \in \mathcal{X}} \Delta(x) \otimes x^* = \sum_{x \in \mathcal{X}}(x \otimes 1 \otimes x^* + 1 \otimes x \otimes x^*)=Y_2.$$
As operators on $M \otimes V \otimes V$, this says that $Y_{1}$ commutes with $Y_2$.
Using $\Delta^j$ to denote the iterated coproduct $\mathfrak{p}(n)\to \mathfrak{p}(n)^{\otimes j}$, by induction we get that
$$Y_j = (\Delta^j \otimes 1) \Omega \quad \text{ commutes with } \quad Y_k = \sum_{i=0}^{k-1} \Omega_{i, k} \text{ for $k<j$},$$
since $\Delta^j(x)$ for $x\in \mathcal{X}$ commutes with $\Omega_{i, k}$ for $i, k < j$. \end{proof}
\remark \label{rem:rel}
There is a quotient map $\sVW \to \sBr$, determined by $y_1\to 0$, $b_i\mapsto b_i$, $b_i^*\mapsto b_i^*$, $s_i\mapsto s_i$. Under this quotient map,
$$y_j\mapsto \sum^{j-1}_{i=1} \left(\hspace{-0.15cm}
\TikZ{[scale=.5]
\draw
(0,0) node{} to (1,1.5) node{}
(1,0) node{} to (0,1.5) node{}
(1.5,0.75) node{+}
(2,0) arc(180:0:0.5) (2,1.5) arc(-180:0:0.5) (2,2) node{};}\right)_{ij}.
$$
These commuting elements of $\sBr_a$ are the analogues of Jucys-Murphy elements for the symmetric group or the Brauer algebra, see \cite{CST} and \cite[Section 2]{Nazarov}. As elements of the superalgebra $\sBr_a$, they were independently defined in \cite[Section 6]{C}, and their eigenvalues are then used, following the approach of \cite{OV}, to study the representation theory of $\sBr_a$ and consequently $\mathfrak{p}(n)$. In terms of the action on $M\otimes V^{\otimes a}$, taking the cyclotomic quotient determined by $y_1\mapsto 0$ corresponds to taking $M$ to be the trivial module (see Lemma~\ref{lem:OmegaOnVV}). This recovers the action of $\sBr$ on $V^{\otimes a}$ from \cite{M}.
\remark \label{rem:rel2}
We have the following relation in $\mathrm{Hom}_{\mathfrak{p}(n)}(M\otimes V^{\otimes a})$, for any $1\leq j<a$, which can be checked directly:
$$\Omega_{i,j+1}=\s_j\;\Omega_{ij}\;\s_j \text{ for } i<j.$$
\subsection{The functor $\Psi^M_n$} \label{sec:funct}
The diagrammatically described supercategory $\sVW$ can be related to $\pn$ and used to study the representation theory of the periplectic Lie superalgebra
Analogous to the notation $\Omega_{ji}$, we will denote by $\s_i, \beta_i$ and $\beta^*_i$ the operators acting as $\s, \beta$ and $\beta^*$ in the $i$th and $(i+1)$st positions of a tensor product $M \otimes V^{\otimes a}$, and identity elsewhere. Here, $M$ is considered as the $0$th factor.
\begin{proposition}
\label{prop:funct}
For any $M \in \pn$, there is a superfunctor $\Psi^M_n: \sVW \longrightarrow \pn$ defined on objects by $a \mapsto M \otimes V^{\otimes a}$ and on morphisms by
$$
s_i \mapsto \sigma_i, \quad\quad
b_i \mapsto \beta_i, \quad\quad
b_i^* \mapsto \beta^*_i, \quad\quad
y_i \mapsto Y_i=\sum_{0\le j <i} \Omega_{ji}.
$$
\end{proposition}
\begin{proof}[Proof of Proposition~\ref{prop:funct}]
From Lemma~\ref{lem:homsbb*s} and Proposition~\ref{prop:Omega-fpn-commute}, we know that $\beta, \beta^*,\s$, and $\Omega$ are morphisms in $\pn$, hence so are the images of $s_i, b_i, b^*_i, y_i$ under $\Psi_n^M$. Furthermore, $\Psi^M_n$ preserves parity, since $\overline{s_i}=\overline{\s_i}=0$, $\overline{b_i}=\overline{\beta_i}=\overline{b^*_i}=\overline{\beta^*_i}=1$, and $\overline{y_i}=\overline{\sum_{0\le j <i} \Omega_{ji}}=0$, see Remark \ref{rem:Omegaparity}. It remains to check that the images of the generating morphisms satisfy the defining relations of $\sVW$. In the calculations we suppress the $0$-th tensor factor $M$.
\begin{enumerate}[(R1)]
\item \begin{enumerate}
\item $\s_i^2=1$.
This follows from
$\s^2(v\otimes w)= (-1)^{\overline{v}\overline{w}}\s(w\otimes v)=(-1)^{2\overline{v}\overline{w}}v\otimes w=v\otimes w$.
\item $ \s_i\s_{i+1}\s_i= \s_{i+1}\s_i\s_{i+1}$.
It is enough to prove this for $i=1, \; a=3$:
\begin{align*}
(\s_{1}\s_2\s_{1})(u\otimes v\otimes w) &=(-1)^{\overline{u}\overline{v}} (\s_{1}\s_2)(v\otimes u\otimes w)=(-1)^{\overline{u}\overline{v}+\overline{u}\overline{w}} \s_{1}(v\otimes w\otimes u) \\
&=(-1)^{\overline{u}\overline{v}+\overline{u}\overline{w}+\overline{v}\overline{w}} w\otimes v\otimes u=(\s_{2}\s_1\s_{2})(u\otimes v\otimes w).
\end{align*}
\end{enumerate}
\item \begin{enumerate}
\item $\beta_i \beta_{i+1}^*=-1$.
It is enough to prove this for $i=1$:
\begin{align*}
\beta_{1} \beta_{2}^*(v)&=(-1)^{\overline v}\beta_{1} (v \otimes \beta^*(1))
=(-1)^{\overline v}\beta_{1} \left(v \otimes
\left( \sum_{i=1}^n v_i\otimes v_{i'} -v_{i'}\otimes v_i
\right)
\right) \\
&=(-1)^{\overline v}\sum_{i=1}^ n (\beta(v,v_i) v_{i'} -\beta(v,v_{i'})v_i)=-v.
\end{align*}
The last equality is easily checked on every $v=v_j$, $j\in [n]\cup[n']$.
\item $\beta_{i+1} \beta_{i}^*=1$.
Similar.
\end{enumerate}
\item \begin{enumerate}
\item $\s_{i+1}\beta_{i}^*=\s_{i}\beta_{i+1}^*$.
It is enough to prove this for $i=1$:
\begin{align*}
\s_{2}\beta_{1}^*(v) &= \s_2\left(\sum_{i=1}^n(v_i\otimes v_{i'} -v_{i'}\otimes v_i )\otimes v\right)= \sum_{i=1}^n ((-1)^{\overline{v}}v_i\otimes v\otimes v_{i'} - v_{i'}\otimes v\otimes v_i),\\
\s_{1}\beta_{2}^*(v)&=\sum_{i=1}^n ((-1)^{\overline v}v_i\otimes v \otimes v_{i'} - (-1)^{\overline v+\overline v}v_{i'}\otimes v \otimes v_i).
\end{align*}
\item
$\s_{i}\beta_{i}^*=-\beta_{i}^*$.
This follows from the fact that $\beta^*(1)$ is skew supersymmetric. Note that this, together with the previous relations, also implies that $\beta_i\s_i=\beta_i$ and $\beta^*_i\beta_i=0$, which will be used in proving (R4)(b).
\end{enumerate}
\item \begin{enumerate}
\item \label{reln-y2=second}
$Y_{i+1}=\s_iY_i\s_i+\s_i+\beta_i^*\beta_i $.
This formula follows via the following computation, using Remarks \ref{rem:rel} and \ref{rem:rel2}, and Lemma~\ref{lem:OmegaOnVV}
\begin{align*}
Y_{i+1}&=\sum_{0\le k <i+1} \Omega_{k,i+1}
=\sum_{0\le k <i} \Omega_{k,i+1}+ \Omega_{i,i+1}
=\sum_{0\le k <i} \s_i\Omega_{k,i}\s_i+ \Omega_{i,i+1} \\
&= \s_i\left(\sum_{0\le k <i} \Omega_{k,i}\right) \s_i+\s_i+\beta_i^*\beta_i =\s_iY_i \s_i+\s_i+\beta_i^*\beta_i
\end{align*}
\item \label{beta(y1-y2)-second}
$\beta_1(Y_1-Y_2)=-\beta_1$.
We have $\beta \circ (x^* \otimes 1 - 1\otimes x^*)=0$ for any $x^* \in \mathfrak{p}(n)^{\perp}$, which can be checked directly on a basis of $V \otimes V$, and hence $\beta_1 \circ (\Omega_{01}-\Omega_{02})=0$. It follows that
$\beta_1(\Omega_{01}-\Omega_{02}-\Omega_{12})=-\beta_1\Omega_{12}=-\beta_1 (\s_1+\beta_1^*\beta_1)=-\beta_1\s_1+0=-\beta_1$.\qedhere
\end{enumerate}
\end{enumerate}
\end{proof}
\section{Linear independence of $S_{a,b}^\bullet$}
\label{linindep}
The purpose of this section is to prove linear independence of the sets $S_{a,b}$ and $S_{a,b}^\bullet$, and thus prove Theorems \ref{Thm1} and \ref{Thm2}. The idea is to exploit a close connection of $\sVW$ and the representation theory of the periplectic Lie superalgebra $\mathfrak{p}(n)$. Namely, as explained in Proposition~\ref{prop:funct}, for every $n$ and every $\mathfrak{p}(n)$-representation $M$, the functor $\Psi_n^M:\sVW \to \pn$ gives a way of interpreting diagrams $d\in \Hom_{\sVW}(a,b)$ as linear $\mathfrak{p}(n)$-homomorphisms $\Psi_n^M(d):M\otimes V^{\otimes a}\to M\otimes V^{\otimes b}$. For given $a,b$, and $k$ in $\mathbb{N}_0$, we will pick $n$ and an appropriate $M\in \pn$ so that the corresponding functor $\Psi_n=\Psi_n^M:\sVW \to \pn$ maps $S_{a,b}^{\le k}$ to a linearly independent set in $\Hom_{\mathfrak{p}(n)}(M\otimes V^{\otimes a}, M\otimes V^{\otimes b})$.
The argument for linear independence is slightly easier in the associated graded setting. For that purpose, we define an auxillary category $\GVW$ and auxilary functors $\Phi_n$, which will turn out to be the associated graded of $\sVW$ and $\Psi_n$. This is analogous to the structure of the main proof in \cite{BCNR}, where a close connection between the affine oriented Brauer category and $\mathcal{W}$-algebras is exploited to construct certain functors, which are then used to prove linear independence. We start with some preliminaries about filtrations and gradings.
\subsection{Graded and filtered supercategories}
An $\NN_0$-filtered superspace is a superspace $U$ with a filtration by subspaces $\{0\}=U^{\leq -1} \subseteq U^{\leq 0} \subseteq U^{\leq 1} \subseteq \hdots \subseteq U$, and $U=\bigcup_{k \geq 0} U^{\leq k}$.
A supercategory $\mathcal{C}$ such that for every $M,N \in \mathcal{C}$, $\Hom_{\mathcal{C}}(M,N)$ has a fixed filtration compatible with composition of morphisms, $\Hom_{\mathcal{C}}(M,N)^{\leq k} \times \Hom_{\mathcal{C}}(N,P)^{\leq \ell} \rightarrow \Hom_{\mathcal{C}}(M,P)^{\leq (k+\ell)}$ is a supercategory $\mathcal{C}$ enriched in the category of filtered superspaces (that is in the category whose objects are filtered superspaces and morphisms are homogeneous linear maps of degree zero). We call such a supercategory a \emph{filtered supercategory}. A \emph{graded supercategory} is a supercategory enriched in graded superspaces; this means its morphism spaces are graded superspaces, and composition is a homogeneous linear map of degree zero.
We say a functor $F:\mathcal{C} \to \mathcal{D}$ between two filtered (respectively, graded) supercategories $\mathcal{C}$ and $\mathcal{D}$ is \emph{filtered} (respectively, \emph{graded}) if it preserves the filtration (respectively, grading) on the morphism spaces.
Now assume we have a filtered supercategory $\mathcal{C}$. Its \emph{associated graded supercategory} $gr\mathcal{C}$ is the graded supercategory with the same objects as $\mathcal{C}$, and morphism spaces the graded superspaces $\Hom_{gr\mathcal{C}}(M,N)=gr (\Hom_{\mathcal{C}}(M,N))= \bigoplus_{k\geq 0}\Hom_{gr\mathcal{C}}(M,N)^k$, where $\Hom_{gr\mathcal{C}}(M,N)^k=\Hom_{\mathcal{C}}(M,N)^{\leq k}/\Hom_{\mathcal{C}}(M,N)^{\leq (k-1)}$.
A filtered functor $F:\mathcal{C} \to \mathcal{D}$ between two filtered supercategories induces a graded functor $gr(F):gr\mathcal{C} \to gr\mathcal{D}$. The functor $gr(F)$ is equal to $F$ on objects, and takes the associated graded map of $F$ on the morphism superspaces.
\subsection{The supercategories $\FVect$ and $\GVect$, and the functor $G$}
Let $\FVect$ be the supercategory with objects $\NN_0$-filtered superspaces, and morphisms given by the filtered superspaces
$\Hom_{\FVect}(M,N)=\bigcup_{k \in \NN_0}\Hom_{\FVect}(M,N)^{\leq k}$, where $\Hom_{\FVect}(M,N)^{\leq k}=\{f:M\rightarrow N \mid f \textrm{ linear, } f(M^{\leq i}) \subseteq N^{\leq (i+k)} \textrm{ for all } i\}$. This is an $\NN_0$-filtered supercategory as above.
Similarly, let $\GVect$ denote the supercategory whose objects are $\NN_0$-graded superspaces, and whose morphisms are superspaces of linear maps equipped with the grading coming from the objects, that is $\Hom_{\GVect}(M,N)=\bigoplus_{k \in \NN_0}\Hom_{\GVect}(M,N)^{k}$, where $\Hom_{\GVect}(M,N)^{k}=\{f:M\rightarrow N \mid f \textrm{ linear, } f(M^{i}) \subseteq N^{(i+k)} \textrm{ for all } i\}$. It is an $\NN_0$-graded supercategory in the above sense.
In particular, we can consider the associated graded category $gr(\FVect)$ described above. (Note that $gr(\FVect)$ and $\GVect$ are not the same categories; objects of $gr(\FVect)$ are filtered while objects of $\GVect$ are graded vector superspaces.)
There is a functor $G:gr(\FVect)\to \GVect$ which associates to a filtered superspace $M=\bigcup_i M^{\leq i}$ its associated graded superspace $G(M)=gr(M)=\bigoplus_i M^{\leq i}/M^{\leq (i-1)}$. On morphisms $G:\Hom_{\FVect}(M,N)^{\leq k}/\Hom_{\FVect}(M,N)^{\leq (k-1)} \to \Hom_{\GVect}(gr(M), gr(N))^k$ is given on $f\in \Hom_{\FVect}(M,N)^{\leq k}$ and $m\in M^{\le i}$
by $$G(f+\Hom_{\FVect}(M,N)^{\leq (k-1)})(m+M^{\le (i-1)})=f(m)+N^{\leq (k+i-1)}.$$
\subsection{$\sVW$ as a filtered supercategory}
The affine VW supercategory $\sVW$ can be viewed as a filtered supercategory, with the filtration on the morphism spaces given by the number of dots. Let $gr(\sVW)$ be its associated graded supercategory, defined as above. In particular, the following relations hold in $gr(\sVW)$:
\begin{equation*} \tag{grR-4}
\begin{aligned}
\TikZ{[scale=.5] \draw
(0,0) node{} to (0,2) node{}
(1,0) node{} to (1,2) node{} (1,1) node[fill,circle,inner sep=1.5pt]{}
;}& =
\TikZ{[scale=.5] \draw
(0,0) node{} to (1,1) node{} to (0,2) node{}
(1,0) node{} to (0,1) node{} to (1,2) node{} (0,1) node[fill,circle,inner sep=1.5pt]{}
;} & \in \Hom_{gr(\sVW)}(2,2)^1=\Hom_{\sVW}(2,2)^{\le 1}/ \Hom_{\sVW}(2,2)^{\le 0},\\
\TikZ{[scale=.5] \draw
(0,1) arc(180:0:0.5) (0,0) node{} to (0,1) node{} (1,0) node{} to (1,1) node{} (1,0.5) node[fill,circle,inner sep=1.5pt]{}
;}&=
\TikZ{[scale=.5] \draw
(0,1) arc(180:0:0.5) (0,0) node{} to (0,1) node{} (1,0) node{} to (1,1) node{} (0,0.5) node[fill,circle,inner sep=1.5pt]{}
;} &\in \Hom_{gr(\sVW)}(2,0)^1=\Hom_{\sVW}(2,0)^{\le 1}/ \Hom_{\sVW}(2,0)^{\le 0}. \nonumber
\end{aligned}
\end{equation*}
It is however not a priori obvious that these, along with (R1)-(R3), are the only defining relations for $gr(\sVW)$. In general, given a filtered algebra or a category, describing its associated graded by generators and relations is a nontrivial problem, and the solution to this problem usually goes most of the way towards proving a basis theorem for the filtered version (as basis theorems for graded versions are usually easier). With that in mind, we define another category $\GVW$ by generators and relations, and prove in Section \ref{final-proof-2} that $gr(\sVW)$ and $\GVW$ are indeed isomorphic as graded supercategories.
\subsection{The category $\GVW$}
Let $\GVW$ be the $\mathbb{C}$-linear monoidal supercategory generated as a monoidal supercategory by a single object $\filledstar$, morphisms $s=\TikZ{[scale=.5] \draw (0,0) node{} to (1,1)node{} (1,0) node{} to (0,1)node{} ;}:\;\filledstar\otimes \filledstar\longrightarrow \filledstar\otimes \filledstar$, $b=\TikZ{[scale=.5] \draw (1,0) node{} (0,0) node{} (0,0) arc(180:0:0.5) ;}:\;\filledstar\otimes \filledstar\longrightarrow\mathbbm{1}$, $b^*=\TikZ{[scale=.5] \draw (1,1) node{} (0,1) node{} (0,1) arc(-180:0:0.5) ;}:\;\mathbbm{1}\longrightarrow \filledstar\otimes \filledstar$ and $
y=\TikZ{[scale=.5] \draw (0,0) node{} to (0,1) node{} (0,0.5) node[fill,circle,inner sep=1.5pt]{};}:\;\filledstar \longrightarrow \filledstar,$ subject to relations (R1)--(R3) and (grR-4). The $\Z/2\Z$ parity is given by $\overline {s}=\overline{y}=0$, $\overline{b}=\overline{b^*}=1$. The $\mathbb{N}_0$-grading is given by $\deg s=\deg b=\deg b^*=0, \deg y=1$. Note that the imposed relations are $\NN_0$-homogeneous and so the category is well-defined. In other words, the objects of $\GVW$ are nonnegative integers, the morphisms are linear combinations of dotted diagrams, and the $\NN_0$-grading is given by the number of dots on the diagram.
The following is analogous to Proposition~\ref{oVW-span}, and proved in exactly the same way.
\begin{lemma}\label{span-GVW}
For any $a,b,k\in \mathbb{N}_0$, the set $S_{a,b}^k$ is a spanning set for $\Hom_{\GVW}(a,b)^k$.
\end{lemma}
\subsection{The functor $\Theta: \GVW \to gr(\sVW)$}\label{Theta}
The tautological assignments $\Theta(\star)=\star$, $\Theta(s)=s$, $\Theta(b)=b$, $\Theta(b^*)=b^*$, $\Theta(y)=y$ define a graded monoidal superfunctor $\Theta: \GVW \to gr(\sVW)$.
It is bijective on objects, and full, i.e.\ surjective on morphisms.
\subsection{The Verma module $M(0)$ and the functor $\Psi_n$}
For $n\in \mathbb{N}$, let $\mathfrak{n}_+$ denote the Lie subalgebra of strictly upper triangular matrices, and $\mathfrak{b}$ the Lie subalgebra of lower triangular matrices in $\mathfrak{gl}(n)$. They can be considered as subalgebras of $\mathfrak{gl}(n)= \mathfrak{g}_0\subseteq \mathfrak{p}(n)$ via the inclusion $E_{ij}\mapsto A^-_{ij}$. Consider $\mathbb{C}$ as the trivial representation of $\mathfrak{b}\oplus \mathfrak{g}_{-1}\subseteq \mathfrak{p}(n)$ by letting $A_{ij}^-$ with $i\ge j$ and $C_{ij}^-$ with $i<j$ act on it by $0$. Consider the $\mathfrak{p}(n)$-module
$M(0)=\mathrm{Ind}_{\mathfrak{b}\oplus \mathfrak{g}_{-1}}^{\mathfrak{p}(n)} \mathbb{C}$, the Verma module of highest weight $0$. Using the PBW theorem we can see that, as a vector superspace, this is
$\mathcal{U}(\mathfrak{p}(n))\otimes_{\mathcal{U}(\mathfrak{b}\oplus \mathfrak{g}_{-1})}\mathbb{C}
\cong \Lambda (\mathfrak{g}_1) \otimes \mathrm{S}(\mathfrak{n}_+).$
Consider the filtration on $M(0)$
coming from the PBW theorem, i.e.\ given by $\deg(B^+_{ij})=\deg(A^-_{ij})=1$. In particular, $M(0) \otimes V^{\otimes a}$ inherits a filtration (by putting $V$ in degree $0$). In this way, $M(0) \otimes V^{\otimes a}$ can be considered, for any $a \in \NN_0$, as an object in $\FVect$.
\begin{lemma}\label{Psi is filtered}
The superfunctor $\Psi^{M(0)}_n:\sVW \to \pn$ induces (by forgetting the action of $\mathfrak{p}(n)$ on the image of $\Psi^{M(0)}_n$)
a filtered superfunctor $\Psi_n: \sVW\to \FVect$.
\end{lemma}
\begin{proof}
The generators $s_i,b_i,b^*_i$ of $\sVW$ have filtered degree $0$, and map under the functor $\Psi_n$ to $\sigma_i, \beta_i, \beta^*_i$ which only act on the $i$-th and $(i+1)$-st tensor factors of $M(0)\otimes V^{\otimes a}$, $1 \leq i \leq a-1$, thus do not change the filtered degree defined on the $0$-th tensor factor $M(0)$.
The generator $y_k$ has filtered degree $1$ in $\sVW$, and its image under $\Psi_n$ is the operator
$$\Psi_n(y_k)=\sum_{i=0}^{k-1}\Omega_{i k}.$$
For $i =1,\ldots, k-1$ the operator $\Omega_{i k}$ does not change the filtered degree.
For $i =0$, the operator $\Omega_{0k}$ acts on $M(0)\otimes V^{\otimes a}$ as
$$\Omega_{0k}=\left( \sum_{i,j} A_{ij}^- \otimes A_{ji}^+
- \frac{1}{2} \sum_{i} B_{ii}^+ \otimes C_{ii}^+
- \sum_{i < j} B_{ij}^+ \otimes C_{ji}^+
+ \sum_{i < j} C_{ij}^- \otimes B_{ji}^-
\right)_{0k}.$$
The summands with $C_{ij}^-$, $i<j$ and $A_{ij}^-$, $i\ge j$ in the $0$-th tensor factor preserve the filtered degree. The summands with $B_{ij}^+$, $i\le j$, and $A_{ij}^-$, $i<j$ in the $0$-th tensor factor increase the filtered degree by $1$. Thus, $\Psi_n(y_k)$ acts by increasing the filtered degree by $1$.
\end{proof}
\subsection{The functor $\Phi_n$}
Next, we define a certain graded superfunctor, which will eventually turn out to be $\gr (\Psi_n)$.
Consider again the vector space $\Lambda (\mathfrak{g}_1) \otimes \mathrm{S}(\mathfrak{n}_+)$, now as a graded superspace with the grading given by $\deg(B_{ij}^+)=\deg(A_{ij}^-)=1$. This gives a grading on $\left( \Lambda (\mathfrak{g}_1) \otimes \mathrm{S}(\mathfrak{n}_+)\right) \otimes V^{\otimes a}$
Define a functor $\Phi_n: \GVW\to \GVect$ on objects by
$\Phi_n(a)=\left(\Lambda (\mathfrak{g}_1) \otimes \mathrm{S}(\mathfrak{n}_+) \right) \otimes V^{\otimes a}$. In the image, we again label $\Lambda (\mathfrak{g}_1) \otimes \mathrm{S}(\mathfrak{n}_+)$ as the $0$-th tensor factor, and $V\otimes \ldots \otimes V$ as factors $1,2,\ldots, a$. With this convention, set
$\Phi_n(s_i)=\sigma_i$, $\Phi_n(b_i)= \beta_i$, $\Phi_n(b_i^*)= \beta_i^*$, and let
$$\Phi_n(y_k)=\left( \sum_{i<j} A_{ij}^- \otimes A_{ji}^+
- \frac{1}{2} \sum_{i} B_{ii}^+ \otimes C_{ii}^+
- \sum_{i < j} B_{ij}^+ \otimes C_{ji}^+ \right)_{0k},$$
with the action of $A_{ij}^-\in \mathfrak{n}_+$ and of $B_{ij}^+\in \mathfrak{g}_1$ on $\Lambda (\mathfrak{g}_1) \otimes \mathrm{S}(\mathfrak{n}_+)$ given by multiplication.
\begin{lemma}\label{Phi is graded}
$\Phi_n:\GVW\to \GVect$ is a well-defined graded superfunctor.
\end{lemma}
\begin{proof}
This is a direct calculation analogous to Proposition ~\ref{prop:funct} and Lemma ~\ref{Psi is filtered}.
\end{proof}
\begin{lemma}\label{square}
With our fixed $n \in \NN$, the following square strictly commutes:
$$
\xymatrix{
\ar @{} [dr] |{=}
\GVW \ar[d]_{\Theta} \ar[r]^{\Phi_n}
& \GVect \\
gr(\sVW) \ar[r]^{gr\Psi_n} & gr(\FVect) \ar[u]_G }$$
That is, $G\circ gr\Psi_n \circ \Theta=\Phi_n$ on all objects and morphisms.
\end{lemma}
\begin{proof}
It clearly strictly commutes on objects, and on the generating morphisms $s_i, b_i,b_i^*$ of degree $0$, so it only remains to check it on $y_k$ of filtered degree $1$. This follows from the proof of Lemma~\ref{Psi is filtered} and from the definition of $\Phi_n$.
\end{proof}
Define a total ordering $\to $ on the set $[n]\cup [n']$ by saying that $i\to j$ if there is a path (of length at least one) from $i$ to $j$ in the graph
\begin{equation}\label{eqn:graph}
1 \to 2 \to \ldots \to n \to n' \to (n-1)' \to \ldots \to 2' \to 1'.
\end{equation}
With this we have the following technical tool:
\begin{lemma}\label{graph}
Let $0\not=m\in M(0)$, $i_1,\ldots, i_a \in [n]\cup [n']$, and $1\le k\le a$ be arbitrary. Then
\begin{align*}
\Phi_n(y_k) (m\otimes v_{i_1}\otimes v_{i_2}\otimes \ldots \otimes v_{i_a})&=
\sum_{i_k\to j} m_j \otimes v_{i_1}\otimes \ldots \otimes v_{i_{k-1}} \otimes v_j \otimes v_{i_{k+1}} \otimes \ldots \otimes v_{i_a}
\end{align*}
for some $m_j\in M(0)$.
Additionally, if $i_k\in [n-1]$, then $m_{i_k+1}=A_{i_k, i_k+1}^- m\ne 0$.
\end{lemma}
\begin{proof} First note that by definition, $\Phi_n(y_k)(m\otimes v_{i_1}\otimes v_{i_2}\otimes \ldots \otimes v_{i_a})$ equals
\begin{align*}
&\left( \sum_{i<j} A_{ij}^- \otimes A_{ji}^+
- \frac{1}{2} \sum_{i} B_{ii}^+ \otimes C_{ii}^+
- \sum_{i < j} B_{ij}^+ \otimes C_{ji}^+ \right)_{0k} (m\otimes v_{i_1}\otimes v_{i_2}\otimes \ldots \otimes v_{i_a})=\\
&=\sum_{i<j} A_{ij}^-m \otimes v_{i_1}\otimes \ldots \otimes A_{ji}^+ v_{i_k} \otimes \ldots \otimes v_{i_a} -\frac{1}{2} \sum_{i} B_{ii}^+ m \otimes v_{i_1}\otimes \ldots \otimes C_{ii}^+ v_{i_k} \otimes \ldots \otimes v_{i_a} - \\
& \quad - \sum_{i < j} B_{ij}^+m \otimes v_{i_1}\otimes \ldots \otimes C_{ji}^+ v_{i_k} \otimes \ldots \otimes v_{i_a}.
\end{align*}
Thus, all summands are of the form $m_j \otimes v_{i_1}\otimes \ldots \otimes v_{j} \otimes \ldots \otimes v_{i_a}$ for $m_j \in M(0)$.
To determine the occuring $v_j$, recall that
$A_{ji}^+ = E_{ji} + E_{i' j'}$ and $C_{ji}^+ = E_{j'i} + E_{i' j}$, and therefore we have
\begin{equation}\label{eqn:Aji-action}
A_{ji}^+v_l =\delta_{il}v_j, \quad A_{ji}^+v_{l'} =\delta_{jl}v_{i'} , \text{ for } i<j \quad \text{ and } \quad C_{ji}^+v_l =\delta_{il}v_{j'}+\delta_{jl}v_{i'}, \quad C_{ij}^+v_{l'} =0.\end{equation}
In either case, $v_j$ is (possibly a constant multiple of) another standard basis vector, whose index appears strictly to the right of $i_k$ in (\ref{eqn:graph}), thus proving the first claim.
For the second, it follows from \eqref{eqn:Aji-action} that the only summand transforming $v_{i_k}$ to $v_{i_{k}+1}$ acts by $A_{i_k+1,i_k}^+$ on the $k$-th tensor factor, and thus acts by $A_{i_k,i_k+1}^-$ in the $0$-th tensor factor, replacing $m$ by $A_{i_k,i_k+1}^-m$.
\end{proof}
\subsection{The key construction}\label{key-construction}
The following construction, associating two vectors $v_d$ and $w_d$ to a diagram $d\in \Hom_{\sVW}(a,b)$, is key to the proof of Theorem~\ref{Thm2} in Section \ref{final-proof-2}. In the special case when $d$ has no dots and has the same number of cups and caps (i.e.\ $d\in \Hom_{\sVW}(a,a)^0$), it specializes to a certain construction from \cite[Section 4]{M}; see Section~\ref{final-proof-1} for details.
Given a diagram $d\in \Hom_{\sVW}(a,b)^k$ and $n\ge\frac{a+b}{2}+k$, define $v_d\in V^{\otimes a}$ and $w_d\in V^{\otimes b}$ by the following algorithm.
\begin{itemize}
\item[\bf{STEP 0.}] Put an ordering on the strings in $d$ so that caps come first, ordered left to right with respect to their left end; then through strings, ordered left to right with respect to their bottom end; then cups, ordered right to left with respect to their right end. (See for instance (\ref{eqn:Numberstrings}), where the strings are ordered using the set $\{\textcircled{1},\textcircled{2},\textcircled{3},\textcircled{4},\textcircled{5},\textcircled{6},\textcircled{7}\}$ with the usual ordering.)
\item[\bf{STEP 1.}]
Starting with the smallest cap label, and repeating along the order, label its left end by the minimal $i \in \left[n\right]$ which is bigger than all the labels already assigned. If the cap has $\ell$ dots, label its right end by $i+\ell$.
\item[\bf{STEP 2.}
Continue with the through strings in the assigned order, and for each, label its bottom end by the minimal $i \in \left[n\right]$ which is bigger than all the labels already assigned. If the through string has $\ell$ dots, label its top end by $i+\ell$.
\item[\bf{STEP 3.}
For each cup in order, label its right end by the minimal element $i$ of the set $\left[n\right]$ which is bigger than all the labels already assigned. If the cup has $\ell$ dots, label its left end by $i+\ell$.
\item[\bf{STEP 4.}] For each cup and cap, change the right end label from $i$ to $i'$.
\item[\bf{STEP 5.}] Now we have assigned to the bottom of the diagram labels $i_1, i_2,\ldots ,i_a$ and to the top $j_1, j_2,\ldots ,j_b$ for some $i_1, \ldots ,i_a, j_1,\ldots , j_b \in [n] \cup [n'] $. Set
$$v_d =v_{i_1}\otimes v_{i_2}\otimes \ldots \otimes v_{i_a}\in V^{\otimes a}, \quad\text{and}\quad
w_d =v_{j_1}\otimes v_{j_2}\otimes \ldots \otimes v_{j_b}\in V^{\otimes b}.$$
\end{itemize}
\begin{example}\label{big-example-vd-wd}
For instance, for $d=y_1^2 s_2 s_6\beta_3^*\beta_1^* s_3 s_2 \beta_1s_2 y_1^2 y_2 y_4^2 y_6 \in \Hom_{\sVW}(6,8)^8,$
\begin{equation}\label{eqn:Numberstrings}
d=\qquad
\TikZ{[scale=0.7]
\draw
(0,5) node{} to (0,3.5) node{}
(1,5) node{} to (2,4) node{}
(2,5) node{} to (1,4) node{}
(1,4) node{} to (1,3.5) node{}
(3,5) node{} to (3,4) node{}
(4,5) node{} to (4,4) node{}
(5,5) node{} to (6,4) node{}
(6,5) node{} to (5,4) node{}
(7,5) node{} to (7,4) node{}
(0,4.7) node[fill,circle,inner sep=1.5pt]{}
(0,4.3) node[fill,circle,inner sep=1.5pt]{}
(0,3.5) arc(-180:0:0.5)
(2,4) arc(-180:0:0.5)
(4,4) node{} to (4,3) node{}
(5,4) node{} to (5,3) node{}
(6,4) node{} to (7,3) node{}
(7,4) node{} to (6,3) node{}
(4,3) node{} to (4,2) node{}
(5,3) node{} to (6,2) node{}
(6,3) node{} to (5,2) node{}
(7,3) node{} to (7,2) node{}
(0,1) arc(180:0:0.5)
(4,2) node{} to (2,1) node{}
(5,2) node{} to (3,1) node{}
(6,2) node{} to (4,1) node{}
(7,2) node{} to (5,1) node{}
(0,1) node{} to (0,0) node{}
(1,1) node{} to (2,0) node{}
(2,1) node{} to (1,0) node{}
(3,1) node{} to (3,0) node{}
(4,1) node{} to (4,0) node{}
(5,1) node{} to (5,0) node{}
(0,0.3) node[fill,circle,inner sep=1.5pt]{}
(0,0.6) node[fill,circle,inner sep=1.5pt]{}
(1.2,0.2) node[fill,circle,inner sep=1.5pt]{}
(3,0.3) node[fill,circle,inner sep=1.5pt]{}
(3,0.6) node[fill,circle,inner sep=1.5pt]{}
(5,0.5) node[fill,circle,inner sep=1.5pt]{}
(0,-0.5) node{1}
(0,-1.2) node{\textcircled{1}}
(2,-0.5) node{3'}
(1,-0.5) node{4}
(1,-1.2) node{\textcircled{2}}
(4,5.5) node{5}
(3,-0.5) node{6}
(3,-1.2) node{\textcircled{3}}
(7,5.5) node{8}
(4,-0.5) node{9}
(4,-1.2) node{\textcircled{4}}
(6,5.5) node{9}
(5,-0.5) node{10}
(5,-1.2) node{\textcircled{5}}
(5,5.5) node{11}
(3,5.5) node{12'}
(1,5.5) node{12}
(1,6.2) node{\textcircled{6}}
(2,5.5) node{13'}
(0,5.5) node{15}
(0,6.2) node{\textcircled{7}}
;}\end{equation}
we get $v=v_1\otimes v_4\otimes v_{3'}\otimes v_6\otimes v_9\otimes v_{10}\in V^{\otimes 6}$, and $w_d=v_{15}\otimes v_{12}\otimes v_{13'}\otimes v_{12'}\otimes v_{5}\otimes v_{11}\otimes v_{9}\otimes v_8 \in V^{\otimes 8}$.
\end{example}
\begin{remark} \label{remark:theKeyconstruction}
The largest label is always $\frac{a+b}{2}+k$. We require $n\ge \frac{a+b}{2}+k$ in order to be able to realize $v_{\frac{a+b}{2}+k}\in V=\mathbb{C}^{n|n}$ in STEP 5.
The ordering in STEP 0 could be changed, as long as all caps come first, then all through strings, then all cups. This changes the vectors $v_d$ and $w_d$, but preserves the important features of the construction.
Observe also that if $i',j'\in [n']$ are labels with $i'$ at the bottom, $j'$ at the top, then $i<j$.
\end{remark}
\subsection{The key lemma}
The proof of linear independence relies on the observation that the vectors $v_d$ and $w_d$ can be used to distinguish diagrams in $S^k_{a,b}$.
Namely, the standard basis $v_1,\ldots , v_n,v_{1'}, \ldots, v_{n'}$ of $V$ induces a standard basis $B_b$ of $V^{\otimes b}$. For any vector $z\in M(0)\otimes V^{\otimes b}$ and any standard basis vector $w\in B_b$ we denote by $\left< w \mid z\right> \in M(0)$ the coefficient of $z$ in this standard basis. In other words,
$$z=\sum_{w\in B_b} \left< w \mid z\right>\otimes w.$$
\begin{lemma}\label{Key lemma}
Let $a, b, k \in \NN_0$. For any $d, d' \in S_{a,b}^k$, we have $\left< w_d \mid \Phi_n(d') v_d\right> \ne 0$ iff $d=d'$.
\end{lemma}
\begin{proof}
\fbox{$\Leftarrow$}
We repeatedly use the second part of Lemma~\ref{graph}.
Consider a cap with $\ell$ dots on it, and the edges labelled $i$ and $(i+\ell)'$. By Lemma ~\ref{graph}, applying the $\ell$ dots replaces $v_d$ by a linear combination of vectors which have the tensor factor $v_i$ of $v_d$ replaced by some $v_{j}$'s with $i\to j$, such that the path in \eqref{eqn:graph} from $i$ to $j$ has length at least $\ell$.
Exactly one such summand will give a non-zero contribution when such a $v_j$ is paired with $v_{(i+\ell)'}$ via $\beta$; namely, the one with $j= i+\ell$. Applying this dotted cap transforms the $0$-th tensor factor, say $m$, into the factor $A_{i,i+1}^-A_{i+1,i+2}^-\ldots A_{i+\ell -1,i+\ell }^- m$.
Next, consider a through string with $\ell $ dots and labels $i$ and $i+\ell$. It prescribes the order of some superswaps of tensor factors of $v_d$. After applying the $\ell$ dots, $v_d$ is replaced by a linear combination of vectors which have the tensor factor $v_i$ of $v_d$ replaced by some $v_{j}$ with $i\to j$, for which the path in \eqref{eqn:graph} from $i$ to $j$ has length at least $\ell$.
Reading off the coefficient of $w_d$ manifests itself in the tensor factor corresponding to this string to reading off the coefficient of $v_{i+\ell}$. The only summand with a non-zero contribution is the one with $j={i+\ell}$; in effect the $0$-th tensor factor got acted on by $A_{i,i+1}^-A_{i+1,i+2}^-\ldots A_{i+\ell -1,i+\ell}^-$.
Finally, consider a cup with $\ell$ dots and labels $i+\ell$ and $i'$. The $\beta^*$ corresponding to this cup produced $\sum_{j}(v_j\otimes v_{j'}-v_{j'} \otimes v_{j})$; applying the $\ell$ dots on the left end of it and reading off the coefficient of $v_{i+\ell}\otimes v_{i'}$ (as prescribed by $\left< w_d \mid\cdot \;\right>$) gives exactly one summand with a non-zero contribution. The effect on the $0$-th tensor factor is action by $A_{i,i+1}^-A_{i+1,i+2}^-\ldots A_{i+\ell-1,i+\ell}^-$. Thus, $\left< w_d \mid \Phi_n(d')v_d \right>$ is, up to a possible sign, equal to
\begin{align*}
\prod_{\TikZ{[scale=0.4]
\draw
(-0.1,0.5) arc(180:0:0.5)
(-0.1,0.5) node[fill,circle,inner sep=1.5pt]{}
(-0.4,0) node{\tiny{i}}
(1.1,-0) node{\tiny{(i+$\ell$)'}}
;}} A_{i,i+1}^-\ldots A_{i+\ell-1,i+\ell}^- \cdot
\prod_{\TikZ{[scale=0.4]
\draw
(0,0) node{} to (0,1) node{}
(0,0) node[fill,circle,inner sep=1.5pt]{}
(-0.4,0) node{\tiny{i}}
(-0.7,1) node{\tiny{i+$\ell$}}
;}} A_{i,i+1}^-\ldots A_{i+\ell-1,i+\ell}^- \cdot
\prod_{\TikZ{[scale=0.4]
\draw
(0,0) arc(-180:0:0.5)
(0,0) node[fill,circle,inner sep=1.5pt]{}
(-0.4,0.5) node{\tiny{i+$\ell$}}
(1.2,0.5) node{\tiny{i'}}
;}} A_{i,i+1}^-\ldots A_{i+\ell-1,i+\ell}^- \ne 0,
\end{align*}
where the factors are given by the shape and the assigned labels of $d$.
\fbox{$\Rightarrow$}
Let $d' \in S_{a,b}^k$ be any diagram for which
$\left< w_d \mid \Phi_n(d')v_d \right> \ne 0.$ We first recover the underlying connector $P(d')$ from the labelling of $d$.
Consider any cap in $d'$. By Lemma~\ref{graph} and the ordering $\to$, the dots increase indices $i\in [n]$ or replace them by $j'\in [n']$, and they decrease $j'\in [n']$. From that, and the facts $\left< w_d \mid \Phi_n(d')v_d\right> \ne 0$ and $\beta(v_i,v_j)=\beta(v_{i'},v_{j'})=0$, $\beta(v_i,v_{j'})=\delta_{ij}$, it follows that a cap in $d'$ can connect two points which are labelled in $d$ by an (unordered) pair of the form $\{i,j\}$ or $\{i,j'\}$ with $i\le j$.
Next, consider any cup in $d'$. Note that $\beta^*(1)=\sum_{i}(v_i\otimes v_{i'}-v_{i'}\otimes v_i)$, and that subsequent application of dots increases $i\in [n]$ or replaces it by $j'\in [n']$, and decreases $j'\in [n']$. Hence $\left< w_d \mid \Phi_n(d')v_d \right> \ne 0$ implies that a cup in $d'$ can only connect those pairs of points in $d$ labelled by $\{i',j'\}$, or by $\{i,j'\}$ with $i\ge j$.
Finally, consider any through string in $d'$. The possibilities for its labels (bottom and top) are then, by Lemma~ \ref{graph} and $\left< w_d \mid \Phi_n(d')v_d \right> \ne 0$, given by ordered pairs of the form $(i,j')$, or of the form $(i,j)$ with $i\le j$, or of the form $(i',j')$ with $i\ge j$. However, the last of these is not possible by Remark \ref{remark:theKeyconstruction}, so the remaining possibilities for the bottom and top labels of a through string are $(i,j')$ and $(i,j)$ with $i\le j$.
For any diagram $d''$, let $\cap(d'')$ denote the number of caps of $d''$; $\cup(d'')$ the number of cups, and $t(d'')$ the number of through strings. By the above analysis,
all labels $i'\in [n']$ on the bottom are on caps in $d'$, so
\begin{align}\label{eqn:ineq1}
\cap(d')\ge \# \textrm{ labels } j' \in [n'] \textrm{ at the bottom}= \cap(d).
\end{align}
As every cup in $d'$ has at least one label of type $ j' \in [n'] $, we also see that
\begin{align}\label{eqn:ineq2}
\cup(d')\le \# \textrm{ labels } j' \in [n'] \textrm{ at the top}= \cup(d).
\end{align}
We get a sequence of inequalities
\begin{align*}
t(d')= a-2\cap(d') \stackrel{(\ref{eqn:ineq1})}{\le} a-2\cap(d) = t(d)=
b-2\cup(d) \stackrel{(\ref{eqn:ineq2})}{\le} b-2\cup(d') = t(d').
\end{align*}
This implies that \eqref{eqn:ineq1} and \eqref{eqn:ineq2} are equalities, and moreover
\begin{align}\label{eqn:ineq3}
\cap(d')= \cap(d), \quad \cup(d')= \cup(d), \quad t(d')=t(d).
\end{align}
So, $d$ and $d'$ have the same number of cups, of caps and of through strings.
Next, we reconstruct the caps of $d'$. We saw in \eqref{eqn:ineq1}, \eqref{eqn:ineq3} that any label $j'\in [n']$ on the bottom of the diagram $d'$ needs to be on a cap, and all caps have exactly one label of type $j'\in [n']$. The other end of that cap is labelled by some $i\in [n]$ with $i\le j$.
Starting from the smallest bottom label of type $j'\in [n']$, there is exactly one label at the bottom of type $i\in [n]$ with $i\le j$, so these two labels must be joined by a cap in $d'$. To get the non-vanishing of the action of the dots composed with $\beta$ prescribed by this cap, this cap needs by Lemma~\ref{graph} to have at most $j-i$ dots in $d'$. (It has exactly $j-i$ dots in $d$). Proceed with the next smallest label of type $j'\in [n']$, noticing that there is exactly one unpaired label $i$ with $i\le j$, and pair them. After doing this for all $j' \in \left[n'\right]$ on the bottom, we see that the connectors $P(d')$ and $P(d)$ have the same pairing of the points given by caps, and every cap in $d'$ has at most as many dots as the corresponding cap in $d$.
Next, we recover the cups. By (\ref{eqn:ineq2}) and (\ref{eqn:ineq3}), every label of type $ j' \in [n']$ needs to be on an end of a cup, whereas the other end is labelled by some $i\in [n]$ with $j\le i$, and which has at most $i-j$ dots. By STEP 0 the cups come last, so there is exactly one such pairing of points on the top. So, $P(d')$ and $P(d)$ also have the same pairing of the points given by cups, and every cup in $d'$ has at most as many dots as the corresponding cup in $d$. Finally, all remaining unassigned labels are of type $i\in [n]$, and there is exactly one pairing such that the bottom label is smaller than the top label. So, the connectors $P(d')$ and $P(d)$ have the same pairing of the points given by through strings, and every through string in $d'$ has at most as many dots as the corresponding through string in $d$.
Therefore, $P(d)=P(d')$. As the underlying undotted diagrams of $d$ and $d'$ are both in $S_{a,b}$, they are the same. Finally, as $d'$ has at most as many dots as $d$ on every string, and they have the same total number of dots, we conclude that $d'=d$.
\end{proof}
\begin{example}
For the diagram $d$ from Example \ref{big-example-vd-wd},
$$\left< w_d \mid \Phi_n(d)v_d \right> =A_{12}^-A_{23}^-A_{45}^-A_{67}^-A_{78}^-A_{10,11}^-A_{13,14}^-A_{14,15}^- \in \mathcal{U}(\mathfrak{n}_-)=M(0).$$
\end{example}
\subsection{Proof of Theorem~\ref{Thm2}}\label{final-proof-2}
In this section we will finally prove the linearly independence of $S_{a,b}^\bullet$, thus proving Theorem~\ref{Thm2}. We start by proving it in the graded setting.
\begin{lemma}\label{Phi_n injecitve}
Given $a, b, k\in \mathbb{N}_0$, and $n\ge \frac{a+b}{2}+k$, the map
$$\Phi_n:\Hom_{\GVW}(a,b)^k\longrightarrow \Hom_{\GVect}(M(0)\otimes V^{\otimes a}, M(0)\otimes V^{\otimes b})^k$$ maps the set $S_{a,b}^k$ to a linearly independent set. Thus, $S_{a,b}^k$ is linearly independent in $\Hom_{\GVW}(a,b)^k$, and $\Phi_n$ is injective on $\Hom_{\GVW}(a,b)^k$.
\end{lemma}
\begin{proof}
Assume there are some $\alpha_{d'} \in \mathbb{C}$ such that
$\sum_{d' \in S_{a,b}^k} \alpha_{d'} \Phi_n(d')=0$.
For any $d \in S_{a,b}^k$, applying both sides of the above equation to the vector $v_d$, reading off the coefficient of $w_d$, and applying Lemma~\ref{Key lemma}, we get
$\alpha_d=0.$
So, the set $\{ \Phi_n(d) \mid d\in S_{a,b}^k\}$ is linearly independent. From that it follows that $S_{a,b}^k$ is linearly independent in $\Hom_{\GVW}(a,b)^k$. It is also a spanning set for $\Hom_{\GVW}(a,b)^k$ by Lemma~\ref{Phi is graded}, so $\Phi_n$ is injective on $\Hom_{\GVW}(a,b)^k$.
\end{proof}
\begin{corollary}\label{Basis-gsVW}
For all $a,b\in \mathbb{N}_0$, the set $S_{a,b}^\bullet$ is a basis of $\Hom_{\GVW}(a,b)$.
\end{corollary}
\begin{lemma}
\label{lemmathm2}
For all $a,b\in \mathbb{N}_0$, the set $S_{a,b}^\bullet$ is linearly independent in $\Hom_{\sVW}(a,b).$
\end{lemma}
\begin{proof}
Assume there is a nontrivial relation among elements of $S_{a,b}^\bullet$ in $\Hom_{\sVW}(a,b).$ As this is a filtered category, the highest order terms (of degree $k$) in this relation give a nontrivial relation among the elements of $S_{a,b}^k$ in $\Hom_{gr(\sVW)}(a,b)$. Thus, it is enough to prove that the set $S_{a,b}^k$ is linearly independent in $\Hom_{gr(\sVW)}(a,b)$ for each $k$.
Set $n=\frac{a+b}{2}+k$ and consider the square
$$
\xymatrix{
\Hom_{\GVW}(a,b)^k \ar@{->>}[d]_{\Theta} \ar@{^{(}->}[r]^<<<<<<{\Phi_n}
& \Hom_{\GVect}(M(0)\otimes V^{\otimes a}, M(0)\otimes V^{\otimes b})^k \\
\Hom_{gr(\sVW)}(a,b)^k \ar[r]^<<<<<{gr\Psi_n} & \Hom_{gr(\FVect)}(M(0)\otimes V^{\otimes a}, M(0)\otimes V^{\otimes b})^k \ar[u]_G }$$
The map $\Phi_n$ is injective by Lemma~\ref{Phi_n injecitve}, and the diagram strictly commutes by Lemma~\ref{square}. Thus, $\Theta$ is injective. It is surjective by Section \ref{Theta}, so it is an isomorphism of superspaces.
In particular, $\Theta$ maps the basis $S_{a,b}^k$ of $\Hom_{\GVW}(a,b)^k$ to a basis in $\Hom_{gr(\sVW)}(a,b)^k$ which by construction is $S_{a,b}^k$.
\end{proof}
\begin{corollary}$\Theta:\GVW\to gr(\sVW)$ is a graded isomorphism. \label{theta-iso}
\end{corollary}
\begin{corollary}
For any $a, b, k$, and $n\ge \frac{a+b}{2}+k$, the map $\Psi_n$ is injective on $\Hom_{\sVW}(a,b)^{\le k }$.
\end{corollary}
Theorem~\ref{Thm2} now follows directly from Proposition~\ref{oVW-span} and Lemma~\ref{lemmathm2}.
\subsection{A basis theorem for $\sBr$ as a special case} \label{final-proof-1}
Theorem~\ref{Thm1} now follows immediately by realizing the supercategory $\sBr$ as the $0$-th filtration piece of the supercategory $\sVW$.
\begin{proof}[Proof of Theorem 1]
Consider the functor $I: \sBr\to \sVW$ which is the identity on objects and interprets undotted diagrams as dotted diagrams with zero dots. For every $a$ and $b$, $I:\Hom_{\sBr}(a,b)\to \Hom_{\sVW}(a,b)$ maps the spanning set $S_{a,b}$ to the set $S_{a,b}^0$, which by Theorem~\ref{Thm2} is a basis of $\Hom_{\sVW}(a,b)^0$. Thus, the set $S_{a,b}$ is a basis of $\Hom_{\sBr}(a,b)$.
\end{proof}
\begin{remark}
The functor $\Psi_n\circ I:\sBr\to \FVect$ can be decomposed as
$\Psi_n\circ I=J_n\circ \Psi_n^{\mathbb{C}}$
where $\Psi_n^{\mathbb{C}}:\sBr\to \mathcal{V}ect$ is given on objects by $\Psi_n^{\mathbb{C}}(a)=V^{\otimes a}$ and the expected map on morphisms,
and $J_n:\mathcal{V}ect \to \FVect$, is given by $J_n(W)=M(0)\otimes W$. The functor $\Psi_n^{\mathbb{C}}$ appears in \cite{M}. It is shown there that when $n\ge a$, $\Psi_n^{\mathbb{C}}:\Hom_{\sBr}(a,a)\to \Hom_{\mathfrak{p}(n)}(V^{\otimes a}, V^{\otimes a})$ maps $S_{a,a}$ to a linearly independent set, thus proving that $S_{a,a}$ is a basis, and that $\Psi_n^{\mathbb{C}}$ is injective on $\Hom_{\sBr}(a,a)$. It is also proved that $\Psi_n^{\mathbb{C}}$ is surjective, so $\mathrm{End}_{\sBr}(a) \cong \mathrm{End}_{\pn}(V^{\otimes a})$ for $a \leq n$ (see \cite[Theorem 4.5]{M}).
\end{remark}
\begin{remark}
Clearly $\Psi^{\C}_n$ is not injective if $n<a$ since it is not injective when restricted to the symmetric group $S_a$. The question of surjectivity of the functors $\Psi_n^M$ for different modules $M$ is interesting and so far not understood. One would need to better understand the combinatorics of decomposition numbers in $\pn$ or category $\mathcal{O}(\mathfrak{p}(n))$. To our knowledge, only the decomposition numbers of the finite dimensional (thick and thin) Kac modules are known, see \cite{BDEHHILNSS}. Even in these cases, a precise surjectivity statement is so far not available. Based on explicitly calculated examples, we expect a more involved behaviour than in the $\fgl(n|n)$ case, see \cite{BS4}.
\end{remark}
\section{The affine VW superalgebra $\sVW_a$ and its centre}
\label{centre}
We fix $a\ge 2 \in \mathbb{N}$ for the whole section, and study the affine VW superalgebra $\sVW_a=\End_{\sVW}(a)$. The results from the previous section show that the algebra $\sVW_a$ is a PBW deformation of the algebra $\GVW_a$, in the sense that $\sVW_a$ is a filtered algebra, and $\gr(\sVW_a)=\GVW_a$. For $\hbar$ a parameter, the Rees construction gives the algebra $A_{\hbar}$ over $\mathbb{C}[\hbar]$, such that its specializations at $\hbar=0$ and $\hbar=1$ are precisely $A_1= \sVW_a$ and $A_0=\GVW_a$. We then use Theorem \ref{Thm2} to describe the center of the $\mathbb{C}[\hbar]$-algebra $A_{\hbar}$, and all its specializations $A_t$ for any $t\in \mathbb{C}$; in particular we find the centre of $\sVW_a$ and $\GVW_a$.
We refer e.g. to \cite{BG}, \cite{HSS}, \cite{Schedler}, \cite{SW} for the general theory.
\subsection{The algebras $A_{\hbar}$}
\label{secAbar}
We first define a $\mathbb{C}[\hbar]$-algebra $A_{\hbar}$ and its specializations $A_t$ at $t\in \mathbb{C}$ directly using generators and relations.
\begin{definition}\label{def:Ahbar}
Let $A_{\hbar}$ be the superalgebra over $\mathbb{C}[\hbar]$
with generators
$$s_i,\; e_i,\; y_j\quad 1 \leq i \leq a-1, \; 1 \leq j\leq a$$
where $\overline{s_i}=\overline{e_i}=\overline{y_j}=0$, subject to the relations:
\begin{multicols}{2}
\begin{enumerate}[(VW1)]
\item Involutions: $s_i^2=1 \text{ for } 1\leq i <a.$
\item Commutation relations:\label{comm}
\begin{enumerate}[label=(\roman*)]
\item $s_ie_j=e_js_i$ if $|i-j|>1$,
\item $e_ie_j=e_je_i$ if $|i-j|>1$,
\item $e_iy_j=y_je_i$ if $j\neq i, i+1$,
\item $y_iy_j=y_jy_i$ for $1\leq i,j\leq a$.
\end{enumerate}
\item Affine braid relations:\label{affbraid}
\begin{enumerate}[label=(\roman*)]
\item $s_is_j=s_js_i$ if $|i-j|>1$,
\item $s_is_{i+1}s_i=s_{i+1}s_is_{i+1}$
for $1\leq i \leq a-1$,
\item $s_iy_j=y_js_i$ if $j\neq i, i+1$.
\end{enumerate}
\item Snake relations:
\begin{enumerate}[label=(\roman*)]
\item $e_{i+1}e_ie_{i+1}=- e_{i+1}$,
\item $e_ie_{i+1}e_i=- e_i$
for $1\leq i\leq a-2$.
\end{enumerate}
\item Tangle and untwisting relations:\label{tang}
\begin{enumerate}[label=(\roman*)]
\item \label{twist}$e_is_i=e_i$ and $s_ie_i=-e_i$
for $1\leq i\leq a-1$,
\item $s_ie_{i+1}e_i= s_{i+1}e_i$,
\item $s_{i+1}e_ie_{i+1}=-s_ie_{i+1}$,
\item $e_{i+1}e_is_{i+1}= e_{i+1}s_i$,
\item $e_ie_{i+1}s_i=-e_is_{i+1}$
for $1\leq i \leq a-2$.
\end{enumerate}
\item \label{idemp}Idempotent relations:
$e_i^2=0$ for $1\leq i \leq a-1$.
\item Skein relations:\label{skein}
\begin{enumerate}[label=(\roman*)]
\item $s_iy_i-y_{i+1}s_i=-\hbar e_i-\hbar$,
\item $y_is_i-s_iy_{i+1}= \hbar e_i-\hbar$
for $1\leq i\leq a-1.$
\end{enumerate}
\item \label{unwrap} Unwrapping relations:
$e_1y_1^ke_1=0 \text{ for }k \in \NN$.
\item\label{sym} (Anti)-Symmetry relations:
\begin{enumerate}[label=(\roman*)]
\item $e_i(y_{i+1}-y_{i})=\hbar e_i$,
\item $(y_{i+1}- y_{i})e_i= -\hbar e_i$
for $1\leq i\leq a-1$.
\end{enumerate}
\end{enumerate}
\end{multicols}
For any $t\in \mathbb{C}$, let $A_{t}$ be the quotient of $A_{\hbar}$ by the ideal generated by $\hbar-t$.
\end{definition}
\begin{remark}
The above set of relations is not minimal. For instance, relations \ref{idemp} and \ref{unwrap} can be deduced from \ref{tang}\ref{twist} and \ref{sym}.
\end{remark}
As a $\mathbb{C}[\hbar]$-algebra, $A_{\hbar}$ is filtered by $\deg(y_i)=1$, $\deg(s_i)=\deg(e_i)=0.$
Considered as a $\mathbb{C}$-algebra, $A_{\hbar}$ can be given a grading by setting $\deg(y_i)=\deg(\hbar)=1$, $\deg(s_i)=\deg(e_i)=0.$ Interpreting $s_i$ , $e_i=b_i*b_i$, $y_i$ as diagrams as in Section~\ref{sectionone}, the elements of $A_{\hbar}$ and $A_t$ can be written as linear combinations of dotted diagrams with $a$ bottom points and $a$ top points.
\begin{lemma}
The set $S_{a,a}^{\bullet}$ is a spanning set for $A_{\hbar}$ and $A_t$ for any $t$.
\end{lemma}
\begin{proof}
Using the braid, snake and untwisting relations (analogous to (R1)-(R4)) in $A_{\hbar}$ or $A_t$ we see that every element of $S_{a,a}^{\bullet}$ gives rise to a well-defined element of $A_{\hbar}$, respectively $A_t$. Then we can repeat the proof that $S_{a,a}^{\bullet}$ spans $\sVW_a$ for these algebras.
\end{proof}
\begin{proposition}
\begin{enumerate}[(a)]
\item The assignments $\varphi_1(y_i)=y_i$, $\varphi_1(s_i)=s_i$ and $\varphi_1(e_i)=b_i^*b_i$ define an isomorphism of algebras $\varphi_1:A_1 \to \sVW_a$.
\item The assignments $\varphi_0(y_i)=y_i$, $\varphi_0(s_i)=s_i$ and $\varphi_0(e_i)=b_i^*b_i$ define an isomorphism of algebras $\varphi_0:A_0 \to \GVW_a$.
\item For any $t\ne 0$, the assignments $\psi_t(y_i)=ty_i$, $\psi_t(s_i)=s_i$ and $\psi_t(e_i)=e_i$ define an isomorphism of algebras $\psi_t:A_t \to A_1$.
\item The set $S^\bullet_{a,a}$ is a $\mathbb{C}$-basis of $A_t$ for any $t$, and a $\mathbb{C}[\hbar]$-basis of $A_{\hbar}$.
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{enumerate}[(a)]
\item One checks directly that $\varphi_1$ can be extended to an algebra homomorphism by checking that all relations from Definition \ref{def:Ahbar} hold in $\sVW_a$. To see surjectivity, consider an arbitrary element $b$ of $\sVW_a$, and let us construct its preimage. Assume without loss of generality that $b=p(y_1,\ldots, y_a) \, d \, q(y_1,\ldots, y_a)$ for some monomials $p,q$, and some undotted diagram $d$. If $d$ has $c$ cups, then it also has $c$ caps, and can be written in the form $d=\sigma_1 (b_1^*b_1)(b_2^*b_2)\ldots (b_c^*b_c)\sigma_2$ for some permutations $\sigma_1, \sigma_2$. Thus, $b=p\sigma_1 (b_1^*b_1)(b_2^*b_2)\ldots (b_c^*b_c)\sigma_2 q=\varphi_1(p \sigma_1 e_1e_2\ldots e_c\sigma_2 q)$. So, $\varphi_1$ is a surjective homomorphism mapping a spanning set to a basis, so it is an isomorphism.
\item Analogous to (a).
\item A direct check of the relations shows that this assignment extends to an algebra homomorphism for any $t\in \mathbb{C}$. For $t\ne 0$, the inverse is given by $\psi_t^{-1}(y_i)=\frac{1}{t}y_i$, $\psi_t^{-1}(s_i)=s_i$ and $\psi_t^{-1}(e_i)=e_i$.
\item
For any $t\ne 0$, $S_{a,a}^\bullet$ is a basis of $\sVW_a$ by Theorem~\ref{Thm2}, so by (a) and (c) above it is also a basis of $A_t\cong A_1\cong \sVW_a$. For $t=0$, $S_{a,a}^\bullet$ is a basis of $\GVW_a\cong A_0$ by Corollary~\ref{Basis-gsVW}. Assume there is a relation among the elements of $S_{a,a}^{\bullet}$ in $A_{\hbar}$, with coefficients in $\mathbb{C}[\hbar]$. Evaluating at some $t\in \mathbb{C}$ for which not all coefficients vanish, we get a relation in $A_t$, which is impossible. So, $S_{a,a}^{\bullet}$ is also a basis of $A_{\hbar}$. \qedhere
\end{enumerate}
\end{proof}
\subsection{The Rees construction}
Let $B=\bigcup_{k\ge 0}B^{\le k}$ be a filtered $\mathbb{C}$-algebra. The \emph{Rees algebra} of $B$ is the $ \mathbb{C}[\hbar]$-algebra $\operatorname{Rees}(B)$, given as a $\mathbb{C}$-vector space by $\operatorname{Rees}(B)=\bigoplus_{k\geq 0}B^{\le k}\hbar^k$, with multiplication and the $\hbar$-action both given by $(a\hbar^i)(b\hbar^j)=(ab)\hbar^{i+j}$ for $a\in B^{\le i}$, $b\in B^{\le j}$, and $ab\in B^{\le i+j}$ the product in $B$. It is graded as a $\mathbb{C}$-algebra by the powers of $\hbar$.
\begin{lemma}\label{lem:Rees-basis}
Let $\bigcup_{i\ge 0} S_i$ be a basis of $B$ compatible with the filtration, in the sense that the $S_i$'s are pairwise disjoint, and $\bigcup_{i= 0}^k S_i$ is a basis of $B^{\le k}$. Then $\bigcup_{i\ge 0} S_i\hbar^i$ is a $\mathbb{C}[\hbar]$-basis of $\operatorname{Rees}(B)$.
\end{lemma}
\begin{proof}
The set $\bigcup_{i= 0}^k S_i$ is a $\mathbb{C}$-basis of $B^{\le k}$, so
$\bigcup_{i= 0}^k S_i\hbar^k$ is a $\mathbb{C}$-basis of $B^{\le k}\hbar^k$, and then $\bigcup_{k\ge 0} \bigcup_{i= 0}^k S_i\hbar^k$ is a $\mathbb{C}$-basis of $\operatorname{Rees}(B)$. On the other hand, $\bigcup_{k\ge 0} \bigcup_{i= 0}^k S_i\hbar^k=\bigcup_{i\ge 0} \bigcup_{k\ge i} S_i\hbar^k=\bigcup_{i\ge 0} \bigcup_{j\ge 0} S_i\hbar^{i+j}=\bigcup_{j\ge 0} \hbar^j (\bigcup_{i\ge 0} S_i\hbar^{i})$. Thus, the set $\bigcup_{i\ge 0}S_i\hbar^i$ is a $\mathbb{C}[\hbar]$-basis of $\operatorname{Rees}(B)$.
\end{proof}
For any algebra $B$, let $Z(B)$ denote the centre of $B$.
\begin{lemma}
\label{ZRees}
$Z(\operatorname{Rees}(B))=\operatorname{Rees}(Z(B))$.
\end{lemma}
\begin{proof}
The centre of $B$ inherits the filtration of $B$, and $\operatorname{Rees}(Z(B))$ embeds naturally into $\operatorname{Rees}(B)$, with the image contained in $Z(\operatorname{Rees}(B))$. To see the other inclusion, assume $c$ is central in $\operatorname{Rees}(B)$. Without loss of generality $c$ is of homogeneous graded degree $i$, so $c=b \hbar^i$ for some $b \in B^{\le i}$. This shows that $b$ is a central in $B$, proving the claim.
\end{proof}
\begin{lemma} \label{algiso}
There is an isomorphism of $\mathbb{C}[{\hbar}]$-algebras
$\operatorname{Rees}(A_1)\cong A_{\hbar}$.
\end{lemma}
\begin{proof}
The map $A_{\hbar}\to \operatorname{Rees}(A_1)$ is given on generators by $y_i\mapsto \hbar y_i$, $s_i \mapsto s_i$, $e_i \mapsto e_i$. It is verified to be a morphism of algebras by directly comparing relations, and it is an isomorphism as it maps the basis $S_{a,a}^\bullet$ to the basis $S_{a,a}^\bullet$.
\end{proof}
\subsection{The centre is a subalgebra of the symmetric polynomials}
We now start computing the centre of $A_{\hbar}$, and show that $Z(A_{\hbar})\subseteq \mathbb{C}[\hbar][y_1,\ldots , y_a]^{S_a}$.
\begin{lemma}\label{centre-poly}
For $f \in A_{\hbar}$, the following are equivalent:
\begin{enumerate}[(a)]
\item $fy_i=y_if$ for all $i\in [a]$;
\item $f\in \mathbb{C}[\hbar][y_1,\ldots , y_a]$.
\end{enumerate}
\end{lemma}
\begin{proof}
Because of relation \ref{comm} (iv), only $(a)\Rightarrow (b)$ requires proof.
Assume that $f\notin\mathbb{C}[\hbar][y_1,\ldots , y_a]$. That means that the expansion of $f$ in the basis $S^\bullet_{a,a}$ contains at least one dotted diagram whose underlying undotted diagram is not the identity $1_a$.
Assume that this expansion of $f$ in the basis $S^\bullet_{a,a}$ contains at least one dotted diagram with a cup. Label the top and bottom endpoints of strings $1,\ldots, a$ from left to right. Among all diagrams with a cup, choose $d$ with a maximal number of dots on a cup; say that this cup is connecting $i$ and $j$, and has $k$ dots on it. Then $y_if$, written in the basis $S^\bullet_{a,a}$, contains at least one diagram with a cup and $k+1$ dots on it (namely, $y_id$). On the other hand, $fy_i$ contains no diagrams with $k+1$ dots on a cup, so $y_if\ne fy_i$.
Now assume that the expansion of $f$ in the basis $S^\bullet_{a,a}$ contains no diagrams with cups, and consequently no diagrams with caps. Then it contains at least one dotted diagram with a through strand connecting differently labelled points at the top and the bottom. Among all such diagrams and all such strings, choose $d$ with a maximal number of dots on such a string; say the string is connecting $i$ at the top of the diagram and $j$ at the bottom, $i\ne j$, and it has $k$ dots on it. Then $y_if$, written in the basis $S^\bullet_{a,a}$, contains at least one diagram with a string connecting $i$ and $j$ and with $k+1$ dots on it, while $fy_i$ contains no such diagrams as $i\ne j$. So, $y_i f\ne fy_i$.
\end{proof}
In particular, $Z(A_{\hbar})\subseteq \mathbb{C}[\hbar][y_1,\ldots , y_a]$. The following lemma shows that $Z(A_{\hbar})$ is in fact a subalgebra of the symmetric polynomials $\mathbb{C}[\hbar][y_1,\ldots , y_a]^{S_a}.$
\begin{lemma}\label{centre-sym}
Let $f \in \mathbb{C}[\hbar][y_1,\ldots , y_a]\subseteq A_{\hbar}$ and $1\le i\le a-1$.
\begin{enumerate}[(a)]
\item
If $fs_i=s_if$, then $f(y_1,\ldots, y_i,y_{i+1},\ldots, y_a)=f(y_1,\ldots, y_{i+1},y_i,\ldots, y_a)$.
\item For the special value $\hbar=0$, the converse also holds: if $f(y_1,\ldots, y_i,y_{i+1},\ldots, y_a)=f(y_1,\ldots, y_{i+1},y_i,\ldots, y_a)$, then $fs_i=s_if$ in $A_0$.
\end{enumerate}
\end{lemma}
\begin{proof} It is enough to prove this for $a=2$.
\begin{enumerate}[(a)]
\item By Lemma~\ref{AMR2.3}, the expansion of $fs_1$ in the basis $S^\bullet_{a,a}$ is
\begin{equation}\label{eqn:fsi}f(y_1,y_2)s_1=s_1f(y_2,y_1)+\hbar \sum_{i,j}\left(\alpha_{ij}y_1^iy_2^j+\beta_{ij}y_1^ie_1y_1^j\right)
\end{equation} for some $\alpha_{ij},\beta_{ij}\in \mathbb{C}$.
On the other hand, $s_1f$ is already a linear combination of normal diagrams. If $fs_1=s_1f$, then using that $S^\bullet_{a,a}$ is a basis, and reading off the terms with the underlying undotted diagram $s_1$, we get $s_1f(y_2,y_1)=s_1f(y_1,y_2)$, and so $f(y_2,y_1)=f(y_1,y_2)$.
\item For $\hbar=0$ and $f$ symmetric in $y_1,y_2$, equation~\eqref{eqn:fsi} turns into the equalities
$f(y_1,y_2)s_1=s_1f(y_2,y_1)=s_1f(y_1,y_2),$
thus $fs_1=s_1f$. \qedhere
\end{enumerate}
\end{proof}
\subsection{Some central elements}
Consider the following elements in $\mathbb{C}[\hbar][y_1,\ldots , y_a]$:
$$z_{ij}=(y_i-y_j)^2, \text{ for }1\leq i\not=j\leq a\quad \text{and}\quad
D_{\hbar}=\prod_{1\le i<j\le a}(z_{ij}-\hbar^2).$$ Notice that the deformed squared Vandermonde determinant $D_{\hbar}$ is symmetric, $D_{\hbar}\in \mathbb{C}[\hbar][y_1,\ldots, y_a]^{S_a}$.
We will use these to produce central elements in $A_{\hbar}$.
\begin{lemma}For any $1\le i\le a-1$, we have in $A_{\hbar}$ the equality
$$e_i\cdot (z_{i,i+1}-\hbar^2)=(z_{i,i+1}-\hbar^2)\cdot e_i=0,$$ and consequently
$D_{\hbar}e_i=e_i D_{\hbar}=0$.\label{centre-ei}
\end{lemma}
\begin{proof}
Using \ref{sym} (i), we get
$
e_i\cdot (z_{i,i+1}-\hbar^2)=e_i (y_{i+1}-y_i)^2 -\hbar^2e_i=\hbar e_i(y_{i+1}-y_i) -\hbar^2e_i=\hbar^2 e_i -\hbar^2 e_i=0,
$
which implies $e_iD_{\hbar}=0$.
The claim $D_{\hbar}e_i=0$ is proved analogously.
\end{proof}
\begin{lemma}For any $1\le k\le a-1$ we have $D_{\hbar}s_k=s_k D_{\hbar}$.\label{Dsi}
\end{lemma}
\begin{proof}We analyze the commutation of $s_k$ with different factors $(z_{ij}-\hbar^2)$ of $D_{\hbar}$ separately.
Assume $i,j \notin \{k,k+1\}$. Then \ref{affbraid}(iii) says that $y_i$ and $y_j$ commute with $s_k$. Therefore,
\begin{align}(z_{ij}-\hbar^2)s_k=s_k(z_{ij}-\hbar^2).\label{commute1}\end{align}
Now assume $i=k, j=k+1$. We claim that
\begin{align}(z_{k,k+1}-\hbar^2)s_k=s_k(z_{k,k+1}-\hbar^2).\label{commute2}\end{align}
To prove it, use \ref{skein} to calculate $(y_k-y_{k+1})s_k=s_k(y_{k+1}-y_k)-2\hbar$, and then
\begin{align*}
(y_k-y_{k+1})^2s_k&=(y_k-y_{k+1})s_k(y_{k+1}-y_{k})-2\hbar(y_k-y_{k+1})\\
&=(s_k(y_{k+1}-y_k)-2\hbar)(y_{k+1}-y_k)-2\hbar(y_k-y_{k+1})\;=\;s_k(y_k-y_{k+1})^2.
\end{align*}
The remaining factors of $D_{\hbar}$ contain $z_{ij}$ with exactly one of $i,j$ in $\{k,k+1\}$. Since $z_{ij}=z_{ji}$, it suffices to consider $j\ne k,k+1$, and further assume $j>k+1$. We claim that
\begin{align}(z_{k,k+1}-\hbar^2)\left( (z_{k,j}-\hbar^2)(z_{k+1,j}-\hbar^2)s_k\right)=(z_{k,k+1}-\hbar^2)\left(s_k(z_{k,j}-\hbar^2)(z_{k+1,j}-\hbar^2)\right).\label{commute3}\end{align}
To prove \eqref{commute3}, let us first calculate
\begin{align}
z_{k,j}s_k&=(y_k-y_j)^2s_k\;=\;(y_k-y_j)s_k(y_{k+1}-y_j)+\hbar(y_k-y_j)(e_k-1)\notag\\
&=s_kz_{k+1,j}+\hbar(e_k-1)(y_{k+1}-y_j)+\hbar(y_k-y_j)(e_k-1).\notag
\end{align}
From this and Lemma~\ref{centre-ei}, we get \
\begin{align}
(z_{k,k+1}-\hbar^2)(z_{k,j} -\hbar^2)s_k&=(z_{k,k+1}-\hbar^2)\left( s_k (z_{k+1,j} -\hbar^2) -\hbar (y_k+ y_{k+1}-2y_j)\right).
\label{zijskstep1}
\end{align}
Similarly,
\begin{align}
(z_{k,k+1}-\hbar^2)(z_{k+1,j} -\hbar^2)s_k&=(z_{k,k+1}-\hbar^2)\left( s_k (z_{k,j} -\hbar^2) +\hbar (y_k+ y_{k+1}-2y_j)\right).
\label{zijskstep2}
\end{align}
Using \eqref{zijskstep1} and \eqref{zijskstep2}, we then obtain \eqref{commute3}, since $(z_{k,k+1}-\hbar^2)\left( (z_{k,j}-\hbar^2)(z_{k+1,j}-\hbar^2)s_k\right)$ equals
$$(z_{k,k+1}-\hbar^2)\left[s_k(z_{k,j}-\hbar^2)(z_{k+1,j}-\hbar^2) +(\hbar (y_k+ y_{k+1}-2y_j)-\hbar (y_k+ y_{k+1}-2y_j))(z_{k,j}-\hbar^2)\right]$$ which is however the same as
$(z_{k,k+1}-\hbar^2)\left(s_k(z_{k,j}-\hbar^2)(z_{k+1,j}-\hbar^2)\right)$. Thus \eqref{commute3} holds. Finally, \eqref{commute1}, \eqref{commute2} and \eqref{commute3} imply $D_{\hbar}s_k=s_k D_{\hbar}$.
\end{proof}
\begin{lemma} Let $1\le i \le a-1$, and let $\tilde{f}\in \mathbb{C}[\hbar][y_1,\ldots , y_a]$ be symmetric in $y_i,y_{i+1}$. Then there exist polynomials $ p_j= p_j(y_1,\ldots , y_a)\in \mathbb{C}[\hbar][y_1,\ldots , y_a]$ such that
$$\tilde{f}s_i=s_i\tilde{f}+\sum_{j=0}^{\deg{\tilde{f}}-1} y_{i}^j \cdot e_i \cdot p_j.$$
\label{centre-tildefsi}
\end{lemma}
\begin{proof}
Analogues of the formulas in Lemma~\ref{lem:dot-on-cup} and~\ref{AMR2.3} imply that for any $k$,
\begin{align*}
(y_i^k+y_{i+1}^k)s_i&=s_i(y_i^k+y_{i+1}^k)+\hbar \sum_{j=0}^{k-1} \left(y_i^{k-1-j}e_i y_{i+1}^j+ y_{i+1}^je_i y_{i}^{k-1-j}\right)\\
&=s_i(y_i^k+y_{i+1}^k)+\hbar \sum_{j=0}^{k-1} y_i^{k-1-j}e_i y_{i+1}^j+ \sum_{j=0}^{k-1}\sum_{\ell=0}^{j} \hbar^{1+j-\ell}(-1)^{j+\ell} y_{i}^\ell e_i y_{i}^{k-1-j}.
\end{align*}
Thus, the claim holds for $\tilde{f}=y_i^k+y_{i+1}^k$. It also trivially holds for $\tilde{f}=y_j$ if $j\ne i,i+1$, as such $y_j$ commute with $s_i$. Finally, note that if the claim holds for $\tilde{f}_1$ and $\tilde{f}_2$, it also holds for $\tilde{f}_1\tilde{f}_2$ and $\tilde{f}_1+\tilde{f}_2$. On the other hand, the algebra of polynomials symmetric in $y_i,y_{i+1}$ is generated by the $y_i^k+y_{i+1}^k$, $k\ge 1$, and $y_j$'s with $ j\ne i,i+1$, and the claim follows.
\end{proof}
\begin{lemma}\label{lem:manycentralelts}
Let $\tilde{f}\in \mathbb{C}[\hbar][y_1,\ldots , y_a]^{S_a}$ be an arbitrary symmetric polynomial, and $c$ a constant. Then $f=D_{\hbar}\tilde{f}+c$ lies in the centre of $A_{\hbar}$.
\end{lemma}
\begin{proof}The element $f$ is in $\mathbb{C}[\hbar][y_1,\ldots, y_a]$ so it commutes with $y_i$ for all $i$. By Lemma~\ref{centre-ei}, $$fe_i=\tilde{f}D_{\hbar}e_i+ce_i=ce_i=e_i D_{\hbar}\tilde{f}+ce_i=e_if.$$
Using Lemma~\ref{centre-tildefsi}, and then Lemmas \ref{Dsi} and \ref{centre-ei} we get
\begin{align*}
fs_i& =D_{\hbar}\tilde{f}s_i+cs_i=D_{\hbar}\left( s_i\tilde{f}+\sum_j y_{i}^j \cdot e_j \cdot p_j \right)+cs_i= s_iD_{\hbar}\tilde{f}+s_ic= s_i f. \qedhere
\end{align*}
\end{proof}
\subsection{The centre of $\sVW_a$ and of $A_{\hbar}$}
\begin{proposition}\label{prop:Z(A0)}
The centre $Z(A_0)$ of the graded VW superalgebra $\GVW_a$ consists of all $f\in \C[y_1,\ldots, y_a]$ of the form
$f=D_{0}\tilde{f}+c$, for $\tilde{f}\in \C[y_1,\ldots, y_a]^{S_a}$ and $c\in \mathbb{C}$.
\end{proposition}
\begin{proof}
We showed in Lemmas \ref{centre-poly} and \ref{centre-sym} that $Z(A_0) \subseteq \C[y_1,\ldots, y_a]^{S_a}$, and that any symmetric polynomial commutes with $s_i$ for $1 \leq i \leq a-1$ and $y_j$ for $1 \leq j \leq a$. It remains to check which symmetric polynomials commute with $e_i$ for all $1 \leq i \leq a-1$. To this end, fix $f\in Z(A_0)$. We will compute a condition on commutation with $e_1$; then the symmetry of $f$ will complete the proof.
Expanding $fe_1$ in the normal dotted diagram basis, the terms appearing with nonzero coefficient all have underlying (undotted) diagrams equal to $e_1$; i.e.\ $fe_1$ is a linear combination of terms of the form $y_1^ke_1p_{k}$ with $p_k \in \C[y_3,\ldots , y_a]$. Similarly, $e_1f$ is a linear combination of terms of the form $e_1y_1^k p_{k}$. Comparing, we get $p_k=0$ for $k>0$, and that $fe_1=p_{0}(y_3,\ldots , y_a)e_1$. Using the presentation of $A_0$ given in Definition \ref{def:Ahbar}, we have that a polynomial in the $y_i$'s is annihilated by $e_1$ if any only if it is a multiple of $(y_1-y_2)$ (see \ref{sym}, specializing to $\hbar = 0$). Thus
$$f = (y_1-y_2) g+ p_0, \quad \text{ with } g \in \mathbb{C}[y_1,\ldots, y_a] \text{ and } p_0\in \mathbb{C}[y_3,\ldots, y_a].$$
We claim that $p_0 \in \C$, which will follow from the symmetry of $f$. For this let $b y_3^{\lambda_3}\cdots y_a^{\lambda_a}$ be a non-zero summand of $p_0$, and write $y^\lambda = y_4^{\lambda_4} \cdots y_a^{\lambda_a}$ for short.
If $\lambda_3\geq 1$,
then symmetry implies
$b y_1^{\lambda_3} y^{\lambda}$ is a term in $f$, so that
$by_1^{\lambda_3-1} y^{\lambda}$ is a term in $g$. So $-by_1^{\lambda_3-1} y_2 y^{\lambda}$, and therefore $-b y_2 y_3^{\lambda_3-1} y^{\lambda}$, are summands in $f$. Going back to $g$, we get that $b y_3^{\lambda_3-1} y^{\lambda}$ is a summand there, so that $b y_1 y_3^{\lambda_3-1} y^{\lambda}$ is a summand of $f$. But comparing the coefficient to that of $y_2 y_3^{\lambda_3-1} y^{\lambda}$, we see that this contradicts the symmetry of $f$. Therefore $\lambda_3 = 0$ for all non-zero summands of $p_0$, and thus by symmetry, $p_0 \in \mathbb{C}$ as claimed.
Next, since $f$ is symmetric (specifically in $y_1$ and $y_2$), we have $g$ is antisymmetric in $y_1$ and $y_2$. Thus $g$ itself is a multiple of $(y_1 - y_2)$, i.e.\ $f-p_0$ is a multiple of $(y_1 - y_2)^2$. But now, since $f - p_0$ is symmetric, it must also be a multiple of $D_{0} = \prod_{1\le i<j\le a}(y_i-y_j)^2$.
So finally, $f$ is of the form
$$f=\prod_{1\le i<j\le a}(y_i-y_j)^2\cdot \tilde{f}+c=D_{0}\tilde{f}+c,$$
for some symmetric polynomial $\tilde{f} \in \C[y_1,\ldots, y_a]^{S_a}$ and constant $c \in \C$.
\end{proof}
\subsection{The centre of $\sVW_a$}
The main result of this section, Theorem~\ref{centreA1}, describes the centre of $\sVW_a$. To do that, we use the fact that the algebra $\sVW_a$ is a PBW deformation of the algebra $\GVW_a$, determine the centre of $\GVW_a$ and find a lift of the appropriate basis elements to $\sVW_a$. This approach differs from the common arguments for diagram algebras, where often the centre is realized as a subring of invariant polynomials satisfying certain cancellation properties, \cite{DRV2}. In our situation the cancellation properties did not appear very manageable, and we therefore omitted them. It would however be nice to know if an explicit result as Theorem~\ref{centreA1} could be achieved for instance for affine VW algebras as in \cite{Nazarov}, \cite{ES2}, BMW-algebras, see e.g. \cite{DRV2}, or walled Brauer algebras, see e.g. \cite{JK}, \cite{Sartori}. Compare also with \cite{C}, where the center of the Brauer superalgebra $\sBr_a$ is described in a similar way.
\begin{theorem}\label{centreA1}
The centre $Z(\sVW_a)$ of the VW superalgebra $\sVW_a=A_1$ consists of all $f\in \C[y_1,\ldots, y_n]$, of the form
$f=D_{1}\tilde{f}+c$, for $\tilde{f}\in \C[y_1,\ldots, y_a]^{S_a}$ an arbitrary symmetric polynomial and $c\in \mathbb{C}$.
\end{theorem}
\begin{proof}
For any filtered algebra $B$ there exists a canonical injective algebra homomorphism $\varphi :\gr Z(B)\hookrightarrow Z(\gr(B))$, given for $f\in Z(B)^{\le k}$ by $\varphi(f+Z(B)^{\le (k-1)})= f+B^{\le (k-1)}$, see~\cite[6.13, 6.14]{MR}. For $B=\sVW_a$ and $\gr(B)=\GVW_a$, by Proposition~\ref{prop:Z(A0)} the centre of $A_0$ consists of elements of the form $f=D_{0}\tilde{f}+c$ for $\tilde{f}$ a symmetric polynomial and $c$ a constant. By Lemma~\ref{lem:manycentralelts}, $D_{1}\tilde{f}+c$ lies in the centre of $\sVW_a$, and we have $\varphi (c)=c$, and for $\tilde{f}$ symmetric and homogeneous of degree $k$, $\varphi(D_{1}\tilde{f}+\sVW_a^{\le a(a-1)+k-1})=D_{0}\tilde{f}$. Using Proposition~\ref{prop:Z(A0)}, we see that every $f\in Z(\GVW_a)$ is in the image of $\varphi$, so $\varphi$ is an isomorphism.
\end{proof}
\begin{remark}
It is interesting to compare the description of the centre of $\sVW_a$ with \cite[Theorem 4.8]{Verapn}. It is shown there that the centre of $\mathcal{U}(\mathfrak{p}(n))/I$, where $I$ is the Jacobson radical of $\mathcal{U}(\mathfrak{p}(n))$, is isomorphic to the subring in the polynomial ring $\mathbb C[z_1,\dots z_n]$ of the form $\mathbb C\oplus\mathbb C[z_1,\dots,z_n]^{S_n}\Theta$, where $\Theta=\prod_{i<j}(z_i-z_j)^2$. In other words, this centre is isomorphic to $Z(\sVW_a)$ when $a=n$.
\end{remark}
\begin{theorem}
The centre $Z(A_{\hbar})$ of the superalgebra $A_{\hbar}$ consists of polynomials $f\in \C[\hbar][y_1,\ldots, y_n]$, of the form
$f=D_{\hbar}\tilde{f}+c$, for $\tilde{f}\in \C[\hbar][y_1,\ldots, y_a]^{S_a}$ an arbitrary symmetric polynomial and $c\in \mathbb{C}[\hbar]$.
\end{theorem}
\begin{proof}
The centre $Z(A_{\hbar})$ is by Lemma~\ref{algiso} isomorphic to
$Z(\operatorname{Rees}(A_{1}))$, which is by Lemma~\ref{ZRees} isomorphic to $\operatorname{Rees}(Z(A_1))$. The centre $Z(A_1)$ consists by Theorem~\ref{centreA1} of elements of the form $f=D_{1}\tilde{f}+c$, with $\tilde{f}\in \C[y_1,\ldots, y_a]^{S_a}$ and $c\in \mathbb{C}$. Assume $\tilde{f}$ is homogeneous of degree $k$. Then $D_{1}\tilde{f}\in A_1^{\le k+a(a-1)}$, which gives an element $D_{1}\tilde{f}\hbar^{k+a(a-1)}$ of $\operatorname{Rees}(Z(A_1))\cong Z(\operatorname{Rees}(A_1))$. Using Lemma~\ref{algiso}, we see that $Z(A_{\hbar})$ is spanned by constants and the preimages under the isomorphism $A_{\hbar}\cong \operatorname{Rees}(A_1)$ of elements $D_{1}\tilde{f}\hbar^{k+a(a-1)}$,
which are equal to $D_{\hbar}\tilde{f}$.
\end{proof}
\addtocontents{toc}{\protect\setcounter{tocdepth}{0}}
|
\section*{Background}
Combined coronary artery disease and ventricular hypertrophy are not uncommon; they both share hypertension, which affects \(\sim 26\%\) of the world population \cite{kearney2005global}, as a risk factor. Accumulation of atheromatous plaques under tunica intima of the epicardial arteries restricts the blood flow to the supplied cardiac tissue. Chronic high-grade narrowing of the coronary arteries induces subendocardial ischemia during the escalation of the myocardial oxygen demand throughout exercise or stress \cite{libby2005pathophysiology}. The strained myocytes release mediators like adenosine and bradykinin \cite{safran2001cardioprotective}\cite{parratt1997bradykinin} which in addition to stimulating coronary vasodilatation, irritate the nerve endings leading to anginal pain \cite{crea1990role}.\\
The treatment strategy for treating CAD aims to improve survival and/or relieve symptoms \cite{wijns2010guidelines} including dyspnea and stable angina pectoris. Said strategy usually involves anti-anginal medications and/or PCI, or CABG in case of complex CAD and/or left main involvement, for achieving those aims. Trials show that revascularization by PCI or CABG is more effective than a strategy of medical therapy alone, in relieving symptoms like angina and dyspnea. Besides; it improves the quality of life by reducing the use of anti-angina drugs and increasing exercise capacity \cite{chamberlain1997coronary}\cite{time2001trial}\cite{weintraub2008effect}\cite{erne2007effects}. However; studies indicate that PCI, as an initial management strategy in patients with stable coronary artery disease, did not reduce the risk of complications as myocardial infarction or other major cardiovascular events when added to optimal medical therapy \cite{boden2007optimal}\cite{hueb2004medicine}\cite{stergiopoulos2012initial}.\\
The guidelines recommend that ad hoc PCI should not automatically be applied after angiography \cite{windecker2014authors}; and emphasize the usefulness of optimal medical treatment for selected patients, which can reduce angina and the risk of myocardial infarction and stroke substantially and prevent progression of atherosclerosis in the entire vasculature. Consequently; the clinical decisions for management of CAD generally need to balance adherence to guidelines against judgments based on specific patient, operator, social, economic, and cultural factors \cite{lamy2011medical}. Generally; PCI and medical therapy should be viewed as complementary, rather than opposing, strategies \cite{blumenthal2000medical}. Patients with stable coronary artery disease and functionally significant stenoses, benefit from the combination therapy of PCI plus optimal medical therapy by showing greater symptomatic improvement \cite{pursnani2012percutaneous} and decreased the need for urgent revascularization. However; in patients without ischemia, the outcome appeared to be favorable with the optimal medical therapy alone \cite{de2012fractional}.\\
For revascularization decisions and recommendations; said significant stenosis has been defined by most studies of CAD revascularization, which have been based on and reported according to angiographic criteria, as \(\geq70\%\) diameter narrowing, and/or \(\geq50\%\) for left main CAD \cite{levine20112011}. Besides; angiophysiological criterion, such as an assessment of fractional flow reserve (FFR), has been used in deciding when revascularization is indicated. Thus, for recommendations about revascularization, coronary stenosis with FFR \(\leq0.8\) is also considered to be significant \cite{pijls1996measurement}. The standard values provided by both methods, and so the revascularization decision, don't consider the relation between the resulting effective flow distal to the stenosis and the demand of a comorbid hypertrophied myocardial tissue.
\section*{Model}
Hagen-Poiseuille law, which is an analytical solution to the Navier-Stokes equation \cite{bruus2007theoretical}, states that; the flow rate \(Q\) through a coronary vessel is directly proportional to the pressure gradient \(\Delta P\) between the aortic root and the right atrium; and inversely proportional to the resistance \(R\) within the vessel. Wherein; the resistance \(R\) is inversely proportional to the radius \(\alpha\) of the vessel elevated to the fourth power, and is directly proportional to the blood viscosity \(\mu\) and the vessel length \(\Delta l\). So, by considering a circular cross-section of the vessel:
\begin{equation*}
Q=\Delta P \frac{\pi \alpha^4}{8 \mu \Delta l}
\end{equation*}
when
\begin{equation*}
R= \frac{8 \mu \Delta l}{\pi \alpha^4}
\end{equation*}
so;
\begin{equation*}
Q \propto \frac{1}{R} \propto \alpha^4 \propto \frac{1}{\Delta l}
\end{equation*}\\
The blood flow, which is a non-Newtonian fluid, within the circulation doesn't imitate precisely this law \cite{klabunde2011cardiovascular}, because said relation is applied on a Newtonian fluid in the steady laminar flow moving through a long cylindrical pipe. Still, the law outlines the dominant determinants which influence the blood flow \(Q\) within the vasculature either in physiological or pathological conditions.\\
Atherosclerosis commonly affects the epicardial coronary vessels leading to narrowing of the vessel caliber \(\alpha_e\) and increase vascular resistance of the supplying vessel \(R_e\).
While:
\begin{equation*}
R_e \propto \frac{1}{\alpha^4_e}
\end{equation*}
\noindent the corresponding supplied myocardial segment doesn't actually suffer this severe blood flow reduction indicated in the above equation. The vasculature of the coronary circulation is arranged in-series, in addition to the in-parallel arrangement, so that the epicardial vascular resistance \(R_e\) is a segmental resistance. The coronary circulation can be divided into two compartments, the large epicardial conduit vessels and the resistance vessels, which are typically less than \(300 \mu m\) in diameter \cite{schelbert2010anatomy}. Whereas the conduit vessels exert little if any resistance to flow, resistance to flow progressively rises as the vessel diameter of the resistance vessels declines from about \(300 \mu m\) in the small arteries to less than \(100 \mu m\) in the arteriolar vessels \cite{muller1996integrated}. Therefore; the total resistance to blood flow \(R\) comprises mainly the pre-capillary resistance \(R_c\), the resistance of microvasculature \(R_m\), and the negligible resistance of the epicardial or conductance vessels \(R_e\).
\begin{equation*}
R = R_c + R_m + R_e
\end{equation*}\\
Narrowing of the radius of the epicardial vessel, due to atheromatous plaque, will increase the resistance in this vessel; but as:
\begin{equation*}
R_c + R_m \gg R_e
\end{equation*}
\noindent the impact of mild to moderate increase of the epicardial resistance \(R_e\) on the overall resistance of the coronary circulation \(R\) is insignificant.\\
However; in case of combined coronary artery disease and ventricular hypertrophy, both \(R_e\) and \(R_c+R_m\) are increased. Microangiogenesis is activated during the pathogenesis of ventricular hypertrophy as a compensatory mechanism to maintain effective blood supply to the hypertrophied tissue. Accordingly; CAD causes an increase in \(R_e\) due to epicardial arterial stenosis and ventricular hypertrophy increases \(R_c+R_m\) due to neomicroangiogenesis, i.e. addition of a new microvascular segment.
\begin{equation*}
R_{c+m} \propto \frac{1}{\alpha^4_m} \propto \Delta l
\end{equation*}
\noindent consequently; the flow rate \(Q\), and so the perfusion of myocardial tissue, diminish significantly upon subcritical stenosis of the supplying epicardial artery during the pathogenesis of CAD.\\
As mentioned; the identification of clinically-significant stenosis of an epicardial artery depends on an angiographic criterion, its radius \(\alpha_e\), and/or an angiophysiologic criterion, the FFR, which is an absolute number result from the ratio of the pressure distal to the lesion, to the pressure proximal to the lesion during induced hyperemia \cite{pijls1996measurement}. The graphical representation of the relation between both these criteria of the supplying artery and the perfusion of the supplied myocardial tissue follows a direct proportional relationship represented by a sigmoid-shaped curve, due to the effect of segmental resistance. Myocardial perfusion \(\xi\) describes the blood flow \(Q\) in ml/min per cubic centimeter of cardiac muscle volume \(V\).
\begin{equation*}
\xi = \frac{Q}{V}
\end{equation*}\\
According to the relation between the radius \(\alpha_e\) of an epicardial coronary artery, as an angiographic criterion, and the perfusion \(\xi\) of the corresponding supplied myocardial tissue represented in Figure 1; the perfusion \(\xi\) doesn't decrease significantly with gradual stenosis in isolated CAD until a critical stenotic value \(\phi_\alpha\) is reached, wherein the perfusion collapses relatively. Clinically; said critical value \(\phi_\alpha\) is defined as \(\geq70\%\) radius \(\alpha_e\) reduction, significant stenosis \cite{levine20112011}. However; in patients with comorbid CAD and ventricular hypertrophy; the curve is shifted to the right indicating an increase in the critical stenotic value \(\phi_\alpha\), so that the perfusion \(\xi\) of the corresponding supplied myocardial tissue collapses relatively at a clinically subsignificant stenosis. The right shift in said patients depends on the degree of ventricular hypertrophy.\\
Additionally; the relation between another angiophysiologic criterion, the fractional flow reserve FFR, within a stenotic epicardial artery and the perfusion \(\xi\) of the corresponding supplied myocardial tissue, as represented in Figure 2, indicates that the perfusion \(\xi\) is not meaningfully reduced with the gradual decrease of FFR until a critical value \(\phi_{FFR}\) is reached, wherein the perfusion \(\xi\) collapses relatively. Clinical trials defined said critical value \(\phi_{FFR}\) as a FFR \(\leq 0.8\) \cite{tonino2009fractional}\cite{pijls2007percutaneous}. Though; in patients with combined CAD and ventricular hypertrophy; the curve shows a right shift, which is directly proportional to the degree of ventricular hypertrophy, indicating an increase in the critical stenotic value \(\phi_{FFR}\), so that the perfusion \(\xi\) of the corresponding supplied myocardial tissue collapses relatively at a clinically subsignificant reduction in the FFR.\\
The proposed model gives a more sensitive formula to detect the critical stenosis, which takes into account the demand of the supplied bulky myocardium. The isolated CAD curve is a logistic function; wherein \(x\) represents the critical stenosis and \(k\) is the curve slope:
\begin{equation*}
f(x) = \frac{1}{1+e^{-kx}}
\end{equation*}
\noindent in patients with comorbid CAD and ventricular hypertrophy; the curve is shifted to the right by \(a\) yielding \(x^\backprime\) as a representation of the critical stenosis:
\begin{equation*}
f(x) = \frac{1}{1+e^{-k(x^\backprime - a)}}
\end{equation*}
\noindent then;
\begin{equation*}
x^\backprime = x+a
\end{equation*}
\noindent wherein the curve shift \(a\) is directly proportional to the difference in muscle bulk \(\Delta M\) which obtained by Echocardiogram;
\begin{equation*}
a \propto \Delta M
\end{equation*}
\begin{equation*}
a = \omega \Delta M
\end{equation*}
\noindent the value of the constant \(\omega\) can be obtained experimentally. So; the percentage of the critical patency in patients with comorbid CAD and ventricular hypertrophy \(x^\backprime\) is:
\begin{equation*}
x^\backprime = x + (\omega \Delta M)
\end{equation*}
\section*{Results}
\textbf{Individuals with pathological ventricular hypertrophy are more sensitive to haemodynamic changes of the coronary circulation or pathologies that reduce the coronary reserve.} Ventricular hypertrophy stresses the subendocardial myotissue due to increase the structural resistance of the coronary circulation. Said stress is ameliorated by compensatory functional changes to sustain the normal coronary blood flow. Although during vigorous exercise, the compensatory capability of the coronary flow reserve is exhausted under the effect of demand upsurge and shortened diastolic period. Occasional hemodynamic disturbances or subclinical pathologies which lessen the maximum coronary reserve may lead to selective subendocardial hypoperfusion.\\
\textbf{Comorbid CAD and ventricular hypertrophy cause the subendocardial tissue to suffer, during exercise or stress, from ischemia at an angiographically subsignificant stenosis in the supplying epicardial artery.} CAD primes the structural resistance of the neomicrovasculature of the hypertrophied tissue. So; subcritical stenosis of the corresponding epicardial artery, mainly due to atherosclerosis, causes the total resistance to rise effectively to reduce the flow rate and exhaust the reactive compensatory mechanisms. The curve shift to the right in said patients doesn't affect the risk of myocardial infarction, \textbf{yet they are more susceptible to and usually presented by NSTEMI; with higher rates of transition from ischemia to necrosis in the affected hypertrophied endocardial tissue.} Increased muscle bulk shifts the endocardium away from the main blood supply. Besides; subjection to higher extravascular pressure depletes the functional vasodilator reserve in long standing pathological hypertrophy.\\
\textbf{Patients with combined CAD and ventricular hypertrophy have a higher risk to develop arrhythmias than their peers who suffer from isolated CAD.} In pathological hypertrophy; the neomicroangiogenesis shows anatomical and architectural dysgenesis in relation to the hypertrophied tissue. Said dysgenesis leads to failure of the coronary bed to uniformly supply the cardiac muscle, rendering foci within the hypertrophied muscle bulk at greater risk of ischemic injury. These stressed foci can be arrhythmogenic upon increased cardiac demand leading to serious arrhythmia and sudden cardiac death.
\section*{Discussion}
During cardiac catheterization; the main determinants of revascularization therapy in CAD patients, are either angiographic or angiophysiological criteria, measured during drug induced hyperemia, to identify the clinically-significant stenosis. Said determinants depend on the relation between the size of the insinuated plaque and the vascular diameter. A stenosis which reduces the radius of the epicardial vessel by \(70\%\) is considered significant angiographically. In another determinant; a small sensor on the tip of the guidewire is used to measure the pressure, temperature and flow to determine the exact severity of the lesion. The ratio between the pressure distal to the lesion and the pressure proximal to the lesion, which is the fraction flow reserve FFR, measures the pressure drop caused by the stenosis. A fraction flow reserve value \(\leq 0.8\) defines the stenosis to be significant. The FFR is a better method to detect the physiological significance of a stenosis \cite{pijls2010fractional}, as it takes into account collateral flow, which can render an anatomical blockage functionally unimportant. Also, standard angiography can underestimate or overestimate narrowing, because it only visualizes contrast inside a vessel. Yet; the standard values provided by both methods to identify a stenotic lesion as significant, don't consider the relation between the resulting effective flow distal to the stenosis and the demand of a hypertrophied myocardial tissue.\\
The pathogenesis of ventricular hypertrophy implicates an increase in the number of force-generating sarcomeres in the myocyte \cite{lorell2000left}. According to La Place law; an increase in pressure can be offset by an increase in wall thickness \cite{gunther1979determinants}. This mechanical input is transduced into a biochemical sequel that modifies gene transcription in the nucleus. The focal adhesion complex, in addition to the G-coupled neurohormonal augmentation \cite{sugden1999signaling}, is proposed as the effector transducers of said mechanical input \cite{borg1992holding}. Integrins connect the internal cytoskeleton of the cell to the extracellular matrix, wherein multiple tyrosine-phosphorylated kinases and serine-threonine kinases that are implicated in the signaling of hypertrophy can be found in the ECM \cite{kuppuswamy1997association}\cite{terracio1991expression}.\\
Angiogenesis is triggered during the pathogenesis of myocardial hypertrophy by increased cardiac work and oxygen demand; in an attempt to normalize maximal myocardial perfusion and capillary domains to sustain oxygen delivery. A limitation of capillary growth will increase diffusion distance for oxygen, while inadequate arteriolar growth will reduce maximal tissue perfusion. Pathogenesis of hypertrophy is categorized into pressure overload-induced, volume overload-induced, thyroxin-induced and exercise-induced models according to the stimulus for increasing muscle bulk. In exercise-induced and thyroxin-induced models; angiogenesis and arteriogenesis are well documented experimentally \cite{tomanek2007angiogenesis}. While in other models; there is a considerable variation in the reports of the literature about the extent and pattern of angiogenesis and the consequential coronary microvascular resistance. The reasons for the discrepancy between these studies are not evident, but the duration of the hypertrophy and the specificity of the stimulus may play a role.\\
Mathematically; angiogenesis increases the coronary microvascular resistance structurally, due to the addition of a new microvascular segment. However, in vivo; structural resistance can be modulated by functional changes, wherein autoregulatory adjustments involving the vasodilator reserve may ameliorate said structural resistance escalation. Well-trained athletes with physiological cardiac hypertrophy show a proportional increase of cardiac myocytes and coronary vasculature with no change in the proportion of extracellular collagen \cite{duncker2008regulation}. These structural modulations are accompanied by functional adaptations resulting in a compensatory exponential coronary reserve and vasodilator capacity. Functional adaptations can include changes in neurohumoral control and changes in local vascular control mechanisms \cite{laughlin1985effects}\cite{laughlin2004physical}. In pathological severe hypertrophy; pathological features of the strained neoangiogenesis halt the functional compensation for the structural increase in the microvascular resistance. Endothelium-dependent vasodilation is markedly impaired in the coronary microvessels of patients with hypertension-induced ventricular hypertrophy \cite{treasure1993hypertension}. Accordingly; severe ventricular hypertrophy is associated with a reduction in coronary vascular reserve \cite{pichard1981coronary}.\\
Myocardial infarction is mainly caused by rupture of vulnerable fibroatheromatous plaque forming a thrombus that interferes with myocardial blood supply leading to excessive ischemia then necrosis \cite{reed2017acute}. Usually; soft non-stenotic plaques are more susceptible to rupture, causing major cardiovascular events \cite{libby2013mechanisms}. The vulnerability of the plaque depends on lesion-specific characteristics like thin fibrous cap, large lipid-rich necrotic core, increased plaque inflammation, positive vascular remodeling, increased vasa-vasorum neovascularization, and intra-plaque hemorrhage \cite{moreno2010vulnerable}. Therefore; the comorbidity between CAD and ventricular hypertrophy doesn't affect the risk of developing MI. However; patients with said comorbid diseases have higher rates of transition from ischemia to necrosis in the affected endocardial tissue, due to increase diffusion distance for oxygen and exhaustion of the functional compensation. They are also more susceptible to and usually presented by NSTEMI, due to the sensitivity of the endocardial myotissue to the equilibrium between the structural resistance of microvasculature and the reactive functional modulation.\\
Myocardial ischemia is characterized by ionic and biochemical alterations, creating an unstable electrical substrate capable of initiating and sustaining arrhythmias \cite{ghuran2001ischaemic}. Theoretically; onerous angiogenesis in pathological hypertrophy shows patterns of anatomical and architectural dysgenesis rendering foci within the hypertrophied muscle bulk at greater risk of ischemic injury. Said stressed foci acquire different electrochemical properties, due to defective function of ATPase-dependent pumps, leading to tissue heterogeneity. Theses foci become arrhythmogenic, especially with increased cardiac demand during above-normal exercise or severe stressful conditions, leading to functional re-entry. Hence; presence of ventricular hypertrophy is associated with a greater risk of sustained arrhythmias \cite{chatterjee2014meta}.
\section*{Conclusion}
The mathematical model establishes that ventricular hypertrophy increases the vascular structural resistance of the coronary circulation due to neomicroangiogenesis. Patients with comorbid CAD and ventricular hypertrophy suffer, due to exhaustion of functional compensation, from complications of myocardial hypoperfusion at angiographically sub-significant coronary artery stenosis. These patients are more susceptible to NSTEMI, serious arrhythmias and sudden cardiac death than patients with isolated CAD. Upon confirmation of such results by large investigational studies, said results should be taken into account during designing the treatment strategy of said patients.
\section*{Acknowledgement}
The author states that there is no conflict of interest regarding this article.
\thispagestyle{plain}
\bibliographystyle{unsrt}
|
\section{Introduction and background}
Networks can be used to represent public transportation systems from various unique perspectives. Traditionally nodes and edges represent the physical infrastructure of a transportation system, e.g., routes and stops/stations \cite{london}, while recent developments in data collection and modeling allow researchers to accurately map contacts between individuals traveling together. Detailed commuting patterns can be recorded from smart card data \cite{sun,sun2} or can be the output of activity-based travel models \cite{l32,l33,l34,l35,l36}. The resulting passenger contact patterns are both spatial and temporal in nature, and include travel times on specific vehicles and contact with other travelers \cite{l41,l42,l64}. Such detailed travel patterns can then be used for various planning objectives including the estimation of the capacity of infrastructure \cite{WANG2011814}, calculating environmental impact \cite{CARLSSONKANYAMA1999405} or designing surveillance and containment strategies during an epidemic outbreak \cite{l49}. While these methods allow researchers to map contacts between known individuals, the data collection and processing required to recreate a real-world contact network presents many challenges in terms of accuracy and computational complexity, among other issues such as privacy \cite{l43, l44, l45, l46, l47, l48}.
In this work we specifically focus on the community structure of public transit ridership patterns. Observations on social interactions reveal that people tend to form groups according to their lines of interest, occupation, etc. This concept is known as homophily in the social sciences \cite{eagle2009inferring,yuan2006homophily,chin2012using}, while in network modeling we refer to this phenomenon as community structure \cite{fortunato}. One of the most frequently used definitions of this concept was proposed by Newman \cite{newman}: ``In an arbitrary network a {\em community} is a set of nodes where the density of connections between the nodes of the community is greater than the density of connections between communities''. Community detection is a diverse field with many applications in various fields of science. Newman's definition has been extended or replaced by newly proposed algorithms, yet it is still the most intuitive way of describing communities.
The majority of community detection algorithms consider communities as disjoint sets of nodes. Popular approaches include modularity maximization methods \cite{newman,blondel}, information theoretic approaches \cite{infomap}, statistical inference \cite{sbm,wsbm} and spectral techniques \cite{krzakala,newmanspectral}.
Other algorithms allow overlaps between neighboring groups \cite{bota2015high,kertesz,wu}. Community detection algorithms can also be applied to weighted networks, where a value on each edge represents similarity (or distance) between nodes \cite{bota2015high,wsbm}. All weighted networks can be transformed into unweighted forms by setting a threshold and omitting edges with weights below the threshold. This technique has been used in real-world applications \cite{palla2005uncovering,bota2014community}. A selection of excellent reviews on community detection can be found in \cite{fortunato,overlapping,leskovec}.
In this work we propose the development of two novel network structures, namely the {\em transfer network}, and the {\em community network}. The transfer network captures the movement patterns of atomic passenger groups, i.e., groups of people who travel together on one or more vehicles, mathematically defined as maximal complete subgraphs of the contact network. The community network captures the
similarity of travel patterns between passengers, is derived directly from the transfer network, and is constructed using a novel link-based metric called the {\em connection strength}. The community network is intended to reveal groups of travelers who may also be more likely to come into contact outside of their daily travel routine e.g.: colleagues traveling to work or children going to school \cite{yuan2006homophily,chin2012using}. A description of each proposed network structure as well as the contact network can be found in Table 1.
We further demonstrate potential applications of each of these novel network structures. First, the transfer network is implemented to detect the most frequent vehicle trip combinations in the system (bus and train transfers). While there are existing methods in the literature to measure the capacity of public transit systems \cite{cepeda2006frequency,lai2011behavioral}, our approach provides a more refined view of passenger movement between vehicle trips by tracking the movements of groups of passengers traveling together. Identifying the most relevant vehicle trip combinations can aid public transit authorities in timetable planning and optimizing vehicle assignments.
Second, the community network is applied to evaluate a diffusion process in a transit network, specifically infectious disease spread among passengers. This application complements our previous work \cite{nets}, to identify the components of the public transportation system most vulnerable to a bio-security threat. The community network proposed here can further improve the performance of such models due to its novel representation of passenger interactions. We use the proposed network structure to simulate how a disease might spread among the vehicles of the public transportation system and the suburbs of the city, and identify the vehicle trips most likely to carry infected passengers. All applications are evaluated on a real-world case study, the public transit system of Twin Cities, MN, using output from an activity based travel demand model \cite{l64}.
The rest of the paper is structured as follows. In Section 2 we introduce the travel demand model, and how it is used to create passenger contact network. In Section 3 we introduce the transfer network, and illustrate how it can be used to detect frequent vehicle trip combinations. In Section 4 we introduce the community network, including the connection strength value, and illustrate how it can be used to model infectious disease spread. Section 5 contains our conclusions and future research directions.
\begin{table}[!h]
\begin{center}
\caption{Network definitions used in this paper, including the contact network and the proposed transfer and community networks}\label{tab:test}
\begin{tabular}{|p{1,2cm}|p{3cm}|p{3cm}|p{3cm}|} \hline
& Contact network & Transfer network & Community network \\ \hline
Nodes & Passengers & Atomic passenger groups & Passengers traveling on at least two vehicle trips \\ \hline
Edges & Passengers are connected of they are physically present on the same vehicle trip at the same time (undirected) & Atomic passenger groups are connected if they share a passenger (directed)
& Passengers are connected if they are traveling together on at least two vehicle trips (undirected) \\ \hline
Attributes & Contact duration and start time & Number of transfer passengers between atomic groups & Connection strength measuring the similarity of the travel patterns between passengers \\ \hline
\end{tabular}
\end{center}
\end{table}
\section{Model Inputs}
The inputs of our work are based on the transit assignment model published in \cite{l64}.
In this section, we present a description of the assignment model and provide a short summary of the properties of the passenger contact network built from it.
\subsection{Travel demand model}
As described in our previous work \cite{nets}, public transportation data in this study was obtained from the transit system in Twin Cities region in Minnesota, where 187 routes serve 13,700 stops in the region. Transit network and schedule data were created from General Transit Feed Specification (GTFS), including near 0.5 Million stop-times on a weekday in 2015. Transit passenger trips were obtained from Metropolitan Councils's activity-based demand model \cite{alireza}, and contained more than 293,000 linked trips (i.e. a passenger trip from an origin to a destination that may include zero or more transfers). The assignment of transit demand to transit network was done using the FAST-TrIPs model \cite{l64}. FAST-TrIPs is a schedule-based transit assignment model that generates hyperpaths using a defined logit route choice model \cite{alireza2}, assigns individual passengers to the paths using the hyperpath probabilities, and simulates them using a mesoscopic transit passenger simulation module. Since a calibrated transit route choice model was not available for the Twin Cities network, a route choice model from Austin, TX was borrowed from a previous study by the authors \cite{alireza3}. The model specifies the following route choice utility function:
$$u = t_{IV} + 1.77t_{WT} + 3.93t_{WK} + 47.73X_{TR}$$
where $t, t_{IV}, t_{WT}, t_{WK}$ and $X_{TR}$ represent path utility, in-vehicle time, waiting time, walking time, and number of transfers in a transit path, respectively. The output of the transit assignment model contains individual passengers' trajectories, including their walking from the origin to the transit stop, boarding the transit vehicle, alighting and walking to the transfer stop, boarding the next transit vehicle, etc. and finally alighting and walking to the destination. By post-processing the transit assignment model's outputs, the amount of time each pair of passengers are on-board the same transit vehicle were calculated on a daily basis.
\subsection{Contact network}
We define a network structure denoted as the {\em contact network} based on the outputs of the travel demand model. The nodes of the contact network are passengers, and edges connect passengers if they were traveling on the same vehicle at the same time. The relationship is symmetric, therefore the network is undirected. All passenger movements take place in a temporal setting, which is indicated by two values assigned to the edges of the network: the contact start time on an edge indicates the start time of the contact between the two connected passengers, while a contact duration value indicates the length of the contact in minutes. Since each edge represents a connection between a pair of passengers traveling on a specific vehicle, the id of the vehicle can also be assigned to the edges of the graph. The vehicle id identifies a vehicle trip: a single vehicle traveling on a specific route with a specific start time. The basic properties of the network were shown in \cite{nets} along with a detailed analysis regarding the networks potential role in the spreading of epidemics. A short review on the description of the network as appeared in \cite{nets} follows.
\begin{figure}[tbp]
\centering
\includegraphics[scale=0.34]{trb1.png}
\caption{The distribution of a) contact start times and b) the degree distribution of the contact network.}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[scale=0.34]{trb3.png}
\caption{A static subgraph of the contact network. A passenger with a high number contacts is represented by the black node in the middle, connected to all other contacts. The colors indicate of the vehicle trips the passengers first met on. The figure was made with Gephi using the Fruchterman-Reingold algorithm.}
\end{figure}
The contact network corresponding to the Twin cities dataset has $94475$ nodes and $6287847$ contacts between them. Figure 1/a shows the density plot of the contact start times. The distribution has two peaks, one around 7 AM and another around 5 PM, which corresponds to the morning and evening weekday commute. The average number of contacts per person is 136 during the observed day, while the maximum is 827. Figure 1/b shows the degree distribution of the graph.
Figure 2 shows a subgraph of the network structure. Nodes represent passengers and edges connect them if they ride on the same vehicle trip. The dynamics of the network are omitted in this example, i.e. edges are aggregated over the entire day. The black node in the middle represents a passenger with a high number of contacts, who traveled together with all the other nodes shown on this subgraph. The other nodes are colored according to the vehicle trip they rode on, while darker edges indicate longer contact durations.
\section{Transfer network}
The first objective of this paper is to detect the movement patterns of passenger groups.
To do this, we define a novel network structure denoted as the {\em transfer network}. The transfer network is a directed network, where nodes represent the atomic passenger groups of the contact network and groups are connected if at least one member of a group transfers from one vehicle to another. The weight of the edges denote the number of transfer passengers.
In order to build the transfer network we identify the subgraph corresponding to each vehicle trip, detect atomic passenger groups -- defined as maximal cliques -- on each of the resulting subgraphs of the contact network and connect the atomic passengers groups according to direction of transfer between vehicle trips. We weight the connections based on the number of transferring passengers. Section 3.1 defines the construction of the transfer network in more detail.
The transfer network proposed in this paper provides a refined way to identify the most frequent vehicle trip combinations in the public transit system, which can aid decision makers in defining timetables and optimizing vehicle assignments. Section 3.2 discusses this application.
\subsection{Transfer network construction}
The three steps required to build the transfer network are discussed in the subsequent subsections. Section 3.1.1 defines how the subgraphs corresponding to the vehicle trips are constructed, 3.1.2 defines atomic passenger groups as cliques and show how they can be detected in an efficient way and 3.3.3 shows how the transfer network can be built from the passenger groups.
\subsubsection{Graph partitioning}
We define the subgraph corresponding to each vehicle trip as follows. Let $G$ be the original contact network, $V(G)$ the vertices and $E(G)$ the edges of the network. We divide the original network into subgraphs along vehicle trips. Let $T$ be the set of all vehicle trips in the network and let $t_i \in T$ denote the $i$-th trip. We define $G_{t_i}$ as the subgraph corresponding to trip $t_i$ where, $V(G_{t_i})$ and $E(G_{t_i})$ are the vertices and the edges of trip $G_{t_i}$. We create a subgraph $G_{t_i}$ for all trips $t_i \in T$. Since a single passenger may travel on multiple trips, the node corresponding to the passenger may appear in multiple subgraphs. Figure 3 shows an example of the partitioning process.
\begin{figure}[h]
\centering
\includegraphics[width=1\textwidth]{partitioning}
\caption{Partitioning of an example graph along vehicle trips. Each vehicle trip has a corresponding subgraph where nodes are passengers who used the given vehicle trip.}
\end{figure}
\subsubsection{Clique detection}
In graph theory, a {\em clique} is defined as a fully connected subgraph of a given graph. A {\em maximal clique} is a clique, that is not a subgraph of any other clique. Finding the set of all maximal cliques is a well-studied NP-hard problem in graph theory. An arbitrary $n$-vertex graph may have up to $3^{n/3}$ maximal cliques, but this number is much lower in many complex networks, including the contact network studied in this paper. Among the available methods in clique detection we adopt the Bron-Kerbosch (BK) algorithm \cite{Bron:1973:AFC:362342.362367}, which has proven its reliability in real-life applications \cite{eppstein2010listing}. We apply the BK algorithm\footnote{We implemented the Bron-Kerbosch algorithm with pivoting and degeneracy ordering in the outer most level of recursion in C++. We ran the algorithm on a PC with I7 4790 CPU (3.6 Ghz) and 16GB of RAM. } to detect the maximal cliques of all subgraphs $G_{t_i}$ corresponding to all vehicle trips $t_i \in T$.
The analysis can be further extended if we consider the graphs as weighted according to the contact durations available on the edges. Introducing a weight threshold $\tau$ we can prune the edges of the graphs by omitting all edges with contact durations below $\tau$. This allows us to redefine what a connection means in the network: passengers are only considered to be connected if they spend at least a certain amount of time on the same vehicle.
We define three thresholds to represent the strength of connection between passengers $\tau_5 = 5$ minutes, $\tau_{15} = 15$ minutes and $\tau_{30} = 30$ minutes. We chose these thresholds to represent short, medium and long duration contacts between passengers.
We prune the edges of the contact network by these values resulting in graphs $G_5$, $G_{15}$ and $G_{30}$. In additional, let $G_0$ and $\tau_0$ represent the original (uncut) graphs and the corresponding threshold. We run the Bron-Kerbosch clique detection algorithm on these graphs and analyze and compare results. The speed of the detection algorithm is amplified by the fact, that all subgraphs corresponding to vehicle trips are interval graphs. In order to show this speedup, we also run the clique detection algorithm on the original contact network and compare results. Table 2 shows graph size, runtime of the clique detection method on the original contact network and the total runtime of the detection method on all subgraphs for $G_0$, $G_5$, $G_{15}$ and $G_{30}$. We can see a significant speedup for all graphs in this analysis. For the larger $G_0$ and $G_5$ graphs the total runtime on all subgraphs corresponding to vehicle trips is 14 times less than on the unpartitioned network.
\begin{table}[!htb]
\begin{center}
\caption{ Runtimes of the Bron-Kerbosch on graphs $G_0, G_5, G_{15}, G_{30}$. Raw denotes runtime in seconds on the original contact network, while partitioned denotes the total runtime on all subgraphs corresponding to vehicle trips.}\label{tab:test}
\begin{tabular}{r|c c c c c c} \hline
\begin{tabular}[x]{@{}c@{}}Graph\end{tabular}& \begin{tabular}[x]{@{}c@{}}Passengers\end{tabular}&\begin{tabular}[x]{@{}c@{}}Edges\end{tabular}&\begin{tabular}[x]{@{}c@{}}Trips\end{tabular}&\begin{tabular}[x]{@{}c@{}}Time\\raw (s)\end{tabular}&\begin{tabular}[x]{@{}c@{}}Time\\partitioned (s)\end{tabular}\\ \hline
$G_0$ & 94475 & 6435482 & 8002 & 5285 & 367 \\
$G_5$ & 91894 & 5203557 & 7934 & 3830 & 258 \\
$G_{15}$ & 63714 & 1630522 & 6969 & 1195 & 50 \\
$G_{30}$ & 26154 & 218233 & 4083 & 129 & 6 \\
\end{tabular}
\end{center}
\end{table}
\subsubsection{Graph building}
Let $F$ denote the {\em transfer network} as a directed graph where every node $v \in V(F)$ is a clique from $G_{t_i}$ for all $t_i$. Edges connect nodes $v$ and $u$ if the corresponding cliques do not lie in the same vehicle trip, yet they have at least one common passenger. More formally, for nodes $u$ and $v$ in the transfer network and their corresponding cliques $c_v$ and $c_u$ in the contact network, $u$ and $v$ is connected if $c_v \cap c_u \geq 1$ and if $c_v \subseteq G_{t_v}, c_u \subseteq G_{t_u}$ then $G_{t_v} \neq G_{t_u}$.
The {\em direction of edges} correspond to the direction of the transfer between $t_u$ and $t_v$, that is whether the passengers of $c_v \cap c_u$ move from $t_u$ to $t_v$ or the opposite. We establish the direction by looking at the contact start times for the individuals in $c_v \cap c_u$ in the following way. For all edges of the transfer network $e_{uv} \in E(F)$, let $c_{uv}$ denote the set of corresponding passengers in $G$ as $c_{uv}=c_u \cap c_v, c_v \subseteq G_{t_v}, c_u \subseteq G_{t_u}$, $G_{t_v} \neq G_{t_u}$. Let $\alpha_{xy}^{t_{i}}$ denote the contact start time between passengers $x$ and $y$ on vehicle trip $t_i$. For all passengers $p,q \in c_{uv},$ if $min(\alpha_{pq}^{t_{u}}) < min(\alpha_{pq}^{t_{v}}) $, then the direction of the edge $e_{uv} \in E(F)$ is from $u$ to $v$, else it is from $v$ to $u$. If $|c_v \cap c_u| = 1$, then let $c_v \cap c_u = \{p_0\}$ and $p \in c_u \setminus \{p_0\}, q \in c_v \setminus \{p_0\}$. Then $min_p(\alpha_{p_{0}p}^{t_{u}}) < min_q(\alpha_{p_{0}q}^{t_{v}})$ decides the direction of the edge as above.
We assign integer values $w_v = |c_v|$ to all $v \in V(F)$ and $w_{u,v} = |c_v \cap c_u|$ for all $e(u,v) \in E(F)$. Values $w_v$ and $w_{u,v}$ represent the amount of passengers corresponding to both the nodes and edges of the transfer network.
\subsection{Detecting frequent vehicle trip combinations}
The {\em transfer network} allows us to identify the most frequent vehicle trip combinations passengers travel on in the public transportation system.
Detecting the most frequent combinations helps decision makers in defining timetables and optimizing vehicle assignments.
As before, let $T$ be the set of vehicle trips, and $t_i \in T$ the i-th vehicle trip. We define vehicle trip pairs as follows: for all $t_i$ and $t_j$, $t_{ij}$ is a vehicle trip pair where $i \neq j $ and $t_i,t_j\in T$, while the set of all vehicle trip pairs is denoted by $T^{p}$. We assign a number $m_{ij}$ to each vehicle trip pair $t_{ij} \in T^{p} $ indicating the amount of passenger traffic between the corresponding trips. To calculate this value let the set $V(C_{t_{ij}})$ contain all nodes of the transfer network corresponding to all cliques $c_{t_{ij}} \in C_{t_{ij}}$ where $c_{t_{ij}} \in G_{t_i} \cup G_{t_j}$, and take the subgraph induced by $V(C_{t_{ij}})$. The number $m_{ij}$ is the sum of all edge weights of the induced subgraph. This approach offers a more refined view of passenger traffic between vehicle trips because it represents the movements of passenger groups as opposed the behavior of individual passengers.
To give another dimension to our analysis we compared the frequent vehicle trip combinations in both $G_0, G_5, G_{15}$ and $ G_{30}$, that is only considering passengers to be in contact with each other if they travel together for more than $\tau_5 = 5$ minutes, $\tau_{15} = 15$ minutes and $\tau_{30} = 30$ minutes in addition to the unfiltered network $G_0$. Table 3 shows the five most frequent vehicle trip combinations $t_{ij}$ and the amount $m_{ij}$ of passenger traffic between them.
\begin{table}[!h]
\begin{center}
\caption{The five most frequent vehicle trip combinations in $G_0, G_5, G_{15}$ and $G_{30}$. The first number indicates the route number followed by the start time of the sepcific vehicle trip.}\label{tab:test}
\begin{tabular}{r|c c c|r|c c c} \hline
Graph & $t_i$ & $t_j$ & $m_{ij}$ & Graph & $t_i$ & $t_j$ & $m_{ij}$ \\ \hline
$G_0 \ 1.$ & 614/5:25 & 675/5:30 & 38 & $G_{15} \ 1.$ & 68/6:19 & 94/7:04 & 21\\
$2.$ & 68/5:46 & 94/6:25 & 35 & $2.$ & 94/18:32 & 68/18:52 & 19 \\
$3.$ & 68/6:19 & 94/7:04 & 27 & $3.$ & 68/5:46 & 94/6:24 & 18 \\
$4.$ & 901/6:18 & 415/7:02 & 25 & $4.$ & 614/5:16 & 675/5:35 & 16 \\
$5.$ & 415/17:12 & 901/17:33 & 23 & $5.$ & 294/5:38 & 94/6:44 & 15 \\ \hline
$G_5 \ 1.$ & 614/5:25 & 675/5:30 & 37 & $G_{30} \ 1.$ & 19/17:22 & 850/18:03 & 8 \\
$2.$ & 68/5:46 & 94/6:25 & 35 & $2.$ & 54/18:02 & 68/18:23 & 4 \\
$3.$ & 68/6:19 & 94/7:04 & 27 & $3.$ & 71/6:07 & 61/6:38 & 4 \\
$4.$ & 901/6:18 & 415/7:02 & 25 & $4.$ & 5/0:30 & 901/2:07 & 4 \\
$5.$ & 415/17:12 & 901/17:33 & 23 & $5.$ & 902/19:25 & 68/19:52 & 4 \\ \hline
\end{tabular}
\end{center}
\end{table}
Almost all trip combinations in Table 3 follow a similar pattern. Results for $G_0$ and $G_5$ are nearly identical. The most frequent combinations for $G_{15}$ connect two additional suburbs (Oakdale and Stillwater) to $G_0$ and $G_5$, but otherwise are identical. The vehicle trips pairs with the greatest amount of passenger traffic between them are mostly in the morning or afternoon peak hours. Almost all of the trip combinations link one of the outlying suburbs to the city center, indicating the daily commuting patterns of workers or students. Just the urban area of Twin Cities covers 2646 $km^2$ with a population of more than three million. This means that commuters trying to reach the city center from one of the outlying suburbs must travel on two or sometimes three separate routes to get to their destination. The pattern -- also shown on Figure 4 -- is the following. Commuters start the journey from one of the {\em smaller outlying suburbs} like Ridgedale (route 614), Inver Grove and other suburbs south of St. Paul (route 68) or Oakdale and Stillwater (route 294). Then they change vehicles in the transport hub of one of the {\em major outlying suburbs} (St. Paul, Minnetonka, Mall of America in Bloomington), and take an express service to the {\em city center} (routes 94, 675, 850, etc..).
While the frequent combinations for $G_{30}$ are similar, there are differences. While the suburbs these routes connect are different, we can see almost the same patterns as before: travel from one of the smaller outlying suburb to a major one and then to the city centre (combinations 71--61 and 5--901). One specific route (850) is one of the longest express bus route in the city and it includes a long section where the bus doesn't stop at all. One exception to the pattern is route 54, which connects St. Paul International airport to St. Paul and route 68 connecting to south St. Paul. In terms of time in addition to peak hours, we see late night services as well.
\begin{figure}[h]
\centering
\includegraphics[width=1.0\textwidth]{Fig4}
\caption{Most frequent vehicle trips combinations in Twin Cities, MN for $G_0, G_5, G_{15}$ and $ G_{30}$. }
\end{figure}
We can summarize, that graphs $G_0, G_5$ and $G_{15}$ behave similarly, showing the movements of passenger groups traveling through the public transit system. The most frequent trip combinations are the ones connecting distant suburbs to the city center during the morning and afternoon peak hours. $G_{30}$ captures an alternative set of routes corresponding to people who travel together for longer periods for different reasons, like a long service (route 850), the scarcity of services late at night (route 5 at 0:30) or traveling from the airport to a major transport hub (route 54).
\section{Community network}
Next we propose a novel network structure, the {\em community network}, which expands the definition of passenger connectivity to be a function of both the number of transfers passengers make together as well as the total amount of time they spend together while traveling. This contrasts the contact network, which simply quantifies passenger connectivity passengers using contact duration on individual vehicle trips. We define a {\em community of passengers} as a set of passengers who have common travel patterns, e.g., vehicle trips and/or transfers.
In order to build communities, we define and quantify a novel {\em connection strength} metric between passengers indicating the similarity of their travel patterns and create a new network structure, the {\em community network} based on this value. Section 4.1 outlines this process in more detail.
This network can serve as the basis of a community detection algorithm, but in this paper we take a different approach. We define communities as those connected by edges whose values lie above a predetermined threshold. In section 4.2 we demonstrate this feature by identifying the commuting patterns of the members of the largest passenger community of the network.
We propose an application of the community network in section 4.3. We show how the connection strength value can be used to model infectious disease transmission among passengers of a public transit system. Tying to our previous work in \cite{nets} we seek to identify the vehicle trips most likely to carry infected passengers during an outbreak.
\subsection{Community network construction}
In order to detect the communities of the public transportation network we construct a weighted network structure called {\em community network}.
The community network connects passengers using on a novel link-based metric, the {\em connection strength}.
The connection strength $s$ defines the edge weights in the community network, and takes into account the number of transfers a pair of passengers makes together. Thus, if the passengers meet on multiple different vehicle trips, their connection strength $s$ will increase. This method is based on the assumption that passengers who not only travel together but also transfer together have a stronger connection than travelers who are simply present on the same vehicle trip at the same time. Below we explain how this link metric is derived.
Let $H$ denote the new network where the nodes are the passengers, and the set of nodes V(H) includes all passengers who traveled on at least two vehicle trips with another passenger. Thus, the nodes in the community network correspond to the edges of the transfer network. We define the connections and weights between passengers as follows. Nodes $u$ and $v$ in the community network are connected if they are both present in at least two different cliques $c_1, c_2$ in two different subgraphs corresponding to vehicle trips, \textit{i.e.} they traveled together on at least two vehicle-trips ($u,v \in c_1 \subseteq G_{t_1}$ and $u,v \in c_2 \subseteq G_{t_2}$). Let $g_{uv}$ denote the number of instances where $u$ and $v$ are members of the same clique, that is $g_{uv} = |C_{uv}|$ where $c_{uv} \in C_{uv}$ if $u,v \in c_{uv}$.
Let $T_{uv}$ be the set of {\em vehicle trips} where both $u$ and $v$ are present: $t_{uv} \in T_{uv}$ if $u,v \in t_{uv}$, and let $g_{uv}^{t_{i}}$ be the number of the cliques in vehicle trip $t_{i}$ where $u$ and $v$ are both present: $g_{uv}^{t_{i}} = |C_{uv}^{t_{i}}|$ where $c_{uv}^{t_{i}} \in C_{uv}^{t_{i}}$ if $u,v \in c_{uv}^{t_{i}} \in G_{t_i}$. Thus, the connection strength $s_{uv}$ between passengers $u$ and $v$ can be formalized as follows:
\begin{equation}
s_{uv}=\frac{g_{uv}*(g_{uv}-1)}{2}-\sum_{t_{i} \ in \ T_{uv}}\frac{g_{uv}^{t_{i}}*(g_{uv}^{t_{i}} -1)}{2}
\end{equation}
Using this definition of connection strength, if a pair of passengers traveled together in $g_{uv}$ different atomic passenger groups, then they would have $\frac{g_{uv}*(g_{uv}-1)}{2}$ different edges between them because the affected nodes form a clique in the transfer network. This way the first part of the equation rewards the movements between different vehicle trips. Since based on our community definition traveling on the same vehicle trip doesn't indicate strong connection between the passengers, in the second part of the equation we penalize any instance where $u$ and $v$ travels together on the same vehicle trip. The value of the penalty will be the sum of the edges in every vehicle trip where the passenger pair appears more than one time, and the number of the edges in a vehicle trip is counted in the same way as in the first part of the equation.
\begin{algorithm}[h]
\caption{ Construction of the {\em community network}}
\renewcommand{\algorithmiccomment}[1]{\hfill {\it #1}}
\begin{algorithmic}[1]
\STATE $V(H) \leftarrow Every \ passengers \ from \ E(F) \ sections $ \
\STATE $E(H) \leftarrow \emptyset$
\STATE \textbf{For} in $E(F) sections$ \textbf{Do}
\STATE\hspace{\algorithmicindent} \textbf{For} $u \ v $ in every passenger pairs \textbf{Do}
\STATE\hspace{\algorithmicindent}\hspace{\algorithmicindent} \textbf{If} {$e(u,v) \notin E(H) $}
\STATE\hspace{\algorithmicindent}\hspace{\algorithmicindent}\hspace{\algorithmicindent} $E(H) \leftarrow e(u,v)$
\STATE\hspace{\algorithmicindent}\hspace{\algorithmicindent} \textbf{Else}
\STATE\hspace{\algorithmicindent}\hspace{\algorithmicindent}\hspace{\algorithmicindent} $s_{uv}++$
\STATE\hspace{\algorithmicindent}\hspace{\algorithmicindent} \textbf{End if}
\STATE\hspace{\algorithmicindent} \textbf{End for}
\STATE \textbf{End for}
\end{algorithmic}
\end{algorithm}
\begin{figure}[h]
\centering
\includegraphics[width=1.0\textwidth]{example1}
\caption{Examples of the connection strength between pairs of passengers in three different travel scenarios.. Rectangles represent vehicle trips and circles represent cliques. Edges marked with black increase the connection strength between highlighted passengers. Red edges penalize connection strength because these are on the same vehicle trip. Figure 5/a: two passengers travel together in two cliques on vehicle trip $t_1$ and one clique on vehicle trip $t_2$, therefore $ g_{uv}^{t_{1}} = 2, \ g_{uv}^{t_{2}} = 1, \ g_{uv}=3$ and $s=2$. Figure 5/b: two passengers travel together in three cliques on $t_3$ and two cliques on $t_4$ making $g_{uv}^{t_{3}} = 3, \ g_{uv}^{t_{4}} = 2, \ g_{uv}=5$ and $s=6$. Figure 5/c: two passengers travel together on a single vehicle trip in four cliques making $ g_{uv}^{t_{1}} = 4, \ g_{uv}=4$ and $s=0$.}
\end{figure}
Algorithm 1 shows the construction of the {\em community network}, while Figure 5 illustrates a few examples for computing $s$ between passengers.
On Figure 5/a two passengers travel together in two cliques on vehicle trip $t_1$ and one clique on vehicle trip $t_2$, therefore $ g_{uv}^{t_{1}} = 2, \ g_{uv}^{t_{2}} = 1, \ g_{uv}=3$ and $s=2$. A different situation is shown on 5/b where two passengers travel together in three cliques on $t_3$ and two cliques on $t_4$ making $g_{uv}^{t_{3}} = 3, \ g_{uv}^{t_{4}} = 2, \ g_{uv}=5$ and $s=6$. Figure 5/c presents a trivial case, when two passengers travel together on a single vehicle trip in four cliques making $ g_{uv}^{t_{1}} = 4, \ g_{uv}=4$ and $s=0$.
\subsection{Passenger communities}
In this section we illustrate examples of passenger communities in the public transportation system of Twin Cities MN. The first example seen on Figure 6 shows subgraphs of the community network constructed from $G_{30}$, \textit{i.e.} the contact network only containing edges where contact duration is above 30 minutes. Figure 6/a depicts the entire community network. Most of the communities on this network are of size two or three, but there are several larger communities with strong connections between the members. Figure 6/b shows a subgraph where edges with weights $s<5$ are omitted as well as all nodes with degrees below two. The remaining subgraph contains the largest group of the network, while the largest individual community is depicted on 6/c.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.9\textwidth]{collage2}
\caption{The community network of Twin Cities, MN. a) the whole community network, b) a subgraph with edge weights greater than 5, c) the largest passenger group of the network.}
\end{figure}
Figure 6/c shows the largest group in the network. The group contains nine passengers who traveled together on two different vehicle trips, while the overall time they spent using the public transportation network was almost 1.5 hours. The passengers embarked on route 805 in the morning between 7:08 and 7:16 near Blaine and traveled together to Northtown. They disembarked at 7:48 and waited together for the second vehicle 852 arriving at 8:12 and traveled together to downtown Minneapolis for almost an hour until 9:12 and 9:16. The travel path of the community, shown on Figure 7, indicates a commuting pattern from one of the suburbs to the city center of Minneapolis.
\begin{figure}[h]
\centering
\includegraphics[width=1.0\textwidth]{Fig7}
\caption{The travel path of the passenger community on Figure 6/c traveling from a suburb to the city center.}
\end{figure}
Both contact and community networks reveal contact patterns between passengers of a public transportation system. In the contact network, the weight between individual passengers is defined based on the contact duration on individual vehicle trips. In contrast, the community network provides a more refined way to represent connection strength which takes into account the amount of transfers passengers take together.
The underlying concept behind community structure in networks is homophily, which is a well-studied concept of the social sciences \cite{eagle2009inferring,yuan2006homophily,chin2012using}. Homophily states that people tend to form groups according to their lines of interest, occupation, etc. Physical proximity -- on public transportation for example -- is one of these indicators. Therefore, strong connections on the community network can help us uncover connections in other areas of life like workplace or school or other common interests. It should be noted, that physical proximity does not guarantee another type of connection, it simply increases the likelihood of occurrence for it.
\subsection{Epidemic spreading risk application}
One application of identifying the communities within the transit network, as described in this work, is infrastructure security. Understanding passenger communities enables more efficient and accurate tracking of infectious disease spread, were one to be naturally or maliciously introduced into the public transit system.
One of the challenges in modeling epidemic spreading is accurately mapping the relationships between individuals traveling on the same vehicle.
A traditional contact network as defined in section 2 as well as in \cite{nets,sun2,trb} simply revelas the set of passengers who were present on the same vehicle and the amount of time they spend on the same vehicle. As shown in \cite{trb}, this limits the options for how infection spreading probabilities can be defined to be simply a function of contact duration, without the possibility to take into account physical proximity, communication etc. between people.
In contrast, the weights of the community network reveals a deeper level of connections between travelers, as passengers with strong connections in this network may also be connected in other areas of life. Passenger groups identified in the community network are more likely to be traveling within close physical proximity of each other and interacting with each other \cite{eagle2009inferring,yuan2006homophily,chin2012using}. As a consequence, the probablity of disease transmission between the travelers belonging to the same community is greater than between two passengers simply sharing the same vehicle without any other connection. This is especially true for large public transportation vehicles like trains or trams, where simply being present on the same vehicle may not imply any kind of connection at all.
In this section we expand upon our previous work in \cite{nets} which examined epidemic spreading risk in the same public transportation network (Twin Cities, MN) with the goal of identifying the vehicle trips most likely to carry infected passengers. The analysis was presented in two parts. First, the passenger contact network, in the same form as in Section 2, was used to model a variety of outbreak scenarios. The scenarios differed in the number and distribution of initially infected passengers and the level of infectiousness, represented as the risk of spreading between passengers. The spreading risk was defined as the contact duration multiplied by a constant value. The scenarios were compared in terms of their impact on the network and confirmed previous observations in the literature, that is that the most central vehicle trips in the public transportation network are also the most susceptible to infection. The second part of the analysis focused on a newly proposed vehicle trip network, which represents the public transit network as a network of vehicle trips instead of passengers. We showed that centrality metrics on the vehicle trip network provided a more efficient way to detect the set of vehicle trips most susceptible to disease, and this estimation can be done much faster than running simulations on the contact network.
In the rest of this section we present an alternative way to model the risk of disease spreading between passengers, which exploits the community network structure introduced in this work. We define the infection transmission probability between a pair of passengers to be a function of both the connection strength value used in the community network and contact duration. We believe this multidimensional transmission probability more accurately represents the level of connectivity between pairs of passengers. Using these new transmission functions, we implement similar spreading scenarios to the ones in \cite{nets} and identify and rank the vehicle trips susceptible to disease spreading. Below we define the new transmission probability function and the infection model, then present the new vehicle-trips identified to be at highest risk.
\subsubsection{Experiment setup}
In order to simulate an epidemic outbreak on the contact network we use the well-known discrete compartmental susceptible-infected (SI) model. In the SI infection process all nodes adopt one of two available states: susceptible (S) or infected (I). Real values denoted as edge infection probabilities are assigned to the links of the network and denoted as $w_e \in [0,1]$. The infection spreads in a network when susceptible nodes adopt an infected state. This is done in discrete time steps in an iterative manner starting from an initially infected set of nodes. In each iteration each infected node tries to infect its susceptible neighbors according to the transmission probability of the link connecting them. If the attempt is successful, the susceptible node is transformed into an infected one in the next iteration. In this work we limit the number of iterations to five, representing a complete work week with recurrent commuting patterns. The inputs of this model are: 1. a contact network of individuals, 2. an assignment of weights to the links of the network which represent the infection transmission probabilities and 3. the set of initially infected nodes, e.g., individuals.
We use the original structure of the contact network, which connects all pairs of passengers that travel on a vehicle-trip together as the basis of this experiment. The link weights are computed in the following way. Since the nodes and links of the community network are a subset of the nodes and links of the contact network, if a link between two nodes is present in the community network we will assign its connection strength value to the corresponding link in the contact network. We do this in all cases where such an assignment can be made. In order to account for extreme outliers of connection strength we cap all values at 100, then rescale all values to the interval of $0.1$ and $0.8$ using a standard feature scaling method. The 0.8 upper bound is set because a transmission probability of 1.0 is assumed to be unrealistic. For all links of the contact network that do not have a corresponding link in the community network we assign a uniform infection value of $0.05$. In contrast to the duration-based value used in \cite{nets}, this enables us to capture the increased spreading risk between passengers sharing similar travel patterns, while still allowing disease to spread between travelers simply sharing a vehicle. Future work will explore the model's sensitivity to the uniform infection assignment using in this work.
Following the procedure in \cite{nets}, we randomly select 100 passengers from the network to be initially infected. Due to the probabilistic nature of the simulation model, we run the SI infection model $k = 10000$ times to quantify the likelihood of each nod being in an infectious state at the end of the fifth time step.
\subsubsection{Vehicle trip ranking}
The infection model constructed above provides the likelihood of infection for each passenger in the contact network at the end of the simulation, i.e., after five days. As in \cite{nets}, we compute a similar infection value for vehicle trips by summing the probability of infection for all passengers on a given vehicle trip. While this does not represent a probability value in a strict sense, this value is proportional to the risk of getting infected on a given vehicle trip.
The routes which contain the highest risk vehicle trips, and corresponding travel times identified in this study are presented in Table 4. Figure 8 shows the routes on the map of the city. Coinciding with the findings in Section 3 and also in \cite{nets}, the vehicle trips are in the morning and afternoon peak hours. Also coinciding with the results in \cite{nets} all of the high risk routes identified go through the city center.
Since the goal of Section 3.2 -- identifying the most frequent trip combinations -- and the goal of this section -- identifying the most risky vehicle trips in the case of an epidemic outbreak -- are different, the difference between the selected vehicle trips are not surprising. We observe some similarities between the "highest risk" vehicle trips identified here, and the most frequent vehicle trip combinations (identified in section 3.2). For example, route 18 connects Bloomington with the city center, while route 61 connects St. Paul to the same destination. We have seen in Section 3.2, that these target-destination pairs are present in the most frequent combinations as well. As these routes connect the most important parts of the city, some similarity is expected.
\begin{table}[!h]
\begin{center}
\caption{Ranking of vehicle trip where infection is most likely to appear.}\label{tab:test}
\begin{tabular}{r|c c} \hline
Rank & Route number & Start time\\ \hline
$1.$ & 25 & 6:52 \\
$2.$ & 17 & 6:13 \\
$3.$ & 25 & 5:11 \\
$4.$ & 14 & 16:54 \\
$5.$ & 5 & 5:36 \\
$6.$ & 25 & 7:16 \\
$7.$ & 18 & 6:07 \\
$8.$ & 11 & 15:52 \\
$9.$ & 11 & 17:14 \\
$10.$ & 61 & 6:07 \\ \hline
\end{tabular}
\end{center}
\end{table}
The set of high risk vehicle trips identified here partially correspond to those identified in \cite{nets}. Specifically, the methodology proposed in \cite{nets} identifies vehicle trips on the routes 7, 9, 11, 14, 17, 25, 61 and 94 as most at risk. These are routes crossing or connecting to the city center in the peak hours, so even though they do not completely match those in Table 4, the pattern they present is the same.
The similarity in the set and type of routes identified in both studies points to the critical role of the network structure in modeling outbreak risk. Future research will continue to build on this application, and further explore the robustness and sensitivity of the proposed methodology. The relevant sensitivity analysis is, however, outside the scope of this work.
\begin{figure}[h]
\centering
\includegraphics[width=1.0\textwidth]{FigX}
\caption{Vehicle trips most likely to carry infected passengers in the public transit system of Twin Cities, MN.}
\end{figure}
\section{Conclusions}
In this paper we proposed two novel network structures to detect and quantify the movement patterns of passengers of a public transit system. The transfer network tracks the movements of atomic passenger group while the community network links passengers with similar travel patterns. We presented applications for both networks, specifically to identify the most frequently used travel paths, i.e., routes and transfers, and model epidemic risk posed by passengers of a public transit network, respectively.
The transfer network was demonstrated as a tool to efficiently detect the most frequently used vehicle trip combinations in the public transportation system. We have shown that the most frequent combinations follow a similar pattern. The trip pairs identified using the transfer network identified peak morning and afternoon trips that connect the outlying suburbs with the CBD, which corresponds to existing literature.
We further illustrated how the community network can be used to augment the traditional contact network structure, and reveal components of the transit system at risk from an epidemic outbreak. Similar to \cite{nets} we sought to identify the vehicle trips most likely to carry infected passengers. However, in contrast to the model presented in \cite{nets} where infection probability between passengers was solely based on contact duration, here we used the connection strength value to estimate physical contact between passengers in an alternative way.
Our results reinforce previous observations, that vehicle trips crossing or connecting to the city center pose the highest risk.
This study is subject to certain limitations. We note, that the transit demand model used as an input of both works was used to generate commuting patterns for single workday. This limitation is most critical for the epidemic application due to the implicit assumption that daily travel patterns remain constant during a five day work week. While this assumption is supported by a recent study \cite{sun2}, more long term observations potentially involving weekends and public holidays would improve the quality of the results presented in this paper.
Further, while we present an alternative metric to quantify disease transmission risk between passengers (compared to contact duration alone), without any real-world observations on an outbreak validating the results of this work remains a challenge, and will be the focus of future research.
\begin{acknowledgements}
The authors are grateful to Metropolitan Council of Twin Cities for sharing the activity-based travel demand model with researchers at the University of Minnesota. Any limitations of this study remains the sole responsibility of the authors.
L\'aszl\'o Hajdu acknowledges the support of the National Research,
Development and Innovation Office - NKFIH Fund No. SNN-117879.
Mikl\'os Kr\'esz acknowledges the European Commission for funding the
InnoRenew CoE project (Grant Agreement \#739574) under the Horizon2020
Widespread-Teaming program and the support of the EU-funded Hungarian grant
EFOP-3.6.2-16-2017-00015
\end{acknowledgements}
\bibliographystyle{spmpsci}
|
\section{Introduction}\label{sec:Intro}
Diversity coding, node repair, and security are three basic ingredients of modern distributed storage systems. The interplay of all three ingredients is captured by a fairly general mathematical model known {\em multilevel diversity coding with secure regeneration (MDC-SR)} \cite{Shao-Arxiv17}.
More specifically, in an $(n,d,\ell)$ MDC-SR problem, a total of $d-\ell$ independent files $\mathsf{M}_{\ell+1},\ldots,\mathsf{M}_d$ of size $B_{\ell+1},\ldots,B_d$, respectively, are to be encoded and stored in $n$ distributed storage nodes, each of capacity $\alpha$. The encoding needs to ensure that:
\begin{itemize}
\item (Diversity coding) the file $\mathsf{M}_j$ can be perfectly recovered by having full access to any $j$ out of the total $n$ storage nodes for any $j\in\{\ell+1,\ldots,d\}$;
\vspace{4pt}
\item (Node repair) when node failures occur and there are $d$ remaining nodes in the system, any failed node can be recovered by downloading data of size $\beta$ from each one of the remaining nodes;
\vspace{4pt}
\item (Security) the files $\mathsf{M}_{\ell+1},\ldots,\mathsf{M}_d$ needs to be kept {\em information-theoretically} secure against an eavesdropper, which can access the repair data for $\ell$ compromised storage nodes.
\end{itemize}
Setting $\ell=0$, the above problem reduces to the problem of {\em multilevel diversity coding with regeneration (MDC-R)} considered in \cite{Tian-IT16,Shao-CISS16}. Setting $B_j=0$ for all $j\neq k$, the above problem reduces to the $(n,k,d,\ell)$ {\em secure regenerating code (SRC)} problem considered in \cite{Pawar-ISIT10,Pawar-IT11,Shah-Globecom11,Goparaju-NetCod13,Rawat-IT14,Tandon-IT16,Ye-ISIT16,Shao-IT17}. The goal is to understand the optimal tradeoffs between the storage capacity and repair bandwidth in satisfying all three aforementioned requirements.
From the code construction perspective, it is natural to consider the so-called {\em separate coding} scheme, i.e., to construct a code for the $(n,d,\ell)$ MDC-SR problem, we can simply use an $(n,j,d,\ell)$ SRC to encode the file $\mathsf{M}_j$ for each $j\in\{\ell+1,\ldots,d\}$, and the coded messages for each file remain separate when stored in the storage nodes and during the repair processes. However, despite being a natural scheme, it was shown in \cite{Tian-IT16} that separate coding is in general {\em suboptimal} in achieving the optimal tradeoffs between the normalized storage-capacity and repair-bandwidth for the MDC-R problem (which is a special case of the MDC-SR problem as mentioned previously). On the other hand, it has been shown \cite{Shao-Arxiv17} that separate coding can, in fact, achieve the {\em minimum-bandwidth-regenerating (MBR)} point of the achievable normalized storage-capacity and repair-bandwidth tradeoff region for the general MDC-SR problem. Nevertheless, the optimal tradeoffs between the storage capacity $\alpha$ and download bandwidth $\beta$, and, the performance of the {\em minimum-storage-regenerating (MSR)} point are still not fully understood. Especially for the MSR point, a code was given in \cite{Shah-Globecom11} for SRC problem by extending the known MSR code without any security constraint. This coding scheme can achieve the MSR point when $d\geq 2k-2$ and the eavesdropper can only observe type \uppercase\expandafter{\romannumeral1} compromised nodes (the definition of type \uppercase\expandafter{\romannumeral1} compromised node will be defined in the following part). However, it is still unknown as to whether this code is optimal for the more general eavesdropper model in our paper.
In this paper, we shall revisit the MDC-SR problem with a more general eavesdropping model. More specifically, instead of assuming that the eavesdropper can access the repair data for all compromised storage nodes, we shall assume that the compromised storage nodes can be divided into two different categories: type \uppercase\expandafter{\romannumeral1} compromised nodes and type \uppercase\expandafter{\romannumeral2} compromised nodes. While for the type \uppercase\expandafter{\romannumeral2} compromised nodes, we assume that the eavesdropper can access the repair data as before, for the type \uppercase\expandafter{\romannumeral1} compromised nodes we assume that the eavesdropper can only access the stored data contents.
Let $\ell_1$ and $\ell_2$ be the number of type \uppercase\expandafter{\romannumeral1} compromised nodes and type \uppercase\expandafter{\romannumeral2} compromised nodes respectively, and $\ell:=\ell_1+\ell_2$ be the total number of compromised nodes. By the node repair requirement, the data contents stored at any node can be fully recovered from its repair data. Therefore, for any fixed $\ell$, the eavesdropper becomes weaker as $\ell_1$ increases, which leads to a potentially larger achievable normalized storage-capacity and repair-bandwidth tradeoff region. A question of fundamental interest is to understand whether increasing $\ell_1$ can lead to a {\em strictly} larger achievable normalized storage-capacity and repair-bandwidth tradeoff region. Our main result of the paper is to show that the MBR point of the achievable normalized storage-capacity and repair-bandwidth tradeoff region remains the {\em same}, as long as $\ell_1 \leq \ell/2$ (or equivalently, $\ell_1 \leq \ell_2$ by the fact that $\ell_2=\ell-\ell_1$). From the technical viewpoint, this is mainly accomplished by establishing two outer bounds (one of them must be ``horizontal", i.e., on the normalized repair-bandwidth {\em only}) on the achievable normalized storage-capacity and repair-bandwidth tradeoff region, which intersect precisely at the MBR point.
The rest of the paper is organized as follows. In Section~\ref{sec:PF} we formally introduce the problem of MDC-SR with the generalized eavesdropping model. The main results of the paper are then presented in Section~\ref{sec:Main}. In Section~\ref{sec:Proof}, we introduce two ``exchange" lemmas and use them to establish the main results of the paper. Finally, we conclude the paper in Section~\ref{sec:Con}.
{\em Notation}. Sets and random variables will be written in calligraphic and sans-serif fonts respectively, to differentiate from the real numbers written in normal math fonts. For any two integers $t \leq t'$, we shall denote the set of consecutive integers $\{t,t+1,\ldots,t'\}$ by $[t:t']$. The use of the brackets will be suppressed otherwise.
\section{The Generalized MDC-SR Problem}\label{sec:PF}
In this paper, we study a distributed storage system that share the same file recovery and node repair function with \cite{Shao-Arxiv17}. Let $(n,d,N_1,\ldots,N_d,K,T,S)$ be a tuple of positive integers such that $d<n$. Formally, an $(n,d,N_1,\ldots,N_d,K,T,S)$ code consists of:
\begin{itemize}
\item for each $i \in [1:n]$, a {\em message-encoding} function $f_i: \left(\prod_{j=1}^{d}[1:N_j]\right)\times[1:K] \rightarrow [1:T]$;
\vspace{4pt}
\item for each $\mathcal{A} \subseteq [1:n]: |\mathcal{A}|\in[1:d]$, a {\em message-decoding} function $g_\mathcal{A}: [1:T]^{|\mathcal{A}|} \rightarrow [1:N_{|\mathcal{A}|}]$;
\vspace{4pt}
\item for each $\mathcal{B}\subseteq [1:n]: |\mathcal{B}|=d$, $i' \in \mathcal{B}$, and $i \in [1:n]\setminus\mathcal{B}$, a {\em repair-encoding} function $f^{\mathcal{B}}_{i' \rightarrow i}: [1:T] \rightarrow [1:S]$;
\vspace{4pt}
\item for each $\mathcal{B}\subseteq [1:n]: |\mathcal{B}|=d$ and $i \in [1:n]\setminus\mathcal{B}$, a {\em repair-decoding} function $g^{\mathcal{B}}_i:[1:S]^d \rightarrow [1:T]$.
\end{itemize}
For each $j\in[1:d]$, let $\mathsf{M}_j$ be a message that is uniformly distributed over $[1:N_j]$. The messages $\mathsf{M}_1,\ldots,\mathsf{M}_d$ are assumed to be mutually independent. Let $\mathsf{K}$ be a random key that is uniformly distributed over $[1:K]$ and independent of the messages $(\mathsf{M}_1,\ldots,\mathsf{M}_d)$. For each $i\in[1:n]$, let $\mathsf{W}_i=f_i(\mathsf{M}_1,\ldots,\mathsf{M}_d,\mathsf{K})$ be the data stored at the $i$th storage node, and for each $\mathcal{B}\subseteq [1:n]: |\mathcal{B}|=d$, $i' \in \mathcal{B}$, and $i \in [1:n]\setminus\mathcal{B}$, let $\mathsf{S}^{\mathcal{B}}_{i'\rightarrow i}=f^{\mathcal{B}}_{i'\rightarrow i}(\mathsf{W}_{i'})$ be the data downloaded from the $i'$th storage node in order to regenerate the data originally stored at the $i$th storage node under the context of repair group $\mathcal{B}$. Obviously,
\begin{align*}
(B_j&=\log{N_j}: j\in[1:d]), \quad \alpha=\log{T}, \quad \mbox{and} \;\; \beta=\log{S}
\end{align*}
represent the message sizes, storage capacity, and repair bandwidth, respectively.
The main deference between our definition in this paper and that in \cite{Shao-Arxiv17} is the model of eavesdropper. The eavesdropper now can observer a more complicated data combination consisted of both stored content and repair content.Let $\ell_1$ and $\ell_2$ be two nonnegative integers such that $\ell:=\ell_1+\ell_2<d$. A normalized message-rate storage-capacity repair-bandwidth tuple
$(\bar{B}_{\ell+1},$ $\ldots,\bar{B}_d,\bar{\alpha},\bar{\beta})$ is said to be {\em achievable} for the $(n,d,\ell_1,\ell_2)$ generalized MDC-SR problem if an
$(n,d,1,\ldots,1,N_{\ell+1},\ldots,N_d,K,T,S)$ code (i.e., $N_j=1$ for all $j\in[1:\ell]$) can be found such that:
\begin{itemize}
\item (rate normalization)
\begin{align}
\frac{\alpha}{\sum_{t=\ell+1}^dB_t}=\bar{\alpha}, \;
\frac{\beta}{\sum_{t=\ell+1}^dB_t}=\bar{\beta}, \; \frac{B_j}{\sum_{t=\ell+1}^dB_t}=\bar{B}_j\label{eq:Rate}
\end{align}
for any $j\in[\ell+1:d]$;
\vspace{4pt}
\item (message recovery)
\begin{align}
\mathsf{M}_{|\mathcal{A}|} =g_\mathcal{A}(\mathsf{W}_i:i \in\mathcal{A})
\label{eq:MessageRecovery}
\end{align}
for any $\mathcal{A} \subseteq [1:n]: |\mathcal{A}|\in[\ell+1:d]$;
\vspace{4pt}
\item (node regeneration)
\begin{align}
\mathsf{W}_i =g^{\mathcal{B}}_i(\mathsf{S}^{\mathcal{B}}_{i' \rightarrow i}:i'\in \mathcal{B})
\label{eq:NodeRegen}
\end{align}
for any $\mathcal{B}\subseteq [1:n]: |\mathcal{B}|=d$ and $i \in [1:n]\setminus\mathcal{B}$;
\vspace{4pt}
\item (repair secrecy)
\begin{align}
I((\mathsf{M}_{\ell+1},\ldots,\mathsf{M}_d);(\mathsf{W}_i :i\in \mathcal{E}_1), (\mathsf{S}_{\rightarrow j}: j\in \mathcal{E}_2))=0\label{eq:RepairSecrecy}
\end{align}
for any $\mathcal{E}_1,\mathcal{E}_2\subseteq [1:n]$ such that $|\mathcal{E}_1|=\ell_1$, $|\mathcal{E}_2|=\ell_2$ and $\mathcal{E}_1\cap \mathcal{E}_2=\emptyset$ (so $\mathcal{E}_1$ and $\mathcal{E}_2$ represent the sets of type \uppercase\expandafter{\romannumeral1} and type \uppercase\expandafter{\romannumeral2} compromised storage nodes, respectively), where $\mathsf{S}_{\rightarrow i} :=(\mathsf{S}^{\mathcal{B}}_{i'\rightarrow i}:\mathcal{B}\subseteq [1:n], \; |\mathcal{B}|=d, \; \mathcal{B}\not\ni i, \; i' \in \mathcal{B})$ is the collection of data that can be downloaded from the other nodes to regenerate node $i$.
\end{itemize}
The closure of all achievable $(\bar{B}_{\ell+1},\ldots,\bar{B}_d,\bar{\alpha},\bar{\beta})$ tuples is the {\em achievable normalized message-rate storage-capacity repair-bandwidth tradeoff region} $\mathcal{R}_{n,d,\ell_1,\ell_2}$ for the $(n,d,\ell_1,\ell_2)$ generalized MDC-SR problem. For a fixed normalized message-rate tuple $(\bar{B}_{\ell+1},\ldots,\bar{B}_d)$, the {\em achievable normalized storage-capacity repair-bandwidth tradeoff region} is the collection of all normalized storage-capacity repair-bandwidth pairs $(\bar{\alpha},\bar{\beta})$ such that $(\bar{B}_{\ell+1},$ $\ldots,\bar{B}_d,\bar{\alpha},\bar{\beta}) \in \mathcal{R}_{n,d,\ell_1\ell_2}$ and is denoted by $\mathcal{R}_{n,d,\ell_1,\ell_2}$ $(\bar{B}_{\ell+1},$ $\ldots,\bar{B}_d)$.
Fixing $\ell$ and setting $\ell_1=0$, the $(n,d,\ell_1,\ell_2)$ generalized MDC-SR problem reduces to the $(n,d,\ell)$ MDC-SR problem considered previously in \cite{Shao-Arxiv17}, where it was shown that any achievable normalized message-rate storage-capacity repair-bandwidth tuple $(\bar{B}_{\ell+1},\ldots,\bar{B}_d,\bar{\alpha},\bar{\beta}) \in \mathcal{R}_{n,d,\ell}$ must satisfy:
\begin{align}
\bar{\beta} & \geq \sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}\bar{B}_j \label{eq:type2-1}\\
\mbox{and} \quad \bar{\alpha}+(d(d-\ell)-\ell)\bar{\beta} & \geq (d-\ell)(d+1)\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}\bar{B}_j\label{eq:type2-2}.
\end{align}
When set as equalities, the intersection of \eqref{eq:type2-1} and \eqref{eq:type2-2} is given by:
\begin{align}
\label{eq:MBR}
\left(\bar{\alpha},\bar{\beta}\right
&=\left(d\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}\bar{B}_j,\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}\bar{B}_j\right)
\end{align}
which can be achieved by separate encoding with a previous scheme proposed by Shah, Rashmi, and Kumar \cite{Shah-Globecom11}. This provides a precise characterization of the MBR point for the $(n,d,\ell)$ MDC-SR problem.
\section{Main Results}\label{sec:Main}
Our main result of the paper is to show that the tradeoff point \eqref{eq:MBR} remains to be the MBR point of $\mathcal{R}_{n,d,\ell_1,\ell_2}$ for the generalized MDC-SR problem as long as $\ell_1 \leq \ell_2$. The results are summarized in the following theorem.
\begin{thm}\label{thm}
For the generalized MDC-SR problem, any achievable normalized message-rate storage-capacity repair-bandwidth tuple $(\bar{B}_{\ell+1},\ldots,\bar{B}_d,\bar{\alpha},\bar{\beta}) \in $
$\mathcal{R}_{n,d,\ell_1,\ell_2}$ must satisfy:
\begin{align}
\bar{\beta} \geq \sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}\bar{B}_j \label{eq:B3}
\end{align}
and in addition, when $\ell_1 \leq \ell_2=\ell-\ell_1$, we also have
\begin{align}
\bar{\alpha}+T_{d,d,\ell_1+1}\bar{\beta} & \geq (T_{d,d,\ell_1}+\ell_1)\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}\bar{B}_j\label{eq:B4}
\end{align}
where $T_{d,k,\ell}:=\sum_{t=\ell+1}^{k}(d+1-t)$. When set as equalities, the intersection of \eqref{eq:B3} and \eqref{eq:B4} is precisely given by \eqref{eq:MBR}. We may thus conclude immediately that \eqref{eq:MBR} is the MBR point of $\mathcal{R}_{n,d,\ell_1,\ell_2}$ for the generalized MDC-SR problem as long as $\ell_1 \leq \ell_2$.
\end{thm}
Note that setting $\ell_1=0$, the outer bound \eqref{eq:B4} reduces to
\begin{align}
\bar{\alpha}+\frac{d(d-1)}{2}\bar{\beta} & \geq \frac{d(d+1)}{2}\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}\bar{B}_j.
\label{eq:B}
\end{align}
So while the outer bound \eqref{eq:B3} coincides with \eqref{eq:type2-1}, the outer bound \eqref{eq:B4} does {\em not} reduce to \eqref{eq:type2-2} when setting $\ell_1=0$. Simple calculations yield that the outer bound \eqref{eq:B} is stronger than \eqref{eq:type2-2} if and only if $\ell\leq d/2$. In particular, when $\ell=0$, the outer bound \eqref{eq:B} reduces to that for the $(n,d)$ MDC-R problem \cite{Shao-CISS16}, while the outer bound \eqref{eq:type2-2} is {\em strictly} weaker. Fig. \ref{fig1} shows the comparison of \eqref{eq:B} and \eqref{eq:type2-2} when $(\bar{B}_1,\bar{B}_2,\bar{B}_3)=(0,1/3,2/3)$ in $(4,3,0,0)$ MDC-SR problem. In this figure, the outer bound \eqref{eq:type2-2} is below outer bound \eqref{eq:B}, though both of them intersect with \eqref{eq:B3} at the MBR point.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.5\textwidth]{geogebra-export}
\caption{The optimal tradeoff region for $(4,3,0,0)$ MDC-SR problem when $(\bar{B}_1,\bar{B}_2,\bar{B}_3)=(0,1/3,2/3)$ . The outer bounds \eqref{eq:B3}, \eqref{eq:B} and \eqref{eq:type2-2} are evaluated as $\bar{\beta} \geq 8/45$, $\bar{\alpha}+3\bar{\beta} \geq 16/15$, and $\bar{\alpha}+9\bar{\beta} \geq 32/15$, respectively. When set as equalities, they intersect precisely at the MBR point $(8/15,8/45)$.}
\label{fig1}
\end{figure}
\section{Proof of the Main Results}\label{sec:Proof}
Let us first outline the main ingredients for proving the outer bounds \eqref{eq:B3} and \eqref{eq:B4}.
\begin{itemize}
\item[1)] {\em Total number of nodes.} To prove the outer bounds \eqref{eq:B3} and \eqref{eq:B4}, let us first note that these bounds are {\em independent} of the total number of storage nodes $n$ in the system. Therefore, in our proof, we only need to consider the cases where $n=d+1$. For the cases where $n>d+1$, since any subsystem consisting of $d+1$ out of the total $n$ storage nodes must give rise to a $(d+1,d,\ell)$ MDC-SR problem. Therefore, these {\em outer} bounds must apply as well. When $n=d+1$, any repair group $\mathcal{B}$ of size $d$ is uniquely determined by the node $j$ to be repaired, i.e., $\mathcal{B}=[1:n]\setminus\{j\}$, and hence can be dropped from the notation $\mathsf{S}^{\mathcal{B}}_{i \rightarrow j}$ without causing any confusion.
\item[2)] {\em Code symmetry.} Due to the built-in {\em symmetry} of the problem, to prove the outer bounds \eqref{eq:B3} and \eqref{eq:B4}, we only need to consider the so-called {\em symmetrical} codes \cite{Tian-JSAC13} for which the joint entropy of any subset of random variables from
\begin{align*}
&\left((\mathsf{M}_1,\ldots,\mathsf{M}_d),\mathsf{K},\right.\\
&\left.\hspace{30pt}(\mathsf{W}_i:i\in[1:n]),(\mathsf{S}_{i \rightarrow j}: i,j\in[1:n],i\neq j)\right)
\end{align*}
remains {\em unchanged} under {\em any} permutation over the {\em storage-node} indices.
\item[3)] {\em Key collections of random variables.} Focusing on the symmetrical $(n=d+1,d,N_1,\ldots,N_d,K,T,S)$ codes, the following collections of random variables play a key role in our proof:
\begin{align*}
& \mathsf{M}_{\mathcal{A}} := (\mathsf{M}_i: i \in \mathcal{A}), \quad \mathcal{A}\subseteq [1:d]\\
& \mathsf{M}^{(m)} := \mathsf{M}_{[1:m]}, \quad m\in[1:d]\\
&\mathsf{W}_{\mathcal{A}} :=\left(\mathsf{W}_i:i\in \mathcal{A}\right), \quad \mathcal{A}\subseteq [1:n]\\
&\mathsf{S}_{i\rightarrow\mathcal{B}} := \left(\mathsf{S}_{i \rightarrow j}: j\in\mathcal{B}\right),\quad i\in [1:n], \; \mathcal{B}\subseteq [1:n]\setminus \{i\}\\
&\mathsf{S}_{\mathcal{B}\rightarrow j} := \left(\mathsf{S}_{i \rightarrow j}: \quad i\in\mathcal{B}\right), j\in[1:n],\; \mathcal{B}\subseteq [1:n]\setminus \{j\}\\
&\mathsf{S}_{\rightarrow j} := \mathsf{S}_{[1:j-1]\cup[j+1:n]\rightarrow j}, \quad j\in[1:n]\\
&\mathsf{S}_{\rightarrow \mathcal{B}} :=\left(\mathsf{S}_{\rightarrow j}:j\in\mathcal{B}\right), \quad \mathcal{B}\subseteq [1:n]\\
&\underline{\mathsf{S}}_{\rightarrow j} := \mathsf{S}_{[1:j-1]\rightarrow j}, \quad j\in[1:n]\\
&\underline{\mathsf{S}}_{\rightarrow \mathcal{B}} :=(\underline{\mathsf{S}}_{\rightarrow j}:j\in\mathcal{B}), \quad \mathcal{B}\subseteq [1:n]\\
&\overline{\mathsf{S}}_{\rightarrow j} := \mathsf{S}_{[j+1:n]\rightarrow j}, \quad j\in[1:n]\\
&\overline{\mathsf{S}}_{\rightarrow \mathcal{B}} :=(\overline{\mathsf{S}}_{\rightarrow j}:j\in\mathcal{B}), \quad \mathcal{B}\subseteq [1:n]\\
&\mathsf{U}^{(t,s)} :=(\mathsf{W}_{[1:t]},\overline{\mathsf{S}}_{\rightarrow[t+1:s]}), \quad s\in[1:n], \; t\in[0:s]\\
&\mathsf{U}^{(s)} := \mathsf{U}^{(0,s)}.
\end{align*}
These collections of random variables have also been used in \cite{Shao-IT17,Shao-CISS16}.
\end{itemize}
An important part of the proof is to understand the relations between the collections of random variables defined above, and to use them to derive the desired converse results. We shall discuss this next.
\subsection{Technical Lemmas}
\begin{lemma}\label{lemma1}
For any $(n=d+1,d,N_1,\ldots,N_d,K,T,S)$ code that satisfies the node regeneration requirement \eqref{eq:NodeRegen}, $(\underline{\mathsf{S}}_{\rightarrow[t+1:s]},\mathsf{W}_{[t+1:s]})$ is a function of $\mathsf{U}^{(t,s)}$ for any $s \in [1:n]$ and $t\in[0:s-1]$.
\end{lemma}
The above lemma, which was first introduced in \cite{Shao-Arxiv17,Shao-IT17}, demonstrates the ``compactness" of $\mathsf{U}^{(t,s)}$ and has a number of direct consequences. For example, for any fixed $s \in [1:n]$, it is clear from Lemma~\ref{lemma1} that $\mathsf{U}^{(t_2,s)}$ is a function of $\mathsf{U}^{(t_1,s)}$ and hence $H(\mathsf{U}^{(t_2,s)}) \leq H(\mathsf{U}^{(t_1,s)})$ for any $0 \leq t_1 \leq t_2 \leq s-1$.
\begin{lemma}[Exchange lemma \uppercase\expandafter{\romannumeral1} \cite{Shao-Arxiv17}]
\label{lemma:exchange}
For any symmetrical $(n=d+1,d,N_1,\ldots,N_d$, $K,T,S)$ code that satisfies the node regeneration requirement \eqref{eq:NodeRegen}, we have
\begin{align}
& \frac{d+1-j}{d-m}H(\mathsf{U}^{(i,m)}|\mathsf{M}^{(m)})+H(\mathsf{U}^{(i',j)}|\mathsf{M}^{(m)})\nonumber\\
& \hspace{15pt} \geq \frac{d+1-j}{d-m}H(\mathsf{U}^{(i,m+1)}|\mathsf{M}^{(m)})+H(\mathsf{U}^{(i',j-1)}|\mathsf{M}^{(m)})\label{eq:exchange}
\end{align}
for any $m\in[1:d-1]$, $i\in[0:m-1]$, $i'\in[0:i]$, and $j\in[i'+1:m-i+i'+1]$.
\end{lemma}
\begin{coro}
\label{coro1}
For any symmetrical $(n=d+1,d,N_1,\ldots, N_d$, $K,T,S)$ code that satisfies the node regeneration requirement \eqref{eq:NodeRegen}, we have
\begin{align}
& \frac{T_{d,j_1,j_2}}{d-j_1}H(\mathsf{U}^{(i,j_1)}|\mathsf{M}^{(j_1)})+H(\mathsf{U}^{(i',j_1)}|\mathsf{M}^{(j_1)})\nonumber\\
& \hspace{15pt} \geq \frac{T_{d,j_1,j_2}}{d-j_1}H(\mathsf{U}^{(i,j_1+1)}|\mathsf{M}^{(j_1)})+H(\mathsf{U}^{(i',j_2)}|\mathsf{M}^{(j_1)})\label{eq:coro1}
\end{align}
for any $j_1\in[1:d-1]$, $i=[0:j_1]$, $i' \in[\max\{0,i-1\}:i]$ and $j_2\in[i':j_1-1]$.
\end{coro}
\begin{IEEEproof}
Set $m=j_1$ in \eqref{eq:exchange}. We have
\begin{align}
& \frac{d+1-j}{d-j_1}H(\mathsf{U}^{(i,j_1)}|\mathsf{M}^{(j_1)})+H(\mathsf{U}^{(i',j)}|\mathsf{M}^{(j_1)})\nonumber\\
& \hspace{10pt}\geq \frac{d+1-j}{d-j_1}H(\mathsf{U}^{(i,j_1+1)}|\mathsf{M}^{(j_1)})+H(\mathsf{U}^{(i',j-1)}|\mathsf{M}^{(j_1)})\label{eq:NewExchange2}
\end{align}
for any $j\in[j_2+1:j_1]$. Add the inequalities \eqref{eq:NewExchange2} for $j\in[j_2+1:j_1]$ and cancel the common term $\sum_{j=j_2+1}^{j_1-1}H(\mathsf{U}^{(i',j)}|\mathsf{M}^{(j_1)})$ from both sides. We have
\begin{align}
&\frac{T_{d,j_1,j_2}}{d-j_1}H(\mathsf{U}^{(i,j_1)}|\mathsf{M}^{(j_1)})+H(\mathsf{U}^{(i',j_1)}|\mathsf{M}^{(j_1)})\nonumber\\
& \hspace{15pt} \geq \frac{T_{d,j_1,j_2}}{d-j_1}H(\mathsf{U}^{(i,j_1+1)}|\mathsf{M}^{(j_1)})+H(\mathsf{U}^{(i',j_2)}|\mathsf{M}^{(j_1)}).\nonumber
\end{align}
\end{IEEEproof}
\begin{coro}
\label{coro2}
For any symmetrical $(n=d+1,d,N_1,\ldots,N_d$, $K,T,S)$ code that satisfies the node regeneration requirement \eqref{eq:NodeRegen}, we have
\begin{align}
T_{d,m,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}^{(m)})&\geq T_{d,m+1,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m+1)}|\mathsf{M}^{(m)})+\nonumber\\
&
(T_{d,m,\ell}^{-1}-T_{d,m+1,\ell}^{-1})H(\mathsf{U}^{(\ell_1,\ell)}|\mathsf{M}^{(m)})\label{eq:coro2}
\end{align}
for any $\ell\in[0:d-1]$, $\ell_1\in[0:\ell]$ and $m\in[\ell+1:d-1]$.
\end{coro}
\begin{IEEEproof}
Set $i=i'=\ell_1$, $j_1=m$ and $j_2=\ell$ in \eqref{eq:coro1}. We have
\begin{align}
& \frac{T_{d,m,\ell}}{d-m}H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}^{(m)})+H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}^{(m)})\nonumber\\
& \hspace{15pt} \geq \frac{T_{d,m,\ell}}{d-m}H(\mathsf{U}^{(\ell_1,m+1)}|\mathsf{M}^{(m)})+H(\mathsf{U}^{(\ell_1,\ell)}|\mathsf{M}^{(m)}),
\end{align}
which can be equivalently written as
\begin{align}
&\frac{T_{d,m+1,\ell}}{d-m}H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}^{(m)})\nonumber\\
& \hspace{30pt} \geq \frac{T_{d,m,\ell}}{d-m}H(\mathsf{U}^{(\ell_1,m+1)}|\mathsf{M}^{(m)})+ H(\mathsf{U}^{(\ell_1,\ell)}|\mathsf{M}^{(m)})\label{eq:Temp4}
\end{align}
by the fact that $T_{d,m,\ell}+(d-m)=T_{d,m+1,\ell}$. Multiplying both sides of \eqref{eq:Temp4} by
$$\frac{d-m}{T_{d,m+1,\ell}T_{d,m,\ell}}=T_{d,m,\ell}^{-1}-T_{d,m+1,\ell}^{-1}$$
completes the proof of \eqref{eq:coro1}.
\end{IEEEproof}
\begin{lemma}[exchange lemma \uppercase\expandafter{\romannumeral2}]
\label{lemma:mutant exchange}
For any symmetrical $(n=d+1,d,N_1$, $\ldots,N_d,K,T,S)$ code that satisfies the node regeneration requirement \eqref{eq:NodeRegen}, we have
\begin{align}
& \frac{d-\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})+H(\mathsf{U}^{(\ell_1,\ell_1+1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})\nonumber\\
& \hspace{15pt} \geq \frac{d-\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})+H(\mathsf{U}^{(\ell_1,\ell_1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})\label{eq:mutant exchange}
\end{align}
for any $\ell \in[1:d-1]$ and $\ell_1 \in[0:\left\lfloor \ell/2 \right\rfloor]$.
\end{lemma}
\begin{IEEEproof}
See the Appendix.
\end{IEEEproof}
We note here that when setting $\ell_1=0$, the above lemma coincides with Lemma~\ref{lemma:exchange} with $i=i'=0$ and $j=1$.
\subsection{The Proof}
Consider a symmetrical $(n=d+1,d,1,\ldots,1,N_{\ell+1},\ldots,N_d$, $K,T,S)$ regenerating code that satisfies the rate normalization requirement \eqref{eq:Rate}, the message recovery requirement \eqref{eq:MessageRecovery}, the node regeneration requirement \eqref{eq:NodeRegen}, and the repair secrecy requirement \eqref{eq:RepairSecrecy}. Let us first prove a few intermediate results. The outer bounds \eqref{eq:B3} and \eqref{eq:B4} will then follow immediately.
\begin{prop}\label{prop1}
\begin{align}
&\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)}) \geq \sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+\nonumber\\
& \hspace{15pt} T_{d,m,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}_{[\ell+1:m]})+\left(\frac{1}{d-\ell}-T_{d,m,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\label{eq:EG}
\end{align}
for any $m\in [\ell+1:d]$. Consequently,
\begin{align}
\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)}) \geq \sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}B_j+\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)}).
\label{eq:prop1}
\end{align}
\end{prop}
\begin{IEEEproof}
To see \eqref{eq:EG}, consider proof by induction. For the base case with $m=\ell+1$, we have
\begin{align*}
&\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\nonumber\\
\stackrel{(a)}{=}& \frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)},\mathsf{M}_{\ell+1})\\
\stackrel{(b)}{=}& \frac{1}{d-\ell}\left(H(\mathsf{M}_{\ell+1})+H(\mathsf{U}^{(\ell_1,\ell+1)}|\mathsf{M}_{\ell+1})\right)\\
\stackrel{(c)}{=}& \frac{1}{d-\ell}\left(B_{\ell+1}+H(\mathsf{U}^{(\ell_1,\ell+1)}|\mathsf{M}_{\ell+1})\right)\\
\stackrel{(d)}{=}& T_{d,\ell+1,\ell}^{-1}B_{\ell+1}+T_{d,\ell+1,\ell}^{-1}H(\mathsf{U}^{(\ell_1,\ell+1)}|\mathsf{M}_{\ell+1})
\end{align*}
where $(a)$ follows from the fact that $\mathsf{M}_{\ell+1}$ is a function of $\mathsf{W}_{[1:\ell+1]}$, which is a function of $\mathsf{U}^{(\ell_1,\ell+1)}$ by Lemma~\ref{lemma1}; $(b)$ follows from the chain rule for entropy; $(c)$ follows from the fact that $H(\mathsf{M}_{\ell+1})=B_{\ell+1}$; and $(d)$ follows from the fact that $T_{d,\ell+1,\ell}=d-\ell$. Assuming that \eqref{eq:EG} holds for some $m\in [\ell+1:d-1]$, we have
\begin{align*}
& \frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
& \stackrel{(a)}{\geq} \sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+T_{d,m,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt} \left(\frac{1}{d-\ell}-T_{d,m,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
& \stackrel{(b)}{\geq} \sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+T_{d,m+1,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m+1)}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt} \left(\frac{1}{d-\ell}-T_{d,m+1,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
& \stackrel{(c)}{\geq} \sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+\nonumber\\
& \hspace{20pt} T_{d,m+1,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m+1)},\mathsf{M}_{m+1}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt} \left(\frac{1}{d-\ell}-T_{d,m+1,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
& \stackrel{(d)}{=} \sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+T_{d,m+1,\ell}^{-1}H(\mathsf{M}_{m+1}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt} T_{d,m+1,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m+1)}|\mathsf{M}_{[\ell+1:m+1]})+\\
& \hspace{20pt} \left(\frac{1}{d-\ell}-T_{d,m+1,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
& \stackrel{(e)}{=} \sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+T_{d,m+1,\ell}^{-1}B_{m+1}+\\
& \hspace{20pt} T_{d,m+1,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m+1)}|\mathsf{M}_{[\ell+1:m+1]})+\\
& \hspace{20pt} \left(\frac{1}{d-\ell}-T_{d,m+1,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
& = \sum_{j=\ell+1}^{m+1}T_{d,j,\ell}^{-1}+T_{d,m+1,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m+1)}|\mathsf{M}_{[\ell+1:m+1]})+\\
& \hspace{20pt} \left(\frac{1}{d-\ell}-T_{d,m+1,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})
\end{align*}
where $(a)$ follows from the induction assumption; $(b)$ follows from Corollary~\ref{coro1}; $(c)$ follows from the fact that $\mathsf{M}_{m+1}$ is a function of $\mathsf{W}_{[1:m+1]}$, which is is a function of $\mathsf{U}^{(\ell_1,m+1)}$ by Lemma~\ref{lemma1}; $(d)$ follows from the chain rule for entropy; and $(e)$ follows from the facts that $\mathsf{M}_{m+1}$ is independent of $\mathsf{M}_{[\ell+1:m]}$ and that $H(\mathsf{M}_{m+1})=B_{m+1}$. This completes the induction step and hence the proof of \eqref{eq:EG}.
To see \eqref{eq:prop1}, simply set $m=d$ in \eqref{eq:EG}. We have
\begin{align}
&\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)}) \geq \sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}B_j+\nonumber\\
& \hspace{20pt} T_{d,d,\ell}^{-1}H(\mathsf{U}^{(\ell_1,d)}|\mathsf{M}_{[\ell+1:d]})+\left(\frac{1}{d-\ell}-T_{d,d,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)}).\label{eq:EG2}
\end{align}
Note that
\begin{align}
H(\mathsf{U}^{(\ell_1,d)}|\mathsf{M}_{[\ell+1:d]}) \geq H(\mathsf{U}^{(\ell_1,\ell)}|\mathsf{M}_{[\ell+1:d]}) = H(\mathsf{U}^{(\ell_1,\ell)})\label{eq:EG3}
\end{align}
where the last equality follows from the fact that $I(\mathsf{U}^{(\ell_1,\ell)};\mathsf{M}_{[\ell+1:d]})=0$ by the repair secrecy requirement \eqref{eq:RepairSecrecy}. Substituting \eqref{eq:EG3} into \eqref{eq:EG2} completes the proof of \eqref{eq:prop1}.
\end{IEEEproof}
\begin{prop}\label{prop2}
\begin{align}
H(\mathsf{S}_{\ell_1+1\rightarrow[1:\ell_1]})+\frac{\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})\geq
\frac{\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)}).\label{eq:prop2}
\end{align}
\end{prop}
\begin{IEEEproof}
First note that for any $m \in [1:\ell_2+1]$ and $k\in[\ell+1:d+1]$, we have
\begin{align}
H&(\mathsf{S}_{\ell_1+1\rightarrow[1:m]})+H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{[\ell+2:k]\rightarrow \ell+1})\nonumber\\
& \stackrel{(a)}{=} H(\mathsf{S}_{k+1\rightarrow[\ell_1+1:\ell_1+m-1]\cup\{\ell+1\}})+\nonumber\\
& \hspace{20pt} H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{\rightarrow [\ell_1+1:\ell]},\mathsf{S}_{[\ell+2:k]\rightarrow \ell+1})\nonumber\\
& \stackrel{(b)}{\geq} H(\mathsf{S}_{k+1\rightarrow[\ell_1+1:\ell_1+m-1]})+H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{[\ell+2:k+1]\rightarrow \ell+1})\nonumber\\
& \stackrel{(c)}{=} H(\mathsf{S}_{\ell_1+1\rightarrow[1:m-1]})+H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{[\ell+2:k+1]\rightarrow \ell+1})\label{eq:STE}
\end{align}
where $(a)$ and $(c)$ follow from the fact that $H(\mathsf{S}_{\ell_1+1\rightarrow[1:m]})=H(\mathsf{S}_{k+1\rightarrow[1:m-1]\cup\{\ell+1\}})$ and $H(\mathsf{S}_{k+1\rightarrow[1:m-1]})=H(\mathsf{S}_{\ell_1+1\rightarrow[1:m-1]})$ due to the symmetrical code that we consider, and $(b)$ follows from the submodularity of the entropy function. Add \eqref{eq:STE} over $m\in[1:\ell_1]$ and cancel $\sum_{m=1}^{\ell_1-1}H(\mathsf{S}_{d+1\rightarrow[1:m]})$ from both sides. We have
\begin{align}
&H(\mathsf{S}_{d+1\rightarrow[1:\ell]})+\ell_1 H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{[\ell+2:k+1]\rightarrow \ell+1}) \geq\nonumber\\
&\quad \ell_1 H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{[\ell+2:k+1]\rightarrow \ell+1}).\label{eq:STE2}
\end{align}
Add \eqref{eq:STE2} over $k\in [\ell+1:d]$ and cancel $\sum_{k=\ell+1}^{d-1} H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{[\ell+2:k+1]\rightarrow \ell+1})$ from both sides. We have
\begin{align}
&(d-\ell)H(\mathsf{S}_{d+1\rightarrow[1:\ell]})+\ell_1 H(\mathsf{U}^{(\ell_1,\ell)})\nonumber\\
\geq& \ell_1 H(\mathsf{U}^{(\ell_1,\ell)},\mathsf{S}_{[\ell+2:d+1]\rightarrow \ell+1})\nonumber\\
= &\ell_1 H(\mathsf{U}^{(\ell_1,\ell+1)}).\label{eq:STE2}
\end{align}
Multiplying both sides by $ (d-\ell)^{-1}$ completes the proof of \eqref{eq:prop2}
\end{IEEEproof}
\begin{prop}\label{prop3}
\begin{align}
&H(\mathsf{U}^{(\ell_1+1,m)})+\frac{d-m}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\nonumber\\
& \geq (d-m)\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+H(\mathsf{U}^{(\ell_1+1,m+1)})+\frac{d-m}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})\label{eq:JH}
\end{align}
for any $m\in[\ell+1:d-1]$. Consequently,
\begin{align}
& H(\mathsf{U}^{(\ell_1+1,\ell+1)})+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\nonumber\\
& \hspace{20pt} \geq T_{d,d,\ell}\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}B_j+\frac{T_{d,d,\ell}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)}).\label{eq:prop3}
\end{align}
\end{prop}
\begin{IEEEproof}
To see \eqref{eq:JH}, note that for any $m\in[\ell+1:d-1]$, we have
\begin{align*}
&H(\mathsf{U}^{(\ell_1+1,m)}|\mathsf{M}_{[\ell+1:m]})+\frac{d-m}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
& \stackrel{(a)}{\geq} H(\mathsf{U}^{(\ell_1+1,m)}|\mathsf{M}_{[\ell+1:m]})+ \\
& \hspace{20pt} (d-m)\left(\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+\right.T_{d,m,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}_{[\ell+1:m]})\\
& \hspace{20pt} \left.+\left(\frac{1}{d-\ell}-T_{d,m,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\right)\\
& = H(\mathsf{U}^{(\ell_1+1,m)}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt}
(d-m)T_{d,m,\ell}^{-1}H(\mathsf{U}^{(\ell_1,m)}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt} (d-m)\left(\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+\left(\frac{1}{d-\ell}-T_{d,m,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\right)\\
& \stackrel{(b)}{\geq} H(\mathsf{U}^{(\ell_1+1,m+1)}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt}
(d-m)T_{d,m,\ell}^{-1}H(\mathsf{U}^{(\ell_1,\ell)}|\mathsf{M}_{[\ell+1:m]})+\\
& \hspace{20pt} (d-m)\left(\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+\left(\frac{1}{d-\ell}-T_{d,m,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1,\ell)})\right)\\
& \stackrel{(c)}{=} H(\mathsf{U}^{(\ell_1+1,m+1)}|\mathsf{M}_{[\ell+1:m]})+(d-m)T_{d,m,\ell}^{-1}H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt} (d-m)\left(\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+\left(\frac{1}{d-\ell}-T_{d,m,\ell}^{-1}\right)H(\mathsf{U}^{(\ell_1\ell)})\right)\\
& = H(\mathsf{U}^{(\ell_1+1,m+1)}|\mathsf{M}_{[\ell+1:m]})+(d-m)\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j+\\
& \hspace{20pt} \frac{d-m}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})
\end{align*}
where $(a)$ follows from \eqref{eq:EG} of Proposition~\ref{prop1}; $(b)$ follows from Corollary~\ref{coro2}; and $(c)$ follows from the fact that $I(\mathsf{U}^{(\ell_1,\ell)};\mathsf{M}_{[\ell+1:m]})=0$ due to the repair secrecy requirement \eqref{eq:RepairSecrecy}. Adding $H(\mathsf{M}_{[\ell+1:m]})$ to both sides and using the facts that
\begin{align*}
&H(\mathsf{U}^{(\ell_1+1,m)}|\mathsf{M}_{[\ell+1:m]})+H(\mathsf{M}_{[\ell+1:m]})\\
&\hspace{40pt}=H(\mathsf{U}^{(\ell_1+1,m)},\mathsf{M}_{[\ell+1:m]})\stackrel{(a)}{=}H(\mathsf{U}^{(\ell_1+1,m)})
\end{align*}
and that
\begin{align*}
&H(\mathsf{U}^{(\ell_1+1,m+1)}|\mathsf{M}_{[\ell+1:m]})+H(\mathsf{M}_{[\ell+1:m]})\\
&\hspace{40pt}=H(\mathsf{U}^{(\ell_1+1,m+1)},\mathsf{M}_{[\ell+1:m]})\stackrel{(b)}{=}H(\mathsf{U}^{(\ell_1+1,m+1)})
\end{align*}
complete the proof of \eqref{eq:JH}. Here, $(a)$ and $(b)$ are due to the facts that $\mathsf{M}_{[\ell+1:m]}$ is a function of $\mathsf{W}_{[1:m]}$, which is a function of both $\mathsf{U}^{(\ell_1+1,m)}$ and $\mathsf{U}^{(\ell_1+1,m+1)}$ by Lemma~\ref{lemma1}.
To see \eqref{eq:prop3}, add \eqref{eq:JH} over $m\in[\ell+1:d-1]$ and cancel $\sum_{m=\ell+2}^{d-1}H(\mathsf{U}^{(\ell_1+1,m)})$ from both sides of the inequality. We have
\begin{align}
& H(\mathsf{U}^{(\ell_1+1,\ell+1)})+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\nonumber\\
& \hspace{40pt} \geq \sum_{m=\ell+1}^{d-1}\left((d-m)\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j\right)+\nonumber\\
& \hspace{80pt} H(\mathsf{U}^{(\ell_1+1,d)})+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)}).\label{eq:JH1}
\end{align}
Note that
\begin{align}
& \sum_{m=\ell+1}^{d-1}\left((d-m)\sum_{j=\ell+1}^{m}T_{d,j,\ell}^{-1}B_j\right)\nonumber\\
& \hspace{10pt} = \sum_{j=\ell+1}^{d-1}T_{d,j,\ell}^{-1}B_j\left(\sum_{m=j}^{d-1}(d-m)\right)
= \sum_{j=\ell+1}^{d-1}T_{d,j,\ell}^{-1}T_{d,d,j}B_j.\label{eq:JH2}
\end{align}
Furthermore,
\begin{align}
H(\mathsf{U}^{(\ell_1+1,d)})& \stackrel{(a)}{=} H(\mathsf{U}^{(\ell_1+1,d)},\mathsf{M}_{[\ell+1:d]})\nonumber\\
& \stackrel{(b)}{=} H(\mathsf{U}^{(\ell_1+1,d)}|\mathsf{M}_{[\ell+1:d]})+H(\mathsf{M}_{[\ell+1:d]})\nonumber\\
& \stackrel{(c)}{=} H(\mathsf{U}^{(\ell_1+1,d)}|\mathsf{M}_{[\ell+1:d]})+\sum_{j=\ell+1}^dB_j\nonumber\\
& \geq H(\mathsf{U}^{(\ell_1,\ell)}|\mathsf{M}_{[\ell+1:d]})+\sum_{j=\ell+1}^dB_j\nonumber\\
& \stackrel{(d)}{=} H(\mathsf{U}^{(\ell_1,\ell)})+\sum_{j=\ell+1}^dB_j\label{eq:JH3}
\end{align}
where $(a)$ follows from the fact that $\mathsf{M}_{[\ell+1:d]}$ is a function of $\mathsf{W}_{[1:d]}$, which is is a function of $\mathsf{U}^{(\ell_1+1,d)}$ by Lemma~\ref{lemma1}; $(b)$ follows from the chain rule for entropy; $(c)$ follows from the fact that
$H(\mathsf{M}_{[\ell+1:d]})=\sum_{j=\ell+1}^dB_j$; and $(d)$ follows from the fact that $I(\mathsf{U}^{(\ell_1,\ell)};\mathsf{M}_{[\ell+1:d]})=0$ due to the repair secrecy requirement \eqref{eq:RepairSecrecy}.
Substituting \eqref{eq:JH2} and \eqref{eq:JH3} into \eqref{eq:JH1} gives:
\begin{align*}
& H(\mathsf{U}^{(\ell_1+1,\ell+1)})+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
& \geq \sum_{j=\ell+1}^{d-1}T_{d,j,\ell}^{-1}T_{d,d,j}B_j+\sum_{j=\ell+1}^dB_j+\\
& \hspace{20pt} \left(1+\frac{T_{d,d,\ell+1}}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
& = \sum_{j=\ell+1}^{d-1}T_{d,j,\ell}^{-1}(T_{d,d,j}+T_{d,j,\ell})B_j+B_d+\frac{T_{d,d,\ell}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})\\
& \stackrel{(a)}{=} T_{d,d,\ell}\sum_{j=\ell+1}^{d-1}T_{d,j,\ell}^{-1}B_j+B_d+\frac{T_{d,d,\ell}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})\\
& = T_{d,d,\ell}\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}B_j+\frac{T_{d,d,\ell}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})
\end{align*}
where $(a)$ follows from the fact that $T_{d,d,j}+T_{d,j,\ell}=T_{d,d,\ell}$. This completes the proof of the proposition.
\end{IEEEproof}
We are now ready to prove the outer bounds \eqref{eq:B3} and \eqref{eq:B4}. To prove \eqref{eq:B3}, note that
\begin{align*}
\beta+\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)}) & \stackrel{(a)}{\geq} \frac{1}{d-\ell}\left(H(\overline{\mathsf{S}}_{\rightarrow \ell+1})+H(\mathsf{U}^{(\ell_1,\ell)})\right)\\
& \stackrel{(b)}{\geq} \frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
& \stackrel{(c)}{\geq} \sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}B_j+\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})
\end{align*}
where $(a)$ follows from the fact that $H(\overline{\mathsf{S}}_{\rightarrow \ell+1}) \leq (d-\ell)\beta$; $(b)$ follows from the union bound on entropy; and $(c)$ follows from \eqref{eq:prop1} of Proposition~\ref{prop1}. Cancelling $\frac{1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})$ from both sides of the inequality and normalizing both sides by $\sum_{t=\ell+1}^{d}B_t$ complete the proof of \eqref{eq:B3}.
To prove \eqref{eq:B4}, note that
\begin{align*}
\alpha&+T_{d,d,\ell_1+1}\beta+\frac{\ell_1+T_{d,d,\ell_1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})\\
&\stackrel{(a)}{=} \alpha+T_{d,d,\ell_1+1}\beta+\\
& \hspace{10pt}\left(\frac{\ell_1}{d-\ell}+\frac{T_{d,\ell,\ell_1+1}}{d-\ell}
+\frac{T_{d,d,\ell+1}}{d-\ell}+\frac{d-\ell_1}{d-\ell}+1\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
&=\frac{T_{d,\ell,\ell_1+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})+H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt}\alpha+T_{d,d,\ell_1+1}\beta+\\
& \hspace{20pt} \left(\frac{\ell_1}{d-\ell}+\frac{T_{d,d,\ell+1}}{d-\ell}+\frac{d-\ell_1}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
&\stackrel{(b)}{\geq} \frac{T_{d,\ell,\ell_1+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})+H(\mathsf{U}^{(\ell_1,\ell_1+1)})+\\
& \hspace{20pt}\alpha+T_{d,d,\ell_1+1}\beta+\\
& \hspace{20pt}
\left(\frac{\ell_1}{d-\ell}+\frac{T_{d,d,\ell+1}}{d-\ell}+\frac{d-\ell_1}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})\\
&= \frac{d-\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})+H(\mathsf{U}^{(\ell_1,\ell_1+1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+\\
& \hspace{20pt}\alpha+T_{d,d,\ell_1+1}\beta+\left(\frac{\ell_1}{d-\ell}
+\frac{T_{d,d,\ell+1}}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt} \frac{T_{d,\ell,\ell_1+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(c)}{\geq} \frac{d-\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})+H(\mathsf{U}^{(\ell_1,\ell_1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+\\
& \hspace{20pt}\alpha+T_{d,d,\ell_1+1}\beta+\left(\frac{\ell_1}{d-\ell}
+\frac{T_{d,d,\ell+1}}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt} \frac{T_{d,\ell,\ell_1+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(d)}{=} \alpha+H(\mathsf{U}^{(\ell_1,\ell_1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+\\
& \hspace{20pt}T_{d,d,\ell_1+1}\beta+\left(\frac{\ell_1}{d-\ell}
+\frac{T_{d,d,\ell+1}}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt} \frac{T_{d,\ell,\ell_1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(e)}{\geq} H(W_{\ell+1})+H(\mathsf{U}^{(\ell_1,\ell_1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+\\
& \hspace{20pt}T_{d,d,\ell_1+1}\beta+\left(\frac{\ell_1}{d-\ell}
+\frac{T_{d,d,\ell+1}}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt} \frac{T_{d,\ell,\ell_1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(f)}{=} H(W_{\ell+1},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+H(\mathsf{U}^{(\ell_1,\ell_1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+\\
& \hspace{20pt}T_{d,d,\ell_1+1}\beta+\left(\frac{\ell_1}{d-\ell}
+\frac{T_{d,d,\ell+1}}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt} \frac{T_{d,\ell,\ell_1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(g)}{\geq} H(\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+H(\mathsf{U}^{(\ell_1+1,\ell_1+1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+\\
& \hspace{20pt}T_{d,d,\ell_1+1}\beta+\left(\frac{\ell_1}{d-\ell}
+\frac{T_{d,d,\ell+1}}{d-\ell}\right)H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt} \frac{T_{d,\ell,\ell_1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(h)}{\geq} H(\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+\frac{\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt}H(\mathsf{U}^{(\ell_1+1,\ell_1+1)})+T_{d,d,\ell_1+1}\beta
+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})+\\
& \hspace{20pt}\frac{T_{d,\ell,\ell_1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(i)}{\geq} \frac{\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})+
H(\mathsf{U}^{(\ell_1+1,\ell_1+1)})+T_{d,d,\ell_1+1}\beta\\
& \hspace{20pt}
+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})
+\frac{T_{d,\ell,\ell_1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(j)}{=} H(\mathsf{U}^{(\ell_1+1,\ell_1+1)})+T_{d,\ell+1,\ell_1+1}\beta\\
& \hspace{20pt}
+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})+T_{d,d,\ell+1}\beta+\\
& \hspace{20pt}
\frac{T_{d,\ell,\ell_1}+\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(k)}{\geq} H(\mathsf{U}^{(\ell_1+1,\ell+1)})+\frac{T_{d,d,\ell+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})+\\
& \hspace{20pt}\frac{T_{d,\ell,\ell_1}+\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})\\
&\stackrel{(l)}{\geq} (T_{d,d,\ell}+T_{d,\ell,\ell_1}+\ell_1)\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}B_j+\\
& \hspace{20pt} \frac{T_{d,d,\ell}+T_{d,\ell,\ell_1}+\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})\\
&\stackrel{(m)}{=} (T_{d,d,\ell_1}+\ell_1)\sum_{j=\ell+1}^{d}T_{d,j,\ell}^{-1}B_j
+\frac{T_{d,d,\ell_1}+\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})
\end{align*}
where $(a)$ follows from the fact that $T_{d,\ell,\ell_1+1}+T_{d,d,\ell+1}+d-\ell_1+d-\ell=T_{d,d,\ell_1} $; $(b)$ follows from Corollary \ref{coro1} that
\begin{align*}
&\frac{T_{d,\ell,\ell_1+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})+H(\mathsf{U}^{(\ell_1,\ell)})\geq \\
&\hspace{40pt}\frac{T_{d,\ell,\ell_1+1}}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell+1)})+H(\mathsf{U}^{(\ell_1,\ell_1+1)})
\end{align*}
by setting $j_1=\ell$, $j_2=\ell_1+1$ and $i=i'=\ell_1$ in \ref{eq:coro1}; $(c)$ follows from \eqref{eq:mutant exchange} in Lemma \ref{lemma:mutant exchange}; $(d)$ follows from the fact that $T_{d,\ell,\ell_1+1}+d-\ell_1=T_{d,\ell,\ell_1} $; $(e)$ follows from the fact that $H(\mathsf{W}_{\ell+1}) \leq \alpha$; $(f)$ and $(h)$ follows from the fact that $\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]}$ is a function of $\mathsf{W}_{\ell_1+1}$; $(g)$ follows from the fact that $H(W_{\ell+1},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+H(\mathsf{U}^{(\ell_1,\ell_1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})\geq H(\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})+H(\mathsf{U}^{(\ell_1+1,\ell_1+1)},\mathsf{S}_{\ell_1+1\rightarrow [1:\ell_1]})$ due to submodularity; $(i)$ follows from \eqref{eq:prop2} in Proposition \ref{prop2}; $(j)$ follows from the fact that $T_{d,d,\ell_1+1}=T_{d,d,\ell+1}+T_{d,\ell+1,\ell_1+1}$; $(k)$ follows from the facts that $T_{d,\ell+1,\ell_1+1}\beta \geq \overline{\mathsf{S}}_{\rightarrow [\ell_1+2:\ell+1]}$ and $(d-\ell)\beta \geq \overline{\mathsf{S}}_{\rightarrow \ell+1}$; $(l)$ follows from \eqref{eq:prop1} and \eqref{eq:prop3} of Proposition
\ref{prop1} and \ref{prop3} respectively; $(m)$ follows from the fact that $T_{d,d,\ell}+T_{d,\ell,\ell_1}=T_{d,d,\ell_1}$. Cancelling $\frac{T_{d,d,\ell_1}+\ell_1}{d-\ell}H(\mathsf{U}^{(\ell_1,\ell)})$ from both sides of the inequality and normalizing both sides by $\sum_{t=\ell+1}^{d}B_t$ complete the proof of \eqref{eq:B4}.
\section{Concluding remarks}\label{sec:Con}
This paper considered the problem of MDC-SR with a generalized eavesdropping model. It was shown that the MBR point of the achievable normalized storage-capacity repair-bandwidth tradeoff region depends on the numbers of type \uppercase\expandafter{\romannumeral1} and type \uppercase\expandafter{\romannumeral2} compromised storage nodes only via their total, as long as the number of type \uppercase\expandafter{\romannumeral1} compromised nodes is less than or equal to the number of type \uppercase\expandafter{\romannumeral2} compromised nodes. Moving forward, it would be interesting to see whether this result extends to the entire achievable normalized storage-capacity repair-bandwidth tradeoff region.
|
\section{Introduction}
We propose a variational method to generalize cubic splines on the space of densities using multimarginal optimal transport. In short, the proposed method consists in minimizing, on the space of measures on the path space, under marginal constraints, the norm squared of the acceleration.
\comment{ In this setting, we show that two numerical approaches, classical in optimal transportation can be applied. One is based on entropic regularization and the Sinkhorn Algorithm, the other relies on the
Semi-Discrete formulation of Optimal Transportation and the computation of Laguerre cells, a classical problem in computationnal geometry.} We showcase our methodology on 1D and 2D data.
\par
In the past few years, higher-order interpolations methods have been investigated for applications in computer vision or medical imaging, for time-sequence interpolation or regression. The most usual setting is when data are modeled as shapes, which can be understood as objects embedded in the Euclidean space with no preferred parametrization: space of unparametrized curves or surfaces, or images are some of the most important examples. These examples are infinite dimensional but the finite dimensional case of a Riemannian manifold was interesting for camera motion interpolation as first introduced in \cite{Noakes1} and further developed in \cite{splinesCk,Crouch}.
Motivated by different applications, the problem of interpolation between two shapes is usually treated via the use of a Riemannian metric on the space of shapes and computing a geodesic between the two shapes. From a mathematical point of view, shape spaces are often infinite dimensional and thus, non-trivial analytical questions arise such as existence of minimizing geodesics or global well-posedness of the initial value problem associated with geodesics. A finite dimensional approximation is still possible such as in \cite{TrVi2010}, in which spline interpolation is proposed for a diffeomorphic group action on a finite dimensional manifold. It has been extended for invariant higher-order lagrangians in \cite{HOSplines1,HOSplines2} on a group, still finite dimensional. A numerical implementation of the variational and shooting splines has been developed in \cite{SinghVN15} with applications to medical imaging. The question of existence of an extremum is not addressed in these publications. An attempt is given in \cite{Vialard2016} where the exact relaxation of the problem is computed in the case of the group of diffeomorphisms of the unit interval. In a similar direction, in \cite{WirthSplines}, the authors discuss the convergence of the discretization of cubic splines in some particular infinite dimensional Riemannian context on the space of shapes.
\par
As a shape space, we are interested in this article in probability measures endowed with the Wasserstein metric. Since the Wasserstein metric shares some similarities with a Riemannian metric on this space of probability densities, it is natural to study further higher-order models in this context. Our motivation is to answer the following practical question of the extension of cubic splines to the Wasserstein space and their numerical computation. \par
We present in Section \ref{SecCubicSplines} the notion of cubic splines on a Riemannian manifold and detail its variational formulation in Hamiltonian coordinates. We then discuss independently in Section \ref{SecHomogeneousSpace} a geometric approach to the Wasserstein space that will be useful for the introduction of our proposed method detailed in Section \ref{modelspline}. Finally in Sections \ref{multimarg} we present the numerical entropic relaxation method and an alternative numerical method based on semi-discrete optimal transport. The reader not interested in geometric interpretation can skip directly to Section \ref{modelspline}. \par
To the best of our knowledge, this question has not been yet addressed in the literature on optimal transport until very recently in two independant and simultaneous preprints~: \cite{georgiouspline} and \cite{ourpreprint} (this paper). Both work share the same idea of relaxing the cubic spline formulation in the space of measure using multi-marginal optimal transport. Our paper however explores a larger hierarchy of models and several numerical methods.
\section{Cubic splines on Riemannian manifolds}\label{SecCubicSplines}
In this section, we present Riemannian cubics, which are the extension of variational splines to a Riemannian manifold $(M,g)$ where $g$ is the Riemannian metric.
Variational cubic splines on a Riemannian manifold are the minimizers of the acceleration; that is, denoting $\frac{D}{Dt}$ the covariant derivative, minimization on the set of curves $x: [0,T] \to M$ of the functional
\begin{equation}
\mathcal{E}(x) = \int_0^1 g(x)\left(\frac{D}{Dt} \dot{x},\frac{D}{Dt}\dot{x}\right) \,\mathrm{d} t\,,
\end{equation}
subject to constraints on the path such as constraints on the tangent space, $(x(t_i),\dot{x}(t_i))$ are prescribed for a collection of times $t_i \in [0,1]$, or constraints on the positions such as $x(t_i) = x_i$.
\par
Under mild conditions on the constraints, if $M$ is complete, minimizers exist, for instance in the case of constraints on the tangent space mentioned above. A pathological case where minimizers might not exist is when the initial speed is not prescribed. Consider for instance the two dimensional torus, where lines of irrational slopes are dense, it is possible to show that for any collection of points which do not lie on a line, the infimum of $\mathcal{E}$ is $0$ while it is never reached, see \cite{WirthSplines}.
The Euler-Lagrange equation associated to the functional $\mathcal{E}$ is
\begin{equation}\label{EqEulerLagrangeEquationSplines}
\frac{D^3}{Dt^3}\dot{x} - R\left(\dot{x},\frac{D}{Dt}\dot{x}\right)\dot{x} = 0\,,
\end{equation}
where $R$ is the curvature tensor of the Riemannian manifold $M$. Note that this equation is similar to a Jacobi field equation.
\par
We now formulate the variational problem in coordinates. In a coordinate chart around a point $x(t) \in M$, the geodesic equations are given by
\begin{equation} \label{RiemannGeodesics}
\frac{D}{Dt} \dot{x} = \ddot{x} + \Gamma(x)(\dot{x},\dot{x}) = 0 \,,
\end{equation}
where $\Gamma$ is a short notation for the Christoffel symbols associated with the Levi-Civita connection. It is a second-order differential equation which is conveniently written as a first-order differential equation, via the Hamiltonian formulation.
Again in local coordinates on $T^*M$ the cotangent bundle of $M$,
the geodesic equation can be written as
\begin{equation}\label{HamiltonianGeodesics}
\begin{cases}
\dot{p} + \partial_x H = 0\\
\dot{x} - \partial_p H = 0\,,
\end{cases}
\end{equation}
where $H(x,p) = \frac 12 g(x)^{-1}(p,p)$. Note that, the ODE \eqref{RiemannGeodesics} can be obtained from the Hamiltonian system using $\dot{x} = g(x)^{-1}p$.
From these two equivalent formulations \eqref{RiemannGeodesics} and \eqref{HamiltonianGeodesics}, it can be shown that $g^{-1}(x)(\dot{p} + \partial_x H) = \frac{D}{Dt} \dot{x} $.
Therefore, it proves that the variational spline problem can be rewritten in Hamiltonian coordinates as follows
\begin{equation*}
\inf_{u} \int_0^1 g(x)^{-1}(a,a) \,\mathrm{d} t \,,
\end{equation*}
under the constraint
\begin{equation*}
\begin{cases}
\dot{x} - g(x)^{-1}p = 0\\
\dot{p} +\partial_x H(x,p) = a\,,
\end{cases}
\end{equation*}
with initial conditions $x(0) = x_0$ and $p(0) = p_0$.
It is natural to ask whether such variational problems carry over in infinite dimensional situations such as the Wasserstein space, which will be discussed in the rest of the paper.
\section{A formal application of spline interpolation to the Wasserstein space}\label{SecHomogeneousSpace}
It is well known that the Hamiltonian formulation of geodesics on the Wasserstein space, define over a riemannian manifold $M$, are
\begin{equation} \label{Geodesics}
\begin{cases}
\dot{\rho} + \nabla \cdot (\rho \nabla \phi) = 0 \\
\dot{\phi} + \frac{1}{2}|\nabla \phi|^2 = 0\, ,
\end{cases}
\end{equation}
where $\rho: M \mapsto {\mathbb R}_{\geq 0}$ and $\phi: M \mapsto {\mathbb R}$ {\em implicitly time dependant } are respectively a probability density and a function. Note that these equations are valid when working with smooth densities. The Hamiltonian is the following,
\begin{equation} \label{Hamiltonian}
H(\rho,\phi) = \frac{1}{2} \int_M |\nabla \phi|^2 \rho \, d\mu_0 \,,
\end{equation}
where $\mu_0$ is a reference measure on $M$.
\begin{rem}
Taking the gradient of the equation governing $\phi$, and denoting $v=\nabla \phi$, we get Burger's equation:
\begin{equation} \label{Burger}
\dot{v} + (v,\nabla)v = 0 \,,
\end{equation}
where in coordinates, the operator $(v,\nabla)$ is defined as $(v,\nabla) w \doteq \sum_{i=1}^n v_i \nabla w_i$ where $v,w$ are vector fields and $n$ is the dimension of the $M$. In Lagrangian coordinates, this equation implies that
\begin{equation}
\ddot{\varphi} = 0 \,,
\end{equation}
where $\varphi (t) :$ $M\mapsto M$ is the Lagrangian flow associated with $v$ ($\dot{\varphi} = v \circ \varphi$), which is well-defined under sufficient regularity conditions.
\end{rem}
\begin{rem}
For the Wasserstein case, the operator is given by $g(\rho)^{-1}\phi = - \nabla \cdot [\rho \nabla \phi] $ so that the (formal) computation of the covariant derivative $\frac{D}{Dt} \dot{\rho}$ on the Wasserstein space is:
\begin{equation} \label{WassersteinConnection}
\frac{D}{Dt} \dot{\rho}= - \nabla \cdot [\rho \, (v + (v,\nabla)v)]\, ,
\end{equation}
where $v = \nabla \phi $ is the horizontal lift associated with $\dot{\rho}$, that is $\dot{\rho} + \nabla \cdot (\rho \nabla \phi) = 0 $. This result is proven rigorously in \cite{lott2008some}.
\end{rem}
From a control viewpoint, we aim at minimizing $\frac{1}{2}\int_0^1H(\rho,a) \, dt$ for the control system:
\begin{equation} \label{ControlledSystem}
\begin{cases}
\dot{\rho} + \nabla \cdot (\rho \nabla \phi) = 0 \\
\dot{\phi} + \frac{1}{2}|\nabla \phi|^2 = a\, ,\\
\end{cases}
\end{equation}
where $a$ is a time dependent function defined on $M$.
Alternatively, in terms of the variables $(\rho,\phi)$, this amounts to minimize
\begin{equation}\label{EqFirstProblem}
\int_0^1 \int_M | \nabla [\dot{\phi} + \frac{1}{2}|\nabla \phi|^2] |^2 \rho \, \,\mathrm{d} \mu_0 \, \,\mathrm{d} t \,,
\end{equation}
under the continuity equation constraint $\dot{\rho} + \nabla \cdot (\rho \nabla \phi) = 0 $.
It is a nonconvex optimization problem in the couple $(\rho,\phi)$. The key issue here is that the variational problem itself is a priori not well-posed since our formulation is valid in a smooth setting and to make it rigorous on the space of measures, the tight relaxation of this problem is needed. However, we do not address this issue in our work and in the next section we turn our attention to a simple relaxation of the problem which is probably not tight.
\section{A hierarchy of relaxed models}\label{modelspline}
\subsection{Context}
We recall the classical optimal transport setting. We have the following well known equivalence \cite{OttoPorousMedium,VilON}
\begin{equation}
\label{equalitystandart}
\begin{array}{ll}
W_2^2\left(\rho_0,\rho_1\right) & \displaystyle = \inf_{\varphi }\int_0^1 \int_M | \dot{\varphi}|^2 \,\mathrm{d} \mu_0 \,\mathrm{d} t =
\inf_{\rho ,v }\int_0^1 \int_M | v |^2 \,\mathrm{d} \rho \,\mathrm{d} t \\[10pt] & = \displaystyle
\inf_{\rho }\int_0^1 \inf_{v } \int_M | v |^2 \,\mathrm{d} \rho \,\mathrm{d} t
=\inf_{\rho ,\nabla \phi }\int_0^1 \int_M | {\nabla \phi }|^2 \,\mathrm{d} \rho \,\mathrm{d} t
\end{array}
\end{equation}
Under constraints that $$ [ \varphi (t) ]_* \mu_0=\rho(t) \mbox{ for }t=0,1$$
($ [ \varphi (t) ]_* \mu_0$ is the image measure of $\mu_0$~:
$\int_M f(y)\,\mathrm{d} [ \varphi (t) ]_* \mu_0 (y) = \int f(T(x)) \,\mathrm{d} \mu(x)$ for every measurable function $f: M \to {\mathbb R}$ ) \\
and the continuity equation
$$\dot{\rho }+\nabla \cdot (\rho v )=\dot{\rho} +\nabla \cdot (\rho \nabla \phi )=0$$
with fixed initial and final conditions
$$ \rho(0) = \rho_0 \mbox{ and } \rho(1) = \rho_1. $$
Moreover, geodesics in the space of densities for the Wasserstein metric are given by \\
$[ \varphi(t) ]_* \mu_0= \rho(t)$ and the associated displacement maps satisfy $v \circ \varphi = \dot{\varphi }$. \\
The last equality in \eqref{equalitystandart} exactly says that the infimum $\inf_{v(t)} \int_M | v(t) |^2 \,\mathrm{d} \rho(t)$ among all $v(t)$ satisfying the continuity equation at each time $t$ is achieved when $v(t)$ is a gradient. This property is a consequence of a Riemannian submersion and $\nabla \phi$ is called the horizontal lift of $\dot{\rho}$. It is this last formulation that formally gives a Riemannian structure on the space of probability measures. See the remark \ref{gs} below for more details on the geometrical structure. \\
For higher-order variational problems, e.g. the minimization of the acceleration, the reduction in the last inequality does not holds true in general, even if the Riemannian submersion structure is present as shown in \cite{HOSplines2}.
It means in the case of acceleration that, a priori, with the same constraint as for \eqref{equalitystandart}~:
\begin{equation}\label{equalitystandart2}
\begin{array}{ll}
\displaystyle \inf_{\varphi } \int_0^1 \int_M | \ddot{\varphi }|^2 \,\mathrm{d} \mu_0 \,\mathrm{d} t & =
\displaystyle \inf_{\rho ,v }\int_0^1 \int_M |\dot{v }+ (v ,\nabla)v |^2 \,\mathrm{d} \rho \,\mathrm{d} t \\[10pt] & \neq
\displaystyle \inf_{\rho ,\nabla \phi }\int_0^1 \int_M | {\dot{\phi} +(\nabla \phi ,\nabla)\nabla \phi }|^2 \,\mathrm{d} \rho \,\mathrm{d} t ,
\end{array}
\end{equation}
where we have used that $\ddot{\varphi }=\dot{v }\circ \varphi + (v \circ \varphi ,\nabla)v \circ \varphi $. \\
\begin{remark}
\label{gs}
From a geometrical point of view, \eqref{equalitystandart} says the Wasserstein space can be seen, at least formally, as a homogeneous space as described in \cite[Appendix 5]{khesin2008geometry} and originally in \cite{OttoPorousMedium}.
Consider the group of (smooth) diffeomorphisms of $M$ a closed manifold, $\on{Diff}(M)$, and the space of (smooth) probability densities $\on{Dens}(M)$.
The space of densities is endowed with a $\on{Diff}(M)$ action defined by the pushforward, that is to a given $\varphi \in \on{Diff}(M)$ and $\rho \in \on{Dens}(M)$, the pushforward of $\rho$ by $\varphi$ is $\on{Jac}(\varphi^{-1})\rho \circ \varphi^{-1}$.
By Moser's lemma, this action is transitive, thus making the space of densities as a homogeneous space.
More importantly, there exists a compatible Riemannian structure between $\on{Diff}(M)$ and $\on{Dens}(M)$.
Once having chosen a reference density $\mu_0$, the $L^2(M,\mu_0)$ metric on the diffeomorphism group descends to the Wasserstein $L^2$ metric on the space of densities, or in other words, the pushforward action $\varphi \mapsto \varphi_*\mu_0$ is a Riemannian submersion.
An important property of Riemannian submersion is that geodesics on $\on{Dens}(M)$ are in correspondence with geodesics on the group, given by horizontal lift.
This property is actually contained in Brenier's polar factorization theorem, which shows that the horizontal lift is the gradient of a convex function.
\end{remark}
\subsection{The Monge formulation}
In Section \ref{SecHomogeneousSpace} we used the formal Riemannian structure on the set of probability measure to define an intrinsic notion of splines, \eqref{EqFirstProblem} is indeed the RHS of inequality \eqref{equalitystandart2}.
In this section we propose a simpler alternative definition of Wasserstein splines based on the LHS of inequality \eqref{equalitystandart2}.
\begin{definition}[Monge formulation]\label{Mongeformulation}
Let $0 = t_0 < \ldots < t_n =1$, $n \geq 2$ and $\rho_1, \ldots, \rho_n $ be $n$ probability measures on $M$.
\par
Minimize, among time dependent maps $\varphi(t): M \mapsto M$,
\begin{equation}\label{EqSecondProblem}
\int_0^1 \int_M | \ddot{\varphi} |^2 \,\mathrm{d} \mu_0 \,\mathrm{d} t \,,
\end{equation}
under the marginal constraints $\varphi(t_i)_* \mu_0 = \rho_i$.
This minimizing problem is denoted by $(MS)$.
\end{definition}
It is a Monge formulation of the variational problem, similar to standard optimal transport.
On a Riemannian manifold $M$, the notation $\ddot{\varphi}$ stands for $\frac{D}{Dt}{\dot{\varphi}}$.
By the change of variable with the map $\varphi$, the problem can be written in Eulerian coordinates, that is using the vector field associated with the Lagrangian map $\varphi$, $\partial_t \varphi = v \circ \varphi$, one aims at minimizing for $(\rho,u)$
\begin{equation}
\int_0^1 \int_M |u|^2 \rho \,\mathrm{d} \mu_0 \,\mathrm{d} t\,
\end{equation}
under the constraints
\begin{equation}
\begin{cases}
\dot{\rho} + \on{div} (\rho v) = 0 \\
\dot{v} + (v,\nabla)v = u\,,
\end{cases}
\end{equation}
with the marginals constraints $\rho(t_i) = \rho_i$.
\begin{remark}
Remark that formally when $v = \nabla \phi$, this new model reduces to the formulation \eqref{EqFirstProblem}. Therefore, it justifies the fact that Problem \eqref{EqSecondProblem} is a relaxation of \eqref{EqFirstProblem}. However, as already mentioned, this relaxation is probably not tight.
\par
Another formal geometric argument in the direction of proving that the two formulations are different is that the Wasserstein space has nonnegative curvature if the underlying space $M$ has nonnegative curvature, but the space of maps in the Euclidean space is flat. Therefore, the two Euler-Lagrange equations \eqref{EqEulerLagrangeEquationSplines} lead to a different evolution equations: for instance, if $M$ is the Euclidean space then the Euler-Lagrange equation for the second model is simply $\ddddot{\varphi} = 0$, which is a priori different from the splines Euler-Lagrange equation in the Wasserstein case.
\end{remark}
\subsection{The Kantorovich relaxation}
Since, as is well-known in standard optimal transport, the Monge formulation is not well-posed for general given margins $\rho_1, \ldots, \rho_n $, we propose instead to solve yet another relaxation of the problem on the space of curves which takes the form:
\begin{definition}[Kantorovich relaxation]\label{ThmKantorovich}
Let $0 = t_1 < \ldots < t_n =1$, $n \geq 3$ and $\rho_1, \ldots, \rho_n $ be $n$ probability measures on $M$.
\par
Minimize on the space of probability measures on the path space $H^2([0,1],M)$ denoted by $\mathcal{H}$ in short,
\begin{equation}\label{EqMultiMarginal}
\min_{\mu} \int_{\mathcal{H}} |\ddot{x}|^2 \,\mathrm{d} \mu(x) \,,
\end{equation}
which is a linear functional of $ \,\mathrm{d} \mu$. The curves of densities is given by its marginals in time
\begin{equation}\label{TimeMarginals}
t \mapsto \rho(t) \mu_0 := [e_t]_*(\mu)\,,
\end{equation}
$e_t$ is the evaluation function at time $t$~:
if $\gamma \in H^2([0,1],M) \subset C^0([0,1],M)$ then $e_t(\gamma) = \gamma(t,.) \in M$. \\
The notation $[e_t]_*\mu$ is the image measure by the map $e_t$ defined by duality~:\\ $\int_M f(y)\,\mathrm{d} [e_t]_* \mu(y) = \int_{\mathcal{H}} f(e_t(x)) \,\mathrm{d} \mu(x)$ for every measurable function $f: M \to {\mathbb R}$.
Note that $x$ is a path on $[0,1]\times M$ while $y$ is a point on $M$.
With these notations, the marginal constraint at given time $t_i$ are
\begin{equation}\label{EqMarginalConstraints}
[e_t]_*(\mu) = \rho_i \, \mu_0\,.
\end{equation}
\end{definition}
By standard arguments, the Kantorovich relaxation admits minimizers under general hypothesis on the manifold $M$, which we do not detail here. It is straightforward to check that existence of minimizers holds when $M = {\mathbb R}^d$.
\par
As expected, the Kantorovich formulation is the relaxation of the Monge formulation in Definition \ref{Mongeformulation}.
\begin{theorem}\label{ThmRelaxation}
Let $M = {\mathbb R}^d$, $0 = t_1 < \ldots < t_n =1$, $n \geq 3$ and $\rho_1, \ldots, \rho_n \in $ be $n$ probability measures on ${\mathbb R}^d$ with compact support and $\rho_1$ being atomless. Then, under the constraints \eqref{EqMarginalConstraints}, the infimums of the variational problem \eqref{EqSecondProblem} and \eqref{EqMultiMarginal} coincide, moreover, the infimum is attained for the latter.
\end{theorem}
\begin{proof}
See the proof of a more general result in Appendix \ref{Appendix}.
\end{proof}
First we remark that we can reformulate both the Monge and Kantorovich problems on the set of cubic splines.
It is the purpose of the following lemmas and corollaries, whose proofs are straightforward.
\begin{definition}[Cubic interpolant]
Let $(x_1,\ldots,x_n) \in {\mathbb R}^d$ be $n$ given points and $(t_1< \ldots<t_n)$ be $n$ timepoints.
There exists a unique cubic spline minimizing the acceleration of the curve $x(t)$ such that $x(t_i) = x_i$.
This unique curve is called cubic interpolant and is denoted by $c_{x_1,\ldots,x_n}$, depending implicitly on the timepoints.
\end{definition}
\begin{lemma}\label{ThmReduction}
When the supports of the measures $\rho_i$ are compact on ${\mathbb R}^d$,
the support of every minimizing $\mu$ in Definition \ref{ThmKantorovich} is included in the set the cubic interpolants $c_{x_1,\ldots,x_n}$ for $(x_1,\ldots,x_n) \in \on{Supp}(\rho_1) \times \ldots \times \on{Supp}(\rho_n)$.
\end{lemma}
\begin{proof}
The constraints are the marginal constraints $[e_{t_i}]_*(\mu) = \rho_i$ for $i \geq 3$ which implies that set of paths charged by an optimal measures satisfies $x(t_i) \in \on{Supp}(\rho_i)$.
In particular, any path in this set can be replaced by its minimal spline energy, the cubic interpolant $c_{x_1,\ldots,x_n}$.
\end{proof}
\begin{corollary}
As a consequence, the set of paths charged by an optimal plan are uniformly $C^2$ and for every smooth function $\eta: {\mathbb R}^d \mapsto {\mathbb R}$ with compact support, the map $t \mapsto \langle \mu(t), \eta \rangle$ is $C^2$.
\end{corollary}
\begin{proof}
The set of cubic interpolants is compact since the map $(x_1,\ldots,x_n) \mapsto c_{x_1,\ldots,x_n}$ is continuous from ${\mathbb R}^{dn}$ to the space of $C^2$ fonctions (solution of an invertible linear system) and $\on{Supp}(\rho_i)$ are compact. Therefore, the set of maps are uniformly $C^1$. The last point follows directly.
\end{proof}
\begin{corollary}\label{ThmMultiMarginal}
The Kantorovich problem in Definition \ref{ThmKantorovich} on ${\mathbb R}^d$ reduces to a multimarginal optimal transport problem, as follows, let $c(x_1,\ldots,x_n)$ be the continuous cost of the cubic interpolant at times $t_1,\ldots,t_n$, the minimization of \eqref{EqMultiMarginal} reduces to the minimization of
\begin{equation}\label{eqK}
\int_{M^n} c(x_1,\ldots,x_n) \,\mathrm{d} \pi(x_1,\ldots,x_n) \phantom{1111111} \textrm{$(K)$}
\end{equation}
on the space of probability measures $\pi \in \mathcal{P}(M^n)$ and under the marginal constraints $(p_i)_*(\pi) = \rho_i$ where $p_i$ is the projection of the $i^{\text{th}}$ factor.
\end{corollary}
\begin{proof}
Direct consequence of Lemma \ref{ThmReduction}.
\end{proof}
Similarly
\begin{corollary}\label{ThmMultiMarginalmonge}
The Monge problem in Definition \ref{Mongeformulation} on ${\mathbb R}^d$ reduces to a Monge multimarginal optimal transport problem, as follows, let $c(x_1,\ldots,x_n)$ be the continuous cost of the cubic interpolant at times $t_1,\ldots,t_n$, the minimization of \eqref{EqSecondProblem} reduces to the minimization of
\begin{equation}\label{eqMmulti}
\int_{M} c\left(x, \varphi(t_1,x),\ldots,\varphi(t_n,x) \right) \,\mathrm{d} \mu_0(x),
\end{equation}
on the space of path $\varphi \in C^2([0,1],M)$ (or even cubic splines) and under the marginal constraints $(\varphi(t_i))_*\mu_0 = \rho_i$.
\end{corollary}
The dual formulation of the minimization problem $(K)$ is also well known \cite[Theorem 2.1]{PassKim2013}
\begin{definition}[Kantorovich dual problem $(KP)$]
Let $ \mathcal{Q}=\left\{ \phi_i \in L^1(\rho_i \, \mu_0)\, , i=1..n \right\}$ be the space of integrable $n$-uplet. Maximize on $ \mathcal{Q}$
\begin{equation}\label{KP}
\sum_{i=1}^{n} \int_{M} \phi_i \rho_i \, \mu_0, \mbox{ under the constraint }\sum_{i=1}^{n} \phi_i (x_i) \leq c(x_1,...,x_n).
\end{equation}
\end{definition}
And the following duality results holds true:
\begin{proposition}
There exists a $n$-uplet $(\phi_i)_{i=1..n}\in \mathcal{Q}$ optimal for $(KP)$. Moreover $(K)$=$(KP)$ and for any $\pi$ optimal in \eqref{eqK} there holds $\sum_{1}^{n} \phi_i (x_i) = c(x_1,...,x_n) $, $\pi$ almost everywhere.
\end{proposition}
A natural question is whether the solution of the Kantorovich problem $(K)$ is admissible in the Monge formulation $(MS)$ (Definition \ref{Mongeformulation}).
With the formulation reduced above the spline, given by \eqref{eqK} and \eqref{eqMmulti}, one can try to apply existing theory to answer to this question, see \cite{ PassKim2013, Passreview} and references therein for precise criterion.
However our cost does not satisfy any of those known criterion.
In fact, we have the following result which proves that the relaxation to plans are necessary even in the context of Theorem \ref{ThmRelaxation}.
\begin{proposition}(Counter Example) \label{ThCounterExample}
Given the three-marginals problems of minimizing the acceleration, there exist data $(\rho_0,\rho_1,\rho_2)$ such that $\rho_0$ is atomless and such that the solution of $(K)$ is not a (measurable) Monge map.
\end{proposition}
\begin{proof}
Consider $\rho_0(x) = \mathbf{1}_{[-1,1]}$ and the Dirac masses $a = \delta_1$ and $b = \delta_{-1}$ and the maps $T_a,T_b$ that respectively pushforward $\rho_0$ onto $a$ and $b$. These maps are uniquely determined and affine. Consider now $\rho_2 = \frac 12 (T_a)_*\rho_0 + \frac 12 (T_b)_*\rho_0 = \frac a2 + \frac b2$. Then, introducing $(T^{1/2}) = \frac12 (\on{Id} + T)$, we consider
$\rho_1 = \frac12 (T^{1/2}_a)_*\rho_0 + \frac12 (T^{1/2}_b)_*\rho_0$, note that it is equal to $\rho_0$ since the maps $T^{1/2}_{a,b}$ are affine.
\par
By construction, the minimization of the acceleration for $(\rho_0,\rho_1,\rho_2)$ is null since it is a mixture of plans supported by straight lines. If there existed an optimal Monge solution it is necessarily supported by only one map denoted by $T$ and since the cost is null, the map at time $1/2$ is necessarily $T^{1/2}$ defined above. The preimage of $1$ (resp. $-1$) by $T$ is a measurable set $A$ (resp. $B$). Then, necessarily, $\rho_1 = (T^{1/2})_*\chi_A + (T^{1/2})_*\chi_B$, and in fact, $T_{|A} = T_a$ and $T_{|B} = T_b$ (since the image of the map is known). Therefore, we have $\rho_1 = 2 \chi_A \circ (T_a^{1/2})^{-1} + 2 \chi_B \circ (T_b^{1/2})^{-1}$ which is not equal to the uniform Lebesgue measure on $[-1,1]$.
\end{proof}
\begin{figure}
\centering
\begin{tikzpicture}[>=triangle 45,font=\sffamily]
\draw[->] (0,0) -- (5,0)node[anchor= north] {\text{time}};
\draw[->] (0,-3) -- (0,3)node[anchor=east] {\text{${\mathbb R}$}};
\draw[thick,,line width=1mm,blue] (-0.05,-2) -- (-0.05,2)
\draw[thick,,line width=1mm,red] (0.05,-2) -- (0.05,2) ;
\draw[fill=red!40,opacity=0.5] (0,-2) --(0,2)--(4,2)--(0,-2);
\draw[fill=blue!40,opacity=0.5] (0,-2) --(0,2)--(4,-2)--(0,-2);
\node at (4,-2) {$\bullet$};
\node at (4,-2.3) {$-1$};
\node at (4,2) {$\bullet$};
\node at (4,2.3) {$1$};
\node at (4,0) {$\bullet$};
\node at (4,-0.3) {$t = 2$};
\node at (2,0) {$\bullet$};
\node at (2,-0.3) {$t = 1$};
\node at (0,0) {$\bullet$};
\node at (-0.6,-0.3) {$t = 0$};
\end{tikzpicture}
\caption{The inital density at time $0$ is described with a mixture of two densities colored in red and blue which are evolving indepently along straight lines in time. The blue density is mapped onto $-1$ and the red density is mapped onto $1$. The acceleration cost is null and the proof of Proposition \ref{ThCounterExample} shows that it is not possible to reproduce the density at time $1/2$ by a map.}
\end{figure}
\begin{remark}
It is an open question to prove or disprove a similar result when the final density $\rho_2$ is atomless.
The counterexample explained above strongly uses the fact that the final density is a sum of Dirac masses and it might not be robust when replacing the final density by a uniform density on a small interval.
\end{remark}
\subsection{The corresponding interpolation problem on the tangent space}\label{SecPhaseSpace}
The relaxed problem on the space of curves can be used to define variational interpolation problem on the phase space, or more precisely on the tangent space $TM$. Since the space $H^2([0,T],M)$ is contained in $C^1([0,T],M)$, one can formulate the optimal transport problem on phase space (identified with the tangent space) for the acceleration cost.
\begin{definition}[Optimal transport on phase space]\label{costphase}
Let $\bar{\rho_0},\bar{\rho_1} $ be two probability measures on $TM$.
Minimize on the space of probability measures on $\mathcal{H}$,
\begin{equation}\label{EqMultiMarginal2}
\min_{\mu} \int_{\mathcal{H}} |\ddot{x}|^2 \,\mathrm{d} \mu(x) \,,
\end{equation}
which is a linear functional of $ \mu$ under the marginal constraints
\begin{align}\label{EqMarginalConstraints2}
[j_0]_*(\mu) = \bar{\rho_0 }\,, \text{ and } [j_1]_*(\mu) =\bar{ \rho_1}\,,
\end{align}
where $j_t: H^2([0,T],M) \to TM$ is defined by $j_t(x) = (x(t),\dot{x}(t))$.
\end{definition}
\begin{proposition}[Optimal interpolation on phase space]\label{ThMongeInPhase}
The support of every optimal solution is contained in the set of cubic splines interpolating between $(x,v) \in \on{Supp}(\bar{\rho_0})$ and $(y,w) \in \on{Supp}(\bar{\rho_1})$.
Moreover if $M = {\mathbb R}^d$ and if $ \bar{\rho_0}$ has density with respect to the Lebesgue measure, then the unique solution to Problem \eqref{EqMultiMarginal2} is characterized by a map $\varphi:$ $TM \mapsto TM$.
\end{proposition}
Remark that the optimal solution in the last part of Proposition \ref{ThMongeInPhase} provides an interpolation on the phase space using $[j_t]_*(\mu)$.
\begin{proof}
The proof of the first part is similar to Lemma \ref{ThmReduction} and the second part follows by application of Brenier's theorem since the total cost of the cubic splines between $(x,v)$ and $(y,w)$ can be explicitly computed as
\begin{equation}\label{EqCostTM}
c_{ph}((x,v),(y,w)) = 12|x-y|^2 + 4(|v|^2 + |w|^2 + \langle v,w \rangle + 3 \langle v+w, x-y \rangle)\,
\end{equation}
and satisfies the twisted condition, so \cite[Theorem 10.28]{VilON} applies.
\end{proof}
Note that this problem is very different from using the Wasserstein distance on $\mathcal{P}(TM)$ where the tangent space $TM$ is endowed with the direct product metric.
Indeed, the cost $c_{ph}$ does not vanish on the diagonal $(x,v) =(y,v)$ contrarily to the quadratic cost on $TM$.
\par
Interestingly, let us remark that the multimarginal problem can be recast as the minimization problem on $\Pi \in \mathcal{P}(\underbrace{TM \times \ldots \times TM}_{\text{n times}})$, denoting $\Pi_{t_i,t_{i+1}}$ the pushforward on $TM \times TM$ at times $(t_i,t_{i+1})$,
\begin{equation}
\min_{\pi} \sum_{i = 1}^{n-1} \langle \Pi_{t_i,t_{i+1}} ,c_{ph}((x_i,v_i),(x_{i+1},v_{i+1})) \rangle
\end{equation}
under the constraints that $[e_{t_i}]_*(\Pi_{i,i+1}) = \rho_i$. From the numerical point of view, this rewriting might be useful since the cost used on the multimarginal problem is now separable in time. This relaxation to the tangent space is used in the semidiscrete algorithm in Section \ref{SecImplementationSemiDiscrete}.
Obviously, up to the minimization on the variables $v_i$, we retrieve the minimization problem $(K)$ since one has a cost $c$ which is defined on $M^n$
\begin{equation}
\label{psc}
c(x_0, \ldots, x_n) = \min_{v_0, \ldots ,v_n} \sum_{i=1}^{n-1} c_{ph}((x_i,v_i),(x_{i+1},v_{i+1}))
\end{equation}
where the index $i$ runs over the marginals.
\section{Numerical Study }\label{multimarg}
\comment{We have discussed several variational relaxation of the classical definition of splines, applied to the Wasserstein space of densities.
At least two different numerical techniques from Optimal Transportation can be used in this setting. We apply the Entropic regularisation and Sinkhorn (briefly
recalled in appendix \ref{App2} first to a simple Hermite interpolation problem (section \ref{Herm}) and then in section to the multimarginal problem (\ref{eqK}).
In section \ref{SD}, we use the semi-discrete Optimal Transportation approach in the spirit of \cite{memi} directly to problem (\ref{EqMultiMarginal})
without the time discretisation in (\ref{eqK}).}
\subsection{Hermite interpolation}
\label{Herm}
In this section, we are interested in the problem of interpolation on the phase space described in the previous.
The marginals $[e_t]_*(\mu)$ are densities defined on the tangent space $TM$. If we only specify
the marginals at time $0$ and $1$ as empirical measures: $[e_0]_*(\mu) = \sum_{i=1}^k \alpha_i \, \delta_{x_i} \delta_{v_i} $ and
$[e_1]_*(\mu) = \sum_{j=1}^k \beta_j \, \delta_{y_j} \delta_{w_j} $, as explained in Section \ref{SecPhaseSpace}, we can simplify the Kantorovich using the exact $L^2$ norm of the acceleration of the spline between
$(x_v)$ and $(y,w$), whose cost is given in Formula \eqref{EqCostTM}.
Again, let us underline that this cost is \emph{not} a Riemannian cost on the tangent space of ${\mathbb R}^d$ since if $v = w$ and $x,y$ are close, the cost is dominated by the term $4(|v|^2 + |w|^2 + \langle v,w \rangle)$ which need not be zero. Then, the Kantorovich problem reduces to the minimization of
\begin{equation}
\sum_{i,j = 1}^{k,l} \pi_{i,j} c((x_i,v_i),(y_j,w_j)) \,,
\end{equation}
under the constraints
\begin{equation}
\begin{cases}
\sum_{i = 1}^k \pi_{i,j} = \beta_j \,\\
\sum_{j = 1}^l \pi_{i,j} = \alpha_i \,.
\end{cases}
\end{equation}
It is straightforward to apply entropic regularization/Sinkhorn in this case
which amounts to add, for a positive parameter $\varepsilon$, $\varepsilon \sum_{i,j} \pi_{i,j} \log(\pi_{i,j})$ to the previous linear functional and to numerically solve the corresponding variational problem with the Sinkhorn algorithm \cite{Sinkhorn67,Cut} (See also appendix \ref{App2} where Sinkhorn algorithm is detailed in the more general multimarginal case).
It is interesting to note that the choice of $\varepsilon$ is more delicate than in the standard case of a quadratic distance cost. \\
In Figure \ref{FigHermite}, we present the convergence rate of this method with respect to two different values of $\varepsilon$ and the most likely deterministic plan given the optimal plan $\pi^\varepsilon$. Note that this entropic regularization method scales with the number of points as $N^2$ and is valid in every dimension.
\begin{figure}
\centering
\begin{tabular}[h]{cc}
\includegraphics[width=8cm]{ConvergenceCurve}
&
\includegraphics[width=8cm]{SimulationHermite}
\end{tabular}
\caption{Convergence (left) and Hermite interpolation problem between Two empirical measure in phase space (right). We represent the most likely splines in the position space. }
\label{FigHermite}
\end{figure}
\par
\subsection{MultiMarginal formulation}
This is the direct discretization of \eqref{EqMultiMarginal} which avoids working in phase space with the cost (\ref{psc}) thus enabling fast computations in 2D.
In what follows, the time cylinder $[0,1] \times M $ is discretized in time as $ \bigotimes_{i=0,N} M_i$, the product space of $N+1$ copies of $M$ at each of the
$N+1$ time steps. We will use a regular time step discretization $\tau_i = i\, d\tau$ where $d\tau =\frac{1}{N}$.
Using a classic finite difference approach, the time discretization of (\ref{EqMultiMarginal}) is
\begin{equation}
\label{JN}
\min_{\mu_{d\tau} } \int_{\bigotimes_{i=0,N} M_i } c_{d\tau} (x_1,...,x_N) \,\mathrm{d} \mu_{d\tau}(x_1,..,x_N) \,,
\end{equation}
where $\mu_{d\tau}$ now spans the space of probability measures on $\bigotimes_{i=0,N} M_i$ representing\
the space of piecewise linear curves passing through $x_0,x_1,...,x_N$ at times $\tau_0, ..., \tau_N$. \\
A straighforward computation gives
\begin{equation}
\label{acost}
c_{d\tau} (x_1,...,x_N) := \sum_{i=1,N-1} \dfrac{ \| x_{i+1}+x_{i-1}-2\, x_i \|^2 }{d\tau^3}
\end{equation}
For all times, marginals (\ref{TimeMarginals}) are computed as :
\begin{equation}
\label{TM}
\tau_j \mapsto \int_{\bigotimes_{i \neq j } M_i } \,\mathrm{d} \mu_{d\tau} (x_1,..x_N)
\end{equation}
In order to simplify the presentation we will assume that the marginal constraints (\ref{EqMarginalConstraints})
are set at times $t_1,..t_n$ which coincide with times steps of the discretization (of course $n < N$, meaning
the number of constraint is not the same as the number of time steps).
In short, there exist $(j_1,..j_n) \in [0,N]$ such that
\[
(t_1,..,t_n) = (\tau_{j_1}, ... , \tau_{j_n}).
\]
The constraint (\ref{EqMarginalConstraints}) becomes for all $k = 1,..n$
\begin{equation}
\label{JC}
\int_{\bigotimes_{i \neq j_k } M_i } \,\mathrm{d} \mu_{d\tau} (x_1,..x_N) = \rho_{j_k} ( x_{j_k})
\end{equation}
where $\rho_{j_k}$ is the prescribed density to interpolate at time $\tau_{j_k} = t_k$. \\
The time discretized problem is the multimarginal problem (\ref{JN} -\ref{JC}). \\
The simplest space discretization strategy is to use a regular cartesian grid. In dimension 2 and for $M = [0,1]^2$ and at time $t_i$, the grid
will be denoted $x_{\alpha_i,\beta_i} = (\alpha_i \, h,\beta_i h)$ for $(\alpha_i,\beta_i) \in [0,N_x]$ and $h = \frac{1}{N_x}$, $ a = \{\alpha_i\}$
and $b = \{\beta_i \}$ will be the vectors of indices.
The time and space discretization of the problem then becomes
\begin{equation}
\label{JND}
\min_{ T} \sum_{a,b} C_{a,b} \, T _{a,b}
\end{equation}
Where $T$ is the $N\times N_x \times N_x$ tensor of grid values $\mu_{d\tau}(x_{\alpha_1,\beta_1} ,..,x_{\alpha_N,\beta_N})$ and
\begin{equation}
\label{acostD}
C_{a,b} = c_{d\tau} (x_{\alpha_1,\beta_1} ,..,x_{\alpha_N,\beta_N})
\end{equation}
The marginals (\ref{TM}) at all times $\tau_j$ are given by
\begin{equation}
\label{TMD}
\sum_{a \setminus \{\alpha_{j} \} , \, b \setminus \{\beta_{j} \} } T _{a,b}
\end{equation}
The constraints (\ref{JC}) therfore becomes for all $k$
\begin{equation}
\label{JCDk}
\sum_{a \setminus \{\alpha_{j_k} \} , \, b \setminus \{\beta_{j_k} \} } T _{a,b} = \rho_{j_k} ( x_{\alpha_{j_k}, \beta_{j_k}} )
\end{equation}
$a \setminus \{\alpha_{j_k} \} $ denotes the set of indices $a$ minus $\alpha_{j_k}$. \\
The Entropic regularized problem is
\begin{equation}
\label{JNDe2}
\min_{ T^\epsilon} \sum_{a,b} \{ C_{a,b} \, T^\epsilon_{a,b} + \epsilon \, T^\epsilon_{a,b} \, \log(T^\epsilon_{a,b}) \}
\end{equation}
and easier to solve. See Appendix \ref{App2} for a description of Sinkhorn algorithm.
\subsection*{Numerical Simulations}
\paragraph{\textbf{1D case: }}We present, figures \ref{FigBasicExample} and \ref{FigBasicExample2}, a 1D test case to highlight some of the qualitative properties of the cubic splines interpolation on the space of densities.
\begin{figure}[htb]
\begin{center}
\begin{tabular}{cc}
\begin{tabular}{c}
$\null$\hspace{-0.35cm}
\includegraphics[width= 6cm]{SecondTest1.png}
\\
Initial time data and targets
\end{tabular}
\\
\begin{tabular}{ccc}
$\null$\hspace{-0.35cm} \includegraphics[width= 5.1cm]{SecondTest3.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{SecondTest6.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{SecondTest7.png}\\
$\null$\hspace{-0.35cm} \includegraphics[width= 5.1cm]{SecondTest9.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{SecondTest13.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{SecondTest16.png}
\end{tabular}
\end{tabular}
\caption{Four interpolation timepoints, $1,6,11,16$ and representation of the four density configurations, as well as $6$ intermediate times.
The doted line represent the reconstructed density curve in time.
This experiment underlines that the spline curve has more smoothness in time and can present some concentration or diffusion effects depending on the data which would not be present for the usual Wassertein geodesic. The entropic regularization parameter is $\varepsilon = 8.10^{-5}$.
}
\label{FigBasicExample}
\end{center}
\end{figure}
We consider four interpolation time points and the corresponding data are mixture of Gaussians of different standard deviations. We use a discretization of $140$ points on the interval $[0,1]$ with $16$ time steps. The doted line represent the reconstructed density curve in time. This experiment shows that the mass can concentrate or diffuse in some situation.
Another important point here is that the entropic regularization parameter has an important impact on this concentration/diffusion effects: we show the simulations for $\varepsilon = 0.002$ and $\varepsilon = 8.10^{-5}$. In the simulation with a large $\varepsilon$, the concentration effect is not present and it is due to the diffusion on the path space.
\begin{figure}[htb]
\begin{tabular}{ccc}
$\null$\hspace{-0.35cm} \includegraphics[width= 5.1cm]{ThirdTest3.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{ThirdTest6.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{ThirdTest7.png}\\
$\null$\hspace{-0.35cm} \includegraphics[width= 5.1cm]{ThirdTest9.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{ThirdTest13.png}&
$\null$\hspace{-0.45cm} \includegraphics[width= 5.1cm]{ThirdTest16.png}
\end{tabular}
\caption{The same experiment with a larger entropic regularization parameter $\varepsilon = 0.002$. As expected, we observe less concentration of mass.
}
\label{FigBasicExample2}
\end{figure}
\paragraph{\textbf{2D case: }}We present a 2D test case which computes a Wasserstein spline in the sense of (\ref{JND}) interpolating four Gaussian identical
densities at time 1, 5, 13, and 17, see figure \ref{f2d1}. We use a time step $d\tau =1$ and 17 $N=17$ time steps. The space discretization is $Nx = 50$.
The entropic regularization parameter is $\epsilon = 0.002$, note that the stability of the method depends on this parameter.
It also generates artificial diffusion as it becomes more costly top concentrate the available mass on fewer Euclidean splines between the points of the support of the
four Gaussians. We can compute the interpolating densities at intermediate times using (\ref{TMD}) but is more
interesting to represent in figure \ref{f2d2} the contour line of the third quartile, i.e. the highest values of the densities representing 1/4 of the total mass.
Comparing with figure \ref{rotation2}, it seems clear that the Entropy diffusion spreading pollutes the solution of the original problem (without entropic regularization). \\
We compare this solution with the classical Quadratic cost Optimal Transport interpolation, i.e. with the speed instead of the acceleration
in the cost. More precisely taking :
\begin{equation}
\label{acost2}
c_{d\tau} (x_1,...,x_N) := \sum_{i=0,N-1} \dfrac{ \| x_{i+1}- x_i \|^2 }{d\tau}
\end{equation}
As expected the mass follows respectively the linear interpolation or the Euclidean spline interpolation of the center of the Gaussians which
are represented as thick red lines in figure \ref{f2d1}. \\
Finally we show the convergence of the Sinkhorn iterate for both simulations in figure \ref{f2d2}.
The convergence is much slower for the speed case but we did not optimize the implementation which does not need tensors and instead
just used a degraded version of the acceleration code. This may be the reason for this strange difference.
\begin{figure}
\centering
\begin{tabular}[h]{cc}
\hspace{-2cm} \includegraphics[width=10cm]{pap2d1}
&
\hspace{-2cm} \includegraphics[width=10cm]{pap2d2}
\end{tabular}
\caption{ Spline interpolation of Four Gaussians with 17 times steps. Left : the data and the linear and classic cubic spline
interpolation of the of Gaussian center point. Right : the level curve of the third quartile of the density every 2 time steps, in solid line for our Spline Wasserstein interpolation and
in dashed line for the classic quadratic cost (speed) interpolation. }
\label{f2d1}
\end{figure}
\begin{figure}
\centering
\begin{tabular}[h]{cc}
\hspace{-1cm}\includegraphics[width=8cm]{convacc}
&
\hspace{-1cm}\includegraphics[width=8cm]{convspeed}
\end{tabular}
\caption{ Convergence, i.e. Infinity norm of the difference of the Dual unknown between to Sinkhorn iteration. This is computed every 10 iterations.
Left :for the acceleration cost, right : for the speed cost . }
\label{f2d2}
\end{figure}
\subsection{Semi-Discrete approach}\label{SD}
We propose another numerical scheme based on the semi-discrete approach introduced by M\'erigot in \cite{Mer} in dimension 2 and developed by Levy \cite{Levy} in dimension 3. Here we approximate the optimal plan $\pi$ in the formulation \eqref{eqK} by a sum of N tensor product of diracs masses. That is $\pi_N =\sum^N_{j=1} \left( \bigotimes^n_{i=1} \frac1N \delta_{X^i_j}\right)= \sum^N_{j=1} \frac1N \delta_{\left( X^1_j,\ldots,X^n_j\right)} $.
\begin{remark}
Since there is a unique corresponds between $n$ points $\left( X^1_j,\ldots,X^n_j\right)$ and the spline $c_{X^1_j,\ldots,X^n_j }$ passing through these points at time $(t_1,\ldots.t_n)$ the measure $\pi_N $ can also be seen as $N$ direct masses defined over the set of splines: $\pi_N = \sum^N_{j=1} \frac1N \delta_{c_{ X^1_j,\ldots,X^n_j}}$.
\end{remark}
We then have to relax the constraint $(p_i)_*(\pi) = \rho_i$ since $(p_i)_*(\pi_N)= \sum^N_{j=1} \frac1N \delta_{X^i_j} $ cannot be absolutely continuous.
It leads to the following variational problem.
\begin{definition}[Semi-discrete variational problem]\label{SDVdef}
Let $\epsilon>0$, $0 = t_1 < \ldots < t_n =1$, $n \geq 3$ and $(\rho_i)_{i=1\ldots n}$ be n absolutely continuous measures. Recall that $c(Y_1,\ldots,Y_n)$ is the cost of the cubic spline passing through the points $(Y_1,\ldots,Y_n)$ at time $(t_1,\ldots.t_n)$.
Let
$$\mathcal{Q}^N = \left\{ \sum^N_{j=1} \frac1N \delta_{\left( X^1_j,\ldots,X^n_j \right) } \middle| (X_j)_{j=1,\ldots,N}\in M^{n} \right\}.$$ Then the semi-discrete variational problem, (SDV), is given by
\begin{equation}\label{SDV}
(SDV) = \min_{\mathcal{Q}^N} \, \frac{1}{N}\sum_{j=1}^N c(X^1_j,\ldots, X^n_j) + \sum^n_{i=1} \frac{1}{2\epsilon^2} W_2^2\left(\sum^N_{j=1} \frac1N \delta_{X^i_j} , \rho_i \right),
\end{equation}
where $W_2$ is the classical Wasserstein distance given by the quadratic cost.
\end{definition}
The main drawback of this method is that, as illustrated in the numerical simulations below, the problem $(SDV)$ is not convex.
\subsubsection{Implementation} \label{SecImplementationSemiDiscrete}
In order to solve numerically the minimization problem $(SDV)$ we use the reformulation of the spline cost in the phase space, that is in ${\mathbb R}^d$, with $t_{i+1}-t_i = \delta_i$:
\begin{equation}\label{equivalencedescout}
c\left(Y_1,\ldots,Y_n\right) = \min_{\left(V_1,..V_n\right)\in ({\mathbb R}^d)^n} \sum_{i=1}^{n-1} \frac{1}{\delta_i^3}
c_{ph}\left[\left(Y_i,\delta_i V_i\right), \left(Y_{i+1},\delta_i V_{i+1}\right)\right]
\end{equation}
where
\begin{equation}\label{EqCostTM2}
c_{ph}[(x,v),(y,w)] = 12|x-y|^2 + 4(|v|^2 + |w|^2 + \langle v,w \rangle + 3 \langle v+w, x-y \rangle).
\end{equation}
The advantage of the formulation \eqref{equivalencedescout} is that the cost is separable in the phase space and the gradient with respect to speeds and positions is easy to compute.
We thus implement a gradient descent in the phase space using the lbfgs function in python. We compute the gradient by automatic differentiation.
The Wasserstein terms in the minimization problem \eqref{SDV} depends only on the positions and are computed thanks to M\'erigot Library \cite{merigotgit} in dimension 2.
To do simulations in dimension 3 one has to use L\'evy Library \cite{levywebpage}. The density constraints $\rho_i$ are given trough linear functions on a triangulation.
\begin{remark}
Other problems can be addressed using similar optimization problem as in Definition \ref{SDVdef}.
For instance the quadratic cost in \eqref{SDV} leads to Wasserstein interpolation.
We can also interpolate with curves as smooth as we want, using for instance the $L^2$ norm of the derivative of order $m$ of the curve or even other classical interpolating curves.
\end{remark}
\subsubsection{Numerical simulations}
We propose three numerical simulations, one to compare the qualitative results with respect to the multi marginal approach and especially Figure \ref{f2d1}.
A second one in order to illustrate the non-convexity issue and a third one for applications in images.
\paragraph{\textbf{The rotation case: Figure \ref{rotation2}.}}
In this case we compute Wasserstein splines passing through four gaussians with variance 15 and center of masses respectively $(0,2),(10,0),(10,6),(0,4)$ with constraint parameter $\epsilon= 10^{-3}$. The number of points is $2000$.
In this case the result is a global minimizer and is not sensible to initialization.
The lack of convexity is not an issue.
Compare to Figure \ref{f2d1}, this approach gives a better a approximation of the intermediate densities especially with less diffusion.
\begin{figure} \centering
\begin{tabular}[h]{cc}
\hspace{-2cm} \includegraphics[width=10cm]{papier_feracheval_initiallocalshuffle_2000_15_1000_40_contraintes}
\hspace{-2cm} \includegraphics[width=10cm]{papier_feracheval_initiallocalshuffle_2000_15_1000_40} \\
\hspace{-2cm} \includegraphics[width=15cm]{papier_feracheval_JD}
\end{tabular}
\caption{Spline interpolation for gaussians with 2000 Dirac masses for each measure, $\epsilon=1O^{-3}$.
Left: sample of each density constraints $\rho_i$, i=1,2,3,4.
Right: Some trajectory of diracs masses randomly chosen, marginals at the constrained time $0,1,2,3$ and marginals at time $0.5,1.2,1.5,1.7,2.5$.
Second Line : the same configuration as in figure \ref{f2d1}.
}
\label{rotation2}
\end{figure}
\paragraph{\textbf{The crossing case: Figure \ref{crossing0}, \ref{crossing} }}
Here we compute Wasserstein splines starting from a mixture of two gaussians with centrer $(0,-1),(0,1)$ and variance $15$ then passing through a gaussian with center $(0,0)$ and variance $15$ and finishing at a translation of the initial mixture. The number of points is $2000$, $\epsilon$ will value $1$ or $1000$.
We expect the global minimizer to be straight lines crossing around the middle constraint and with a low cost.
Numerically depending on the initial conditions, we can recover different local minimizers,
the local minimum which is reached is extremely correlated with the initial coupling.
In Figure \ref{crossing0} we observe that changing $\epsilon$ but keeping a similar initial coupling, all points are given by a quantization of the middle density with a random enumeration and $0$ initial speed, yields to a similar local minimum.
\begin{figure} \centering
\begin{tabular}[h]{cc}
\hspace{-2cm} \includegraphics[width=9cm]{papier_croisement_esp1tr1} &
\hspace{-0.5cm}\includegraphics[width=9cm]{papier_croisement_esp1000tr1}
\end{tabular}
\caption{Spline interpolation for a mixture of gaussians with 2000 Dirac masses. Same initial coupling for both figure.
Left: $\epsilon=1$.
Right: $\epsilon=1000$.
}
\label{crossing0}
\end{figure}
Finding a good initial coupling is the hard part in order reach the global maximum.
One solution is to initialize with points close to each other and a very large $\epsilon$.
Then one as to add some noise in the gradient and decreases slowly $\epsilon$.
Unfortunately we didn't find a systematic approach for this random multi-scale method and one as to fit the parameters case by case.
In Figure \ref{crossing} the global minimizer is achieved by first computing the spline with a relaxed constraint, i.e. large $\epsilon$, only for the final time ( in pratice $\epsilon =[1000,1000,1]$.
Then we use this result, which has the good initial coupling, as and initial condition and set $\epsilon=1000$ for all the constraints.
We also compare this results with the interpolation with a different initial condition and the Wasserstein geodesics.
In all these simulations we clearly observe that particles can cross along the dynamic appart from the optimal transport inthis situation.
\begin{figure} \centering
\begin{tabular}{cc}
\hspace{-2cm} \includegraphics[width=10cm]{papier_croisement_esp1000_initialdroit}
\hspace{-0.5cm} \includegraphics[width=10cm]{papier_croisement_esp1000_initiallocal} \\
\includegraphics[width=10cm]{papier_croisement_esp1000_initiatplat_transport}
\end{tabular}
\caption{Spline interpolation for a mixture of gaussians with 2000 Dirac masses for each measure. $\epsilon=1000$.
Top Left: Initialization with a good coupling, total cost $=302$.
Top Right: Initialization with a quantization of the middle density and no speed, total cost $=804$ (local minima).
Bottom: Interpolation with the Wasserstein geodesic. $\epsilon=1000$, cost $=930$.
}
\label{crossing}
\end{figure}
Note that this spline approach is related to the problem of finding minimal geodesics along volume preserving maps done by M\'erigot and Mirebeau \cite{merigot2015minimal} : in their work the constraints $\rho_i$ are the Lebesgue measure, the cost is changed by the quadratic cost between two points and they have a coupling constraint.
Therefore their minimization problem is also non convex but the coupling is given as a constraint so the non convexity issue didn't rise as clearly as in this spline problem.
\paragraph{\textbf{Image interpolation: }}
pour l'instant c'est pas presentable, ca passe vraiment au milieu.
Je vais relancer dans la semaine mais je propose de faire une version sans.
\begin{remark}[Extrapolation]
The minimization of the acceleration can be used to provide time extrapolation of Wasserstein geodesic in a natural way: particles follow straight lines. This can be implemented in a $3$-marginal problem with the acceleration cost $c(x_1,x_2,x_3) = \frac{1}{\lambda^2}|x_3 - 2x_2 + x_1|^2 + \frac{1}{\lambda}|x_2 - x_1|^2$ under marginal constraints at time $1$ and $2$.
Note that, in the spline model, the formulation we proposed does not prevent particles from crossing each other. They are completely independent. Therefore, the particles following simply geodesic lines and after a shock, the evolution is not geodesic in the Wasserstein sense (since shocks do not occur but at initial and final times). The implementation of time extrapolation using entropic regularization is straightforward. Figures \ref{Extrapolations1} and \ref{Extrapolations2} show some experiments on $[0,1]$ discretized with $100$ points and $\varepsilon = 0.015$. The translation experiment recovers what is expected however the effect of the diffusion can be seen with a twice larger $\varepsilon$.
\begin{figure}
\centering
\begin{tabular}[h]{cc}
\hspace{0cm}\includegraphics[width=8cm]{TranslationExtrapolation}
&
\hspace{0cm}\includegraphics[width=8cm]{TranslationDiffusion}
\end{tabular}
\caption{Extrapolation of a translation with two different $\varepsilon = 0.015$ and $\varepsilon = 0.03$}
\label{Extrapolations1}
\end{figure}
We also show two other simulations, one is a splitting simulation and the last one is a merging of two "bumps" into a single one. The extrapolation shows an other bimodal distribution which is explained by particle crossings.
\begin{figure}
\centering
\begin{tabular}[h]{cc}
\hspace{0cm}\includegraphics[width=8cm]{SplittingExtrapolation}
&
\hspace{0cm}\includegraphics[width=8cm]{MergingExtrapolation}
\end{tabular}
\caption{On the left, a splitting experiment and on the right, a merging experiment.}
\label{Extrapolations2}
\end{figure}
Note that this extrapolation scheme may proven useful in the development of higher-order schemes for the JKO algorithm.
\end{remark}
\section{Perspectives}
In this paper, we presented natural approaches to define cubic splines on the space of probability measures. We have presented a Monge formulation and its Kantorovich relaxation on the path space as well as their corresponding reduction on minimal cubic spline interpolation. We leave for future work theoretical questions such as the study of conditions under which the existence of a Monge map as a minimizer occurs, as well as the relaxation of cubic spline in the Wasserstein metric.
Our main contributions focus on the numerical feasibility of the minimization of the acceleration on the path space with marginal constraints. We have developed the entropic regularization scheme for the acceleration and shown simulations in 1D and 2D. Future work will address the 3D case which is out of reach with the methods presented in the first sections of this paper but possibly tackled with the semi-discrete method presented en Section \ref{SD}. In a similar direction, the application of this approach to the unbalanced case in the spirit of \cite{GeneralizedOT1} seems challenging due to the this dimensionality constraint and could be achieved within the semi-discrete setting.
In the Lagrangian setting, i.e. semi-discrete method, the extrapolation of a Wasserstein geodesic between $\rho_0$ and $\rho_1$ is obtained using three positions with the following formulation : let
$$\mathcal{Q}^N = \left\{ \sum^N_{j=1} \frac1N \delta_{\left(X^1_j,X^2_j,X^3_j\right) } \middle| (X_j)_{j=1,\ldots,N}\in M^{n} \right\},$$ then
\begin{equation}\label{SDVextr}
(SDextra) = \min_{\mathcal{Q}^N} \,\frac{1}{N}\sum_{j=1}^N \frac{d^2}{2}(X^1_j,X^2_j) + \frac{1}{N}\sum_{j=1}^N c(X^1_j,X^2_j,X^3_j) + \sum^2_{i=1} \frac{1}{2\epsilon^2} W_2^2\left(\sum^N_{j=1} \frac1N \delta_{X^i_j} , \rho_i \right),
\end{equation}
where $d$ is the distance on $M$ and $c(X^1_j,X^2_j,X^3_j)$ the cost of the cubic spline.
In particular this formulation forces the curve to be a Wasserstein geodesic between $\rho_1$ and $\rho_2$, using the quadratic cost, and let free the final marginal. The implementation is completely similar as in Section \ref{SD} and the trajectory of each dirac masses is a straight line.
|
\section{Introduction}
A new class of $\gamma$-ray loud active galactic nuclei (AGNs) has been emerging
since the first detection of $\gamma$-ray emission from the radio-loud (RL)
narrow-line Seyfert 1 galaxies (NLS1s) by the Large Area Telescope (LAT) onboard the {\it Fermi} $\gamma$-ray Space Telescope
(hereinafter, {\it Fermi}) satellite \citep{2009ApJ...699..976A}.
Distinctly different from the well-known paradigm that powerful relativistic jets are generally
associated with elliptical galaxies in typical RL AGNs like blazars and radio galaxies,
$\gamma$-ray emitting NLS1s have drawn great attention from the AGN community
in the past decade.
NLS1s are conventionally defined as type 1 AGNs with
the full width at half-maximum (FWHM) of the broad H\ensuremath{\beta}\ line less than 2000\,km\,s$^{-1}$,
weak [O\,{\footnotesize III}]\ emission ([O\,{\footnotesize III}]/H\ensuremath{\beta}\,< 3), and usually strong {\rm Fe\,{\footnotesize II}}\ emission lines
\citep[][]{1985ApJ...297..166O}.
Several correlations have been found among optical emission lines and X-ray properties
referred to as the eigenvector 1 (EV1) correlations
\citep{1992ApJS...80..109B,2002ApJ...565...78B}.
Given their relatively small widths of the broad lines, NLS1s tend to have lower
black hole (BH) masses and higher Eddington ratios near or above their Eddington limits
\citep{2004AJ....127.3168B,2004ApJ...606L..41G,2012AJ....143...83X,2015AJ....150...23Y,2017MNRAS.468.3663J,2017MNRAS.471..706J}.
However, we note in passing that it has been occasionally suggested that
BH masses in NLS1s could be higher \citep[e.g.][]{2013MNRAS.431..210C,2016MNRAS.458L..69B,2017MNRAS.469L..11D}
while other studies continue to favour low BH masses \citep[e.g.][]{2009ApJ...707L.142A,2015AJ....150...23Y,2016MNRAS.463.4469D,2017MNRAS.464.2565L}.
Furthermore, NLS1s are found to preferably be hosted in disc-like galaxies with pseudo-bulges
\citep{2006AJ....132..321D,2012ApJ...754..146M}.
Evidence that indicates the presence of relativistic jets in some
RL NLS1s has been accumulated in recent years \citep{2006AJ....132..531K}, especially at the highest radio loudness regime,
due to their blazar-like characteristics \citep{2007ApJ...658L..13Z,2008ApJ...685..801Y}.
RL NLS1 usually shows a compact radio morphology with a one-sided core-jet structure
detected at parsec scale \citep[][]{2015ApJS..221....3G}.
In some objects, the radio emission extends up to kpc scales \citep[e.g.][]{2008A&A...490..583A,2012ApJ...760...41D,2015ApJ...800L...8R}.
A significant fraction of RL NLS1s presents a flat/inverted radio spectrum,
usually with a very high brightness temperature \citep{2008ApJ...685..801Y,2015ApJS..221....3G}.
Some RL NLS1s are found to have intraday infrared variability
\citep[e.g.][]{2012ApJ...759L..31J,2015MNRAS.454L..16Y}.
Besides, superluminal motions have also been found in some RL NLS1s
(e.g. SBS~0846$+$513: \citealt{2013MNRAS.436..191D};
1H~0323$+$342: \citealt{2016RAA....16..176F}).
Since the launch of the {\it Fermi} satellite in 2008, $\gamma$-ray detections from some
genuine RL NLS1s have confirmed the existence of relativistic jets in this new class of
$\gamma$-ray loud AGNs \citep{2009ApJ...699..976A,2009ApJ...707L.142A,2011nlsg.confE....F,2012MNRAS.426..317D,2015MNRAS.452..520D,2015MNRAS.454L..16Y,2018ApJ...853L...2P}.
With their extreme distributions in AGN parameter space and different host environments,
RL NLS1s allow us to readdress some of the key questions regarding the formation
and evolution of relativistic jets under extreme physical conditions
as well as the coupling of jets and accretion flows.
However, further investigation is hindered by the scarcity of this particular class of objects.
In this paper, we report on the discovery of a new $\gamma$-ray source detected by
the {\it Fermi}-LAT associated with a RL NLS1, SDSS J211852.96$-$073227.5 (hereinafter J2118$-$0732).
This $\gamma$-ray-emitting NLS1 was found in our ongoing study on a sample of RL NLS1s
\citep{2007ApJ...658L..13Z,2008ApJ...685..801Y,2015A&A...574A.121K,2015MNRAS.454L..16Y}.
While we were analysing this NLS1, it was independently reported as
a new $\gamma$-ray source by \citet{2018ApJ...853L...2P} and
in The Preliminary {\it Fermi}-LAT 8-yr Point Source List (FL8Y)\footnote{\url{https://fermi.gsfc.nasa.gov/ssc/data/access/lat/fl8y/}},
and classified as an NLS1 by \citet{2017ApJS..229...39R}.
It was also included in the Combined Radio All-Sky Targeted Eight GHz Survey
\citep[CRATES;][]{2007ApJS..171...61H} catalogue
as a flat-spectrum radio source due to its flat spectrum below 4.8 GHz.
The object was detected in the ROSAT All Sky Survey (RASS), with the total counts
$<$ 20 \citep{2007AJ....133..313A}.
The X-ray spectral and timing properties are essential for linking the postulated high energy
$\gamma$-ray emission with other wavelength data of this AGN,
and for understanding its radiation mechanism.
To this end, we proposed to observe J2118$-$0732 with the {\it XMM-Newton}\ at two epochs
separated by about five months to study its X-ray spectral and timing properties.
We report the results of these observations in this paper.
The paper is organized as follows.
In Section~\ref{sec:spectroscopy}, we report the optical spectroscopic analysis and results.
The {\it Fermi}-LAT data selection and analysis are included in Section~\ref{sec:fermi}.
We report the {\it XMM-Newton}\ observations and results in Section~\ref{sec:xmm}.
The radio properties, infrared variability, and optical/ultraviolet (UV) characteristics are discussed
in Section~\ref{sec:radio}, Section~\ref{sec:wise}, and Section~\ref{sec:opt/uv}, respectively.
The broad-band spectral energy distribution (SED) modelling is given in Section~\ref{sec:sed},
followed by discussions in Section~\ref{sec:discussion} and a summary in Section~\ref{sec:summary}.
Throughout this paper, we assume a cosmology with
$H_0$ = 67.8 km\,s$^{-1}$\,Mpc$^{-1}$, $\Omega_\Lambda$ = 0.692,
and $\Omega_m$ = 0.308 \citep[][]{2016A&A...594A..13P}.
\section{Optical Spectroscopy and the NLS1 Classification}
\label{sec:spectroscopy}
\begin{figure}
\includegraphics[width=\columnwidth]{./figure/optical_spectrum_lorentz.eps}
\caption{The rest-frame SDSS spectrum of J2118$-$0732 after correction
for Galactic extinction. The solid red line shows the best-fitting model of total fluxes.
The dashed blue line represents a broken power-law continuum and the solid green line below the continuum accounts for the broad components of the H\ensuremath{\alpha}\ and H\ensuremath{\beta}\ lines.
The zoom-in H\ensuremath{\beta}\ line is displayed in the inset panel.}
\label{fig:sdss_spectrum}
\end{figure}
The optical spectrum was obtained by the Sloan Digital Sky Survey (SDSS) on 2001 August 25, with an exposure time of 911\,s.
After being corrected for Galactic extinction with $E(B-V)$=0.18\,mag
\citep[][]{2011ApJ...737..103S} and an $R_V$= 3.1 extinction law,
the spectrum is transformed into the rest frame with a redshift of $z$=0.26
(see Fig.~\ref{fig:sdss_spectrum}).
In the spectrum, the host galaxy contribution is negligible.
Thus, a similar strategy as in \citet{2008MNRAS.383..581D} and \citet{2015MNRAS.454L..16Y}
was adopted to fit the spectrum.
We fitted simultaneously the continuum, the {\rm Fe\,{\footnotesize II}}\ multiplets, and other emission lines
in the range of 4200--6900\,\AA.
A broken power law (PL) with a break wavelength of 5600\,\AA\ and
the optical {\rm Fe\,{\footnotesize II}}\ emission multiplets modelled with the templates from
\citet{2008MNRAS.383..581D} were used to fit the so-called pseudo-continuum.
The emission lines identified from the composite SDSS quasar spectrum
\citep[see table 2 in][]{2001AJ....122..549V} from H\ensuremath{\gamma}\ to
[S\,{\footnotesize II}]\,$\lambda\lambda6716,6731$\footnote{\label{foot:lines}
Several emission lines were masked out, for either they were too weak to constrain in the fit
or had little effect on the results. H\ensuremath{\gamma}\,$\lambda$4340, [O\,{\footnotesize III}]\,$\lambda$4363,
He\,{\footnotesize II}\,$\lambda4686$, H\ensuremath{\beta}\,$\lambda4861$, [O\,{\footnotesize III}]\,$\lambda\lambda4959,5007$,
He\,{\footnotesize I}\,$\lambda4861$, [N\,{\footnotesize II}]\,$\lambda \lambda6548,6583$, H\ensuremath{\alpha}\,$ \lambda6563$,
and [S\,{\footnotesize II}]\,$\lambda\lambda6716,6731$ were included in our spectroscopic analysis.}
were modelled as follows.
The broad H\ensuremath{\alpha}\ and H\ensuremath{\beta}\ were assumed to have the same redshift and profile fitted by
either a single Lorentzian or concentric double Gaussians.
Other broad emission lines were modelled with one Gaussian.
One single Gaussian with a constraint of FWHM < 1000 km\,s$^{-1}$ was used
for modelling each narrow emission line.
Both flux ratios of the [N\,{\footnotesize II}]\ doublets and the [O\,{\footnotesize III}]\ doublets
were fixed to their theoretical values of 3.
\begin{table}
\caption{The spectroscopic fitting results and basic parameters of J2118$-$0732.}
\label{tab:spectroscopic}
\begin{threeparttable}
\begin{tabular}{lrr}
\hline
Model$^a$ & L & 2G \\
\hline
FWHM(H\ensuremath{\beta})/km\,s$^{-1}$ & $1585 \pm 182$ & $1882 \pm 375$ \\
FWHM([O\,{\footnotesize III}])/km\,s$^{-1}$ & $364 \pm 15$ & $327 \pm 14$ \\
[O\,{\footnotesize III}]/H\ensuremath{\beta}\ & 1.7 & 2.0 \\
$R_{4570}$ & 0.06 & 0.06 \\
$M_{\rm BH}$/10$^7$\,\ensuremath{M_{\sun}}\ & 3.4 & 3.7 \\
Eddington ratio & 0.15 & 0.11\\
\hline
\end{tabular}
\begin{tablenotes}
\item[]
$^a$ Models used for fitting the broad H\ensuremath{\beta}\ emission lines are:
one Lorentzian profile (L) or a double-Gaussian profile (2G).\\
\end{tablenotes}
\end{threeparttable}
\end{table}
The broad H\ensuremath{\beta}\ line was equally well fitted with either a Lorentzian profile or
a concentric double-Gaussian profile.
In the former case, the width of the broad component is
FWHM(H$\beta_{\rm broad}$) = 1585 $\pm$ 182\,km\,s$^{-1}$, while in the latter case it is
FWHM(H$\beta_{\rm broad}$) = 1882 $\pm$ 375\,km\,s$^{-1}$
after subtracting the effect of instrumental broadening.
\citet{2011ApJS..194...45S} also obtained a similar
FWHM(H$\beta_{\rm broad}$) = 1597 $\pm$ 441\,km\,s$^{-1}$
using a single Gaussian or multiple Gaussians model.
In the following, we adopted the fitting results using the Lorentzian profile
since the Lorentzian model is better suited for the broad-line profiles of NLS1s
\citep[e.g.][]{2001A&A...372..730V,2006ApJS..166..128Z}.
The line width of [O\,{\footnotesize III}]\,$\lambda5007$ is 364 $\pm$ 15\,km\,s$^{-1}$
and there is no obvious evidence of blue wings in [O\,{\footnotesize III}]\ lines.
The flux ratio of [O\,{\footnotesize III}]\,$\lambda5007$ to H$\beta_{\rm total}$ is $\approx1.7$
and $R_{4570} \equiv$ {\rm Fe\,{\footnotesize II}}\,$\lambda4570$/H$\beta_{\rm total}\approx0.06$
({\rm Fe\,{\footnotesize II}}\ is calculated by integrating from 4434 to 4684\,\AA).
The small broad-line width and the relatively weak [O\,{\footnotesize III}]\ of J2118$-$0732
fulfil the conventional definition of an NLS1.
However, the {\rm Fe\,{\footnotesize II}}\ emission is rather weak compared to most NLS1s
and the optical continuum is red with
$\alpha_{\rm opt}$ = -1.5 ($S_{\nu}\propto \nu^{\alpha}$) for J2118$-$0732.
Nevertheless, there is no quantitative definition for the {\rm Fe\,{\footnotesize II}}\ strength,
and some NLS1s do emit weak {\rm Fe\,{\footnotesize II}}\ emission
\citep{2006ApJS..166..128Z,2012AJ....143...83X,2016MNRAS.462.1256C,2017ApJS..229...39R}.
Under the assumption of a virialized broad-line region (BLR), we can estimate
the BH mass of J2118$-$0732 using FWHM(H$\beta_{\rm broad}$) = 1585\,km\,s$^{-1}$
and the H\ensuremath{\beta}-based estimator from \citet{2009ApJ...707.1334W},
with a 1$\sigma$ uncertainty of 0.5 dex.
Considering the contamination from the jet contribution to the continuum, a surrogate for the
5100\,\AA\ luminosity ($\lambda L_{5100}$) estimated from the broad H\ensuremath{\beta}\ line luminosity
using equation (5) from \citet{2006ApJS..166..128Z} was applied.
We found $M_{\rm BH} \approx 3.4 \times 10^7\,M_{\sun}$
($M_{\rm BH} \approx 3.7 \times 10^7\,M_{\sun}$ if the line parameters fitted from
the double-Gaussian profile were used), which is consistent with the estimations from
\citet{2007ApJ...667..131G} and \citet{2008ApJ...688..826L} based on
other commonly used single-epoch formalisms.
The bolometric luminosity can be calculated assuming $L_{\rm bol} = 9\lambda L_{5100}$
(estimated from H\ensuremath{\beta}), as suggested by \citet{2000ApJ...533..631K},
which results in $L_{\rm bol} \approx 6.2 \times 10^{44}$\,erg\,s$^{-1}$.
The Eddington ratio of this object is $\lambda \approx 0.15$,
relatively low compared to other NLS1s but higher than typical broad-line Seyfert 1 galaxies
\citep{2012AJ....143...83X,2016MNRAS.462.1256C,2017ApJS..229...39R}.
The bolometric luminosity and the Eddington ratio of J2118$-$0732 estimated by \citet[][]{2011ApJS..194...45S} are
$L_{\rm bol} \approx 1.1 \times 10^{45}$\,erg\,s$^{-1}$ and $\lambda \approx 0.39$,
slightly larger than our estimations.
This discrepancy has mainly resulted from the overestimation of the bolometric luminosity
using the directly measured luminosity at 5100~\AA,
where the jet contribution cannot be ignored, especially for those RL AGNs.
The fitting results and the basic parameters estimated for J2118$-$0732 are summarized
in Table~\ref{tab:spectroscopic}.
\section{$\gamma$-ray data analysis}
\label{sec:fermi}
\subsection{Observations and data reduction}
The {\it Fermi}-LAT is a pair-conversion $\gamma$-ray telescope operating
from 20\,MeV to > 300\,GeV.
It has a large peak effective area ($\sim$8000\,cm$^2$ for 1\,GeV photons),
allowing a large field of view ($\sim$2.4\,sr) with an angular resolution
(68 per cent containment radius) better than 1\degr\ for energies above 1\,GeV
and an energy resolution of typically $\sim$10 per cent.
Further details about the {\it Fermi}-LAT are given in \citet{2009ApJ...697.1071A}.
We analysed the Pass 8 data collected by the LAT from 2008 August 4th (Modified Julian Day, MJD 54 682) to 2017 July 14 (MJD 57 948).
During this period, the LAT instrument operated in survey mode,
scanning the entire sky every 3\,h.
The data analysis was performed with the standard \texttt{ScienceTools}
version v10r0p5 software package.
We used the standard binned maximum likelihood method implemented in
the Science tool \texttt{gtlike} to analyse
each time bin longer than or equal to 1\,yr.
For smaller time bins, we used the unbinned method where the number of events
was expected to be small.
The LAT data were extracted from a region of interest (ROI) defined as a circle
of 30\degr\ radius centred at the location of J2118$-$0732 for the binned method
and a circle of 20\degr\ radius for the unbinned method.
Only events belonging to the 'Source' class from 100\,MeV to 300\,GeV
were used with \texttt{P8R2\_SOURCE\_V6} Instrument Response Functions.
In addition, a cut on the zenith angle (< 90\degr) was applied to
minimize the contamination from Earth limb $\gamma$-rays.
Isotropic (iso\_P8R2\_SOURCE\_V6\_v06.txt) and Galactic diffuse emission (gll\_iem\_v06.fits) components were used to model the background.
The normalization of each component was set to be free during the spectral fitting.
The significance of the $\gamma$-ray signal from the source was evaluated by means of a
maximum-likelihood test statistic (TS) that results in TS = 2 (log$L_1$ - log$L_0$),
where $L$ is the likelihood of the data given the model with ($L_1$) or without ($L_0$)
a point source at the position of J2118$-$0732 \citep[e.g.][]{1996ApJ...461..396M}.
Because our data cover a 9-yr period, the source model included
all the point sources from The Third {\it Fermi}-LAT source catalogue \citep[3FGL;][]{2015ApJS..218...23A}
that fall within the ROI and an addition of 10\degr\ annular radius around it,
and new sources detected in FL8Y within 10\degr\ of J2118$-$0732.\footnote{Before
the recent publication of FL8Y, we have discovered the $\gamma$-ray detection from the optical
position of J2118$-$0732, just using 3FGL sources for source model in the $\gamma$-ray data analysis.}
The spectra of these sources were parameterized by a PL, a log-parabola,
or a super exponential cut-off, as in the 3FGL catalogue,
and by a PL for 18 new sources in FL8Y.
Sources within 10\degr\ of J2118$-$0732 and with their significances greater than 5
calculated from 3FGL and FL8Y were included with the normalization factors
and the photon indices left as free parameters.
For other sources in the source model, we fixed the normalizations and the photon indices
to the values in the catalogues.
The uncertainties correspond to 68 per cent confidence for the {\it Fermi}-LAT results.
\subsection{Results}
\label{sec:fermi-results}
\begin{figure}
\includegraphics[width=\columnwidth]{./figure/gamma_lc.eps}
\caption{Integrated flux light curve of J2118$-$0732 obtained by the {\it Fermi}-LAT
in the 0.1--300\,GeV energy range during 2008 August 4--2017 July 14
(MJD 54 682--57 948) with 1-yr bins in black, 60-d bins in blue and 20-d bins in red.
Diamonds refer to the detections with higher TS values.
Arrows refer to 95 per cent upper limits on the source flux. Upper limits are computed when
TS < 9 for 1-yr bins and TS < 5 for 60-d bins and 20-d bins. The two dashed
vertical lines correspond to the epochs of our two {\it XMM-Newton}\ Observations,
X-2017-05-10 and X-2017-10-27.}
\label{fig:gamma_lc}
\end{figure}
\begin{table*}
\centering
\caption{Results of the {\it Fermi}-LAT data analysis.}
\label{tab:lat}
\begin{threeparttable}
\begin{tabular}{lcccccccc}
\hline
& \multicolumn{2}{c}{Optimized position (J2000)} & Separation from& $R_{95 \%}$ &
$F_{0.1-300 {\rm GeV}}$ & $\Gamma_\gamma$ & $L_\gamma$ & TS \\
& hh mm ss.s & dd mm ss.s & optical position (\degr) & (\degr)&
(10$^{-9}$\,ph\,cm$^{-2}$\,s$^{-1}$) & & (10$^{45}$\,erg\,s$^{-1}$) & \\
\hline
(1) & & & & &
6.6 $\pm$ 1.8 & 2.65 $\pm$ 0.16 & 0.7 $\pm$ 0.3 & 27 \\
(2) & 21 18 45.4 & -07 25 02.1 & 0.128 & 0.161 &
15.4 $\pm$ 3.1 & 2.79 $\pm$ 0.16 & 1.5 $\pm$ 0.4 & 50 \\
(3) & 21 18 37.1 & -07 28 42.7 & 0.091 & 0.264 &
43.7 $\pm$ 1.6 & 2.62 $\pm$ 0.01 & 4.6 $\pm$ 0.2 & 24 \\
(4) & 21 18 58.9 & -07 30 04.5 & 0.047 & 0.14 &
9.2 $\pm$ 2.0 & 2.80 $\pm$ 0.15 & 0.9 $\pm$ 0.3 & 40 \\
(5) & 21 19 14.1 & -07 28 31.1 & 0.109 & 0.20$\times$0.16 &
13.2 $\pm$ 2.5 & 2.81 $\pm$ 0.12 & 1.3 $\pm$ 0.3 & 47 \\
\hline
\end{tabular}
\begin{tablenotes}
\item[]
(1) our results analysed from the whole period during 2008 August 4--2017 July 14. \\
(2) our results analysed from the active 4 yr during 2009 August 5--2013 July 20. \\
(3) our results analysed from the flare during 2010 December 2--2011 February 1. \\
(4) \citet[][]{2018ApJ...853L...2P} results analysed from the period during 2008 August 5--2017 January 5. \\
(5) FL8Y results analysed from the period during 2008 August 4--2016 August 2. \\
\end{tablenotes}
\end{threeparttable}
\end{table*}
First, we tested whether a $\gamma$-ray source at the optical position of J2118$-$0732
was detected by the {\it Fermi}-LAT over the whole duration from 2008 August 4 to
2017 July 14 with a binned maximum likelihood analysis using a PL model.
A source was clearly detected with a TS value of 27 ($\sim$5\,$\sigma$) for J2118$-$0732,
an average $\gamma$-ray integrated photon flux from 0.1 to 300\,GeV of
$F_{0.1-300 {\rm GeV}}$ = ($6.6 \pm 1.8$)\,$\times\,10^{-9}$\,ph\,cm$^{-2}$\,s$^{-1}$,
and a photon index $\Gamma_{\gamma}=2.65\pm0.16$.
This flux corresponds to an isotropic $\gamma$-ray luminosity of
$L_\gamma$ = (0.7 $\pm$ 0.3)\,$\times\,10^{45}$\,erg\,s$^{-1}$
in the 0.1--300\,GeV rest-frame energy band.
The isotropic $\gamma$-ray luminosity is given as $L_\gamma$ = $4\pi$\,$d_L^2$\,$S_{\gamma} $,
where $d_L$ is the luminosity distance and $S_{\gamma}$ is the flux density.
Next, the integrated flux of each year during the whole 9-yr {\it Fermi}-LAT observation
was calculated to investigate the long-term $\gamma$-ray variability of J2118$-$0732.
For each 1-yr bin, the spectral parameters of J2118$-$0732 and all sources within
10\degr\ were frozen to the values obtained from the likelihood analysis over the entire period.
J2118$-$0732 was relatively active during 2009 August 5--2013 July 20 with higher flux
measurements and significant TS values, but remained undetected in other years.
The 95 per cent upper limits were provided if TS < 9 during 2008 August 4--2009 August 5 and 2013 July 20--2017 July 14 with 1-yr bins in the long-term $\gamma$-ray light
curve of J2118$-$0732 in Fig.~\ref{fig:gamma_lc}.
A further likelihood analysis of the data taken during the relatively active 4 yr
resulted in a TS of 50 ($\sim$7\,$\sigma$).
We obtained a photon flux index of $\Gamma_{\gamma}=2.79\pm 0.16$ and a flux of
$F_{0.1-300 {\rm GeV}}$ = (15.4 $\pm$ 3.1)\,$\times\,10^{-9}$\,ph\,cm$^{-2}$\,s$^{-1}$,
corresponding to an isotropic $\gamma$-ray luminosity of
$L_\gamma$ = (1.5 $\pm$ 0.4)\,$\times\,10^{45}$\,erg\,s$^{-1}$ during the 4-yr period.
We performed a $\gamma$-ray point source localization using the \texttt{gtfindsrc} tool
over the photons in the energy band from 1 to 300\,GeV extracted during this period.
The fit resulted in R.A. = 21$^{\rm h}$18$^{\rm m}$45.4$^\mathrm{s}$,
Dec. = $-$7$^{\circ}$ 25$\arcmin$02.1$\arcsec$ with a 95 per cent error circle radius of
$R_{95 \%}$ = 0.161\degr.
Here, we excluded the low-energy photons because of their poor angular resolutions.
The $\gamma$-ray source is located 0.128\degr\ away from the optical position of J2118$-$0732,
which is within the 95 per cent error circle of the $\gamma$-ray position.
We also calculated the $\gamma$-ray light curve of J2118$-$0732 during 2009--2013 using
60-d time bins and 20-d time bins and added it into Fig.~\ref{fig:gamma_lc}.
For each time bin, the spectral parameters of J2118$-$0732 and all sources within 10\degr\
were frozen to the values obtained from the 4-yr likelihood analysis.
The 95 per cent upper limit was also calculated if TS $<$ 5.
The systematic uncertainty in the flux is dominated by the systematic uncertainty in the
effective area \citep[][]{2012ApJS..203....4A}, which amounts to 10 per cent at 100\,MeV,
decreasing to 5 per cent at 560\,MeV and increasing to 10 per cent above 10\,GeV.
J2118$-$0732 was detected sporadically by the {\it Fermi}-LAT, with an increase in flux and a high
TS value of 24 ($\sim$5\,$\sigma$) during 2010 December 2--2011 February 1.
The flux was doubled within 20\,d and quickly dropped to a lower state below the
{\it Fermi}-LAT detection limit.
During this flaring period, we obtained a photon index of $\Gamma_{\gamma}$ = 2.62 $\pm$ 0.01 and a flux of
$F_{0.1-300 {\rm GeV}}= (43.7 \pm 1.6)\,\times$\,10$^{-9}$\,ph\,cm$^{-2}$\,s$^{-1}$,
corresponding to an isotropic $\gamma$-ray luminosity of
$L_\gamma$ = (4.6 $\pm$ 0.2)\,$\times$\,10$^{45}$\,erg\,s$^{-1}$,
a factor of $\sim$7 higher than the average value during the whole period.
A point source localization using full energy band photons was also performed during this 60-d high state.
The fit resulted in R.A. = 21$^{\rm h}$18$^{\rm m}$37.1$^{\rm s}$,
Dec. = $-$7$^{\circ}$28$\arcmin$42.7$\arcsec$ which is 0.091\degr\ away from the
optical position of J2118$-$0732 with a 95 per cent error circle radius of $R_{95 \%}$ = 0.264\degr.
The $\gamma$-ray detection of this source has also been independently
reported by \citet[][]{2018ApJ...853L...2P} and FL8Y.
A brief comparison between their results and ours regarding the $\gamma$-ray properties
is summarized in Table~\ref{tab:lat}.
The $\gamma$-ray detected source suffers from a large spatial uncertainty given
the limited resolution of the {\it Fermi}-LAT.
So the association of the $\gamma$-ray source with J2118$-$0732
will be discussed further in Section~\ref{sec:association}.
\section{X-ray observations and data analysis}
\label{sec:xmm}
\subsection{XMM observations and data reduction}
\begin{table}
\caption{XMM EPIC observations for J2118$-$0732.}
\label{tab:x_obs}
\begin{threeparttable}
\begin{tabular}{lcccc}
\hline
Data Set & \multicolumn{2}{c}{X-2016-05-10} & \multicolumn{2}{c}{X-2016-10-27} \\
\hline
Observation ID & \multicolumn{2}{c}{0784090201} & \multicolumn{2}{c}{0784090301} \\
Start Time & \multicolumn{2}{c}{2016 May 10} & \multicolumn{2}{c}{2016 October 27} \\
(UTC) & \multicolumn{2}{c}{12:36:12} & \multicolumn{2}{c}{10:55:03} \\
Duration (ks) & \multicolumn{2}{c}{31.9} & \multicolumn{2}{c}{32.8} \\
Detectors & PN & MOS1/2 & PN & MOS1/2 \\
Good Exposure & 18.6 & 29.3 & 16.2 & 29.4 \\
(ks) & & & & \\
Net Count Rate & 0.193 & 0.117 & 0.069 & 0.043 \\
(count s$^{-1}$) & & & & \\
\hline
\end{tabular}
\end{threeparttable}
\end{table}
We conducted two observations on J2118$-$0732 with the {\it XMM-Newton}\ \citep{2001A&A...365L...1J}
in 2016: for 31.9 ks on 2016 May 10 and for 32.8 ks on 2016 October 27
(Observation IDs 0784090201 and 0784090301, PI: Su Yao), hereinafter X-2016-05-10 and X-2016-10-27.
All observations were performed with the European Photon Imaging Camera (EPIC) using
PN \citep[][]{2001A&A...365L..18S} and MOS \citep[][]{2001A&A...365L..27T} CCD arrays
as well as the Optical Monitor \citep[OM;][]{2001A&A...365L..36M}.
OM data analysis will be discussed in Section~\ref{sec:opt/uv}.
All EPIC observations operated in prime large window mode with the thin filter.
The {\it XMM-Newton}\ EPIC observations are summarized in Table~\ref{tab:x_obs}.
Observation data files were processed to create calibrated event lists and full frame images
following standard procedures with the {\it XMM-Newton}\ Science Analysis System (\textsc{sas} Version 16).
Source regions were extracted from a circular aperture around J2118$-$0732
with an optimized radius of 35$\arcsec$.
The background regions were chosen from either a source-free circle for PN or a source-centre
annulus for MOS with similar readout distances to ensure similar background noise levels.
A weak X-ray source near J2118$-$0732 positioned at
R.A. = 21$^{\rm h}$18$^{\rm m}$53.76$^{\rm s}$, Dec. = $-$7$^{\circ}$32$\arcmin$2.4$\arcsec$
was detected by the \textsc{sas} algorithm \texttt{edetect\_chain} and was excluded during the region selections.
Pile-up was negligible in all observations and flaring particle background was filtered.
The resulting net (source minus background) PN count rates and good exposures are
0.193\,cts\,s$^{-1}$ and 18.6\,ks (0.117\,cts\,s$^{-1}$ and 29.3\,ks for the combined MOS spectrum)
and 0.069\,cts\,s$^{-1}$ and 16.2\,ks (0.043\,cts\,s$^{-1}$ and 29.4\,ks
for the combined MOS spectrum) for the first and the second observations, respectively.
The uncertainties correspond to 90 per cent confidence for EPIC results.
\subsection{X-ray spectra and variability}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{./figure/xmm_spectrum.eps}
\caption{The {\it XMM-Newton}\ EPIC spectra and best-fitting results of a single power-law model modified by
the total Galactic absorption for X-2016-05-10 and X-2016-10-27.
Upper panel (a): X-2016-05-12 PN (white circles), X-2016-05-12 MOS (blue hexagons),
X-2016-10-27 PN (black dots), and X-2016-10-27 MOS data (red squares),
as well as the corresponding best fits for an absorbed PL evaluated over 0.3--10\,keV
(black solid line) and 2--10\,keV (black dashed line). Mid panel (b): residuals for fitting the 2--10\,keV
energy alone, extrapolated to lower energies. Lower panel (c): residuals for the best fits over
the whole 0.3--10\,keV energy range.}
\label{fig:x_spectra}
\end{figure}
\begin{table}
\caption{Spectral fits to XMM EPIC observations for J2118$-$0732}
\label{tab:x_spectra_para}
\begin{threeparttable}
\begin{tabular}{lcc}
\hline
Data Set & X-2016-05-10 & X-2016-10-27 \\
Detectors & PN$+$MOS & PN$+$MOS \\
\hline
\multicolumn{3}{l}{\textsc{tbabs}$\times$\textsc{ztbabs}$\times$\textsc{zpow},
fixed $N_{\rm H}$ = $N_{\rm H}^{\rm Gal}= 1.15 \times\,10^{21}$\,cm$^{-2}$} \\
\multicolumn{3}{l}{ at $z$ = 0, free $N_{\rm H}$ at $z$ = 0.26,
fitted over 0.3--10\,keV} \\
$\Gamma_{\rm X}^{a}$ & $1.78^{+0.05}_{-0.03}$ & $1.69^{+0.10}_{-0.09}$ \\
$N_{\rm H}$/10$^{21}$ & $0^{+0.2}$ & $0.3^{+0.5}_{-0.3}$ \\
$\chi^{2}/$ d.o.f. & $133.0/150$ & $87.1/97$ \\
\hline
\multicolumn{3}{l}{\textsc{tbabs}$\times$\textsc{zpow},
fixed $N_{\rm H}$ = $N_{\rm H}^{\rm Gal}= 1.15 \times\,10^{21}$\,cm$^{-2}$ at $z$ = 0,} \\
\multicolumn{3}{l}{fitted over 2--10\,keV and extrapolated to 0.3--10\,keV} \\
$\Gamma_{\rm X}^{a}$ & $1.63^{+0.12}_{-0.11}$ & $1.70 \pm 0.19$ \\
$\chi^{2}/$ d.o.f. & $62.2/73$ & $31.5/34$ \\
\hline
\multicolumn{3}{l}{\textsc{tbabs}$\times$\textsc{zpow},
fixed $N_{\rm H}$= $N_{\rm H}^{\rm Gal}= 1.15 \times\,10^{21}$\,cm$^{-2}$ at $z$ = 0,} \\
\multicolumn{3}{l}{fitted over 0.3--10\,keV} \\
$\Gamma_{\rm X}^{a}$ & $1.78\pm 0.03$ & $1.66 \pm 0.06$ \\
Flux$^b$ & 7.55 $\pm$ 0.25 & 2.97 $\pm$ 0.19 \\
$L_{\rm X}^c$ & 15.8 $\pm$ 0.6 & 6.1 $\pm$ 0.5 \\
$\chi^{2}/$ d.o.f. & $137.5/154$ & $94.4/101$ \\
\hline
\end{tabular}
\begin{tablenotes}
\item[]
$^a$ zpower-law photon index; $^b$ unabsorbed flux integrated from 0.3 to 10\,keV in units of
10$^{-13}$\,erg\,cm\,$^{-2}$\,s$^{-1}$; $^c$ isotropic X-ray luminosity in the 0.3 -- 10\,keV
rest-frame energy band in units of 10$^{43}$\,erg\,s$^{-1}$.
\end{tablenotes}
\end{threeparttable}
\end{table}
The MOS1 and MOS2 spectra were co-added to produce a single combined spectrum.
And all the spectra were binned to contain at least 25 counts per bin needed for $\chi^2$ analysis.
We used \textsc{xspec} (v.12.9.1) to fit PN and the combined MOS spectra simultaneously
over the 0.3--10\,keV energy range.
We first adopted a single PL modified by the Galactic hydrogen absorption
plus a free hydrogen column at the redshift of the source using the Tuebingen-Boulder model
(\textsc{tbabs} and \textsc{ztbabs} in \textsc{xspec}).
We also assumed element abundances from \citet[][]{2000ApJ...542..914W} and photoelectric cross-sections from \citet[][]{1996ApJ...465..487V}.
The Galactic hydrogen column was fixed to the total (H I plus H$_2$ ) Galactic column $N_{\rm H}^{\rm Gal}$
of 1.15 $\times \, 10^{21}$\,cm$^{-2}$ from \citet[][]{2013MNRAS.431..394W}.
The intrinsic absorption is negligible as it was fitted close to zero.
So, the absorption was fixed at the Galactic value to reduce the number of free parameters.
We checked for the existence of a soft excess by modelling the spectra only above 2\,keV and
extrapolating the best fit down to 0.3\,keV.
No significant soft X-ray excess emission was observed in both observations [see panel (b) in Fig.~\ref{fig:x_spectra}].
The X-ray spectra were then fitted with a single PL modified by the total Galactic absorption which resulted in good fits.
The best-fitting indices are $\Gamma_{\rm X}$ = $1.78\pm 0.03$ and
$\Gamma_{\rm X}$ = $1.66\pm 0.06$ for X-2016-05-10 and X-2016-10-27, respectively, which are harder than
normal radio-quiet NLS1s \citep[$\Gamma_{\rm X} \sim$2--4; e.g. ][]{2010ApJS..187...64G,2011ApJ...727...31A},
but similar to some other $\gamma$-ray-emitting NLS1s \citep{2009ApJ...699..976A,2013MNRAS.436..191D}.
The fact that the X-ray spectrum is well modelled by a hard PL suggests that it is typical
of the X-ray emission from jets in blazar-like AGNs and any contribution from a possible corona may be small.
The spectral fit results reported above are shown in Table~\ref{tab:x_spectra_para} and the best-fitting spectra
of a single power-law model modified by the total Galactic absorption are shown in Fig.~\ref{fig:x_spectra}.
The X-ray flux in the 0.3--10\,keV band decreased by a factor of $\sim$2.5 between the two observations.
However, no significant variability was found within each observation.
\section{Other data}
\subsection{Radio properties}
\label{sec:radio}
\begin{figure}
\includegraphics[width=\columnwidth]{./figure/radio_spectrum.eps}
\caption{The radio spectrum of J2118$-$0732 collected from NED and VizieR. The simultaneous data across 72$-$231\,MHz indicated with blue cross marks were obtained from the GLEAM survey.}
\label{fig:radio_spectrum}
\end{figure}
\begin{table*}
\caption{The GHz radio data of J2118$-$0732 retrieved from NED and VizieR.}
\label{tab:radio_spec}
\begin{threeparttable}
\begin{tabular}{lccccc}
\hline
Frequency & Flux & Polarized & Resolution & Observation & Surveys and \\
(GHz) & (mJy) & flux (mJy) &(arc sec) & date & references \\
\hline
0.074 & $750 \pm 130 $ & & 80 & 2003-09-20& (1) \\
0.15 & $ 383.5\pm 39$ & & 25 & 2016-03-15 & (2) \\
0.17--0.231 & $307.3 \pm 9.1^a$ & & $\sim100$ & 2013-08-13 & (3) \\
0.365& $277 \pm 30$ & & 6 & & (4) \\
1.4& $ 98.7 \pm 20 $ & & 5 & 1997-05-26 & (5) \\
1.4& $ 96.1 \pm 2.9$ &3.16 & 45 &1993-09-20 & (6) \\
4.85 & $102 \pm 12 $ & & 168 &1990-11-06 & (7) \\
8.4 & $97.5 $ & & & & (8) \\
19.9 & $113 \pm 6 $ & 6 & 34 & 2007 Oct 26-30 & (9) \\
\hline
\end{tabular}
\begin{tablenotes}
\item[] Surveys and References:
(1) The VLA Low-Frequency Sky Survey \citep[VLSS,][]{2007AJ....134.1245C},
(2) The Giant Metrewave Radio Telescope (GMRT) 150 MHz All-sky Radio Survey
\citep[][]{2017A&A...598A..78I}, (3) GaLactic and Extragalactic All-sky Murchison Widefield Array
(GLEAM) survey \citep[][]{2017MNRAS.464.1146H},
(4) The Texas Survey of radio sources at 365MHz \citep[][]{1996AJ....111.1945D},
(5) The Faint Images of the Radio Sky at 20 cm Survey \citep[FIRST,][]{1997ApJ...475..479W},
(6) The NRAO VLA Sky Survey \citep[NVSS,][]{1998AJ....115.1693C},
(7) The Parkes-MIT-NRAO Surveys \citep[PMN surveys,][]{1995ApJS...97..347G},
(8) The Combined Radio All-Sky Targeted Eight GHz Survey \citep[CRATES,][]{2007ApJS..171...61H},
(9) The Australia Telescope 20 GHz Survey \citep[AT20G survey,][]{2010MNRAS.402.2403M} \\
$^a$ integrated flux between 170 and 231 MHz.
\end{tablenotes}
\end{threeparttable}
\end{table*}
J2118$-$0732 was detected as a single unresolved source within 5$\arcsec$ of the SDSS position in several surveys
from the NASA/IPAC Extragalactic Database website (NED)\footnote{\url{http://ned.ipac.caltech.edu}} and
VizieR website\footnote{\url{http://vizier.u-strasbg.fr/index.gml}} listed in Table~\ref{tab:radio_spec}.
The Faint Images of the Radio Sky at 20 cm Survey \citep[FIRST;][]{1997ApJ...475..479W} observation
indicates a compact radio structure at FIRST resolution (5$\arcsec$).
The NRAO VLA Sky Survey \citep[NVSS;][]{1998AJ....115.1693C} 1.4 GHz flux density is
$96.1 \pm 2.9$\,mJy, corresponding to a radio power of
$P_{\rm 1.4GHz} \sim (1.98 \pm 0.06)\,\times 10^{25}$\,W\,Hz$^{-1}$, and the polarized flux density is
$3.16 \pm 0.49$\,mJy, thus a fractional polarization of $\sim3.3$ per cent.
The source shows a steep spectrum between 74\,MHz and 1.4\,GHz with
$\alpha_{\rm rad} = -0.7$ \citep[][]{2010A&A...511A..53V}, however,
a flat/inverted spectrum above 1.4\,GHz (see Fig.~\ref{fig:radio_spectrum}).
We refitted the radio spectrum using a broken PL with a break frequency at 1.4\,GHz and obtained
$\alpha_\mathrm{rad} = -0.66$ in the lower frequency band and $\alpha_{\rm rad} = 0.15$ in the higher
frequency band without considering the variability of the radio flux as the data were not all simultaneous.
However, the synchronous observations across 72--231\,MHz from the GaLactic and Extragalactic All-sky Murchison Widefield Array (GLEAM) survey \citep[][]{2017MNRAS.464.1146H} indicate
the sincerity of a steep radio spectrum at lower frequency.
To estimate the radio loudness parameter defined as
$R_{5 \,{\rm GHz}} \equiv f_{\nu}(5\,{\rm GHz})/f_\nu$(4400\,\AA),
we used the 4.85\,GHz flux from the Parkes-MIT-NRAO surveys \citep[PMN surveys;][]{1995ApJS...97..347G}
and the 4400 \AA\ flux from the SDSS g band point-spread function (PSF)-model magnitude with Galactic extinction considered.
The $k$-correction with the radio spectral index $\alpha_{\rm rad} = 0.15 $ and
the optical index $\alpha_{\rm opt} = -1.5$ was adopted.
It results in $R_{5\, {\rm GHz}} \sim920$, which is consistent with
$R = f_\nu(5\,{\rm GHz})/f_\nu$(2500\,\AA) $\approx1227$ derived from \citet{2011ApJS..194...45S}.
However, it cannot be ruled out, though unlikely,
that a strong starburst may contribute to the radio emission.
Such an estimation of radio luminosity from a starburst contribution can be obtained
using the observed far-infrared (FIR)-radio luminosity correlation from \citet[][]{2001ApJ...554..803Y} and
the extrapolated FIR luminosity [using the Wide-field Infrared Survey Explorer \citep[{\it WISE};][]{2010AJ....140.1868W} data and a power-law assumption in the infrared band].
This gives $P_{1.4\,{\rm GHz}} \sim 8.9 \times\,10^{22}$\,W\,Hz$^{-1}$, which is about two orders of magnitude
smaller than the observed radio luminosity, excluding a strong starburst contribution to the radio emission.
Besides, there is also a possibility that intrinsic reddening may affect the optical continuum slope.
Nevertheless, just using the radio power itself, $P_{1.4\, {\rm GHz}} \sim 2\times 10^{25}$\,W\,Hz$^{-1}$, makes J2118$-$0732 formally radio-loud.
\subsection{{\it \textbf{WISE}} infrared data and intraday variability}
\label{sec:wise}
\begin{figure}
\includegraphics[width=\columnwidth]{./figure/wise_lightcurve.eps}
\caption{$w1$ (upper panel) and $w2$ (lower panel) light curves of J2118$-$0732 obtained from the {\it WISE} data base.
The numbers in the upper panel refer to the epoch number of each intraday epoch.
The data showing significant intraday variability are differentiated with diamond marks.
Epoch (6) in blue and (7) in red correspond to X-2016-05-10 and X-2016-10-27, respectively.}
\label{fig:wise_lc}
\end{figure}
J2118$-$0732 was observed with the {\it WISE}
in four bands $w1$, $w2$, $w3$, $w4$ centred at 3.4, 4.6, 12 and 22\,$\mu$m.
Its long-term light curves of $w1$ and $w2$ bands were constructed from the PSF profile-fit photometric magnitudes
via the near-Earth objects {\it WISE} (NEOWISE) Reactivation Single-exposure data base \citep[][]{2014ApJ...792...30M}
and the AllWISE Multi-epoch Photometry data base (see Fig.~\ref{fig:wise_lc}),
while $w3$ and $w4$ light curves are not available due to the lack of the data in the NEOWISE data base.
Those data with poor signal-to-noise ratios (S/N $<$ 10 or marked as 'null')
or the reduced $\chi^2$ of the profile-fit photometries larger than 2 were excluded.
Significant intraday variability was detected in epoch 4, 5 in the $w1$ band and epoch 5, 6 in the $w2$ band,
with $p$-values $< 0.1$ per cent using the $\chi^2$-test against the null hypothesis of no variation.
This variability time-scale sets an upper limit on the size of the emitting region to
$\lse 8 \times 10^{-4}$ \,pc,
which is much smaller than the scale of a putative torus but consistent with that of the jet-emitting region.
The last two epochs of $w1$ and $w2$ light curves coincided with our two {\it XMM-Newton}\ observations (see
Section~\ref{sec:xmm}) which provided the quasi-simultaneous infrared fluxes together with the X-ray and optical/UV fluxes.
The average fluxes during these two epochs were calculated to construct the broad-band SEDs in Section~\ref{sec:sed}
and are reported in Table~\ref{tab:opt/uv}.
A 1$\sigma$ error of 0.2 magnitude was assumed due to the significant intraday variability of the infrared fluxes.
\subsection{Optical/UV photometry and the morphology of the host galaxy}
\label{sec:opt/uv}
\begin{table*}
\caption{Infrared/optical/UV photometric results before corrected for Galactic extinction.}
\label{tab:opt/uv}
\begin{tabular}{lcccccccccccccccccccccccccccccc}
\hline
& \multicolumn{30}{c}{{\it WISE} profile-fit photometric magnitudes} \\
Date & \multicolumn{15}{c}{$w1$ (3.4 $\mu$m)} & \multicolumn{15}{c}{$w2$ (4.6 $\mu$m)} \\
\hline
2016-05-10 & \multicolumn{15}{c}{13.529 $ \pm$ 0.2 } & \multicolumn{15}{c}{12.516 $ \pm$ 0.2} \\
2016-10-27 & \multicolumn{15}{c}{14.054 $ \pm$ 0.2} & \multicolumn{15}{c}{12.876 $ \pm$ 0.2} \\
\hline
& \multicolumn{30}{c}{OM aperture magnitudes} \\
Date & \multicolumn{5}{c}{UVW2 (2120 \AA)} & \multicolumn{5}{c}{UVM2 (2310 \AA)} &
\multicolumn{5}{c}{UVW1 (2910 \AA)} & \multicolumn{5}{c}{U (3440 \AA)} &
\multicolumn{5}{c}{B (4500 \AA)} & \multicolumn{5}{c}{V (5430 \AA)} \\
\hline
2016-05-10 & \multicolumn{5}{c}{} & \multicolumn{5}{c}{} &
\multicolumn{5}{c}{19.07 $ \pm$ 0.11} & \multicolumn{5}{c}{19.34 $ \pm$ 0.08} &
\multicolumn{5}{c}{19.82 $ \pm$ 0.05} & \multicolumn{5}{c}{19.40 $ \pm$ 0.09} \\
2016-10-27 & \multicolumn{5}{c}{} & \multicolumn{5}{c}{} &
\multicolumn{5}{c}{20.18 $ \pm$ 0.17} & \multicolumn{5}{c}{20.24 $ \pm$ 0.10} &
\multicolumn{5}{c}{20.61 $ \pm$ 0.10} & \multicolumn{5}{c}{} \\
\hline
& \multicolumn{30}{c}{SDSS PSF-model magnitudes} \\
Date & \multicolumn{5}{c}{u (3543 \AA)} & \multicolumn{5}{c}{g (4770 \AA)} &
\multicolumn{5}{c}{r (6231 \AA)} & \multicolumn{5}{c}{i (7625 \AA)} &
\multicolumn{5}{c}{z (9097 \AA)} \\
\hline
2000.6741 & \multicolumn{5}{c}{20.144 $ \pm$ 0.056} & \multicolumn{5}{c}{19.770 $ \pm$ 0.019} &
\multicolumn{5}{c}{19.072 $ \pm$ 0.023} & \multicolumn{5}{c}{18.785 $ \pm$ 0.021} &
\multicolumn{5}{c}{18.339 $ \pm$ 0.032} \\
2004.6988 & \multicolumn{5}{c}{20.101 $ \pm$ 0.048} & \multicolumn{5}{c}{19.691 $ \pm$ 0.023} &
\multicolumn{5}{c}{19.118 $ \pm$ 0.024} & \multicolumn{5}{c}{18.892 $ \pm$ 0.022} &
\multicolumn{5}{c}{18.399 $ \pm$ 0.036} \\
2008.7315 & \multicolumn{5}{c}{19.887 $ \pm$ 0.052} & \multicolumn{5}{c}{19.255 $ \pm$ 0.021} &
\multicolumn{5}{c}{18.601 $ \pm$ 0.018} & \multicolumn{5}{c}{18.274 $ \pm$ 0.020} &
\multicolumn{6}{c}{17.827 $ \pm$ 0.029} \\
\hline
& \multicolumn{30}{c}{GALEX calibrated magnitudes} \\
Date & \multicolumn{15}{c}{FUV (1516 \AA)} & \multicolumn{15}{c}{NUV (2267 \AA)} \\
\hline
2003-08-06 & \multicolumn{15}{c}{21.58 $ \pm$ 0.12} & \multicolumn{15}{c}{21.38 $ \pm$ 0.10} \\
\hline
\end{tabular}
\end{table*}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{./figure/SDSS_image.eps}
\caption{SDSS r band image of J2118$-$0732. The '$\times$' mark shows the optical position of the source
and the colour bar is in units of nanomaggy.}
\label{fig:sdss_image}
\end{figure}
The source was also observed with the OM detector onboard the {\it XMM-Newton}\ configured in 'imaging' mode
with an exposure of 5\,ks for each filter both on X-2016-05-10 and X-2016-10-27.
We extracted the background-subtracted photometric data using the \texttt{omichain} processing pipeline
with the default parameter settings, as recommended by the \textsc{sas} threads.
The source was not detected in the $UVM2$ and $UVW2$ bands in both observations, and in the $V$ band
in the second observation utilizing the \texttt{omdetect} searching algorithm with a minimum significance of 3.
We also collected non-simultaneous data from the SDSS Photometric Catalog Data Release 9
\citep[][]{2012ApJS..203...21A} and the Medium Imaging Survey (MIS) observed by
the Galaxy Evolution Explorer \citep[GALEX;][]{2005ApJ...619L...1M}.
The SDSS performed photometric measurements three times on J2118$-$0732 from 2000 to 2008
which showed variations of $\sim$0.5\,mag on time-scale of years.
The source was also detected in the GALEX Far-UV ($FUV$) and Near-UV ($NUV$) bands with magnitudes FUV=21.58
and NUV=21.38 centred at an effective wavelength of 1516\,\AA\ and 2267\,\AA\
in the AB magnitude system\footnote{\url{http://galex.stsci.edu/GalexView/}}.
The collection of the optical/UV magnitudes from the OM, the SDSS, and the GALEX observations
is also reported in Table~\ref{tab:opt/uv}.
All the optical/UV magnitudes were corrected for Galactic extinction using extinction law described in
Section~\ref{sec:spectroscopy} before further analysis. Particularly for the GALEX magnitudes, we used
$A_{\rm FUV}/E(B-V) =$ 8.376 and $A_{\rm NUV}/E(B-V) =$ 8.741,
following \citet{2005ApJ...619L..15W}.
Of particular interest, the optical image of J2118$-$0732 obtained from the SDSS (see Fig.~\ref{fig:sdss_image}) shows an extended galaxy structure and likely shows a disturbed morphology, possibly suggesting a recent merger.
\section{broad-band Spectral Energy Distribution}
\label{sec:sed}
\begin{figure*}
\centering
\subfigure[the EC/torus model]{
\includegraphics[width=\columnwidth]{./figure/sed_rest_torus.eps}
}
\subfigure[the EC/BLR model]{
\includegraphics[width=\columnwidth]{./figure/sed_rest_blr.eps}
}
\caption{The rest-frame broad-band SEDs of J2118$-$0732 and their best-fitting
EC/torus models (a) and EC/BLR models (b).
The black squares are the data from 2016-05-10 and the red dots are the data from 2016-10-27
covering the infrared, optical/UV, X-ray, and $\gamma$-ray bands.
Other non-simultaneous data marked as black crosses are also included into the SED fitting.
Besides, the GALEX data denoted as purple stars are also included in the high state
to constrain the disc component.
The solid grey/black lines show the best-fitting jet/jet$+$disc models for 2016-05-10 data
and the dashed magenta lines show the best-fitting jet models for 2016-10-27 data.
The solid/dashed orange lines, cyan lines, and spring green lines indicate the synchrotron (Syn),
synchrotron-self-Compton (SSC), and External Compton (EC) emissions for the high/low state.
And the solid blue lines indicate the disc models for the high state.}
\label{fig:bbsed}
\end{figure*}
\begin{table}
\caption{Parameters of the best-fitting SED models in the EC/torus and EC/BLR cases.}
\label{tab:fit_para}
\begin{threeparttable}
\begin{tabular}{lcccc}
\hline
& \multicolumn{2}{c}{2016-05-10 (high state)} &\multicolumn{2}{c} {2016-10-27 (low state)} \\
& EC/torus & EC/BLR & EC/torus & EC/BLR \\
\hline
$\gamma_{\rm min}^a$ & 1 & 78 & 1 & 78 \\
$\gamma_{\rm b}^b$ & 575 $\pm$ 146 & 269 $\pm$ 81 & 617 $\pm$ 185 & 372 $\pm$ 111 \\
$\gamma_{\rm max}^c$ & 20000 & 6000 & 15000 & 10000 \\
$p_{1}^d$ & 2.02 & 1.54 & 1.92 & 1.2 \\
$p_{2}^e$ & 3.8 & 3.8 & 3.92 & 3.9 \\
$N_{0}^f$ & 120 $\pm$ 61 & 28 $\pm$ 8 & 19 $\pm$ 11 & 2.6 $\pm$ 0.7 \\
$\delta^g$ & 8.9 $\pm$ 1.5 & 5.9 $\pm$ 0.7 & 10.3 $\pm$ 1.5 & 5.8 $\pm$ 0.7 \\
$B^h$ & 1.0 $\pm$ 0.3 & 3.7 $\pm$ 0.8 & 0.9 $\pm$ 0.3 & 2.6 $\pm$ 0.5 \\
${\chi^2_{\rm red}}^i$ & 0.5 & 0.7 & 0.3 & 0.5 \\
$L_{\rm bol}^j$ & 2.46 & 2.88 & & \\
$\lambda^k$ & 0.06 & 0.07 & & \\
\hline
\end{tabular}
\begin{tablenotes}
\item[]
$^a$ minimum Lorentz factor of the injected electrons. \\
$^b$ break Lorentz factor of the injected electrons. \\
$^c$ maximum Lorentz factor of the injected electrons. \\
$^d$ low-energy electron spectral index. \\
$^e$ high-energy electron spectral index. \\
$^f$ electron density parameter in units of 10$^{2}$\,cm$^{-3}$. \\
$^g$ Doppler boosting factor which equals to bulk Lorentz factor assumed in \citet{2015ApJ...807...51Z}.\\
$^h$ magnetic field in units of Gauss. \\
$^i$ reduced $\chi^2$ of the SED fittings with one-zone leptonic jet model. \\
$^j$ bolometric luminosity $L_\mathrm{bol}$ = $L_\mathrm{disc}$ in units of 10$^{44}$\, erg\,s$^{-1}$. \\
$^k$ Eddington ratio calculated from $\lambda = L_\mathrm{bol}/L_\mathrm{Edd}$.
\end{tablenotes}
\end{threeparttable}
\end{table}
For $\gamma$-ray loud NLS1s like J2118$-$0732, which exhibit the hybrid properties of both NLS1s and blazars, it would be revealing to study their broad-band SEDs.
Here, we made use of the {\it XMM-Newton}\ data and the {\it WISE} data to construct the quasi-simultaneous SEDs
in a relatively high state (on 2016 May 10) and low state (on 2016 October 27).
The radio spectrum was taken from the NED and VizieR websites as presented in Section~\ref{sec:radio}.
We also made use of the GALEX data in the $UV$ band to constrain the disc component in the high state because of the similar UV flux levels of the GALEX data and the OM data.
The broad-band SEDs are shown in Fig.~\ref{fig:bbsed}.
The source was not detected in $\gamma$-rays at the epochs around the {\it XMM-Newton}\ observations within a time span of 60 d (see the study on the $\gamma$-ray light curve in Section~\ref{sec:fermi-results}).
The flux limits derived above are denoted as arrows.
As a first-order approximation, the average $\gamma$-ray spectrum obtained from all the data over the total 9-yr time span was used instead, which is consistent with the upper limits.
We assume that the $\gamma$-ray emission in the SED of J2118$-$0732 at the epochs concerned can be approximated by the averaged spectrum.
This allows us to perform spectral modelling of the SED, which is expected to be still revealing, though the best-fitting model and parameters may suffer from uncertainties to some extent due to possible spectral variations.
To fit the broad-band SED, we first used a simple one-zone leptonic jet model, which consists of
synchrotron, synchrotron self-Compton (SSC), and external-Compton (EC) processes
\citep[e.g.][]{2009ApJ...707L.142A,2012MNRAS.426..317D,2013ApJ...768...52P,2015ApJ...798...43S}.
The $\chi^2$ minimization technique is applied to perform the fitting.
The radio emission may come from radio lobes at larger scales and/or superposition of multiple jet components
and thus does not fit to the synchrotron component which is self-absorbed below 10$^{11}$\,Hz.
Two scenarios were considered for the EC scattering process:
a seed photon field dominated from the torus and from the BLR, respectively.
The energy density of the former was assumed to be $3 \times\,10^{-4}$\,erg\,cm$^{-3}$
\citep[see more details in][]{2007ApJ...660..117C, 2008MNRAS.387.1669G} and the latter
was estimated from the broad H\ensuremath{\beta}\ line luminosity derived in Section~\ref{sec:spectroscopy}
\citep[see more details in][]{2015ApJ...798...43S,2015ApJ...807...51Z}.
The energy distribution of the injected relativistic electrons was assumed to be a broken PL in the range of [$\gamma_{\rm min}$, $\gamma_{\rm max}$] with a break energy $\gamma_{\rm b}$.
The value of $\gamma_{\rm min}$ was taken as $\gamma_{\rm min} = 1$ in the EC/torus model
which was constrained from the soft X-ray slope of the SSC bump and $\gamma_{\rm min} = 78$ in the EC/BLR model
which is the mean value of the $\gamma_{\rm min}$ distribution for a GeV-NLS1 sample \citep[][]{2015ApJ...798...43S}.
The $\gamma_{\rm max}$ is usually poorly constrained and
does not significantly affect our results, so it was fixed at a large value.
The two indices $p_1$, $p_2$ of the broken PL were derived from the spectral indices of the observed SEDs.
Given the prominent broad emission lines in the optical spectrum of J2118$-$0732,
a thermal emission contributed by an accretion disc is very likely present in J2118$-$0732.
This is particularly true when J2118$-$0732 was in the high state, where the UV fluxes calculated
from the OM data and the GALEX data started to rise towards the frequency range near the peak of the big blue bump.
We used a multitemperature blackbody model of a standard thin disc to account for this component in the high state.
But for the low state, we did not add a disc model due to the lack of available data points
covering the frequency range where the disc is the strongest.
A black hole mass of 3.4 $\times$ 10$^7$\,$M_{\sun}$ was employed.
We also note that the optical fluxes should be treated as upper limits (marked as arrows in Fig.~\ref{fig:bbsed})
caused by the contamination from the host galaxy, as revealed by the SDSS optical image (see Fig.~\ref{fig:sdss_image}) which shows an
extended galaxy structure, while the UV fluxes\footnote{\label{foot:uvbands}The observational $U$ band of the OM will be
blueshifted to the UV band at the rest frame considering the redshift of J2118$-$0732.} are not.
Since the hard X-ray spectrum is more flat than the average of radio-quiet NLS1s,
we assume that it is dominated by jet emission (in addition to fainter emission from the corona).
For simplicity, in the SED modelling, we therefore do not include coronal emission.
Both the EC/torus and the EC/BLR models can adequately fit the broad-band SEDs and
the disc component can be roughly constrained in the high state using the additional GALEX data.
The parameters of the fitting are given in Table~\ref{tab:fit_para}, together with their 1$\sigma$ errors.
As can be seen in Fig.~\ref{fig:bbsed}, the infrared to ultraviolet is dominated by the synchrotron radiation
and the $\gamma$-ray band is dominated by the EC radiation, respectively.
The EC peak of the EC/torus model is broader than that of the EC/BLR model and extends to the X-ray band
while only the SSC peak dominates the X-ray band in the EC/BLR model.
The best-fitting parameters show that the EC/BLR model contains a higher magnetic field and
the Doppler boosting factor is smaller compared with that of the EC/torus model.
Noticing that the main difference of the broad-band SEDs of these two models lies in the luminosity from 10\,keV up to
100\,MeV, hard X-ray and/or soft $\gamma$-ray data would help to distinguish between the EC/torus and the EC/BLR models.
However, only the upper limit of the source in the hard X-ray and/or soft $\gamma$-ray band can
be constrained from surveys like the latest 105-month Swift-BAT all-sky hard X-ray survey in the 14--195\,keV
\citep[][]{2018ApJS..235....4O} which is not sensitive enough to discriminate between the two models for the SED yet.
Further data are needed to distinguish between the two models and give better understandings of the EC process in astrophysical jets.
The multiwavelength SEDs of J2118$-$0732 changed from a high state on 2016 May 10 to
a low state on 2016 October 27 with a synchronous drop of the broad-band fluxes from infrared to X-rays in five months.
The electron density parameter varied by about an order of magnitude between the two states for both models.
The Eddington ratio was estimated to be $\approx0.07$ in the high state.
For the low state, a reliable estimate of the Eddington ratio cannot be made,
but it is probably lower than that of the high state, as indicated by its lower UV fluxes.
The best-fitting parameter values, including the magnetic fields, the Doppler boosting factors, and the electron densities are within the typical range of some other GeV NLS1s
\citep[e.g.][]{2009ApJ...707L.142A,2011MNRAS.413.1671F,2012MNRAS.426..317D,2013ApJ...768...52P,2015ApJ...798...43S,2015AJ....150...23Y}
and similar to those of blazars \citep[e.g.][]{2010MNRAS.401.1570T,2015ApJ...807...51Z}.
However, the exact values of these parameters may be subject to relatively large uncertainties, given the degeneracies among some of the parameters \citep[e.g.][]{1998ApJ...509..608T,2012ApJ...752..157Z} as well as the non-simultaneity of the multiwavelength data.
\section{Discussions}
\label{sec:discussion}
\subsection{Blazar-like properties of J2118$-$0732}
\label{sec:blazar-like}
Blazar-like characteristics have been found in some RL NLS1s \citep{2008ApJ...685..801Y},
suggesting the presence of relativistic jets which was later confirmed by
the detection of $\gamma$-ray emission from some of these objects with the {\it Fermi}.
The similar multiwavelength properties of J2118$-$0732 make it a new member of this interesting and rare class of objects.
The flat radio spectrum above 1.4\,GHz of J2118$-$0732 is probably the result of a superposition of several jet
components \citep{1981ApJ...243..700K} while the steep spectrum below 1.4\,GHz is likely coming from radio lobes.
The short time-scale of infrared emission is consistent with the size of the jet-emission region.
Besides, the hard X-ray spectrum of J2118$-$0732 is similar to those of some blazar-like NLS1s,
hinting perhaps at a jet origin.
Furthermore, \citet{2014ApJS..215...14D} found that the {\it WISE} mid-infrared colours of J2118$-$0732 are similar to those
of the confirmed $\gamma$-ray emitting blazars that occupy the so-called {\it WISE} blazar locus
\citep[][]{2013ApJS..206...12D} in the {\it WISE} colour space,
which makes J2118$-$0732 a '{\it WISE} blazar-like radio-loud source'.
The blazar-like characteristics of J2118$-$0732 are also reminiscent of another $\gamma$-ray-emitting NLS1s -- PKS~2004$-$447.
PKS~2004$-$447 has a very high radio loudness (1710 $< R <$ 6320) and a low BH mass
($M_{\rm BH} = 10^{6.7}\,M_{\sun}$) at a redshift of 0.24.
The jet emission is also suggested to dominate at multiple wavelengths in PKS~2004$-$447 which is supported by the
multiwavelength analysis and the SED modelling
\citep{2009ApJ...707L.142A,2013ApJ...768...52P,2015MNRAS.453.4037O}.
Moreover, both J2118$-$0732 and PKS~2004$-$447 show weak {\rm Fe\,{\footnotesize II}}\ emission
and lack an obvious soft X-ray excess\footnote{We note a tentative soft excess
has been reported in only one of the three {\it XMM-Newton}\ observations of PKS~2004$-$447
\citep[][]{2006MNRAS.370..245G,2016A&A...585A..91K}.}, while a soft X-ray excess has been detected in most
other $\gamma$-loud NLS1s that have good-quality X-ray spectra
\citep[e.g.][]{2008MNRAS.388L..54D,2014MNRAS.438.3521D,2015A&A...575A..13F,2018MNRAS.475..404K,2018MNRAS.476...43L}.
\subsection{Association of the $\gamma$-ray source with J2118$-$0732}
\label{sec:association}
Even though the blazar-like characteristics make the RL NLS1 J2118$-$0732 a highly possible counterpart for the
$\gamma$-ray source detected by the {\it Fermi}-LAT and both \citet{2018ApJ...853L...2P}
and FL8Y reported a high association probability of the source, we still make a sanity check of the association
between the two within the relatively large error circle of the $\gamma$-ray source detection.
There are 11 radio sources detected in the FIRST survey within the gamma-error circle above the 1.5\,mJy flux threshold
and J2118$-$0732 is the brightest one with its radio flux higher than the combined flux of the rest 10 sources.
Its monochromatic radio luminosity ($2.7 \times 10^{41}$\,erg\,s$^{-1}$) at 1.4\,GHz
and the 9-yr averaged monochromatic $\gamma$-ray luminosity at 1\,GeV ($0.9 \times 10^{44}$\,erg\,s$^{-1}$) of the $\gamma$-ray source
follow the correlation between the two luminosities as claimed by \citet{2016ApJS..226...20F}.
The source detection algorithm on the EPIC images found 45 X-ray sources within the gamma-error circle and
J2118$-$0732 is the brightest X-ray source with its X-ray flux in accordance with other $\gamma$-ray emitting {\it Fermi} blazars from a sample in \citet{2016ApJS..226...20F}.
Although a small part of the $\gamma$-ray error circle is not covered by the EPIC field of view,
the only detection of J2118$-$0732 within this error circle by the RASS can set an upper limit on the fluxes of other X-ray sources.
Furthermore, J2118$-$0732 is the only X-ray bright source in the EPIC image which has a radio counterpart among 11
radio sources detected in the FIRST catalogue, making it most likely associated with the $\gamma$-ray source.
No known blazar is discovered by checking every X-ray source and radio source found above.
However, an unambiguous and independent confirmation should come from future monitoring observations
which search for the correlated variability of $\gamma$-ray emission with emission from other wavelengths.
\subsection{$\gamma$-ray Variability}
\label{sec:gamma_vaviability}
So far, only a limited number of $\gamma$-loud NLS1s have been detected by {\it Fermi}-LAT including J2118$-$0732.
However, the $\gamma$-ray properties are different from source to source.
Three objects, PMN~J0948$+$0022, SBS~0846$+$513, and 1H~0323$+$342, have shown short-term strong
$\gamma$-ray flares emerging and dying out within one month \citep[][]{2012MNRAS.426..317D,2015MNRAS.446.2456D}.
Other $\gamma$-ray sources have not shown strong flares so far.
J2118$-$0732 was in a relatively high-flux state during 2009--2013
and dimmed below the detection limit of {\it Fermi}-LAT after 2013.
During the active 4 yr, at least one flare was detected.
The long-term activity with relatively high fluxes lasting for years of J2118$-$0732 may be a common feature
since similar behaviours have been observed among some other $\gamma$-ray loud NLS1s: a nine-month flare
was observed in a long time variability analysis of FBQS~J1644$+$2619 \citep{2015MNRAS.452..520D},
and a relatively high-flux state was also observed in SDSS~J144246.29$+$474129.4
in late 2012 lasting for several months \citep{2015arXiv151005584L}.
This may indicate that a number of RL NLS1s may radiate in $\gamma$-rays
but remain undetected by the {\it Fermi}-LAT because they spend more of the time in a low-flux state.
Therefore, because of this characteristic of long-term variability, more RL NLS1s, especially those with the highest radio loudness, are expected to be observed in the $\gamma$-ray band in the future.
\section{Summary}
\label{sec:summary}
$\gamma$-ray-emitting NLS1s have been providing new insights into the formation and evolution of relativistic jets
and the jet-disc coupling mechanism under the extreme conditions of low BH masses and high accretion rates.
Here, we present our discovery of a new $\gamma$-ray-emitting NLS1-- SDSS~J211852.96$-$073227.5
by analysing the SDSS spectrum, the {\it Fermi}-LAT, and the {\it XMM-Newton}\ observational data.
The multiwavelength observations presented here allow a comprehensive study of the properties of
its broad-band radiation. The main results are summarised below:
\begin{enumerate}
\item[1.] The optical spectrum of J2118$-$0732 from the SDSS confirms its NLS1 nature: a broad H\ensuremath{\beta}\ emission line with a width (FWHM) of
1585\,km\,s$^{-1}$ and a flux ratio [O\,{\footnotesize III}]/H\ensuremath{\beta}\ $\approx1.7$ even though the {\rm Fe\,{\footnotesize II}}\ emission is rather weak.
The estimated BH mass is $\sim$3.4 $\times 10^7\,M_{\sun}$ and the Eddington ratio is $\sim$0.15.
\item[2.] The {\it Fermi}-LAT observations centred at J2118$-$0732 detected a new $\gamma$-ray source
with a 9-yr averaged isotropic luminosity of 0.7 \,$\times\,10^{45}$\,erg\,s$^{-1}$
in the 0.1--300\,GeV rest-frame energy band.
The new $\gamma$-ray source experienced a relatively high state for about 4\,yr
and remained in a quiescent state in other {\it Fermi}-LAT observations.
At least one flare was emerging during the long-term activity.
\item[3] Our two {\it XMM-Newton}\ observations covering a significant part of the $\gamma$-ray source error circle
reveal that J2118$-$0732 is the brightest X-ray source in the field, and the one with large X-ray variability.
The X-ray spectrum of J2118$-$0732 is flat, with no significant evidence of a soft excess.
\item[4.] Other blazar-like properties of J2118$-$0732 include a large radio loudness with a flat/inverted radio
spectrum and remarkably rapid infrared variability of less than a day, detected by {\it WISE}.
This short time-scale implies a compact emission region much more compact than the torus.
\item[5.] The broad-band SED can be modelled by emission from a simple one-zone leptonic jet, and the flux variation from infrared to X-rays in five months can be explained by changes of the jet parameters.
Besides, there may exist a relatively strong disc component, at least in the high state.
We note the caveat that, however, the exact values of the fitted parameters in our model may be subject to relatively large uncertainties, given the degeneracies among some of the model parameters as well as the non-simultaneity of the multiwavelength data.
\end{enumerate}
\section*{Acknowledgements}
We thank the anonymous referee for helpful comments and suggestions.
This work is supported by the Natural Science Foundation of China (grants No. 11473035, 11573034, and 11773037). SY and YL acknowledge the support by the KIAA-CAS Fellowship, which is jointly supported by Peking University and Chinese Academy of Sciences.
SY thanks for the support by Boya fellowship.
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions,
the National Science Foundation, and the U.S. Department of Energy Office of Science.
The SDSS-III web site is \url{http://www.sdss3.org/}.
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III
Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory,
Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group,
Harvard University, the Instituto de Astrofisica de Canarias, the Michigan State/Notre Dame/JINA Participation Group,
Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics,
Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University,
Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group,
University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington,
and Yale University.
The {\it Fermi}-LAT Collaboration acknowledges generous ongoing support from a number of agencies and institutes
that have supported both the development and the operation of the LAT as well as scientific data analysis.
These include the National Aeronautics and Space Administration and the Department of Energy in the United States,
the Commissariat \`{a} l'Energie Atomique and the Centre National de la Recherche Scientifique / Institut National de
Physique Nucl\'{e}aire et de Physique des Particules in France, the Agenzia Spaziale Italiana and the Istituto Nazionale di
Fisica Nucleare in Italy, the Ministry of Education, Culture, Sports, Science and Technology (MEXT), High Energy
Accelerator Research Organization (KEK) and Japan Aerospace Exploration Agency (JAXA) in Japan, and the K. A.
Wallenberg Foundation, the Swedish Research Council and the Swedish National Space Board in Sweden.
Additional support for science analysis during the operations phase is gratefully acknowledged from the Istituto
Nazionale di Astrofisica in Italy and the Centre National d'Etudes Spatiales in France.
Based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA Member States and NASA.
This publication makes use of data products from the Wide-field Infrared Survey Explorer,
which is a joint project of the University of California, Los Angeles, and the Jet Propulsion Laboratory/California Institute
of Technology, funded by the National Aeronautics and Space Administration.
This work is based in part on observations made with the Galaxy Evolution Explorer (GALEX).
GALEX is a NASA Small Explorer, whose mission was developed in cooperation with the Centre National d'Etudes
Spatiales (CNES) of France and the Korean Ministry of Science and Technology. GALEX is operated for NASA by the
California Institute of Technology under NASA contract NAS5-98034.
This research has made use of the NASA/IPAC Infrared Science Archive and Extragalactic Database (NED),
which are operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration.
This research has also made use of the VizieR catalogue access tool, CDS, Strasbourg, France.
|
\section{Introduction}
\label{Intro}
In this work, we study the dynamics and fractal geometry of the minimal sets for generic Kuperberg flows on 3-manifolds.
The minimal sets resemble, in many ways, the strange attractors that arise in physics, and one of the outstanding open problems is to understand the dimension theory of Kuperberg minimal sets, and its dependance on the dynamics.
Krystyna Kuperberg showed in the work \cite{Kup1} that every closed 3-manifold admits a smooth flow with no periodic orbits.
Her proof was based on the construction of a smooth aperiodic flow in a \textit{plug}, which is a compact three-manifold with boundary.
This plug is inserted in flows to break open periodic orbits.
It is known that the flow in the plug preserves a unique minimal set $ \mathcal{M} $, and that under generic assumptions, $ \mathcal{M} $ is a codimension one lamination with a Cantor transversal, as was shown in \cite{Hur}.
The dynamics of Kuperberg flows have been previously studied in \cite{Kup2}, \cite{Ghy}, \cite{Hur}, \cite{Hur2}, \cite{Hur3}, and \cite{Mat}, and it is known that the topology of $ \mathcal{M} $ is particularly complicated.
There have been many notable contributions to the dimension theory of limit sets of dynamical systems in dimensions higher than two; see \cite{Bot}, \cite{Sim}, \cite{Sim2}, \cite{Sim3} for some examples.
A common theme in these works is hyperbolicity in the dynamics and a reduction to one dimension via stable manifolds.
However, as stated in the survey \cite{Sch},
\vspace{0.1cm}
\begin{displayquote}
``Even for the simplest examples of higher dimension [than 2] we are far from a general theory of the Hausdorff dimension of limit sets.''
\end{displayquote}
\vspace{0.1cm}
The Kuperberg flow does not resemble these systems, because by a theorem of Katok \cite{Kat}, an aperiodic flow cannot preserve a hyperbolic measure.
Though the flow is not hyperbolic and has zero entropy, arbitrarily small perturbations of it are hyperbolic and have positive entropy (see \cite{Hur2}). For this reason, the dynamics of the Kuperberg flow are said to lie ``at the boundary of hyperbolicity."
In two dimensions, this type of behavior is present in H\'{e}non-like families and Kupka-Smale diffeomorphisms (see \cite{Bed3},\cite{Cao},\cite{Lep}).
Studying the fractal geometry and dimension theory of the Kuperberg minimal set makes a new contribution to a general dimension theory for limit sets in dimension three, in the absence of hyperbolicity.
Fortunately, the characterization of the minimal set as a codimension one lamination reduces the dimension theory to that of the transverse Cantor set.
Without this, the study of its fractal geometry and dimension theory would be completely intractable.
The dimension theory of Cantor sets in the the line has a vast literature, particularly for limit sets of iterated function systems, graph directed systems, and their generalizations.
However, the transverse Kuperberg minimal set poses new challenges in this direction as well.
These come from the complicated symbolic dynamics of the action of the holonomy pseudogroup associated to the flow, which is not semiconjugate to a subshift.
In this paper we propose a general framework for treating the symbolic dynamics of limit sets of pseudogroups, and apply this to a transverse section of the Kuperberg minimal set.
We build a symbolic model of this transverse Cantor set and extract a graph directed subspace by analyzing the pseudogroup.
This allows the application of results from one-dimensional thermodynamic formalism to obtain dimension estimates, which are then extended to the minimal set via the product structure.
\subsection{The Kuperberg flow}
Kuperberg's construction is the first-- and only currently known-- smooth flow on $ S^3 $ with no periodic orbits.
This was discovered as a counterexample to Seifert's conjecture.
\subsubsection{Seifert's conjecture}
A vector field on a manifold is said to have a closed orbit if one of its integral curves is homeomorphic to $ S^1 $.
The Hopf vector field on $ S^3 $, whose integral curves form the Hopf fibration, has all orbits closed.
In 1950, Seifert \cite{Sei} showed that every nonsingular vector field on $ S^3 $ sufficiently close to the Hopf vector field also has a closed orbit, and then asked if every continuous vector field on $ S^3 $ does.
The generalized Seifert conjecture asked this question for any compact orientable $ n $-manifold with Euler characteristic zero.
Counterexamples in dimension four and greater were discovered in 1966 by Wilson \cite{Wil}, who constructed the first \textit{plug}, the product of a closed rectangle with a torus, which carries a smooth vector field satisfying certain properties.
A plug is a manifold with boundary, together with a smooth vector field.
If this local vector field satisfies some symmetry conditions, the plug can be inserted into a manifold carrying a global vector field, in such a way that the local dynamics in the plug are compatible with the global dynamics.
If the plug intersects a periodic orbit, the plug's interior dynamics can break it.
Using this method, Wilson constructed smooth counterexamples to Seifert's conjecture in dimension greater than or equal to four.
Seifert's conjecture is trivial in dimension two, so Seifert's conjecture only remained unsolved in dimension three, although Wilson did succeed in showing that on every closed connected three-manifold there exists a smooth vector field with only \textit{finitely many} closed orbits.
The first counterexample to Seifert's conjecture in dimension three was constructed in 1972 by Schweitzer \cite{Schw}.
This counterexample used a plug supporting an aperiodic vector field of class $ C^1 $.
In 1988, Harrison \cite{Har} modified Schweitzer's construction to class $ C^2 $, but serious obstructions remained in extending to $ C^{\infty} $.
For an account of Schweitzer's and Harrison's constructions, see \cite{Ghy}.
\subsubsection{Kuperberg's plug}
In 1994, Kuperberg \cite{Kup1} constructed a $ C^{\infty} $ counterexample to Seifert's conjecture in dimension three.
This construction began with a modified Wilson plug embedded in $ \mathbb{R}^3 $ containing two periodic orbits.
Kuperberg then used self-intersections to break the periodic orbits inside Wilson's plug without creating new periodic orbits.
See \cite{Kup1}, \cite{Ghy}, \cite{Hur}, and \cite{Mat} for descriptions of Kuperberg's construction.
\subsubsection{Kuperberg's minimal set}
Ghys \cite{Ghy} showed that Kuperberg's plug contains a unique minimal set.
Using a numerical simulation due to B. Sevannec, he obtained an image of this minimal set on a transverse section of the plug.
See Figure \ref{matfig}.
\begin{figure}[h]
\includegraphics[width=0.6\linewidth, trim={2.95cm 1.4cm 0 0}, clip]{matsumoto}
\caption{Cross-section of Kuperberg minimal set, from Ghys (\cite{Ghy})}
\label{matfig}
\end{figure}
Ghys encouraged an investigation into the properties of this minimal set, and asked how such properties depend on Kuperberg's construction.
A closer study of the topology and dynamics of the minimal set was carried out by Hurder and Rechtman \cite{Hur}.
To answer Ghys' question, they defined a special class of flows called \textit{generic Kuperberg flows} that preserves a unique minimal set with the following characterization.
\begin{theorem}(\cite{Hur}, Theorem 17.1)
\label{minchar3}
Let $ K $ be the Kuperberg plug, $ \psi_t : K \rightarrow K $ a generic Kuperberg flow, and $ \mathcal{M} \subset K $ the minimal set.
Then $ \mathcal{M} $ is a codimension one lamination with a Cantor transversal $ \tau $.
Furthermore, there exists a closed surface $ \mathcal{R}' \subset K $ such that
$$
\mathcal{M} = \overline{\bigcup_{-\infty < t < \infty} \psi_t(\mathcal{R}')}.
$$
\end{theorem}
The surface $ \mathcal{R}' $ is called the \textit{notched Reeb cylinder}. Because of Theorem \ref{minchar3}, the fractal geometry of $ \mathcal{M} $ can be studied by analyzing the orbit of the $ \mathcal{R}' $.
Perhaps the first question in this direction is the Hausdorff dimension of the minimal set.
Because of the local product structure implied by this theorem, the dimension theory of $ \mathcal{M} $ reduces to that of $ \tau $.
The study of dynamically defined Cantor sets and their dimension theory has a long history.
\subsection{Iterated function systems and limit sets of group actions}
A large class of fractals are the limit sets of \textit{iterated function systems}, which were introduced by Hutchinson \cite{Hut}.
\subsubsection{Iterated function systems}
Let $ X $ be a compact space, and $ E = \{1,\ldots,p\} $ a finite alphabet.
An \textit{iterated function system} is a collection $ \{ \phi_i : X \rightarrow X \}_{i \in E} $ of injective contracting maps, with a common Lipschitz constant $ 0 < s < 1 $.
Each iterated function system has an invariant \textit{limit set}:
$$
J = \bigcap_{n = 1}^{\infty} \bigcup_{(i_1,\ldots,i_n) \in E^n} \phi_{i_1} \circ \cdots \circ \phi_{i_n} (X).
$$
With appropriate separation conditions, $ J $ is a Cantor set.
There is a $ p $-to-$ 1 $ expanding map $ S $ whose inverse branches are $ \phi_i $, and the dynamics of $ S |_J $ is conjugate to the one-sided shift on $ E^{\mathbb{N}} $.
For an introduction to iterated function systems, see chapter 9 of \cite{Fal2}.
There are many generalizations of iterated function systems, including graph-directed Markov systems.
\subsubsection{Graph directed Markov systems}
Let $ (V,E) $ be directed graph with finite vertex and edge sets $ V $ and $ E $, respectively.
Each edge $ e $ has an initial vertex $ i(e) \in V $ and terminal vertex $ t(e) \in V $.
Let $ A: E \times E \rightarrow \{0,1\} $ be the edge incidence matrix of this directed graph, so if $ A_{ee'} = 1 $, then $ t(e) = i(e') $.
For each $ v \in V $, let $ X_v $ be a metric space, and for each $ e \in E $ let $ \phi_e : X_{t(e)} \rightarrow X_{i(e)} $ be an injective contraction map.
If the maps $ \{ \phi_e \}_{e \in E} $ have a common Lipschitz constant $ 0 < s < 1 $, the collection is called a \textit{graph directed Markov system}.
For each $ n \geq 1 $, the matrix $ A $ determines the following space of \textit{admissible words} of length $ n $:
$$
E_A^n = \{ \omega \in E^{\mathbb{N}} : A_{\omega_i, \omega_{i+1}} = 1 \text{ for all } 1 \leq i \leq n-1 \}.
$$
In terms of these, the system has an invariant limit set:
$$
J = \bigcap_{n = 1}^{\infty} \bigcup_{(i_1,\ldots,i_n) \in E_A^n} \phi_{i_1} \circ \cdots \circ \phi_{i_n} \left(X_{t(\omega_n)} \right).
$$
As with iterated function systems, these limit sets are often Cantor sets, and their dynamics are conjugate to a subshift of finite type over the alphabet $ E $.
In some cases, the limit set of a discrete group $ \Gamma = \langle g_1, \ldots, g_n \rangle $ acting on a compact space $ X $ can be realized as the limit set of an graph-directed system defined by the generators $ g_i $ and their images $ g_i(X) $.
Here are some examples.
\begin{itemize}
\item \textit{Expanding maps}: A distance expanding map $ f: X \rightarrow X $ of a metric space $ X $ defines a semigroup action of $ \mathbb{N} $ on $ X $.
Such a map has a Markov partition of arbitrarily small diameter (see \cite{Rue}).
Defining the iterated function system to be the inverse branches of $ f $, the limit set of this action is the limit set of the graph directed system whose incidence matrix is the matrix defining the Markov partition.
\vspace{0.1cm}
\item \textit{Fuchsian groups}: Let $ \Gamma $ be a Fuchsian group acting on the hyperbolic disc $ \mathbb{H}^2 $. Bowen \cite{Bow20} related the action of $ \Gamma $ on its boundary circle $ \partial \mathbb{H}^2 = S^1 $ to an expanding Markov map $ f : S^1 \rightarrow S^1 $.
This correspondence is called the \textit{Bowen-Series coding}; via this correspondence, these actions are orbit equivalent.
As above, the inverse branches of $ f $ form a graph directed system with admissible words coded by the matrix defining the Markov map $ f $.
\vspace{0.1cm}
\item \textit{Schottky groups}: Another example is the limit set of a finitely generated Kleinian group of Schottky type, acting on the Riemann sphere.
It can be shown that such a limit set is the limit set of an appropriately defined graph directed system.
For details, see Chapter 5 of \cite{Mau3}.
\vspace{0.1cm}
\end{itemize}
\subsubsection{Infinitely generated function systems and pseudo-Markov systems}
There are many generalizations of iterated function systems and graph directed systems.
These include the infinite iterated function systems of Mauldin and Urba\'{n}ski \cite{Mau1} and the pseudo-Markov systems of Stratmann and Urba\'{n}ski (\cite{Str}).
The former can be used to describe sets of complex continued fractions (see \cite{Mau2}), and the latter are models of limit sets of infinitely generated Schottky groups (see \cite{Str}), among many other applications.
The dynamics of a graph directed function system on its limit set is semiconjugate to a shift over a sequence space of admissible words.
This is the domain of symbolic dynamics, and the ergodic properties of such systems is well studied.
One of the advantages of relating the limit set of a group to the limit set of a function system, is that the symbolic dynamics of the function system can then be used to study the symbolic dynamics of the group action.
Once such a connection has been made, the fractal geometry of the limit set of the group can be studied using techniques from iterated function systems.
The patterns that emerge when ``zooming in" to the fractal by applying maps in the function system, are the same as those that emerge by applying the generators of the group to a fundamental domain.
These regular patterns are captured by the incidence matrix determining the admissible words in the coding of the limit set.
\subsection{General function systems and limit sets of pseudogroup actions}
Pseudogroups are a generalization of groups of transformations of metric spaces (see \cite{Hae}).
A primary application of pseudogroups is in the dynamics of foliations and laminations.
Compositions of transition maps of a foliation or lamination comprise its \textit{holonomy} pseudogroup.
For a flow that does not admit a global section, the collection of first-return maps to a section also forms a pseudogroup.
For an exposition of the dynamics of pseudogroups see \cite{Hur0} and \cite{Wal}.
Limit sets of pseudogroup actions have a similar definition to those of group actions, but are generally more difficult to study.
They can be fractals, but they need not exhibit the same self-similarity evident in limit sets of groups.
In Chapter \ref{genfun}, we define the notion of a \textit{general function system}.
The limit set of such a system is a fractal that need not be self-similar.
This provides a framework to relate the limit sets of pseudogroups to those of function systems.
The transverse Cantor set of the Kuperberg minimal set is the limit set of a pseudogroup action on the transversal.
The pseudogroup here is the holonomy of the foliation by flowlines of the Kuperberg flow.
In Chapter \ref{Transcant}, we will relate this set to the limit set of a general function system.
\subsection{Symbolic dynamics and thermodynamic formalism}
Let $ E $ be an alphabet (finite or infinite).
The dynamics of the shift map on invariant subspaces of the sequence space $ E^{\mathbb{N}} $ is well studied.
The shift map has an associated \textit{topological pressure} that is related to ergodic properties of measures supported on the space.
This is part of the thermodynamic formalism developed by Sinai, Ruelle, and Bowen (see \cite{Sin}, \cite{Rue}, and \cite{Bow1}).
For generalized systems such as infinite iterated function systems and pseudo-Markov systems, there are extensions of the thermodynamic formalism (see \cite{Mau3}).
In Chapter \ref{Thermo}, we will define the topological pressure in an appropriate context.
\subsubsection{Symbolic dynamics of limit sets of graph directed systems}
For graph directed systems, there is a bijective coding map $ \pi : \Sigma \rightarrow J $, where $ \Sigma \subset E^{\mathbb{N}} $ is a compact shift-invariant subset, and $ J $ is the limit set of the system.
This map intertwines the system's dynamics on $ J $ with the shift on $ \Sigma $.
Following Barriera \cite{Bar1} we say that the function system is \textit{modeled} by the subshift $ \Sigma $.
In this way, symbolic quantities such as pressure have natural analogues defined entirely in terms of the function system.
If the function system is assumed to have regularity $ C^{1+\alpha} $ for some $ \alpha > 0 $, the pressure has additional uniformity properties that makes its definition particularly transparent.
In Chapter \ref{Conf} we will present the pressure in this context, and study these properties.
\subsubsection{Symbolic dynamics of limit sets of general function systems}
General function systems are coded by more general sequence spaces, including spaces that are not shift-invariant.
These are also introduced in Chapters \ref{Thermo} and \ref{Conf}.
In later chapters we will equate the transverse Kuperberg minimal set to the limit set $ J $ of a general function system, and show that there is a bijective correspondence $ \pi : \Sigma \rightarrow J $, where $ \Sigma \subset \mathbb{N}^{\mathbb{N}} $ is a sequence space that is not shift-invariant.
As with subshifts, we say that such a general function system is \textit{modeled} by this general symbolic space $ \Sigma $.
The definition of limit sets of general function systems resembles that of graph directed systems.
However, their fractal geometry is \textit{a priori} more complicated than their graph directed counterparts, and exhibits less self-similarity.
Applying the maps in the function system, we ``zoom in" on the fractal, but the regular patterns present in graph directed systems do not emerge, because the underlying dynamics are those of a pseudogroup rather than those of a group.
The limit sets of actions of pseudogroups is not as widely studied as those of groups and can exhibit substantially more pathology.
The ergodic theory and symbolic dynamics of these systems is still being developed (see \cite{Wal}).
Progress in this direction includes the entropy theory of Ghys, Langevin, and Walczak \cite{Ghy2}.
However, it is not at all clear how to develop a thermodynamic formalism or to define quantities such as pressure for limit sets of pseudogroups and of function systems that are coded by these general symbolic spaces.
\subsubsection{Dual symbolic spaces}
In his study of differentiable structures on Cantor sets, Sullivan \cite{Sul} defined the notion of a \textit{dual} Cantor set.
The symbolic description of the dual is given by simply reversing the coding and reading the words in the opposite order.
The distortion of a fractal in a metric space can be quantified by its \textit{ratio geometry}.
The ratio geometry is a sequence of real numbers that measure the self-similarity defect of the fractal; if the sequence is constant, the fractal is self-similar and its similarity coefficient is equal to this constant.
The asymptotic ratio geometry is called the \textit{scaling function} and is viewed as a function on the symbolic space coding the fractal.
Sullivan proved that for Cantor sets defined by $ C^{1+\alpha} $ function systems, the scaling function on the dual is an invariant of the differential structure.
In Chapter \ref{Kupmin} we will see that dual Cantor sets arise naturally in our study of the symbolic dynamics of the Kuperberg minimal set.
We present the dual of a symbolic space in Chapter \ref{dualcant} in the context of general symbolic spaces appropriate for coding the limit sets of general function systems.
For references on Sullivan's theorem and dual Cantor sets, see \cite{Bed2}, \cite{Prz1}, and \cite{Prz2}.
\subsection{Symbolic dynamics of the Kuperberg minimal set}
We now return to the Kuperberg flow, its minimal set, and the fractal geometry of the minimal set.
In Chapter \ref{Wilsflow} we briefly present the general theory of plugs, and summarize Wilson's construction \cite{Wil} of a vector field on a mirror-image plug with two periodic orbits.
In Chapter \ref{Kupsection}, we summarize Kuperberg's construction of a plug $ K $ \cite{Kup1}, using self-insertions to modify Wilson's plug.
The flow of the resulting vector field on $ K $ is called the Kuperberg flow $ \psi_t $.
The images of these periodic orbits under the quotient map are called the \textit{special orbits}.
To simplify the problem, it is necessary to make additional assumptions on the construction $ K $ and $ \psi_t $.
These assumptions are listed in Chapter \ref{InsertAssume}, and are compatible with the generic hypotheses on Kuperberg flows given in \cite{Hur}.
Under these assumptions, we can write the insertion maps in coordinates and explicitly integrate the Kuperberg vector field.
The dynamics of $ \psi_t $ are complicated, but there are several important notions that allow us to relate these to the simpler dynamics of the Wilson flow.
These notions are called \textit{transition} and \textit{level}; they were defined by Kuperberg in \cite{Kup1} and used extensively in \cite{Hur}, \cite{Ghy}, and \cite{Mat}.
We can decompose orbits of points in $ K $ by level, and relate each level set to an orbit in Wilson's plug.
We make this precise in Chapter \ref{levo}.
\subsubsection{The Kuperberg pseudogroup}
In Chapter \ref{Kuppseudo} we commence the study of the holonomy pseudogroup associated to $ \psi_t $.
This flow does not admit a global section, so we choose a convenient local section defined in Chapter \ref{InsertAssume}.
It is the union of two rectangles transverse to the flow, that lie in the entrances to the two insertion regions.
In Chapters \ref{Kupmin} through \ref{FunctionSystems}, we restrict to just one which we refer to as $ S $.
The map taking a point $ x \in S $ to its first return under $ \psi_t $ generates a pseudogroup $ \Psi $.
Using the theory of levels from Chapter \ref{Kupsection}, we first show that this pseudogroup is generated by the first-return maps of the Wilson flow, together with the insertion maps.
The intersection of the notched Reeb cylinder $ \mathcal{R}' $ with $ S $ is a curve $ \gamma $.
In view of Theorem \ref{minchar3}, the intersection $ \mathcal{M} \cap S $ is the closure of the orbit of the curve $ \gamma $ under this pseudogroup.
Because our assumptions in Chapter \ref{InsertAssume} allowed us to integrate the Kuperberg flow and write the insertion maps in coordinates, we then set out to explicitly parametrize the transition curves in the intersection $ \mathcal{M} \cap S $.
We carry this out in Chapter \ref{Kupmin}.
See Figure \ref{TheMin} for a picture of some of these curves.
\begin{figure}[h]
\includegraphics[width=0.9\linewidth, trim={0 0.5 0 0.8cm}, clip]{TheMinimalSet}
\caption{The first two iterations in the recursive construction of the transverse minimal set. Compare with Figure \ref{matfig}.}
\label{TheMin}
\end{figure}
\subsubsection{Interlaced Cantor sets}
Through Chapter \ref{FunctionSystems}, we only consider the first return to $ S $, one of the two rectangular regions defined by the insertions.
To account for the entire minimal set, we must also consider points that enter the other insertion region, before intersecting $ S $.
These points also form a Cantor set in $ S $, and because of the symmetry of the plug, these Cantor sets are identical.
In Chapter \ref{Transcant} we will prove that these two Cantor sets are \textit{interlaced}, and that $ \mathcal{M} \cap S $ is equal to this interlaced Cantor set.
The symbolic dynamics of two interlaced Cantor sets modeled by sequence spaces $ \Sigma $ and $ \Xi $, is defined naturally by the induced dynamics on a \textit{joint} sequence space $ \Sigma \ast \Xi $.
These terms will be defined precisely in Chapter \ref{sectioninter}.
\subsubsection{Symbolic dynamics of the Kuperberg minimal set}
Using the theory of levels, we prove that each curve in $ \mathcal{M} \cap S $ is coded by a word $ \omega $ in an appropriate general sequence space, whose word length corresponds to the level of the curve.
These can be used to code the points in $ \tau \subset S $, the Cantor transversal of $ \mathcal{M} $.
The space $ \Sigma $ of admissible words is not shift-invariant, and depends delicately on the symbolic dynamics of the Kuperberg pseudogroup.
The number of words in each level depends on the \textit{escape times} of curves in $ \mathcal{M} \cap S $ under the pseudogroup.
In general, it is impossible to predict the exact escape times of all curves in $ \mathcal{M} \cap S $.
However, in Chapter \ref{Transcant} we give an iterative construction of the sequence space $ \Sigma $ in terms of these escape times.
In Chapter \ref{FunctionSystems}, we use the Kuperberg pseudogroup and projection maps along the leaves of the lamination $ \mathcal{M} $ to define a general function system on the transversal.
Using the symbolic dynamics developed in Chapters \ref{Kuppseudo} and \ref{Kupmin}, we show that this general function system is modeled by the dual of the sequence space $ \Sigma $, in the sense of Sullivan.
This allows us to prove the following theorem.
\begin{theorem*}[\textbf{A}]
Let $ \mathcal{M} $ be the Kuperberg minimal set with Cantor transversal $ \tau $.
There is a sequence space $ \Sigma \subset \mathbb{N}^{\mathbb{N}} $ and a $ C^{1+\alpha} $ general function system on $ [0,1] $ modeled by the dual $ \widetilde{\Sigma} $, with limit set $ \tau $.
\end{theorem*}
As we show in Chapter \ref{Conf}, limit sets of general function systems modeled by a sequence space have a bijective coding to the space.
Then as an immediate corollary to Theorem \textbf{A}, we obtain
\begin{corollary*}[\textbf{B}]
Let $ \mathcal{M} $ be the Kuperberg minimal set with Cantor transversal $ \tau $.
Then there exists a sequence space $ \Sigma \subset \mathbb{N}^{\mathbb{N}} $ and bijective coding map
$$
\pi : \Sigma \rightarrow \tau.
$$
\end{corollary*}
This coding of $ \tau $ by $ \Sigma $ will be crucial later, when estimating the dimension.
\subsection{Dimension theory of limit sets}
In his study of the limit sets Fuchsian groups, Bowen \cite{Bow2} related the thermodynamic formalism to dimension theory.
In this setting, the pressure defined by the symbolic dynamics depends only on a parameter $ t \in \mathbb{R} $, and can thus be viewed as a function $ p : \mathbb{R} \rightarrow \mathbb{R} $.
Bowen proved that this function has a unique zero that coincides with the Hausdorff dimension of the limit set.
This relation is known as \textit{Bowen's equation} for dimension.
This equation-- and its subsequent generalizations in other settings-- is now ubiquitous in the dimension theory of dynamical systems.
There is an immediate analogue of Bowen's equation for limit sets of graph directed Markov systems.
Similarly, there is an analogue for each generalization, including graph directed and pseudo-Markov systems.
In Chapter \ref{Dime} we will present the pressure function and Bowen's equation in the appropriate generality.
For a proof of Bowen's equation for limit sets of finite iterated function systems, see \cite{Bed1}.
For generalizations of Bowen's equation, see \cite{Mau1}, \cite{Mau3}, and \cite{Str}, in increasing order of generality.
For general expositions of applications of thermodynamic formalism to dimension theory see \cite{Pes2}, \cite{Fal1}, \cite{Prz1} and \cite{Sch}.
In the survey \cite{Sch}, Schmeling and Weiss point out how pervasive Bowen's ideas are in the dimension theory of dynamical systems.
\vspace{0.1cm}
\begin{displayquote}
``One of the most useful techniques in the subject is to obtain a \textit{Bowen formula} for the Hausdorff dimension of a set, i.e. to obtain the Hausdorff dimension as the zero of an expression involving the thermodynamic pressure. Most dimension formulas for limit sets of dynamical systems and geometric constructions in the literature are obtained, or can be viewed, as Bowen formulas.''
\end{displayquote}
\vspace{0.1cm}
For this reason, to study the dimension theory of a set as complicated as the transverse minimal set $ \tau $ in the Kuperberg plug, it seems necessary to have the full power of the thermodynamic formalism at our disposal.
However, we have already noted that for limit sets of pseudogroups and general function systems modeled by sequence spaces that are not shift-invariant, such a formalism does not exist.
Thus, it is necessary to relate $ \tau $ to a more tractable function system, for instance the pseudo-Markov systems of Stratmann and Urba\'{n}ski (\cite{Str}).
\subsection{A graph directed subspace of $ \Sigma $}
We carry out this analysis in Chapters \ref{Kupmin} and \ref{GDPMsub}.
For each $ \epsilon > 0 $, let $ S_{\epsilon} \subset S $ be a sub-rectangle of width $ \epsilon $.
By analyzing the parametrizations of these curves and their images under the generators of $ \Psi $, we obtain bounds (with error) on the escape times of curves in $ \mathcal{M} \cap S_{\epsilon} $.
The error in these bounds decreases as $ \epsilon \rightarrow 0 $.
Because $ \Sigma $ is defined in terms of escape times, we thus extract a subspace $ \Sigma_{\epsilon} \subset \Sigma $ that we can determine explicitly for small $ \epsilon $.
We then show that the bijective coding $ \pi : \Sigma \rightarrow \tau $ restricts to a bijective coding $ \pi : \Sigma_{\epsilon} \rightarrow \tau_{\epsilon} $, where $ \tau_{\epsilon} $ is the intersection of $ \tau $ with an $ \epsilon $-neighborhood of the critical orbit in $ K $.
Fortunately, for small enough $ \epsilon > 0 $, the fractal $ \tau_{\epsilon} $ exhibits much more self-similarity than is evident in $ \tau $.
The next theorem exploits this self-similarity.
\begin{theorem*}[\textbf{C}]
Let $ \mathcal{M} $ be the Kuperberg minimal set, with Cantor transversal $ \tau $.
Let $ \tau_{\epsilon} $ be the intersection of $ \tau $ with an $ \epsilon $-neighborhood of the critical orbit in $ K $.
For sufficiently small $ \epsilon > 0 $ there is a $ C^{1+\alpha} $ graph directed pseudo-Markov system on $ [0,\epsilon] $ with limit set $ \tau_{\epsilon} $.
\end{theorem*}
\subsection{Dimension theory of the Kuperberg minimal set}
Theorem \textbf{C} shows that the general function system modeled by $ \Sigma $ from Theorem \textbf{B} has a graph-directed subsystem modeled by $ \Sigma_{\epsilon} \subset \Sigma $.
Thus for small enough $ \epsilon $, we can invoke the dimension theory developed in Chapter \ref{Dime} for graph directed systems to obtain results about the dimension theory of $ \tau_{\epsilon} $.
\subsubsection{Properties of the dimension}
To relate this to the dimension theory of $ \tau $, we first state the following global-to-local result.
\begin{lemma*}[\textbf{D}]
Let $ \tau $ be the transverse Cantor set of the Kuperberg minimal set, and let $ \tau_{\epsilon} $ be the intersection of $ \tau $ with an $ \epsilon $-neighborhood of the critical orbit in $ K $.
Then for any $ \epsilon > 0 $,
$$
\text{dim}_H(\tau) = \text{dim}_H(\tau_{\epsilon}).
$$
\end{lemma*}
We prove this lemma in Chapter \ref{Dimtrans}.
Applying the thermodynamic formalism for graph directed systems from Chapter \ref{Dime}, we obtain the following theorem.
\begin{theorem*}[\textbf{E}]
Let $ \tau $ be the transverse Cantor set of the Kuperberg minimal set. Then the Lebesgue measure of $ \tau $ is zero, and $ 0 < \dim_H(\tau) < 1 $.
\end{theorem*}
\subsubsection{Numerical estimates for dimension}
Finally we turn to numerical dimension results.
The Kuperberg flow is defined in terms of several external parameters, the most important being its angular speed $ a > 0 $.
To numerically estimate dimension using Bowen's equation, it is necessary to calculate the pressure function and its zero explicitly.
Besides calculating the dimension, we are interested in its dependence on the parameter $ a > 0 $.
As we show in Chapter \ref{Dime}, the pressure function depends on the symbolic dynamics and the derivatives of the maps comprising the function system.
Both of these quantities depend on external parameters, including $ a $.
The symbolic dynamics are determined by the space $ \Sigma_{\epsilon} $, which we have calculated by virtue of Theorem \textbf{C}.
However, the function system on $ [0,1] $ from Theorem \textbf{B} is defined in terms of the Kuperberg pseudogroup and projection maps along leaves.
Explicit calculation of the derivatives of these maps seems impossible.
Fortunately, in regularity $ C^{1+\alpha} $, the derivatives of the maps can be related to ratio geometry of the limit set.
This is the \textit{bounded distortion property} from one-dimensional dynamics, used by Shub and Sullivan (\cite{Shu}), and is presented in Chapter \ref{Dime}.
This reduces the pressure calculation to the estimation of the ratio geometry of the transverse Cantor set $ \tau $.
A detailed study of this ratio geometry is carried out in Chapter \ref{Transversal}.
In this chapter, we use the parametrizations of the curves calculated in Chapter \ref{Kupmin} and study their intersections with the transversal.
As with the symbolic dynamics, by restricting to a suitably small $ \epsilon $-neighborhood of the critical orbit, we obtain explicit bounds on the ratio geometry.
The simplest type of ratio geometry is that of \textit{stationary} systems, such as iterated function systems whose maps are similarities.
Such systems have a clean numerical dimension theory that depends on the \textit{ratio coefficients} of the system (see \cite{Pes2}).
In this direction, we define in Chapter \ref{Dime} an \textit{asymptotically stationary function system with error $ a_{\delta} $} for some $ \delta $.
This error is a function $ a_{\delta} : \Sigma \rightarrow \mathbb{R}_{\geq 0} $ that decreases to zero as $ \delta $ does.
The ratio geometry of the limit set of such a function system differs from that of a stationary system by this error.
As long as the error satisfies a natural summability condition, the pressure function for an asymptotically stationary system approaches that of a stationary system and allows for numerical estimates.
In Chapter \ref{Transversal}, we show that for any $ \delta > 0 $, there exists $ \epsilon > 0 $ such that pseudo-Markov system whose limit set is $ \tau_{\epsilon} $ is asymptotically stationary with summable error $ a_{\delta} $.
This can be used to obtain the following dimension estimates.
\begin{theorem*}[\textbf{F}]
Let $ \tau $ be the Cantor transversal of the Kuperberg minimal set.
Let $ t = \text{dim}_H(\tau) $ be its Hausdorff dimension, and $ a > 0 $ the angular speed of the Kuperberg flow.
\begin{itemize}
\item $ t = \text{dim}_H(\tau) $ is the unique zero of a dynamically defined pressure function,
\item $ t $ depends continuously on $ a $,
\item For any $ a $ we may compute $ t $ to a desired level of accuracy.
\end{itemize}
\end{theorem*}
We conclude Chapter \ref{Dimtrans} by extending the results of Theorems \textbf{E} and \textbf{F} to the entire minimal set $ \mathcal{M} $, using the product structure from Theorem \ref{minchar}.
In Chapter \ref{Furth}, we survey some remaining open questions related to the dimension theory of the Kuperberg minimal set.
\subsection{Acknowledgements}
The author owes a debt of gratitude to Steve Hurder for his guidance and support for the duration of this project.
\vfill
\eject
\section{Symbolic spaces over an infinite alphabet}
\label{Thermo}
In this chapter we will fix some important notation that will be used throughout the paper.
The notation of graph-directed symbolic spaces is standard and we follow some commonly observed conventions.
The main reference here is \cite{Mau1} (see also \cite{Bow1}, \cite{Mau1} \cite{Rue}).
We then introduce general symbolic spaces and symbolic spaces of infinite type, which are natural generalizations of graph-directed symbolic spaces.
We conclude by presenting dual symbolic spaces.
\subsection{Countable alphabets}
Let $ E \subset \mathbb{N} $ be a countable alphabet, and let $ E^{\ast} = \bigcup_{n \geq 1} E^n $ and $ E^{\infty} = E^{\mathbb{N}} $ be the finite and infinite words in $ E $, respectively.
If $ \omega \in E^{\ast} $ then $ \omega \in E^n $ for some $ n $ and we say $ |\omega| = n $ is the word length of $ \omega $.
If $ \omega \in E^{\infty} $, we set $ |\omega|=\infty $.
If $ \omega \in E^{\ast} \cup E^{\infty} $ and $ n \leq |\omega| $, we denote by $ \omega |_n $ the truncated word $ (\omega_1, \ldots, \omega_n) $.
If $ \omega \in E^{\ast} $ is a finite word, we denote
$$
[\omega] = \{ \tau \in E^{\infty} : \tau |_{|\omega|} = \omega \}.
$$
We have a countable-to-one left shift map $ \sigma: E^{\infty} \rightarrow E^{\infty} $.
With the convention $ \frac{1}{2^{\infty}} = 0 $, the space $ E^{\ast} \cup E^{\infty} $ is metrizable in the usual metric
$$
d(\omega, \tau) = \frac{1}{2^{|c(\omega, \tau)|}},
$$
where $ c(\omega, \tau) $ is the longest common initial subword of $ \omega $ and $ \tau $.
\subsection{General and infinite type symbolic spaces}
\subsubsection{General symbolic spaces}
Let $ \Sigma \subset E^{\ast} $ be a collection of finite words.
Because the alphabet $ E $ is countable, in general $ \Sigma $ is infinite.
For each such $ \Sigma $ and $ n \geq 1 $, let
$$
\Sigma_n = \{ \omega \in \Sigma : |\omega| = n \}.
$$
The symbolic spaces that arise naturally in our applications will satisfy the following property.
\begin{definition}[Extension admissibility property]
\label{extadm}
We say that $ \Sigma \subset E^{\ast} $ satisfies the \textit{extension admissibility property} if $ \Sigma_n \neq \emptyset $ for all $ n \geq 1 $, and for all $ (\omega_1, \ldots, \omega_n) \in \Sigma_n $ with $ n > 1 $, we have $ (\omega_1, \ldots, \omega_{n-1}) \in \Sigma_{n-1} $.
\end{definition}
We will refer to spaces $ \Sigma \subset E^{\ast} $ satisfying the extension admissibility property as \textit{general symbolic spaces}.
These spaces have words of arbitrary length, and each word is comprised of admissible subwords.
Such spaces need not be shift-invariant, and the spaces we will consider in our applications will not be.
\subsubsection{Symbolic spaces of infinite type}
Let $ \Sigma \subset E^{\infty} $ be a closed subspace.
For each $ n \geq 1 $ define
$$
\Sigma_n = \{ \omega |_n : \omega \in \Sigma \}.
$$
This definition is compatible with the one given above for spaces of finite words.
There is a natural analogue of Definition \ref{extadm} for these spaces.
\begin{definition}[Restriction admissibility property]
\label{restadm}
We say that $ \Sigma \subset E^{\infty} $ satisfies the \textit{restriction admissibility property} if for all $ \omega \in \Sigma $ and for all $ n > 1 $ with $ \omega |_n \in \Sigma_n $, we have $ \omega |_{n-1} \in \Sigma_{n-1} $.
\end{definition}
We will refer to spaces $ \Sigma \subset E^{\infty} $ satisfying the restriction admissibility property as \textit{symbolic spaces of infinite type}.
There is a natural way of obtaining a space of infinite type from a general symbolic space, and vice versa, called extension and restriction.
There are versions of these notions for sequences of words, and those of spaces.
\subsubsection{Extension and restriction of words}
\label{extrestwords}
Fix a general symbolic space, and consider a sequence of finite words
$$
(\omega_1, \ldots, \omega_n) \in \Sigma_n,
$$
defined for all $ n \in \mathbb{N} $.
In terms of this, we define $ \omega \in E^{\infty} $ by
$$
\omega = (\omega_1, \omega_2, \ldots),
$$
so that $ \omega |_n = (\omega_1, \ldots, \omega_n) $.
The word $ \omega \in E^{\infty} $ is called the \textit{infinite extension} of the sequence $ (\omega_1, \ldots, \omega_n) $.
Similarly, if $ \Sigma \subset E^{\infty} $ is a symbolic space of infinite type, for each word $ \omega \in \Sigma $ we obtain a sequence $ \omega |_n \in \Sigma_n $ by truncating.
This is naturally a sequence in $ E^{\ast} $, and we call it the \textit{finite restriction} of $ \omega $.
Extension and restriction are naturally dual to each other.
If $ (\omega_1, \ldots, \omega_n) \in \Sigma_n $ is a sequence in a general symbolic space, it is equal to the restriction of its extension.
If $ \omega \in \Sigma $ is a word in a space of infinite type, it is equal to the extension of its restriction.
\subsubsection{Extension and restriction of spaces}
For general symbolic spaces, we have the following analogue of the above notion, which we also refer to as infinite extension.
\begin{definition}[Infinite extension]
\label{infext}
Let $ \Sigma \subset E^{\ast} $ be a general symbolic space. The \textit{infinite extension} $ \Sigma^{\infty} $ is
$$
\Sigma^{\infty} = \{ \omega \in E^{\infty} : \omega |_n \in \Sigma_n \text{ for all } n \in \mathbb{N} \}.
$$
\end{definition}
Thus the infinite extension $ \Sigma^{\infty} $ of a general symbolic space $ \Sigma $ consists of the infinite words whose finite truncations lie in $ \Sigma $.
Notice that $ \Sigma^{\infty} $ satisfies the restriction admissibility property because $ \Sigma $ is assumed to satisfy the extension admissibility property, so $ \Sigma^{\infty} $ is in fact a space of infinite type.
Similarly, we obtain a general space from a space of infinite type by \textit{finite restriction}.
\begin{definition}[Finite restriction]
\label{finrest}
Let $ \Sigma \subset E^{\infty} $ be symbolic space of infinite type. The \textit{finite restriction} $ \Sigma^{\ast} $ is
$$
\Sigma^{\ast} = \bigcup_{n \geq 1} \Sigma_n
$$
\end{definition}
Thus the finite restriction $ \Sigma^{\ast} $ of a space of infinite type $ \Sigma $ consists of all the finite truncations of words in $ \Sigma $.
Notice that $ \Sigma^{\ast} $ satisfies the extension admissibility property because $ \Sigma $ is assumed to satisfy the restriction admissibility property, so $ \Sigma^{\ast} $ is in fact general symbolic space.
As with words and sequences, extension and restriction are naturally dual to each other.
If $ \Sigma $ is a general symbolic space then $ (\Sigma^{\infty})^{\ast} = \Sigma $.
If $ \Sigma $ is a symbolic space of infinite type then $ (\Sigma^{\ast})^{\infty} = \Sigma $.
\subsection{Graph directed symbolic spaces}
Let $ (V,E) $ be a directed graph with countable vertex and edge sets $ V $ and $ E $.
For each edge $ e \in E $ let $ i(e) $ and $ t(e) \in V $ be its initial and terminal vertex, respectively.
Let $ A: E \times E \rightarrow \{0,1\} $ be the edge incidence matrix of this directed graph, i.e. if $ A_{ee'} = 1 $ then $ t(e)=i(e') $.
For $ n \geq 1 $, the admissible words of length $ n $ are
\begin{equation}
\label{admis}
E^n_A = \{ \omega \in E^n : A_{\omega_i \omega_{i+1}} = 1 \text{ for all } 1 \leq i \leq n-1 \}.
\end{equation}
Let $ E^{\ast}_A = \bigcup_{n \geq 1} E^n_A $ be the collection of all finite admissible words, and $ E^{\infty}_A $ the one-sided infinite admissible words.
It is easy to see that $ E_A^{\ast} $ satisfies the extension admissibility property, so it is a special case of a general symbolic space.
Because $ E_A^{\infty} $ is closed, it is a special case of a symbolic space of infinite type.
The infinite extension of $ E_A^{\ast} $ is $ E_A^{\infty} $ and the finite restriction of $ E_A^{\infty} $ is $ E_A^{\ast} $.
The left shift restricts to $ \sigma: E_A^{\infty} \rightarrow E_A^{\infty} $ because the admissible words $ E_A^{\infty} $ are invariant.
\subsection{Dual symbolic spaces}
\label{dualcant}
In this chapter we will define the dual of a symbolic space (see \cite{Sul}).
Consider the case $ E = \mathbb{N} $ so that $ E^{\infty} = \prod_{i=1}^{\infty} E $.
We define the space $ \widetilde{E}^{\infty} \subset \prod_{i=-\infty}^{-1} E $ as follows.
$$
\widetilde{E}^{\infty} = \{(\ldots, \omega_2, \omega_1) : (\omega_1,\omega_2,\ldots) \in E^{\infty} \}
$$
There is a natural bijection $ E^{\infty} \rightarrow \widetilde{E}^{\infty} $ given by
$$
(\omega_1, \omega_2, \ldots,) \mapsto (\ldots, \omega_2, \omega_1)
$$
This map is an isometry in the above metric.
It is also an involution, so we say that $ \widetilde{E}^{\infty} $ is the \textit{dual space} to $ E^{\infty} $.
Similarly, we define
$$
\widetilde{E}^n = \{ (\omega_n,\ldots,\omega_1) : (\omega_1,\ldots,\omega_n) \in E^n \},
$$
and $ \widetilde{E}^{\ast} = \bigcup_{n \geq 1} \widetilde{E}^n $.
For a graph directed symbolic space $ E_A^{\infty} $ as defined in Chapter \ref{Thermo}, we have a dual $ \widetilde{E}_A^{\infty} $ defined by
$$
\widetilde{E}_A^{\infty} = \{ (\ldots, \omega_2, \omega_1) : A_{\omega_{i-1} \omega_i} = 1 \text{ for all } i \},
$$
and similarly for $ \widetilde{E}_A^n $ and $ \widetilde{E}_A^{\ast} $.
Finally, general symbolic spaces, spaces of infinite type, and their subspaces have duals defined in an analogous way.
\vfill
\eject
\section{$ C^{1+\alpha} $ function systems}
\label{Conf}
In this chapter we will present graph-directed pseudo-Markov systems, their limit sets, and some of their associated thermodynamic formalism.
This theory is parallel to that of Stratmann and Urba\'{n}ski \cite{Str}, but altered to account for the symbolic dynamics of the Kuperberg pseudogroup, which will be studied in detail in Chapter \ref{Kupmin}.
We assume that each space is a compact subinterval of $ [0,1] $ and that the maps have regularity $ C^{1+\alpha} $.
From this we will deduce the important properties of bounded variation and distortion in this context, which are analogues of the corresponding properties in the setting of the cookie-cutter Cantor sets of Sullivan \cite{Sul}, \cite{Bed1}.
We will then introduce general function systems-- a natural generalization of pseudo-Markov systems-- and their limit sets.
We conclude by presenting interlaced limit sets of two general function systems satisfying a disjointness condition.
\subsection{Graph directed pseudo-Markov systems}
\label{GDPM}
Let $ X $ be a bounded metric space.
Let $ E $ be a countable alphabet and $ A: E \times E \rightarrow \{0,1\} $ an incidence matrix determining the admissible words $ E_A^{\infty} $.
Assume that for each $ i \in E $ we have injective maps $ f_i : X \rightarrow X $ with a common Lipschitz constant $ 0 < s < 1 $.
We denote $ \Delta_i = f_i(X) $, and further assume that these images satisfy the separation condition
$$
\Delta_i \cap \Delta_j = \emptyset \; \text{ if } \; i \neq j.
$$
The following definition is given in terms of the above notation.
\begin{definition}
\label{GDPMdef}
A \textit{graph directed pseudo-Markov system}-- or pseudo-Markov system for short-- is a set
$$
\bigcup_{\substack{i,j \in E \\ A_{ij} = 1}} \{ \phi_{i,j} : \Delta_j \rightarrow X \}
$$
of injective maps satisfying the following properties.
\begin{itemize}
\item \textit{Lipschitz}: For each $ i $, the maps $ \phi_{i,j} : \Delta_j \rightarrow X $ have a common Lipschitz constant $ 0 < s < 1 $.
\vspace{0.2cm}
\item \textit{Separation}: For each $ i,j \in E $ with $ A_{ij} = 1 $ we have
$$
\phi_{i,j}(\Delta_j) \cap \phi_{i',j'}(\Delta_{j'}) = \emptyset
$$
when $ i \neq i' $ or $ j \neq j' $.
\vspace{0.2cm}
\item \textit{Graph directed property}: For all $ i,j \in E $ with $ A_{ij} = 1 $, we have
$$
\phi_{i,j}(\Delta_j) \subset \Delta_i.
$$
\vspace{0.2cm}
\end{itemize}
\end{definition}
By the graph directed property and Equation \ref{admis}, for each $ n \geq 1 $ and $ \omega \in E_A^n $ we have a map $ \phi_{\omega} : X \rightarrow X $ given by the composition
\begin{equation}
\label{comp}
\phi_{\omega} = \phi_{\omega_1,\omega_2} \circ \phi_{\omega_2,\omega_3} \circ \cdots \circ \phi_{\omega_{n-1},\omega_n} \circ f_{\omega_n}.
\end{equation}
For convenience, define
\begin{equation}
\label{comp2}
\Delta_{\omega} = \phi_{\omega}(X).
\end{equation}
In this notation, we deduce the \textit{nesting property} $ \Delta_{\omega,i} \subset \Delta_{\omega} $ for all $ \omega \in E_A^{\ast} $ and $ i \in E $ such that $ (\omega,i) \in E_A^{\ast} $.
Since each map $ \phi_{\omega_i,\omega_{i+1}} $ and $ f_i $ has Lipschitz constant $ 0 < s < 1 $, we have for each $ n \geq 1 $ that
$$
\text{diam}\left(\Delta_{\omega|_n}\right) \leq s^n \: \text{diam}(X).
$$
From the nesting property we see $ \Delta_{\omega|_n} \supset \Delta_{\omega|_{n+1}} $.
By this and the above equation, $ \bigcap_{n=1}^{\infty} \Delta_{\omega|_n} $ is necessarily a singleton.
This defines a bijective coding map $ \pi : E_A^{\infty} \rightarrow X $ given by
$$
\pi(\omega) = \bigcap_{n=1}^{\infty} \phi_{\omega |_n}(X) .
$$
The \textit{limit set} $ J $ of the pseudo-Markov system $ \{ \phi_{i,j} \} $ is
\begin{align}
\label{J}
J &= \pi(E_A^{\infty}) \\
&= \bigcup_{\omega \in E_A^{\infty}} \bigcap_{n=1}^{\infty} \Delta_{\omega|_n} \nonumber \\
&= \bigcap_{n=1}^{\infty} \bigcup_{\omega \in E_A^n} \Delta_{\omega}. \nonumber
\end{align}
Note: the above description of $ J $ is only true when the pseudo-Markov system is of \textit{finite multiplicity}, which is a consequence of our separation condition.
For a definition of this term and details, see Lemma 3.2 of \cite{Str}.
\subsection{Topological pressure}
\subsubsection{Pressure of continuous potentials}
Fix an alphabet $ E $ and incidence matrix $ A $, and let $ f : E_A^{\infty} \rightarrow \mathbb{R} $ be a continuous function; we will refer to such as a \textit{potential}.
For any $ n \geq 1 $, denote by $ S_n f : E^n_A \rightarrow \mathbb{R} $ the sum
$$
S_n f(\omega) = \sup_{\tau \in [\omega]} \sum_{j=0}^{n-1} f(\sigma^j \tau),
$$
and from this we form the $ n $th \textit{partition function}
$$
Z_n(f) = \sum_{\omega \in E_A^n} \exp S_n f(\omega).
$$
From the cocycle relation $ S_{m+n}f(\omega) = S_m f(\omega) + S_n f(\sigma^m \omega) $ we deduce that $ Z_{m+n}(f) \leq Z_n(f) Z_m(f) $ and so the following limit exists, which we call the \textit{topological pressure} of the potential $ f $
$$
P(f) = \lim_{n \to \infty} \frac{1}{n} \log Z_n(f).
$$
There is a natural generalization of this notion, to families of potentials.
\subsubsection{Pressure of summable H\"{o}lder families of potentials}
We use the notation
$$
F = \{ g_i : X \rightarrow \mathbb{R}, h_{i,j} : \Delta_j \rightarrow \mathbb{R} \}
$$
to denote a family of H\"{o}lder continuous functions of the same H\"{o}lder order.
Also assume that $ F $ satisfies the summability conditions
$$
\sum_{i \in E} \left\| e^{g_i} \right\| < \infty, \; \text{ and } \; \sum_{\substack{i,j \in E \\ A_{ij}=1}} \left\| e^{h_{i,j}} \right\| < \infty.
$$
We refer to such a family as a \textit{summable H\"{o}lder family}.
For any $ n \geq 1 $, word $ \omega \in E_A^n $, and summable H\"{o}lder family $ F $, denote by $ S_n F(\omega) : X \rightarrow \mathbb{R} $ the function
$$
S_n F(\omega) = \sum_{j=1}^n h_{\omega_j, \omega_{j+1}} \circ \phi_{\sigma^j \omega} + g_{\omega_n}.
$$
Similar to above, the following cocycle relation holds:
\begin{align*}
S_{m+n} F(\omega) &= \sum_{j=1}^{n+m} h_{\omega_j, \omega_{j+1}} \circ \phi_{\sigma^j \omega} + g_{\omega_{n+m}} \\
&= \sum_{j=1}^m h_{\omega_j, \omega_{j+1}} \circ \phi_{\sigma^j \omega} + \sum_{j=m+1}^{m+n} h_{\omega_j, \omega_{j+1}} \circ \phi_{\sigma^j \omega} + g_{\omega_{n+m}} \\
&= \sum_{j=1}^m h_{\omega_j, \omega_{j+1}} \circ \phi_{\sigma^j \omega} + \sum_{j=1}^n h_{\omega_{j+m}, \omega_{j+m+1}} \circ \phi_{\sigma^{j+m} \omega} + g_{\omega_{n+m}} \\
&= S_m F(\omega) + S_n F(\sigma^m \omega).
\end{align*}
This implies that the following limit exists:
\begin{equation}
\label{fampres}
P(F) = \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in E_A^n} \left\| \exp S_n F(\omega) \right\|.
\end{equation}
This is called the \textit{topological pressure} of the family $ F $.
\subsection{$ C^{1+\alpha} $ graph directed systems in dimension one}
The pseudo-Markov formalism outlined above is very general.
To apply this formalism to the Kuperberg minimal set, we will make the following assumptions on $ X $, the images $ \Delta_i = f_i(X) $, and maps $ \phi_{i,j} : \Delta_j \rightarrow X $.
\subsubsection{Dimension one.}
From now on we assume that $ X $ is an interval in $ [0,1] $, and that each $ \Delta_i $ is a closed subinterval.
Let $ | \cdot | $ be usual distance on $ [0,1] $, and set $ |U| = \text{diam}(U) $ when $ U \subset [0,1] $.
For any function $ f: X \rightarrow X $ or $ X \rightarrow \mathbb{R} $, we denote its uniform norm in this distance by
$$
\| f \|_{\infty} = \sup_{x \in X} | f(x) |.
$$
From the condition $ \lim_{n \to \infty} |\Delta_{\omega|_n}| = 0 $ for all $ \omega \in E_A^{\infty} $ we see that the limit set $ J $ from Equation \ref{J} is perfect.
From the separation condition on pseudo-Markov systems, $ J $ is totally disconnected.
By these facts and our above assumption on $ X $ and $ \Delta_i $, we see that $ J $ is a Cantor set in the line.
See Figure \ref{GDMSLimit} for a picture of a limit set of pseudo-Markov system in the line satisfying these conditions.
\begin{figure}[h]
\includegraphics[width=1.7\linewidth, trim={6.2cm 19.5cm 0 3cm}, clip]{PicTemplate9};
\caption{The first two steps of the recursive construction of $ J $ in the notation of Equation \ref{J}. The alphabet is $ E=\{1,2,3,\ldots\} $, and the incidence matrix is $ A_{ij} = 1 $ for all $ i,j \in E $. Note the separation condition $ \Delta_i \cap \Delta_j = \emptyset $ and nesting property $ \Delta_{\omega, i} \subset \Delta_{\omega} $.}
\label{GDMSLimit}
\end{figure}
\subsubsection{$C^{1+\alpha} $ regularity}
In general, to develop thermodynamic formalism we need a conformality condition.
Since we are assuming $ \Delta_i \subset X \subset [0,1] $, this can be replaced by the weaker condition of $ C^{1+\alpha} $ regularity.
\begin{definition}
\label{conf}
A pseudo-Markov system $ \{ \phi_{i,j} : \Delta_j \rightarrow X \} $ is said to be $ C^{1+\alpha} $ if there exists an $ \alpha > 0 $ such that
\begin{itemize}
\item for all $ i \in E $, the map $ f_i: X \rightarrow X $ defining $ \Delta_i $ has regularity $ C^{1+\alpha} $.
\item For all $ i,j \in E $ such that $ A_{ij}=1 $, the map $ \phi_{i,j}: \Delta_j \rightarrow X $ has regularity $ C^{1+\alpha} $.
\end{itemize}
\end{definition}
A pseudo-Markov system satisfying this assumption is referred to as a $ C^{1+\alpha} $ pseudo-Markov system.
Henceforth we will assume this regularity.
The following lemmas are standard in one-dimensional dynamics (see \cite{Shu}, \cite{Bed1}, or the appendix to \cite{Sul}).
Our proofs are based on their analogues for iterated function systems.
\begin{lemma}[Bounded variation]
\label{BV}
Let $ F= \{ g_i, h_{i,j} \} $ be a summable H\"{o}lder family of potentials.
Then there exists a constant $ M>0 $ such that for any $ n \geq 1 $ and all $ \omega \in E_A^n $ we have
$$
\left|S_n F(\omega) (x) - S_n F(\omega)(y) \right| < M
$$
for all $ x,y \in X $.
\end{lemma}
\begin{proof}
Let $ \alpha>0 $ be the H\"{o}lder order of each $ g_i $ and $ h_{i,j} $.
Since these maps have Lipschitz constant $ 0 < s < 1 $, we know for all $ x,y \in X $ that
$$
|\phi_{\omega}(x) - \phi_{\omega}(y)| \leq s^{|\omega|} |X|.
$$
By this and the H\"{o}lder continuity of each potential we have
\begin{align*}
\left|S_n F(\omega) (x) - S_n F(\omega)(y) \right| &\leq \sum_{j=1}^{n-1} \left| h_{\omega_j, \omega_{j+1}} \left( \phi_{\omega_{j+1},\ldots,\omega_n}(x)\right) - h_{\omega_j, \omega_{j+1}} \left( \phi_{\omega_{j+1},\ldots,\omega_n}(y)\right) \right| \\
&\qquad \qquad \qquad \qquad \qquad +|g_{\omega_n}(x)-g_{\omega_n}(y)| \\
&\leq C \sum_{j=1}^{n-1} \left| \phi_{\omega_{j+1}, \ldots, \omega_n}(x) - \phi_{\omega_{j+1}, \ldots, \omega_n}(y) \right|^{\alpha} + C|x-y|^{\alpha} \\
&\leq C \sum_{j=0}^{n-1} s^{(n-j-1) \alpha} |X| \\
&< \frac{C|X|}{1-|s|^{\alpha}}.
\end{align*}
\end{proof}
For $ C^{1+\alpha} $ pseudo-Markov systems in dimension one, we obtain the important \textit{bounded distortion property} from the bounded variation property.
\begin{lemma}[Bounded distortion of derivatives]
\label{BD1}
Let $ \{ \phi_{i,j} \} $ be a $ C^{1+\alpha} $ pseudo-Markov system. Then there exists a constant $ K>1 $ such that for all $ n \geq 1 $ and $ \omega \in E_A^n $,
$$
K^{-1} < \frac{|\phi'_{\omega}(x)|}{|\phi'_{\omega}(y)|} < K
$$
for all $ x, y \in X $.
\end{lemma}
\begin{proof}
Consider the family $ F = \{ g_i, h_{i,j} \} $, where
$$
g_i(x) = \log | f'_i(x) |, \; \text{ and } \; h_{i,j}(x) = \log | \phi'_{i,j}(x) |.
$$
By our $ C^{1+\alpha} $ assumption in Definition \ref{conf}, each $ f_i $ and $ \phi'_{i,j} $ is H\"{o}lder continuous on a compact set and bounded away from zero, so $ F $ is a H\"{o}lder family.
Note that the summability conditions on $ F $ are
$$
\sum_{i \in E} \| f'_i \| < \infty, \; \text{ and } \; \sum_{\substack{i,j \in E \\ A_{ij}=1}} \| \phi'_{i,j} \| < \infty.
$$
The first is a consequence of the mean value theorem and the separation conditions on the images $ \Delta_i = f_i(X) $.
The second is a consequence of that, together with the nesting property $ \Delta_{i,j} \subset \Delta_i $ when $ A_{ij}=1 $.
Since $ F $ is a summable H\"{o}lder family, we may apply Lemma \ref{BV} to say there exists a constant $ M>0 $ such that for all $ n \geq 1 $ and all $ \omega \in E_A^n $ we have $ | S_n F(\omega)(x) - S_n F(\omega)(y) | < M $ for all $ x,y \in X $.
For our choice of $ F $, by the chain rule we have
$$
S_n F(\omega)(x) = \sum_{j=1}^{n-1} \log \left| \phi'_{\omega_j,\omega_{j+1}}(\phi_{\omega_{j+1}, \ldots, \omega_n}(x)) \right| + \log \left| f'_{\omega_n}(x) \right| = \log |\phi'_{\omega}(x)|,
$$
so the conclusion of Lemma \ref{BV} states that
$$
e^{-M} < \frac{|\phi'_{\omega}(x)|}{|\phi'_{\omega}(y)|} < e^M.
$$
Let $ K = e^M > 1 $.
\end{proof}
From the bounded distortion of derivatives and the mean value theorem, we obtain bounded distortion of the intervals $ \Delta_{\omega} $.
\begin{lemma}[Bounded distortion of intervals]
\label{BD2}
Let $ K \geq 1 $ be the constant defined in Lemma \ref{BD1}.
Then for all $ n \geq 1 $ and $ \omega \in E_A^n $ we have
$$
K^{-1}|X| < \frac{|\Delta_{\omega}|}{|\phi'_{\omega}(x)|} < K |X|
$$
for all $ x \in X $.
\end{lemma}
\begin{proof}
By the mean value theorem applied to $ \phi_{\omega}: X \rightarrow X $ we have
$$
\inf_{x \in X} |\phi'_{\omega}(x)| \leq \frac{|\Delta_{\omega}|}{|X|} \leq \sup_{x \in X} |\phi'_{\omega}(x)|.
$$
Let $ x^-, x^+ \in X $ be the points on which $ \phi'_{\omega} $ takes its infimum and supremum respectively, and let $ x \in X $ be arbitrary.
By Lemma \ref{BD1} and the above inequality,
$$
K^{-1} |\phi'_{\omega}(x)| < |\phi'_{\omega}(x^-)| \leq \frac{|\Delta_{\omega}|}{|X|} \leq |\phi'_{\omega}(x^+)| < K |\phi'_{\omega}(x)|.
$$
\end{proof}
\subsection{Asymptotically stationary pseudo-Markov systems}
\label{statmark}
In the last chapter, we showed that pseudo-Markov systems with regularity $ C^{1+\alpha} $ have bounds on the distortion of their derivatives and intervals.
In this chapter, we will introduce a simpler class of pseudo-Markov systems with zero distortion, called stationary systems.
Then we will introduce asymptotically stationary systems, a simple generalization of these.
\begin{definition}[Ratio geometry]
Let $ \{ \phi_{i,j} \} $ be a pseudo-Markov system.
For each $ i \in E $ let $ R_i : E_A^{\ast} \rightarrow \mathbb{R}_{\geq 0} $ be given by
$$
R_i(\omega) = \frac{|\Delta_{\omega,i}|}{|\Delta_{\omega}|}.
$$
The function $ E_A^{\ast} \rightarrow \mathbb{R}_{\geq 0}^{\mathbb{N}} $ defined by $ \omega \mapsto \{ R_i(\omega) \}_{i \in E} $ is called the \textit{ratio geometry} of the pseudo-Markov system.
\end{definition}
The simplest pseudo-Markov systems are those whose ratio geometry is constant.
Following Pesin and Weiss (see \cite{Pes1}, \cite{Pes2}, \cite{Bar2}) we refer to such systems as \textit{stationary}.
\begin{definition}
\label{statdef}
Let $ \{ \phi_{i,j} \} $ be a pseudo-Markov system with ratio geometry $ R_i $.
Suppose that there exist positive real constants $ \{ r_i \}_{i \in E} $ such that for all $ \omega \in E_A^{\ast} $ with $ |\omega| > 1 $, we have
$$
R_i(\omega) = r_i.
$$
Such a pseudo-Markov system is called \textit{stationary}, and the numbers $ \{r_i\}_{i \in E} $ are called the \textit{ratio coefficients} of the system.
\end{definition}
For example, consider a pseudo-Markov system for which $ f_i $ and $ \phi_{i,j} $ are similarities for all $ i,j \in E $ (i.e. $ f'_i $ and $ \phi'_{i,j} $ are everywhere constant); this is a stationary system.
For each $ i \in E $ let $ s_i = |\Delta_i| $.
Then for each $ \omega \in E_A^n $, by Equations \ref{comp} and \ref{comp2}, the lengths of the intervals $ \Delta_{\omega} $ of a stationary pseudo-Markov system are simply a product of the ratio coefficients.
\begin{equation}
\label{ratcoef}
|\Delta_{\omega}| = s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n}
\end{equation}
In Chapter \ref{Dime} we will see that stationary systems have a particularly simple dimension theory, in terms of their ratio coefficients.
We now introduce a class of pseudo-Markov systems whose ratio geometry differs from that of a stationary system by some explicit error functions.
\begin{definition}
\label{asympstat1}
Let $ \{ \phi_{i,j} \} $ be a pseudo-Markov system.
Suppose that there exist positive real constants $ \{ r_i \}_{i \in E} $ and functions $ a^{\pm}: E_A^{\ast} \rightarrow \mathbb{R}_{\geq 0} $ such that for all $ n \geq 1 $ and $ \omega \in E_A^n $,
\begin{equation}
\label{asympratcoef}
s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n} - a^-(\omega) < |\Delta_{\omega}| < s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n} + a^+(\omega)
\end{equation}
Such a pseudo-Markov system is called \emph{asymtotically stationary with error $ a^{\pm} $}.
\end{definition}
To relate these systems to their simpler stationary counterparts, it is necessary to impose some conditions on the error functions $ a^{\pm} $.
With these conditions, we will see later that the dimension theory of limit sets of asymptotically stationary systems can also be analyzed using their ratio coefficients.
\begin{itemize}
\item \textit{Summability}: Assume for all $ n \geq 1 $ that
$$
\sum_{\omega \in E_A^n} a^{\pm}(\omega) < \infty.
$$
\vspace{0.1cm}
\item \textit{Monotonicity}: Assume that the error functions $ a^{\pm} $ depend on an external parameter $ \delta \in \mathbb{R}_{\geq 0} $-- which we notate as $ a^{\pm} = a_{\delta}^{\pm} $-- such that the following holds.
$$
\lim_{\delta \to 0} a_{\delta}^{\pm} = 0.
$$
\vspace{0.1cm}
\end{itemize}
Henceforth when referring to an asymptotically stationary pseudo-Markov system with summable monotone error, we mean a system in the sense of Definition \ref{asympstat1} satisfying these two properties.
\subsection{General function systems}
\label{genfun}
We will now present general function systems and their limit sets.
These are generalizations of graph-directed systems, and their dynamics are not necessarily conjugate to a shift.
Let $ E $ be a countable alphabet and let $ \Sigma \subset E^{\infty} $ be a symbolic space of infinite type as defined in Chapter \ref{Thermo}.
This implies that $ \Sigma_n \neq \emptyset $ for all $ n \geq 1 $.
Let $ X $ be a bounded metric space, and for each $ i \in E $ assume that there exist injective maps $ f_i : X \rightarrow X $ with a common Lipschitz constant $ 0 < s < 1 $.
We denote $ \Delta_i = f_i(X) $ and assume the separation condition
$$
\Delta_i \cap \Delta_j = \emptyset \; \text{ when } \; i \neq j.
$$
In terms of this notation, we give the following definition.
\begin{definition}
\label{genfundef}
A \textit{general function system modeled by $ \Sigma $} is a set
$$
\{ \phi_{i,j} : \Delta_j \rightarrow X \}_{(i,j) \in \Sigma_2}
$$
of injective maps satisfying the following properties.
\begin{itemize}
\item \textit{Lipschitz}: For each $ (i,j) \in \Sigma_2 $, the maps
$$
\{ \phi_{i,j} : \Delta_j \rightarrow X \}
$$
have a common Lipschitz constant $ 0 < s < 1 $.
\vspace{0.2cm}
\item \textit{Separation}: For each $ (i,j) \in \Sigma_2 $ we have
$$
\phi_{i,j}(\Delta_j) \cap \phi_{i',j'}(\Delta_{j'}) = \emptyset
$$
when $ i \neq i' $ or $ j \neq j' $.
\vspace{0.2cm}
\item \textit{Nesting property}: For all $ n \geq 1 $ and $ \omega \in \Sigma_n $ we have
$$
\phi_{\omega_i, \omega_{i+1}}(\Delta_{\omega_{i+1}}) \subset \Delta_{\omega_i}
$$
for all $ 1 \leq i \leq n-1 $.
\vspace{0.2cm}
\end{itemize}
\end{definition}
By the nesting property, for any $ n \geq 1 $ and $ \omega \in \Sigma_n $ we have a map $ \phi_{\omega} : X \rightarrow X $ given by the composition
$$
\phi_{\omega} = \phi_{\omega_1,\omega_2} \circ \phi_{\omega_2,\omega_3} \circ \cdots \circ \phi_{\omega_{n-1},\omega_n} \circ f_{\omega_n}.
$$
Setting $ \Delta_{\omega} = \phi_{\omega}(X) $, we have the following consequence of the nesting property.
$$
\Delta_{\omega,i} \subset \Delta_{\omega}, \; \text{ and } \; \Delta_{\omega,i} \cap \Delta_{\omega,j} \neq \emptyset
$$
for all $ \omega \in \Sigma \cap E^{\ast} $ and $ i \neq j \in E $ such that $ (\omega,i) $ and $ (\omega,j) \in \Sigma \cap E^{\ast} $.
Because the maps $ \phi_{i,j} $ have global Lipschitz constant $ 0 < s < 1 $, we have for each $ n \geq 1 $ that
$$
\text{diam} \left( \Delta_{\omega |_n} \right) \leq s^n \: \text{diam}(X).
$$
As with the graph-directed systems, the compact sets $ \Delta_{\omega|_n} $ are nested, so $ \bigcap_{n=1}^{\infty} \Delta_{\omega|_n} $ is necessarily a singleton and nonempty by our assumption on $ \Sigma $.
This defines a bijective coding map $ \pi : \Sigma \rightarrow X $ given by
$$
\pi(\omega) = \bigcap_{n=1}^{\infty} \Delta_{\omega |_n}.
$$
The \textit{limit set} $ J $ of the general function system $ \{ \phi_{i,j} \} $ is
\begin{align}
\label{J}
J &= \pi(\Sigma) \\
&= \bigcup_{\omega \in \Sigma} \bigcap_{n=1}^{\infty} \Delta_{\omega|_n} \nonumber \\
&= \bigcap_{n=1}^{\infty} \bigcup_{\omega \in \Sigma_n} \Delta_{\omega}. \nonumber
\end{align}
As with graph-directed systems, for our applications we will only consider the case when $ X \subset [0,1] $ is compact and each $ \Delta_i \subset X $ is a closed subinterval.
We will impose the same regularity conditions on general function systems as we did on graph-directed systems.
Namely, we assume that there exists $ \alpha > 0 $ such that the maps $ f_i : X \rightarrow X $ and $ \phi_{i,j} : \Delta_j \rightarrow X $ have regularity $ C^{1+\alpha} $.
We call such a function system a \textit{$ C^{1+\alpha} $ general function system modeled by $ \Sigma $}.
If $ A : E \times E \rightarrow \{0,1\} $ is an incidence matrix, the space of admissible words $ E_A^{\infty} $ defined in Chapter \ref{Thermo} is a symbolic space of infinite type, so for the choice $ \Sigma = E_A^{\infty} $, the general function system is a graph directed pseudo-Markov system as in Chapter \ref{GDPM}.
\subsection{Interlaced limit sets}
\label{sectioninter}
Suppose we have general function systems modeled by two disjoint copies of the same symbolic space, with a mutual disjointness condition on their images.
These two systems can naturally combined to create a function system modeled by a ``joint" sequence space.
The limit set of this function system is said to be the \textit{interlacing} of the limit sets of the two original systems.
In this chapter we will give a precise definition of these terms in the context of limit sets of the $ C^{1+\alpha} $ general function systems from Chapter \ref{genfun}, and then the special case of pseudo-Markov systems from Chapter \ref{GDPM}.
\subsubsection{Interlaced limit sets of general function systems}
Let $ E $ be a countable alphabet, and $ \Sigma \subset E^{\infty} $ a symbolic space of infinite type.
Let $ X \subset [0,1] $ be compact, and consider two $ C^{1+\alpha} $ general function systems $ \{\phi_{i,j} : X_j \rightarrow X\} $ and $ \{ \psi_{i,j} : Y_j \rightarrow X \} $, modeled by $ \Sigma $.
To distinguish between the maps in the two function system, define $ E $ and $ E' $ to be disjoint copies of $ E $, define $ \Sigma \subset E^{\infty} $ and $ \Sigma' \subset {E'}^{\infty} $ two disjoint copies of the same symbolic space, and say that $ \{\phi_{i,j} : X_j \rightarrow X\} $ and $ \{ \psi_{i,j} : Y_j \rightarrow X \} $ are modeled by $ \Sigma $ and $ \Sigma' $, respectively.
Separation conditions on $ X_i = f_i(X) $ and $ Y_j = g_j(X) $ are implicit in the definition presented in Chapter \ref{genfun}.
Assume further that $ X_i $ and $ Y_j $ satisfy the \textit{joint} separation property
$$
X_i \cap Y_j = \emptyset \; \text{ when } \; i \in E \text{ and } j \in F.
$$
For each $ n \geq 1 $ and $ \omega \in \Sigma_n $, $ \tau \in \Sigma'_n $ we have composition maps $ \phi_{\omega}, \psi_{\tau} : X \rightarrow X $ with images $ X_{\omega} = \phi_{\omega}(X) $ and $ Y_{\tau} = \phi_{\tau}(Y) $.
The nesting property satisfied by each function system, together with this joint separation condition, ensures that
$$
X_{\omega} \cap Y_{\tau} = \emptyset \; \text{ for all } \; \omega \in \Sigma_n \text{ and } \tau \in \Sigma'_n
$$
for all $ n \geq 1 $.
These two function systems have Cantor limit sets $ J_{\Sigma} $, $ J_{\Sigma'} $, respectively.
See Figure \ref{JEJF} for a picture of two such limit sets.
\begin{figure}[h]
\minipage{0.51\textwidth}
\includegraphics[width=1.6\linewidth, trim={7cm 20.8cm 3cm 2cm}, clip]{PicTemplate10}
\caption*{}
\endminipage
\minipage{0.51\textwidth}%
\includegraphics[width=1.6\linewidth, trim={7cm 20.8cm 3cm 2cm}, clip]{PicTemplate11}
\caption*{}
\endminipage
\caption{Two Cantor limit sets $ J_{\Sigma} $ and $ J_{\Xi} $ of general function systems in the line, satisfying the joint separation condition.}
\label{JEJF}
\end{figure}
Let $ \Sigma \ast \Sigma' \subset (E \cup E')^{\infty} $ be the set of all infinite words on the alphabet $ E \cup E' $ comprised of admissible subwords of $ \Sigma $ and $ \Sigma' $.
This is called the \textit{joint} sequence space of $ \Sigma $ and $ \Sigma' $.
From the general function systems $ \{ \phi_{i,j} \} $ and $ \{ \psi_{i,j} \} $ modeled by $ \Sigma $ and $ \Sigma' $, we will now construct a general function system modeled by $ \Sigma \ast \Sigma' $.
For $ i \in E $ and $ j \in E' $, assume we have an extension
$$
\widetilde{\psi}_{i,j} : Y_j \rightarrow X \; \text{ satisfying } \; \widetilde{\psi}_{i,j}(Y_j) \subset X_i.
$$
Similarly, for $ i \in E' $ and $ j \in E $ assume an extension
$$
\widetilde{\phi}_{i,j} : X_j \rightarrow X \; \text{ satisfying } \; \widetilde{\phi}_{i,j}(X_j) \subset Y_i.
$$
Now consider the function system
$$
\{ \gamma_{i,j} : Z_j \rightarrow X\}_{(i,j) \in (\Sigma \ast \Sigma')_2}
$$
modeled by $ \Sigma \ast \Sigma' $, where
$$
Z_j =
\left\{
\begin{array}{lr}
X_j & : j \in E \\
Y_j & : j \in E'
\end{array}
\right.
$$
and
$$
\gamma_{i,j} =
\left\{
\begin{array}{lr}
\phi_{i,j} & : i,j \in E \\
\widetilde{\phi}_{i,j} & : i \in E', j \in E \\
\widetilde{\psi}_{i,j} & : i \in E, j \in E' \\
\psi_{i,j} & : i,j \in E'
\end{array}
\right.
$$
Then for any $ \omega \in (\Sigma \ast \Sigma')_n $ we have a composition map $ \gamma_{\omega} : X \rightarrow X $ given by
$$
\gamma_{\omega} = \gamma_{\omega_1, \omega_2} \circ \gamma_{\omega_2, \omega_3} \circ \cdots \circ \gamma_{\omega_{n-1},\omega_n} \circ h_{\omega_n},
$$
where $ h_{\omega_n} = f_{\omega_n} $ if $ \omega_n \in E $, and $ h_{\omega_n} = g_{\omega_n} $ if $ \omega_n \in E' $.
As in Chapter \ref{genfun}, we let $ \Delta_{\omega} = \gamma_{\omega}(X) $ so that the Cantor limit set of $ \{ \gamma_{i,j} \} $ is
$$
J_{\Sigma \ast \Sigma'} = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in (\Sigma \ast \Sigma')_n} \Delta_{\omega}.
$$
We say the Cantor set $ J_{\Sigma \ast \Sigma'} $ is the \textit{interlacing} of the Cantor sets $ J_{\Sigma} $ and $ J_{\Sigma'} $.
See Figure \ref{interlacefig}.
\begin{figure}[h]
\includegraphics[width=1.5\linewidth, trim={5.5cm 20.5cm 0cm 2.5cm}, clip]{PicTemplate12};
\caption{The interlacing $ J_{\Sigma \ast \Xi} $ of the Cantor sets $ J_{\Sigma} $ and $ J_{\Xi} $ from Figure \ref{JEJF}.}
\label{interlacefig}
\end{figure}
\subsubsection{Interlaced limit sets of pseudo-Markov systems}
If we have incidence matrices $ A^E : E \times E \rightarrow \{0,1\} $ and $ A^{E'} : E' \times E' \rightarrow \{0,1\} $ such that $ E_{A^E}^n = \Sigma_n $ and $ {E'}_{A^{E'}}^n = \Sigma'_n $ for all $ n \geq 1 $, the above general function systems are graph directed pseudo-Markov systems as studied in Chapter \ref{GDPM}.
Then the joint sequence space $ \Sigma \ast \Sigma' $ defined above is $ (E \cup E')_{A^{E \cup E'}}^{\infty} $, where $ A^{E \cup E'} : (E \cup E') \times (E \cup E') \rightarrow \{0,1\} $ is the joint incidence matrix given by $ A^{E \cup E'}(i,j) = A^E(i,j) $, i.e. the joint words are admissible according to $ E $.
For each $ n \geq 1 $, consider the intervals $ X_{\omega} = \phi_{\omega}(X) $ where $ \omega \in E_{A^E}^n $, and $ Y_{\tau} = \psi_{\tau}(X) $ where $ \tau \in {E'}_{A^{E'}}^n $.
Then by Equation \ref{J} we have the following descriptions of the limit sets of the respective pseudo-Markov systems.
$$
J_E = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in E_{A^E}^n} X_{\omega}, \; \text{ and } \; J_{E'} = \bigcap_{n=1}^{\infty} \bigcup_{\tau \in {E'}_{A^{E'}}^n} Y_{\omega}.
$$
The interlacing $ J_{E \cup E'} $ of $ J_E $ and $ J_{E'} $ is the limit set of the joint pseudo-Markov system, and is given by
$$
J_{E \cup E'} = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in (E \cup E')_{A^{E \cup E'}}^n} \Delta_{\omega},
$$
where $ \Delta_{\omega} = \gamma_{\omega}(X) $ and $ \gamma_{\omega} $ is the composition of the maps $ \phi_{i,j} $ and $ \psi_{i,j} $ indexed by admissible words $ \omega $ in the joint sequence space $ (E \cup E')_{A^{E \cup E'}}^n $.
Each point in $ J_{E \cup E'} $ corresponds to a unique word in $ (E \cup E')^{\infty}_{A^{E \cup E'}} $.
\vfill
\eject
\section{Dimension theory of limit sets}
\label{Dime}
The Hausdorff dimension of a limit set is related to the pressure by Bowen's equation.
In regularity $ C^{1+\alpha} $, the pressure has uniformity properties that can be deduced from the bounded variation and distortion properties in Lemmas \ref{BV} and \ref{BD2}.
We present these properties for pseudo-Markov systems and then state Bowen's equation in this context.
We then apply this to the dimension theory of the asymptotically stationary pseudo-Markov systems of Chapter \ref{statmark}.
\subsection{The pressure function}
Let $ E $ be a countable alphabet, $ A $ an incidence matrix, and $ \{ \phi_{i,j} : \Delta_j \rightarrow X \} $ a $ C^{1+\alpha} $ pseudo-Markov system as in Chapter \ref{GDPM}.
For any $ t \in (0,\infty) $ consider the family $ F_t = \{ g_i, h_{i,j} \} $, where
$$
g_i(t) = t \log |f'_i(x)|, \; \text{ and } \; h_{i,j}(x) = t \log |\phi'_{i,j}(x)|.
$$
This is a summable H\"{o}lder family of potentials as defined in Chapter \ref{Conf}, and as such has a well-defined topological pressure $ P(F_t) $.
We define $ p(t) = P(F_t) $ and call $ p $ the \textit{pressure function} determined by the system $ \{\phi_{i,j}\} $.
From the proof of Lemma \ref{BD1}, for all $ \omega \in E_A^n $ we have
$$
S_n F_t(\omega)(x) = t \log | \phi'_{\omega}(x) |.
$$
Substituting this into Equation \ref{fampres} we obtain
\begin{equation}
\label{pressurefunction}
p(t) = \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in E_A^n} \| \phi'_{\omega} \|^t.
\end{equation}
Notice that $ p = \lim_{n \to \infty} \frac{1}{n} \log p_n $, where
$$
p_n(t) = \sum_{\omega \in E_A^n} \| \phi'_{\omega} \|^t.
$$
Because $ p_{m+n}(t) \leq p_m(t) p_n(t) $ for all $ t \in [0,\infty) $, we have that $ p_n(t) < \infty $ if and only if $ p_1(t) < \infty $.
Let $ \theta = \inf \{ t: p(t) < \infty \} $, so that the set of finiteness of $ p $ is $ (\theta, \infty) $.
A summary of the properties of $ p $ are collected below.
\begin{proposition}[Proposition 4.10 from \cite{Str}]
\label{presprop}
The topological pressure function $ p(t) $ is non-increasing on $ [0,\infty) $, and is continuous, strictly decreasing, and convex on $ (\theta, \infty) $.
\end{proposition}
The definition of topological pressure given in Equation \ref{pressurefunction} can be difficult to use in practice.
Fortunately, the assumption of $ C^{1+\alpha} $ regularity and its consequences yields a more useful definition.
Applying Proposition \ref{BD1},
$$
p(t) = \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in E_A^n} \left|\phi'_{\omega}(x)\right|^t
$$
for any $ x \in X $.
By Proposition \ref{BD2},
\begin{equation}
\label{pressurefunction2}
p(t) = \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in E_A^n} |\Delta_{\omega} |^t.
\end{equation}
\subsection{Bowen's equation for pressure}
A generalization of Bowen's equation (\cite{Bow1}) is proved in \cite{Str} for what are termed ``weakly thin" pseudo-Markov systems.
Weak thinness is a general notion, but in our setting it is equivalent to $ p_1(1) = \sum_{i \in E} |\Delta_i| < \infty $, which is a consequence of the separation and compactness conditions from Chapter \ref{Conf}.
\begin{theorem}[Proposition 4.13 of \cite{Str}]
\label{Bow}
Let $ \{ \phi_{i,j} \} $ be a $ C^{1+\alpha} $ pseudo-Markov system with limit set $ J $ and associated pressure function $ p(t) $.
Then the Hausdorff dimension $ \text{dim}_H(J) $ satisfies
$$
\text{dim}_H(J) = \inf \{ t \geq 0 : p(t) < 0 \},
$$
and if $ p(t) = 0 $ then $ t $ is the only zero of $ p(t) $ and $ t = \text{dim}_H(J) $.
\end{theorem}
\subsection{Dimension of limit sets of asymptotically stationary pseudo-Markov systems}
In Chapter \ref{statmark} we introduced the asymptotically stationary pseudo-Markov systems, with error $ a_{\delta}^{\pm} $.
We assume that this error is summable and monotone, as specified in that chapter.
The dimension theory of stationary systems is particularly simple and goes back to Moran (\cite{Mor}).
The dimension theory of asymptotically stationary systems is similar.
\begin{theorem}
\label{statGDMS}
Let $ \{ \phi_{i,j} \} $ be an asymptotically stationary pseudo-Markov system, with summable monotone error $ a_{\delta}^{\pm} $, and let $ J_{\delta} $ be its limit set. Then the Lebesgue measure of $ J_{\delta} $ satisfies
$$
\lim_{\delta \to 0} \mu(J_{\delta}) = 0,
$$
and the Hausdorff dimension $ \text{dim}_H(J_{\delta}) $ satisfies
$$
0 < \lim_{\delta \to 0} \text{dim}_H(J_{\delta}) < 1.
$$
\end{theorem}
\begin{proof}
Let $ \mu $ be Lebesgue measure on $ [0,1] $.
By the nesting and separation conditions on $ \Delta_{\omega} $,
$$
\mu(J_{\delta}) = \mu \left( \bigcap_{n=1}^{\infty} \bigcup_{\omega \in E_A^n} \Delta_{\omega} \right)
= \lim_{n \to \infty} \mu \left( \bigcup_{\omega \in E_A^n} \Delta_{\omega} \right)
= \lim_{n \to \infty} \sum_{\omega \in E_A^n} |\Delta_{\omega}|.
$$
We then substitute Equation \ref{asympratcoef} to obtain
\begin{align*}
\mu(J_{\delta}) &\leq \lim_{n \to \infty} \sum_{\omega \in E_A^n} s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n} + \lim_{n \to \infty} \sum_{\omega \in E_A^n} a_{\delta}^+(\omega) \\
&\leq \lim_{n \to \infty} \sum_{\omega \in E^n} s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n} + \lim_{n \to \infty} \sum_{\omega \in E_A^n} a_{\delta}^+(\omega) \\
&= \lim_{n \to \infty} \left( \sum_{i \in E} s_i \right) \left( \sum_{i \in E} r_i \right)^{n-1} + \lim_{n \to \infty} \sum_{\omega \in E_A^n} a_{\delta}^+(\omega)
\end{align*}
By the separation condition in Definition \ref{GDPM} we know $ \sum_{i \in E} r_i < 1 $.
By the summability and monotonicity conditions on $ a_{\delta}^+ $ the right term decreases to $ 0 $ as $ \delta \rightarrow 0 $, so $ \lim_{\delta \to 0} \mu(J_{\delta}) = 0 $, as desired.
We now turn to the Hausdorff dimension.
Let $ p_{\delta}(t) $ be the pressure function associated to this pseudo-Markov system.
Substituting Equation \ref{asympratcoef} into the pressure function in Equation \ref{pressurefunction2} we obtain that $ p_{\delta}^- < p_{\delta} < p_{\delta}^+ $, where
$$
p_{\delta}^{\pm}(t) = \lim_{n \to \infty} \frac{1}{n} \log \left( \sum_{\omega \in E_A^n} (s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n})^t \pm \sum_{\omega \in E_A^n} a_{\delta}^{\pm}(\omega) \right)
$$
Then by the monotonicity of $ a_{\delta}^{\pm} $ we have that $ \lim_{\delta \to 0} p_{\delta}^{\pm} = p $, where
$$
p(t) = \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in E_A^n} (s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n})^t
$$
For the upper bound, we calculate
\begin{align*}
p(t) &< \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in E^n} s_{\omega_1}^t r_{\omega_2}^t \cdots r_{\omega_n}^t \\
&= \lim_{n \to \infty} \frac{1}{n} \log \left( \sum_{i \in E} s_i^t \right) \left( \sum_{i \in E} r_i^t \right)^{n-1} \\
&= \log \sum_{i \in E} r_i^t.
\end{align*}
Let $ t_{\ast} $ be the unique solution to $ \sum_{i \in E} r_i^t = 1 $, and notice that $ t_{\ast} < 1 $.
Applying Bowen's theorem (\ref{Bow}), we have $ \dim_H(J) < t_{\ast} < 1 $.
For the lower bound, recall that for all $ n \geq 1 $, $ E_A^n $ contains more than one word, say $ \omega = (\omega_1, \ldots, \omega_n) $.
$$
p(t) > \lim_{n \to \infty} \frac{1}{n} \log \left(s_{\omega_1} r_{\omega_2} \cdots r_{\omega_n} \right)^t = \lim_{n \to \infty} \frac{1}{n} \left( \log s_{\omega_n}^t + \sum_{j=1}^{n-1} \log r_{\omega_j}^t \right).
$$
Setting the right hand side $ =0 $, we see that $ t_{\ast} = 0 $ is a solution.
So again by Bowen's theorem, we have $ \text{dim}_H(J) > t_{\ast} = 0 $.
\end{proof}
\vfill
\eject
\section{The Wilson flow}
\label{Wilsflow}
Wilson's flow (\cite{Wil}) is defined on a \textit{plug}, a closed manifold that traps orbits.
First we will define general plugs, and then present the construction of Wilson's plug.
Then we will introduce Wilson's vector field, and study its dynamics in the plug.
\subsection{Plugs}
\label{plugs}
Let $ M $ be a compact orientable manifold with nonempty boundary.
A \textit{plug} is a product $ M \times [0,1] $, supporting a vector field $ \mathcal{V} $ with flow $ \phi_t $.
For the plugs we consider, $ M $ will have dimension two, so $ M \times [0,1] $ is an oriented three-manifold with boundary $ \partial M \times [0,1] $. Let $ (x,z) $ be a coordinate system on $ M \times [0,1] $. We will orient the plug vertically, so that $ M \times \{0\} $ is the ``bottom" of the plug, and $ M \times \{1\} $ the ``top."
If $ (x,0) \in M \times \{0\} $ and $ (x',1) \in M \times \{1\} $ satisfy $ x=x' $, then these two points are said to be \textit{facing}.
A plug is a local dynamical system designed to be inserted into a global one.
For the plug to be inserted into a manifold with a flow there are several important assumptions it must satisfy.
These ensure that the dynamics inside the plug are compatible with the dynamics outside, and that the plug traps a set of orbits of the flow on the manifold.
\begin{itemize}
\item \textit{Matched ends property}: If a flowline of $ \phi_t $ passes through the points $ (x, 0) $ and $ (x', 1) $, then these points are facing, i.e. $ x=x' $.
\item \textit{Trapped orbit property}: There exists a flowline of $ \phi_t $ passing through $ (x,0) $ but not intersecting $ M \times \{1\} $.
\end{itemize}
If a plug satisfies the following additional symmetry condition, we call it a \textit{mirror-image plug}.
\begin{itemize}
\item \textit{Mirror-image property}: The reflection of the field $ \mathcal{V} $ over the center $ M \times \left\{\frac{1}{2}\right\} $ is the negative of $ \mathcal{V} $.
\end{itemize}
Flowlines in a mirror-image plug are symmetric over $ M \times \left\{\frac{1}{2}\right\} $.
Notice that the mirror-image property implies the matched-ends property.
The Wilson plug is a mirror-image plug with $ M $ a closed annulus. For the original description see \cite{Wil}, and for subsequent descriptions \cite{Kup1}, \cite{Ghy}, \cite{Hur}.
Our notation does not differ much from this literature's.
\subsection{The Wilson plug}
Define the closed rectangle $ E = [1,3] \times [-2,2] $ in coordinates $ (r, z) $, and the closed rectangular solid $ E \times [0,2\pi] $, in coordinate $ (r, \theta, z) $.
Denote by $ c_1, c_2 \in E $ the points $ (2,-1) $ and $ (2,+1) $, respectively.
Then $ l_i = c_i \times [0,2\pi] $ are two line segments in the rectangular solid.
Finally, define closed neighborhoods $ B_i $ of $ c_i $, so that $ B_i \times [0,2\pi] $ is a tubular neighborhood of each $ l_i $.
The Wilson plug $ W $ is the image of the region $ E \times [0,2\pi] $ under the embedding $ (r, \theta, z) \mapsto (r \cos \theta, r \sin \theta, z) $.
\begin{figure}[!h]
\includegraphics[width=\linewidth, trim={2cm 14.5cm 4cm 4cm}, clip]{WilsonRectangle};
\caption{The rectangular region $ E \times [0,2\pi] $ with the coordinates and special regions indicated}
\label{rectreg}
\end{figure}
See Figures \ref{rectreg} and \ref{plug1} for a picture of the rectangle and the embedded plug, respectively.
Notice that the lines $ l_i $ map to circles under the embedding, and the tubes $ B_i \times [0,2\pi] $ map to torii containing the corresponding circles $ l_i $.
\begin{figure}[!h]
\includegraphics[width=\linewidth, trim={2cm 13.5cm 2cm 2.6cm}, clip]{PicTemplate8};
\caption{The embedded Wilson plug. The lines $ l_i $ from Figure \ref{rectreg} map to periodic orbits of the Wilson flow.}
\label{plug1}
\end{figure}
Under this embedding, $ M=\{(r,\theta) : 2 \leq r \leq 3, 0 \leq \theta \leq 2\pi \} $ is an annulus, and $ W=M \times [-2,2] $ is a plug in the notation of Chapter \ref{plugs}.
The bottom of the plug is $ M \times \{-2\} $ and the top is $ M \times \{2\} $.
\subsection{The Wilson vector field}
For convenience, we will describe the dynamics in the coordinates $ (r, \theta, z) $ and suppress the embedding.
On $ E \times [0,2\pi] $, we define a vector field $ \mathcal{W} $.
\begin{equation}
\mathcal{W} = f \: \frac{\partial}{\partial \theta} + g \: \frac{\partial}{\partial z},
\label{W}
\end{equation}
where $ f $ and $ g $ are $ C^{\infty} $ real-valued functions of the rectangle $ E $, constructed as follows. First, fix $ a > 0 $, and define $ f : E \rightarrow \mathbb{R} $ by
\begin{equation}
f(r,z)=
\left\{
\begin{array}{lr}
a & : z < 0 \\
-a & : z \geq 0
\end{array}
\right.
\label{f}
\end{equation}
Notice that this function is not $ C^{\infty} $-- not even continuous-- but can be made so by adjusting it in an arbitrarily small neighborhood of $ \{z=0\} \subset R $.
To construct $ g $, for $ i=1,2$ let $ p_i : B_i \rightarrow [0,1] $ be $ C^{\infty} $ functions satisfying
\begin{equation}
\label{p}
p_i(c_i) = 0, \quad \quad p_i \equiv 1 \text{ on } \partial B_i, \quad \quad p_i(x) > 0 \text{ for all } x \in B_i \setminus \{c_i\}
\end{equation}
Then we define $ g : E \rightarrow [0,1] $ by
\begin{equation}
\label{g}
g(x) = \left\{
\begin{array}{lr}
p_i(x) & : x \in B_i, \; i=1, 2 \\
1 & : x \in E \setminus (B_1 \cup B_2)
\end{array}
\right.
\end{equation}
Notice that $ g \equiv 1 $ outside the regions $ B_i $. Inside each $ B_i $, $ g $ decreases smoothly to zero, reaching zero (by definition of $ p_i $) at precisely $ c_i $.
Since $ g \equiv 0 $ at the two points $ c_i \in E $, the $ z $ component of the Wilson field $ \mathcal{W} $ (equation \ref{W}) is singular on the circles $ l_i $.
The field $ \mathcal{W} $ preserves these circles, forming two periodic orbits inside the plug.
These are referred to as the \textit{special orbits}, and are illustrated in Figure \ref{plug1}.
The torii $ B_i \times [0,2\pi] $ that contain them are referred to as the \textit{critical torii}.
Finally, we define the \textit{Reeb cylinder} as $ \mathcal{R} = \{r=2\} $ and for any $ \epsilon > 0 $ we define the \textit{critical region} $ C_{\epsilon} $ as an $ \epsilon $-neighborhood of $ \mathcal{R} $-- explicitly, $ C_{\epsilon} = \{ 2 \leq r \leq 2+\epsilon \} \subset W $.
All the interesting dynamics will occur inside this critical region.
\subsection{Dynamics of the Wilson flow}
\subsubsection{Orbits of points: Helices}
Let $ \phi_t $ be the flow of $ \mathcal{W} $.
By definition of $ \mathcal{W} $, the radial coordinate of each orbit is preserved, so that flowlines are helical in shape.
At the base annulus $ \{z=-2\} $ we have $ f \equiv a $ and $ g \equiv 1 $ in equation \ref{W}, so the orbit spirals upward counter-clockwise from the base annulus to the central annulus $ \{z=0\} $.
At this point, $ f \equiv -a $, so the $ \theta $ component of the flow direction is reversed; now the orbit spirals upward clockwise until it reaches the upper annulus $ \{z=2\} $ and escapes the plug.
Since $ f $ is anti-symmetric across the line $ \{z=0\} \subset E $, flowlines are symmetric about the annulus $ \{z=0\} \subset W $.
This implies that $ W $ is a mirror-image plug. In particular, it satisfies the matched-ends property (See Chapter \ref{plugs}).
Wilson orbits that originate in the base $ \{z=-2\} $ of the plug have three orbit types, as shown in Table \ref{wilsontable}.
The third orbit type shows that $ W $ satisfies the trapped orbit property.
\begin{table}
\begin{tabular}{ m{7cm} m{11cm} }
\begin{itemize}
\item \textbf{Disjoint from critical torii}: In this case $ f \equiv \pm a $ and $ g \equiv 1 $ in equation \ref{W}, and the orbit helix spirals at a constant speed.
The orbit takes a short time to escape the plug.
\end{itemize}
&
\includegraphics[width=0.8\linewidth, trim={1cm 13cm 1cm 1cm}, clip]{MiddleWilson}
\\
\begin{itemize}
\item \textbf{Intersecting critical torii}: Inside the critical torii, $ g \equiv p_i $ which is zero at $ c_i $.
Thus the vertical component of the orbit slows dramatically inside the torii, at a speed depending on the orbit's radial proximity to the special orbit.
The orbit takes a long time to escape the plug.
\end{itemize}
&
\includegraphics[width=0.8\linewidth, trim={1cm 13cm 1cm 1cm}, clip]{OutsideWilson}
\\
\begin{itemize}
\item \textbf{In the Reeb cylinder $ \mathcal{R} $}: Upon entering the first torus, $ g \equiv p_1 $ and the orbit spirals towards the special orbit $ c_1 $.
As the orbit approaches $ c_1 $ the speed of its vertical component approaches zero.
The orbit is trapped and remains in the plug for infinite time.
\end{itemize}
&
\includegraphics[width=0.8\linewidth, trim={1cm 13cm 1cm 1cm}, clip]{WilsonCrit}
\\
\end{tabular}
\caption{Classification of Wilson orbits originating in the base $ \{z=-2\} $}
\label{wilsontable}
\end{table}
\subsubsection{Orbits of curves: Propellers}
\label{propellers}
Following \cite{Hur} we make the following definition.
\begin{figure}[h]
\minipage{0.5\textwidth}
\includegraphics[width=0.9\linewidth]{MathPropeller}
\caption*{Section of single propeller}
\endminipage
\minipage{0.5\textwidth}%
\includegraphics[width=0.9\linewidth]{MathDoublePropeller}
\caption*{Section of double propeller}
\endminipage
\caption{The intersections of a single and double propeller with a transverse section $ \{\theta= \text{constant}\}$ of the Wilson plug. The inside edge at $ r=2 $ is trapped and limits on the intersection of the special orbit $(r,z)=(2,-1)$, while the outside edge(s) at $ r>2 $ escapes.}
\label{prop}
\end{figure}
\begin{definition}[\textit{Single propellers}]
\label{propdef}
Let $ \eta : [s_1,s_2] \rightarrow W $ be a continuous curve such that the radial coordinate of $ \eta(s_1) $ is $ 2 $, and for all $ s_1 < s \leq s_2 $, the radial coordinate of $ \eta(s) $ is strictly greater than $ 2 $.
A \textit{single propeller} is $ \bigcup_{t \geq 0} \phi_t(\eta) $ for such an $ \eta $.
\end{definition}
\begin{definition}[\textit{Double propellers}]
\label{doublepropdef}
Let $ \eta : [s_1,s_2] \rightarrow W $ be a continuous curve such that there exists $ s_1 < s_c < s_2 $ with $ \eta(s_c) $ having a radial coordinate of $ 2 $, and for all $ s_1 \leq s < s_c $ and $ s_c < s \leq s_2 $ the radial coordinate of $ \eta(s) $ is strictly greater than $ 2 $.
A \textit{double propeller} is $ \bigcup_{t \geq 0} \phi_t(\eta) $ for such an $ \eta $.
\end{definition}
Notice that a single propeller can be obtained from a double propeller by restricting the parametrization of the generating curve $ \eta $.
We will see later that the minimal set of the Kuperberg flow can be decomposed into a union of single propellers, so understanding how propellers are embedded in $ W $ is the key to understanding the embedding of the minimal set.
A propeller forms a ``helical ribbon'' winding around the Wilson plug.
Its outside edge has an $ r $-coordinate bounded away from $ 2 $, so it forms a helix, the first orbit type. Its inside edge has an $ r $-coordinate of $ 2 $ and thus is trapped in the plug, the third orbit type.
Thus each propeller contains curve that is trapped for infinite time, resulting in a complicated embedding in the plug.
This complexity is illustrated in a cross-section of the Wilson plug shown in Figure \ref{prop}.
\subsection{The Wilson minimal set}
Let $ x \in W $. By Table \ref{wilsontable}, if the radial coordinate of $ x $ is $ >2 $, its orbit escapes through the top $ \{z=2\} $ in finite forward time, and escapes through the bottom $ \{z=-2\} $ in finite backward time.
If the radial coordinate of $ x $ is $ =2 $, its orbit limits on one of the special orbits $ l_i $ in forward and/or backward time, depending on its vertical position in the plug.
Thus the minimal invariant set in $ W $ is the union $ l_1 \cup l_2 $.
See Figure \ref{wilmin}.
\begin{figure}[h]
\includegraphics[width=\linewidth, trim={2cm 13.5cm 2cm 2.6cm}, clip]{WilsonMinimal}
\caption{The minimal set in the Wilson plug $ W $ is the two special orbits $ l_1 \cup l_2 $.}
\label{wilmin}
\end{figure}
\subsection{The Wilson pseudogroup}
\label{Wilpseudo}
Let $ S \subset W $ be a surface tranverse to the Wilson flow $ \phi_t $.
For our purposes, it will suffice to consider a small rectangle with a constant $ \theta $-coordinate.
Consider the first return map $ \Phi : S \rightarrow S $ of $ \phi_t $ to $ S $.
Explicitly, $ \Phi(x) = \phi_T(x) $ where $ T>0 $, $ \phi_T(x) \in S $, and $ T $ is minimal with respect to these properties.
Each such map has a natural inverse, by first-return under the backward orbit.
Notice that $ \Phi $ is not defined on all of $ S $, nor are successive compositions of $ \Phi $ necessarily defined, even where $ \Phi $ is.
Thus $ \Phi $ does not generate a group, but does generate a \textit{pseudogroup} (see \cite{Hur0}, \cite{Wal}) of local homeomorphisms.
This is the holonomy pseudogroup of the foliation of $ W $ by flowlines of $ \phi_t $.
\vfill
\eject
\section{The Kuperberg flow}
\label{Kupsection}
The Kuperberg plug is constructed by performing two operations of \textit{self-insertion} on the Wilson plug.
We will summarize this below, but the construction is delicate and we refer to \cite{Kup1} for the details.
\subsection{Kuperberg's construction and theorem}
\label{KupThm}
First we define two closed disjoint regions $ L_1, L_2 \subset W $, intersecting the outside boundary $ \{r=3\} $ of the plug, the top and bottom of the plug, and the two special orbits.
For $ i=1,2 $ we denote by $ L_i^+ $ the intersection of these regions with the top of the plug, and by $ L_i^- $ the bottom.
We then re-embed the Wilson plug in $ \mathbb{R}^3 $ in a folded figure-eight.
See Figure \ref{kup1}.
\begin{figure}[htp]
\subfloat{
\includegraphics[width=0.8\linewidth, trim={4.5cm 16.6cm 0cm 4.5cm}, clip]{PicTemplate13}
}
\subfloat{
\includegraphics[width=0.8\linewidth, trim={4.5cm 16.6cm 0cm 4.5cm}, clip]{PicTemplate15}
}
\vspace{0.4cm}
\caption{The re-embedded Wilson plug with the closed regions $ L_1 $ and $ L_2 $, and the quotient Kuperberg plug $ K = \widehat{W}/\sim $}
\label{kup1}
\end{figure}
Now for each $ i = 1,2 $ we define diffeomorphisms $ \sigma_i : L_i \rightarrow W $, called \textit{insertion maps}.
Denote $ D_i = \sigma_i(L_i) \subset W $, and let $ D_i^{\pm} = \sigma_i(L_i^{\pm}) $.
We make several assumptions about the images $ D_i $.
\begin{itemize}
\item We choose each $ D_i $ to intersect a short segment of the special orbit $ l_i $.
\item The neighborhoods $ D_i $ intersect the inside boundary $ \{r=1\} $ of the plug.
\item The regions $ L_i $ are ``twisted" under $ \sigma_i $ so that special orbits $ l_i $ \textit{enter} through $ D_i^- $ and \textit{exit} through $ D_i^+ $.
\item There is a single angle $ \alpha_i \in [0,2\pi] $ such that the vertical arc $ \{ r=2, \theta=\alpha_i, -2 \leq z \leq 2\} \subset \mathcal{R} \cap L_i $ maps onto the horizontal special orbit segment $ D_i \cap l_i $.
\end{itemize}
We will use the insertion maps to define a new plug as follows.
First we remove the images $ D_i $ of the insertion maps from $ W $, denoting $ \widehat{W} = W \setminus (D_1 \cup D_2) $.
Then, we define an equivalence relation $ \sim $ on $ \widehat{W} $ by setting $ x \sim y $ if $ x $ lies in either $ L_i^+ \cup L_i^- $ or the outside boundary $ L_i \cap \{r=3\} $, and $ y $ lies in the images of these regions under $ \sigma_i $, for both $ i =1,2 $.
The \textit{Kuperberg plug} $ K $ is the quotient $ \widehat{W}/\sim $, a manifold with boundary (See Figure \ref{kup2}).
Let $ \tau: \widehat{W} \rightarrow K $ be the quotient map.
\begin{figure}[h]
\includegraphics[width=\linewidth, trim={4cm 17cm 0cm 4cm}, clip]{PicTemplate14}
\caption{The self-insertions defining the Kuperberg plug. The special orbits enter the bottom faces $ D_i = \sigma_i(L_i) $ where $ L_i $ are shown in Figure \ref{kup1}}
\label{kup2}
\end{figure}
The set $ \mathcal{R} \cap \{ |z| \leq 1 \} $ is the sub-cylinder of the Reeb cylinder $ \mathcal{R} $ lying between the two special orbits.
Let $ \mathcal{R}' $ be the closure of $ \mathcal{R} \cap \{ |z| \leq 1 \} \setminus \widehat{W} $.
This is the sub-cylinder with the two ``notches" $ L_i \cap \mathcal{R} $ removed.
We refer to $ \mathcal{R}' $ as the \textit{notched Reeb cylinder}.
Now, for each $ i =1,2 $, we define a rectangular region $ S_i \subset D_i^- $.
We will assume that the the radial coordinate of the inner edge of each $ S_i $ is constant $ =2 $.
Thus $ S_i \cap \mathcal{R} $ is a vertical line segment, which we denote by $ \gamma_i $, and $ S_i \cap \mathcal{R}' $ is the upper half of $ \gamma_i $, which we denote by $ \gamma_i^u $.
Further, each rectangle $ S_i $ is foliated by vertical line segments $ \{\gamma_{c,i}\}_c $, where $ \gamma_{0,i} = \gamma_i $.
We will write each $ S_i $, $ \gamma_i $, and $ \gamma_i^u $ in coordinates in Chapter \ref{InsertAssume}.
For now, we need only specify that each $ S_i $ intersects the special orbit $ l_i $, which is consistent with Kuperberg's construction outlined above.
Using this notation, there are two important assumptions we must make about the insertions $ \sigma_i $ defining $ K $.
The first is important for proving that the dynamics inside $ K $ are aperiodic.
The second will prove to be crucial for determining properties of the minimal set.
\begin{itemize}
\item \textbf{Radius Inequality}: For $ i=1,2 $, the radial coordinate of each point in $ L_i $ is strictly greater than that of its image under $ \sigma_i $, with one exception.
That is, for points in the inverse image $ \{ r=2, \theta=\alpha_i, -2 \leq z \leq 2\} $ under $ \sigma_i $ of the special orbit $ c_i $, where the radial coordinates agree.
\item \textbf{Quadratic Insertion}: For $ i=1,2 $, the inverse image under $ \sigma_i $ of $ \gamma_i $ is a parabola with vertex $ (2,\alpha_i, -2) $.
Furthermore, the inverse image under $ \sigma_i $ of the rectangular region $ S_i $ is a ``parabolic strip" with vertex $ (2,\alpha_i, -2) $.
More precisely, the inverse image under $ \sigma_i $ of each vertical line segment $\gamma_{c,i} $ in the vertical foliation of $ S_i $ is a parabola with vertex $ (2+c, \alpha_i, -2) $.
\end{itemize}
See Figure \ref{kup2} for an illustration of the quadratic insertion property.
\begin{figure}[htp]
\subfloat{
\includegraphics[width=0.7\linewidth, trim={5.2cm 18.2cm 8cm 4cm}, clip]{PicTemplate16}
}
\vspace{-0.4cm}
\subfloat{
\includegraphics[width=0.8\linewidth, trim={5cm 18cm 6cm 4.5cm}, clip]{PicTemplate17}
}
\vspace{-0.5cm}
\caption{The quadratic insertion property}
\label{kup2}
\end{figure}
If a closed manifold carries the dynamics of a smooth vector field, we may insert a plug-- supporting a separate smooth vector field-- into the interior of this manifold.
Assume that the plug has the \textit{matched ends property}, and that the ends of the plug are transverse to the field on the manifold.
Then the theory of plugs and insertions developed in \cite{Wil}, \cite{Sch}, \cite{Kup1} and \cite{Kup2} show that a smooth global field on the plugged manifold, compatible with the dynamics of both the manifold and the plug, can be defined by smoothly altering the dynamics in a tubular neighborhood of the boundary of the plug.
The construction is delicate and we refer to \cite{Kup1} for the details.
By these facts, the Wilson field $ \mathcal{W} $ induces a smooth vector field $ \mathcal{K} $ on the Kuperberg plug, which we call the \textit{Kuperberg field}.
Kuperberg proved that the self-insertions defining $ K $ break the periodic orbits $ l_i $, without creating new periodic orbits.
\begin{theorem} \emph{(Theorem 4.4 from \cite{Kup1})}
\label{Kuptheorem}
The $ C^{\infty} $ vector field $ \mathcal{K} $ defined on $ K $ has no closed orbits.
\end{theorem}
Kuperberg's theorem is true under very flexible assumptions; in fact, the proof uses only the radius inequality and does not require the quadratic insertion property.
However, to determine finer aspects of the dynamics of the Kuperberg flow on its minimal set, we will need to make several more assumptions.
\subsection{Further insertion assumptions}
\label{InsertAssume}
In this chapter, we will impose more restrictive versions of the assumptions we have already made, to obtain explicit formulas for the insertion maps $ \sigma_i $ and the Wilson flow $ \phi_t $.
To simplify the exposition, we will write these formulas only for $ \sigma_1 $, the lower insertion map.
In the following chapter, we denote by $ \sigma $, $ D^- $, $ B $, $ p $, $ \gamma $, $ \alpha $, $ S $, $ \gamma_{c} $ and $ l $ the quantities $ \sigma_1 $, $ D_1^- $, $ B_1 $, $ p_1 $, $ \gamma_1 $, $ \alpha_1 $, $ S_1 $, $ \gamma_{c, 1} $ and $ l_1 $ respectively.
Identical assumptions will be made (but not written down) for the upper insertion $ \sigma_2 $.
\subsubsection{Rectangular intersection}
First, we assume that the rectangular region $ S $ has a constant angular coordinate $ \theta = \beta $, width $ 0 < b < 1 $, and height $ 2R $ for some $ R>0 $.
Explicitly,
\begin{equation}
\label{Sstrip}
S = \{ (r,\beta, z) : 0 \leq r-2 \leq b, |z+1| \leq R \}.
\end{equation}
The upper and lower boundaries of this rectangle are
\begin{equation}
\label{Spm}
S^{\pm} = \{ (r,\beta, -1 \pm R) : 0 \leq r-2 \leq b \}
\end{equation}
Both intervals $ S^{\pm} $ can be identified with $ [0,b] $ and will be used extensively later when describing the transverse minimal set.
The inner edge of this rectangle is the intersection $ S \cap \mathcal{R} $, the vertical line $ \gamma $ we defined earlier:
\begin{equation}
\label{gammadef}
\gamma = \{ (2,\beta,z) : |z+1|\leq R \}
\end{equation}
Also, we define $ \gamma^u $ and $ \gamma^l $ to be the upper and lower half of $ \gamma $, so $ \gamma = \gamma^u \cup \gamma^l $.
By definition of $ \mathcal{R}' $, we have $ \mathcal{R}' \cap S = \gamma^u $.
See Figure \ref{rectfig}.
\begin{align}
\label{gammauldef}
\gamma^u &= \{ (2,\beta,z) : 0 \leq z+1 \leq R\} \\
\gamma^l &= \{ (2,\beta,z) : -R \leq z+1 \leq 0\} \nonumber
\end{align}
\begin{figure}[h]
\includegraphics[width=1.2\linewidth, trim={4.7cm 15.7cm 1cm 3.1cm}, clip]{PicTemplate34}
\caption{The rectangle $ S $, with the inner edge $ \gamma = \gamma^u \cup \gamma^l $ and the upper and lower boundaries $ S^{\pm} $.}
\label{rectfig}
\end{figure}
Additionally, we assume that $ (B \times [0,2\pi]) \cap (S \times [0,2\pi]) = (S \times [0,2\pi])$.
Recall that the vertical component $ g $ (defined in Equations \ref{W} and \ref{g}) of the Wilson flow changes from $ g=1 $ to $ g=p $ precisely at $ \partial B $.
This assumption will simplify the boundary conditions that arise when integrating $ \mathcal{W} $, since the upper and lower boundaries of the critical torus $ B \times [0,2\pi] $ must now coincide with the two annuli
$$
C^{\pm} = \{(r, \theta, -1 \pm R): 0 \leq r-2 \leq b, \; 0 \leq \theta \leq 2\pi \}.
$$
The intersection of the annuli $ C^{\pm} $ with $ S $ are the upper and lower boundary intervals $ S^{\pm} $ of the strip $ S $.
\subsubsection{Quadratic decay}
Recall the function $ p $ defined in Equation \ref{p}.
We now assume that $ p $ decays quadratically inside the critical strip $ S $.
\begin{equation}
\label{quadraticp}
p |_S(r,z) = \frac{1}{R^2}((r-2)^2+(z+1)^2)
\end{equation}
By the rectangular intersection assumption, this is compatible with the boundary condition $ p = 1 $ on $ \partial B $ from Equation \ref{p}.
\subsubsection{Quadratic insertion formula}
Recall the quadratic insertion assumption made in Chapter \ref{KupThm}. In this chapter, we will make these assumptions more specific; in particular we will write the inverse of the insertion map $ \sigma $ in coordinates.
By equation \ref{Sstrip}, any point in the rectangle $ S $ can be written as $ (2+r, \beta, -1+z) $, where $ 0 \leq r \leq b $ and $ -R \leq z \leq R $.
We will assume that $ \sigma^{-1} $ takes $ S $ to a parabolic strip in the base $ {z=-2} $, its vertex having a constant $ \theta $ coordinate of $ \alpha $, in the following way:
\begin{equation}
\label{sigmainverse}
\sigma^{-1}(2+r, \beta, -1+z) = (2+r+z^2, \alpha-z, -2)
\end{equation}
See Figure \ref{kup2}.
In light of Equation \ref{gammadef}, we can parametrize $ \gamma $ as
\begin{equation}
\label{gamma}
\gamma: [-R,R] \rightarrow S \quad \quad \gamma(s) = (2, \beta, -1+s), \\
\end{equation}
and by Equation \ref{gammauldef}, $ \gamma^u $ and $ \gamma^l $ are parametrized as $ \gamma^u = \gamma |_{[0,R]} $ and $ \gamma^l = \gamma |_{[-R,0]} $.
Referring to equation \ref{sigmainverse}, we can parametrize parabolic the curve $ \sigma^{-1} \gamma $ as follows.
\begin{equation}
\label{sigmagamma}
\sigma^{-1}\gamma(s) = (2+s^2, s+\alpha, -2)
\end{equation}
Observe that $ S = \bigcup_{0 \leq c \leq b} \gamma_c $, where
$$
\gamma_c = \{ (2+c, \beta, z) : |z+1| \leq R \}.
$$
The collection $ \{ \gamma_c \}_{0 \leq c \leq b} $ is the foliation of $ S $ by vertical lines, introduced in the statement of the quadratic insertion property from Chapter \ref{KupThm}.
We parametrize each vertical line $ \gamma_c $ as follows.
\begin{equation}
\label{gammac}
\gamma_c: [-R,R] \rightarrow S \quad \quad \gamma_c(s) = (2+c,\beta,-1+s)
\end{equation}
Equation \ref{sigmainverse} implies that for each $ c \in [0,b] $, the curve $ \sigma^{-1} \gamma_c $ is parabolic in the base $ \{z=-2\} $ of the plug, with the parametrization
\begin{equation}
\label{sigmagammac}
\sigma^{-1}\gamma_c(s) = (2+c+s^2, s+\alpha, -2).
\end{equation}
Since $ \gamma_0 = \gamma $, this parametrization is compatible with the above parametrization of $ \gamma $.
\subsection{Integrals of $ \mathcal{W} $}
Our quadratic decay assumption allows us to integrate $ \mathcal{W} $ explicitly.
At points $ (r, \theta, -2) \in \{z=-2\} $, the Wilson vector field $ \mathcal{W} $ has $ f\equiv +a $ and $ g \equiv 1 $, resulting in the simple expression
\begin{equation}
\label{wilout}
\phi_t(r, \theta, z) = (r, \theta+at, z+t) \text{ when } 0 \leq z(t) \leq -1-R
\end{equation}
A flowline looks like the first case in Table \ref{wilsontable}, a helix rising with constant vertical speed $ \frac{2\pi}{a} $.
The upper bound on $ z $ in Equation \ref{wilout} is the point at which the orbit intersects the lower annulus $ C^- $.
At this point, we have $ g = p $ by our rectangular intersection assumption, and use Equation \ref{quadraticp} to integrate $ \mathcal{W} $.
\begin{multline}
\label{wilin}
\phi_t(r, \theta, z) = \left( r, \theta+At, -1+(r-2) \tan \left( \frac{r-2}{R^2}t+\tan^{-1}\left(\frac{z+1}{r-2}\right)\right)\right) \\
\text{ when } |z(t)+1| \leq R.
\end{multline}
In this region, a flowline looks like the second case in Table \ref{wilsontable}, a helix rising at a variable speed depending on its radial proximity to the Reeb cylinder $ \{ r=2 \} $ and its vertical proximity to $ z=-1 $.
\subsection{Transition and level}
\label{levo}
Let $ \psi_t $ be the flow of the Kuperberg vector field.
Flowlines of $ \psi_t $ are very complicated and do not admit a classification as simple as those of the Wilson flow given in Table \ref{wilsontable}.
However, since the $ K $ is a quotient of $ W $, the dynamics of $ \psi_t $ resemble the dynamics of $ \phi_t $.
To see this resemblance, we begin by embedding $ K $ in $ \mathbb{R}^3 $ as we did $ W $ in Figure \ref{plug1}, suppressing the more complicated embedding as in Figure \ref{kup1}, but retaining the interior self-insertions defining $ K $.
See Figure \ref{kupwil} for this embedding.
\begin{figure}[h]
\includegraphics[width=1.1\linewidth, trim={3cm 13.5cm 2cm 2.6cm}, clip]{PicTemplate18}
\caption{The Kuperberg plug embedded as the Wilson plug.}
\label{kupwil}
\end{figure}
Each orbit of the Kuperberg flow $ \psi_t $ contains \textit{transition points}.
These are intersections of the orbit with an insertion region.
Between these transition points, the flowline coincides with one of the flowlines of the Wilson flow $ \phi_t $.
The hierarchy of \textit{levels} will be used to keep track of these transition points.
By studying levels and the dynamics of the Wilson flow, we can understand the dynamics of the Kuperberg flow.
\subsubsection{Transition points and the level function for orbits}
\begin{definition}[\textit{Orbit segments and orbits}]
\label{orbpoint}
For any $ x \in K $, we denote its \textit{closed orbit segment} for time $ t_2-t_1>0 $ by
$$
\mathcal{O}[x,t_1, t_2] = \bigcup_{t_1 \leq t \leq t_2} \psi_t(x).
$$
Its \textit{open orbit segment} is
$$
\mathcal{O}(x,t_1, t_2) = \bigcup_{t_1 < t < t_2} \psi_t(x),
$$
and its \textit{half-open orbit segment} is
$$
\mathcal{O}(x,t_1, t_2] = \bigcup_{t_1 < t \leq t_2} \psi_t(x).
$$
Its \textit{orbit} $ \mathcal{O}(x) $, \textit{forward orbit} $ \mathcal{O}^+(x) $, and \textit{backward orbit} $ \mathcal{O}^-(x) $ are
$$
\mathcal{O}(x) = \bigcup_{-\infty < t < \infty} \psi_t(x), \qquad
\mathcal{O}^+(x) = \bigcup_{t \geq 0} \psi_t(x), \qquad
\mathcal{O}^-(x) = \bigcup_{t \leq 0} \psi_t(x).
$$
\end{definition}
Depending on the location of $ x $ in the plug, its orbit $ \mathcal{O}(x) $ may be finite or infinite (see Table \ref{wilsontable}).
An orbit's intersection with the bottom $ \{z=-2\} $, the top $ \{z=+2\} $, or either of the four insertion faces $ D_i^{\pm} $ ($i=1,2$), is called a \textit{transition point}.
There are four types of transition points.
\begin{itemize}
\item \textit{primary entry points} are transition points in $ \{z=-2\} $.
\item \textit{primary exit points} are transition points in $ \{z=2\} $.
\item \textit{secondary entry points} are transition points in $ D_i^+ $ for $ i=1,2 $.
\item \textit{secondary exit points} are transition points in $ D_i^- $ for $ i=1,2 $.
\end{itemize}
For each $ x \in K $, there is a natural orbit decomposition
\begin{equation}
\label{orbitdecomp}
\mathcal{O}(x) = \bigcup_{i \in I} \mathcal{O}(x, t_i, t_{i+1}],
\end{equation}
into disjoint half-open orbit segments, where for all $ i \in I $, $ \psi_{t_i}(x) $ is a transition point and $ \mathcal{O}(x, t_i, t_{i+1}) $ contains no interior transition points.
The indexing set $ I $ is countable if $ x $ has an infinite orbit, and is finite if the orbit is.
The \textit{level function} $ n_x(t) $ along the orbit of $ x $ indexes how many insertions an orbit has passed through at time $ t $, measured from zero.
\begin{definition}[\textit{Level function along orbits}]
\label{leveldef}
Let $ x \in K $, let $ n_x^+(t) $ be the number of secondary entry points in $ \mathcal{O}(x,0,t] $, and let $ n_x^-(t) $ be the number of secondary exit points in $ \mathcal{O}(x,0,t] $.
Define the \textit{level function} $ n_x : \mathcal{O}(x) \rightarrow \mathbb{N} $ by $ n_x(\psi_t(x)) = n_x^+(t) - n_x^-(t) $.
\end{definition}
For a fixed $ x \in K $, we say that $ y \in \mathcal{O}(x) $ \textit{has level k} if $ y=\psi_T(x) $ with $ n_x(T) = k $.
The following lemma appears in \cite{Ghy} (Lemme, pg. 300) and is formulated more precisely in Lemma 6.5 of \cite{Hur}; the only secondary entrance points that are trapped have a radial coordinate $=2$; the rest escape the insertion in finite time.
\begin{lemma}
\label{escapeorbit}
Suppose $ x $ has a radial coordinate $>2 $, and the orbit $ \mathcal{O}(x) $ contains a secondary entrance point $ \psi_T(x) $ for some $ T>0 $. Then there exists $ S>T $ such that $ \psi_S(x) $ is a secondary exit point, $ \psi_T(x) $ and $ \psi_S(x) $ are facing, and $ n_x(T) = n_x(S) $.
\end{lemma}
The next lemma appears in various forms in the literature (Proposition 4.1 of \cite{Kup1}, Lemma 5.1 of \cite{Hur}, and Lemma 7.1 of \cite{Ghy}) and is crucial in relating orbits of the Kuperberg flow to orbits of the Wilson flow.
Recall that $ \tau: \widehat{W} \rightarrow K $ is the quotient map defining the Kuperberg plug.
\begin{lemma}[\textit{short-cut lemma}]
\label{shortcutorbit}
Suppose that a secondary entrance point $ x_- \in D_i^- $ and a secondary exit point $ x_+ \in D_i^+ $ are facing.
Then there exists a point $ y_- $ in the base $ \{z=-2\} \subset W $ and $ y^+ $ in the top $ \{z=2\} \subset W $ of the Wilson plug such that $ \tau(y_{\pm}) = x_{\pm} $, and a finite time $ T>0 $ such that $ y_+ = \phi_T(y_-) $.
\end{lemma}
In this way, the dynamics of a Kuperberg orbit segment between secondary entrance and exit points reduces to the dynamics of a finite Wilson orbit from the base to the top of the plug.
Finally, for orbits of curves we have an analogous definition to that of Definition \ref{orbpoint}.
\begin{definition}[\textit{Orbit strips and surfaces}]
\label{orbsurf}
For any $ \eta $ be a curve with image in $ K $.
Its \textit{closed orbit strip} for time $ t_2-t_1>0 $ is
$$
\mathcal{O}[\eta,t_1, t_2] = \bigcup_{t_1 \leq t \leq t_2} \psi_t(\eta).
$$
Its \textit{open orbit strip} is
$$
\mathcal{O}(\eta,t_1, t_2) = \bigcup_{t_1 < t < t_2} \psi_t(\eta),
$$
and its \textit{half-open orbit strip} is
$$
\mathcal{O}(\eta,t_1, t_2] = \bigcup_{t_1 < t \leq t_2} \psi_t(\eta).
$$
Its \textit{orbit surface} $ \mathcal{O}(\eta) $, \textit{forward orbit surface} $ \mathcal{O}^+(\eta) $, and \textit{backward orbit surface} $ \mathcal{O}^-(\eta) $ are
$$
\mathcal{O}(\eta) = \bigcup_{-\infty < t < \infty} \psi_t(\eta), \qquad
\mathcal{O}^+(\eta) = \bigcup_{t \geq 0} \psi_t(\eta), \qquad
\mathcal{O}^-(\eta) = \bigcup_{t \leq 0} \psi_t(\eta).
$$
\end{definition}
As we will see in Section \ref{Kupmin}, the minimal set of the Kuperberg flow is the closure of a union of \textit{propellers}, and each propeller is an orbit surface in this sense.
\vfill
\eject
\section{The Kuperberg pseudogroup}
\label{Kuppseudo}
In Chapter \ref{Wilpseudo} we introduced the Wilson pseudogroup generated by $ \Phi $, the first-return map of the Wilson flow to a tranverse section.
In this chapter, we will study the Kuperberg pseudogroup $ \Psi $, defined in the same way using the Kuperberg flow.
The domains of the generators of the pseudogroup we define will be subsets of the two transverse rectangles $ S_i $, $ i=1,2 $, defined in Chapter \ref{InsertAssume}.
\begin{figure}[h!]
\includegraphics[width=1.1\linewidth, trim={3cm 13.5cm 2cm 2.6cm}, clip]{PicTemplate35}
\caption{The rectangular regions $ S_i \subset D_i^- $ and their inverse images $ \sigma^{-1}(S_i) \subset \{z=-2\} $. Compare with Figures \ref{kup2} and \ref{kupwil}.}
\label{kup3}
\end{figure}
In Chapter \ref{Kupsection}, we showed that every orbit decomposes into segments whose endpoints are transition points, having no interior transition points.
At a secondary transition point, the orbit intersects an insertion region $ D_i^{\pm} $, which is identified via $ \sigma^{-1} $ with $ L_i^{\pm} \subset \{ z = \pm 2 \} $ in the base or the top of the plug.
The dynamics of the orbit changes drastically at transition points, and these dynamics are determined by $ \sigma^{-1} $.
The interior of the orbit segment follows the helical Wilson flow $ \phi_t $ studied in Chapter \ref{Wilsflow}.
The transverse rectangles $ S_i $ lie in $ D_i^- $, so the Kuperberg first-return of a point to $ S_i $ is a secondary entrance point, by definition.
This first-return map follows the Wilson flow.
At the transition point, it is mapped via $ \sigma^{-1} $ into a parabolic strip in the base of the plug, which then follows the Wilson flow up to more intersections with $ S_i $.
So the Kuperberg pseudogroup $ \Psi $ of first-return maps to the rectangles $ S_1 \cup S_2 $ is generated by the Wilson pseudogroup to $ S_i $ from $ S_i $ or the base, and the insertion maps from $ S_i $ to the base, for $ i=1,2 $.
In this section, we will construct these generators for $ \Psi $.
In Chapter 9 of \cite{Hur}, the full Kuperberg pseudogroup to a larger transverse section was studied.
This pseudogroup is very complicated, and in subsequent chapters of \cite{Hur} its properties were used to study the dynamics of the Kuperberg flow on the entire plug $ K $.
In this paper we are concerned only with the dynamics of the Kuperberg flow in small neighborhoods of the special orbits $ l_i $, which is why we choose the sections $ S_i $.
The pseudogroup $ \Psi $ we consider is a restriction of the full pseudogroup studied in \cite{Hur}.
In the second part of this chapter, we explore the symbolic dynamics of the $ \Psi $ on an orbit.
For simplicity, we focus on the lower rectangle $ S_1 $, by considering a suitable sub-pseudogroup of $ \Psi $.
For any $ x \in K $, the intersection $ \mathcal{O}(x) \cap S $
is a sequence of points ordered by the flow direction, on which the Kuperberg pseudogroup acts faithfully.
Using the notion of \textit{level} introduced Chapter \ref{Kupsection}, we will decompose this intersection into level sets, and show that the pseudogroup generators permute this level decomposition.
Finally, we will construct a sequence space $ \Sigma \subset \mathbb{N}^{\mathbb{N}} $ and a bijective coding map $ \Sigma \rightarrow \mathcal{O}(x) \cap S $, and study the induced dynamics of the pseudogroup on this space.
This is the symbolic dynamics of the Kuperberg pseudogroup, which will be instrumental later when studying the minimal set.
\subsection{Generators of the pseudogroup}
Recall the rectangular regions $ S_i \subset D_i^- $ defined in Equation \ref{Sstrip} of Chapter \ref{InsertAssume}.
In the quotient $ K $, these regions are identified with the parabolic regions $ \sigma^{-1}(S_i) \subset \{z=-2\} $ in the base of the plug.
See Figure \ref{kup3}.
We now list the generators of the Kuperberg pseudogroup restricted to the rectangles $ S_1 \cup S_2 $.
\begin{figure}[h!]
\includegraphics[width=1.1\linewidth, trim={3cm 13.5cm 2cm 2.6cm}, clip]{PicTemplate36}
\caption{The map $ \Phi_1 $ follows the Wilson flow $ \phi_t $ through the lower insertion region $ D_1 $ and around to its first return to $ S_1 $.
There is a similar picture for $ \Phi_2 $.}
\label{kup4}
\end{figure}
\subsubsection{The Wilson maps $ \Phi_i : D_{\Phi_i} \rightarrow R_{\Phi_i} $}
Consider a point $ x \in S_i \subset D_i^- $ for $ i = 1,2 $.
We assume that $ x $ is not the intersection point of the special orbit $ l_i $ with $ S_i $, i.e. $ x \neq (2, \beta_i, \pm 1) $.
We define $ \Phi_i(x) $ as the first return to $ S_i $ under the Wilson flow $ \phi_t $.
Explicitly, $ \Phi_i(x) = \phi_T(x) $,
where $ T>0 $, $ \phi_T(x) \in S_i $, and $ T $ is minimal with respect to these properties.
In the Kuperberg plug, $ x $ is identified with $ \sigma^{-1}(x) $ in the base.
By the assumption that $ x $ is not the intersection point of $ l_i $ with $ S_i $, we know by the radius inequality that $ \sigma^{-1}(x) $ has radius $ >2 $.
Applying the short-cut lemma (Lemma \ref{shortcutorbit}), there exists a facing point $ x' \in D_i^+ $, and the flow from $ x $ to $ x' $ is a finite union of Wilson flow segments.
From $ x' $, the orbit follows the Wilson flow around the plug and back to $ \Phi_i(x) $, its first-return to $ S_i $.
See Figure \ref{kup4} for an illustration of $ \Phi_1 $.
\begin{figure}[h!]
\includegraphics[width=1.2\linewidth, trim={4.7cm 13.5cm 1cm 3.1cm}, clip]{PicTemplate37}
\caption{The image $ R_{\Phi_i} = \Phi_i(D_{\Phi_i}) $ is the orange region. The points $ x \in S_i \setminus D_{\Phi_i} $ are near the top of $ S_i $; the Wilson flow of these points does not return to $ S_i $.}
\label{kup5}
\end{figure}
Note that $ \Phi_i(x) $ is not defined for all $ x \in S_i $.
This is because the Wilson flow of a point has a monotonically increasing $ z $-coordinate, so there are points near the top of $ S_i $ that never return to $ S_i $ under $ \Phi_i $.
However, there is a subset $ D_{\Phi_i} \subset S_i $ of points $ x $ for which $ \Phi_i(x) $ is defined.
Denote the image by $ R_{\Phi_i} = \Phi_i(D_{\Phi_i}) \subset S_i $ (see Figure \ref{kup5}).
\begin{figure}[h!]
\includegraphics[width=1.1\linewidth, trim={3cm 13.5cm 2cm 2.6cm}, clip]{PicTemplate38}
\caption{The map $ \Phi_{1,2} $ follows the Wilson flow $ \phi_t $ from the upper region of $ S_1 $, through the insertion region $ D_1 $ and up to its first return to $ S_2 $.}
\label{kup6}
\end{figure}
\subsubsection{The Wilson map $ \Phi_{1,2} : D_{\Phi_{1,2}} \rightarrow R_{\Phi_{1,2}} $}
As discussed in the previous paragraph, there is a set of points near the upper boundary of $ S_1 $ that do not return to $ S_1 $ under the Wilson flow.
However, the Wilson flow of these points does intersect the upper rectangle $ S_2 $.
This defines a map $ \Phi_{1,2} : D_{\Phi_{1,2}} \rightarrow R_{\Phi_{1,2}} $ given by $ \Phi_{1,2}(x) = \phi_T(x) $, where $ T>0 $, $ \phi_T(x) \in S_2 $, and $ T $ is minimal with respect to these properties (See Figure \ref{kup6}).
\begin{figure}[h!]
\includegraphics[width=1.1\linewidth, trim={3cm 13.5cm 2cm 2.6cm}, clip]{PicTemplate39}
\caption{The map $ \Theta_1 $ follows the Wilson flow from $ \sigma^{-1}(S_1) $ in the base $ \{ z=-2 \} $ up to its first return to $ S_1 $.
There is a similar picture for $ \Theta_2 $.}
\label{kup7}
\end{figure}
\subsubsection{The insertion maps $ \Theta_i : D_{\Theta_i} \rightarrow R_{\Theta_i} $}
In the Kuperberg plug $ K $, the quotient map $ \tau $ identifies each rectangle $ S_i $ with the parabolic strip $ \sigma^{-1}(S_i) $ in the base (See Figure \ref{kup3}).
Thus an orbit that intersects a rectangle $ S_i $ is identified via $ \sigma^{-1} $ with the base, after which the orbit follows the Wilson flow back up to the lower rectangle $ S_1 $.
See Figure \ref{kup7} for an illustration of this.
We will now define a map $ \Theta_i : D_{\Theta_i} \rightarrow R_{\Theta_i} $, where $ D_{\Theta_i} \subset S_i $, and $ R_{\Theta_i} \subset S_1 $.
\begin{figure}[h!]
\includegraphics[width=1.2\linewidth, trim={4.7cm 15.7cm 1cm 3.1cm}, clip]{PicTemplate40}
\caption{The tip of the twisted parabolic strip $ \Theta_i(S_i) $ in $ S_i $. The vertex $ \phi_{T_k}(x_i) $ of the strip lies in the Reeb cylinder $ \{ r=2 \} $.}
\label{kup8}
\end{figure}
The vertex of a parabolic strip $ \sigma^{-1}(S_i) $ is $ x_i = (2, \alpha_i, -2) $ (See Chapter \ref{InsertAssume} and Figure \ref{kup2}).
Because the radial coordinate of $ x_i $ is $ 2 $, its Wilson orbit $ \phi_t(x_i) $ is trapped in $ K $.
The orbit $ \phi_t(\sigma^{-1}(S_i)) $ of the entire parabolic strip $ \sigma^{-1}(S_i) $ intersects the lower rectangle $ S_1 $ at a sequence of times $ T_j $, $ j=1,2,\ldots $.
These intersections are ``twisted" parabolic strips, resembling the propellers' cross-sections in Figure \ref{prop}.
The vertices of these parabolic regions are the intersections of the orbit of $ x_i $ with $ S_1 $, which is the ordered sequence of points $ \phi_{T_j}(x_i) $ limiting on the special orbit intersection, whose $ z $-coordinate monotonically increases with $ j $.
Because $ S $ contains the special orbit intersection $ (2, \beta, -1) $, there is a critical time $ T_k $ such that this sequence of points remains in $ S_i $ for all $ j \geq k $.
In other words, let $ k $ be the minimal value of $ j $ such that $ \phi_{T_j}(x_i) \in S_i $ for all $ j \geq k $.
In terms of this fixed $ k $, define
$$
\Theta_i(x) = \phi_{T_k} (\sigma^{-1}(x)).
$$
Define $ D_{\Theta_i} \subset S_i $ the set of points $ x $ for which the above equation is defined, and define the image $ R_{\Theta_i} = \Theta_i(D_{\Theta_i}) \subset S_1 $ (See Figure \ref{kup8}).
\subsection{Restriction to a sub-pseudogroup}
To summarize, we have constructed the Kuperberg pseudogroup $ \Psi $ on $ S_1 \cup S_2 $, generated by five elements:
\begin{equation}
\label{Psi}
\Psi = \langle \Phi_1, \Phi_2, \Phi_{1,2}, \Theta_1, \Theta_2 \rangle
\end{equation}
The dynamics of the full pseudogroup $ \Psi $ are complicated.
To simplify the study, we will consider the sub-pseudogroup generated by the two maps $ \Phi_1 , \Theta_1 : S_1 \rightarrow S_1 $.
To save on notation, we denote $ \Phi = \Phi_1 $ and $ \Theta = \Theta_1 $.
In terms of these, we define
\begin{equation}
\label{Psi1}
\Psi_1 = \langle \Phi, \Theta \rangle.
\end{equation}
Following the shorthand used in Section \ref{InsertAssume}, we will refer to $ S_1 $, $ D_1 $, $ \sigma_1 $, $ l_1 $, $ \gamma_1 $, $ \gamma_{c,1} $, and $ \beta_1 $ simply by $ S $, $ D $, $ \sigma $, $ l $, $ \gamma $, $ \gamma_c $, and $ \beta $, respectively.
These conventions will be observed for the remainder of this chapter, and throughout Chapters \ref{Kupmin} -- \ref{FunctionSystems}.
We will return to the dynamics of the full pseudogroup in Chapter \ref{Transcant} when we discuss \textit{interlacing}.
\subsection{Orbit intersections with a transversal}
In this and subsequent sections we will introduce the intersection of a forward Kuperberg orbit $ \mathcal{O}^+(x) $ with $ S $.
We then study its level decomposition and the symbolic dynamics of the pseudogroup action on this intersection.
Let $ x \in S $ be any point \textit{other than} the intersection $ (2,\beta,-1) $ of $ S $ with the special orbit $ l $, and consider its forward orbit $ \mathcal{O}^+(x) $ in $ K $.
In Section \ref{levo}, we introduced the level function along an orbit.
This induces a level decomposition of the intersection $ \mathcal{O}^+(x) \cap S $.
\begin{equation}
\label{orbint1}
\mathcal{O}^+(x) \cap S = \bigcup_k \mathcal{O}^+(x)_k \cap S, \; \text{ where } \; \mathcal{O}^+(x)_k \cap S = \{ y \in \mathcal{O}^+(x) \cap S : n_x(y) = k\}.
\end{equation}
\subsubsection{The pseudogroup action}
The next two lemmas show that $ \Phi $ preserves level, while $ \Theta $ increases level by one.
\begin{lemma}
\label{Phipermute}
If $ x \in S $ is not in the special orbit, the map $ \Phi $ restricted to $ \mathcal{O}^+(x) \cap S $ maps $ \mathcal{O}^+(x)_k \cap S $ into $ \mathcal{O}^+(x)_k \cap S $.
\end{lemma}
\begin{proof} Let $ y \in \mathcal{O}^+(x)_k \cap S $, so there exists $ T>0 $ such that $ y =\psi_T(x) $ and $ n_x(T) = k $.
Since $ S \subset D^- $, $ y $ is necessarily a secondary entrance point.
In the quotient plug $ K $, $ y $ is identified with $ \sigma^{-1}(y) $.
Because we assumed that $ \mathcal{O}^+(x) $ is not the special orbit, $ y $ is not in the special orbit's intersection with $ S $.
Applying the radius inequality (Section \ref{Kupsection}), we see that $ y $ has a radial coordinate $ >2 $, and so we may apply Lemma \ref{escapeorbit} to say there exists $ S > T $ with $ \psi_S(x) \in D^+ $ a secondary exit point, $ y $ and $ \psi_S(x) $ are facing, and $ n_x(T) = n_x(S) = k $.
The orbit of $ \psi_S(x) $ now follows the Wilson flow forward to its first return to $ S $ (if it exists), which by definition is $ \Phi(y) $ (if it is defined).
Because the Wilson flow preserves the level, $ \Phi(y) $ has level $ k $, so $ \Phi(y) \in \mathcal{O}^+(x)_k \cap S $, which concludes the proof.
\end{proof}
\begin{lemma}
\label{Thetapermute}
If $ x \in S $ is not in the special orbit, the map $ \Theta $ restricted to $ \mathcal{O}^+(x) \cap S $ maps $ \mathcal{O}^+(x)_k \cap S $ into $ \mathcal{O}^+(x)_{k+1} \cap S $.
\end{lemma}
\begin{proof} Let $ y \in \mathcal{O}^+(x)_k \cap S $, so there exists $ T>0 $ such that $ y =\psi_T(x) $ and $ n_x(T) = k $.
As in Lemma \ref{Phipermute}, $ y $ is a secondary entrance point.
If it exists, $ \Theta(y) $ is the first return of $ \sigma^{-1}(y) $ to $ S $.
Notice that points on the forward orbit of $ \sigma^{-1}(y) $ have level $ k $.
Because $ S \subset D^- $, $ \Theta(y) $ is a secondary entrance point and thus has level $ k+1 $.
This shows that $ \Theta(y) \in \mathcal{O}^+(x)_{k+1} \cap S $, which concludes the proof.
\end{proof}
These lemmas demonstrate that the pseudogroup $ \langle \Phi, \Theta \rangle $ acts faithfully on the intersection $ \mathcal{O}^+(x) \cap S $ by permuting the level.
\subsection{Symbolic dynamics of orbits}
\label{symborbits}
In this section, we will define a natural sequence space coding the points in an intersection $ \mathcal{O}^+(x) \cap S $.
This space will consist of finite words, whose word length is equal to the level of the corresponding point.
The action of the pseudogroup $ \langle \Phi, \Theta \rangle $ on the level decomposition of $ \mathcal{O}^+(x) \cap S $ will induce a faithful action on this sequence space.
Fix $ x \in K $ with $ \mathcal{O}^+(x) \cap S \neq \emptyset $, and let $ y \in \mathcal{O}^+(x) \cap S $ be a point of level zero, i.e. $ y = \psi_T(x) $ with $ n_x(T) = 0 $.
Then by Lemma \ref{Thetapermute}, $ \Theta(y) $ has level one.
\begin{enumerate}[(1)]
\vspace{0.1cm}
\item \textit{Points of level one}:
For $ 1 \leq i_1 \leq M(x) $, let
$$
y_{i_1} = (\Phi^{i_1-1} \Theta)(y),
$$
where $ M(x) $ is the minimum positive integer such that $ (\Phi^{M(x)} \Theta)(y) $ is not defined, i.e. $ (\Phi^{M(x)}\Theta)(y) $ does not return to $ S $ under the Kuperberg flow.
We call $ M(x) $ the \textit{escape time} of $ \Theta(y) $ from $ S $.
By Lemma \ref{Phipermute}, each point $ y_{i_1} $ has level one.
\vspace{0.2cm}
\item \textit{Points of level two}: For each $ 1 \leq i_1 \leq M(x) $, let
$$
y_{i_1,i_2} = (\Phi^{i_2-1} \Theta)(y_{i_1}),
$$
where $ 1 \leq i_2 \leq M_{i_1}(x) $ and $ M_{i_1}(x) $ is the minimum positive integer such that $ (\Phi^{M_{i_1}(x)+1} \Theta)(y_{i_1}) $ is not defined.
Note that $ M_{i_1}(x) \neq \infty $, because $ \Theta(y_{i_1}) $ has a radial coordinate $ >2 $, so the Wilson orbit of the points $ \Theta(y_{i_1}) $ escape in finite time.
By Lemmas \ref{Phipermute} and \ref{Thetapermute}, each point $ y_{i_1, i_2} $ has level two.
\vspace{0.2cm}
\item \textit{Points of level k}: For each $ 1 \leq i_{k-1} \leq M_{i_1,\ldots,i_{k-1}}(x) $, let
$$
y_{i_1,\ldots,i_k} = (\Phi^{i_k-1} \Theta) (y_{i_1,\ldots,i_{k-1}}),
$$
where $ 1 \leq i_n \leq M_{i_1,\ldots,i_{k-1}}(x) $ and $ M_{i_1,\ldots,i_{k-1}}(x) $ is the minimum positive integer such that $ (\Phi^{M_{i_1,\ldots,i_{k-1}}(x)+1} \Theta)(y_{i_1,\ldots,i_{k-1}}) $ is not defined.
By Lemmas \ref{Phipermute} and \ref{Thetapermute}, each point $ y_{i_1,\ldots,i_k} $ has level $ k $.
\end{enumerate}
We have recursively defined the symbolic dynamics of a forward orbit.
For a finite orbit, this process must terminate, resulting in a finite sequence space.
Naturally, the sequence space for an infinite orbit is infinite.
We now make this precise.
\begin{figure}[h]
\includegraphics[width=1.4\linewidth, trim={4.7cm 15.7cm 0 3.1cm}, clip]{PicTemplate19}
\caption{Symbolic dynamics of finite orbit of $ \Theta(y) $ on the rectangle $ S $. The points are labeled according to Equation \ref{kuppermute}. The map $ \Phi $ moves points up along the Wilson flow, preserving the radial coordinate. The map $ \Theta $ moves points outward, through the insertion. The radius inequality implies that $ \Theta $ increases the radius.}
\label{symborbit}
\end{figure}
\subsubsection{Symbolic dynamics of a finite orbit}
Assume that $ O^+(x) $ is finite, and let $ y \in \mathcal{O}^+(x) \cap S $ be a point of level zero as above.
Since a finite orbit must intersect $ S $ at most a finite number of times, there exists $ N \in \mathbb{N} $ such that
$$
\mathcal{O}^+(\Theta(y)) \cap S = \bigcup_{j=1}^N \bigcup_{i_j=1}^{M_{i_1,\ldots,i_{j-1}}(x)} y_{i_1,\ldots,i_j},
$$
where $ M_{i_1,i_0}(x) = M(x) $, and $ M_{i_1,\ldots,i_{N-1}}(x) < \infty $ for all $ i_1,\ldots, i_{N-1} $.
From Lemmas \ref{Phipermute} and \ref{Thetapermute}, we have that for all $ 1 \leq j \leq N $, the maps $ \Phi $ and $ \Theta $ in the Kuperberg pseudogroup permute these points in the following way.
\begin{align}
\label{kuppermute}
\Phi(y_{i_1,\ldots,i_j}) &= y_{i_1,\ldots,i_j + 1} \\
\Theta(y_{i_1,\ldots,i_{j-1}}) &= y_{i_1,\ldots,i_j, 1} \nonumber
\end{align}
See Figure \ref{symborbit} for a picture of part of a finite orbit's intersection with $ S $ and its permutation by $ \Phi $ and $ \Theta $.
We now have a sequence space $ \Sigma \subset \mathbb{N}^N $ given by
$$
\Sigma = \bigcup_{j=1}^N \bigcup_{i_j=1}^{M_{i_1,\ldots,i_{j-1}}(x)} (i_1, \ldots, i_j).
$$
The Kuperberg pseudogroup acts faithfully on this space by Equation \ref{kuppermute} and we have a bijective coding map
$$
\pi: \Sigma \rightarrow O^+(\Theta(y)) \cap S,
$$
given by $ \pi(\omega) = y_{\omega} $.
In this correspondence, the length of $ \omega $ is equal to the level of $ \pi(\omega) $.
\subsubsection{Symbolic dynamics of an infinite orbit}
We have similar symbolic dynamics for an infinite orbit, but the orbit now has points of arbitrary level so $ N = \infty $.
The sequence space is now $ \Sigma \subset \mathbb{N}^{\mathbb{N}} $, given by
\begin{equation}
\label{inforbsigma}
\Sigma = \bigcup_{j=1}^{\infty} \bigcup_{i_j=1}^{M_{i_1,\ldots,i_{j-1}}(x)} (i_1, \ldots, i_j).
\end{equation}
The coding map $ \pi :\Sigma \rightarrow O^+(\Theta(y)) \cap S $ is still bijective and the Kuperberg pseudogroup acts on $ \Sigma $ as defined in Equation \ref{kuppermute}.
For finite or infinite orbits, the sequence space $ \Sigma $ coding the points in the transverse section was constructed iteratively; we added symbols to the right of words of length $ k-1 $ to define the words of length $ k $.
This implies that the sequence space $ \Sigma $ satisfies the \textit{extension admissibility condition} from Definition \ref{extadm}, and thus is a general symbolic space as defined in Chapter \ref{Thermo}, over the alphabet $ E = \mathbb{N} $.
\vfill
\eject
\section{The Kuperberg minimal set}
\label{Kupmin}
The Kuperberg flow $ \psi_t $ preserves a unique minimal set $ \mathcal{M} \subset K $, with the following characterization.
\begin{theorem}(\cite{Hur}, Theorem 17.1)
\label{minchar}
Let $ \mathcal{M} \subset K $ be the Kuperberg minimal set, and let $ \mathcal{R}' $ be the notched Reeb cylinder.
Then $ \mathcal{M} $ is a codimension one lamination with Cantor transversal $ \tau $, and
$$
\mathcal{M} = \overline{\bigcup_{-\infty < t < \infty} \psi_t(\mathcal{R}')}.
$$
\end{theorem}
This theorem is proved under \textit{generic} assumptions on the insertions and flow, detailed in Chapter 17 of \cite{Hur}.
The assumptions we made in Section \ref{Kupsection} are special cases of these generic assumptions, so the above theorem applies to the plug $ K $ that we have constructed.
We will use this theorem as a point of departure in studying $ \mathcal{M} $.
\subsection{The level decomposition}
First, define
\begin{equation}
\label{MN0}
\mathcal{N}_0 = \bigcup_{-\infty < t < \infty} \psi_t(\mathcal{R}), \; \text{ and } \; \mathcal{M}_0 = \bigcup_{-\infty < t < \infty} \psi_t(\mathcal{R}'),
\end{equation}
so that $ \mathcal{M} = \overline{\mathcal{M}_0} $.
In the notation of orbit surfaces from Definition \ref{orbsurf}, we have $ \mathcal{N}_0 = \mathcal{O}(\mathcal{R}) $ and $ \mathcal{M}_0 = \mathcal{O}(\mathcal{R}') $.
An important part of the analysis in \cite{Hur} that we will require is the level decomposition.
In Definition \ref{leveldef}, for any $ x \in K $ we defined the level function $ n_x : \mathcal{O}(x) \rightarrow \mathbb{N} $ along the orbit of $ x $.
We extend $ n_x $ to a level function $ n_0 $ on $ \mathcal{M}_0 $ in the following way.
\begin{definition}[\textit{level function of $ \mathcal{M}_0 $}]
Let $ x \in \mathcal{M}_0 $.
By Equation \ref{MN0}, there exists $ T \geq 0 $ and $ y \in \gamma^u $ such that $ x = \psi_T(y) $.
In terms of the level function $ n_y : \mathcal{O}(y) \rightarrow \mathbb{N} $, define
$$
n_0(x) = n_y(T).
$$
\end{definition}
The following proposition appears as Proposition 10.1 in \cite{Hur}.
\begin{proposition}
\label{levelfunction}
The function $ n_0 : \mathcal{M}_0 \rightarrow \mathbb{N} = \{0, 1, 2, \ldots \} $ is well-defined.
\end{proposition}
As a consequence, the following level decomposition is well-defined.
\begin{equation}
\label{leveldecompM0}
\mathcal{M}_0 = \bigcup_{k \geq 0} \mathcal{M}_0^n, \text{ where } \mathcal{M}_0^k = \{ x \in \mathcal{M}_0 : n_0(x) = k \}
\end{equation}
Note that the level function $ n_0 $ extends to $ \mathcal{N}_0 $, hence $ \mathcal{N}_0 $ also has a well-defined level decomposition.
\subsection{The intersections $ \mathcal{N}_0 \cap S $ and $ \mathcal{M}_0 \cap S $}
The lower insertion rectangle $ S $ is transverse to the Kuperberg flow $ \psi_t $.
Recall from Equation \ref{Psi} that $ \Psi $ is the Kuperberg pseudogroup of first-return maps of the flow $ \psi_t $.
By construction $ \gamma = S \cap \mathcal{R} $ and $ \gamma^u = S \cap \mathcal{R}' $.
From this and Equation \ref{MN0} we have
\begin{equation}
\label{N0S}
\mathcal{N}_0 \cap S = \bigcup_{g \in \Psi} g(\gamma), \; \text{ and } \; \mathcal{M}_0 \cap S = \bigcup_{g \in \Psi} g(\gamma^u).
\end{equation}
In the quotient plug $ K $, $ \gamma $ is identified with $ \sigma^{-1} \gamma $, a parabolic curve parametrized in Equation \ref{sigmagamma}.
By this equation, we see that $ \sigma^{-1} \gamma(0) $ is the only point of $ \sigma^{-1} \gamma $ with a radial coordinate of $ 2 $; every other point in $ \sigma^{-1} \gamma $ has a radial coordinate $ >2 $.
Thus the Wilson orbit of $ \gamma $ is a double propeller (see Definition \ref{doublepropdef}), and by restricting the parametrization of $ \gamma $ to $ \gamma^u $, we obtain a single propeller (see Definition \ref{propdef}).
Under the Kuperberg flow, each intersection of the orbit of $ \gamma $ with $ S $ is identified with a curve in the base of the plug.
The Kuperberg flow of these curves then follows the Wilson flow up into the interior of the plug, and each of these orbits is a propeller.
Each of these surfaces then may intersect the insertion regions $ S_i $ again, and the process repeats, creating an infinitely branching union of propellers with a complicated embedding in $ K $.
As Kuperberg orbits of $ \gamma $ and $ \gamma^u $, the surfaces $ \mathcal{N}_0 $ and $ \mathcal{M}_0 $ are unions of double and single propellers, respectively.
These branching surfaces are termed ``choux-fleurs" in \cite{Ghy}, and are extensively studied in \cite{Hur}.
Each single propeller in $ \mathcal{M}_0 $ is a restriction of a double propeller in $ \mathcal{N}_0 \supset \mathcal{M}_0 $.
The embedding and transverse dynamics of $ \mathcal{N}_0 $ and $ \mathcal{M}_0 $ are complicated.
By Equation \ref{N0S}, to study $ \mathcal{N}_0 \cap S $ we must compute the image of $ \gamma $ under the full Kuperberg pseudogroup $ \Psi $ as defined in Equation \ref{Psi}.
However, determining the admissible compositions of the generators of $ \Psi $ is difficult.
So we will begin by focusing on the dynamics of $ \Psi_1 = \langle \Phi, \Theta \rangle $ (see Equation \ref{Psi1}), defining
\begin{equation}
\label{N01S}
\mathcal{N}_{0,1} \cap S = \bigcup_{g \in \Psi_1} g(\gamma).
\end{equation}
Of course, we have $ \mathcal{M}_{0,1} $ defined similarly.
\subsection{The pseudogroup action on the level decomposition}
\label{pseudomin}
In Chapter \ref{Kuppseudo} we studied the action of the pseudogroup $ \langle \Phi, \Theta \rangle $ on the transverse section $ \mathcal{O}^+(\Theta(y)) \cap S $ of an orbit.
We defined a general sequence space $ \Sigma $ and a bijective coding map $ \pi : \Sigma \rightarrow \mathcal{O}^+(\Theta(y)) \cap S $, and studied the dynamics of the pseudogroup on the section and the induced dynamics on the sequence space.
In this section, we will develop similar symbolic dynamics for $ \langle \Phi, \Theta \rangle $ acting on $ \mathcal{N}_{0,1} \cap S $.
In addition to labeling the curves in $ \mathcal{N}_{0,1} \cap S $ according to a sequence space, we will explicitly parametrize these curves in coordinates $ (r, \theta, z) $.
To do this, we return to the assumptions made in Chapter \ref{InsertAssume}.
These will allow us to explicitly parametrize the propellers $ \mathcal{O}^+(\gamma, 0, t) $ for $ t \geq 0 $.
The images of $ \gamma $ under the maps $ \Phi $ and $ \Theta $ can be explicitly calculated from these parametrizations, by studying their intersections with the lower insertion rectangle $ S $.
As we studied in Chapter \ref{Kuppseudo}, the sequence space coding an orbit is defined in terms of escape times.
By analyzing the parametrizations of the curves in $ \mathcal{N}_{0,1} \cap S $, we will obtain precise estimates on their escape times, and thus the sequence space coding these curves.
The level decomposition of $ \mathcal{N}_0 $ (identical to that of $ \mathcal{M}_0 $ given in Equation \ref{leveldecompM0}) induces the following level decomposition of $ \mathcal{N}_{0,1} \cap S $.
\begin{equation}
\label{leveldecompNS}
\mathcal{N}_{0,1} \cap S = \bigcup_{k \geq 0} \mathcal{N}_{0,1}^k \cap S, \; \text{ where } \; \mathcal{N}_{0,1}^k \cap S = \{ x \in \mathcal{N}_{0,1} \cap S : n_0(x) = k \}
\end{equation}
By Propositions \ref{Phipermute} and \ref{Thetapermute}, the pseudogroup $ \langle \Phi, \Theta \rangle $ acts on this level decomposition in the following way.
\begin{align}
\label{kuppermute2}
\Phi : \mathcal{N}_{0,1}^k \cap S &\mapsto \mathcal{N}_{0,1}^k \cap S \\
\Theta : \mathcal{N}_{0,1}^k \cap S &\mapsto \mathcal{N}_{0,1}^{k+1} \cap S \nonumber
\end{align}
\subsubsection{The level-zero curve $ \mathcal{N}_{0,1}^0 \cap S $}
By Equation \ref{MN0}, $ \mathcal{N}_0 = \mathcal{O}(\mathcal{R}) $.
Points in the Reeb cylinder $ \mathcal{R} $ have level zero, as do points in the intersection $ \mathcal{R} \cap S = \gamma $, hence $ \mathcal{N}_{0,1}^0 \cap S = \{ \gamma \} $.
Recall the parametrization of $ \gamma $ given in Equation \ref{gamma}:
$$
\gamma(s) = (2, \beta, -1+s), \; \text{ with } \; s \in [-R, R].
$$
We refer to the midpoint $ \gamma(0) $ as the \textit{vertex} of $ \gamma $, which is the intersection $ l \cap S $ of the special orbit with $ S $.
\subsubsection{The level-one curves $ \mathcal{N}_{0,1}^1 \cap S $}
For all $ i_1 \in \mathbb{N} $ let
\begin{equation}
\label{gammai1}
\gamma_{i_1} = (\Phi^{i_1-1} \Theta)(\gamma).
\end{equation}
Because $ \gamma $ has level zero, by Equation \ref{kuppermute2} $ \gamma_{i_1} $ has level one for all $ i_1 $.
We now use the assumptions we made in Chapter \ref{InsertAssume} to parametrize each $ \gamma_{i_1} $.
\begin{proposition}
\label{gamma1param}
For all $ i_1 \in \mathbb{N} $ there exist $ s_{i_1}^{\pm} $ with $ -R < s_{i_1}^- < 0 < s_{i_1}^+ < R $ such that the parametrization of $ \gamma_{i_1}:[s_{i_1}^-, s_{i_1}^+]\setminus 0 \rightarrow S $ is
\begin{align}
\label{gamma1}
\displaystyle
\gamma_{i_1}(s) &= \left(2+s^2, \beta, -1+q_{i_1}(s)\right) \text{, where } \\
& \displaystyle q_{i_1}(s) = s^2 \tan \left(\frac{s^2}{R^2} T_{i_1}(s)- \tan^{-1}\left(\frac{R}{s^2}\right)\right), \nonumber \\
& \displaystyle T_{i_1}(s) = a^{-1}(2\pi i_1+ \beta-\alpha+s)+R-1. \nonumber
\end{align}
\end{proposition}
\begin{proof}
By definition $ \gamma_{i_1} $ is the $ i_1 $-th return time of $ \sigma^{-1} \gamma $ to $ S $.
Recall from Equation \ref{sigmagamma} the the parametrization:
$$
\sigma^{-1}\gamma (s) = (2+s^2, \alpha-s, -2), \; \text{ with } \; s \in [-R,R]
$$
From $ \{z=-2\} $ to $ \{z=-1-R\} $, the Kuperberg flow is given by Wilson's flow in Equation \ref{wilout}.
Applying this to the above parametrization, we obtain a parametrization of the following orbit strip:
$$
\mathcal{O}^+(\sigma^{-1}\gamma, 0, 1-R) = (2+s^2, \alpha-s+at, -2+t), \text{ where } s \in [-R, R] \text{ and } t \in [0,1-R].
$$
As the Wilson orbit of the parabolic curve $ \sigma^{-1} \gamma $, we see that $ \mathcal{O}^+(\sigma^{-1}\gamma, 0, 1-R) $ is a double propeller parametrized by $ s $ and $ t $.
Its intersection with the bottom annulus $ C^- $ is the curve $ \psi_{1-R} \sigma^{-1} \gamma $.
For $ |z+1| \leq R $ the Kuperberg flow is now given by Wilson's flow in Equation \ref{wilin}.
Applying this to the parametrization of $ \psi_{1-R} \sigma^{-1} \gamma $, we obtain a parametrization of the double propeller inside this region.
\begin{multline*}
\mathcal{O}^+(\psi_{1-R} \sigma^{-1} \gamma, 0, T) = \left(2+s^2, \alpha-s+a(1-R+t), -1+\left(\frac{s^2}{R^2} t-\tan^{-1} \left(\frac{r}{s^2}\right)\right) \right), \text{ where } \\
s \in [-R,R] \setminus \{0\} \text{ and } t \in [0,T].
\end{multline*}
By Equation \ref{gammai1} and the definition of $ \Phi $, each curve $ \gamma_{i_1} $ is the $ i_1 $-th intersection of this double propeller with $ S $ as $ T $ increases.
To find parametrizations of these curves, recall by Equation \ref{Sstrip} that $ S $ has a constant angular coordinate $ \theta=\beta $.
Setting the $ \theta $ coordinate in the parametrization of $ \mathcal{O}_1^+(\psi_{1-R} \sigma^{-1} \gamma, 0, T) $ to $ \beta+2\pi i_1 $ and solving for $ t>0 $, we find that the $ i_1 $-th return time of $ \psi_{1-R} \sigma^{-1} \gamma $ to $ S $ is
$$
T_{i_1}(s) = a^{-1}(2\pi i_1 +\beta-\alpha +s) + R-1
$$
Substituting this back into the parametrization of $ \mathcal{O}^+(\psi_{1-R} \sigma^{-1} \gamma, 0, T) $ we obtain the desired formula given in Equation \ref{gamma1}.
However, these parametrizations are not valid for all $ s \in [-R,R] $ or $ i_1 \in \mathbb{N} $; because $ \gamma_{i_1} $ is defined by $ \Phi, \Theta : S \rightarrow S $ we must restrict to values of $ s $ and $ i_1 $ such that $ \gamma_{i_1}(s) \in S $.
The upper boundary $ S^+ $ of $ S $ has a constant $ z $-coordinate $ z=-1+R $.
Thus in the notation of Equation \ref{gamma1}, our restriction should be such that $ q_{i_1}(s) \leq R $.
Define $ s_{i_1}^+ $ and $ s_{i_1}^- $ as the unique solutions to the equation $ q_{i_1}(s) = R $ on the domains $ s>0 $ and $ s<0 $, respectively.
Using the parametrization for $ q_{i_1} $ given in Equation \ref{gamma1}, it is easy to show that by the intermediate value theorem that these exist, and that by monotonicity of $ q_i $ they are unique.
In Chapter \ref{Transversal}, we will prove that the radial coordinates of the endpoints $ \gamma_{i_1}(s_{i_1}^{\pm}) $ decrease monotonically as $ i_1 \to \infty $.
Referring to Equation \ref{Sstrip}, we see that $ S $ has a fixed radial width of $ b>0 $, so there exists a minimal $ N_b \in \mathbb{N} $ such that $ \gamma_{i_1}(s_{i_1}^{\pm}) \in S $ for all $ i_1 \geq N_b $.
In terms of $ N_b $, we define
\begin{equation}
\label{Sigma1}
\Sigma_{b,1} = \{ N_b, N_b+1, \ldots, \} \subset \mathbb{N}.
\end{equation}
The index $ i_1 $ ranges through all $ \Sigma_{b,1} $ because the double propeller $ \mathcal{O}^+(\psi_{1-R} \sigma^{-1} \gamma, 0, T) $ is trapped.
We conclude by restricting our parametrization to $ \gamma_{i_1}: [s_{i_1}^-, s_{i_1}^+] \setminus 0 \rightarrow S $, and indices to $ i_1 \in \Sigma_{b,1} $.
\end{proof}
As with the level-zero curve $ \gamma $, for all $ i_1 \in \Sigma_{b,1} $, we call $ \gamma_{i_1}(0) $ the \textit{vertex} of $ \gamma_{i_1} $.
Defining $ v_{i_1} = \lim_{s \to 0} q_{i_1}(s) $, we see by Equation \ref{gamma1} that the vertex of $ \gamma_{i_1} $ is
$$
\lim_{s \to 0} \gamma_{i_1}(s) = (2, \beta, -1+v_{i_1})
$$
Notice that the vertices of the level-one curves $ \gamma_{i_1} $ lie on the level-zero curve $ \gamma $.
Explicitly, $ \gamma(v_{i_1}) = (2, \beta, -1+v_{i_1}) $ using Equation \ref{gamma}.
This relation will imply a nesting property for higher-level curves.
Using the parametrization given in Equation \ref{gamma1}, it can be shown that $ v_{i_1}<0 $ for all $ i_1 $ and that $ \lim_{i_1 \to \infty} v_{i_1} = 0 $.
So as $ i_1 \rightarrow \infty $, these vertices limit on intersection $ l \cap S = (2, \beta, -1)$, the vertex of $ \gamma $.
See Figure \ref{levelone} for a plot of these curves.
\begin{figure}[h]
\includegraphics[width=0.9\linewidth, trim={3.2cm 10.5cm 1cm 2.8cm}, clip]{Levelone}
\caption{A plot of the level-one curves $ \gamma_{i_1} \subset S $ for $ i_1 = 1,2, \ldots, 20 $, and $ a=R=1 $, $ \alpha = \beta = 0 $. The vertices $ v_{i_1} $ form a vertical sequence on the Reeb cylinder $ \{r=2\} $, limiting on the special point $ (2,\beta,-1) $.}
\label{levelone}
\end{figure}
\subsubsection{The level-two curves $ \mathcal{N}_{0,1}^2 \cap S $}
For each $ i_1 \in \Sigma_{b,1} $ define
\begin{equation}
\label{gammai1i2}
\gamma_{i_1, i_2} = (\Phi^{i_2-1} \Theta)(\gamma_{i_1}).
\end{equation}
Because each $ \gamma_{i_1} $ has level one, by Equation \ref{kuppermute2} each $ \gamma_{i_1, i_2} $ has level two.
We can parametrize each $ \gamma_{i_1,i_2} $ as follows.
{\small
\begin{align}
\label{gamma2}
\gamma_{i_1,i_2}(s) &= \left(2+(s^2+q_{i_1}^2(s)), \beta, -1+q_{i_1,i_2}(s)\right) \text{, where } \\
& \displaystyle q_{i_1,i_2}(s) = (s^2+q_{i_1}^2(s)) \tan \left(\left(\frac{s^2+q_{i_1}^2(s)}{R^2}\right) T_{i_1,i_2}(s)- \tan^{-1}\left(\frac{R}{s^2+q_{i_1}^2(s)}\right)\right), \nonumber \\
& \displaystyle T_{i_1,i_2}(s) = a^{-1}(2\pi i_2+ \beta-\alpha+q_{i_1}(s))+R-1 \nonumber
\end{align}
}
Here $ \gamma_{i_1,i_2} : [s_{i_1,i_2}^-, s_{i_1,i_2}^+]\setminus 0 \rightarrow S $, where $ s_{i_1,i_2}^- < 0 < s_{i_1,i_2}^+ $ are the solutions to the equation $ q_{i_1,i_2}(s) = R $.
The derivation of the parametrization in Equation \ref{gamma2} goes exactly like the proof of Proposition \ref{gamma1param}; we follow the orbit surface of each $ \gamma_{i_1} $ through the insertion and calculate its $ i_2 $-th intersection with $ S $.
We omit the details.
See Figure \ref{leveltwo} for a plot of these curves.
It remains to determine the admissible words $ (i_1,i_2) $ coding the level-two curves in $ \mathcal{N}_{0,1}^2 $.
To determine these words, we will need to estimate the escape times of the vertices of the level-two curves $ \gamma_{i_1,i_2} $, which we now define.
As with the level-one curves, we call $ \gamma_{i_1,i_2}(0) $ the \textit{vertex} of $ \gamma_{i_1,i_2} $, and define $ v_{i_1,i_2} = \lim_{s \to 0} q_{i_1, i_2}(s) $, so that the vertex is
$$
\lim_{s \to 0} \gamma_{i_1,i_2}(s) = \left(2+v_{i_1}^2, \beta, -1+v_{i_1,i_2} \right),
$$
using Equation \ref{gamma2}.
The level-two curves satisfy an important nesting property that we now describe.
\begin{definition}
Let $ \eta $ be a curve in $ S $, and suppose that $ \eta \cup S^+ $ bounds a closed region in $ S $. If $ \zeta $ is another curve in $ S $, we say that $ \zeta $ is \textit{nested in} $ \eta $ if the image of $ \zeta $ is contained in this closed region.
\end{definition}
Notice in Figure \ref{levelone} that each $ \gamma_{i_1} \cup S^+ $ bounds a closed region.
\begin{proposition}
\label{nesting2}
For each $ (i_1, i_2) $, the level-two curve $ \gamma_{i_1, i_2} $ is nested in $ \gamma_{i_2} $.
\end{proposition}
\begin{proof}
Recall that $ v_{i_1} = \lim_{s \to 0} q_{i_1}(s) $.
Using Equations \ref{gamma1} and \ref{gamma2}, it is easy to show that $ \lim_{s \to 0} T_{i_1,i_2}(s) = T_{i_2}(v_{i_1}) $ and $ \lim_{s \to 0} q_{i_1, i_2}(s) = q_{i_2}(v_{i_1}) $.
From this we obtain that
\begin{align*}
\lim_{s \to 0} \gamma_{i_1,i_2}(s) &= (2+v_{i_1}^2, \beta, -1+q_{i_2}(v_{i_1})) \\
&= \gamma_{i_2}(v_{i_1}).
\end{align*}
This shows that the vertex of $ \gamma_{i_1, i_2} $ is located on the image of $ \gamma_{i_2} $.
By the radius inequality, $ \gamma_{i_1, i_2} $ is nested in $ \gamma_{i_2} $.
\end{proof}
\begin{figure}[h]
\includegraphics[width=0.9\linewidth, trim={3.2cm 10.5cm 1cm 2.8cm}, clip]{Leveltwo}
\caption{A plot of the level-two curves $ \gamma_{i_1,i_2} $ in $ S_{\epsilon} $ where $ i_1 = 1 $. Note that each $ \gamma_{i_1,i_2} $ is nested in $ \gamma_{i_2} $.}
\label{leveltwo}
\end{figure}
Inspecting the parametrization in Equation \ref{gamma2}, we see that for a fixed $ (i_1, i_2) $, the radial coordinate of each $ \gamma_{i_1,i_2} $ is bounded away from 2.
Each $ \gamma_{i_1,i_2} $ is the $i_2$-th intersection of the orbit surface $ \mathcal{O}^+(\psi_{1-R} \sigma^{-1} \gamma_{i_1}, 0, T) $ with $ S $, so this surface is \textit{not} a double propeller and escapes the plug in finite time.
In particular it has finitely many intersection curves, thus for a fixed $ i_1 \in \Sigma_{b,1} $ there are only finitely many values of $ i_2 $ such that $ (\Phi^{i_2-1} \Theta) (\gamma_{i_1}) \cap S \neq \emptyset $.
The minimal value of $ i_2 $ is $ N_b $, because $ \gamma_{i_1,i_2} $ is nested in $ \gamma_{i_2} $ by Proposition \ref{nesting2}.
So for each $ i_1 \in \Sigma_{b,1} $ there exists $ M_{i_1} $ such that $ N_b \leq i_2 \leq M_{i_1} $, hence the admissible words defining $ \gamma_{i_1,i_2} $ are
\begin{equation}
\label{Sigma2}
\Sigma_{b,2} = \bigcup_{i_1 \in \Sigma_{b,1}} \bigcup_{i_2=N_b}^{M_{i_1}} (i_1,i_2) = \bigcup_{i_1=N_b}^{\infty} \bigcup_{i_2=N_b}^{M_{i_1}} (i_1,i_2).
\end{equation}
Using the parametrization in Equation \ref{gamma2}, we can show that the vertex of each curve is its point of minimal $ z $-coordinate.
Recall that $ \gamma_{i_1,i_2} $ is defined (i.e. the parametrization in Equation \ref{gamma2} is valid) if and only if $ q_{i_1,i_2}(s) = R $ has a solution; equivalently, if $ q_{i_1,i_2}(s) \leq R $ for some $ s $.
Since $ -1+q_{i_1,i_2}(s) $ is the $ z $-coordinate of $ \gamma_{i_1,i_2}(s) $, we see that $ \gamma_{i_1,i_2} $ is defined if and only if $ v_{i_1,i_2} \leq R $.
Thus $ M_{i_1} $ coincides with the \textit{escape time} of the vertex as defined in Chapter \ref{Kuppseudo}; for a fixed $ i_1 \in \Sigma_{b,1} $, it is the maximal $ i_2 $ such that $ v_{i_1,i_2} \leq R $.
Using this, we can find explicit bounds on $ M_{i_1} $ by estimating these escape times.
\begin{proposition}
\label{escape2}
For each $ i \in \Sigma_{b,1} $, let $ M_i $ be the greatest positive integer such that $ v_{i, M_i} \leq R $.
Then there exist constants $ C,K>0 $ such that $ M_i $ is asymptotic to $ C+Ki^2 $.
More precisely, for any $ \delta>0 $ there is an integer $ N_1>0 $ with
$$
C+(K-\delta)i^2 < M_i < (C+\delta)+Ki^2
$$
for all $ i\geq N_1 $.
\end{proposition}
\begin{proof}
We will prove the upper bound; the lower bound is similar.
Recall from the proof of Proposition \ref{nesting2} the nesting property $ v_{i_1,i_2} = q_{i_2}(v_{i_1}) $.
In particular, $ v_{i,M_i} = q_{M_i}(v_i) $.
Since the upper boundary $ S^+ $ of $ S $ has a constant $ z $-coordinate of $ -1+R $, we have that $ M_i $ is the greatest positive integer such that $ q_{M_i}(v_i) \leq R $.
Referring to the parametrization in Equation \ref{gamma1param}, this inequality is equivalent to
$$
2\pi M_i \leq \alpha-\beta+a(1-R)-v_i + \frac{2aR^2}{v_i^2} \tan^{-1} \left(\frac{R}{v_i^2}\right).
$$
Recall that $ \lim_{i \to \infty} v_i = 0 $. Thus for any $ \delta>0 $, there exists $ N>0 $ such that $ 0 < -v_i < 2\pi \delta $ for all $ i\geq N $.
Also note that $ \displaystyle \tan^{-1}\left(\frac{R}{v_i^2}\right) < \frac{\pi}{2} $ for all $ i $.
Substituting these into the above inequality, we obtain
\begin{equation}
\label{ineq}
2\pi M_i < \alpha-\beta+a(1-R)+2\pi \delta + \frac{\pi aR^2}{v_i^2}.
\end{equation}
Using the definition $ v_i = \lim_{s \to 0} q_i(s) $, it is easy to show that there exists a constant $ p>0 $ such that $ \displaystyle v_i = -\frac{p}{i} $.
We define the following constants.
\begin{equation}
\label{defCK}
C = \frac{\alpha-\beta+a(1-R)}{2\pi}, \hskip 2cm K = \frac{aR^2}{2p^2}
\end{equation}
Substituting these into Equation \ref{ineq}, we obtain
$$
M_i < (C+\delta)+K i^2.
$$
\end{proof}
Recall that the vertices $ v_{i_1} $ of $ \gamma_{i_1} $ limit on the special orbit intersection $ l \cap S $, the vertex of $ \gamma $.
From the recursive definition in Equation \ref{gammai1i2} and the parametrizations in Equation \ref{gamma2}, one can show that the vertices of $ \gamma_{i_1,i_2} $ limit on the vertices of $ \gamma_{i_1} $.
See Figure \ref{leveltwo2} for another picture of the level-two curves inside the level-one curves, and observe this limiting behavior.
\begin{figure}[h]
\includegraphics[width=0.9\linewidth, trim={0 0.5 0 0.8cm}, clip]{TheMinimalSet}
\caption{The level-two curves limiting on the level-one curves.}
\label{leveltwo2}
\end{figure}
\subsubsection{The level-k curves $ \mathcal{N}_{0,1}^k \cap S $}
Let $ \Sigma_{b,k-1} $ denote the admissible words of level $k-1$ defining the curves $ \gamma_{i_1,\ldots,i_{k-1}} $.
As before, we define
\begin{equation}
\label{gammai1ik}
\gamma_{i_1, \ldots, i_k} = \Phi^{i_k-1} \Theta(\gamma_{i_1, \ldots, i_{k-1}}),
\end{equation}
and observe that $ \gamma_{i_1, \ldots, i_k} \in \mathcal{N}_{0,1}^k \cap S $ by Equation \ref{kuppermute2}.
As with levels one and two, we can explicitly parametrize these curves.
{\Small
\begin{align}
\label{gammak}
\gamma_{i_1, \ldots, i_k}(s) &= \left(2+\left(s^2+\sum_{j=1}^{k-1} q_{i_1, \ldots,i_j}^2(s)\right), \beta, -1+q_{i_1,\ldots,i_k}(s)\right) \text{, where } \\
&\displaystyle q_{i_1,\ldots,i_k}(s) = \left(s^2+\sum_{j=1}^{k-1} q_{i_1, \ldots,i_j}^2(s) \right) \tan \left(\left(\frac{s^2+\sum_{j=1}^{k-1} q_{i_1, \ldots,i_j}^2(s)}{R^2} \right) T_{i_1,\ldots,i_k}(s)-\tan^{-1}\left(\frac{R}{s^2+\sum_{j=1}^{k-1} q_{i_1, \ldots,i_j}^2(s)}\right)\right), \nonumber \\
&\displaystyle T_{i_1,\ldots,i_k}(s) = a^{-1}(2\pi i_k+ \beta-\alpha+q_{i_1,\ldots,i_{k-1}}(s))+R-1 \nonumber
\end{align}
}
For each $ \omega = (i_1,\ldots,i_k) $, we have $ \gamma_{\omega} : [s_{\omega}^-, s_{\omega}^+]\setminus 0 \rightarrow S $, where $ s_{\omega}^{\pm} $ are the unique solutions to the equation $ q_{\omega}(s) = R $.
As with levels one and two, we call $ \gamma_{\omega}(0) $ the \textit{vertex} of $ \gamma_{\omega} $.
The $ z $-coordinate of the vertex is $ -1+v_{\omega} $, where
$$
v_{\omega} = \lim_{s \to 0} q_{\omega}(s).
$$
The proof of the following proposition is identical to the proof of Proposition \ref{nesting2}.
\begin{proposition}
\label{nestingk}
For each $ (i_1,\ldots,i_k) \in \Sigma_{b,k} $, the level-$k$ curve $ \gamma_{i_1,\ldots,i_k} $ is nested in the level-$ (k-1) $ curve $ \gamma_{i_2, \ldots, i_k} $.
\end{proposition}
It remains to recursively determine the admissible words $ \Sigma_{b,k} $ from $ \Sigma_{b,k-1} $.
For fixed values of $ (i_1, \ldots, i_{k-1}) \in \Sigma_{b,k-1} $, the curve $ \gamma_{i_1,\ldots,i_k} $ is defined for finitely many $ N_b \leq i_k \leq M_{i_1,\ldots, i_{k-1}} $, resulting in a sequence space
\begin{align}
\label{Sigmak}
\Sigma_{b,k} &= \bigcup_{(i_1,\ldots,i_{k-1}) \in \Sigma_{b,k-1}} \bigcup_{i_k = N_b}^{M_{i_1,\ldots,i_{k-1}}} (i_1,\ldots,i_k) \\
&= \bigcup_{i_1 = N_b}^{\infty} \bigcup_{i_2 = N_b}^{M_{i_1}} \cdots \bigcup_{i_k=N_b}^{M_{i_1,\ldots,i_{k-1}}} (i_1,\ldots,i_k) \nonumber.
\end{align}
As in Proposition \ref{escape2} we will estimate $ M_{i_1,\ldots,i_{k-1}} $ via escape times of vertices $ v_{\omega} $.
Recall from Equation \ref{defCK} the constants $ C,K>0 $ determining the admissible words of level two.
\begin{proposition}
\label{escapek}
For each $ (i_1, \ldots, i_{k-1}) \in \Sigma_{b,k-1} $ let $ M=M_{i_1,\ldots,i_{k-1}} $ be the greatest positive integer such that $ v_{i_1,\ldots,i_{k-1}, M} \in S $.
Then for large values of $ i_1,\ldots,i_{k-1} $, $ M_{i_1,\ldots,i_{k-1}} $ is asymptotic to $ C+K i_{k-1}^2 $.
More precisely, for any $ \delta>0 $ there is an integer $ N_{k-1} >0 $ with
$$
C+(K-\delta)i_{k-1}^2 < M_{i_1,\ldots,i_{k-1}} < (C+\delta)+K i_{k-1}^2
$$
when $ i_1,\ldots,i_{k-1} \geq N_{k-1} $.
\end{proposition}
\begin{proof}
We will prove the upper bound; the lower bound is similar.
By Proposition \ref{nestingk} we have the nesting property $ v_{i_1,\ldots,i_k} = q_{i_2,\ldots,i_k}(v_{i_1}) $.
In particular, $ v_{i_1,\ldots,i_{k-1},M} = q_{i_2,\ldots,i_{k-1},M}(v_{i_1}) $.
Then $ M=M_{i_1,\ldots,i_{k-1}} $ is the greatest positive integer such that $ q_{i_2,\ldots,i_{k-1},M}(v_{i_1}) \leq R $.
By Equation \ref{gammak} this is equivalent to
{\small
$$
2\pi M_{i_1,\ldots,i_{k-1}} < \alpha-\beta+a(1-R) - q_{i_2,\ldots,i_{k-1}}(v_{i_1}) + \frac{2R^2}{v_{i_1}^2+\sum_{j=2}^{k-1}q_{i_2,\ldots,i_j}^2(v_{i_1})} \tan^{-1} \left(\frac{R}{v_{i_1}^2+\sum_{j=2}^{k-1}q_{i_2,\ldots,i_j}^2(v_{i_1})} \right)
$$
}
Recall that $ \lim_{i \to \infty} v_i = 0 $ and $ v_{\omega} = \lim_{s \to 0} q_{\omega} $.
Combining this with the nesting property we obtain that
$$
\lim_{i_1 \to \infty} q_{i_2,\ldots,i_{k-1}}(v_{i_1}) = v_{i_2,\ldots,i_{k-1}} = q_{i_3,\ldots,i_{k-1}}(v_{i_2}),
$$
and by induction, for all $ 1 \leq j \leq k-1 $ that
$$
\lim_{i_1,\ldots,i_{j-1} \to \infty} q_{i_2,\ldots,i_j}(v_{i_1}) = v_{i_j}
$$
Then for a sufficiently large integer $ N_{k-1} $, we have for all $ i_1, \ldots, i_{k-1} \geq N_{k-1} $ that
$$
0 < -q_{i_2,\ldots,i_{k-1}}(v_{i_1}) < 2\pi \delta, \; \text{ and } \;
0 < |q_{i_2,\ldots,i_j}(v_{i_1})| < \frac{\sqrt{\delta}}{k}.
$$
Substituting this into the first inequality and using that $ \displaystyle \tan^{-1}(\cdot) < \frac{\pi}{2} $, we obtain for $ i_1, \ldots, i_{k-1} \geq N_k $ that $ M_{i_1,\ldots,i_{k-1}} $ is the greatest positive integer such that
$$
2\pi M_{i_1,\ldots,i_{k-1}} < \alpha-\beta+a(1-R)+2\pi \delta + \frac{\pi aR^2}{v_{i_{k-1}}^2+\delta}.
$$
Since $ \displaystyle v_{i_{k-1}} = -\frac{p}{i_{k-1}} $ from the proof of Proposition \ref{escape2}, this is equivalent to
$$
M_{i_1,\ldots,i_{k-1}} < (C+\delta)+K i_{k-1}^2.
$$
\end{proof}
Finally, from the recursive definition in Equation \ref{gammai1ik}, and Equation \ref{kuppermute2}, the pseudogroup $ \langle \Phi, \Theta \rangle $ permutes the curves $ \gamma_{i_1,\ldots,i_k} $ in the following way.
\begin{align}
\label{kuppermutegamma}
\Phi(\gamma_{i_1,\ldots,i_k}) &= \gamma_{i_1,\ldots,i_k + 1} \\
\Theta(\gamma_{i_1,\ldots,i_{k-1}}) &= \gamma_{i_1,\ldots,i_k, 1} \nonumber
\end{align}
\subsection{Symbolic dynamics of the rectangle $ S $}
\label{symbrect}
In the previous section we described the level decomposition of the intersection $ \mathcal{N}_{0,1} \cap S $, parametrized the curves in each level set, and labeled the curves by words in a sequence space.
The parametrization of the parabolic curve $ \sigma^{-1} \gamma $ in Equation \ref{sigmagamma} was crucial in this analysis.
In this section, we recall the additional assumption from Chapter \ref{InsertAssume} that $ \sigma^{-1}(S) $ is a parabolic strip (see Figure \ref{kup2}).
More specifically, we assumed that the foliation of $ S $ by the vertical lines $ \{ \gamma_c \}_{0 \leq c \leq b} $ (where $ \gamma_c $ is parametrized in Equation \ref{gammac}) is mapped under $ \sigma^{-1} $ into the parabolic foliation $ \{ \sigma^{-1} \gamma_c \}_{0 \leq c \leq b} $ of $ \sigma^{-1}(S) $,
where $ \sigma^{-1}\gamma_c $ is parametrized in Equation \ref{sigmagammac}.
See Figure \ref{stripfol}.
\begin{figure}[h]
\includegraphics[width=0.9\linewidth, trim={-1cm -0.5cm 4cm -0.5cm}, clip]{Stripfol}
\caption{The vertical curves $ \gamma_c $ and their images $ \sigma^{-1}\gamma_c $.}
\label{stripfol}
\end{figure}
For each $ 0 \leq c \leq b $, define
\begin{equation}
\label{Nc}
\mathcal{N}_c = \bigcup_{-\infty < t < \infty} \psi_t(\gamma_c).
\end{equation}
Since $ \gamma_0 = \gamma $, this agrees with our definition of $ \mathcal{N}_0 $ given in Equation \ref{MN0}.
For $ 0 \leq c \leq b $, and for $ i=1,2 $ we have a definition of $ \mathcal{N}_{c,1} \cap S $ identical to the definition of $ \mathcal{N}_{0,1} \cap S $ given in Equation \ref{N01S}.
As in Equation \ref{leveldecompNS}, for each $ c $ there is a well-defined level decomposition of $ \mathcal{N}_{c,1} \cap S $.
\begin{equation}
\label{leveldecompc}
\mathcal{N}_{c,1} \cap S = \bigcup_{k \geq 0} \mathcal{N}_{c,1}^k \cap S.
\end{equation}
Each level set $ \mathcal{N}_{c,1}^k \cap S $ is comprised of curves $ \gamma_{c,(i_1,\ldots,i_k)} $ recursively defined by pseudogroup elements as
\begin{equation}
\label{gammack}
\gamma_{c,(i_1,\ldots,i_k)} = \Phi^{i_k-1} \Theta\left(\gamma_{c,(i_1,\ldots,i_{k-1})}\right),
\end{equation}
exactly as we defined $ \gamma_{i_1,\ldots,i_k} $ in Equation \ref{gammai1ik}.
The symbolic dynamics of the action of $ \langle \Phi, \Theta \rangle $ on $ \mathcal{N}_{c,1} \cap S $ is similar to that of its action on $ \mathcal{N}_{0,1} \cap S $.
For $ c=0 $ and each $ k \geq 1 $ we recover the sequence space $ \Sigma_{b,k} $ from Equation \ref{Sigmak} coding the curves $ \gamma_{i_1,\ldots,i_k} \in \mathcal{N}_{0,1} \cap S $.
For $ 0 < c \leq b $, there is a similar sequence space $ \Sigma_{c,k} $ coding the curves $ \gamma_{c,(i_1,\ldots,i_k)} \in \mathcal{N}_{c,1} \cap S $, but this sequence space has fewer admissible words because the escape times of $ \gamma_c $ under the action of $ \Phi $ decrease as $ c \rightarrow b $.
This is evident from Figure \ref{stripfol}.
Finally, let $ A_i = \Phi^{i-1} \Theta(S) $, and recursively define
$$
A_{i_1,\ldots,i_k} = \Phi^{i_k-1} \Theta \left(A_{i_1,\ldots,i_{k-1}}\right).
$$
Notice that the admissible words $ \omega $ coding the sets $ A_{\omega} $ are the same as those coding the curves $ \gamma_{\omega} $ because $ \gamma_{\omega} \subset \partial A_{\omega} $, so their escape times are equal.
This is an important point that we will return to later, when defining function systems on the transversal.
The nesting property for curves $ \gamma_{\omega} $ established in Proposition \ref{nestingk} implies that the sets $ A_{\omega} $ are nested.
\begin{proposition}
\label{nestingstrip}
For each $ (i_1,\ldots,i_k) \in \Sigma_k $, we have
$$
A_{i_1,\ldots,i_k} \subset A_{i_2,\ldots,i_k}.
$$
\end{proposition}
See Figure \ref{transstripfig} for a picture of these sets $ A_{\omega} $ for level-one $ \omega $.
\begin{figure}[h]
\includegraphics[width=0.9\linewidth, trim={3.2cm 10.5cm 1cm 2.8cm}, clip]{PicTemplate22}
\caption{The sets $ A_{\omega} $ for $ \omega \in \Sigma_{b,1} $ of level one. Notice that each curve $ \gamma_{\omega} $ is the lower boundary of each $ A_{\omega} $. Compare with Figure \ref{levelone}.}
\label{transstripfig}
\end{figure}
\subsection{Summary of symbolic dynamics}
From Equation \ref{leveldecompNS}, the transverse intersection $ \mathcal{N}_{0,1} \cap S $ has a level decomposition
\begin{equation}
\label{leveldecompNS2}
\mathcal{N}_{0,1} \cap S = \bigcup_{k \geq 0} \mathcal{N}_0^k \cap S.
\end{equation}
Each level set is a collection of curves
\begin{equation}
\label{Ok}
\mathcal{N}_{0,1}^k \cap S = \bigcup_{\omega \in \Sigma_{b,k}} \gamma_{\omega},
\end{equation}
where $ \Sigma_{b,k} \subset \mathbb{N}^k $ is the space of admissible words of length $ k $ (see Equation \ref{Sigmak}) depending on $ b $, the width of the transverse section $ S $.
Each curve $ \gamma_{\omega} \in \mathcal{N}_{0,1} \cap S $ corresponds to a word $ \omega $ whose length $ |\omega| $ is the level of $ \gamma_{\omega} $.
Define the space of all finite admissible words as
\begin{equation}
\label{Sigmab}
\Sigma_b = \bigcup_{k=0}^{\infty} \Sigma_{b,k},
\end{equation}
where $ \Sigma_{b,0} $ is a singleton (because there is only one curve of level zero, namely $ \gamma $).
Referring to Equation \ref{Sigmak}, a word $ (i_1,\ldots,i_k) $ is in $ \Sigma_{b,k} $ only if $ (i_1,\ldots,i_{k-1}) $ is in $ \Sigma_{b,k-1} $.
By Definition \ref{extadm}, $ \Sigma_b $ satisfies the extension admissibility property, and thus is a general symbolic space as defined in Chapter \ref{Thermo}.
Substituting Equations \ref{Ok} and \ref{Sigmab} into Equation \ref{leveldecompNS2}, we obtain
\begin{equation}
\label{leveldecompgamma}
\mathcal{N}_{0,1} \cap S = \bigcup_{k \geq 0} \bigcup_{\omega \in \Sigma_{b,k}} \gamma_{\omega} = \bigcup_{\omega \in \Sigma_b} \gamma_{\omega}.
\end{equation}
From Equation \ref{MN0}, $ \mathcal{M}_0 $ is the orbit of the curve $ \gamma^u $ which is obtained by restricting the parametrization of $ \gamma $.
By restricting the parametrization in Equation \ref{leveldecompgamma} we obtain
\begin{equation}
\label{leveldecompgammau}
\mathcal{M}_{0,1} \cap S = \bigcup_{\omega \in \Sigma_b} \gamma_{\omega}^u.
\end{equation}
The faithful action of the pseudogroup $ \langle \Phi, \Theta \rangle $ on $ \mathcal{N}_{0,1} \cap S $ given in Equation \ref{kuppermutegamma} induces a faithful action on $ \Sigma_b $:
\begin{align}
\label{KupSigmak}
& \Phi : \Sigma_{b,k} \rightarrow \Sigma_{b,k} \qquad \Phi(i_1,\ldots,i_k) = (i_1,\ldots,i_k+1) \\
& \Theta : \Sigma_{b,k} \rightarrow \Sigma_{b,k+1} \qquad \Theta(i_1,\ldots,i_k) = (i_1,\ldots,i_k,1) \nonumber
\end{align}
For each $ 0 \leq c \leq b $ and each curve $ \gamma_c $ in the vertical foliation of $ S $, we have a similar level decomposition of $ \mathcal{N}_{c,1} \cap S $ as a collection of curves coded by a smaller space $ \Sigma_c $ of admissible words.
Together, this gives a level decomposition of $ \bigcup_{t \geq 0} \psi_t(S) \cap S $ in terms of the sets $ A_{\omega} $.
\begin{equation}
\label{leveldecompS}
\bigcup_{-\infty < t < \infty} \psi_t(S) \cap S = \bigcup_{\omega \in \Sigma_b} A_{\omega}.
\end{equation}
\subsection{Dual symbolic dynamics}
\label{dualsymb}
In Chapter \ref{dualcant} we introduced the dual $ \widetilde{\Sigma} $ of a symbolic space $ \Sigma $.
In this section we will compute the admissible words in the dual space $ \widetilde{\Sigma}_{b,k} $.
We first recall the conventions; if $ \omega = (i_1,\ldots,i_k) \in \Sigma_{b,k} $ is an admissible word, then we denote its dual by $ \widetilde{\omega} = (i_k,\ldots,i_1) $.
For any $ k \geq 1 $, the dual of $ \Sigma_{b,k} $ is
$$
\widetilde{\Sigma}_{b,k} = \{ \widetilde{\omega} : \omega \in \Sigma_{b,k} \}.
$$
By Equation \ref{Sigmak},
\begin{equation}
\label{dualSigmak}
\widetilde{\Sigma}_{b,k} = \bigcup_{i_1=N_b}^{\infty} \bigcup_{i_2=N_b}^{M_{i_1}} \cdots \bigcup_{i_k=N_b}^{M_{i_1,\ldots,i_{k-1}}} (i_k,\ldots,i_1).
\end{equation}
The space of all finite dual words is
\begin{equation}
\label{dualSigmab}
\widetilde{\Sigma}_b = \bigcup_{k=0}^{\infty} \widetilde{\Sigma}_{b,k},
\end{equation}
For every $ \omega \in \Sigma_b $ there is a corresponding curve $ \gamma_{\omega} $.
The curve dual to $ \gamma_{\omega} $ is $ \widetilde{\gamma}_{\omega} = \gamma_{\widetilde{\omega}} $.
From the action of $ \langle \Phi, \Theta \rangle $ on $ \Sigma_{b,k} $ shown in Equation \ref{KupSigmak}, we obtain an obviously defined action on $ \widetilde{\Sigma}_b $.
Also, the nesting property for curves $ \gamma $ given in Proposition \ref{nestingk} implies a nesting property for dual curves $ \widetilde{\gamma} $.
\begin{proposition}
\label{dualnestingk}
For each $ (i_1,\ldots,i_k) \in \widetilde{\Sigma}_{b,k} $, the level-$k$ curve $ \gamma_{i_1,\ldots,i_k} $ is nested in the level-$(k-1)$ curve $ \gamma_{i_1,\ldots,i_{k-1}} $.
\end{proposition}
Finally, recall the sets $ A_{\omega} $ coded by $ \omega \in \Sigma_b $ introduced in Chapter \ref{symbrect}.
For each set $ A_{\omega} $ there is a corresponding dual set $ \widetilde{A}_{\omega} = A_{\widetilde{\omega}} $.
These dual sets satisfy a nesting property similar to that in Proposition \ref{nestingstrip}.
\begin{proposition}
\label{dualnestingstrip}
For each $ (i_1,\ldots,i_k) \in \widetilde{\Sigma}_{b,k} $ we have
$$
A_{i_1,\ldots,i_k} \subset A_{i_1,\ldots,i_{k-1}}.
$$
\end{proposition}
\vfill
\eject
\section{Transverse dynamics}
\label{Transversal}
In the previous chapter we coded the curves in the intersection of the surface $ \mathcal{N}_{0,1} $ with the transverse rectangle $ S $, and studied the pseudogroup dynamics on this intersection.
In this chapter we choose a one-dimensional transversal in $ S $, and study the induced pseudogroup dynamics on its intersection with $ \mathcal{N}_{0,1} $.
Our choice of transversal is the upper boundary $ S^+ $ of $ S $, as defined in Section \ref{InsertAssume}:
\begin{equation}
\label{Sp}
S^+ = \{ (r,\beta,-1+R): 0 \leq r-2 \leq b \}.
\end{equation}
Note that $ S^+ $ can be identified with $ [0,b] $.
We will introduce the \textit{transverse distances} of the curves $ \gamma_{\omega} $ measured along $ S^+ $.
Then we will use the parametrizations of the curves derived in Section \ref{Kupmin} to asymptotically estimate these transverse distances.
These will be important for later estimates of the Hausdorff dimension of the minimal set.
\subsection{The transverse set $ \mathcal{N}_{0,1} \cap S^+ $}
Recall from Equation \ref{gammak} and the remarks afterwards that for each $ k \geq 1 $ and $ \omega \in \Sigma_{b,k} $ there exist unique $ s_{\omega}^{\pm} $ with $ s_{\omega}^- < 0 < s_{\omega}^+ $ such that $ q_{\omega}(s_{\omega}^{\pm}) = R $.
By the definition of $ S^+ $ above and the parametrizations of $ \gamma_{\omega} $ in Equation \ref{gammak}, this is equivalent to $ \gamma_{\omega}(s_{\omega}^{\pm}) \in S^+ $, so each curve $ \gamma_{\omega} $ has two unique points of intersection with $ S^+ $.
Because $ \gamma = \gamma^l \cup \gamma^u $ as defined in Equation \ref{gammauldef}, we see that for all $ k \geq 1 $ and each $ \omega \in \Sigma_{b,k} $, $ \gamma_{\omega}^l $ and $ \gamma_{\omega}^u $ each have one unique intersection point with $ S^+ $.
We define $ a_{\omega}^{\pm} $ as the radial distances of these points from the Reeb cylinder, measured along $ S^+ $.
In coordinates,
\begin{align}
\label{awdef}
\gamma_{\omega}^u \cap S^+ = (2+a_{\omega}^-, \beta, -1+R) \\
\gamma_{\omega}^l \cap S^+ = (2+a_{\omega}^+, \beta, -1+R). \nonumber
\end{align}
With this choice, it is easy to see from the parametrization in Equation \ref{gammak} that $ a_{\omega}^- < a_{\omega}^+ $ for each $ \omega $.
From Equation \ref{leveldecompgamma} we have
\begin{equation}
\label{leveldecompgamma1}
\mathcal{N}_{0,1} \cap S^+ = \bigcup_{\omega \in \Sigma_b} a_{\omega}^{\pm},
\end{equation}
and by Equation \ref{leveldecompgammau} we have
\begin{equation}
\label{transverse1}
\mathcal{M}_{0,1} \cap S^+ = \bigcup_{\omega \in \Sigma_b} a_{\omega}^-.
\end{equation}
From the parametrization in Equation \ref{gammak},
\begin{equation}
\label{a}
a_{i_1,\ldots,i_k}^{\pm} = \left(s_{i_1,\ldots,i_k}^{\mp}\right)^2+\sum_{j=1}^{k-1} q_{i_1, \ldots,i_j}^2\left(s_{i_1,\ldots,i_k}^{\mp}\right)
\end{equation}
for each $ \omega = (i_1,\ldots,i_k) \in \Sigma_{b,k} $.
See Figure \ref{transa} for a picture of $ a_{\omega}^{\pm} $ for words $ \omega \in \Sigma_{b,1} $ of level one.
\begin{figure}[h!]
\includegraphics[width=0.9\linewidth, trim={6cm 7.1cm 1cm 6cm}, clip]{PicTemplate24}
\caption{The points $ a_i^{\pm} $ as intersections of the level-one curves $ \gamma_i $ with the upper boundary $ S^+ $ of $ S $. In this case $ N_b = 2 $, the minimal value of $ i $ such that $ q_i(s) = R $ has a solution.}
\label{transa}
\end{figure}
In Figures \ref{levelone} and \ref{leveltwo} it appears that $ \gamma_{\omega} $ becomes radially narrower as $ |\omega| \rightarrow \infty $, as does $ \gamma_{i,\omega} $ as $ i \rightarrow \infty $ for $ \omega $ fixed.
In the next section we measure the asymptotics of these widths more precisely.
\subsection{Transverse distances}
Define the function $ a: \Sigma_b \rightarrow \mathbb{R}^+ $ by
$$
a(\omega) = |a_{\omega}^+ - a_{\omega}^-|.
$$
This function gives the transverse width of the curve $ \gamma_{\omega} $ measured along $ S^+ $.
We say that $ \{ a(\omega) \}_{\omega \in \Sigma_{b,k}} $ are the \textit{transverse distances} of level $ k $.
We will now estimate the transverse distances of each level.
\subsubsection{Transverse distances of level one}
By Equation \ref{a}, we have
\begin{equation}
\label{ai}
a(i) = \left| (s_i^+)^2 - (s_i^-)^2 \right|,
\end{equation}
where $ s_i^{\pm} $ are the unique solutions to $ q_i(s) = R $.
Recall the constants $ C $ and $ K $ from Equation \ref{defCK}.
\begin{proposition}
\label{trans1}
For all $ \delta>0 $ there exists $ L_1 \in \mathbb{N} $ such that for all $ i \geq L_1 $,
$$
\left| a(i)-\left(\frac{\pi^{-1} K^{\frac{3}{2}}}{i^{\frac{5}{2}}} \right) \right| < \frac{\delta}{i^2}.
$$
\end{proposition}
\begin{proof}
Using the parametrization given in Equation \ref{gamma1}, the equation $ q_i(s) = R $ is equivalent to $ f_i(s) = 0 $, where
$$
f_i(s) = 2\pi C + s +\frac{4K}{s^2} \tan^{-1} \left( \frac{R}{s^2} \right) - 2\pi i.
$$
So $ s_i^{\pm} $ are the unique roots of $ f_i $.
We now claim that for any $ \delta>0 $ there exists $ N \in \mathbb{N} $ such that for all $ i \geq N $,
$$
s_i^+, -s_i^- \in \left[ \sqrt{ \frac{K(1-\delta)}{i} } , \sqrt{ \frac{K(1+\delta)}{i-1} }, \; \right].
$$
We will prove this for $ s_i^+ $; the proof for $ -s_i^- $ is identical.
First, restrict parameter values $ s $ to the interval
$$
\sqrt{ \frac{K(1-\delta)}{i-C} } < s < \sqrt{ \frac{K(1+\delta)}{i-1-C} }.
$$
We will show that $ f_i $ has a root on this interval; by uniqueness it must be $ s_i^+ $.
Notice as $ i \rightarrow \infty $ that $ s \searrow 0 $ on this interval, so for large enough $ i $, $ \displaystyle \tan^{-1}\left(\frac{R}{s^2}\right) \sim \frac{\pi}{2} $.
From this, we can show for sufficiently large $ i $ that $ f_i^-(s) < f_i(s) < f_i^+(s) $ for all $ s $ on this interval, where
$$
f_i^{\pm}(s) = 2\pi C + s +\frac{2\pi K (1\pm \delta)}{s^2} - 2\pi i.
$$
Note that $ f_i^{\pm} $ are monotonically decreasing, and that
$$
f_i^- \left( \sqrt{\frac{K(1-\delta)}{i-C}} \right)>0, \quad \text{ and } \quad f_i^+ \left( \sqrt{\frac{K(1+\delta)}{i-1-C}} \right) < 0,
$$
so $ f_i $ must have a root on this interval and we obtain the desired bounds on $ s_i^+ $, after absorbing $ C $ into the constant $ N $.
As an immediate corollary, notice that for any $ \delta>0 $ and sufficiently large $ i $,
\begin{equation}
\label{sisi}
\frac{2\sqrt{K(1-\delta)}}{i^{\frac{1}{2}}} < \left| s_i^+ - s_i^- \right| < \frac{2\sqrt{K(1+\delta)}}{i^{\frac{1}{2}}}
\end{equation}
We now turn to the proof of the proposition.
Substituting the equations $ f_i(s_i^{\pm}) = 0 $ into Equation \ref{ai} yields
$$
a(i) = \left| \frac{4K}{2\pi (i-C) - s_i^+} \tan^{-1} \left(\frac{R}{(s_i^+)^2}\right) - \frac{4K}{2\pi(i-C) - s_i^-} \tan^{-1} \left(\frac{R}{(s_i^-)^2}\right) \right|.
$$
It is easy to show using the parametrization in Equation \ref{gamma1} that $ (s_i^+)^2 > (s_i^-)^2 $.
Applying this to the above expression for $ a(i) $ we obtain
$$
\tan^{-1} \left(\frac{R}{(s_i^+)^2}\right) < \frac{a(i)}{\left|\frac{4K}{2\pi (i-C) - s_i^+} - \frac{4K}{2\pi(i-C) - s_i^-}\right|} < \tan^{-1} \left(\frac{R}{(s_i^-)^2}\right).
$$
In light of the bounds we established on $ s_i^{\pm} $ we know that $ s_i^{\pm} \rightarrow 0 $ and therefore that
$$
\tan^{-1} \left(\frac{R}{(s_i^{\pm})^2}\right) \rightarrow \frac{\pi}{2} $$
as $ i \to \infty $.
Then for any $ \delta>0 $,
$$
\frac{K(1-\delta)}{2\pi i^2}\:|s_i^+-s_i^-| < a(i) < \frac{K(1+\delta)}{2\pi i^2} \: |s_i^+-s_i^-|
$$
for sufficiently large $ i $. Combining this with Equation \ref{sisi}, we obtain the desired result.
\end{proof}
From this proof we deduce the following corollary.
\begin{corollary}
\label{limita1}
The following limit exists
$$
\lim_{i \to \infty} a_i^- = 0.
$$
\end{corollary}
\begin{proof}
By Equation \ref{a}, $ a_i^- = (s_i^+)^2 $. From the proof of Proposition \ref{trans1}, $ \lim_{i \to \infty} s_i^+ = 0 $.
\end{proof}
Corollary \ref{limita1} is analytic confirmation of one of the heuristic facts evident in Figure \ref{levelone}; that the level-one curves $ \gamma_i $ limit on the Reeb cylinder $ r=2 $ as $ i \to \infty $.
From this and the nesting properties, we will later deduce that the level-two curves limit on the level-one curves, and inductively that the level-$ k $ curves limit on the level-$ (k-1) $ curves.
\subsubsection{Transverse distances of level two}
By Equation \ref{a}, for all $ (i,j) \in \Sigma_{b,2} $ we have
\begin{equation}
\label{aij}
a(i,j) = \left| (s_{i,j}^+)^2+q_i^2(s_{i,j}^+) - (s_{i,j}^-)^2 - q_i^2(s_{i,j}^-) \right|,
\end{equation}
where $ s_{i,j}^{\pm} $ are the unique solutions to $ q_{i,j}(s) = R $.
\begin{proposition}
\label{trans2}
For all $ \delta>0 $ there exists $ L_2 \in \mathbb{N} $ such that for all $ (i,j) \in \Sigma_{b,2} $ with $ i,j \geq L_2 $,
$$
\left| a(i,j)- \left(\frac{\pi^{-1} K^{\frac{3}{2}}}{j^{\frac{5}{2}}} \cdot \frac{(2\pi)^{-2} aR^2}{i^2} \right)\right| < \frac{\delta}{i^2 j^2}.
$$
\end{proposition}
\begin{proof}
Using the parametrization given in Equation \ref{gamma2}, the equation $ q_{i,j}(s) = R $ is equivalent to $ f_{i,j}(s) = 0 $, where
$$
f_{i,j}(s) = 2\pi C+ q_i(s) + \frac{4K}{s^2+q_i^2(s)} \tan^{-1} \left( \frac{R}{s^2+q_i^2(s)} \right) -2\pi j.
$$
The unique roots of $ f_{i,j} $ are $ s_{i,j}^{\pm} $.
Recall that $ \lim_{s \to 0} q_i(s) = v_i $ by definition and that $ \lim_{i \to \infty} v_i = 0 $.
Applying this fact to the above expression for $ f_{i,j} $ and using a method similar to that in the proof of Proposition \ref{trans1}, we can show for any $ \delta $ that
$$
s_{i,j}^+, -s_{i,j}^- \in \left[ \sqrt{\frac{K(1-\delta)}{j+1-C}}, \sqrt{\frac{K(1+\delta)}{j-1-C}} \right]
$$
for sufficiently large $ i,j $.
As a corollary we obtain that
\begin{equation}
\label{sij-sij}
\frac{2\sqrt{K(1-\delta)}}{j^{\frac{1}{2}}} < | s_{i,j}^+ - s_{i,j}^- | < \frac{2\sqrt{K(1+\delta)}}{j^{\frac{1}{2}}}
\end{equation}
for sufficently large $ i,j $.
We now claim that for small enough parameter values $ s $ we have
\begin{equation}
\label{qibound}
\frac{-aR^2-\frac{\delta}{i^2}}{2\pi(i-C)-s +aR+\delta} < q_i(s) < \frac{-aR^2+\frac{\delta}{i^2}}{2\pi(i-C)-s +aR-\delta},
\end{equation}
for sufficiently large $ i $.
To prove this we use the parametrization from Equation \ref{gamma1}:
$$
q_i(s) = \frac{s^2 \tan\left(\frac{s^2}{R^2}T_i(s)\right)-R}{1+\frac{R}{s^2}\tan\left(\frac{s^2}{R^2}T_i(s)\right)}.
$$
For small $ x $, $ \tan x \sim x $. Then for any $ \delta>0 $ we have
$$
\frac{-R-\frac{\delta}{i}}{1+\frac{1}{R}T_i(s)+\delta} < q_i(s) < \frac{-R+\frac{\delta}{i}}{1+\frac{1}{R}T_i(s)-\delta}
$$
for large enough $ i $ and small enough $ s $.
Multiplying the top and bottom of each fraction by $ aR $, substituting $ a T_i(s) = 2\pi(i-C)-s $ from Equation \ref{gamma1} and re-scaling $ \delta $ we obtain the desired bounds.
As an immediate corollary we obtain for any $ \delta>0 $ and small enough parameter values $ u,v $ that
\begin{equation}
\label{qi-qi}
\frac{(2\pi)^{-2} aR^2}{i^2}\: |u-v|-\frac{\delta}{i^2} < |q_i(u)-q_i(v)| < \frac{(2\pi)^{-2} aR^2}{i^2}\: |u-v|+\frac{\delta}{i^2}
\end{equation}
for large enough $ i $.
We now turn to the proof of the proposition.
Substituting the equations $ f_{i,j}(s_{i,j}^{\pm}) = 0 $ into Equation \ref{aij} yields
{\small
$$
a(i,j)=\left| \frac{4K}{2\pi (j-C) - q_i(s_{i,j}^+)} \tan^{-1} \frac{R}{(s_{i,j}^+)^2+q_i^2(s_{i,j}^+)} - \frac{4K}{2\pi(j-C) - q_i(s_{i,j}^-)} \tan^{-1} \frac{R}{(s_{i,j}^-)^2+q_i^2(s_{i,j}^-)} \right|,
$$
}
and since $ (s_{i,j}^+)^2 > (s_{i,j}^-)^2 $ this implies
{\small
$$
\tan^{-1} \left(\frac{R}{(s_{i,j}^+)^2+q_i^2(s_{i,j}^+)}\right) < \frac{a(i,j)}{\left|\frac{4K}{2\pi (j-C) - q_i(s_{i,j}^+)} - \frac{4K}{2\pi(j-C) - q_i(s_{i,j}^-)}\right|} < \tan^{-1} \left(\frac{R}{(s_{i,j}^-)^2+q_i^2(s_{i,j}^-)}\right).
$$
}
But from the bounds we established on $ s_{i,j}^{\pm} $ we know that $ s_{i,j}^{\pm} \rightarrow 0 $ and thus that
$$
\tan^{-1} \left(\frac{R}{(s_{i,j}^{\pm})^2+q_i^2(s_{i,j}^{\pm})}\right) \rightarrow \frac{\pi}{2}
$$
as $ i,j \rightarrow \infty $.
Then we can eliminate the inverse tangent terms and simplify to
$$
\frac{K(1-\delta)}{2\pi j^2} \: \left| q_i(s_{i,j}^+)-q_i(s_{i,j}^-) \right| < a(i,j) < \frac{K(1+\delta)}{2\pi j^2} \: \left| q_i(s_{i,j}^+)-q_i(s_{i,j}^-) \right|.
$$
Substituting in Equation \ref{qi-qi}, we improve the bounds to
{\small
$$
\frac{K(1-\delta)}{2\pi j^2} \: \left( \frac{(2\pi)^{-2} aR^2}{i^2}\: | s_{i,j}^+ - s_{i,j}^- |-\frac{\delta}{i^2} \right) < a(i,j) < \frac{K(1+\delta)}{2\pi j^2} \: \left( \frac{(2\pi)^{-2} aR^2}{i^2}\:| s_{i,j}^+ - s_{i,j}^- |+\frac{\delta}{i^2} \right).
$$
}
Finally, we substitute in Equation \ref{sij-sij} and re-scale $ \delta $ to obtain the desired bounds.
\end{proof}
Using the estimates in the above proof, we now show that the level-two points $ a_{i,j} $ limit on the level-one points $ a_i $ in the following way.
\begin{corollary}
\label{limita2}
For $ (i,j) \in \Sigma_{b,2} $ the limit exists
$$
\lim_{i,j \to \infty} a_{i,j}^- = 0,
$$
and for $ j $ sufficiently large,
$$
\lim_{i \to \infty} a_{i,j}^- = a_j^-.
$$
\end{corollary}
\begin{proof}
By Equation \ref{a},
$$
a_{i,j}^- = (s_{i,j}^+)^2 + q_i^2(s_{i,j}^+).
$$
From the proof of Proposition \ref{trans1} we know that $ \lim_{i,j \to \infty} s_{i,j}^+ = 0 $.
Using this, furthermore we have
$$
\lim_{i,j \to \infty} q_i(s_{i,j}^+) = \lim_{i \to \infty} v_i = 0,
$$
which proves the first statement.
To prove the second statement, we first claim that for a sufficiently large $ j $,
$$
\lim_{i \to \infty} s_{i,j}^+ = s_j^+.
$$
To prove this, recall from the proof of Proposition \ref{trans1} that $ s_j^+ $ is the unique root of $ f_j $ on $ s>0 $, so
$$
f_j(s_j^+) = 2\pi C + s_j^+ +\frac{4K}{(s_j^+)^2} \tan^{-1} \left( \frac{R}{(s_j^+)^2} \right) - 2\pi j = 0,
$$
and from the proof of Proposition \ref{trans2} that $ s_{i,j}^+ $ is the unique root of $ f_{i,j} $, so
$$
f_{i,j}(s_{i,j}^+) = 2\pi C+ q_i(s_{i,j}^+) + \frac{4K}{(s_{i,j}^+)^2+q_i^2(s_{i,j}^+)} \tan^{-1} \left( \frac{R}{(s_{i,j}^+)^2+q_i^2(s_{i,j}^+)} \right) -2\pi j=0.
$$
For sufficiently large $ j $, $ \lim_{i \to \infty} q_i(s_{i,j}^+) =0 $ from the proof of Proposition \ref{trans2}.
Using this and comparing the above parametrizations, we see that $ \lim_{i \to \infty} s_{i,j}^+ $ is a root of $ f_j $ for sufficiently large $ j $.
Since the root of $ f_j $ is unique on $ s>0 $ and equals $ s_j^+ $, we obtain the desired result.
\end{proof}
\subsubsection{Transverse distances of level $ k $}
\begin{proposition}
\label{transk}
For all $ \delta>0 $ there exists $ L_k \in \mathbb{N} $ such that for all $ (i_1,\ldots,i_k) \in \Sigma_{b,k} $ with $ i_1,\ldots,i_k \geq L_k $,
$$
\left| a(i_1,\ldots,i_k)- \left(\frac{\pi^{-1} K^{\frac{3}{2}}}{i_k^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{k-1}}{i_1^2 \cdots i_{k-1}^2} \right)\right| < \frac{\delta}{i_1^2 \cdots i_k^2}.
$$
\end{proposition}
\begin{proof}
By the parametrization given in Equation \ref{gammak}, the equation $ q_{i_1,\ldots,i_k}(s) = R $ is equivalent to $ f_{i_1,\ldots,i_k}(s) = 0 $, where
{\small
$$
f_{i_1,\ldots,i_k}(s) = 2\pi C+q_{i_1,\ldots,i_{k-1}}(s) + \frac{4K}{s^2+\sum_{j=1}^{k-1} q_{i_1, \ldots,i_j}^2(s)} \tan^{-1} \left(\frac{R}{s^2+\sum_{j=1}^{k-1} q_{i_1, \ldots,i_j}^2(s)} \right)- 2\pi i_k,
$$
}
Recall that $ \lim_{s \to 0} q_{i_1,\ldots,i_{k-1}}(s) = v_{i_1,\ldots,i_{k-1}} $ and that $ \lim_{i_1,\ldots,i_{k-1} \to \infty} v_{i_1,\ldots,i_{k-1}} =0 $.
Applying this to the above expression for $ f_{i_1,\ldots,i_k} $ we can show that
$$
s_{i_1,\ldots,i_k}^+, -s_{i_1,\ldots,i_k}^- \in \left[ \sqrt{\frac{K(1-\delta)}{i_k+1-C}}, \sqrt{\frac{K(1+\delta)}{i_k-1-C}} \right]
$$
for sufficiently large $ i_1,\ldots,i_k $.
As a corollary we obtain that
\begin{equation}
\label{sik-sik}
\frac{2\sqrt{K(1-\delta)}}{i_k^{\frac{1}{2}}} < | s_{i_1,\ldots,i_k}^+ - s_{i_1,\ldots,i_k}^- | < \frac{2\sqrt{K(1+\delta)}}{i_k^{\frac{1}{2}}}
\end{equation}
for sufficently large $ i_1,\ldots,i_k $.
We now claim that for sufficiently small parameter values $ s $ and sufficiently large values of $ i_1,\ldots,i_k $ we have
{\small
$$
\frac{-aR^2-\frac{\delta}{i_1^2 \cdots i_k^2}}{2\pi(i_k-C)-q_{i_1,\ldots,i_{k-1}}(s)+aR+\delta} < q_{i_1,\ldots,i_k}(s) < \frac{-aR^2+\frac{\delta}{i_1^2 \cdots i_k^2}}{2\pi(i_k-C)-q_{i_1,\ldots,i_{k-1}}(s)+aR-\delta},
$$
}
for sufficiently large $ i_1,\ldots,i_k $.
The proof of this uses the expression for $ q_{i_1,\ldots,i_k} $ in terms of $ q_{i_1,\ldots,i_{k-1}} $ given in Equation \ref{gammak}, together with precisely the same method of proof as the corresponding claim in the proof of Proposition \ref{trans2}.
As a corollary, we have for sufficiently large $ i_1, \ldots, i_k $ and small enough parameter values $ u,v $ that
{\small
\begin{multline}
\frac{(2\pi)^{-2} aR^2}{i_k^2}\: |q_{i_1,\ldots,i_{k-1}}(u)-q_{i_1,\ldots,i_{k-1}}(v)|-\frac{\delta}{i_1^2 \cdots i_k^2} < |q_{i_1,\ldots,i_k}(u)-q_{i_1,\ldots,i_k}(v)| \nonumber \\
< \frac{(2\pi)^{-2} aR^2}{i_k^2}\: |q_{i_1,\ldots,i_{k-1}}(u)-q_{i_1,\ldots,i_{k-1}}(v)|+\frac{\delta}{i_1^2 \cdots i_k^2}.
\end{multline}
}
From this recursive expression we can prove by induction on $ k $ for sufficiently large $ i_1,\ldots,i_k $ and small $ u,v $ that
{\small
\begin{equation}
\label{qikinduct}
\frac{\left((2\pi)^{-2} aR^2\right)^k}{i_1^2 \cdots i_k^2}\: |u-v|-\frac{\delta}{i_1^2 \cdots i_k^2} < |q_{i_1,\ldots,i_k}(u)-q_{i_1,\ldots,i_k}(v)| \\
< \frac{\left((2\pi)^{-2} aR^2\right)^k}{i_1^2 \cdots i_k^2}\: |u-v|+\frac{\delta}{i_1^2 \cdots i_k^2}
\end{equation}
}
We now turn to the proof of the proposition.
Substituting the equations $ f_{i_1,\ldots,i_k}(s_{i_1,\ldots,i_k}^{\pm}) = 0 $ into Equation \ref{a} and eliminating the inverse tangent terms as in the proof of Proposition \ref{trans2} yields
{\small
\begin{multline}
\frac{K(1-\delta)}{2\pi i_k^2} \: \left| q_{i_1,\ldots,i_{k-1}}(s_{i_1,\ldots,i_k}^+)-q_{i_1,\ldots,i_{k-1}}(s_{i_1,\ldots,i_k}^-) \right| < a(i_1,\ldots,i_k) \nonumber \\
< \frac{K(1+\delta)}{2\pi i_k^2} \: \left| q_{i_1,\ldots,i_{k-1}}(s_{i_1,\ldots,i_k}^+)-q_{i_1,\ldots,i_{k-1}}(s_{i_1,\ldots,i_k}^-) \right| \nonumber.
\end{multline}
}
Substituting Equation \ref{qikinduct} we obtain
{\small
\begin{multline}
\frac{K(1-\delta)}{2\pi i_k^2} \: \left( \frac{\left((2\pi)^{-2} aR^2\right)^{k-1}}{i_1^2 \cdots i_{k-1}^2}\: |u-v|-\frac{\delta}{i_1^2 \cdots i_{k-1}^2} \right) < a(i_1,\ldots,i_k) \nonumber \\
< \frac{K(1+\delta)}{2\pi i_k^2} \: \left( \frac{\left((2\pi)^{-2} aR^2\right)^{k-1}}{i_1^2 \cdots i_{k-1}^2}\: |u-v|+\frac{\delta}{i_1^2 \cdots i_{k-1}^2} \right) \nonumber.
\end{multline}
}
Finally, we substitute in Equation \ref{sik-sik} and re-scale $ \delta $ to obtain the desired bounds.
\end{proof}
The proof of the following corollary is a straightforward generalization of the proof of Corollary \ref{limita2}.
\begin{corollary}
\label{limitak}
For $ (i_1,\ldots,i_k) \in \Sigma_{b,k} $ the limit exists
$$
\lim_{i_1,\ldots,i_k \to \infty} a_{i_1,\ldots,i_k}^- = 0,
$$
and for $ \omega = (i_1,\ldots,i_k) $ with $ i_1,\ldots,i_k $ sufficiently large,
$$
\lim_{j \to \infty} a_{j, \omega}^- = a_{\omega}^-.
$$
\end{corollary}
\subsection{The projection action}
\label{projaction}
In Chapter \ref{Kupmin}, we exhibited a faithful action of $ \Psi_1 = \langle \Phi, \Theta \rangle $ on $ \mathcal{N}_{0,1} \cap S $.
This does not restrict to an action on the transversal $ \mathcal{N}_{0,1} \cap S^+ $, because the maps $ \Phi $ and $ \Theta $ do not preserve $ S^+ $.
Nevertheless, we will show that the $ \langle \Phi, \Theta \rangle $ acts faithfully on $ \mathcal{N}_{0,1} \cap S^+ $ in this section.
By Theorem \ref{minchar}, $ \mathcal{N} \cap S $ is a codimension-one lamination in $ S $, and $ \mathcal{N}_{0,1} \cap S $ is a collection of leaves of this lamination.
By Equation \ref{leveldecompgamma}, the leaves in $ \mathcal{N}_{0,1} \cap S $ are the collection of curves $ \gamma_{\omega} $ indexed by $ \omega \in \Sigma_b $.
To obtain a faithful action of $ \langle \Phi, \Theta \rangle $ on $ \mathcal{N}_{0,1} \cap S^+ $, we will project to $ S^+ $ along these leaves.
We will call this the \textit{projection action} of $ \langle \Phi, \Theta \rangle $ on $ S^+ $.
To define this action, we will first define the projection maps along the leaves.
\subsubsection{The projection maps}
By Equation \ref{awdef}, each curve $ \gamma_{\omega} $ has two unique intersections with $ S^+ $, whose radial coordinates are $ 2+a_{\omega}^{\pm} $.
Furthermore, each $ \gamma_{\omega} $ has a vertex $ v_{\omega} = \gamma_{\omega}(0) $, as defined in Section \ref{Kupmin}.
For each $ \omega \in \Sigma_b $ we have maps
\begin{equation}
\label{pmaps}
p_{\omega}^{\pm} : a_{\omega}^{\pm} \mapsto v_{\omega},
\end{equation}
and each map $ p_{\omega}^{\pm} $ has a well-defined inverse.
Each leaf $ \gamma_{\omega} $ is the intersection with $ S $ of the orbit of the smooth curve $ \gamma $ under the $ C^{\infty} $ flow $ \psi_t $.
The surface $ S $ is transverse to the flow, so each $ \gamma_{\omega} $ is a $ C^{\infty} $ submanifold of codimension one in $ S $.
Each $ \gamma_{\omega} $ is covered by a finite number of charts of the lamination in $ S $, and each map $ p_{\omega}^{\pm} $ is a finite composition of transition maps of these charts, which are $ C^{\infty} $.
As a consequence, the maps $ p_{\omega}^{\pm} $ are in the holonomy of this lamination and are smooth projections along its leaves.
See Figure \ref{projfig} for a picture of the projection along curves $ \gamma_i $ of level one.
\begin{figure}[h]
\includegraphics[width=1.1\linewidth, trim={4.5cm 8.5cm 0 5.8cm}, clip]{PicTemplate31}
\caption{The projection maps $ p_2^{\pm} $ projecting the points $ a_2^{\pm} $ along the curve $ \gamma_2 $ to its vertex $ v_2 $.}
\label{projfig}
\end{figure}
\subsubsection{The projection action}
Notice that the vertices $ \bigcup_{\omega \in \Sigma_b} v_{\omega} $ are preserved by $ \langle \Phi, \Theta \rangle $.
By Equation \ref{KupSigmak}, the action on the vertices is
\begin{align}
\label{projvert}
\Phi(v_{i_1,\ldots,i_k}) &= v_{i_1,\ldots,i_k+1} \\
\Theta(v_{i_1,\ldots,i_k}) &= v_{i_1,\ldots,i_k,1}. \nonumber
\end{align}
To define the projection action of $ \langle \Phi, \Theta \rangle $ on $ \mathcal{N}_0 \cap S^+ = \bigcup_{\omega \in \Sigma_b} a_{\omega}^{\pm} $, we conjugate the above action by the projection maps.
\begin{align*}
\Phi \cdot a_{i_1,\ldots,i_k}^{\pm} &= \left( p_{i_1,\ldots,i_k+1}^{\pm} \right)^{-1} \Phi \: p_{i_1,\ldots,i_k}^{\pm} (a_{i_1,\ldots,i_k}^{\pm}) \\
\Theta \cdot a_{i_1,\ldots,i_k}^{\pm} &= \left( p_{i_1,\ldots,i_k,1}^{\pm} \right)^{-1} \Theta \: p_{i_1,\ldots,i_k}^{\pm} (a_{i_1,\ldots,i_k}^{\pm}) \nonumber
\end{align*}
Combining this with Equations \ref{pmaps} and \ref{projvert} we see that
\begin{align}
\label{paction}
\Phi \cdot a_{i_1,\ldots,i_k}^{\pm} &= a_{i_1,\ldots,i_k+1}^{\pm} \\
\Theta \cdot a_{i_1,\ldots,i_k}^{\pm} &= a_{i_1,\ldots,i_k,1}, \nonumber
\end{align}
so the symbolic dynamics of this action on the transversal $ S^+ $ is the same as that of the action on the section $ S $ given in Equation \ref{kuppermutegamma}.
\subsection{Dual transverse distances}
\label{dualtrans}
In Section \ref{dualsymb} we defined the dual space $ \widetilde{\Sigma}_b $ and obtained nesting properties for the curves $ \gamma_{\omega} $ and sets $ A_{\omega} $ when $ \omega \in \widetilde{\Sigma}_b $.
Thus from any statement about $ a_{\omega}^{\pm} $ or $ a(\omega) $ we have a dual version of the statement.
For later use we record two such versions below.
The first is the dual version of Proposition \ref{transk}.
\begin{proposition}
\label{dualtransk}
For all $ \delta>0 $ there exists $ L_k \in \mathbb{N} $ such that for all $ (i_1,\ldots,i_k) \in \widetilde{\Sigma}_{b,k} $ with $ i_1,\ldots,i_k \geq L_k $,
$$
\left| a(i_1,\ldots,i_k)- \left(\frac{\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{k-1}}{i_2^2 \cdots i_k^2} \right)\right| < \frac{\delta}{i_1^2 \cdots i_k^2}.
$$
\end{proposition}
The second is the dual version of Corollary \ref{limitak}.
\begin{corollary}
\label{duallimitak}
For $ (i_1,\ldots,i_k) \in \widetilde{\Sigma}_{b,k} $ the limit exists
$$
\lim_{i_1,\ldots,i_k \to \infty} a_{i_1,\ldots,i_k}^- = 0,
$$
and for $ \omega = (i_1,\ldots,i_k) $ with $ i_1,\ldots,i_k $ sufficiently large,
$$
\lim_{j \to \infty} a_{\omega,j}^- = a_{\omega}^-,
$$
and by induction,
$$
\lim_{j_{k+1},\ldots,j_{k+n} \to \infty} a_{\omega, j_{k+1}, \ldots,j_{k+n}} = a_{\omega}^-
$$
for any $ n \geq 1 $.
\end{corollary}
\vfill
\eject
\section{$ C^{1+\alpha} $ function systems on the transversal}
\label{FunctionSystems}
In this chapter we use the pseudogroup $ \Psi_1 = \langle \Phi, \Theta \rangle $ and the projection maps from Chapter \ref{projaction} to define a function system on the transversal $ S^+ $.
By Equation \ref{Sp}, $ S^+ $ can be identified with $ [0,b] $ via the map
$$
(r, \beta, -1+R) \mapsto r-2.
$$
In this coordinate system, Equation \ref{awdef} reads simply
$$
\gamma_{\omega} \cap S^+ = a_{\omega}^{\pm},
$$
so for ease of notation we will frequently use this coordinate system.
The function system we will define will be a $ C^{1+\alpha} $ general function system on $ [0,b] $ modeled by a general symbolic space in the sense of Chapter \ref{genfun}.
Furthermore, we will prove that for sufficiently small $ \epsilon>0 $, this function system has a pseudo-Markov subsystem on $ [0,\epsilon] \subset [0,b] $, as studied in Chapter \ref{GDPM}.
These function systems will be related to the transverse Kuperberg minimal set in Chapter \ref{Transcant}.
\subsection{A $ C^{1+\alpha} $ function system on $ [0,b] $}
The domain of the projection maps defined in Chapter \ref{projaction} is $$
\bigcup_{\omega \in \Sigma_b} a_{\omega}^{\pm} \subset [0,b].
$$
To define a function system on $ [0,b] $, we will need to project along curves $ \gamma_{c,\omega} $ in the parabolic foliations of $ A_{\omega} $ studied in Chapter \ref{symbrect}.
\subsubsection{Extension of the projection maps}
Recall the foliation of $ S $ by vertical lines $ \{ \gamma_c \}_{0 \leq c \leq b} $ parametrized in Equation \ref{gammac}.
Then by Equation \ref{leveldecompc} and the subsequent remarks, we have a level decomposition
$$
\mathcal{N}_{c,1} \cap S = \bigcup_{\omega \in \Sigma_c} \gamma_{c,\omega}.
$$
For each $ \omega \in \Sigma_c $, the curve $ \gamma_{c,\omega} $ has intersection points $ a_{c,\omega}^{\pm} \in [0,b] $, and vertex $ v_{c,\omega} = \gamma_{c,\omega}(0) $.
We extend the projections $ p_{\omega}^{\pm} $ to projections $ p_{c,\omega}^{\pm} $ along the curves $ \gamma_{c,\omega} $ in the same way as in Equation \ref{pmaps}.
\begin{equation}
\label{pcmaps}
p_{c,\omega}^{\pm} : a_{c,\omega}^{\pm} \mapsto v_{c,\omega}
\end{equation}
\subsubsection{Preliminary steps}
The definition of a general function system modeled on a symbolic space, as given in Section \ref{genfun}, has some preliminary steps.
Namely, a compact space $ X \subset [0,1] $, a countable alphabet $ E $, and for each $ i \in E $ a $ C^{1+\alpha} $ map $ f_i : X \rightarrow X $ with Lipschitz constant $ < 1 $ and images $ \Delta_i = f_i(X) $ satisfying the separation property
$$
\Delta_i \cap \Delta_j = \emptyset \; \text{ when } \; i \neq j.
$$
In our setting, we let $ X = S^- \cup S^+ $, where $ S^{\pm} $ are the upper and lower boundaries of $ S $, from Equation \ref{Spm}.
Because $ S^{\pm} $ are both identified with $ [0,b] $, the space $ X $ is naturally identified with two disjoint copies of $ [0,b] $.
Let $ N_b \in \mathbb{N} $ be the constant defined in the proof of Proposition \ref{gamma1param}, and let
$$
E = \Sigma_{b,1} = \{ N_b, N_b+1, \ldots \}
$$
as defined in Equation \ref{Sigma1}.
To define $ f_i : X \rightarrow X $, we define $ f_i(c) $ for $ c $ on each interval $ S^{\pm} $ separately.
If $ c \in S^+ $ then by Equation \ref{gammac}, $ c= a_c^- $, the unique upper endpoint of $ \gamma_c $.
If $ c \in S^- $ then $ c= a_c^+ $, the unique lower endpoint of $ \gamma_c $.
We now define
\begin{equation}
\label{fi}
f_i(a_c^{\pm}) = (p_{c,i}^{\pm})^{-1} \: \Phi^{i-1} \: \Theta \: p_c^{\pm} (a_c^{\pm}).
\end{equation}
In words, $ f_i $ first projects $ a_c^{\pm} $ to the vertex $ v_c $ of $ \gamma_c $, then follows the orbit of $ v_c $ through the insertion to its first intersection $ \Theta(v_c) $ with $ S $.
It then follows the orbit of $ \Theta(v_c) $ to its its $ (i-1) $-th return to $ S $ under $ \Phi $.
By construction, this is the vertex $ v_{c,i} $ of $ \gamma_{c,i} $ which is then inversely projected back along $ \gamma_{c,i} $ to its intersection $ a_{c,i}^{\pm} $ with $ S^+ = [0,b] $.
For any $ i \in \Sigma_{b,1} $, recall from Chapter \ref{symbrect} that $ A_i = (\Phi^{i-1} \Theta) (S) $.
For each $ i $, the curves $ \{\gamma_{c,i} \}_c $ form a parabolic foliation of $ A_i $ (See figures \ref{stripfol} and \ref{transstripfig}).
From this and the definition of the extended projection maps given in Equation \ref{pcmaps}, we see that
$$
f_i(X) = A_i \cap S^+.
$$
Denote $ \Delta_i = f_i(X) $ and note that each $ \Delta_i $ is a closed interval.
Since $ \gamma_i \subset \partial A_i $, and $ a_i^{\pm} $ are the unique intersection points of $ \gamma_i $ with $ [0,b] $, we have
$$
\Delta_i = [a_i^-, a_i^+],
$$
so that $ |\Delta_i| = a(i) $, the transverse distances of level one studied in Section \ref{Transversal}.
See Figure \ref{adeltafig}.
\begin{figure}[h]
\includegraphics[width=1\linewidth, trim={5.5cm 6.8cm 0 6cm}, clip]{PicTemplate21}
\caption{The sets $ A_i $ and their intersection intervals $ \Delta_i $ with $ S^+ = [0,b] $.}
\label{adeltafig}
\end{figure}
We now show that $ f_i $ satisfies the properties we imposed in Section \ref{genfun}.
\begin{itemize}
\item \textit{Uniform contraction}:
For all $ i \in \Sigma_{b,1} $, the maps $ f_i $ have a uniform Lipschitz constant $ 0 < s < 1 $.
\vspace{0.1cm}
In fact, more is true.
First, note that as $ C^1 $ maps of a compact space, each $ f_i $ is individually Lipschitz by the mean value theorem.
Let $ c, c' \in S^+ $.
Then
$$
|f_i(c) - f_i(c')| \leq |\Delta_i| = a(i).
$$
By Proposition \ref{trans1}, $ a(i) \sim i^{-\frac{5}{2}} \rightarrow 0 $ as $ i \rightarrow \infty $.
So as $ i $ increases, the Lipschitz constant of $ f_i $ becomes arbitrarily small.
See Figure \ref{fifig} for a picture of this.
Thus setting $ s $ to be the Lipschitz constant of $ f_1 $ suffices for our purposes.
\vspace{0.2cm}
\begin{figure}[h]
\minipage{0.55\textwidth}
\hspace{-1.5cm}
\includegraphics[width=1.1\linewidth, trim={2cm 7.7cm 4cm 1.8cm}, clip]{PicTemplate32}
\caption*{Two points $ a_c^-, a_{c'}^- \in S^+ = [0,b] $ as endpoints of the vertical segments $ \gamma_c, \gamma_{c'} $.}
\endminipage
\minipage{0.55\textwidth}%
\hspace{-1cm}
\includegraphics[width=1.1\linewidth, trim={2cm 7cm 4cm 3.2cm}, clip]{PicTemplate33}
\caption*{The images $ a_{c,i}^-, a_{c',i}^- \in [a_i^-, a_i^+] $ under $ f_i $ of $ a_c^-, a_{c'}^- $, respectively.}
\endminipage
\caption{}
\label{fifig}
\end{figure}
\item \textit{$ C^{1+\alpha} $ regularity}: There exists $ \alpha > 0 $ such that for all $ i \in \Sigma_{b,1} $, the maps $ f_i $ have regularity $ C^{1+\alpha} $.
\vspace{0.1cm}
Recall from Section \ref{projaction} that the projection maps $ p_{\omega}^{\pm} $ are $ C^{\infty} $; this argument also holds for the maps $ p_{c,\omega}^{\pm} $.
The maps $ \Theta, \Phi $ are in the holonomy of the Kuperberg flow and as such are also $ C^{\infty} $.
By Equation \ref{fi}, the maps $ f_i $ are compositions of these and as such are $ C^{\infty} $.
By the mean value theorem, a $ C^{\infty} $ map of a compact interval has a uniform bound on its second derivative.
This demonstrates that for all $ i \in \Sigma_{b,1} $, $ f_i $ are uniformly $ C^{1+\alpha} $ for $ \alpha = 1 $.
\vspace{0.2cm}
\item \textit{Separation property}:
As $ \Delta_i = A_i \cap S^+ $, the sets $ \Delta_i $ are pairwise disjoint because the sets $ A_i $ are.
\vspace{0.2cm}
\end{itemize}
\subsubsection{The function system on $ [0,b] $}
In this section we will use the spaces $ \Delta_i $ defined above to define a general function system modeled by a symbolic space of infinite type, in the sense of Definition \ref{genfundef}.
The dual space $ \widetilde{\Sigma}_b $ defined in Equation \ref{dualSigmab} is a general symbolic space, and thus has an infinite extension $ \widetilde{\Sigma}_b^{\infty} $ (see Definition \ref{infext}) which is a symbolic space of infinite type.
We now define a general function system
$$
\{ \phi_{i,j} : \Delta_j \rightarrow X \}_{(i,j) \in \widetilde{\Sigma}_{b,2}}
$$
modeled by $ \widetilde{\Sigma}_b^{\infty} $.
The collection of curves $ \{ \gamma_{c,j} \}_{0 \leq c \leq b} $ forms a parabolic foliation of each $ A_j $ (see figure \ref{stripfol}).
Each point $ x \in \Delta_j $ is the unique intersection point $ a_{c,j}^{\pm} $ of one of these curves $ \gamma_{c,j} $ with the upper boundary $ S^+ = [0,b] $.
For any $ i,j \geq N_b $ we define maps $ \phi_{i,j} : \Delta_j \rightarrow [0,b] $ as follows.
\begin{equation}
\label{phiij}
\phi_{i,j}(a_{c,j}^{\pm}) = \left( p_{c,(j,i)}^{\pm} \right)^{-1} \: \Phi^{i-1} \: \Theta \: p_{c,j}^{\pm} (a_{c,j}^{\pm})
\end{equation}
The definition resembles that of $ f_i $ given in Equation \ref{fi}.
Each point $ a_{c,j}^{\pm} \in \Delta_j $ is projected down to the vertex $ v_{c,j} $ of the parabola $ \gamma_{c,j} $.
It then follows the orbit of $ v_{c,j} $ through the insertion $ \Theta $ and the $ (i-1) $-th return to $ S $ under $ \Phi $, which by definition is the vertex $ v_{c,(j,i)} $, and is then inversely projected back along $ \gamma_{c,(j,i)} $ to its intersection point $ a_{c,(j,i)}^{\pm} \in [0,b] $.
We need to show that this is well-defined for $ (i,j) \in \widetilde{\Sigma}_{b,2} $.
Recall from Equation \ref{Sigma2} that $ \Sigma_{b,2} $ is the sequence space indexing the level-two curves $ \gamma_{c,(i,j)} $ defined in Equation \ref{gammack} by
$$
\gamma_{c,(i,j)} = (\Phi^{j-1} \Theta)(\gamma_{c,i}).
$$
Comparing with Equation \ref{phiij}, we see that $ \phi_{i,j} $ is well-defined with image in $ S $ when $ (i,j) \in \widetilde{\Sigma}_{b,2} $.
We can now state the following theorem.
\begin{theorem}
\label{functhm}
Let $ \Sigma_b $ be the general symbolic space given in Equation \ref{Sigmab}, and let $ \Sigma_b^{\infty} $ be its infinite extension.
Then the collection $ \{ \phi_{i,j} : \Delta_j \rightarrow [0,b] \}_{(i,j) \in \widetilde{\Sigma}_{b,2}} $ is a $ C^{1+\alpha} $ general function system modeled by the dual $ \widetilde{\Sigma}_b^{\infty} $.
\end{theorem}
\begin{proof}
We will show that $ \{ \phi_{i,j} \} $ satisfies the requirements of a $ C^{1+\alpha} $ general function system given in Definition \ref{genfundef}.
\begin{itemize}
\item \textit{Uniform contraction}: For each $ (i,j) \in \widetilde{\Sigma}_{b,2} $ the maps $ \{ \phi_{i,j} : \Delta_j \rightarrow [0,b] \} $ have a common Lipschitz constant $ 0 < s < 1 $.
\vspace{0.1cm}
Recall the above proof that the maps $ f_i $ are uniformly Lipschitz; a similar argument holds here.
Consider the dual transverse distances $ a(i,j) $ of level two defined in Chapter \ref{dualtrans}.
Then for all $ 0 \leq c, c' \leq b $ we have
$$
| \phi_{i,j}(a_{c,j}^{\pm}) - \phi_{i,j}(a_{c',j}^{\pm}) | < a(i,j)
$$
for all $ (i,j) \in \widetilde{\Sigma}_{b,2} $.
Since by $ a(i,j) \rightarrow 0 $ as $ i,j \rightarrow \infty $ by Proposition \ref{trans2}, the Lipschitz constant decreases as $ i,j \rightarrow \infty $.
Thus for a fixed $ i $, we have that $ \phi_{i,j} $ is uniformly Lipschitz for all $ j $, with Lipschitz constant equal to the Lipschitz constant of $ \phi_{i,1} $.
Let $ K_f $ be the uniform Lipschitz constant of the maps $ f_i $, and let $ K_{\phi} $ be the uniform Lipschitz constant of the maps $ \phi_{i,1} $. Taking $ K = \max \{ K_f, K_{\phi} \} $ suffices.
\vspace{0.2cm}
\item \textit{Separation}:
For each $ (i,j), (i',j') \in \widetilde{\Sigma}_{b,2} $ we have
$$
\phi_{i,j}(\Delta_j) \cap \phi_{i',j'}(\Delta_{j'}) = \emptyset
$$
when $ i \neq i' $ or $ j \neq j' $.
\vspace{0.1cm}
This is a consequence of the separation of $ \Delta_i $ and the following nesting property.
\vspace{0.2cm}
\item \textit{Nesting property}:
For all $ k \geq 1 $ and $ \omega \in \widetilde{\Sigma}_{b,k} $ we have
$$
\phi_{\omega_i,\omega_{i+1}} (\Delta_{\omega_{i+1}}) \subset \Delta_{\omega_i}
$$
for all $ 1 \leq i \leq k-1 $.
\vspace{0.1cm}
The dual curves $ \{\gamma_{c,(i,j)}\}_{(i,j) \in \widetilde{\Sigma}_{b,2}} $ form a parabolic foliation of the dual sets $ \{ A_{i,j} \}_{(i,j) \in \widetilde{\Sigma}_{b,2}} $.
By Proposition \ref{dualnestingk} we know that $ A_{i,j} \subset A_i $ for dual words $ (i,j) $.
By definition, $ \phi_{i,j} $ maps the endpoints $ a_{c,j}^{\pm} $ of each curve $ \gamma_{c,j} \subset A_j $ to the endpoints $ a_{c,(i,j)}^{\pm} $ of the curve $ \gamma_{c,(i,j)} \subset A_{i,j} \subset A_i $.
Since $ \Delta_i = A_i \cap [0,b] $, this can be rewritten as
$$
\phi_{i,j}(\Delta_j) \subset \Delta_i.
$$
The desired statement then follows by induction on the word length $ k = |\omega| $.
\vspace{0.2cm}
\item \textit{$ C^{1+\alpha} $ regularity}:
There exists $ \alpha > 0 $ such that for all $ (i,j) \in \widetilde{\Sigma}_{b,2} $, the maps $ \phi_{i,j} $ are of class $ C^{1+\alpha} $.
\vspace{0.1cm}
This is identical to the previous argument for the regularity of the maps $ \{f_i\}_{i \in \Sigma_{b,1}} $.
\vspace{0.2cm}
\end{itemize}
\end{proof}
\subsection{A graph directed pseudo-Markov subsystem}
\label{GDPMsub}
The previous section defined a general function system on $ [0,b] $ modeled by $ \widetilde{\Sigma}_b $.
The sequence space $ \widetilde{\Sigma}_b $ is not completely determined, because we do not know the escape times $ M_{i_1,\ldots,i_k} $ defining it.
However, in Proposition \ref{escapek} we obtained explicit estimates on those escape times for large values of $ i_1,\ldots,i_k $.
In this section we will extract a subspace that uses these estimates.
We will then show that the function system modeled by the subspace is a graph directed pseudo-Markov system, as defined in Section \ref{GDPM}.
\subsubsection{The sequence space $ \widetilde{\Sigma}_{\epsilon} $}
For each $ 0 < \epsilon \leq b $, let $ S_{\epsilon} \subset S $ be the rectangle intersecting the Reeb cylinder $ \{ r=2 \} $, and the top $ S^+ $ and bottom $ S^- $ of $ S $, with width $ \epsilon $.
Let $ S_{\epsilon}^+ \subset S^+ $ be the upper boundary of this rectangle, which can be identified with $ [0,\epsilon] $.
See Figure \ref{Sepsilonfig}.
\begin{figure}[h]
\includegraphics[width=\linewidth, trim={1cm 8.3cm 2cm 3cm}, clip]{PicTemplate27}
\caption{The small rectangle $ S_{\epsilon} $ inside the larger rectangle $ S $. Here $ N_b=2 $, the smallest integer such that $ \gamma_i $ intersects $ S^+ $ for all $ i \geq N_b $, and $ N_{\epsilon} = 6 $, the smallest integer such that $ \gamma_i $ intersects $ S_{\epsilon}^+ $ for all $ i \geq N_{\epsilon} $.}
\label{Sepsilonfig}
\end{figure}
As for $ N_b $, let $ N_{\epsilon} $ be the smallest integer such that $ \gamma_i $ intersects $ S_{\epsilon}^+ $ for all $ i \geq N_{\epsilon} $.
This defines a sequence space $ \Sigma_{\epsilon} $ as in Equation \ref{Sigmab}, with dual $ \widetilde{\Sigma}_{\epsilon} $ as in Equation \ref{dualSigmab}.
Since $ \epsilon \leq b $ we have $ N_{\epsilon} \geq N_b $.
Furthermore, $ \lim_{\epsilon \to 0} N_{\epsilon} = \infty $.
We claim that for sufficiently small $ \epsilon > 0 $, the sequence space $ \Sigma_{\epsilon} $ is a graph directed symbolic space in the sense of Chapter \ref{Thermo}; there is a countable alphabet $ E $ and incidence matrix $ A : E \times E \rightarrow \{0,1\} $ such that for each $ n \geq 1 $, $ \Sigma_{\epsilon, n} = E_A^n $ in the notation of Equation \ref{admis}.
By Proposition \ref{escapek}, for small $ \delta>0 $ and for large $ i_1,\ldots,i_n $ we have
$$
C+(K-\delta) i_{n-1}^2 < M_{i_1,\ldots,i_n} < (C+\delta) + K i_{n-1}^2,
$$
where $ C $ and $ K $ are defined in Equation \ref{defCK}.
Let $ \lfloor \cdot \rfloor $ be the integer floor.
Since $ \lim_{\epsilon \to 0} N_{\epsilon} = \infty $, for small enough $ \epsilon $ we may substitute the above estimate into Equation \ref{Sigmak} to obtain
\begin{equation}
\label{Sigmaepn}
\Sigma_{\epsilon, n} = \bigcup_{i_1 = N_{\epsilon}}^{\infty} \bigcup_{i_2=N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_1^2} \cdots \bigcup_{i_n = N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_{n-1}^2} (i_1, \ldots, i_n).
\end{equation}
Let $ E = \Sigma_{\epsilon,1} = \{ N_{\epsilon}, N_{\epsilon}+1, \ldots \} $ and define the matrix $ A : E \times E \rightarrow \{0,1\} $ by
\begin{equation}
\label{Aij}
A(i,j) = \left\{
\begin{array}{lr}
1 & : j \leq \lfloor C \rfloor + \lfloor K \rfloor i^2 \\
0 & : j > \lfloor C \rfloor + \lfloor K \rfloor i^2
\end{array}
\right.
\end{equation}
Then the admissible words $ E_A^n $ defined in Equation \ref{admis} are
\begin{align*}
E_A^n &= \{ (i_1, \ldots, i_n) \in E^n : A_{i_j i_{j+1}} = 1 \text{ for all } 1 \leq j \leq n-1 \} \\
&= \{ (i_1, \ldots, i_n) \in \{ N_{\epsilon}, N_{\epsilon}+1, \ldots \}^n : i_{j+1} \leq \lfloor C \rfloor + \lfloor K \rfloor i_j^2 \text{ for all } 1 \leq j \leq n-1 \} \\
&= \Sigma_{\epsilon, n},
\end{align*}
by comparing with Equation \ref{Sigmaepn}.
Taking the dual, we have $ \widetilde{\Sigma}_{\epsilon, n} = \widetilde{E}_A^n $ for each $ n \geq 1 $.
\subsubsection{The limit set $ J_{\epsilon} $}
By Theorem \ref{functhm}, for any $ 0 < \epsilon \leq b $ the function system
$$
\{ \phi_{i,j} : \Delta_j \rightarrow [0,\epsilon] \}_{(i,j) \in \widetilde{\Sigma}_{\epsilon,2}}
$$
is a well-defined $ C^{1+\alpha} $ subsystem of $ \{ \phi_{i,j} : \Delta_j \rightarrow [0,b] \}_{(i,j) \in \widetilde{\Sigma}_{b,2}} $ modeled by the dual space $ \widetilde{\Sigma}_{\epsilon} $.
By the above discussion, for sufficiently small $ \epsilon > 0 $ there exists an incidence matrix $ A $ such that
$$
\widetilde{\Sigma}_{\epsilon,n} = \widetilde{E}_A^n.
$$
Let $ J_{\epsilon} \subset J_b $ be the limit set of this subsystem.
By definition,
\begin{equation}
\label{Jep}
J_{\epsilon} = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in \widetilde{E}_A^n} \Delta_{\omega}.
\end{equation}
\vfill
\eject
\section{The transverse Cantor set}
\label{Transcant}
In this section we will relate this transverse Cantor set $ \tau $ of the Kuperberg minimal set $ \mathcal{M} $ to limit sets of the function systems defined in Section \ref{FunctionSystems}.
We will use the previous symbolic dynamics developed in Section \ref{Kupmin} for the sets $ \mathcal{N}_0 \cap S $ and $ \mathcal{M}_0 \cap S $ to define bijective coding maps between these Cantor sets and the appropriate symbolic spaces.
\subsection{Sections of the minimal set}
Re-stating Equation \ref{leveldecompgammau},
\begin{equation}
\label{min1char}
\mathcal{M}_{0,1} \cap S = \bigcup_{\omega \in \Sigma_b} \gamma_{\omega}^u.
\end{equation}
Define $ \mathcal{M}_1 = \overline{\mathcal{M}_{0,1}} $,
so that
\begin{equation}
\label{M1}
\mathcal{M}_1 \cap S = \overline{\mathcal{M}_{0,1} \cap S} = \overline{ \bigcup_{\omega \in \Sigma_b} \gamma_{\omega}^u }.
\end{equation}
Recall the subspace $ \Sigma_{\epsilon} \subset \Sigma_b $ defined in Section \ref{GDPMsub}.
Replacing $ S $ with $ S_{\epsilon} $ and $ b $ with $ \epsilon $ in Equation \ref{leveldecompgammau}, we obtain
\begin{equation}
\label{min1char2}
\mathcal{M}_1 \cap S_{\epsilon} = \overline{ \bigcup_{\omega \in \Sigma_{\epsilon}} \gamma_{\omega}^u }.
\end{equation}
\subsection{The transverse Cantor set in $ [0,b] $}
In this section we will prove preliminary versions of Theorem $ \mathbf{A} $ and Corollary $ \mathbf{B} $ from Chapter \ref{Intro}.
The full versions will require the notion of \textit{interlacing} and will be given in the subsequent section.
In the lamination charts for $ \mathcal{M} $ constructed in Chapter 19 of \cite{Hur}, the transverse Cantor set $ \tau $ has a variable radial coordinate.
By Equation \ref{Sp}, our choice of transversal $ S^+ $ is compatible with these lamination charts.
Then with $ \tau $ defined in Theorem \ref{minchar}, we may set
\begin{equation}
\label{tau}
\tau = \mathcal{M} \cap S^+.
\end{equation}
As in the previous section we denote
\begin{equation}
\label{tau1}
\tau_1 = \mathcal{M}_1 \cap S^+.
\end{equation}
Combining Equations \ref{awdef}, \ref{transverse1} and \ref{min1char2} we obtain
$$
\tau_1 = \overline{ \bigcup_{\omega \in \Sigma_b} a_{\omega}^- }.
$$
Re-indexing the points $ a_{\omega}^- $ using the bijection $ (\omega_1, \omega_2, \ldots) \mapsto (\ldots, \omega_2, \omega_1) $ yields
\begin{equation}
\label{min1trans}
\tau_1 = \overline{ \bigcup_{\omega \in \widetilde{\Sigma}_b} a_{\omega}^- }.
\end{equation}
See Figure \ref{transa2} for an illustration the intersections of the level-one curves in $ \mathcal{M}_1 \cap S $ with $ S^+ $.
\begin{figure}[h!]
\includegraphics[width=\linewidth, trim={5cm 9cm 1cm 5.8cm}, clip]{PicTemplate26}
\caption{The points $ a_i^- $ as intersections of the level-one curves $ \gamma^u_i $ with the upper boundary $ S^+ $ of $ S $. We obtain this from Figure \ref{transa} by restricting the parametrization of $ \gamma_i $.}
\label{transa2}
\end{figure}
In Section \ref{FunctionSystems}, we showed that the general symbolic space $ \Sigma_b $ has the extension admissibility property.
Since its dual $ \widetilde{\Sigma}_b $ also does, it has a well-defined infinite extension $ \widetilde{\Sigma}_b^{\infty} $.
We can now state the following theorem, which will be used later to prove Theorem \textbf{A} from Chapter \ref{Intro}.
\begin{theorem*}[$\mathbf{A}_0$]
There is a $ C^{1+\alpha} $ general function system on $ [0,b] $ modeled by $ \widetilde{\Sigma}_b^{\infty} $ with limit set $ \tau_1 $.
\end{theorem*}
\begin{proof}
In Theorem \ref{functhm} we defined a $ C^{1+\alpha} $ general function system on $ [0,b] $ modeled by $ \widetilde{\Sigma}_b^{\infty} $ and proved that its limit set is
$$
J_b = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in \widetilde{\Sigma}_{b,n}} \Delta_{\omega}.
$$
Thus it suffices to show that $ \tau_1 = J_b $.
By Equation \ref{min1trans}, this is equivalent to
$$
\overline{ \bigcup_{\omega \in \widetilde{\Sigma}_b} a_{\omega}^- } = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in \widetilde{\Sigma}_{b,n}} \Delta_{\omega}.
$$
We will show both containments.
First, let
$$
x \in \overline{ \bigcup_{\omega \in \widetilde{\Sigma}_b} a_{\omega}^- } = \overline{ \bigcup_{n=1}^{\infty} \bigcup_{\omega \in \widetilde{\Sigma}_{b,n}} a_{\omega}^- }.
$$
Then there is a sequence of finite words $ \omega_n \in \widetilde{\Sigma}_{b,n} $ with $ a_{\omega_n}^- \rightarrow x $.
Note that $ J_b $ contains each point $ a_{\omega_n}^- $, because by construction, $ a_{\omega_n}^- $ is the left endpoint of the interval $ \Delta_{\omega_n} $.
Because $ J_b $ is a Cantor set it must contain all its limit points.
In particular it must contain $ x $, which concludes the forward containment.
For the reverse containment, let
$$
x \in J_b = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in \widetilde{\Sigma}_{b,n}} \Delta_{\omega}.
$$
By Theorem \ref{functhm}, $ J_b $ is the limit set of a general function system modeled by $ \widetilde{\Sigma}_b^{\infty} $, so $ x $ corresponds to a unique word $ \omega \in \widetilde{\Sigma}_b^{\infty} $ via the coding map $ \pi $:
$$
x = \pi(\omega) = \bigcap_{n=1}^{\infty} \Delta_{\omega |_n}.
$$
For details, see Section \ref{genfun}.
Consider the finite restriction of $ \omega $; this is the sequence $ \omega_n = \omega |_n \in \widetilde{\Sigma}_{b,n} $ (see Section \ref{extrestwords}).
By definition of the sets $ \Delta_{\omega} $ we have $ a_{\omega_n}^- \in \Delta_{\omega_n} $ and thus
$$
\lim_{n \to \infty} a_{\omega_n}^- = \lim_{n \to \infty} \bigcap_{k=1}^n \Delta_{\omega_k} = x.
$$
Then $ x $ is a limit point of a sequence $ a_{\omega_n}^- $ with $ \omega_n \in \widetilde{\Sigma}_b $, so
$$
x \in \overline{ \bigcup_{\omega \in \widetilde{\Sigma}_b} a_{\omega}^- }
$$
as desired.
\end{proof}
In the above proof, we used that the limit set of a general function system modeled by a symbolic space of infinite type has a bijective coding to that space.
For details, see Section \ref{genfun}.
As an immediate corollary to Theorem $ \mathbf{A}_0 $ we obtain the following.
\begin{corollary*}[$ \mathbf{B}_0 $]
There is a symbolic space $ \Sigma_1 $ of infinite type and a bijective coding
$$
\pi_1 : \Sigma_1 \rightarrow \tau_1
$$
\end{corollary*}
\begin{proof}
By Theorem $ \mathbf{A}_0 $, $ \tau_1 $ is the limit set of a general function system on $ [0,b] $ modeled by $ \widetilde{\Sigma}_b^{\infty} $.
By the results of Section \ref{genfun}, there is a bijective coding
$$
\pi_1 : \widetilde{\Sigma}_b^{\infty} \rightarrow \tau_1.
$$
So it suffices to take $ \Sigma_1 = \widetilde{\Sigma}_b^{\infty} $.
\end{proof}
\subsection{The transverse Cantor set $ \tau $ as an interlaced Cantor set}
Recall from Equation \ref{Psi} that the full Kuperberg pseudogroup $ \Psi $ on $ S_1 \cup S_2 $ is
$$
\Psi = \langle \Phi_1, \Phi_2, \Phi_{1,2}, \Theta_1, \Theta_2 \rangle.
$$
Up to now, we have only considered the sub-pseudogroup $ \Psi_1 = \langle \Phi_1, \Theta_1 \rangle $, using the shorthand notation $ \Theta = \Theta_1 $ and $ \Phi = \Phi_1 $.
The results of Chapters \ref{Kupmin} -- \ref{FunctionSystems} gave us a complete description of the transverse set $ \mathcal{M}_{0,1} \cap S $, its closure $ \mathcal{M}_1 \cap S $, and its transverse Cantor set $ \tau_1 $.
However, to account for the full transverse minimal set $ \mathcal{M} \cap S $ it is necessary to incorporate the other maps $ \Phi_2, \Theta_2 $, and $ \Phi_{1,2} $.
In this section we will use the symmetry of the plug $ K $ to show that these maps generate a Cantor set $ \tau_2 $ identical to $ \tau_1 $, and that the interlacing of $ \tau_1 $ and $ \tau_2 $ (see Section \ref{sectioninter}) is the transverse Kuperberg minimal set $ \tau $ from Theorem \ref{minchar}.
In this section, we will dispense with our shorthand notation $ S, D, \gamma $ for $ S_1, D_1, \gamma_1 $, and return to considering $ S_i, D_i $, and $ \gamma_i $ for $ i=1,2 $, as we did in the first half of Chapter \ref{Kuppseudo}.
\subsubsection{The Cantor set $ \tau_2 $ in $ S_1^+ $}
The intersection of the notched Reeb cylinder $ \mathcal{R}' $ with the upper insertion rectangle $ S_2 $ is $ \gamma_2 $, a vertical line with a parametrization similar to that of $ \gamma_1 $ given in Equation \ref{gamma}.
Consider the curves
$$
\overline{\gamma}_i = \Phi_1^{i-1} \Theta_2 (\gamma_2).
$$
The images of $ \overline{\gamma}_i $ lie in $ S_1 $, and have a similar parametrization to those of $ \gamma_i $ given in Proposition \ref{gamma1param}.
A derivation of this fact closely resembles the proof there, which we will not repeat.
The only quantitative difference appearing in the parametrization is the constant $ 0 \leq \alpha < 2\pi $, which is the angular coordinate of the vertex of the parabola $ \sigma^{-1} \gamma_1 \subset \{ z=-2 \} $ (see Equation \ref{sigmagamma}).
If $ \alpha_1 $ and $ \alpha_2 $ are the vertices of the parabolas $ \sigma^{-1} \gamma_1 $ and $ \sigma^{-1} \gamma_2 $ respectively, then $ 0 \leq \alpha_1 < \alpha_2 < 2\pi $.
We can then adjust the proof of Proposition \ref{gamma1param} to show that the $ i $th return time of the Wilson orbit of $ \sigma^{-1} \gamma_2 $ to $ S_1 $ is strictly between the $ i $th and $ (i+1) $th return times of $ \sigma^{-1} \gamma_1 $, which define the curves $ \gamma_i $ and $ \gamma_{i+1} $.
As $ i $ increases, the $ r $ and $ z $-coordinate of the curves $ \gamma_i $ and $ \overline{\gamma}_i $ increases.
Thus the curve $ \overline{\gamma}_i $ is between $ \gamma_i $ and $ \gamma_{i+1} $ for all $ i \in \Sigma_{b,1} $, hence these curves alternate as $ i $ increases.
For an illustration of this, see Figure \ref{curveintfig}.
\begin{figure}[h]
\includegraphics[width=\linewidth, trim={3.2cm 10.5cm 1cm 2.8cm}, clip]{Interlace}
\caption{A plot of the level-one curves $ \gamma_i $ generated by $ \Phi_1 $ and $ \Theta_1 $, shown in blue, and the level-one curves $ \overline{\gamma}_i $ generated by $ \Phi_1 $ and $ \Theta_2 $, shown in red.
Compare with Figure \ref{levelone}.}
\label{curveintfig}
\end{figure}
For any $ (i_1,\ldots,i_k) \in \Sigma_{b,k} $ we recursively define
$$
\overline{\gamma}_{i_1,\ldots,i_k} = \Phi^{i_k -1} \: \Theta_1 (\overline{\gamma}_{i_1,\ldots,i_{k-1}}),
$$
exactly as we defined $ \gamma_{i_1,\ldots,i_k} $.
Because their construction is identical to that of $ \gamma_{i_1,\ldots,i_k} $, the curves $ \overline{\gamma}_{i_1,\ldots,i_k} $ are also coded by $ \Sigma_{b,k} $, and their transverse distances $ \overline{a}(i_1,\ldots,i_k) $ are identical to the transverse distances $ a(i_1,\ldots,i_k) $ estimated in Chapter \ref{Transversal}.
The following definitions are similar to Equations \ref{min1char2} and \ref{tau1}:
$$
\mathcal{M}_2 \cap S = \overline{ \bigcup_{\omega \in \Sigma_b} \overline{\gamma}_{\omega} }, \qquad \qquad \tau_2 = \mathcal{M}_2 \cap S^+.
$$
We then define a function system using conjugations of pseudogroup elements by projections along $ \overline{\gamma}_{\omega} $, and show that $ \tau_2 $ is a Cantor limit set of a general function system modeled by $ \widetilde{\Sigma}_b^{\infty} $, exactly as we did for $ \tau_1 $ in Theorem $ \mathbf{A}_0 $.
\subsubsection{The level-one curves in $ S_1 \cup S_2 $}
Consider the collection $ \{ \gamma_i, \overline{\gamma}_j \}_{i,j \in \Sigma_{b,1}} $ of interlaced curves in $ S_1 $.
We may naturally re-index this collection to $ \eta_i $, where $ i \in (\Sigma_{b,1} \ast \Sigma_{b,1})_1 $, the joint sequence space (see Section \ref{sectioninter}).
There is an identical family of curves of interlaced curves $ \eta'_i $ in $ S_2 $.
The reason for this is the symmetry of $ K $; we may invert the plug, reverse time, and obtain a similar analysis to Chapter \ref{Kupmin}.
Visually, there is an inverted Figure \ref{curveintfig} in the upper insertion rectangle $ S_2 $.
\subsubsection{The level-two curves in $ S_1 \cup S_2 $}
For each $ i_1 \in (\Sigma_{b,1} \ast \Sigma_{b,1})_1 $, define
$$
\eta_{i_1,i_2} = \Phi_1^{i_2-1} \: \Theta_1(\eta_{i_1}), \qquad \qquad \overline{\eta}_{i_1,i_2} = \Phi_1^{i_2-1} \: \Theta_2 (\eta'_{i_1}).
$$
An argument identical to the construction of $ \Sigma_{b,2} $ shows that the admissible words coding $ \eta_{i_1,i_2} $ and $ \overline{\eta}_{i_1,i_2} $ are two disjoint copies of $ \Sigma_{b,2} $.
The generator $ \Theta_1 $ maps $ S_1 $ into $ A_1 $, the boundary of which is $ \gamma_1 $.
Because they are defined using $ \Theta_1 $, the curves $ \eta_{i_1,i_2} $ are nested in $ \gamma_{i_2} $.
Similarly, $ \overline{\eta}_{i_1,i_2} $ are nested in $ \overline{\gamma}_{i_2} $ because they are defined using $ \Theta_2 $.
A proof of these facts follows Proposition \ref{nesting2}.
Consider the collection $ \{ \eta_{i_1,i_2}, \overline{\eta}_{j_1, j_2} \} $ of all level-two curves in $ S_1 $, where $ (i_1, i_2) $ and $ (j_1, j_2) $ range through $ \Sigma_{b,2} $.
As with the level-one curves, we re-index this to a single collection $ \{ \eta_{i_1,i_2} \} $, where $ (i_1,i_2) $ range through the joint sequence space $ (\Sigma_{b,1} \ast \Sigma_{b,1})_2 $.
Finally, note that we have identical level-two interlaced curves $ \eta'_{i_1, i_2} $ in the upper rectangle $ S_2 $, also indexed by $ (\Sigma_{b,1} \ast \Sigma_{b,1})_2 $.
\subsubsection{The level-$k$ curves in $ S_1 \cup S_2 $}
We continue recursively defining the interlaced curves $ \eta_{i_1,\ldots,i_k} $ and $ \overline{\eta}_{i_1,\ldots,i_k} $ in $ S_1 $ as in Equation \ref{gammai1ik}:
\begin{align*}
\eta_{i_1,\ldots,i_k} &= \Phi_1^{i_k-1} \: \Theta_1(\eta_{i_1,\ldots,i_{k-1}}) \\
\overline{\eta}_{i_1,\ldots,i_k} &= \Phi_1^{i_k-1} \: \Theta_2(\eta'_{i_1,\ldots,i_{k-1}})
\end{align*}
After each such definition, we re-index the individual collections $ \eta_{\omega}, \overline{\eta}_{\omega} $ to $ \eta_{\omega} $ by the joint sequence space.
We then note the identical families $ \eta'_{\omega} \in S_2 $, and continue.
\subsubsection{The function system on the interlaced curves}
Recall the transverse Cantor set $ \tau $ defined in Theorem \ref{minchar} and Equation \ref{tau}.
By the interlacing of curves studied above, $ \tau $ is the interlacing of the Cantor sets $ \tau_1 $ and $ \tau_2 $, in the sense of Section \ref{sectioninter}.
Using the interlaced curves, we define a general function system modeled by the infinite extension of the joint sequence space $ \Sigma_{b,1} \ast \Sigma_{b,1} $, whose limit set is $ \tau $.
This is the content of the following theorem, whose proof follows that of Theorem $ \mathbf{A}_0 $.
\begin{theorem*}[\textbf{A}]
Let $ \tau $ be the Cantor transversal of the Kuperberg minimal set.
Then there is a $ C^{1+\alpha} $ general function system on $ [0,b] $ modeled by $ \widetilde{\Sigma}_b^{\infty} \ast \widetilde{\Sigma}_b^{\infty \prime} $ with limit set $ \tau $.
\end{theorem*}
In the notation of Section \ref{sectioninter}, this limit set is
$$
\tau = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in (\widetilde{\Sigma}_b \ast \widetilde{\Sigma}'_b)_n} \Delta_{\omega}.
$$
Just as with Corollary $ \mathbf{B}_0 $, we obtain the following from Theorem \textbf{A}.
\begin{corollary*}[\textbf{B}]
Let $ \tau $ be the Cantor transversal of the Kuperberg minimal set. Then there is a bijective coding map
$$
\pi : \widetilde{\Sigma}_b^{\infty} \ast \widetilde{\Sigma}_b^{\infty \prime} \rightarrow \tau.
$$
\end{corollary*}
\subsection{The transverse Cantor set in $ [0,\epsilon] $}
As in Equation \ref{tau}, we define $ \tau_{\epsilon} \subset \tau $ by
\begin{equation}
\label{tauep}
\tau_{\epsilon} = \mathcal{M} \cap S_{\epsilon}^+.
\end{equation}
For $ i=1,2 $, let $ \tau_{i, \epsilon} $ be the intersection of $ \tau_i $ with an $ \epsilon $-neighborhood of the critical orbit in the Kuperberg plug.
By applying Theorem $ \textbf{A}_0 $ to a suitably small transversal $ [0,\epsilon] $, we can prove Theorem $ \mathbf{C}_0 $.
This will be used to prove Theorem \textbf{C}.
\begin{theorem*}[$ \mathbf{C}_0 $]
For each $ i=1,2 $ and sufficiently small $ \epsilon > 0 $, there is a $ C^{1+\alpha} $ graph directed pseudo-Markov system on $ [0,\epsilon] $ with limit set $ \tau_{i,\epsilon} $.
\end{theorem*}
\begin{proof}
We will give the proof only for $ \tau_{1,\epsilon} $.
By Theorem $ \mathbf{A}_0 $, $ \tau_{1,\epsilon} = J_{\epsilon} $, the limit set of a $ C^{1+\alpha} $ function system modeled by the dual $ \widetilde{\Sigma}_{\epsilon}^{\infty} $.
By the results of Section \ref{GDPMsub}, this is a $ C^{1+\alpha} $ graph directed pseudo-Markov system.
\end{proof}
Again by considering the interlacing of $ \tau_{1,\epsilon} $ with $ \tau_{2,\epsilon} $ we obtain $ \tau_{\epsilon} $ with the following characterization.
\begin{theorem*}[\textbf{C}]
Let $ \tau $ be the Cantor transversal of the Kuperberg minimal set, and let $ \tau_{\epsilon} $ be the intersection of $ \tau $ with an $ \epsilon $-neighborhood of the critical orbit.
For sufficiently small $ \epsilon > 0 $ there is a $ C^{1+\alpha} $ graph directed pseudo-Markov system on $ [0,\epsilon] $ with limit set $ \tau_{\epsilon} $.
\end{theorem*}
We conclude by displaying the limit set of this pseudo-Markov system.
In Equation \ref{Aij} we defined the incidence matrix $ A $ defining the admissible words $ E_A^n $ of length $ n $.
For two copies $ E $ and $ E' $ of the sequence space $ \Sigma_{b,1} $, we have a joint incidence matrix $ A^{E \cup E'} $ coding the admissible words in the interlaced Cantor set, as defined in Section \ref{sectioninter}.
This joint matrix is
\begin{equation}
\label{jointE}
A^{E \cup E'}(i,j) =
\left\{
\begin{array}{lr}
1 & : i,j \in E \text{ or } i,j \in E' \text{ and } j \leq \lfloor C \rfloor + \lfloor K \rfloor i^2 \\
0 & : i,j \in E \text{ or } i,j \in E' \text{ and } j > \lfloor C \rfloor + \lfloor K \rfloor i^2 \\
1 & : i \in E \text{ and } j \in E' \text{ or } i \in E' \text{ and } j \in E
\end{array}
\right.
\end{equation}
This matrix defines admissible words $ (E \cup E')_{A^{E \cup E'}}^n $, and by Theorem \textbf{C} we have
\begin{equation}
\label{minep}
\tau = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in (E \cup E')_{A^{E \cup E'}}^n} \Delta_{\omega}.
\end{equation}
\vfill
\eject
\section{Dimension of the Cantor set}
\label{Dimtrans}
In this section we will apply the dimension theory developed in Section \ref{Dime} to study the Hausdorff dimension of $ \tau $, the transverse minimal set in Kuperberg's plug.
We will then use the product structure of the lamination to extend this to the dimension of $ \mathcal{M} $.
By Theorem \textbf{A} we know that $ \tau $ is the limit set of a $ C^{1+\alpha} $ general function system.
By Theorem \textbf{C}, for sufficiently small $ \epsilon > 0 $, $ \tau_{\epsilon} $ is the limit set of a pseudo-Markov subsystem.
Limit sets of pseudo-Markov systems have a well-developed dimension theory as exposed in Section \ref{Dime}.
We wish to apply this theory to the transverse minimal set $ \tau $, but to do this we must first relate the dimension of $ \tau $ to that of $ \tau_{\epsilon} $.
\subsection{The Hausdorff dimension of $ \tau $}
The next lemma uses minimality of $ \mathcal{M} $ to show that the Hausdorff dimension of $ \tau $ can be calculated inside a small neighborhood of an arbitrary point.
For any $ x \in \tau $, let $ B_{\epsilon}(x) \subset M$ denote the closed ball of radius $ \epsilon $ centered at $ x $, and let $U_{\epsilon}(x) \subset B_{\epsilon}(x)$ denote its open interior.
\begin{lemma}[\textbf{D}]
\label{constdim}
Let $ \tau $ be the Cantor transversal of the Kuperberg minimal set.
For $ \epsilon, \epsilon' > 0 $ sufficiently small, and any $ x,y \in \tau $, we have
$$
{\rm dim}_H(\tau \cap B_{\epsilon}(x)) = {\rm dim}_H(\tau \cap B_{\epsilon'}(y)).
$$
\end{lemma}
\begin{proof}
We show that $\displaystyle {\rm dim}_H(\tau \cap B_{\epsilon'}(y)) \geq {\rm dim}_H(\tau \cap B_{\epsilon}(x))$, with the reverse inequality following by the same method, which proves the claim.
\medskip
For each $z \in \tau \cap B_{\epsilon}(x)$ there exists $T_z > 0$ such that $ \psi_{T_z}(z) \in B_{\epsilon'}(y)$. Let $\varepsilon_z> 0$ be sufficiently small so that
$\displaystyle \psi_{T_z}(B_{\varepsilon_z}(z)) \subset B_{\epsilon'}(y)$.
\medskip
The collection of open balls $\{U_{\varepsilon_z}(z) \mid z \in \tau \cap B_{\epsilon}(x)\}$ is an open covering of the compact set $\tau \cap B_{\epsilon}(x)$ so there exists a finite subcovering, centered at points $\{z_1, \ldots, z_k\} \subset \tau \cap B_{\epsilon}(x)$.
By standard properties of Hausdorff dimension, we have that
$${\rm dim}_H(\tau \cap B_{\epsilon}(x)) = \max ~ \{{\rm dim}_H(B_{\varepsilon_{z_i}}(z_i) \cap \tau)) \mid 1 \leq i \leq k \}. $$
The flow $\psi$ is $C^{\infty}$ so for each $i$ we have that
$${\rm dim}_H(\psi_{T_z}(B_{\varepsilon_{z_i}}(z_i) \cap \tau)) = {\rm dim}_H( B_{\varepsilon_{z_i}}(z_i) \cap \tau) .$$
Now assume that $\epsilon'> 0$ is sufficiently small so that the projection $\Pi_{\mathcal F}$ along the leaves of the foliation is 1-1 when restricted to the ball $ B_{\epsilon'}(y)$,
$$\Pi_{\mathcal F} \colon \psi_{T_z}(B_{\varepsilon_{z_i}}(z_i) \cap \tau) \to \tau \cap B_{\epsilon'}(y) .$$
The value of $\epsilon'> 0$ depends only on the construction of the flow, and not on the choice of the point $y'$.
The holonomy projection map $\Pi_{\mathcal F}$ is $C^1$ by results in HR, so we then have
$${\rm dim}_H(\psi_{T_z}(B_{\varepsilon_{z_i}}(z_i) \cap \tau)) = {\rm dim}_H(\Pi_{\mathcal F}(\psi_{T_z}(B_{\varepsilon_{z_i}}(z_i) \cap \tau))) \leq {\rm dim}_H(\tau \cap B_{\epsilon'}(y)) . $$
The claim follows.
\end{proof}
Now consider the point $ x = (2, \beta, -1) $ in the Kuperberg plug.
This is the intersection of the lower critical orbit with the rectangle $ S $.
By the definition in Section \ref{GDPMsub}, $ x $ is the left endpoint of the transversal $ S_{\epsilon}^+ $.
Then for any $ \epsilon > 0 $, $ \tau_{\epsilon} = \tau \cap B_{\epsilon}(x) $.
Taking $ \epsilon' = b $ in the statement of Lemma \textbf{D}, we obtain that
\begin{equation}
\label{tauepeq}
\text{dim}_H(\tau) = \text{dim}_H (\tau_{\epsilon})
\end{equation}
for any $ \epsilon > 0 $.
This reduces the calculation of the Hausdorff dimension of $ \tau $ to that of $ \tau_{\epsilon} $.
We now combine this with the estimates from Chapter \ref{Transversal} on the transverse distances, to prove the following theorem.
\begin{theorem*}[\textbf{E}]
Let $ \tau $ be the Cantor transversal of the Kuperberg minimal set.
Then the Lebesgue measure of $ \tau $ is zero, and $ 0 < \dim_H (\tau) < 1 $.
\end{theorem*}
\begin{proof}
By Equation \ref{tauepeq}, it suffices to prove the statement for $ \tau_{\epsilon} $ for any $ \epsilon > 0 $.
By Theorem \textbf{C} and Equation \ref{minep} we know that for sufficiently small $ \epsilon > 0 $,
$$
\tau_{\epsilon} = \bigcap_{n=1}^{\infty} \bigcup_{\omega \in (E \cup E')_{A^{E \cup E'}}^n} \Delta_{\omega}.
$$
By construction of the pseudo-Markov system from Section \ref{GDPMsub}, $ |\Delta_{\omega}| = a(\omega) $, the transverse distances studied in Section \ref{Transversal}.
By Proposition \ref{dualtransk}, for any $ \delta > 0 $ there exist $ L_n \in \mathbb{N} $ such that for all $ i_1,\ldots,i_n \geq L_n $ we have
\begin{equation}
\label{ain}
\left| a(i_1,\ldots,i_n)- \left(\frac{\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{n-1}}{i_2^2 \cdots i_n^2} \right)\right| < \frac{\delta}{i_1^2 \cdots i_n^2}
\end{equation}
Taking the dual of Equation \ref{Sigmaepn} yields
\begin{equation}
\label{dualSigmaepn}
\widetilde{E}_{A^E}^n = \bigcup_{i_1 = N_{\epsilon}}^{\infty} \bigcup_{i_2=N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_1^2} \cdots \bigcup_{i_n = N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_{n-1}^2} (i_n, \ldots, i_1).
\end{equation}
By the definition of $ N_{\epsilon} $ given in the proof of Proposition \ref{gamma1param}, we know that $ N_{\epsilon} \rightarrow \infty $ as $ \epsilon \rightarrow 0 $.
So taking a sequence $ \epsilon_n \rightarrow 0 $ with $ N_{\epsilon_n} \geq L_n $ for all $ n $, we have that Equation \ref{ain} holds for all $ (i_1,\ldots,i_n) \in \widetilde{E}_{A^E}^n $ for small enough $ \epsilon $, and $ \delta \rightarrow 0 $ as $ \epsilon \rightarrow 0 $.
Substituting $ |\Delta_{\omega}| = a(\omega) $ into Equation \ref{ain} and rewriting, we have that for any $ \delta>0 $ and small enough $ \epsilon > 0 $,
\begin{equation}
\label{another}
\frac{\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{n-1}-\delta}{i_2^2 \cdots i_n^2}
< |\Delta_{i_1,\ldots,i_n}|
< \frac{\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{n-1}+\delta}{i_2^2 \cdots i_n^2}
\end{equation}
for all $ (i_1,\ldots,i_n) \in (E \cup E')_{A^{E \cup E'}}^n $ and $ \delta \rightarrow 0 $ as $ \epsilon \rightarrow 0 $.
To simplify notation, let
$$
s_i = \frac{\pi^{-1} K^{\frac{3}{2}}}{i^{\frac{5}{2}}}, \; \text{ and } \; r_i = \frac{(2\pi)^{-2} a R^2}{i^2}.
$$
Referring to Section \ref{statmark}, we see that for $ \delta > 0 $ there exists sufficiently small $ \epsilon > 0 $ such that $ \tau_{\epsilon} $ is the limit set of an asymptotically stationary pseudo-Markov system with ratio coefficients $ r_i $ given above, and summable monotone error
$$
a_{\delta}^{\pm}(i_1,\ldots,i_n) = \pm \frac{\delta}{i_1^2 \cdots i_n^2}. $$
By Theorem \ref{statGDMS}, we obtain that the Lebesgue measure of $ \tau_{\epsilon} $ is zero, and that $ 0 < \text{dim}_H(\tau_{\epsilon}) < 1 $.
\end{proof}
\subsection{Estimating the dimension via the pressure}
The following theorem is an application of the thermodynamic formalism developed in Section \ref{Dime} to the dimension theory of $ \tau $.
\begin{theorem*}[\textbf{F}]
Let $ \tau $ be the Cantor transversal of the Kuperberg minimal set.
Let $ t = \text{dim}_H(\tau) $ be its Hausdorff dimension, and $ a > 0 $ the angular speed of the Kuperberg flow.
\begin{itemize}
\item $ t = \text{dim}_H(\tau) $ is the unique zero of a dynamically defined pressure function,
\item $ t $ depends continuously on $ a $,
\item For any $ a $ we may compute $ t $ to a desired level of accuracy.
\end{itemize}
\end{theorem*}
\begin{proof}
By Equation \ref{tauepeq}, it suffices to prove the statement for $ \tau_{\epsilon} $ for any $ \epsilon > 0 $.
By Theorem \textbf{C}, we know that for small enough $ \epsilon $, $ \tau_{\epsilon} $ is the interlaced limit set of a pseudo-Markov system, with limit set given in Equation \ref{minep}.
From Section \ref{Dime}, the pressure function determined by this pseudo-Markov system is
$$
p(t) = \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in (E \cup E')_{A^{E \cup E'}}^n} | \Delta_{\omega} |^t.
$$
Since $ E $ and $ E' $ are equal, for each interval $ \Delta_{\omega} $ coded by a word $ \omega \in (E \cup E')_{A^{E\cup E'}}^n $ there are two intervals $ \Delta_{\omega} $ for $ \omega \in E_{A^E}^n $, and these two intervals have equal lengths.
From this we obtain
$$
p(t) = \lim_{n \to \infty} \frac{1}{n} \log \sum_{\omega \in E_A^n} | 2 \Delta_{\omega} |^t.
$$
Applying Equations \ref{dualSigmaepn} and \ref{another} we obtain that $ p^-(t) < p(t) < p^+(t) $, where
\begin{align}
\label{another2}
p^{\pm}(t) &= \lim_{n \to \infty} \frac{1}{n} \log \sum_{(i_1,\ldots,i_n) \in E_A^n} \left| \frac{2\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{n-1} \pm \delta}{i_2^2 \cdots i_n^2} \right|^t \\
&= \lim_{n \to \infty} \frac{1}{n} \log \sum_{i_1 = N_{\epsilon}}^{\infty} \sum_{i_2=N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_1^2} \cdots \sum_{i_n=N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_{n-1}^2} \left| \frac{2\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{n-1} \pm \delta}{i_2^2 \cdots i_n^2} \right|^t. \nonumber
\end{align}
By Bowen's theorem (Theorem \ref{Bow}),
$$
\text{dim}_H(\tau_{\epsilon}) = \inf \{ t \geq 0 : p(t) \leq 0 \}.
$$
It is easy to see that $ p^{\pm}(t) $ have the same properties as $ p(t) $ specified in Theorem \ref{presprop}; in particular they are strictly decreasing and have unique zeros on $ (0,1) $.
Then $ t = \text{dim}_H(\tau_{\epsilon}) $ is bounded between these zeros, by Bowen's theorem.
Furthermore, as $ \epsilon \rightarrow 0 $ in the sequence space $ E_{A^E}^n $, we have $ \delta \rightarrow 0 $, so these two zeros approach $ \text{dim}_H(\tau_{\epsilon}) $.
From Equation \ref{another2}, the zeros of $ p^{\pm}(t) $ vary continuously with $ a $, and thus $ \text{dim}_H(\tau_{\epsilon}) $ also does.
For the final statement, we refer to the explicit formula for $ p^{\pm}(t) $ given in Equation \ref{another2}.
For a specific choice of $ \epsilon $, $ \delta $, and $ a $, we can estimate the roots of $ p^{\pm}(t) $.
These are upper and lower bounds on $ \text{dim}_H(\tau_{\epsilon}) $, which improve as $ \epsilon \rightarrow 0 $.
\end{proof}
\subsection{Numerical results for dimension}
\label{Numres}
Finally, we turn to the numerical problem of estimating the Hausdorff dimension of $ \tau $.
As before, by Equation \ref{tauepeq} it suffices to estimate the Hausdorff dimension of $ \tau_{\epsilon} $ for any $ \epsilon > 0 $.
In this section, we will make specific choices of $ \epsilon $ and $ a $, and derive explicit upper and lower estimates on $ \text{dim}_H(\tau_{\epsilon}) $.
Consider $ p^+(t) $ as defined in Equation \ref{another2}.
The following establishes an upper bound on $ p^+(t) $ and hence on $ p(t) $.
\begin{align*}
p^+(t) &< \lim_{n \to \infty} \frac{1}{n} \log \sum_{i_1 = N_{\epsilon}}^{\infty} \sum_{i_2 = N_{\epsilon}}^{\infty} \cdots \sum_{i_n = N_{\epsilon}}^{\infty} \left| \frac{2\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{n-1} + \delta}{i_2^2 \cdots i_n^2} \right|^t \\
&= \lim_{n \to \infty} \frac{1}{n} \log \sum_{i=N_{\epsilon}}^{\infty} \left( \frac{2\pi^{-1} K^{\frac{3}{2}}}{i^{\frac{5}{2}}} \right)^t + \lim_{n \to \infty} \frac{1}{n} \log \left( \sum_{j=N_{\epsilon}}^{\infty} \left( \frac{(2\pi)^{-2} aR^2 + \delta}{j^2} \right)^t \right)^{n-1} \\
&=\log \sum_{j=N_{\epsilon}}^{\infty} \left( \frac{(2\pi)^{-2} aR^2 + \delta}{j^2} \right)^t
\end{align*}
Let $ t=t^{\ast} $ be the unique zero of this upper bound.
Since $ p(t) $ and $ p^+(t) $ are strictly decreasing, we have that $ \text{dim}_H(\tau_{\epsilon}) < t^{\ast} $ by Bowen's theorem.
A lower bound for $ p^-(t) $ is more delicate.
For a given $ \epsilon $, choose $ M \in \mathbb{N} $ with $ M > N_{\epsilon} $.
Then we have the following lower bound.
$$
p^-(t) > \lim_{n \to \infty} \frac{1}{n} \log \sum_{i_1 = N_{\epsilon}}^M \sum_{i_2=N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_1^2} \cdots \sum_{i_n=N_{\epsilon}}^{\lfloor C \rfloor + \lfloor K \rfloor i_{n-1}^2} \left| \frac{2\pi^{-1} K^{\frac{3}{2}}}{i_1^{\frac{5}{2}}} \cdot \frac{\left((2\pi)^{-2} aR^2\right)^{n-1} - \delta}{i_2^2 \cdots i_n^2} \right|^t
$$
Let $ t=t_{\ast} $ be the unique zero of the right and side.
Again since $ p(t) $ and $ p^-(t) $ are strictly decreasing, we have that $ t_{\ast} < \text{dim}_H(\tau_{\epsilon}) $.
Recall that the constants $ C,K $ are defined in terms of $ a $ in Equation \ref{defCK}.
The constant $ N_{\epsilon} $ is defined in the proof of Proposition \ref{gamma1param}, and from the proof of Proposition \ref{trans2} we can show that $ N_{\epsilon} \sim \lceil \frac{K}{\epsilon} \rceil $.
Let $ \delta > 0 $ be small, and choose $ \epsilon > 0 $ small enough that Equation \ref{another2} holds.
Substituting the values of the constants $ C,K $ and $ N_{\epsilon} $ into this equation, we can use a computer algebra system to numerically estimate $ t^{\ast} $ and $ t_{\ast} $.
For example, choose the following numerical values:
$$
\delta = \epsilon = 0.01, \quad a=10, \quad R=0.5.
$$
Substituting these into the values of $ C,K, N_{\epsilon} $ in Equation \ref{another2}, and numerically estimating $ t_{\ast} $ and $ t^{\ast} $ in Mathematica, we obtain
$$
0.40105 < \text{dim}_H(\tau) < 0.51826.
$$
The lower bound can be improved by choosing larger values of $ M $ and $ n $ in the lower approximation of $ p^-(t) $ above.
\subsection{The Hausdorff dimension of $ \mathcal{M} $}
From the dimension results for $ \tau $ we obtain results for $ \mathcal{M} $.
First, we have a corollary of Theorem \textbf{E}.
\begin{corollary*}
Let $ \mathcal{M} $ be the Kuperberg minimal set.
Then the three-dimensional Lebesgue measure of $ \mathcal{M} $ is zero, and $ 2 < \text{dim}_H(\mathcal{M}) < 3 $.
\end{corollary*}
\begin{proof}
By Theorem \ref{minchar}, $ \mathcal{M} $ has a local product structure of $ \mathbb{R}^2 \times \tau $.
As a consequence of Theorem \textbf{E}, the product Lebesgue measure is zero.
A standard result in dimension theory (see \cite{Matt} or \cite{Fal1}) states that if $ X $ and $ Y $ are subsets of Euclidean space, and the Hausdorff dimension of $ Y $ is equal to its upper box dimension, then
$$
\text{dim}_H(X \times Y) = \text{dim}_H(X) + \text{dim}_H(Y).
$$
Applying this to the product structure we obtain
$$
\text{dim}_H(\mathcal{M}) = 2 + \text{dim}_H(\tau),
$$
and the result follows from Theorem \textbf{E}.
\end{proof}
Using the product structure in the above proof, we have the following corollary of Theorem \textbf{F}.
\begin{corollary*}
Let $ \mathcal{M} $ be the Kuperberg minimal set.
Let $ t = \text{dim}_H(\mathcal{M}) $ be its Hausdorff dimension, and $ a > 0 $ the angular speed of the Kuperberg flow.
\begin{itemize}
\item $ t = \text{dim}_H(\mathcal{M}) $ is the unique zero of a dynamically defined pressure function,
\item $ t $ depends continuously on $ a $,
\item For any $ a $ we may compute $ t $ to a desired level of accuracy.
\end{itemize}
\end{corollary*}
Because of this corollary, for the choice of $ \delta, \epsilon, a $ and $ R $ above we have
$$
2.40105 < \text{dim}_H(\mathcal{M}) < 2.51826.
$$
\vfill
\eject
\section{Further questions}
\label{Furth}
There are many remaining open questions about Kuperberg flows.
Some of these are surveyed in \cite{Hur3}.
In this section, we will state some open questions that pertain to the dimension theory of minimal sets of Kuperberg flows.
\subsection{Efficient algorithms for dimension estimates}
The method that yields the numerical results from Theorem \textbf{F} is not particularly efficient.
The the zeros of the upper and lower bounds on $ p^{\pm}(t) $ are computationally expensive to estimate.
For this reason, we cannot fully explore the possible range of the Hausdorff dimension over the parameter space.
\begin{question*}
Design a more efficient algorithm for computing the Hausdorff dimension of the transverse Cantor set of the Kuperberg minimal set.
\end{question*}
In the course of the proof of Theorem \textbf{E}, we showed that the ratio geometry of the transverse Cantor set $ \tau_{\epsilon} $ for $ \epsilon > 0 $ is asymptotically stationary.
The dimension theory for limit sets of iterated function systems whose symbolic dynamics are semiconjugate to a subshift of finite type is classical.
For stationary systems, Bowen's equation for dimension reduces to an equation involving the spectral radius of the incidence matrix (see Chapter 7 of \cite{Pes2}).
The proof of this result relies on a theorem of Ruelle relating the pressure to the spectral radius of the Perron-Frobenius operator.
Solving the spectral radius equation is more computationally efficient than calculating the zeros of the pressure, so an answer to this question might be along these lines.
\subsection{Hausdorff measure of the minimal set}
A more delicate problem than determining Hausdorff dimension is proving that the Hausdorff measure at dimension is finite.
In general, for the limit set of a finitely generated iterated function system or a geometric construction to have finite Hausdorff measure at dimension, the dynamics on the sequence space must be topologically mixing.
For subshifts of finite type this is equivalent to transitivity of the incidence matrix (see \cite{Pes2} or \cite{Bar2}).
\begin{question*}
Let $ \mathcal{M} $ be the Kuperberg minimal set, let $ t=\text{dim}_H(\mathcal{M}) $ be its Hausdorff dimension, and let $ H^t $ be the $ t $-dimensional Hausdorff measure.
Show that $ 0 < H^t(\mathcal{M}) < \infty $.
\end{question*}
In \cite{Pes1}, Pesin and Weiss showed that the limit set of a geometric construction has finite Hausdorff measure at dimension, provided that the eigenmeasure of the Perron-Frobenius operator is a Gibbs state.
A Perron-Frobenius operator in the context of pseudo-Markov systems is studied in \cite{Str}.
By transferring the definition there to the notation developed in Sections \ref{Thermo} and \ref{Conf}, it seems possible to prove an analogue of this result for $ \tau $, and then extend to $ \mathcal{M} $ by the product structure.
\subsection{Ergodic properties of invariant measures}
The ergodic theory of measures invariant under the Kuperberg flow appears to be very difficult. However, Theorems \textbf{A} and \textbf{C} appear to offer a foothold onto this problem.
For small $ \epsilon > 0 $ the transverse minimal set $ \tau_{\epsilon} $ is a limit set of a function system modeled on a sequence space $ \Sigma $ that is invariant under the Kuperberg pseudogroup.
Let $ \mu $ be a measure on $ \Sigma $ invariant under the pseudogroup.
Then the pushforward $ \pi_{\ast} \mu $ through the coding map is a measure on the transverse minimal set, invariant under the Kuperberg flow.
From the product structure given in Theorem \ref{minchar}, a global measure on $ \mathcal{M} $ can be disintegrated along the leaves to obtain the product of a measure on the leaves with a measure on the transversal.
As long as the conditional measures on the leaves are absolutely continuous, one can reduce to studying the ergodic properties of the transverse measures on $ \tau $ and therefore to those on $ \Sigma $, which seems more tractable.
A measure invariant under the Kuperberg flow must have zero entropy as a consequence of a theorem of Katok (\cite{Kat}); this was pointed out by Ghys (\cite{Ghy}).
The Kuperberg plug contains an open set of wandering points and as such cannot preserve a measure supported on open sets; this was pointed out by Matsumoto (\cite{Mat}).
Any other question related to the ergodic properties of invariant measures of the Kuperberg flow appears to be wide open.
\subsection{Dimension of minimal sets of perturbations of Kuperberg flows}
Kuperberg flows are not structurally stable.
In \cite{Hur2}, Hurder and Rechtman defined a class of plugs $ K_{\epsilon} $ supporting a $ C^{\infty} $ flow, for which $ K_0 $ is the Kuperberg plug with no periodic orbits, but $ K_{\epsilon} $ for $ \epsilon > 0 $ has infinitely many periodic orbits.
They showed further that the minimal set of $ K_{\epsilon} $ has embedded horseshoes.
It would be simple to construct such a class $ K_{\epsilon} $ compatible with the assumptions we have made in Section \ref{InsertAssume}.
For $ K_0 $ we would recover the symbolic dynamics and dimension results from this paper, and for $ K_{\epsilon} $ with $ \epsilon > 0 $ we would obtain more standard results (positive entropy, uniform hyperbolicity, etc.).
The dimension theory of horseshoes is well studied (see \cite{Man}, \cite{Sim2}, \cite{Sim3}).
It would be interesting to see how the dimension and symbolic dynamics change as $ \epsilon \rightarrow 0 $.
\subsection{Dimension of minimal sets of generic Kuperberg flows}
The minimal set we have studied is that of a very particular Kuperberg flow.
To simplify our calculations, we have made numerous assumptions on the flow, insertion maps, and insertion regions.
These are listed in Section \ref{InsertAssume}.
However, Theorem \ref{minchar} is true under much weaker assumptions, the axioms of a \textit{generic} Kuperberg flow defined by Hurder and Rechtman.
These are listed in Chapter 12 of \cite{Hur}.
It would be interesting to see what results from Theorems \textbf{A} -- \textbf{F} survive in this generality.
\vfill
\eject
|
\section{\label{sec:level1}Introduction}
Heavy fermion $f$-electron intermetallics continue to attract interest because many of them exhibit complex phase diagrams with diverse phenomena including nematic electronic states, charge and spin instabilities, and unconventional superconductivity.~\cite{Stewart01,Rosch07,Gegenwart08,Pfleiderer09,brando16} In many cases this is related to a competition between the RKKY and Kondo interactions, which mediate magnetism and compensate localized spins, respectively.~\cite{Doniach_77,kondo,ruderman,kasuya,yosida} The fine balance between these interactions can cause a magnetic ordering temperature to be continuously suppressed towards zero temperature at a quantum critical point (QCP). As this occurs novel behaviors often emerge including the breakdown of Fermi liquid behavior and the emergence of superconductivity.~\cite{Pfleiderer09,mathur98,thompson12} This has led to a viewpoint that quantum critical fluctuations of an order parameter are key for producing novel phenomena, and this phenomenology even spans diverse families of materials that are distinct from $f$-electron intermetallics: e.g., cuprate and iron-based superconductors,~\cite{ramshaw15,paglione10} organic superconductors,~\cite{leyraud09} charge density wave systems,~\cite{gruner17} and others.
There nonetheless remain many open questions, including how the specific type of magnetism being suppressed influences a quantum critical region. The earliest theories of quantum phase transitions focused on ferromagnetism,~\cite{stoner} and the seminal work of Hertz and Millis predicted that a ferromagnetic phase transition would remain continuous to zero temperature.~\cite{hertz,millis} More recent work by Belitz-Kirkpatrick-Vojta (BKV) demonstrates instead that for clean materials in two and three dimensions a zero temperature transition from ferromagnetism to paramagnetism is discontinuous.~\cite{brando16, Belitz12, belitz99,belitz05,belitz15} The first order change at the phase boundary prevents diverging fluctuations of the magnetic order parameter. This is in contrast to what is seen near second order antiferromagnetic QCPs where the order parameter diverges, and may interfere with phenomena such as unconventional superconductivity. The BKV theory also predicts that there is a tricritical point that separates a high temperature line of second order phase transitions from a low temperature line of first order phase transitions, where the application of a magnetic field produces wing-like second order phase boundaries that intercept zero temperature. In disordered systems the tricritical point is pushed below zero temperature and the second order phase boundary extends to zero temperature. This has spurred interest in ferromagnetic quantum criticality in disordered metals, where an intriguing possibility is that they might host anomalous metallic states and even unconventional superconductivity.~\cite{belitz99, belitz05, belitz15, brando16,huang13, Nakatsuji08} It is noteworthy that while there are U- and Yb- based ferromagnetic superconductors~\cite{saxena2001,huy,aoki2001,akazawa_2004,ner11}, there are no cerium-based analogues despite some electronic similarities between the associated 5$f$ and 4$f$ states.
CePd$_2$P$_2$ was recently reported to be a correlated electron ferromagnet crystallizing in the well-known ThCr$_2$Si$_2$-type structure~\cite{tran14,ikeda15}, while its isoelectronic volume compressed analogue CeNi$_2$P$_2$ exhibits a nonmagnetic ground state.~\cite{jeitschko,chen,Dra} This suggests that the Pd $\rightarrow$ Ni alloy series could host a ferromagnetic QCP. We synthesized single crystal specimens of Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ for 0 $<$ $x$ $<$ 1, where the contracting unit cell volume applies a chemical pressure. X-ray diffraction and magnetic susceptibility measurements show that the cerium ions remain nearly trivalent up to $x$ $\approx$ 0.66, where the rate of unit cell volume contraction increases, signaling a change in the $f$-electron valence. X-ray absorption spectroscopy measurements for CePd$_2$P$_2$ and CeNi$_2$P$_2$ reinforce this view by revealing trivalent and trivalent with a small fraction of tetravalent $f$-electron character, respectively. Features associated with the ferromagnetic ordering are evident for 0 $<$ $x$ $\lesssim$ 0.69 in the magnetic susceptibility, heat capacity, and electrical resistivity, where the ordering temperature is continuously suppressed towards zero at an extrapolated critical value of $x_{\rm{cr}}$ $\approx$ 0.7. In the critical region there is chemical disorder which allows the phase transition to remain second order, even as it approaches zero temperature. This results in a putative ferromagnetic QCP, around which there are indications for a breakdown of Fermi liquid behavior: in particular, the heat capacity divided by temperature $C/T$ diverges nearly logarithmically with decreasing $T$. There is also evidence that the disorder contributes to the unusual temperature dependences by producing a quantum Griffiths phase that extends over a broad $x$-range.~\cite{Griffiths1969,vojta,sereni07, Westerkamp09} We furthermore find that for clean CePd$_2$P$_2$, applied pressure initially suppresses the ferromagnetism in a manner similar to that of Pd $\rightarrow$ Ni substitution. Therefore, this system offers the opportunity to study behavior at a disordered ferromagnetic quantum critical point and eventually to compare to the ordered analogue, as is needed to test the BKV theory and to ultimately design new QCP materials.
\section{\label{sec:level1}Experimental Methods}
Single crystals of Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ were grown from elements with purities \(>\) 99.9\% in a molten flux of Pd, Ni and P. The reaction ampoules were prepared by loading the elements in the ratio Ce:Pd:Ni:P $;$ 1:11(1-$x$):11$x$:8 into a 2 mL alumina crucible for each of the nominal Ni concentrations. The crucibles were sealed under vacuum in quartz ampoules and heated to $300\,^{\circ}\mathrm{C}$ at a rate of $50\,^{\circ}\mathrm{C}$/hour, held at $300\,^{\circ}\mathrm{C}$ for 6 hours, heated to $500\,^{\circ}\mathrm{C}$ at a rate of $50\,^{\circ}\mathrm{C}$/hour, held at $500\,^{\circ}\mathrm{C}$ for 6 hours, heated to $1180\,^{\circ}\mathrm{C}$ at a rate of $50\,^{\circ}\mathrm{C}$/hour, kept at $1180\,^{\circ}\mathrm{C}$ for 3 hours, and then cooled at a rate of $2\,^{\circ}\mathrm{C}$/hour to $1000\,^{\circ}\mathrm{C}$. At this temperature, the remaining flux was separated from the crystals by centrifuging. Single-crystal platelets with typical dimensions of several millimeters on a side and several millimeters in thickness were collected.
The crystal structure and chemical composition were verified by powder x-ray-diffraction (XRD) and energy dispersive spectrometer (EDS) analysis. EDS results are shown in Fig.~\ref{xrd}a, where the measured concentration $x_{\rm{act}}$ is compared to the nominal concentration $x_{\rm{nom}}$. Throughout the rest of the manuscript we use $x_{\rm{act}}$ unless otherwise specified. Magnetization $M$($T,H$) measurements were carried out for single crystals at temperatures $T$ $=$ 1.8 $-$ 300 K under an applied magnetic field of $H$ $=$ 5 kOe for $H$ applied both parallel ($\parallel$) and perpendicular ($\perp$) to the $c$ axis using a Quantum Design VSM Magnetic Property Measurement System. The AC magnetic susceptibility $\chi$'($T$) for selected concentrations was also measured using the same apparatus. Electrical resistivity $\rho$ measurements for temperatures $T$ $=$ 0.5 $-$ 300 K were performed in a four-wire configuration and the heat capacity $C$ was measured for $T$ $=$ 0.39 $-$ 20 K using a Quantum Design Physical Property Measurement System. $\rho$($T$) measurements under applied pressure were performed using a piston cylinder pressure cell with Daphne 7474 oil as the pressure transmitting medium. The pressure is determined by the shift in ruby flourescence peaks and are the values determined below $T$ $=$ 10 K. These measurements were performed at the National High Magnetic Field Laboratory DC field User facility using standard He3 cryostats.
Samples were analyzed using Ce L$_3$-edge X-ray absorption spectroscopy (XANES) at the Stanford Synchrotron Radiation Lightsource (SSRL) on beam line 11-2. Single crystals of the compounds were ground and diluted with boron nitride and painted onto 0.5 mil Kapton tape, the tape was attached to an aluminum sample plate and loaded into a liquid helium cryostat. A single energy was selected using a liquid-N2-cooled double-crystal monochromator utilizing Si(220) ($\phi$ = 0) crystals. The crystals were detuned by 70\% at 6100 eV to remove higher order harmonics. Spectra were measured in fluorescence mode using a Lytle detector equipped with a Ti filter (3 absorption lengths) at two different temperatures, 85 K and 10 K. A Cr-calibration foil was placed downstream of the sample and spectra were calibrated to the first reflection point of Cr(5989.0 eV). Using the Athena11 software package, L$_3$-edge spectral were background subtracted and normalized at $E_0$ (5723 eV). A deconvoluted model for the Ce L$_3$-edge XANES data was obtained using a modified version of EDG-FIT~\cite{nolas} in IGOR 6.0. Using this least-squares algorithm spectra were modeled with a minimum number of pseudo-Voight functions (50:50 Lorentzian:Gaussian) and a 1:1 ratio of arctangent and error function. The areas under the pre-edge peaks (hereafter defined as the intensity) were equal to the FWHM x peak height.
Single energy images, elemental maps, and Ce M$_{5,4}$-edge x-ray absorption spectra (XANES)~\cite{ev07} were acquired using the scanning transmission x-ray microscope (STXM) instrument at the spectromicroscopy beamline 10ID-1 at the Canadian Light Source (CLS) according to data acquisition methodology described previously.~\cite{min17,man18, ha18}
\section{\label{sec:level1}Results}
Powder X-ray diffraction measurements show that the ThCr$_2$Si$_2$-type structure persists across entire the Pd $\rightarrow$ Ni substitution series, while the tetragonal lattice constants ($a$ and $c$) and the unit cell volume ($V$) decrease with increasing $x$ (Fig.~\ref{xrd}). Up to $x$ $\approx$ 0.66 the trends are consistent with Vegard's law, where the linear lattice contraction is due to the smaller size of Ni by comparison to Pd. This suggests that over this $x$-range the room temperature Ce valence remains constant. For $x$ $\gtrsim$ 0.66 the unit cell volume continues to decrease linearly, but with a larger slope, signaling a change in the cerium valence. The volume contraction results in a chemical pressure which is estimated to be near $P_{\rm{ch}}$ $=$ 7.5 GPa for $x$ $=$ 0.66 and $P_{\rm{ch}}$ $=$ 13.6 GPa at $x$= 0.96. These values are calculated using the Birch-Murnaghan equation $P_{\rm{ch}}$ $=$ $B_{\rm{0}}$$\Delta$$V(x)/V(0)$, where the value of the bulk modulus for CeCu$_2$Si$_2$ ($B_{\rm{0}}$ $=$ 110 GPa) is used.~\cite{spain86}
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=1\linewidth]{xrd.pdf}
\caption{ (a) Comparison between the measured phosphorous concentration $x_{\rm{act}}$ and the nominal concentration $x_{\rm{nom}}$, where $x_{\rm{act}}$ was determined using energy dispersive spectrometer analysis. Throughout the rest of the manuscript we use $x_{\rm{act}}$ $=$ $x$ unless otherwise specified. (b) The lattice constants, $a(x)$ (left axis) and $c(x)$ (right axis). (c) the unit cell volume $V(x)$ (left axis) and chemical pressure $P_{\rm{ch}}$(right axis), calculated using the Birch-Murnaghan equation as described in the text with the bulk modulus $B_{0}$ $=$ 110 GPa.
}
\label{xrd}
\end{center}
\end{figure}
The magnetic susceptibility $\chi$ $=$ $M/H$ vs temperature and magnetization $M$ vs $H$ for $H$ $\parallel$ $c$ data are shown in Fig.~\ref{chi}. As previously reported for polycrystalline specimens,~\cite{tran14} ferromagnetic ordering appears in $\chi(T)$ for $x$ $=$ 0 as a sharp increase at $T_{\rm{C}}$ $=$ 28.5 K, which we define as the peak in $\partial$$\chi$/$\partial$$T$ (not shown). For $T$ $\leq$ $T_{\rm{C}}$, $M(H)$ rapidly saturates towards $M_{\rm{sat}}$ $=$ 1.93 $\mu_{\rm{B}}$. The evolution of the ferromagnetic order with $x$ is determined using these quantities, where $T_{\rm{C}}$ decreases linearly and is extrapolated to intercept zero temperature near $x_{\rm{cr}}$ $\approx$ 0.7. The persistence of the ferromagnetism into the critical $x$-region is seen in the $M(H)$ curves, which remain hysteretic even as $M_{\rm{sat}}$ is smoothly suppressed (Figs.~\ref{chi}d and e). For $x$ $\gtrsim$ $x_{\rm{cr}}$, the magnitudes of $\chi(T)$ and $M(H)$ continue to decrease and become similar to that of paramagnetic CeNi$_2$P$_2$ as $x$ approaches 1.~\cite{chen}
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=1\linewidth]{mag.pdf}
\caption{(a) Magnetic susceptibility $\chi$ $=$ $M/H$ vs temperature $T$ for $H$ $=$ 0.1 T applied parallel $\parallel$ to the $c$-axis for Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$. (b) The inverse of Magnetic susceptibility $\chi_{\rm{avg}}$$^{-1}$($T$) for representative curves, where $\chi_{\rm{avg}}$ = (2$\chi_{ab}$ + $\chi_c$)/3. The dotted lines are Curie-Weiss fits to the data for 150 K $<$ $T$ $<$ 300 K. (c) The magnetic anisotropy $\chi_{\perp}$/$\chi_{\parallel}$ vs $x$ at $T$ $=$ 2 K. (d) Magnetization field dependence for different $x$-concentrations measured at 1.8K. (e) Zoom in low field of $M$ vs $H$ for different $x$-concentrations measured at 1.8K.
}
\label{chi}
\end{center}
\end{figure}
The high temperature Curie-Weiss behavior provides further insight into the $x$-evolution of the $f$-electron state, magnetocrystalline anisotropy, and strength of the Kondo hybridization. There is a strong evolution in the magnetic anisotropy $\chi_{\perp}$/$\chi_{\parallel}$ between the $c$ and $a$$b$ directions: while it decreases by roughly a factor of 10 to approach $\chi_{\perp}$/$\chi_{\parallel}$ $\approx$ 20 near $x$ $\approx$ 0.35, it recovers to nearly the $x$ $=$ 0 value at $x$ $\approx$ 0.6 and finally decreases to become isotropic at $x$ $=$ 1 (Fig.~\ref{chi}c). In order to analyze this data using Curie-Weiss fits, we calculate the average susceptibility, defined as $\chi_{\rm{avg}}$ = (2$\chi_{ab}$ + $\chi_c$)/3 (Fig.~\ref{chi}b). For $x$ = 0, $\chi_{\rm{avg}}$ is consistent with earlier results for polycrystalline specimens~\cite{tran14}, yielding an effective magnetic moment $\mu_{\rm{eff}}$ = 2.4 $\mu_{\rm{B}}$ (trivalent cerium) and a Curie-Weiss temperature $\theta$ = 2 K. Pd $\rightarrow$ Ni substitution causes $\theta$ to increase to a value near -193 K at $x$ = 0.66, and afterwards to even larger negative values. This is a common feature in Ce-based materials with strong hybridization between the $f$- and conduction electrons and indicates that the Kondo interaction strengthens with increasing $x$.~\cite{sereni07}
To further evaluate the effect of Pd $\rightarrow$ Ni substitution on the Ce $f$-electron state, Ce L$_3$-edge X-ray absorption spectra (XANES) were obtained from single crystals of CeNi$_2$P$_2$ and CePd$_2$P$_2$, the end members of the series. Ce$^{3+}$ usually show a single absorption peak at ca. 5723-5725 eV, whereas Ce$^{4+}$ exhibits a ``double-white line" feature with maxima at ca. 5724-5728 and 5736-5739 eV. As shown in Fig.~\ref{xas}, the Ce L$_3$-edge spectrum from CePd$_2$P$_2$ is typical of Ce$^{3+}$, containing a single pronounced absorption peak with a maximum at 5725.2 eV (FWHM = 6.60 eV). Although substituting Ni for Pd has essentially no impact on the main absorption peak energy (maximum = 5725.1 eV), subtle spectral changes emerges; most notably, the main absorption peak broadens by 1.24 eV (FWHM = 7.84 eV) and a minor post-edge feature emerges near 5735 eV. Additionally, spectra obtained from CeNi$_2$P$_2$ and CePd$_2$P$_2$ are unchanged as a function of temperature between 85 and 10 K.
To characterize the origin of these changes, the Ce L$_3$-edge XANES spectra are modeled using a least-squares algorithm between 5705 and 5741 eV. Although the entire fitted region for CePd$_2$P$_2$ is easily modeled by the combination of a single peak at 5725.2 eV and a step-function at 5724.4 eV, an analogous model for CeNi$_2$P$_2$ does not adequately represent the data. In particular, there is substantial misfit associated with the post-edge feature that is absent in the CePd$_2$P$_2$ spectrum. Hence, three peaks and a step-function are needed to model the CeNi$_2$P$_2$ spectrum. The main absorption peak is at 5725.1 eV, the step function at 5723.0 eV, and two post-edge peaks are fit at 5732.6 and 5735.7 eV. We believe it is of no coincidence that the highest-energy post-edge peak (at 5735.7 eV) occurs at a similar energy to the higher energy peak of the ``double-white line" feature typically observed for Ce$^{4+}$ (between 5736 and 5739 eV).~\cite{
bianconi1987,kaindl1988,sham2005,walter2009,loble2015,bogart2015,cary2016,kratsova2016,antonio2017,toscani2016} Hence, these data are interpreted as indicating that CeNi$_2$P$_2$ contains a mixture of Ce$^{3+}$ and Ce$^{4+}$. Comparison of the intensities (FWHM x peak height) of the main absorption peak (intensity = 8.8) with the small post-edge peaks (intensities = 0.2 and 0.4) suggests that CeNi$_2$P$_2$ contains on the order of 7(1)\% Ce.
XANES data at the M$_{5,4}$-edge were obtained to corroborate the L$_3$-edge measurements. The spectroscopic approach can be advantageous for probing 4$f$ orbital occupation and mixing, especially for systems with mixed valence or multiconfigurational ground states.~\cite{bianconi1987,dex87,kaindl1988,ani11, ani12,hu00,kaindl84,le85,ra92} The background subtracted and normalized M$_{5,4}$-edge spectra for CeNi$_2$P$_2$ and CePd$_2$P$_2$ are provided in Fig.~\ref{XANES}. The M$_{5,4}$-edge spectra are split into low energy M$_5$ (3$d_{5/2}$) and high energy M$_4$ (3$d_{3/2}$) edges due to spin-orbit coupling with the 3$d$ core hole. For both CeNi$_2$P$_2$ and CePd$_2$P$_2$ as well as the CeCl$_6$$^{3-}$ reference, the M$_{5,4}$-edge exhibits a characteristic ``sawtooth" pattern with fine structure that closely resembles expectations from theory for a 3$d^9$4$f^2$ final state and an isolated Ce$^{3+}$ ion.~\cite{le85} For the Ce$^{4+}$ reference, CeCl$_6$$^{2-}$, both the M$_5$- and M$_4$-edges are split into intense main peaks and additional satellite features about 5 eV higher in energy.~\cite{le85, am16, hu97, jean98, dong03} Upon close inspection, weak satellite features are also present 887.3 and 905.4 eV in the Ce M$_{5,4}$-edge spectrum for CeNi$_2$P$_2$. Previous calculations have attributed the presence of satellite features in the M$_{5,4}$-edge XANES spectra of formally Ce$^{4+}$ compounds to interaction of 3$d^9$4$f^1$ and 3$d^9$4$f^2$ configurations in the final state.~\cite{ra92, ani13} In this general sense the M$_{5,4}$-edge XANES spectrum of CeNi$_2$P$_2$ resembles that of the molecular compound (Et$_4$N)$_2$CeCl$_6$ in addition to extended solids and intermetallics such as CeO$_2$ and CeRh$_{3}$.~\cite{kaindl84} Because transitions associated with the 3$d^9$4$f^1$ and 3$d^9$4$f^2$ final states are not well-resolved in the M$_{5,4}$-edge spectra, the intensity of the satellite features cannot be directly related to the amounts of Ce$^{4+}$, or 4$f^0$ character in the ground state for CeNi$_2$P$_2$. However, the presence of small satellite features in the Ce M$_{5,4}$-edge spectrum for CeNi$_2$P$_2$, and the lack thereof for CePd$_2$P$_2$, is consistent with the observation of $<$ 10\% Ce$^{4+}$ character in the ground state of CeNi$_2$P$_2$.
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=1\linewidth]{XAS.pdf}
\caption{The experimental data (dots) and the curve fitted model (blue trace) for the Ce L$_3$-edge X-ray absorption spectra of CePd$_2$P$_2$ (top) and CeNi$_2$P$_2$ (bottom). The pre-edge pseudo-Voigt functions (green, blue, and red traces) used to generate the model and the step function (grey dashed lines) are shown.
}
\label{xas}
\end{center}
\end{figure}
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=1\linewidth]{Stosh.pdf}
\caption{Cerium M$_{5,4}$-edge XANES spectra for CeNi$_2$P$_2$ (blue trace) and CePd$_2$P$_2$ (red trace) along with Ce$^{3+}$ and Ce$^{4+}$ reference compounds (dashed black and gray traces). The [Ph$_4$P]$_3$CeCl$_6$ and [Et$_4$N]$_2$CeCl$_6$ data are adapted with permission from ref.~\cite{loble2015}.
}
\label{XANES}
\end{center}
\end{figure}
The heat capacity $C_{4f}$ divided by $T$ vs. $T$ data are shown in Fig.~\ref{HC}a, which further expose the ordered state and underlying electronic behavior. The $x$ $=$ 0 ferromagnetism appears as a lambda-like feature near $T_{\rm{C}}$ $=$ 28.5 K, consistent with a second order phase transition. With increasing $x$, $T_{\rm{C}}$ moves to lower temperatures and up to $x$ $\approx$ 0.35 the shape of phase transition is preserved but its overall size grows. This indicates that even as $T_{\rm{C}}$ is suppressed the associated entropy is conserved. Between 0.35 $<$ $x$ $\lesssim$ 0.69, the ferromagnetic feature broadens and is superimposed on an increasing background. The broadening of the phase transition is attributed to chemical/structural disorder which is maximal near the middle of the substitution series. As $T_{\rm{C}}$ approaches zero near $x_{\rm{cr}}$ $\approx$ 0.7, $C_{4f}/T$ diverges nearly continuously down to 0.5 K. This is a common feature of `non-Fermi-liquid' behavior near a quantum critical point in correlated $f$-electron materials and may be associated with quantum critical fluctuation of the magnetic order parameter.~\cite{Stewart01,Rosch07,Gegenwart08,Pfleiderer09,brando16} For larger $x$ the divergence weakens and finally tends to saturate at low temperature for $x$ = 0.96 in a manner that is similar to CeNi$_2$P$_2$, indicating the recovery of the paramagnetic Fermi liquid state.
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=1\linewidth]{HC.pdf}
\caption{ (a) The heat capacity $C_{4f}$ divided by temperature $T$ vs $T$ following the phonon background subtraction for select concentrations of Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$, $C_{4f}/T$ is calculated by subtracting $C/T$ for the nonmagnetic analogue La(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ from that of Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$. (b) 4$f$ entropy $S_{\rm{4f}}$ vs $x$. $S_{\rm{4f}}$ is obtained from the heat capacity data as described in the text. }
\label{HC}
\end{center}
\end{figure}
The 4$f$ contribution to the entropy $S_{4f}$ vs. $T$ is shown in Fig.~\ref{HC}b. $S_{4f}$ was calculated by subtracting $C/T$ for the nonmagnetic analogue La(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ from that of Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ and subsequently integrating from 0.5 K. The nonmagnetic lattice term was approximated by summing the heat capacities of LaPd$_2$P$_2$ and LaNi$_2$P$_2$ in the ratios (1-$x$):$x$. While this approach slightly underestimates the total 4$f$ entropy and only approximates the lattice contribution to the heat capacity, it provides a consistent way to assess the evolution of $S_{4f}$ with $x$. $S_{4f}$ reaches 0.88$R$ln2 at $T_{\rm{C}}$ for $x$ $=$ 0. This is slightly reduced from the full entropy of a a doublet ground state and indicates weak Kondo screening of the $f$-moment by the conduction electrons.~\cite{tran14,ikeda15} In the $x$-region where the phase transition remains sharp (0 $\leq$ $x$ $\lesssim$ 0.35), $S_{4f}$ consistently recovers to similar values at $T_{\rm{C}}$, suggesting that the strength of the hybridization changes little over this range. For specimens with larger concentrations that still show ferromagnetism but have broadened phase transitions (0.35 $\lesssim$ $x$ $\lesssim$ 0.69), the entropy recovered at $T_{\rm{C}}$ grows smaller with increasing $x$, revealing strengthening hybridization. For concentrations in the no order region ($x$ $\gtrsim$ $x_{\rm{cr}}$) $S_{4f}$ is significantly reduced from that seen at lower $x$ and increases smoothly with increasing $T$ in a manner consistent with there being strong Kondo hybridization between the $f$- and conduction electrons.\cite{Doniach_77,kondo}
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=3.5in]{rho.pdf}
\caption{(a) and (b) The electrical resistivity normalized to the room temperature value $\rho/\rho_{\rm{300K}}$ vs. temperature $T$ for Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ at values $x$ $=$ 0 $-$ 1. (c) The residual resistivity $\rho_0$ vs $x$. (d) $\rho/\rho_{\rm{300K}}$($T$) collected under applied pressure $P$ $\lesssim$ 19 kbar for CePd$_2$P$_2$.
}
\label{rho}
\end{center}
\end{figure}
The temperature dependences of the electrical resistivity normalized to the room temperature value $\rho/\rho_{300K}$ vs. $T$ for the entire substitution series are shown in Fig.~\ref{rho}. The behavior for $x$ $=$ 0 is consistent with earlier results, where the resistivity decreases with decreasing $T$ and evolves through a kink near $T_{\rm{C}}$ $=$ 28.5 K that further reduces the electronic scattering due to the removal of magnetic fluctuations.~\cite{tran14} Here, the residual resistivity ratio $RRR$ $=$ $\rho/\rho_{\rm{300K}}$ $\approx$ 12. For 0 $\leq$ $x$ $\lesssim$ 0.35 $RRR$ decreases due to increasing disorder, but the reduction in $\rho$/$\rho_{300K}$ at $T_{\rm{C}}$ remains sharp. The $x$ - dependence of the residual resistivity $\rho_0$ is shown in Fig.~\ref{rho}c, where the doping introduces a substantial amount of disorder which results in a large residual resistivity near the critical region ($\rho_0$ $\approx$ 300 \si{\micro \ohm}cm). Based on this, we estimate that the specimens in this concentration range belong to the second regime as described in the BKV theory,~\cite{brando16, Belitz12} where $\rho_0$ is several hundred \si{\micro \ohm}cm. Over this $x$-range the phase transition is preceded in temperature a growing upturn in $\rho$/$\rho_{300K}$ which indicates a gradual strengthening hybridization between the $f$- and conduction electron states. For larger $x$ the phase transition broadens due to increasing disorder and continues to be suppressed until it is no longer visible near $x_{\rm{cr}}$ $\approx$ 0.7.
In order to further examine the tuning mechanisms that control the ordered state in CePd$_2$P$_2$, we performed measurements of the electrical resistivity under hydrostatic pressure (Fig.~\ref{rho}d). We find that the ferromagnetic phase transition is monotonically suppressed with increasing pressure at a rate of 2.4 K/GPa, and from this we estimate that an applied pressure of 12 GPa would be needed to fully suppress the ferromagnetism to zero temperature in the parent compound. In order to directly compare this result to what is seen for Pd $\rightarrow$ Ni substitution, we convert the applied pressure to change in unit cell volume and then associate this value with a Ni concentration. Results for a typical bulk modulus $B_0$ $\approx$ 110 GPa (similar to what is observed for CeCu$_2$Si$_2$~\cite{spain86}) are shown as open stars in Fig.~\ref{phase}a, where the slope of $T_{\rm{C}}$ is weaker than that seen in the substitution series.
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=3.5in]{griffith.pdf}
\caption{(a) AC susceptibility $\chi$' of selected selected $x$ of Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ series. Solid lines are fits to the data using the expression $\chi$' $\propto$ $T^{\lambda-1}$. (b) Magnetic susceptibility $\chi$ with $H$ $=$ 0.1 T applied parallel $\parallel$ $c$ of selected selected $x$ of Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ series. (c) $C_{4f}/T$ for selected $x$ at low temperatures. Solid lines are fits to the data using the expression $C_{4f}/T$ $\propto$ $T^{\lambda-1}$.
}
\label{griffith}
\end{center}
\end{figure}
Finally, in order to assess whether chemical disorder produces magnetic clustering behavior that impacts the low temperature behavior in the large $x$-region, AC magnetic susceptibility $\chi$' measurements were performed for selected concentrations (Fig~\ref{griffith}a). For $x$ $>$ $x_{\rm{cr}}$ the data can be fit using the formula $\chi$' $\propto$ $T^{\lambda-1}$, where $\lambda$ becomes less negative with increasing $x$ and changes sign to become positive for $x$ $=$ 0.83. This type of behavior is expected if chemical disorder produces cluster regions with short range magnetic correlations while the bulk state remains paramagnetic: e.g., as for a quantum Griffiths phase (QGP).~\cite{Griffiths1969,vojta,sereni07, Westerkamp09} Similar fits were carried out for $C/T$ over a broader temperature range, which also reveal a systematic evolution in $\lambda$ that might be consistent with a QGP. Over the same $x$-range, we find that there is a weak and hysteretic increase in $\chi$($T$) that disappears before $x$ = 1 (Fig. 7b). While this feature indicates the persistence of short range ferromagnetic interactions for $x$ $>$ $x_{\rm{cr}}$, it does not appear in other bulk measurements such as the heat capacity. We furthermore point out that the power-law behavior extends over a broad $x$-range, which is in contrast to the contained v-shaped region that is often seen for ordered materials with quantum critical behavior~\cite{Stewart01,Rosch07,Gegenwart08,Pfleiderer09}. While these measurements are suggestive of QGP behavior, further work is still needed: e.g., the trends at lower temperatures should be established.
\begin{figure}[!tht]
\begin{center}
\includegraphics[width=3.5in]{phase.pdf}
\caption{(a) left axis: Temperature $T$ vs. concentration $x$ phase diagram for Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ for $x$ $=$ 0 $-$ 1 constructed from magnetic susceptibility $\chi$ $=$ $M/H$, electrical resistivity $\rho$ and heat capacity $C$ data. Also shown is the ferromagnetic phase boundary that was observed for CePd$_2$P$_2$ under applied pressure. The Birch-Murnaghan equation of state was used to convert from pressure to unit cell volume, and then to the corresponding $x$. The solid and open stars are for $B_0$ $=$ 110 and 220 GPa, respectively. Right axis: Averaged Curie-Weiss temperature $|$ $\theta$ $|$ vs $x$. (b) Left axis: The value of the background subtracted heat capacity divided by temperature $C_{4f}/T$ at $T$ $=$ 0.5 K. Right axis: The 4$f$ contribution to the entropy at $T_{\rm{C}}$ vs $x$. (c) Electrical resistivity normalized to the room temperature value $\rho$/$\rho_{\rm{300K}}$ at $T$ $=$ 0.5 K vs $x$.}
\label{phase}
\end{center}
\end{figure}
\section{\label{sec:level1}Discussion}
Fig.~\ref{phase} shows the $T$-$x$ phase diagram and the evolution of several quantities vs. $x$. $T_{\rm{C}}$ is suppressed linearly with $x$, and is extrapolated to approach zero temperature near $x_{\rm{cr}}$ $\approx$ 0.7. For $x$ $=$ 0 we also plot results from measurements under applied pressure, where chemical and applied pressure both continuously suppress $T_{\rm{C}}$. If we assume $B_0$ $=$ 220 GPa (closed stars) then both chemical and applied pressure suppress $T_{\rm{C}}$ at the same rate. A more realistic value of $B_0$ $=$ 110 GPa (open stars) results in a more gradual suppression of $T_{\rm{C}}$. Regardless of which $B_0$ is chosen, it is clear that the main tuning parameter that controls $T_{\rm{C}}$ is the unit cell volume, which likely changes the relative strengths of the Kondo and RKKY interactions in a Doniach-like scenario.\cite{Doniach_77} This argument is strengthened by considering that for Kondo lattice systems an estimate of the Kondo energy scale can be made using the expression $\chi(T)$ = $C/(T-2T_K)$.~\cite{thy75} As shown Fig.~\ref{phase}a, $\theta$ increases with increasing $x$, suggesting a strengthening Kondo energy scale. For $x$ $\gtrsim$ 0.7 there is also evidence that the $f$-electron state is distinct from what is seen for $x$ $<$ $x_{\rm{cr}}$: i.e., the cerium $f$-valence evolves away from a purely trivalent state. This is revealed through: (1) a deviation from Vegard's law and (2) in the XANES measurements of the end-member compounds, which show that CePd$_2$P$_2$ has a 3+ $f$-electron valence, while CeNi$_2$P$_2$ shows an admixture of 3+ and 4+. A more detailed study of the region near $x_{\rm{cr}}$ is still needed to determine whether or not this change in the substitution series is abrupt or gradual.
An important feature of this substitution series is that the phase transition remains second order across the entire ferromagnetic $x$-regime. This is likely because disorder influences the intermediate substitution region: e.g., as seen in the broadening of the phase transition in heat capacity and the growing residual resistivity that peaks near $x$ $\approx$ 0.6. This provides the conditions that are expected from the BKV theory for a disordered ferromagnetic quantum critical point near $x_{\rm{cr}}$ $\approx$ 0.7. At the same time, there is some evidence that the disorder leads to magnetic clustering behavior that might be consistent with a quantum Griffiths phase. Further work is needed to verify this scenario and its impact on the possible quantum critical behavior, such as measurements at lower temperatures to establish the powerlaw behavior. In future work, it will also be useful to compare to clean and lightly substituted CePd$_2$P$_2$ under applied pressure, where the evolution from first order to second order behavior can be systematically followed.
Finally, we make a comparison to some related materials. For instance, the prototypical antiferromagnetic quantum critical point alloy series CeCu$_2$Si$_{2-x}$Ge$_x$,~\cite{Knebel} shows a qualitatively similar evolution of the low temperature phenomena. In particular, $C/T$ for $x_{\rm{cr}}$ follows a logarithmic in temperature divergence with a value near 0.9 J/mol K$^2$ at low temperatures that is replaced by antiferromagnetic order with increasing unit cell volume. An important difference is that for this system the critical region is near $x$ $=$ 0 and the amount of disorder is small by comparison to what is seen in our series. Another closely related alloy series is CePd$_2$As$_{2-x}$P$_x$,~\cite{shang} which features a transformation from ferromagnetic to antiferromagnetic order going from P $\rightarrow$ As but does not have obvious features associated with quantum criticality or a quantum Griffiths phase. In this series, it appears that the nonisoelectronic chemical substitution mainly tunes the sign of the magnetic exchange. We finally point out that although the alloy series CeRh$_{1-x}$Pd$_x$~\cite{sereni07} crystallizes in a different structure, there are remarkable similarities to what we have observed for Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$: including that the ferromagnetic order is suppressed towards zero temperature in the disordered $x$-region. In the same area of the phase diagram there is evidence for a change of the cerium valence and the presence of a quantum Griffiths phase. It will be useful to study these systems together to assess the universality of model for ferromagnetic quantum criticality, including the BKV theory.
\section{\label{sec:level1}Conclusions}
These results reveal that Ce(Pd$_{1-x}$Ni$_x$)$_2$P$_2$ is a useful example of a cerium-based intermetallic with a disordered ferromagnetic QCP and accompanying breakdown of Fermi liquid behavior. We also find that for CePd$_2$P$_2$ a pressure of $P_{\rm{c}}$ $\approx$ 12 GPa would likely be sufficient to access the tricritical point and first order quantum phase transition that is expected in the clean limit. It will be interesting to compare the electronic states that appear near $x_{\rm{cr}}$ and $P_{\rm{cr}}$ to test expectations from BKV theory. Further comparison to more conventional antiferromagnetic QCPs is also of interest, where an important question is whether unconventional superconductivity can occur near a disordered ferromagnetic QCP.~\cite{belitz99, belitz05, belitz15, brando16,huang13, Nakatsuji08,Griffiths1969,vojta,sereni07, Westerkamp09}
\vspace{3mm}
\section{\label{sec:level1}Acknowledgements}
A portion of this work was performed at the National High Magnetic Field Laboratory (NHMFL), which is supported by National Science Foundation Cooperative Agreement No. DMR-1157490 and DMR-1644779, and the State of Florida. Research of RB, YL, DG, KH, WP, WLN, SAK and TAS were supported in part by the Center for Actinide Science and Technology, an Energy Frontier Research Center funded by the U.S. Department of Energy (DOE), Office of Science, Basic Energy Sciences (BES), under Award Number DE-SC0016568. MC was supported by DOE-BES through Award No. DE-339SC0002613. Portions of this work were supported by the LANL named fellowship program; the Agnew National Security Fellowship (SEB) and the Glenn T. Seaborg Institutes Postdoctoral Fellowship program at LANL (MGF). Use of the Stanford Synchrotron Radiation Lightsource, SLAC National Accelerator Laboratory, was supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences under Contract No. DE-AC02-76SF00515. The SSRL Structural Molecular Biology Program is supported by the DOE Office of Biological and Environmental Research, and by the National Institutes of Health, National Institute of General Medical Sciences (including P41GM103393). This work was also supported by the Joint Plasma Physics Program of the National Science Foundation and the Department of Energy under grant DE-SC0016251 (GTS). SM was supported by the Director, Office of Science, Office of Basic Energy Sciences, Division of Chemical Sciences, Geosciences, and Biosciences Heavy Element Chemistry Program of the U.S. Department of Energy (DOE) at LBNL under Contract No. DE-AC02-05CH11231. M$_{5,4}$-edge spectra described in this paper was measured at the Canadian Light Source, which is supported by the Canada Foundation for Innovation, Natural Sciences and Engineering Research Council of Canada, the University of Saskatchewan, the Government of Saskatchewan, Western Economic Diversification Canada, the National Research Council Canada, and the Canadian Institutes of Health Research.
|
\section{Introduction}
Self-propelled particles represent a system inherently out of equilibrium as they take energy from the environment, convert it into directed motion
and dissipate it to move in a viscous medium \cite{bechinger2016active,ramaswamy2010mechanics,marchetti2013hydrodynamics}.
In recent years, a variety of models have been proposed in order to capture both the
stationary and the time-dependent properties of these systems. Among these proposals, we mention the Run and Tumble \cite{Tailleur2009Sedimentation, Nash2010Run, cates2013active}, the active Brownian particle
(ABP) model \cite{solon2015active, romanczuk2012active, Hagen2011} and the Gaussian colored noise model also termed active Ornstein-Uhlenbeck particle (AOUP) model \cite{szamel2014self,flenner2016nonequilibrium,maggi2015multidimensional}. They all describe the overdamped motion of particles subjected to a drag force, due to the solvent, proportional to the particle's velocity, to a deterministic force, $ F$,
due to an external driving or to interparticle interactions and to the so-called active force or self-propulsion. In the ABP the active force is modeled by a vector of constant norm and whose orientation performs a Brownian
motion on the unit sphere. The orientation of the self-propulsion has a typical persistence time,$\tau$ i.e.
it decorrelates with respect to its initial value exponentially as $\exp(-t/\tau)$.
The existence of such correlation accounts for the persistence of the trajectories
which is the distinguishing feature between the standard model of colloidal particles and the one describing self-propelled particles.
Interestingly,
in the presence of deterministic forces either due to external fields, such as confining walls or to particle-particle interactions self-propelled particles manifest novel phenomena such as a tendency to cluster \cite{fily2012athermal, buttinoni2013dynamical, bialke2015active} and correlations between the positions and the velocities reflecting their non-equilibrium nature.
The AOUP originates from the necessity of reducing the mathematical complexity of the ABP and is constructed by assuming the same deterministic forces as in the ABP but replacing
the ABP active force by an active force whose components have a Gaussian distribution \cite{fily2012athermal,farage2015effective}.
The matching between ABP and AOUP is enforced by requiring that
the active forces of each model have the same variance and the same exponential time-correlation,
but the AOUP admittedly neglects the non-Gaussian nature of the ABP self-propulsion statistics.
Apart from this non trivial aspect,
the AOUP model has the advantage of lending itself to a simpler analysis and to the possibility of determining the steady state probability distribution function (pdf) of the active particle for small activity \cite{fodor2016far, marconi2017heat}.
Since the AOUP model is formulated as an overdamped particle subject to colored noise, it can be mapped into a new Markovian system, by adding a degree of freedom for each component of the noise. This new enlarged representation allows for the study the problem by using a standard approach based on the Kramers equation for which several approximate methods
of solution are well-known \cite{Gardiner, risken}.
However, the choice of this enlarged space is not unique since the microstate of a single particle at a given instant can be identified by its position and velocity or by its position and by the value of the active force acting on it. The two descriptions are
equivalent and for both one can write the corresponding Fokker-Planck equations and the associated approximate steady state
distribution functions obtained by means of a perturbative analysis in terms of the parameter $\tau^{-1}$.
As far as only the steady state configurational properties are concerned it is possible to derive a closed, non-perturbative expression for the
distribution function by means of the so-called unified colored noise approximation (UCNA) put forward by Hanggi and Jung
\cite{hanggi1995colored,marconi2015towards}. The static properties of the UCNA have been tested with success in the case
of persistence times not too large, but very little is known about its dynamical properties.
The present study aims to fill this gap by considering the response to a small external perturbation of a self-propelled particle
driven by colored noise in the presence of a trapping potential. We shall compare both exact analytic and numerical results obtained by applying the fluctuation dissipation relation (FDR) \cite{marconi2008fluctuation} to the AOUP model for the response
to an initial displacement of the particle's position with the corresponding quantity computed within the UCNA.
As a byproduct of this study, we obtain and explain a result which contradicts the naive expectation that the positional response
function should not depend on the choice of the enlarged representation.
The paper is organized as follows. The description of the model and the theoretical results are presented in Section \ref{Theory}. In Sec.~\ref{Numerical Analysis} we report the results of some numerical simulations obtained in the case of anharmonic potentials. Finally, we draw some conclusions in Sec. \ref{Summary and Conclusions}.
\section{Models and theory}
\label{Theory}
We model the effective dynamics for the space coordinates of an assembly of non-interacting active Brownian particles \cite{marconi2015towards,maggi2015multidimensional}, as:
\begin{eqnarray}
\label{eq:dim_AOUP}
\dot{x}& =& \frac{f(x)}{\gamma}+a, \qquad f(x)=-\frac{d}{dx}\phi(x), \nonumber \\
\dot{a}&=&-\frac{a}{\tau} +\frac{\sqrt{2D}}{\tau}\eta,
\end{eqnarray}
where $x(t)$ is the position of the particle,
$\tau$ is the correlation time, $\gamma$ the drag coefficient, $\phi(x)$ the potential acting on the system
and the term $a(t)$, also called \emph{active bath}, evolves as an Ornstein-Uhlenbeck process. The stochastic force $\eta(t)$ is a white noise, i.e. a Gaussian process with zero mean and $\left<\eta(t)\eta(t') \right>=\delta(t-t')$.
The parameter $D$ is the diffusive coefficient due to the activity related
and fixes the amplitude of the Ornstein-Uhlenbeck process, via:
\begin{equation}
\left<a(t)a(t') \right> = \frac{D}{\tau} \exp{\left[-\frac{t-t'}{\tau}\right]}.
\end{equation}
In order to proceed further, we will adopt non-dimensional variables for position, velocity, and time
\begin{equation}
X=\frac{x}{l},\qquad V=\frac{v}{v_T},\qquad \bar{t}=t\frac{v_T}{l}
\end{equation}
where
$l$ is a suitable length and $v_T=\sqrt{\frac{D}{\tau} } $ is a reference velocity.
We rescale forces and potential as follows:
\begin{equation}
\qquad F(X) =f(x)\frac{l}{D\gamma},\qquad U(X)=\frac{\phi(x)}{D\gamma} , \qquad A=\frac{a}{v_T},
\end{equation}
and $\zeta =\frac{l}{\tau v_T}$ can be seen as the ratio
between the characteristic length of the problem, $l$, such as the typical length-scale of the external potential $U(x)$, and
the mean square diffusive displacement due to the active bath in a time interval $\tau$.
Rewriting Eq.\eqref{eq:dim_AOUP} in terms of these new variables we have:
\begin{eqnarray}
\label{eq:AOUPstoc}
&&\dot{X}=-\frac{U'(X)}{\zeta} + A, \\
&&\dot{A}=-\zeta A + \sqrt{2\zeta}\xi(\bar t) ,
\end{eqnarray}
where $\langle \xi(\bar t) \xi(\bar t')\rangle=\delta(\bar t- \bar t')$.
In the following, we will use the symbol $t$ for the non-dimensional time.
If the particle is confined to a region of space by a potential $U(X)$,
$\zeta^{-1}$ represents the ratio between the persistence length
and size of the potential well and the amplitude of the fluctuating force in reduced units is
$\lim_{ t \to \infty} \langle A(t) A( t)\rangle= 1$.
For the pdf ${\tilde P}(X, A, t)$ of the $(X,A)$ variables we have the following Fokker-Planck equation:
\begin{equation}
\pder{t} {\tilde P} - \pder{X}\left(\frac{U'(X)}{\zeta} - A\right) {\tilde P} =\zeta \frac{\partial}{\partial A} \left[ \frac{\partial}{\partial A} + A \right] {\tilde P}
\label{fpeactive}
\end{equation}
whose stationary solution $\tilde{P_s}(X, A)$ is in general unknown a part from simple potentials \cite{risken}.
In order to apply techniques developed for the Kramers equation it is sometimes convenient
to use instead of the $(X, A)$ variables the phase-space variables $(X, V)$ (see for instance
\cite{maggi2015multidimensional,fodor2016far,marconi2017heat}), through the following change of variables
\begin{eqnarray}
\label{eq:changevariables}
&&V \equiv \dot{X}=-\frac{U'(X)}{\zeta} + A,
\\
&&X'=X.
\end{eqnarray}
In this way we recast
the stochastic differential equation \eqref{eq:AOUPstoc} as:
\begin{eqnarray}
\label{eq:kramer}
&& \dot{X}=V\\
&&\dot{V}= - U'(X) - \zeta g(X) V+ \sqrt{2T\zeta}\eta
\end{eqnarray}
and the associated Kramers equation for the phase-space distribution $P(X, V, t)$:
\begin{equation}
\label{eq:one_kramers}
\pder{t} P + V \pder{X} P -U'(X) \pder{V} P = \zeta \pder{V} \left[\pder{V} + g(X)V \right] P ,
\end{equation}
which means that the activity can be mapped into a space dependent friction term $g(X)=1 + \frac{1}{\zeta^2}\frac{d^2}{dX^2} U(X) $ which depends on $\zeta$.
The second and third term on the left-hand side represent the streaming terms in the
evolution of the phase-space distribution, whereas the right-hand side describes the dissipative part. Again the stationary distribution $P_s(X, V)$ is unknown, in general. Because the Jacobian of the transformation $(X, A)\rightarrow(X', V)$ is unitary, we have:
\begin{equation}
P_s(X, V) = \tilde{P}_s(X, A(V, X)).
\end{equation}
We would like to stress that the $X,V$ representation is relevant because it allows us to obtain the distribution function in terms of these variables and to develop an efficient perturbative scheme in powers of the non equilibrium parameter $1/\zeta$.
\subsection{Steady state probability distributions in the extended space}
Among the few cases whose stationary solution of the Fokker-Plank equation is known, one has the harmonic potential, $U(X)
=\lambda X^2/2$. The steady state distribution
in the $(X, V)$ variables is a Gaussian, $P_s(X,V)\propto e^{-\frac{\beta}{2}\left(\lambda X^2+V^2 \right)}$ with inverse "effective temperature" $\beta=(1+\lambda/\zeta^2)$, while in the variables $(X, A)$ is the following multivariate Gaussian $\tilde{P_s}(X,A)\propto e^{-\frac{\beta}{2}\left(\lambda X^2+\left(A-\lambda X/\zeta \right)^2 \right)}$.
For a generic potential the stationary pdf in the limit $1/\zeta \ll 1$, has the following approximated
form (see \cite{fodor2016far,marconi2017heat}):
\begin{eqnarray}
\label{eq:Kdistribution}
&&P_s(X, V)\propto e^{-U(X)-\frac{V^2}{2} } \Biggr\{1- \frac{1}{2\zeta^2}\Bigl[U'(X)^2 +V^2U''(X)-3U''(X)\Bigr]
\nonumber\\&&
+ \frac{1}{6\zeta^3} U'''(X) V^3 - \frac{1}{2\zeta^3}U'''(X) V \Biggr\} +O(\frac{1}{\zeta^4}),
\end{eqnarray}
showing that positions and velocities are correlated for any finite $\zeta$.
In the $(X, A)$ variables the stationary pdf reads:
\begin{eqnarray}
\label{eq:AOUPdistribution}
&&
\tilde{P_s}(X, A)\propto e^{-U(X)-\frac{\left( A - U'/\zeta \right)^2}{2} } \Biggl\{1- \frac{1}{2\zeta^2}\Bigl[U'(X)^2 +\left(A-\frac{U'(X)}{\zeta} \right)^2U''(X)-3U''(X)\Bigr]
+\nonumber\\&&
+ \frac{1}{6\zeta^3} U'''(X) \left(A-\frac{U'(X)}{\zeta}\right)^3 - \frac{1}{2\zeta^3}U'''(X) \left(A-\frac{U'(X)}{\zeta}\right)\Biggr\} +O(\frac{1}{\zeta^4})\nonumber\\
\end{eqnarray}
Actually, there are no results in the opposite limit $\zeta \ll 1$, where the persistence time is large.
\subsection{Reduced descriptions: distribution in positional space}
\label{section:Overdampedregimes}
Since in general the analytic solutions of eqs. \eqref{fpeactive} or \eqref{eq:one_kramers}
are not known, it is common practice to resort to a reduced description involving only the coordinate $X$ for which it is possible to
develop an efficient approximation method. This is the idea behind the reduction of the Kramers equation onto the Smoluchowski
equation and it can be realized by different procedures such as multiple time-scale methods, functional integral techniques or adiabatic procedures.
The unified color noise approximation (UCNA) was developed the first time by Hanggi et Jung by using an adiabatic elimination procedure to study the behavior of particles driven by colored noise~ \cite{hanggi1995colored,h1989colored} and then recently extended \cite{marconi2015towards, maggi2015multidimensional, marconi2015velocity} for systems of active particles. In the following, we consider two types of approximations: UCNA and an overdamped limit performed
directly on the equation \eqref{eq:AOUPstoc}, with the idea of making a comparison among them. We shall study different regimes: $\zeta\gg1$ and $\zeta\ll1$. The first regime corresponds to a small departure from the equilibrium situation determined by the presence of a small $\tau$, while the second regime to the case in which the persistent time is large and is more interesting, because it shows the peculiar features of the active particles, for instance the accumulation of active particles close to confining walls \cite{maggi2015multidimensional}.
In order to gain some insight, we consider the distribution of positions in two special limits corresponding to short
persistence time $\zeta\gg 1$ and to long persistence time $\zeta \ll 1$.
\begin{enumerate}[(i)]
\item $\zeta\gg1$. Let's consider the system of Eq. \eqref{eq:AOUPstoc}; a first approximation consists in neglecting $\dot{A}$. This means that $A$ is well approximated by a white noise and we have:
\begin{equation}
\label{eq:case1_pdf}
\dot{X} =-\frac{U'(X)}{\zeta} + \sqrt{\frac{2}{\zeta}} \xi, \qquad A= \sqrt{\frac{2}{\zeta}} \xi \qquad \Longrightarrow \qquad P_1(X) \propto e^{-U(X)},
\end{equation}
where $P_1(X)$ is the configurational stationary probability distribution of an equilibrium system.
\item $\zeta \ll1$. In this case, we can neglect $\dot{X}$, therefore the variable $X$ is related to $A$, so that the evolution is given by:
\begin{equation}
\dot{A}=-\zeta A +\sqrt{2\zeta} \xi, \qquad \Longrightarrow \qquad \tilde{P}_2(A) \propto e^{-A^2/2},
\end{equation}
being $\tilde{P}_2(A)$ the stationary pdf for $A$.
We can obtain a new probability distribution, $P_2(X)$, for the variable $X$, since $dA=\frac{U''(X)}{\zeta}dX$
and we have:
\begin{equation}
\label{eq:case2_pdf}
P_2(X) \propto \frac{U''(X)}{\zeta^2}\exp{\left(-\frac{U'(X)^2}{2\zeta^2}\right)} .
\end{equation}
Let us notice that for small $\zeta$, $P_2(X)$ is a very peaked probability distribution.
\end{enumerate}
Let us, now, consider the description based on the variables $(X,V)$ given by Eq.~\eqref{eq:changevariables} and perform the
adiabatic elimination of the $V$ variable, i.e. neglecting the acceleration $\dot V$, both for $\zeta \gg 1$ and $\zeta \ll 1$.
We have the following single first order stochastic equation, the well known UCNA equation:
\begin{equation}
\dot{X}=-g(X)^{-1}\frac{U'(X)}{\zeta} + g(X)^{-1}\sqrt{\frac{2}{\zeta}} \xi .
\end{equation}
whose stationary pdf of positions reads:
\begin{equation}
\label{eq:UCNAdistribution}
P_{U}(X) \propto g(X)\exp{\left(-U(X) - \frac{1}{2\zeta^2} U'(X)^2 \right)}.
\end{equation}
Let's remark that
\begin{itemize}
\item when $\zeta \gg 1$ and $g \sim 1$ we have $P_U \sim P_1$.
\item when $\zeta \ll 1$: since $g \sim U''(X)/\zeta^2$ we have $P_U \sim P_2$ .
\end{itemize}
We note that $P_U$ and $P_1$ are the approximations of the marginal distribution of the system described by the probability distribution given by Eq.~\eqref{eq:Kdistribution}. Indeed, by calling $P_M(X)$ the marginal distribution, with respect to $V$, associated to Eq.~\eqref{eq:Kdistribution}, we have:
\begin{equation}
P_M(X)\equiv\int dV P_s(X, V) = P_U(X) + O\left(\frac{1}{\zeta^4} \right) = P_1(X) +O\left(\frac{1}{\zeta^2} \right).
\end{equation}
Therefore, we can say that when $\zeta \gg 1$, the UCNA-model is a better approximation than the model (i)
and there is no reason to use the model (i) instead of the UCNA-model. In particular in the harmonic case the marginal distribution is exactly reproduced by the UCNA.
\subsection{Linear response function}
In this subsection, we shall study the response of our system when we slightly perturb the initial position of the particle
and show that such a procedure yields different results, depending on the variables chosen
to describe the system. In order to solve this apparent paradox, we first apply the well-known general fluctuation-dissipation relations \cite{falcioni1990correlation,marconi2008fluctuation}, in both representations $(X, V)$ and $(X, A)$.
We show that notwithstanding the Jacobian of the transformation is unitary, a perturbation of the position, $X$, in the $A$ representation corresponds to a perturbation involving both the variables ($X$ and $V$) in the $V$ representation.
The response function $R$ of the AOUP model was studied by Szamel and Fodor et al. in \cite{fodor2016far,szamel2014self}. In particular Fodor et al. numerically measured the susceptibility, defined as the time integral of the response and studied the system using the $(X, V)$ coordinates, in the regime of small persistence time. In the present study, instead we directly measure the response of the system, both in the small and in the large persistence time limit. Let's call $R_A(t)$ the mean response of the system that we will compute numerically by adding a
small impulsive force $h(t)=h_0 \delta(t)$, in the first equation of the system \eqref{eq:AOUPstoc}:
\begin{eqnarray}
\label{eq:AOUPstoc_R}
&&\dot{X_h}=-\frac{U'(X_h)}{\zeta} + A_h + h(t), \\
&&\dot{A_h}=-\zeta A_h + \sqrt{2\zeta} \xi,
\end{eqnarray}
where $X_h$ and $A_h$ are the variables of the system in presence of the perturbation. In a similar way we call $R_V(t)$ the response obtained by adding the small force to equations \eqref{eq:kramer} in the $(X,V)$ representation. The study of $R_A(t)$ and $R_V(t)$ can be numerically performed by computing the following normalized quantity:
\begin{equation} R(t) = \frac{\left<\delta X(t)\right>}{\delta X(0)}=\frac{\left<X_h(t)-X(t)\right>}{X_h(0)- X(0)}.
\end{equation}
Although there are different versions of the FDR for the prediction of $R(t)$ \cite{kubo2012statistical, hanggi1982stochastic}, we focus the attention on the first version of the FDR, independently developed in \cite{A1972} (for a recent work see \cite{Chadauri2012Mod}) and \cite{falcioni1990correlation}.
According to this version of FDR~\cite{marconi2008fluctuation}, we have the response in terms of an average, which involves the stationary pdf:
\begin{equation}
\label{eq:FDR}
R_{V(A)}(t) = -\left<X(t) \left( \pder{X}\log {Prob} \right) \Biggl |_{t=0} \right>_{V(A)},
\end{equation}
where depending on the choice of the variables one sets $Prob$ equal to $P_s(X, V)$ or $\tilde P_s (X, A)$ and the average $\left<\cdot\right>$ is performed by using the corresponding stationary pdf
and the symbol $t=0$ means that the function is computed for the variables at $t=0$.
By inserting Eq.~\eqref{eq:Kdistribution} in relation \eqref{eq:FDR}, we obtain the response, $R_V(t)$ in the $(X, V)$ representation:
\begin{eqnarray}
&&
R_V(t)=-\left<X(t) \Biggl(\pder{X}\log{P_s(X, V)} \Biggr)\Biggl |_{t=0}\right>_V =
\nonumber \\&&
\Biggl<X(t)\Biggl(U'(X)+\frac{1}{\zeta^2}\left[ U''(X)U'(X)+\frac{V^2}{2}U'''(X)-\frac{3}{2}U'''(X)\right]
\nonumber \\ &&
-\frac{1}{\zeta^3}\left[\frac{V^3}{6}U''''(X)-\frac{1}{2}VU''''(X) \right] \Biggr) \Bigg|_{t=0} \Biggr>_V ,
\label{RXV}
\end{eqnarray}
up to the order $O(1/\zeta^4)$, while using Eq.~\eqref{eq:AOUPdistribution} we obtain the response, $R_A(t)$ in the $(X, A)$ representation:
\begin{eqnarray}
&&
R_A(t)=-\left<X(t)\left(\pder{X}\log{\tilde P_s(X, A)}\right)\Bigg|_{t=0}\right>_A =\nonumber \\
&&
\Biggl<X(t) \Biggl(U'(X) - \frac{1}{\zeta} A U''(X)+\frac{1}{\zeta^2} \left(2 U'(X)U''(X)+\frac{A^2}{2}U'''(X) -\frac{3}{2}U'''(X) \right) +
\nonumber\\ &&
\frac{1}{\zeta^3}\left(-A \Bigl[U''(X)^2+U'(X)U'''(X) \Bigr] -\frac{A^3}{6}U''''(X)+\frac{A}{2}U''''(X) \right) \Biggr)\Bigg|_{t=0}\Biggr>_A,
\label{RXA}
\end{eqnarray}
where $\left< \cdot \right>_A$ denotes the average with respect to $\tilde{P}(X, A)$. By expressing the response $R_A$ in the variables $(X, V)$ we obtain:
\begin{equation}
\label{eq:response_AOUPXV}
R_A(t)=R_V(t)-\frac{1}{\zeta}\Biggl<X(t) V(0) U''(X(0))\Biggr>_V -\frac{1}{\zeta^3}\left<X(t) V(0) U''(X(0))^2\right>_V \, ,
\end{equation}
showing that $R_A$ and $R_V$ differ by terms of order $1/\zeta$, which vanish in the limit $1/\zeta \ll 1$.
Such a result seems somehow counterintuitive:
how is it possible that the response of the system to an initial perturbation in the $X$ variable depends on the choice of the coordinates that we use? To explain that,
let's introduce the pdf of the perturbed system $\tilde{P}_s'(X, A)=\tilde{P}_s(X-\delta X_0, A)$ and
$P_s'(X, V)=P_s(X-\delta X_0, V)$, and let's call $\tilde{W}\left((X_0, A_0)\rightarrow(X, A)\right)$ and $W\left((X_0, V_0)\rightarrow(X, V)\right)$ the transition probabilities from the state at time zero to the one at time $t$ in the coordinates $(X, A)$ and $(X, V)$, respectively. Under the usual hypothesis for the probability distribution it is easy \cite{marconi2008fluctuation} to show that the response of the position at time $t$, $R_A(t)$, in the variables $(X, A)$, is:
\begin{equation}
<\delta X(t) >_A=\int X \left(\tilde{P}_s'(X_0, A_0) - \tilde{P}_s(X_0, A_0)\right) \tilde{W}\left( (X_0, A_0)\rightarrow (X, A)\right) dX_0 dA_0 dX dA. \nonumber\\
\label{deltax}
\end{equation}
Since the Jacobian of the transformation is unitary , we can switch from the variables $(X, A)$ to $(X, V)$:
\begin{eqnarray}
&&<\delta X(t) >_A= \nonumber\\
&&\int X \Bigl[P_s'(X_0, V_0(X_0 ,A_0)) - P_s(X_0, V_0(X_0, A_0))\Bigr] W\left( (X_0, V_0)\to (X, V)\right) dX_0 dV_0 dX dV.
\end{eqnarray}
For a small increment $\delta X_0$ we have:
\begin{eqnarray}
&&
P_s'(X, V(X ,A)) - P_s(X, V(X, A)) = -\delta X_0\left[ \frac{\partial}{\partial X}P_s(X,V(X ,A))+\frac{\partial V(X,A)}{\partial X}\frac{\partial P_s}{\partial V} \right]
\nonumber\\&&
=-\delta X_0 \frac{d}{dX}P_s(X, V(X, A)).
\end{eqnarray}
This version of the FDR involves a total derivative, which acts also on the velocity. This means that
the response $R_A(t)=\frac{<\delta X(t) >_A}{\delta X_0}$ is given by:
\begin{eqnarray}
\label{eq:GeneralFDR}
&&
R_A(t)=-\left<X(t)\left(\frac{d}{dX}\log{P_s(X, V(A, X)})\right)\Bigg|_{t=0}\right>_V
\nonumber\\&&
\neq-\left<X(t)\left(\pder{X}\log{P_s(X, V(A, X)})\right)\Bigg|_{t=0}\right>_V=R_V(t),
\end{eqnarray}
where the averages are performed by using $P_s(X, V)$.
In other words we can say that, since $A$ depends on $X$ and $V$, the perturbation $(X, V) \rightarrow (X+\delta X_0, V)$ is not equivalent to the perturbation $(X, A) \rightarrow (X+\delta X_0, A)$.
\section{Results}
\label{Numerical Analysis}
In the following, we shall present some results illustrating the predictions of the theory
in some simple cases. We have performed the simulations for three different potentials $U(X)$: (a) harmonic potential $U(X) = \lambda X^2/2$, (b) quartic potential $U(X) = \lambda X^4/4$, (c) double well potential $U(X) =\lambda ( X^4/2-X^2/2)$.
The numerical computations of $R(t)$, both from data and FDR Eq.\eqref{eq:FDR}, were performed using the Euler-Maruyama method \cite{ToralColen_book}, neglecting order $(\Delta t)^{5/2}$.
\subsection{Response in the limit $\zeta \ll 1$}
In the case (a) the probability distribution of the system can be computed exactly and therefore we have an exact expression for the response: $R_A(t)\sim e^{-t \lambda/\zeta}$.
In the cases (b) and (c)
we know the probability distribution as a series in powers of $1/\zeta\ll 1$ so that we can obtain the FDR only perturbatively. Therefore, the numerical approach is necessary when the limit $\zeta\gg 1$ doesn't hold. In Fig.~\ref{fig:threepotentials} we show $R_A(t)$ for the three different potentials and different values of $\zeta$. Let us first discuss the case (a) and (b):
\begin{figure}[!h]
\centering
\subfloat[]{\includegraphics[width=0.33\linewidth,keepaspectratio]{H_responses_linear.png}\label{fig:1KvsA}}
\subfloat[]{\includegraphics[width=0.33\linewidth,keepaspectratio]{Q_responses_linear.png}\label{fig:3KvsA}}
\subfloat[]{\includegraphics[width=0.33\linewidth,keepaspectratio]{responses_double.png}\label{fig:5KvsA}}
\caption{Responses $R_A(t)$ for the AOUP model computed, via numerical simulation, for different values of $\zeta$
by setting $\lambda=1$.
In figure (a) are plotted the responses for the harmonic potential $\lambda X^2/2$. The plots (b) and (c) correspond to a quartic potential $\lambda X^4/4$ and a double well potential $\lambda (X^4/4 - X^2/2)$, respectively.}
\label{fig:threepotentials}
\end{figure}
when $\zeta$ is large, we are near the delta correlated noise, closed to the equilibrium situation. Therefore the shape of the potential does not change the form of the response, which decays roughly as an exponential.
When $\zeta\sim 1$ or $\zeta\ll 1$ the results relative to the two potentials display large differences
increasing as $\zeta$ decreases.
In particular, when the attractive force becomes stronger, the response becomes slower, as we can see in Fig.~\ref{fig:threepotentials}.
This is a consequence of the departure of the system from thermodynamic equilibrium: indeed the detailed balance holds only in the harmonic case \cite{marconi2017heat} when the drag coefficient in the $(X, V)$ variables is constant. Otherwise, $g(X)$ is not constant, and decreasing $\zeta$ the system goes far from equilibrium.
Indeed, where $\zeta$ is small, in the harmonic case the shape obtained is the one predicted by the theoretical computations $\sim e^{-\lambda t/\zeta}$. In the other cases we can distinguish between two regimes: up to $t\sim 1$ (in
dimensional units this corresponds to $t\sim l/v_T$) there is \emph{fast} relaxation, while for $t\gg1$ there is a relaxation with an effective characteristic time $t_{slow} \gg \zeta$.
The presence of these two regimes is a non-equilibrium effect and is more evident when the activity is large ($\zeta$ small).
The presence of two times scales, when the system is far from equilibrium, is a clear consequence of the accumulation of particles near the confining walls \cite{maggi2015multidimensional}. This means that, even if the potential applied has a single well, the particle in the steady state experiences an effective double well potential.
Such an observation is confirmed by the shape of the stationary pdf $P_U(X)$ in the UCNA-approximation.
Phenomenologically, the drift term takes different values depending on the position of the particle: when $X$ is near the minimum of the potential the effective drift force is proportional to $\sim \zeta V\ll V$, being $U''(X)/\zeta \sim 0$, and the particle moves just because of the deterministic force. For $X$ far from the minimum, the drift force is proportional to $\sim V U''(X)/\zeta \gg V$, which means that the particle experiences a big Stokes force and moves very slowly. For $X$ far from the minimum, the deterministic force is very big and steadily pushes the particle towards the minimum, preventing the particle from going too far. The balance between these two effects leads to a situation where the most probable value of the position does not coincide with the minima of the potential. This fact explains why the decay of the response function displays two different time-regimes, even in the presence of a single well potential $U(X)\propto X^{2n}$ with $n>1$ .
Let us remark that this mechanism acts only when the detailed balance does not hold, as in the case where the curvature
$U''(X)$ is not constant \cite{marconi2017heat}.
Finally, we consider the the double well potential (case (c)). In order to gain some insight, let us consider the situation where the noise is delta correlated: the response function displays two different decay behaviors (roughly exponential), in the first stage the typical decay time is associated with the relaxation in one of the two wells and is determined by
the curvature of the potential, $U^{''}(X_{min})$. For longer times the jumps of the particle between the two minima are relevant and the mean first passage time is determined by Kramers' formula \cite{Gardiner}. If the persistence time, $\tau$, is not very small is not easy to extend the above argument, however, in Fig.~\ref{fig:threepotentials} (c) which displays the behavior of the response function versus $t$, it is quite evident the presence of two different characteristic time scales.
When the persistence time becomes larger the second relaxation becomes slower as clearly indicated by
the plot of Fig.~\ref{fig:threepotentials} (c), as if the effective barrier becomes higher.
\subsection{The UCNA response function}
It is known that the UCNA model well describes
all the stationary properties of the system both for $1/\zeta \ll1$ and $\zeta \ll1$.
This state of affairs is no longer true for the time-dependent dynamical properties such as the response to a small perturbation.
By using the FDR for a system under the action of a generic potential $U(X)$, we easily obtain the following expressions for the responses in the three cases, denoted by a subscript.
\begin{enumerate}
\item if $\zeta \gg 1$ from eq. \eqref{RXV} we have:
\begin{equation}
R_1 = \left<X(t) U'(X(0)) \right>_V.
\end{equation}
\item while for $\zeta \ll 1$ the response is
\begin{equation}
R_2(t) =\frac{1}{\zeta^2}\left<X(t) U'(X(0))U''(X(0)) \right>_V - \left<X(t) \frac{U'''(X(0))}{U''(X(0))}\right>_V.
\end{equation}
\item
and within the UCNA we have:
\begin{equation}
R_U(t) = -\left<X(t)\left(\pder{X}\log{P_U(X)}\right)\Bigg|_{X=X(0)} \right>_U,
\label{eq:RU}
\end{equation}
and explicitly:
\begin{equation}
R_U(t) = \frac{1}{\zeta^2}\left<X(t) U'(X(0))U''(X(0)) \right>_U + \left<X(t) U'(X(0)) \right>_U - \left<X(t) \frac{U'''(X(0))}{\zeta^2+U''(X(0))}\right>_U.
\end{equation}
\end{enumerate}
where the subscript $U$ means that the average is with respect to the UCNA steady state distribution, $P_U(X)$
given by \eqref{eq:UCNAdistribution}.
\subsection{Response function in the presence of a quadratic potential}
In the harmonic case, $U(X)=\lambda X^2/2$, we can apply the FDR for all values of $\zeta$ without approximations
and obtain from Eq.~\eqref{RXV}:
\begin{equation}
R_V(t) = \beta \lambda \left<X(t) X(0) \right>_V
\label{eq:xx}
\end{equation}
and from Eq.~\eqref{eq:response_AOUPXV}:
\begin{equation}
R_A(t)=\beta \lambda \Bigl(\left<X(t) X(0) \right>_V- \frac{1}{\zeta}\left<X(t)V(0) \right>_V \Bigr) = R_V(t) -\beta \frac{\lambda}{\zeta}\left<X(t)V(0) \right>_V .
\label{eq:xv}
\end{equation}
In general, the two responses are not the same, except in the limit $\zeta \rightarrow \infty$ which corresponds to the $\delta$-correlated case.
As shown in the Appendix the correlation functions appearing in r.h.s. of Eqs. \eqref{eq:xx} and \eqref{eq:xv} are given by:
\begin{equation}
\label{eq:XcorrelationF}
\beta\left<X(t)X(0)\right>_V = \frac{1}{\zeta -\lambda/\zeta} \left[\frac{\zeta}{\lambda} e^{-\lambda t/\zeta} - \frac{1}{\zeta}e^{-t\zeta} \right],\\
\end{equation}
and
\begin{equation}
\label{eq:VcorrelationF}
\beta\left<X(t)V(0)\right>_V = \frac{1}{\zeta -\lambda/\zeta} \left[ e^{-\lambda t/\zeta} - e^{-t\zeta} \right].
\end{equation}
Finally, the response functions read:
\begin{equation}
\label{eq:V_response}
R_V(t) = \frac{\lambda}{\zeta-\frac{\lambda}{\zeta}} \left( \frac{\zeta}{\lambda} e^{-\lambda t/\zeta} - \frac{1}{\zeta}e^{-t \zeta} \right)\\
\end{equation}
and
\begin{equation}
\label{eq:exact_response}
R_A(t) = e^{-\lambda t/\zeta},\\
\end{equation}
Perhaps contrary to intuition, the two responses are different for $\zeta$ not too large. For large $\zeta$ the two responses are very close. This is an effect of the memory: indeed, small $\zeta$ means big correlation time, being $\zeta\sim 1/\tau$.
\begin{figure}[!h]
\centering
\subfloat[]{\includegraphics[width=0.35\linewidth,keepaspectratio]{1KA.png}\label{fig:1KA}}
\subfloat[]{\includegraphics[width=0.35\linewidth,keepaspectratio]{3KA.png}\label{fig:3KA}}
\caption{
Response function $R_A(t)$ computed via numerical simulation for a harmonic potential $\lambda X^2/2$ with $\lambda=1$ (black line). The red diamonds (C) represent the sum of the correlation functions given by the Eq.~\eqref{eq:xv}, i.e.
$R_A(t)=\beta \lambda \Bigl(\left<X(t) X(0) \right>_V- \frac{1}{\zeta}\left<X(t)V(0) \right>_V \Bigr)$, a test employed
in order to verify FDR.
The green triangles represent the correlation $\beta\left<X(t)V(0)\right>_V$ (Eq.~\eqref{eq:VcorrelationF}) and
the violet inverse triangles represent the correlation $\beta\left<X(t)X(0)\right>_V$ (Eq.~\eqref{eq:XcorrelationF}) and .
Panel (a) corresponds to $\zeta=1$ and panel (b) to $\zeta=3$.}
\label{fig:KR}
\end{figure}
Consider now the response function as predicted by the UCNA theory in the harmonic case.
By using the FDR given by Eq. \eqref{eq:RU} together with the UCNA stationary probability distribution
\eqref{eq:UCNAdistribution} and the correlation function
$$
<X(t)X(0)>_U=\frac{1}{\beta\lambda} \, \exp{\left(-\frac{\lambda t}{\zeta + \lambda/\zeta} \right)} ,
$$ the response function turns out to be:
\begin{equation}
R_U(t) = \exp{\left(-\frac{\lambda t}{\zeta + \lambda/\zeta} \right)}.
\end{equation}
Let's observe that this response is invariant for $\zeta \rightarrow \lambda/\zeta$. Then:
\begin{itemize}
\item $\zeta\gg1\quad \Longrightarrow \quad R_U(t) \sim \exp{\left(-\frac{\lambda t}{\zeta}\right)}$, which is consistent with the response $R_A(t)$, given by Eq~\eqref{eq:exact_response}.
\item $\zeta\ll1\quad \Longrightarrow \quad R_U(t) \sim \exp{\left(-t\,\zeta\right)}$, which is not correct.
Smaller $\zeta$ means a slower response, in disagreement with the result for $R_A(t)$, given by the Eq.~\eqref{eq:exact_response}.
\end{itemize}
\subsection{Response function with varying $\zeta$}
Let us show the responses, numerically computed, for harmonic and quartic potentials, $U(X)=\lambda X^2/2$ and $U(X)=\lambda X^4/4$, respectively.
In the harmonic case, the simulations are only intended as a check of the numerical codes. In the quartic case, we have only a perturbative result in power of $1/\zeta \ll1$ for the probability distribution function (see eq. \eqref{eq:Kdistribution}). In general, it is difficult to predict the
response in the small-$\zeta$ regime and we need a numerical study.
In fig. \ref{fig:10e05} we show a comparison between the response of the AOUP-model and the UCNA. For $\zeta \gg 1$ the UCNA is a good approximation of the AOUP both in the case of a harmonic potential (Fig. \ref{fig:AOUPUCNA_10e05}) and of a quartic potential (Fig. \ref{fig:four_AOUPUCNA_05e10}). When $\zeta$ becomes smaller, the situation is completely different: in the harmonic case, as predicted by the theoretical computation, the response $R_A(t)$ becomes slower, according to the invariance $\zeta \rightarrow \lambda/\zeta$. Even in this simple case, the UCNA is not able to reproduce the response of the AOUP system for $\zeta \ll 1$. The scenario is similar in the case of the quartic potential $U(X) = \lambda X^4/4$.
\begin{figure}[!h]
\centering
\subfloat[]{\includegraphics[width=0.45\linewidth,keepaspectratio]{AOUPUCNA_10e05.png}\label{fig:AOUPUCNA_10e05}}
\subfloat[]{\includegraphics[width=0.45\linewidth,keepaspectratio]{four_AOUPUCNA_05e10.png}\label{fig:four_AOUPUCNA_05e10}}
\caption{Responses functions computed via numerical simulations from the AOUP ($R_A(t)$) model and the UCNA-model
($R_U(t)$), for different values of $\zeta$: $\zeta=10, 0.5$ and $\lambda=1$. The graphs are obtained for systems under the action of a harmonic potential $U=\lambda X^2/2$ ( panel (a)) and a potential $U=\lambda X^4/4$ (panel (b)).}\label{fig:10e05}
\end{figure}
\begin{figure}[!h]
\centering
\subfloat[]{\includegraphics[width=0.33\linewidth,keepaspectratio]{two_02AOUPkramerUCNA.png}\label{fig:two_02AOUPkramerUCNA.png}}
\subfloat[]{\includegraphics[width=0.33\linewidth,keepaspectratio]{four02_AOUPkramerUCNA.png}\label{fig:four_02AOUPkramerUCNA.png}}
\subfloat[]{\includegraphics[width=0.33\linewidth,keepaspectratio]{05kramerAOUP.png}\label{fig:05kramerAOUP.png}}
\caption{Response functions $R_A(t)$ and $R_V(t)$ computed via numerical simulations from the AOUP model
and response $R_U(t)$ of the UCNA in the case of small $\zeta$: $\zeta=0.2$ and $\lambda=1$. The graphs (a) and (b) are obtained for systems under the action of a harmonic potential $U=\lambda X^2/2$ and a quartic one $U=\lambda X^4/4$, respectively. The graph (c) is obtained for the double well potential $U=\lambda (X^4/4- X^2/2)$, for $\zeta=0.5$. }\label{fig:02AOUPkramerUCNA}
\end{figure}
Moreover, we observe that the response $R_U(t)$ computed within the UCNA can be seen only as an approximation of the response $R_V(t)$ computed from Eq.\eqref{RXV}.
In Fig. \ref{fig:02AOUPkramerUCNA} we show a comparison between the response functions $R_V$, $R_A$ and $R_U$.
We have that the responses $R_V(t)$ and $R_A(t)$ display marked differences when $\zeta/\lambda \ll 1$ and $R_A(t)$ decays much faster.
The explanation in the harmonic case comes from the FDR Eq.\eqref{eq:xv}: Indeed the correlation between $X$ and $V$ plays an important role only for $\zeta$ small enough, giving a non-vanishing negative contribution. This means that the coupling between the $X$ and $V$ in the Eq.\eqref{eq:xv} is responsible for the faster decay of $R_A(t)$ with respect to $R_V(t)$, as is shown in Fig.\ref{fig:KR}.
On the other hand, $R_U(t)$ in the harmonic case is very close to $R_V(t)$, but in the initial stage
the UCNA response is non-monotonic and only in a later stage behaves very closely to $R_V(t)$.
In fact, $R_U(t)$ is only an approximation of $R_V(t)$ and fails in the limit $\zeta \ll 1$. This has been directly verified both for the harmonic potential (Fig. \ref{fig:two_02AOUPkramerUCNA.png}) and the quartic potential (Fig. \ref{fig:four_02AOUPkramerUCNA.png}).
In Fig.\ref{fig:02AOUPkramerUCNA} we compare $R_V$ and $R_A$ in the case of a double well potential when $\zeta$ is small and so the system is far from equilibrium. As expected by the previous cases, $R_V(t)$ is slower than $R_A(t)$. Both responses show a first exponential relaxation for $t\sim 1$ and a relaxation slower than an exponential for a much longer time.
\section{Summary and Conclusions}
\label{Summary and Conclusions}
In this paper, we studied the response of a one-dimensional system of non-interacting AOUP under the action of an external potential. We have
shown that, at variance with the equilibrium case which applies to passive particles, the standard formula connecting the response function after an initial perturbation in the particle position, $X$, to the partial derivative of the stationary
phase-space distribution function, $P(X,V)$, with respect to $X$ has to be modified in the case of active particles.
Such a modification is necessary due to the dependence of the velocity of the particle, $V$, on the position $X$, a distinguishing feature of the
active dynamics.
The relevance of the derived formula is important when the persistence time $\tau$ is large.
In order to validate our claims, we studied the analytically solvable case of a quadratic potential and by numerical methods the case of non quadratic potentials
and compared the response of the AOUP system with the response in the overdamped regime corresponding to the UCNA both for small and large persistence time.
This analysis shows that although the stationary properties are well approximated by the UCNA in both cases, this is not true regarding the dynamical properties. In particular, in the case of the response, the UCNA is a good approximation only when the persistence time is small.
Finally, the present study has shown that when the persistence time is large enough, even in the case of a single well $U(X)\propto X^{2n}$ with $n>1$, the response function relaxation is characterized by two time scales. This result is a clear manifestation of the
non equilibrium nature of the system and appears only when the detailed balance does not hold.
\ack
We thank A. Puglisi for useful discussions and the anonymous referees for constructive criticisms.
\section{Appendix: Exact computation of correlations and response functions for the harmonic potential}\label{Appendix:HarmonicOscillator}
In the case of the harmonic potential the exact stationary probability distribution, $P_s(X,V)$, is well known, being a Gaussian
with respect to both variables so that we can compute the time-dependent correlations and the responses $R_{A(V)}$
using the FDR.
Indeed, by first multiplying by $X(0)$ the system of equations \eqref{eq:AOUPstoc}
and then taking their average with respect to the steady distribution $P_s(X,V)$ , we obtain a system of ordinary differential equation for the correlation functions.
Let's start from the evolution equations for the averages
\begin{eqnarray}
&&\frac{d}{dt}\left<X(t)X(0)\right>_V = \left<A(t)X(0)\right>_V - \frac{\lambda}{\zeta}\left<X(t)X(0)\right>_V, \\
&&\frac{d}{dt}\left<A(t)X(0)\right>_V =-\zeta\left<A(t) X(0) \right>_V .
\end{eqnarray}
We can solve
\begin{equation}
\left<A(t)X(0)\right>_V = B e^{-\zeta t},
\end{equation}
where $B$ is a constant to be fixed by the initial conditions. By substituting we get:
\begin{equation}
\frac{d}{dt}\left<X(t)X(0)\right>_V = - \frac{\lambda}{\zeta}\left<X(t)X(0)\right>_V + B e^{-\zeta t},
\end{equation}
whose solution is:
\begin{equation}
\left<X(t)X(0)\right>_V = C e^{-\lambda t/\zeta} - \frac{B}{\zeta-\frac{\lambda}{\zeta}}e^{-\zeta t},
\end{equation}
where $C$ is a second costant to be determined. Let us choose the following steady state initial conditions:
\begin{eqnarray}
&&\lambda \left<X(0)X(0)\right>_V= \frac{1}{\beta },\\
&&\left<V(0)X(0)\right>_V = \left(\left<A(0)X(0)\right>_V - \frac{\lambda}{\zeta}\left<X(0)X(0)\right>_V \right) = 0,
\end{eqnarray}
which mean that initially the "potential" energy $\lambda X^2/2$ obeys an equipartition principle and
the velocity is not correlated with the position. In this way we can easily determine the constants: $B=1/(\zeta \beta)$ and $\beta C=(\zeta/\lambda)/(\zeta-\lambda/\zeta)$. Finally, we have:
\begin{equation}
\beta\left<X(t)X(0)\right>_V = \frac{1}{\zeta -\lambda/\zeta} \left[\frac{\zeta}{\lambda} e^{-\lambda t/\zeta} - \frac{1}{\zeta}e^{-\zeta t} \right]=\frac{1}{\lambda} R_V(t).
\end{equation}
We can, now, easily compute $\left<V(t)X(0)\right>$:
\begin{equation}
\beta\left<V(t)X(0)\right>_V = \beta\frac{d}{dt}\left<X(t)X(0)\right>_V = - \frac{1}{\zeta -\lambda/\zeta} \left[ e^{-\lambda t/\zeta} - e^{-\zeta t} \right].
\end{equation}
By using the reversibility condition $\left<V(t)X(0)\right>_V=-\left<X(t)V(0)\right>_V$, the response of the AOUP system reads:
\begin{equation}
R_A(t)= \beta\lambda \left<X(t)X(0)\right>_V - \beta\frac{\lambda}{\zeta}\left<X(t)V(0)\right>_V = e^{-\lambda t/ \zeta},
\end{equation}
which is our exact result.
By the same methods we can compute the correlation functions $\left<A(t)X(0)\right>_V$ and $\left<X(t)A(0)\right>_V$.
Since the harmonic oscillator driven by colored noise obeys the detailed balance condition, if the variable $A(t)$ had a
a well defined parity under time-reversal one would obtain the relation $$\left<A(t)X(0)\right>_V = \pm \left<X(t)A(0)\right>_V .$$
Indeed, this is not the case and as a matter of fact the result is:
\begin{eqnarray}
&&\left<A(t)X(0)\right>_V=\left<V(t)X(0)\right>_V +\frac{\lambda}{\zeta}\left<X(t)X(0)\right>_V=\frac{1}{\beta}\frac{e^{-t\zeta}}{\zeta}, \nonumber\\
&&\left<X(t)A(0)\right>_V
=\left<X(t)V(0)\right>_V +\frac{\lambda}{\zeta}\left<X(t)X(0)\right>_V
=\frac{1}{\beta} \frac{1}{\zeta-\lambda/\zeta} \left( 2\,e^{-\lambda t/\zeta } - (1+\lambda/\zeta^2) e^{-\zeta t}\right). \nonumber
\end{eqnarray}
\section*{References}
|
\section{Introduction}
The discovery of superconformal theories (SCFTs) in six and five dimensions has been one of the most surprising results emerging from string theory in the past few decades. There are two types of 6d SCFTs, both of which are classified in terms of singular geometries: $\mathcal N = (2,0)$ theories \cite{Witten:1995zh} and $\mathcal N =(1,0)$ theories \cite{Heckman:2013pva,Heckman:2015bfa,Bhardwaj:2015xxa}. Given the surprising effectiveness of geometry in describing 6d SCFTs, a natural next step is to attempt to classify 5d SCFTs in terms of singular geometries.
In some ways, 5d SCFTs are more rigid as there is only a single type of 5d SCFT corresponding to the 5d ${\cal N}=1$ (i.e. eight supercharges) superconformal algebra. Many examples of 5d SCFTs have been realized in string theory using brane probes \cite{Seiberg:1996bd}, M-theory on local Calabi-Yau 3-folds \cite{Morrison:1996xf,Douglas:1996xp,Intriligator:1997pq}, and type IIB $(p,q)$ 5-brane webs \cite{Aharony:1997bh,Aharony:1997ju,Leung:1997tw,Bergman:2014kza}.
The classification of 6d $\mathcal N = (1,0)$ theories led to a picture involving generalized `quiver-like' theories whose structures could by and large be anticipated from field theoretic reasoning.
There are of course exceptions to this idea and explicit geometric constructions in F-theory clarified which possible exceptions arise that evade field theoretic analysis \cite{Heckman:2015bfa,Heckman:2013pva}. Similarly, in the 5d case, one might expect field theoretic reasoning to be a powerful, albeit incomplete guide. Indeed, as spearheaded in \cite{Intriligator:1997pq} it has been clear for a long time that field theoretic tools combined with the constraints of supersymmetry provide an unexpectedly powerful method for deducing the existence of interacting UV fixed points. More recently it was found in \cite{Jefferson:2017ahm} that relaxing some of the assumptions in \cite{Intriligator:1997pq} can resolve the conflict between the gauge theoretic classification described in \cite{Intriligator:1997pq} with low energy descriptions of some known stringy constructions, leading to a set of necessary (as opposed to sufficient) conditions for a 5d gauge theory to have a UV fixed point. However, it is unclear whether or not there are additional conditions needed to guarantee the existence of gauge theories
as consistent 5d SCFTs. Moreover, there are known cases in which a 5d SCFT is not a gauge theory (for example, M-theory on a local $\mathbb P^2$ embedded in a Calabi-Yau 3-fold)\footnote{Despite the fact that these cases do not admit a Lagrangian description, they can nevertheless be obtained from a gauge theory by passing through phases where some non-perturbative degrees of freedom become massless.}.
A reasonable follow-up to the field theoretic approach, then, is to try to check if the necessary gauge theoretic consistency conditions described in \cite{Jefferson:2017ahm} are in fact also sufficient, by using other string constructions to engineer the same theories. The main aim of this paper is to use geometric constructions of 5d SCFTs, realized as M-theory compactified on local Calabi-Yau (CY) 3-fold (and cross checked with dual constructions involving ($p,q$) 5-brane webs), to devise a classification scheme for 5d SCFTs. As a byproduct of our efforts, we are led to either validate or exclude various candidate 5d SCFTs predicted by the perturbative gauge theoretic analysis.
The basic mathematical setup leading to 5d SCFTs from M-theory on CY 3-folds involves studying how all compact 4-cycles (compact complex surfaces) inside a non-compact 3-fold can be shrunk to a point at a finite distance in moduli space; we call CY 3-folds engineering 5d SCFTs in this manner `shrinkable' 3-folds. This geometric picture can be schematically represented by a graph whose nodes are 4-cycles (surfaces) and whose edges denote the resulting intersecting 2-cycles (curves).
We note that a systematic study of the consistency conditions needed to construct such geometries has not been undertaken in the mathematics literature. Starting from a collapsed set of 4-cycles, the condition that one can resolve the singularities and thereby bring the 4-cycles to finite volume restricts the admissible types of K\"ahler surfaces (i.e.\ the nodes of the graph). We call the number of nodes of such a graph the \emph{rank} of the 5d SCFT. In particular, we show that the nodes of the graph must be rational or ruled surfaces (possibly blown up at a positive number of points)\footnote{Rational and ruled surfaces are equivalent to (respectively) $\mathbb P^2$ and ruled surfaces over genus $g$ curves (which we argue can be restricted to $g=0$)---see Section~\ref{sec:gtrans} for additional details.} in the rank 2 case, and further conjecture this to be true for arbitrary rank.
The Calabi-Yau condition and the requirement of positive volumes place further restrictions on the allowed intersections of the surfaces (i.e.\ the edges of the graph; see Figure \ref{fig:graph}). We thus devise a set of necessary critieria which must be satisfied for a 3-fold to engineer a 5d SCFT and conjecture that these criteria are sufficient to guarantee the existence of a 5d SCFT; this conjecture is supported by various cross checks using ($p,q$) 5-brane webs. Furthermore, we conjecture that all 5d SCFTs can be realized in M-theory on CY 3-folds satisfying these criteria. Similar to the 6d case, where F-theory compactified on elliptic 3-folds was used to classify $\mathcal N = (1,0)$ theories and it was subsequently found that for a few exotic cases frozen singularities are necessary to realize $\text{O7}^+$ planes in F-theory \cite{Tachikawa:2015wka,Bhardwaj-progress}, we find that in the M-theory case it is also necessary to include frozen singularities to obtain a complete classification of 5d SCFTs.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node[draw,circle] (a) at (0,0) {$S_5$};
\node[draw,circle] (b) at (2,0) {$S_3$};
\node[draw,circle] (e) at (2,2) {$S_4$};
\node[draw,circle] (f) at (2,-2) {$S_2$};
\node[] (c) at (4,0) {$\cdots$};
\node[draw,circle] (g) at (4,-2) {$S_1$};
\node[draw,circle] (d) at (6,0) {$S_r$};
\draw (b) --(c) -- (d);
\draw (a) -- (b);
\draw (b) -- (e);
\draw (e) -- (a);
\draw (b)--(f);
\draw (f) -- (g) -- (d);
\end{tikzpicture}
\end{center}
\caption{Graphical representation of a rank $r$ K\"ahler surface $S = \cup S_i \subset X$ embedded in local Calabi-Yau 3-fold $X$. The nodes of the graph correspond to 4-cycles $S_i$, while the edges $C_{i,i+1} = S_i \cap S_{i+1}$ correspond to 2-cycles along which the nodes intersect.}
\label{fig:graph}
\end{figure}
A complete classification of such CY 3-folds appears to be a rather daunting task. For example, it is unknown whether or not the list of possible 5d SCFTs is finite for a given rank. Luckily, it turns out that the rank 2 case is finite, permitting an exhaustive classification of physically distinct SCFTs.
By classifying rank 2 SCFTs in terms of Calabi-Yau geometry, we learn that all rank 2 gauge theories predicted in \cite{Jefferson:2017ahm}, except for one family, are realized.\footnote{We conjecture that all SCFTs admit at least one Coulomb branch parameter at the CFT point. The missing family which is represented by $SU(3)$ at Chern-Simons level $k=8$ has no Coulomb branch parameter at the would-be CFT point
and that is why we rule it out. This family would have led to a putative CFT which allows a Coulomb branch deformation only after a mass deformation (i.e. turning on $1/g^2$).} Additionally, we are also able to pinpoint the non-perturbative physics missing in the gauge theoretic approach of \cite{Jefferson:2017ahm} responsible for excluding this family of SCFTs. Furthermore, the geometric approach allows us to identify additional non-Lagrangian SCFTs whose existence motivates the existence of dual ($p,q$) 5-brane web configurations.
Given the significant practical challenges presented by this classification program, it is natural to ask if the insight we have gained from the rank 2 case can be used to streamline the classification of higher rank cases. Indeed, a careful examination of the list of rank 2 theories reveals a beautifully simple picture: rank 2 SCFTs in 5d can be organized into four distinct families, related and interconnected by RG flows triggered by mass deformations---see Figure \ref{tree}. Each family of 5d SCFTs has a parent 6d SCFT, where the parent 6d SCFT is related to a 5d descendant by circle compactification, up to a choice of automorphism twist (see \cite{Apruzzi:2017iqe} for work on classifying such automorphism twists, and see \cite{DelZotto:2015isa} for a discussion of additional discrete data characterizing circle compactifications of 6d SCFTs.) Thus the rank 2 classification could have been anticipated entirely from the 6d perspective! This result echoes a well-known property of rank 1 SCFTs: rank 1 5d SCFTs belong to a single family which descends from the 6d E-string theory via circle compactification.
We thus conjecture that {\it all 5d SCFTs arise from 6d SCFTs compactified on a circle, possibly up to an automorphism twist}.
More precisely, we anticipate that all 5d SCFTs can be organized into distinct families, each of which arises from a 6d theory. For a fixed rank in 5d, the possible 6d SCFT parents are rather limited. For example (ignoring the possible automorphism twist), the 6d SCFTs leading to rank $r$ 5d SCFTs will have $r-k$ dimensional tensor branches with rank $k$ gauge algebra. This suggests a practical method to classify 5d SCFT families starting with the 6d classification: compactifying a 6d SCFT on a circle produces a 5d theory with a Kaluza Klein (KK) tower of states. We call such theories `5d KK theories'; these theories are in some sense analogous to 6d little string theories. To obtain non-trivial 5d SCFTs from 5d KK theories we need to turn on holonomies suitably tuned to trigger an RG flow to a nontrivial 5d SCFT in the infrared. Aspects of the phase structure of 5d theories arising from circle compactifications of 6d SCFTs were analyzed in \cite{DelZotto:2017pti}.
The organization of this paper is as follows.
In Section \ref{sec:review} we discuss the preliminaries of 5d SCFTs, their effective gauge theory descriptions on the Coulomb branch, and their realizations in M-theory.
In Section \ref{sec:algorithm} we discuss the mathematics of shrinkable 3-folds and explain the basic approach of our geometric classification program.
In Section \ref{sec:classification} we repeat the classification of rank 1 5d SCFTs and extend the same methods to the rank 2 case. We also discuss the connection to 6d $\mathcal N = (1,0)$ SCFTs. Some mathematical results essential for the rank 2 classification are collected in the appendices: Appendix \ref{app:AG} contains an explicit description of the Mori cones of blowups of Hirzebruch surfaces; Appendix \ref{app:bound} contains some numerical bounds constraining rank 2 shrinkable 3-folds; finally, Appendix \ref{app:smooth} contains a detailed discussion of some smoothness assumptions which simplify the classification program.
\section{Effective Description of 5d SCFTs}
\label{sec:review}
In this section we discuss some of the preliminaries that set the stage for the classification of 5d SCFTs later in this paper. The following discussion involves two perspectives on 5d $\mathcal N =1$ theories:
the gauge theoretic perspective, and the geometric perspective of M-theory compactified on a Calabi-Yau 3-fold.
5d superconformal field theories (SCFTs) are strongly interacting systems with no marginal deformations \cite{Cordova:2016emh} and no known Lagrangian description at the CFT fixed point. In order to study the physics of these conformal theories, one needs to use rather indirect approaches. 5d SCFTs admit supersymmetric relevant deformations which lead to several weakly interacting effective descriptions while preserving some amount of supersymmetry. Surprisingly, these effective descriptions can be powerful tools for studying the dynamics of the conformal point. There exist some CFT observables which are rigidly protected under the renormalization group (RG) flow triggered by these deformations. Many BPS quantities are such observables: for example, the spectrum of BPS operators, supersymmetric partition functions, effective Lagrangians on the Coulomb branch, the Coulomb branch of moduli space, etc. In particular, BPS observables are protected by supersymmetry and thus we expect BPS quantities appearing in the effective theories to be a reliable description of the corresponding observables at the CFT fixed point.
String theory provides many effective descriptions of 5d SCFTs. Multiple D4-brane systems in Type IIA string theory and ($p,q$) 5-brane webs in Type IIB string theory can engineer various 5d SCFTs as singularities. Away from the singularity, when mass parameters and gauge couplings are turned on, these brane systems often permit a gauge theory description of the corresponding 5d theories.
5d SCFTs can also be engineered in M-theory: M-theory on a singular non-compact Calabi-Yau 3-fold is described at long distances by an SCFT living on the five-dimensional spacetime transverse to the 3-fold. In familiar cases, the Calabi-Yau singularity can be resolved by means of various K\"ahler deformations, which correspond to mass and Coulomb branch deformations in the corresponding gauge theory.
\subsection{Gauge theory description}
Gauge theories in five dimensions are non-renormalizable and flow to free fixed points at low energy. As a result, these theories are typically believed to be `trivial' theories. However, a large class of 5d gauge theories, mostly engineered in string theory, turn out to have interacting CFT fixed points in the UV \cite{Seiberg:1996bd}. In such cases, 5d gauge theories are rather interesting since they can provide low energy effective descriptions of the CFT.
In this paper, we focus primarily on gauge theories which have 5d SCFTs as their UV completions. These theories preserve $\mathcal{N}=1$ supersymmetry, and their massless field content consists of vector multiplets with gauge algebra $G$ and hypermultiplets in a representation $\textbf{R} = \oplus \textbf{R}_j$ of $G$. These gauge theories might be further specified by topological data $k$ corresponding to classical Chern-Simons level, as in the case of $G = SU(N \geq 3)$, or discrete $\theta$-angle as in the cases $G= Sp(N)$. We can also consider the cases with product gauge algebra $G=\prod_i G_i$. Once the data $G,\textbf{R},k$ is fixed, the low energy gauge theory Lagrangian is uniquely determined by supersymmetry. Our notation for describing 5d gauge theories is
\begin{align}
G_k + \sum_j N_{\textbf{R}_j} \textbf{R}_j,
\end{align}
where $\textbf{R}_j$ is the representation under which the $j$-th matter hypermultiplet is charged, $N_{\textbf{R}_j}$ is the number of hypermultiplets in the representation $\textbf{R}_j$.
5d $\mathcal{N}=1$ gauge theories possesses a rich vacuum structure. The moduli space of vacua is parametrized by expectation values of various local operators. In particular, we are interested in the Coulomb branch of vacua parametrized by vacuum expectation values of scalar fields $\phi$ in the vector multiplets. Here the scalar field $\phi$ takes values in the Cartan subalgebra of the gauge group $G$. So the dimension of the moduli space of the Coulomb branch is given by the rank of group $G$, $r={\rm rank}(G)$.
By abuse of notation, we will denote both a scalar field in the vector multiplet and its expectation value by $\phi$ from now on.
There are global symmetries acting on the hypermultiplets. The classical Lagrangian has global symmetry algebra $F$ rotating the perturbative hypermultiplets and also a topological $U(1)_I$ symmetry for each gauge group. The objects charged under the $U(1)_I$ are non-perturbative particles called `instantons'. Surprisingly, this classical global symmetry is often enhanced in the CFT fixed point by non-perturbative instanton dynamics \cite{Seiberg:1996bd,Douglas:1996xp}. The flavor symmetry of the perturbative hypermultiplets can combine with the topological $U(1)_I$ instanton symmetry and enhance to an even larger symmetry algebra in the UV CFT. One can turn on mass parameters $m_i$ associated to the global symmetry. Doing so breaks some of the global symmetry. In particular, the mass deformation with parameter $g^{-2}$ along the $U(1)_I$ instanton symmetry leads to a gauge theory description with gauge coupling $g$ at low energy.
At a generic point in the Coulomb branch, the gauge symmetry $G$ is broken to the maximal torus $U(1)^{r}$. Thus the low energy dynamics on the Coulomb branch can be effectively described by abelian gauge theories.
The low energy abelian action is determined by a prepotential $\mathcal{F}$. The prepotential is 1-loop exact and the full quantum result is a cubic polynomial of the vector multiplet scalar $\phi$ and mass parameters $m_j$, given by \cite{Witten:1996qb,Intriligator:1997pq}:
\begin{equation}
\label{eqn:pre}
\mathcal{F} = \frac{1}{2g^2}h_{ij}\phi_i \phi_j + \frac{k}{6} d_{ijk}\phi_i\phi_j\phi_k + \frac{1}{12}\left(\sum_{e\in {\rm root}} |e\cdot \phi|^2 -\sum_j \sum_{w\in {\bf R}_j}|w\cdot\phi+m_j|^3\right) \ ,
\end{equation}
where by abuse of notation ${\bf R}_j$ denotes the set of weights of the $j$-th hypermultiplet representation of $G$, $h_{ij}={\rm Tr}(T_iT_j)$, and $d_{ijk}=\frac{1}{2}{\rm Tr}_{\bf F}(T_i\{T_j,T_k\})$ with ${\bf F}$ in the fundamental representation. The first two terms in the prepotential are from the classical Lagrangian and the last two terms are 1-loop corrections coming from integrating out charged fermions in the Coulomb branch. We remark that the prepotential may have different values in the different sub-chambers (or phases) of the Coulomb branch due to the absolute values in the 1-loop contributions.
The 1-loop correction to the prepotential renormalizes the gauge coupling. The effective coupling in the Coulomb branch is simply given by a second derivative of the quantum prepotential which also fixes the exact metric on the Coulomb branch:
\begin{equation}
\label{eqn:der}
(\tau_{\rm eff})_{ij} = (g^{-2}_{\rm eff})_{ij} = \partial_i\partial_j\mathcal{F} \ , \qquad ds^2 = (\tau_{\rm eff})_{ij}d\phi_id\phi_j \ .
\end{equation}
Interestingly, the exact spectrum of magnetic monopoles on the Coulomb branch can be easily obtained from the quantum prepotential. Since monopoles are magnetically dual to electric gauge bosons, tensions of magnetic monopole strings can be computed as
\begin{equation}
\label{eqn:mono}
\phi_{Di} = \partial_i\mathcal{F} \ , \quad i=1,\cdots r \ .
\end{equation}
One can also compute Chern-Simons couplings:
\begin{align}
k_{ijk} = \partial_ i \partial_j \partial_k \mathcal F.
\end{align}
Therefore, we can use $\mathcal F $ to exactly compute some quantum observables such as the Coulomb branch metric and monopole spectrum.
In \cite{Intriligator:1997pq,Jefferson:2017ahm}, the above supersymmetry protected data is used to attempt a classification of possible 5d SCFTs admitting low energy gauge theory descriptions. The main idea in these classification programs is that the quantum metric on the Coulomb branch should be positive semi-definite in the CFT limit, as required by unitarity. In \cite{Intriligator:1997pq}, the positivity condition of the metric was imposed throughout the `perturbative' Coulomb branch and all sensible gauge theories were subsequently identified using this constraint. In this classification, the `perturbative' Coulomb branch is determined by forcing only \emph{perturbative} particles to have positive masses. Under this condition, the number and type of hypermultiplets are strictly constrained and quiver type gauge theories are ruled out; see \cite{Intriligator:1997pq} for details. We refer to this classification as the `IMS classification'.
However, it was pointed out later works \cite{Aharony:1997ju,Bergman:2014kza,Hayashi:2015fsa,Gaiotto:2015una,Yonekura:2015ksa} that string theory can engineer many 5d gauge theories with non-trivial CFT fixed points not included among the theories in the IMS classification. It turns out that the condition of metric positivity throughout the entire perturbative Coulomb branch is too strong \cite{Jefferson:2017ahm} and unnecessarily excludes many non-trivial 5d gauge theories. This suggests that the IMS classification is incomplete, and the gauge theories exceeding the IMS bounds lead us to revisit the problem of classifying 5d SCFTs.
Let us briefly review the classification of \cite{Jefferson:2017ahm}. One of the main results of this analysis is the observation that the `perturbative' Coulomb branch receives quantum corrections by light non-perturbative states \cite{Aharony:1997ju}. It is possible that some of non-perturbative states can become massless somewhere in the perturbative Coulomb branch. These hyperplanes in the Coulomb branch where these light states become massless can be thought of as `non-perturbative' walls. Beyond such walls, the perturbative Coulomb branch breaks down. One way to see this is to note that the signature of the quantum metric on the Coulomb branch changes beyond these non-perturbative walls, which implies the metric cannot be trusted in these regions. However, the classification in \cite{Intriligator:1997pq} imposes metric positivity on the whole perturbative Coulomb branch, even beyond non-perturbative walls. The result is that some theories are excluded because of the unreliability of the metric in these regions, and this leads to an incomplete classification. In order to obtain a complete classification, metric positivity should be applied only on the `physical' Coulomb branch, which can be computed by accounting for restrictions introduced by non-perturbative states.
In general, it is difficult to identify the correct physical Coulomb branch after taking into account non-perturbative effects since this necessarily involves studying the full non-perturbative spectrum. In particular, it is not easy to analyze the spectrum of gauge theory instantons. Only when we know a precise UV completion of the instanton moduli space, such as the ADHM construction, can we compute the exact spectrum using localization. For most gauge theories, such a convenient construction of the instanton moduli space is lacking.
Fortunately, the perturbative prepotential contains part of the exact spectrum of non-perturbative states. As noted in (\ref{eqn:mono}), the full monopole spectrum can be obtained from the prepotential. We can use this information to identify some of the non-perturbative walls in the perturbative Coulomb branch. By relaxing the metric positivity constraint to apply only to the region interior to such non-perturbative walls, it was conjectured in \cite{Jefferson:2017ahm} that all gauge theories having interacting CFT fixed points satisfy the metric positivity condition in the sub-locus of Coulomb branch where perturbative particles and monopole strings have positive masses. In \cite{Jefferson:2017ahm}, it was also shown that a large class of known 5d gauge theories satisfy this criterion. It may be true that all the known 5d gauge theories having 5d SCFT fixed points satisfy this refined condition.
In addition, there are two more conjectures in \cite{Jefferson:2017ahm} used to carry out the classification of 5d gauge theories with simple gauge algebras. The first conjecture is that if all perturbative particles and monopoles have positive masses \emph{somewhere} in the Coulomb branch, the gauge theory has a UV CFT fixed point. The second conjecture is that perturbative prepotentials of all gauge theories with UV CFT fixed points are positive \emph{everywhere} in the perturbative Coulomb branch.
Note that the first conjecture is not sufficient to guarantee that all instanton particles have positive mass and also that the metric is positive in the same region. So this is simply a necessary condition. We will see later that certain theories predicted by this approach must be excluded because some non-perturbative particles acquire negative masses in the CFT limit.
The second conjecture is based on the convergence of the 1-loop sphere partition function of 5d CFTs, but there is neither physical nor mathematical motivation for this conjecture beyond its practical implications.
Using these two conjectures, non-trivial gauge theories with single gauge node were fully classified in \cite{Jefferson:2017ahm}. This classification includes all known single gauge node theories and additionally predicts a large number of new gauge theories.
In this paper, we construct rank 1 and rank 2 CFTs using Calabi-Yau geometry. Rank 1 gauge theories arising from SCFTs were classified in \cite{Seiberg:1996bd,Morrison:1996xf,Intriligator:1997pq,Katz:1996fh}; these theories have gauge algebra $SU(2)$ with $N_\textbf{F}\leq 7$. Geometrically, the rank 1 SCFTs can be engineered by del Pezzo surfaces embedded in a non-compact $3$-fold.
The families of rank 2 gauge theories predicted by the classification of \cite{Jefferson:2017ahm} are displayed in Table \ref{tb:rank2-gauge-theory-clssification}. The UV completions of the theories shown in Table \ref{tb:rank2-gauge-theory-clssification} are all expected to be 6d theories, rather than 5d SCFTs; on the other hand, their descendants obtained by mass deformations are expected to have 5d CFT fixed points.
Many of these theories in Table \ref{tb:rank2-gauge-theory-clssification} are new theories, for example $SU(3)$ with $(N_{\bf F},|k|)=(6,4),(3,\frac{13}{2}),(0,9)$ in $(a)$.
One of the purposes of this paper is to check if the new rank 2 CFTs predicted in \cite{Jefferson:2017ahm} (or descendants of theories in Table \ref{tb:rank2-gauge-theory-clssification}) can be constructed geometrically. We will see that, surprisingly, almost all new theories in Table \ref{tb:rank2-gauge-theory-clssification} admit geometric constructions, therefore their descendants indeed have interacting CFT fixed points. However, some theories do not correspond to geometries in their conformal limits due to subtle non-perturbative effects. Therefore, the geometric constructions of this paper indicate that the criteria described in \cite{Jefferson:2017ahm} require additional non-perturbative corrections in order to be complete.
We hope to revisit the field theoretic approach of \cite{Jefferson:2017ahm} in the near future with the benefit of our improved understanding.
\begin{table}
\centering
\begin{subtable}[t]{0.45\linewidth}
\centering
\vspace{0pt}
\begin{tabular}{|c|c|c|}
\hline
$N_{\textbf{Sym}}$ & $N_{\textbf F}$ & $|k|$ \\
\hline
$1$ & $0$ & $\frac{3}{2}$\\
\hline
$1$ & $1$ & $0$ \\
\hline
$0$ & $10$ & $0$\\
\hline
$0$ & $9$ & $\frac{3}{2}$\\
\hline
$0$ & $6$ & $4$\\
\hline
$0$ & $3$ & $\frac{13}{2}$\\
\hline
$0$ & $0$ & $9$\\
\hline
\end{tabular}
\caption{Marginal $SU(3)$ theories with CS level $k$, $N_{\textbf{Sym}}$ symmetric and $N_{\textbf F}$ fundamental hypermultiplets.}
\label{tb:SU3-classification}
\end{subtable}\hfill
\begin{subtable}[t]{0.45\linewidth}
\centering
\vspace{0pt}
\begin{tabular}{|c|c|}
\hline
$N_{\textbf{AS}}$ & $N_{\textbf F}$ \\
\hline
$3$ & $0$\\
\hline
$2$ & $4$\\
\hline
$1$ & $8$\\
\hline
$0$ & $10$\\
\hline
\end{tabular}
\caption{Marginal $Sp(2)$ gauge theories with $N_{\textbf{AS}}$ anti-symmetric, $N_{\textbf F}$ fundamental hypermultiplets. The theory with $N_{\textbf{AS}}=3$ can have $\theta=0,\pi$.}
\label{tb:Sp2-classification}
\vspace{0.5cm}
\begin{tabular}{|c|c|c|}
\hline
$N_{\textbf F}$ \\
\hline
$6$ \\
\hline
\end{tabular}
\caption{A marginal $G_2$ gauge theory with $N_{\textbf F}$ fundamental matters.}
\label{tb:G2-classification}
\end{subtable}
\caption{Rank 2 gauge theories.}\label{tb:rank2-gauge-theory-clssification}
\end{table}
\subsection{M-theory compactifications}
\label{sec:Mth}
String compactifications are an extraordinarily useful tool for realizing local, non-perturbative models of gauge sector physics in terms of brane dynamics. Consider in particular M-theory on a non-compact singular Calabi Yau variety $Y$, which is conjectured to be described at low energies by a 5d $\mathcal N = 1$ SCFT. We are specifically interested in studying the Coulomb branch deformations of these 5d SCFTs. The heart of this analysis is the correspondence between the Coulomb branch $\mathcal C$ and the extended K\"ahler cone $\mathcal K(Y)$ of the singular threefold $Y$ \cite{Witten:1996qb}:
\begin{align}
\mathcal C =\mathcal K(Y).
\end{align}
The above correspondence is made more precise by establishing a dictionary between the geometry of the threefold and the BPS spectrum of the associated 5d theory, which we now describe in detail. Consider a smooth non-compact 3-fold $X$. The K\"ahler metric of $X$ depends on $h^{1,1}(X)$ moduli controlling the sizes of complex $p$ cycles in $X$. In order to decouple gravitational interactions, it is necessary to scale the volume of $X$ to be infinitely large while keeping the volumes of all 2- and 4-cycles at finite size; this has the effect of sending the 5d Planck mass to infinity. Given a basis $D_i \in H^{1,1}(X)$, one may therefore express the K\"ahler form $J$ as the linear combination
\begin{align}
J = \phi_i D_i,~~ i = 1, \dots, h^{1,1}(X),
\end{align}
where the K\"ahler moduli $\phi_{i=1,\dots, r}$ associated to (cohomology classes dual to) compact 4-cycles $D_i = S_i$ are identified with Coulomb branch moduli, while the K\"ahler moduli $\phi_{r+j,\dots, r+M}=m_{j=1,\dots, M}$ associated to non-compact 4-cycles $D_{r+j} = N_j$ are interpreted as mass parameters of the 5d theory. To align the discussion with the 5d field theoretic interpretation, we find it useful to partition the K\"ahler moduli into $r$ Coulomb branch parameters and $M$ mass parameters:
\begin{align}
h^{1,1}(X) = r + M.
\end{align}
Note that when the associated 5d field theory admits a description as a gauge theory, $r$ coincides with the rank of the gauge group.
The BPS states of the 5d theory include electric particles and (dual) magnetic strings. Geometrically these states correspond to M2 branes wrapping holomorphic 2-cycles and magnetic dual M5 branes wrapping holomorphic 4-cycles, and the masses and tensions of these BPS degrees of freedom are proportional to the volumes of the corresponding holomorphic cycles. At a generic point $\phi \in \mathcal C$ the spectrum of BPS states is massive, and this is reflected by the fact that the 2- and 4-cycles of $Y$ have finite volume. Since the conformal point $\phi = 0$ is characterized by the appearance of interacting massless and tensionless degrees of freedom, we interpret the threefold $Y$ as a singular limit of the smooth threefold $X$ in which some collection of compact 4-cycles have collapsed to a point. Said differently, $X$ is a desingularization of $Y$.
The above discussion suggests that the data of the massive BPS spectrum is encoded in the geometry of $X$. Indeed this is the case, the main connection to geometry being the interpretation of the 5d prepotential (\ref{eqn:pre}) as the cubic polynomial of triple intersection numbers of 4-cycles in $X$:
\begin{align}
\mathcal F = \text{vol}(X) =\frac{1}{3!} \int_X J^3 =\frac{1}{3!} \phi_i \phi_j \phi_k\int_X D_i \wedge D_j \wedge D_k.
\end{align}
In the previous section, we saw that various data characterizing the massive BPS spectrum can be expressed as derivatives of $\mathcal F$. This data equivalently characterizes the geometry of $X$. In particular, the tensions (\ref{eqn:der}) of elementary monopole strings are the volumes of the compact 4-cycles $S_i$:
\begin{align}
\phi_{Di} = \partial_i \mathcal F =\text{vol}(S_i)= \frac{1}{2!} \int_X J^2 \wedge S_i,~~ 1 \leq i \leq r,
\end{align}
the matrix of effective couplings has as its components the volumes of various 2-cycles:
\begin{align}
\tau_{ij} = \partial_i \partial_j \mathcal F= \text{vol}(S_i \cap S_j) = \int_X J \wedge S_i \wedge S_j,~~ 1 \leq i,j \leq r,
\end{align}
and the effective Chern-Simons couplings $k_{ijk}$ are triple intersection numbers:
\begin{align}
k_{ijk} = \partial_i \partial_j \partial_k \mathcal F = \int_X D_i \wedge D_j \wedge D_k.
\end{align}
The K\"ahler cone $\mathcal K$ of the singularity $Y$ can also be specified quite easily; $\mathcal K$ is simply the set of all positive K\"ahler forms (parametrized by the moduli $\phi$):
\begin{align}
\mathcal K(X \backslash Y) = \{ J = \phi_i D_i ~|~\int_{C} J > 0 ~~\text{for all holomorphic curves $C \subset X$} \}.
\end{align}
Thus, it is possible to study Coulomb branch deformations of 5d SCFTs purely in terms of the geometry of a smooth 3-fold $X$. Generically there are multiple smooth 3-folds $X_i$ which share a common singular limit $Y$, so the extended K\"ahler cone is simply the closure of the union of K\"ahler cones,
\begin{align}
\mathcal K(Y) = \overline{\cup \mathcal K(X_i \backslash Y)}.
\end{align}
The extended K\"ahler cone has the structure of a fan, with pairs of cones separated by hypersurfaces in the interior of $\mathcal K(Y)$. The boundaries of $\mathcal K(X_i \backslash Y)$ correspond to loci where the 3-fold $X_i$ develops a singularity. The interior boundaries are regions where a holomorphic curve collapses to zero volume and formally develops negative volume in the adjacent K\"ahler cone, signaling a flop transition (see Section (\ref{sec:gtrans}) for further discussion.) By contrast, the boundaries of $\mathcal K(Y)$ are loci where one of the 4-cycles can collapse to a 2-cycle or a point. The SCFT point is the origin of $\mathcal K(Y)$, and corresponds to the singularity $Y$ which is characterized by a connected union of 4-cycles shrinking to a point.
In some cases the 5d theory associated to a 3-fold $X$ admits a description as a gauge theory. In such cases, the abelian gauge algebra is $H^2(X,\mathbb R) / H^2(X ,\mathbb Z)$ and enhances to a non-abelian gauge algebra in the singularity $Y$. The simple coroots of the gauge algebra correspond to the classes $S_i \in H^2(X,\mathbb Z)$, whereas the simple roots are generic fibers $f_j$ contained in $H_2(X,\mathbb Z)$. More precisely, the W-bosons of the 5d theory correspond to M2-branes wrapping holomorphic curves $f_j$, and so the Cartan matrix $A_{ij}$ is the matrix of charges
\begin{align}
\label{eqn:Cartan}
A_{ij} = - \int_{f_j} S_i.
\end{align}
In practice, we work in an algebro-geometric setting in which volumes of holomorphic cycles can be computed as intersection products. Thus the volumes of 2-cycles $C_i \subset H_2(X,\mathbb Z)$ and 4-cycles $S_i \subset H_4(X,\mathbb Z)$ are expressed in terms of the intersection products of numerical classes of (resp.) complex curves $[C]$ and surfaces $[D]$. That is, $\text{vol}(C) = (J \cdot [C])_X$ and $\text{vol}(S) = (J \cdot J \cdot S_i)_{X}$. We abuse notation and use the same symbols to denote $p$-cycles, their homology classes, and their numerical equivalence classes whenever the context is clear.
\section{Classification Program}
\label{sec:algorithm}
\subsection{Physical equivalence classes of 3-folds}
\label{subsec:constructshrink}
In this section we propose a classification of CY 3-folds defining 5d SCFTs via M-theory compactification. One way to approach this problem is to study singular 3-folds for which there exist desingularizations that preserve the Calabi-Yau condition (i.e. \emph{crepant resolutions}.) However, the problem of classifying singular 3-folds admitting crepant resolutions is notoriously difficult. Rather than attempting to classify singularities, we instead classify \emph{physical equivalence classes} of singularities. We define a pair of 3-folds to be physically equivalent (i.e. leading to the same SCFT, up to decoupled sectors) if they are related by a finite change in K\"ahler and complex parameters.
There is a conjectural aspect to this definition which we now clarify.
It is immediate from the above definition that normalizable K\"ahler and complex deformations do not change the physical equivalence class of a 3-fold, since these deformations do not change the singular limit (and hence do not change the SCFT). However, we also find it useful to identify 3-folds that differ by non-dynamical large complex deformations. While the singular limits of such 3-folds are not identical, we claim they are nevertheless closely related in that their SCFTs differ at most by decoupled free states
As we will see, the notion of physical equivalence dramatically simplifies the problem of classification.
\subsection{Shrinkable 3-folds}
\label{sec:shrinkable}
In this section we specify the necessary criteria a smooth 3-fold must satisfy in order to define a 5d SCFT. Note that we assume all 5d SCFTs have a \emph{maximal} Coulomb branch, meaning that there exists a phase in which the 5d theory has no dynamical massless hypermultiplets, possibly after turning on some mass parameters. Geometrically this means that we assume there exists a smooth 3-fold which has no normalizable (dynamical) complex structure deformations. The geometry of such a 3-fold is thus controlled by three types of parameters: normalizable K\"ahler (i.e. Coulomb branch) parameters, non-normalizable K\"ahler (i.e. mass) parameters, and non-dynamical non-normalizable complex structure deformation parameters (see Section \ref{sec:transitions} for an example).
Before spelling out the necessary criteria, we recall the key features of the geometries which are the subject of our analysis. We are interested in smooth, non-compact CY 3-folds $X$ containing a finite number of compact 4-cycles $S_i$ and non-compact 4-cycles $N_j$. As discussed in the previous section the number of independent compact 4-cycles is equal to the number of Coulomb branch parameters, while the number of mass parameters is identified with the number of non-normalizable K\"ahler deformations. The 4-cycles $S_i \subset X$ are irreducible projective algebraic surfaces, hence K\"ahler. Moreover, $X$ also contains compact 2-cycles which can either be isolated or part of a family of compact 2-cycles belonging to one of the 4-cycles.
From the physics perspective the natural condition for CY 3-folds to lead to SCFTs is that we can tune non-normalizable K\"ahler parameters (mass parameters) so that at a finite distance in normalizable K\"ahler moduli space we can reach a singular CY 3-fold which has no finite volume cycles or surfaces. However, formulating this in algebro-geometric terms is not simple. Instead we formulate it in a somewhat different way which we believe is equivalent to this. Namely,
in order for a 3-fold $X$ to define a 5d SCFT, $X$ must satisfy the property of being \emph{shrinkable}, which we define below:
\medskip\noindent
{\bf Definition.}
\label{def:shrinkability}
Let $X$ be a smooth CY 3-fold modeled locally as the neighborhood of a connected union of compact K\"ahler surfaces $S = \cup S_i$. We say $X$ is \emph{shrinkable} if there exists an intersecting (possibly empty) union of non-compact surfaces $N=\cup N_j$ and a limit $Y$ of K\"ahler metrics such that:
\begin{enumerate}
\item $S$ (and all curves $C \subset S$) have zero volume in $Y$;
\item $Y$ is at finite distance from a metric $X_0$ for which $N$ has zero volume while $S$ has positive volume.
\end{enumerate}
By abuse of terminology, we say the surface $S$ is shrinkable if $S$ is contained in a shrinkable 3-fold $X$ as a maximal compact algebraic surface.
\medskip
Let us now translate the above definition of shrinkability into a set of necessary geometric conditions. We consider first the limit where all non-normalizable K\"ahler moduli have been set to zero. In this limit we may have a singular 3-fold which is described by the K\"ahler class $J=\phi_iS_i$. Our convention is to assume $\phi_i\ge0$ and compute volumes with respect to $-J$; thus, the volume of a curve $C$ is given by $\mathrm{vol}(C)= -J\cdot C$ and the volume of a divisor $D$ is $\text{vol}(D) = J^2 \cdot D$.\footnote{This choice of sign is consistent with
the description of K\"ahler classes $J$ on compact CY 3-folds, as the expansion of $J$ (or any other ample divisor class) in terms of $S_i$ will have non-positive coefficients. A simple example illustrating this point is the rank~1 case, for which $S$ is a del Pezzo surface. Since $J\cdot C = \phi K_S \cdot C$, it follows that $J$ has non-positive intersection with all curves $C \in S$. We therefore have to change the sign in order for $J$ to be a limit of K\"ahler classes on $X$.} Since we require $-J$ to define a K\"ahler metric which assigns postive volumes to complex $p$-cycles in $X$, a necessary condition for shrinkablity is
\begin{equation}
\label{eq:shrinkability}
\mathrm{vol}(C)= -J\cdot C\ge0,~~\forall C\subset S.
\end{equation}
What happens when the inequality (\ref{eq:shrinkability}) is saturated? Suppose there exists a curve $C$, with $\text{vol}(C)=0$. So far, we have only considered the case in which all non-normalizable K\"ahler moduli are set to zero. To give finite volume to $C$ requires a non-normalizable K\"ahler deformation, which in turn implies the existence of a non-compact 4-cycle $N$ attached to $S$ along $C$. Notice that since $C$ belongs to $N$, there may also be other compact curves $C'$ which are homologous to $C$ in $N$; in particular, the full set of curves homologous to $C$ can fiber over $N$. For each of these curves $C'$ it must be that $\text{vol}(C')=0$, and thus $N$ can be said to have degenerated to a non-compact 2-cycle along its fibers.\footnote{It would interesting to compare this defintion of shrinkability with the conjecture of \cite{Xie:2017pfl} that canonical 3-fold singularities give 5d SCFTs, since it is known that the only noncompact 4-cycles in a Calabi-Yau (crepant) resolution of a canonical 3-fold singularity are ADE fibrations. However, we do not need this for the description in our classification.} By making a non-normalizable K\"ahler deformation, we can bring the curve $C = S \cap N$ to finite volume, and we expect that we are again in a situation where the surface $S$ is contractible.
We believe that the above necessary criteria are in fact sufficient to define a shrinkable 3-fold:
\medskip
\noindent\emph{Conjecture}. Let $X$ be a smooth CY 3-fold modeled locally as the neighborhood of a connected union of compact K\"ahler surfaces $S= \cup S_i$. Then $S$ is shrinkable provided that $- J \cdot C \geq 0$ for all curves $C \subset S$ and that there is one $S_i$ with positive volume and the rest should have non-negative (possibly zero) volume.
\medskip
Elliptic Calabi-Yau 3-folds are immediately ruled out by these criteria. F-theory on an elliptic 3-fold engineers a 6d theory. In a 6d theory, cubic terms in the prepotential $\mathcal{F}$ are trivial; they are non-trivial only when we compactify the 6d theory on a circle and turn on holonomies for gauge symmetries where the circle size is inversely proportional to a mass parameter (or a non-compact K\"ahler parameter). This means that the volumes of all 4-cycles in the associated 3-fold are zero when we turn off mass parameters (or equivalently, in the 6d limit). Therefore elliptic 3-folds are not shrinkable.
\subsection{Building blocks for shrinkable 3-folds}
\label{sec:buildingblocks}
We now argue in favor of a series of simplifying assumptions we make concerning the surfaces $S$ which are instrumental for our proposed classification of shrinkable rank 2 surfaces modulo physical equivalence. Observe that when the inequalities of (\ref{eq:shrinkability}) are all strict, then $S$ is \emph{contractible} \cite{grauert}, so that $S$ can be contracted to an isolated singular point $p$ of a singular 3-fold $Y$. In more precise mathematical terms, this means there exists a holomorphic map $f:X \to Y$ with $f(S)=p$ such that $f$ restricts to an isomorphism away from $S$, i.e. $f|_{X-S}:X-S \cong Y-p$. Since $X$ is at finite distance from $Y$ in moduli space, it is evident that contractibility of $S \subset X$ implies shrinkability of $X$. When a curve has zero volume, we expect that we can obtain a contractible surface by means of a non-normalizable K\"ahler deformation which involves bringing non-compact 4-cycles to finite volume. Hence, we conjecture that a holomorphic map $f$ exists when $S$ is shrinkable, as well:
\medskip
\noindent\emph{Conjecture}. Let $X$ be a shrinkable CY 3-fold modeled locally as a neighborhood of a connected union of compact K\"ahler surfaces $S= \cup S_i$ meeting a (possibly empty) collection of non-compact surfaces $N = \cup N_j$. Then there exists a holomorphic map $f:X \to Y$ sending $S$ to a point $p$ and $N$ to a collection of curves $C$ such that $\left. f\right|_{X - S - N} : X - S - N \to Y - C$ is an isomorphism.
\medskip
The existence of a holomorphic map $f$ as described above permits a number of simplifying assumptions for the following reasons. Replacing the singular 3-fold $Y$ by its normalization if necessary, we can assume that the singularities of $Y$ are normal. It follows that $Y$ has ``canonical singularities'', and moreover that $X$ is a crepant resolution of $Y$. But it is known the components of the resolutions of canonical threefold singularities $Y$ are rational or ruled \cite{can3f}.
We next argue that we can further restrict the types of possible building blocks by exploiting physical equivalence:
\medskip
\noindent\emph{Conjecture}. Shrinkable surfaces are physically equivalent to a shrinkable surface $S=\cup S_i$, where the irreducible components $S_i$ are either equal to $\mathbb P^2$ or a blowup $\text{Bl}_{p} \mathbb F_n$ of a Hirzebruch surface at $p$ points intersecting one another (or self-intersecting) transversally. Moreover, there exist non-negative integers $p_{\text{max}}(n)$ such that $p \leq p_{\text{max}}(n)$.
\medskip
We briefly discuss the content of the above conjecture, deferring a more detailed discussion of the first two points to Section \ref{sec:transitions}.
In that section, we describe the rank~2 case only. For higher rank, we have to also consider the situation where three surfaces can intersect transversally.\footnote{Since four or more surfaces in a threefold cannot intersect nontrivially and transversally, we only need to consider intersections of three surfaces at a time.} At such a point of intersection, called a triple point, the three intersecting surfaces have local equation $xyz=0$. As part of the argument in Section \ref{sec:transitions}, we blow up a point where two surfaces intersect, at which the intersecting surfaces have local equation $xy=0$, so our construction will not apply at a triple point. To handle triple points, we simply supplement the argument in Section~\ref{sec:gtrans} by noting that a complex structure deformation will keep a point to be blown up distinct from any of the triple points.
\smallskip\noindent
\begin{enumerate}
\item Using a combination of complex structure and K\"ahler deformations, it is possible
to map a 3-fold containing a ruled surface over a genus $g$ to a 3-fold containing a Hirzebruch surface. We defer a detailed discussion to Section~\ref{sec:transitions}.
\item In all examples that we have investigated, we have been able to bypass non-transverse intersections in one of two ways: either by a complex structure deformation, or by a K\"ahler deformation in the form of a flop. The idea is that when we flop a curve (in $S_1$, say) which passes through a point of non-transversal intersection, the result is to blow up $S_2$ at that point, simplifying the singularity of the intersection curve and rendering it more transverse.
We therefore assume that a combination of complex and K\"ahler deformations will always suffice to produce a 3-fold containing transversally intersecting surfaces $S_i$.
\item We prove in Appendix \ref{app:Mori} that if $p>p_{\text{max}}(n)$ there are infinitely many generators for rational curves. The presence of infinitely many generators is expected to indicate the presence of an infinite dimensional global symmetry group. An example of this is $\text{dP}_9$ (note $p_{\text{max}}(1)=7$), in which case the symmetry group permuting these generators is the affine $E_8$ Weyl group. In such a case, the Weyl group is infinite dimensional, and can be interpreted as a finite symmetry group of a 6d theory viewed from the 5d perspective. As we discussed above, geometries associated to 6d theories are not shrinkable. Since a CFT should not have an infinite dimensional global symmetry group, we claim that surfaces $S_i$ with an infinite number of Mori cone generators cannot be building blocks for 5d SCFTs and are thus excluded.
\end{enumerate}
\subsection{Consistency conditions for shrinkable 3-folds}
\label{sec:consistency}
The condition that $S$ is contained in a CY 3-fold imposes constraints on the curves of intersection of the components of $S$, which will be exploited in a crucial way in our classification program.
Let $S_1$ and $S_2$ be two smooth surfaces glued along a curve $C = S_1 \cap S_2$. Now suppose that $S_1\cup S_2$ is contained in a 3-fold $X$, and that the intersection of $S_1$ and $S_2$ is transverse in $X$.
Then the normal bundle of $C$ in $X$ is given by $N_{C,X}=N_{C,S_1}\oplus N_{C,S_2}$. The Calabi-Yau condition then implies
\begin{equation}
\label{eq:gluingcond}
C^2_{S_1}\oplus C^2_{S_2}=2g-2,
\end{equation}
where $g$ is the genus of $C$ and the subscripts on the right-hand side denote the irreducible surface in which the self-intersection takes place. The gluing curves must satisfy the adjunction formula for each surface $S_i$:
\begin{align}
\label{eq:adjunction}
(K \cdot C)_{S_i} + C_{S_i}^2 = 2g - 2,
\end{align}
where $K_{S_i}$ is the canonical class of the surface $S_i$. For the rank 2 case, which is the primary focus of this paper, we argue in Section \ref{sec:rank2} that it suffices for our classification to assume that $g=0$.
Suppose a compact connected holomorphic surface $S$ satisfies the above constraints on its curves of intersection. These constraints immediately imply that a CY 3-fold can be found containing a neighborhood in $S$ of the curves of intersection (for example, the total space of the normal bundle of $S_1 \cap S_2$ in $X$ works, as the complement of $S_1 \cap S_2 \subset S$ is smooth). Moreover, we can also find local CY 3-folds containing the complement of the intersection curves $S_1 \cap S_2$ in $S$ (for example, just take the total space of the canonical bundle as before). Therefore, it seems reasonable to expect that above two types of local models can be glued to form a local model of a CY 3-fold. In other words,
given smooth holomorphic surfaces $S_1$ and $S_2$ glued along a smooth curve $C$ and satisfying (\ref{eq:gluingcond}),
a smooth CY 3-fold $X$ can be found containing $S=S_1\cup S_2$.
While we have not proven that such an $X$ can always be found if (\ref{eq:gluingcond}) and (\ref{eq:adjunction}) are satisfied, these conditions are consistent with all known examples and it is presumably not too difficult to rigorously prove this.
We emphasize here that the above gluing condition is a local condition that has no bearing on the overall topology of the surface $S$, and therefore permits a variety of interesting configurations. In principle there is nothing preventing, for example, gluing two surfaces together along multiple irreducible curves. Another interesting configuration involves two curves belonging to a single surface $S_i$ being glued together. However, we will see that the only gluing configurations which play a role in the rank 2 classification are pairwise transverse intersections between the irreducible components $S_1$ and $S_2$.
The above discussion plays an essential role in our classification because we do not need to actually construct $X$ to proceed; rather, we only require the existence of $X$ and the existence of a surface $S$ can be used as a proxy for the existence of a local 3-fold. Thus the problem of classifying shrinkable 3-folds can be reduced to the problem of classifying embeddable, shrinkable surfaces $S$.
\subsubsection*{A simple example: $S = \mathbb F_0 \cup \mathbb F_2$}
An illustrative example of this construction is a simple complex surface $S=S_1 \cup S_2$ with $S_1= \mathbb F_0, S_2 = \mathbb F_2$ as depicted in Figure \ref{fig:F0F2}. Our rank 2 ansatz gives us
\begin{align}
\begin{split}
\label{eqn:geotrip}
J^3 &= S_1^3 \phi_1^3 + S_2^3 \phi_2^3 + 3 \phi_1 \phi_2 (J \cdot S_1 \cdot S_2) =K_{S_1}^2 \phi_1^3 + K_{S_2}^2 \phi_2^3 - 3 \phi_1 \phi_2 \text{vol}(S_1 \cap S_2).
\end{split}
\end{align}
The first order of business is to determine an appropriate gluing. Gluing these two surfaces together requires us to identify an irreducible, smooth curve $C = S_1 \cap S_2$ belonging to the Mori cone of both surfaces, satisfying (\ref{eq:gluingcond}). In the case of Hirzebruch surfaces $\mathbb F_{n_i}$, the Mori cones are the positive linear spans $\langle E_{i}, F_{i} \rangle$, where the curve classes satisfy the intersections $F_i^2= 0, E_i \cdot F_i =1, E_i^2 = -n_i$, so the range of possibilities is severely restricted. The gluing condition (\ref{eq:gluingcond}) implies that the self intersection of one of the two gluing curves must be negative. Since the curve $E$ is the unique rational curve with negative self intersection \cite{GH}, it therefore follows that we must select $C_{S_i} = E_{i}$ for one of the two surfaces, say $C_{S_2} =E_{2}$. The other curve must then satisfy
\begin{align}
C_{S_1}^2 = 0.
\end{align}
As a trial solution let us take $C_{S_1} = a F_{1} + b E_{1}$, so that $C_{S_1}^2 = 2 ab = 0$. Therefore, either $a = 0$ or $b = 0$. From the adjunction formula (\ref{eq:adjunction}), we know that $(C \cdot E_1 + C \cdot F_1)_{S_1}= a+b= 1$, and therefore the remaining nonzero coefficient must be set equal to unity. To be concrete, we choose
\begin{align}
C_{S_1} = F_{1},~~~ C_{S_2} = E_{2}.
\end{align}
Now that we have constructed the surface $S$, we must check that the local 3-fold $X$ associated to this surface is shrinkable. We parametrize a K\"ahler class $J$ as follows:
\begin{align}
J = \phi_1 [\mathbb F_0] + \phi_2 [ \mathbb F_2],
\end{align}
where $[\mathbb F]$ is the class associated to the 4-cycle $\mathbb F \subset X$. The Mori cone of $X$ is the union of the Mori cones of the component surfaces $S_i$, namely the positive span $\langle E_{1}, E_{2}, F_{2}
\rangle$ (we omit $F_{1}$ because the gluing identifies $F_1$ and $E_2$.) Therefore, the shrinkability condition (\ref{eq:shrinkability}) implies
\begin{align}
(\text{vol}(E_1), \text{vol}(E_2), \text{vol}(F_2) ) = (2 \phi_1 -\phi_2 , 2 \phi_1, -\phi_1 + 2\phi_2 ) \geq 0.
\end{align}
Since that the above conditions can be satisfied for a nontrivial set of Coulomb branch parameters $\phi_i$, we conclude that the geometry $X$ corresponds to a 5d SCFT on the Coulomb branch.
\begin{figure}
\begin{center}
\includegraphics[scale=.5]{F0-F2.pdf}
\end{center}
\caption{Example of a gluing construction of the K\"ahler surface $S = \mathbb F_0 \cup \mathbb F_2$. The gluing curves in both surfaces, $C_1, C_2$, are encircled by dashed lines in the left figure. The final geometry (on the right) is the result of identifying these two curves subject to the conditions described in Section \ref{sec:algorithm}.}
\label{fig:F0F2}
\end{figure}
\subsection{Geometry of physical equivalences}
\label{sec:transitions}
In this section we discuss some important types of physical equivalences upon which our classification relies. Many of these equivalences identify 3-folds related by geometric transitions, i.e.\ maps between smooth geometries which involve passing through an intermediate singularity. Another type of physical equivalence identifies 3-folds related by a ``large" change in the complex structure of non-dynamical modes, which interpolates between two singular geometries---this is a Hanany-Witten transition \cite{Hanany:1996ie}. We illustrate these two types of maps in turn.
\subsubsection{Geometric transitions}
\subsubsection*{Flop transitions}
\label{sec:gtrans}
One of the simplest and most thoroughly studied types of geometric transitions is a \emph{flop transition}, which is a topology-changing transition $X \rightarrow X'$ between two 3-folds $X, X'$ that is in practice typically realized by blowing down a $-1$ curve $C \subset X$ and blowing up a different $-1$ curve $C' \subset X'$ (see Figure \ref{fig:flop}). A flop is a birational map $X\dashrightarrow X'$ which is an isomorphism away from curves $C,C'$, with $K_X\cdot C=K_{X'}\cdot C'=0$. If $C$ and $C'$ are both isomorphic to ${\mathbb P}^1$, the flop is called a simple flop. Simple flops were classified in \cite{km}.
In field theoretic terms, a flop transition corresponds to a continuous change of the mass of a particular state in the matter hypermultiplet from positive to negative values; this change corresponds to a singular phase transition on the Coulomb branch.
\begin{figure}
\begin{center}
\includegraphics[scale=.7]{flop-trans.pdf}
\end{center}
\caption{A local illustration of a flop transition $X \rightarrow X'$ between two CY 3-folds. The red lines in both diagrams correspond to the $-1$ curves in (respectively) $X$ and $X'$.}
\label{fig:flop}
\end{figure}
\subsubsection*{Genus reduction}
We saw in Section~\ref{sec:buildingblocks} that the $S_i$ can be ruled surfaces over higher genus curves as well as genus 0. Here we argue that by our notion of physical equivalences we can restrict to $g=0$ using geometric transitions. This can be obtained by composing a complex structure deformation of a surface $S_i$ with a flop transition. This provides a map from a ruled surface over a curve of
genus $g$ to a self-glued Hirzebruch surface.
This type of geometric transition is particularly important because it exhibits the non-normalizable K\"ahler moduli of the local 3-fold defined by a ruled surface over a curve of genus $g$ as blowup parameters of the 3-fold defined by a self-glued surface $\text{Bl}_{2g} \mathbb F_n$. While we have not proven that the transition can always be achieved in the higher rank case due to the requirement that additional compact surfaces remain glued throughout the transition, we nevertheless believe this construction can be extended to higher rank surfaces with at most minor modifications.
\begin{figure}
\begin{center}
\begin{tikzpicture}
\node(a) at (0,0) {$
\begin{tikzpicture}[yscale=.8,xscale=1.1]
\draw [thick] (0,0) to [out=90,in=180] (1,1) to [out=0,in=180] (2,.85) to [out=0,in=180] (3,1) to [out=0,in=90] (4,0) to [out=270,in=0] (3,-1) to [out=180,in=0] (2,-.85) to [out=180,in=0] (1,-1) to [out=180,in=270] (0,0);
\draw[thick] (.7,0) to [out=-20,in=180] (1.1,-.13) to [out=1,in=200] (1.5,0);
\draw[thick] (2.5,0) to [out=-20,in=180] (2.9,-.13) to [out=1,in=200] (3.3,0);
\draw[thick] (2.6,-.05) -- (3.2,-.05);
\draw[thick] (.8,-.05) -- (1.4,-.05);
\end{tikzpicture}
$};
\node(b) at (6,0) {$
\begin{tikzpicture}[yscale=.8,xscale=1.1]
\draw [thick] (0,0) to [out=90,in=180] (1,1) to [out=0,in=180] (2,.85) to [out=0,in=180] (3,1) to [out=0,in=90] (4,0) to [out=270,in=0] (3,-1);
\draw[thick] (.8,-1) to [out=180,in=270] (0,0);
\draw[thick] (2.5,0) to [out=-20,in=180] (2.9,-.13) to [out=1,in=200] (3.3,0);
\draw[thick] (2.6,-.05) -- (3.2,-.05);
\draw[thick] (.9,-.09) -- (1.45,-.09);
\draw[thick](.7,-.05) to [out=-10,in=90] (1.2,-.45) to [out=270,in=0] (.8,-1);
\draw[thick] (1.6,-.03) to (1.4,-.13) to [out=200,in=90] (1.2,-.5) to [out=270,in=180] (1.6,-1) to (3,-1);
\end{tikzpicture}
$};
\draw[big arrow] (a) -- (b);
\end{tikzpicture}
\end{center}
\caption{A genus $g= 2$ Riemann surface degenerating into a $g= 1$ Riemann surface with a nodal singularity as the result of identifying two points. By identifying $g$ pairs of points in this manner, it is possible for a smooth curve of genus $g$ to degenerate into a rational curve with $g$ nodal singularities.}
\label{fig:degen}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=.6]{self-gluing.pdf}
\end{center}
\caption{A transition from a ruled surface over a $g=1$ curve to a Hirzebruch surface. The red point in the second figure is a blowup point on a nodal curve and the red lines in the third figure are the exceptional curves. Two proper transforms of the fiber $F$ in a blown up Hirzebruch surface are glued together along the nodal curve.}
\label{fig:selfgluing}
\end{figure}
Before giving a detailed description of this geometric transition, we recall that by the irreducibility of the moduli space $\overline{M}_g$
of stable curves of genus $g$ the complex structure of a smooth curve $C$ of genus $g$ can be
degenerated to a rational curve $C_0$ with $g$ nodes (see Figure \ref{fig:degen}.) The curve $C_0$ can be
constructed directly by identifying $g$ pairs of points of ${\mathbb P}^1$. Note that this construction immediately extends to give a degeneration of a ruled surface
$S$ over $C$ to a ruled surface $S_0$ over the singular curve $C_0$. Conversely,
the degeneration of the ruled surface can be described by starting with ${\mathbb P}^1$-bundle over
${\mathbb P}^1$ (i.e.\ a Hirzebruch surface ${\mathbb F}_n$) and identifying $g$ pairs of fibers
$F \subset \mathbb F_n$.
However, this description of $S_0$ is not completely satisfactory, as $S_0$ cannot be embedded into a CY 3-fold for the following reason. Let $F\subset S_0$ be one of the singular fibers obtained by identifying $g$ pairs of fibers. Locally, $S_0$ has two branches near $F$ with equation $xy=0$ (pulled back
from the local equation $xy=0$ of a node of $C_0$). Being a fiber, $F$ has self-intersection 0 in each branch, So if $S_0$ were contained in a smooth
threefold, the normal bundle of $F$ would be ${\cal O}_F\oplus{\cal O}_F$. Fortunately, the geometric transition naturally rectifies this problem by introducing blowups, in a manner which we describe below.
Consider again the degeneration point of view, which can be described by a holomorphic map $\pi:{\cal S}\to \Delta$. Here ${\cal S}$ is a smooth\footnote{Requiring ${\cal S}$ to be smooth is not a problem; its local equation near a point of $F$ can be taken as $xy=t$, which is smooth. This is the same local calculation which shows that $\overline{M}_g$ is smooth at the nodal curves (in the orbifold sense).} threefold, $\Delta$ is a disk, $\pi^{-1}(0)\simeq S_0$, and $\pi^{-1}(t)$ is diffeomorphic to
$S$ for $t\ne0$. We now pick a point $p\in F\subset S_0\subset {\cal S}$ and blow up $p$ to get $\phi:\widetilde{{\cal S}}\to {\cal S}$. Via $\pi\circ\phi$ we can view $\widetilde{\cal S}$ as a family over $\Delta$. However, $\widetilde{\cal S}$ and ${\cal S}$ are isomorphic over $\Delta-0$, so this gives another degeneration of $S$. The singular limit is $(\pi\circ\phi)^{-1}(0)$, which we now describe.
Blowing up a point $p$ in a smooth threefold creates an exceptional divisor $E$ isomorphic to ${\mathbb P}^2$, and blows up $S_0$ to a surface $\widetilde{S_0}$. We have $(\pi\circ\phi)^{-1}(0)=\widetilde{S}_0\cup {\mathbb P}^2$. It remains to describe $\widetilde{S}_0$ and how ${\mathbb P}^2$ is attached to it.
Since $S_0$ has local equation $xy=0$ at $p$, the exceptional curve of $\widetilde{S}_0\to S_0$ has $xy=0$ as its equation. In this latter instance, the equation $xy=0$ is understood as a homogeneous equation in the exceptional
${\mathbb P}^2$ of the blown-up threefold. In other words, ${\mathbb P}^2$ meets $\widetilde{S_0}$ in two intersecting projective lines $L,L'$; each of these ${\mathbb P}^1$'s can be thought of as arising from the blowup of $p$ in a corresponding branch of $S_0$ near $p$.
The point of intersection $q = L \cap L'$ also intersects the proper transform $\widetilde{F}$ of the original singular fiber $F$. The curve $\widetilde{F}$ is still singular in $\widetilde{S_0}$ and still has two branches in a local
description, but now the blowup has reduced the self-intersection from $0$ to $\widetilde F^2 = -1$ in each branch. So if $\widetilde{S_0}$ is contained in a smooth threefold, then the normal bundle of $\widetilde{F}$ is ${\cal O}_F(-1)\oplus {\cal O}_F(-1)$ and the threefold can be Calabi-Yau!
We can apply this construction to all of the $g$ singular fibers. Since $\widetilde{F}$ has self-intersection $-1$ in each branch, we can view it as the gluing of a pair of exceptional ${\mathbb P}^1$'s.
Therefore the
resulting $\widetilde{S}_0$ is a blown up Hirzebruch surface with $g$
pairs of exceptional curves identified. Each singular fiber consists of a double curve with self-intersection $-1$ in each branch, glued at a common point $q$ to curves $L,L'$ of self-intersection $-1$ in each of the respective local branches
(the surface $\widetilde{S}_0$ is smooth along $L\cup L'-\{q\}$).
In the degeneration described above, we also need to attach $g$ copies of ${\mathbb P}^2$. However, we are only concerned with the rank~2 case, so in our examples
these ${\mathbb P}^2$'s can replaced by noncompact cycles containing $L\cup L'$ and safely ignored.
The final step is to flop the $g$ curves $\widetilde{F}_1,\ldots\widetilde{F}_g$, where we have added a subscript to $\widetilde{F}$ to distinguish
these curves. Let us investigate the birational transform of $\widetilde{S_0}$ after the flops.
When the curves $\widetilde{F}_i$ are contracted, the points of intersection $q_i = L_i \cap L_i'$ become conifolds. When we complete the flops, new ${\mathbb P}^1$'s appear in place of the $q_i$ and the curves $L_i,L'_i$ get separated. These curves
become identified with fibers of a ruled surface over the desingularization $\widetilde{C}_0$ of $C_0$, the fibers over the pairs of points of $\widetilde{C}_0$ which get identified to form a node of $C_0$. Since $\widetilde{C}_0$ is isomorphic to ${\mathbb P}^1$, the result
is a Hirzebruch surface in general with blowups.
\subsubsection*{An example of genus reduction: $G_2 + N_\textbf{F} \textbf{F}$ }
An illustrative example of complex deformations that exchange ruled surfaces over a curve of genus $g >0$ for self-glued Hirzebruch surfaces blown up at $2g$ points is the family of shrinkable 3-folds engineering $G_2 + N_{\textbf{F}} \textbf{F}$, as described in \cite{Diaconescu:1998cn}.
We begin by recalling the form of the gauge theoretic 1-loop prepotential for $G_2 + N_{\textbf{F}} \textbf{F} + N_{\textbf{adj}} \textbf{adj}$:
\begin{align}
\begin{split}
\label{eqn:G2nomass}
6 \mathcal F_{\text{1-loop}} &= ( 8 - 8 N_\textbf{F} -8 N_\textbf{adj}) \phi_1^3 + ( 8 - 8 N_{\textbf{adj}}) \phi_2^3\\
&~+ 3 \phi_1 \phi_2 [ (6 +3 N_\textbf{F} - 6 N_\textbf{adj} )\phi_1 + (8 N_{\textbf{adj}} - N_\textbf{F} - 8 ) \phi_2 ].
\end{split}
\end{align}
We set $N_{\textbf{adj}} =0$ to be consistent with $\mathcal N = 1$ supersymmetry. By giving a nonzero value to mass parameters in the hypermultiplet contributions to the prepotential, one can study the RG flow from $N_{\textbf{F}}$ to $N_{\textbf{F}}-1$ flavors. In order to decouple a massive hypermultiplet, the theory must pass through three phase transitions. These four phases have the following prepotentials (we omit mass parameter terms for brevity):
\begin{align}
\begin{split}
\label{eqn:G2RG}
6 \mathcal F_{}^{(1)} &=(8- 8N_\textbf{F}) \phi_1^3 + 8 \phi_2^3 + 3 \phi_1 \phi_2 [\phi _1 \left(3 N_{\textbf F}+6\right)-\phi _2 \left(N_{\textbf F}+8\right)]\\
6 \mathcal F_{}^{(2)} &=(16-8N_\textbf{F}) \phi_1^3 + 7 \phi_2^3 + 3 \phi_1 \phi_2 [ \phi _1 \left(3 N_{\textbf F}+2\right)-\phi _2 \left(N_{\textbf F}+6\right)]\\
6 \mathcal F_{}^{(3)} &=(15 - 8 N_\textbf{F}) \phi_1^3 + 8 \phi_2^3 + 3 \phi_1 \phi_2 [ \phi _1 \left(3 N_{\textbf F}+3\right)-\phi _2 \left(N_{\textbf F}+7\right) ]\\
6 \mathcal F_{}^{(4)} &=6 \mathcal F_{N_\textbf{F}-1}^{(1)}.
\end{split}
\end{align}
We determine a shrinkable K\"ahler surface $S$ that engineers this theory by setting the triple intersection polynomial (\ref{eqn:geotrip}) equal to prepotential (\ref{eqn:G2nomass}) and demanding that there exist an intersection matrix $f_i \cdot S_j = (A_{G_2})_{ij}$ for some choice of fiber classes $f_i \subset S_i$. Restricting the possible building blocks to be blowups of rational and ruled surfaces \emph{without self-gluing}, the only solutions to these conditions are the geometries shown in Table \ref{tab:G2geo}. For all of these surfaces we have $9n_2+6a=2g-2+n_1$, as required by (\ref{eq:gluingcond}). A key point here is that the surface $S_1$ must be a ruled surface of a curve of genus $g = N_{\textbf{F}}$. This is precisely the geometric setup described in \cite{Diaconescu:1998cn}.
\begin{table}
\begin{center}
$
\begin{array}{|c|c|c|c|}
\hline
g & a & (n_1,n_2) \\\hline
0 & 1 & (8,0) \\\hline
1 & 0 & (9,1) \\\hline
2 & 2 & (10,0) \\\hline
3 & 1 & (11,1) \\\hline
4 & 0 & (12,2) \\\hline
4 & 3 & (12,0) \\\hline
5 & 2 & (13,1) \\\hline
6 & 4 & (14,0) \\\hline
\end{array}
$
\end{center}
\caption{Shrinkable surfaces $S = \mathbb F^{g}_{n_1} \cup \mathbb F_{n_2}$ engineering $G_2 + N_\textbf{F} \textbf{F}$ gauge theories. The surface $\mathbb F^g_{n_1}$ is a ruled surface over a curve $E$ with $g(E) = N_{\textbf{F}}$ and satisfying $E^2 = -n_1$. The gluing curve $C = S_1 \cap S_2$ is given by $C_{S_1} = E$ and $C_{S_2} = a F + 3 H$. The fiber classes are given by are $f_i = F_i$.}
\label{tab:G2geo}
\end{table}
We now demonstrate that we can engineer the same family of theories described above by replacing $S_1$ with the surface $S_1' = \text{Bl}_{ 2g} \mathbb F_{n_1}^{(g)} $, where again $g = N_{\textbf{F}}$ and the superscript notation indicates $S_1'$ is obtained by identifying $g$ pairs of exceptional curves in $\text{Bl}_{2g} \mathbb F_{n_1}$ (i.e. self-gluing; see Appendix \ref{app:math} for some mathematical background.) This shrinkable surface not only reproduces the prepotential (\ref{eqn:G2nomass}) and $G_2$ Cartan matrix, but also has the merit of exhibiting the RG flow (\ref{eqn:G2RG}) in a very natural manner. The four phases, related by flops, have the following geometries:
\begin{enumerate}
\item $ \text{Bl}_{2g} \mathbb F^{(g)}_{8-g} \cup \mathbb F_{n_2}$, where the blowups are all at special points\footnote{Note that while we consider blowups at special points $F \cap E \subset \mathbb F_n$ here for convenience, since we do not introduce any additional irreducible curves with self intersection less than $-1$, we can without loss of generality view a blowup of $\mathbb F_n$ at $p$ special points as a blowup of $\mathbb F_{n+p}$ at $p$ general points. We explore the distinction between special and general points in more depth in Section \ref{sec:rank2}.} $F \cap E$.
\item $\text{Bl}_{2g-2} \mathbb F^{(g-1)}_{8- g} \cup \text{Bl}_1 \mathbb F_{n_2}$.
\item $\text{Bl}_{2g-1} \mathbb F^{(g-1)}_{8- g} \cup \mathbb F_{n_2\pm{} 1}$.
\item $\text{Bl}_{2g-2} \mathbb F^{(g-1)}_{9- g} \cup \mathbb F_{n_2 \pm{} 1}$.
\end{enumerate}
The first phase is $ \text{Bl}_{2g} \mathbb F^{(g)}_{8-g} \cup \mathbb F_{n_2}$, where we introduce $g$ self-gluings of $\text{Bl}_{2g} \mathbb F_p$ along the pairs of exceptional divisors $X_{2i}, X_{2i-1}, i = 1, \dots, g$,\footnote{Here and in the sequel, we use the notation $X_i$ to denote the exceptional divisor of the $i$-th blowup, since we reserve the more standard notation $E_i$ for sections of Hirzebruch surfaces.}the where the gluing curve is defined by $C_{S_1} = E - \sum_{i=1}^{2 g} X_i$ and $C_{S_2} = F + 3 H$, so that $a=1$ in the notation adopted in the caption of Table~\ref{tab:G2geo}. Since the canonical class\footnote{More precisely, the dualizing sheaf of the singular surface $ \text{Bl}_{2g} \mathbb F^{(g)}_{8-g}$, pulled back to its natural desingularization $ \text{Bl}_{2g} \mathbb F_{8-g}$.} is given by $K_{\mathbb F_{8 - g}} + 2\sum_{i=1}^{N_\textbf{F}} (X_{2i-1}+X_{2i})$, we find a perfect match with the first line of (\ref{eqn:G2RG}), using the adjunction relation $9n_2+6-(8+g)=2g-2$.
We now describe the flop to the second phase. The matter curve with volume $2\phi_1 - \phi_2$ which shrinks is one of the self-gluing exceptional divisors, say $X_1$. Blowing down $X_1$ forces us to also blow down $X_2$.
We can blow up $\mathbb F_{n_2}$ at a generic point $F_2 \cap H_2$ if we eventually want to decrease $n_2$ to $n_2 -1$, or at a special point $F_2 \cap E_2$ if we want to increase $n_2$ to $n_2 +1$ in the third phase.
The geometry of the second phase is $\text{Bl}_{2g-2} \mathbb F^{(g-1)}_{8- g} \cup \text{Bl}_1 \mathbb F_{n_2}$, where $C_{S_1} = E -\sum_{i=1}^{2g -2} X_i$ and $C_{S_2}=a F + 3 H - 2 Y_1$. Since the blowup of ${\mathbb F}_{n_2}$ is at the double point of $E$ introduced by gluing $X_{2g-1}$ to $X_{2g}$, the coefficient of $Y$ in $C_{S_2}$ is $-2$.
The matter curve with volume $\phi_2 - \phi_1$ which we blow down is $F_2 - Y_1 \subset \text{Bl}_1 \mathbb F_{n_2}$. Because $F - Y_1$ meets $C$ in one point, we must introduce an exceptional divisor $Y_2$ in the surface $S_1$, leading us to the third phase.
The geometry of the third phase is $\text{Bl}_{2g-1} \mathbb F^{(g-1)}_{8- g} \cup \mathbb F_{n_2\pm{} 1}$, where $C_{S_1} = E - \sum_{i=1}^{2g-2} X_i - Y_2$. Concerning the gluing curve class $C \subset \mathbb F_{n_2 \pm{} 1}$, there are two possible cases. In the case of a generic blowup, the proper transforms of $H, F \subset S_2$ are $H- Y_1, Y_1$, so we set $C_{S_2} = (a +1) F + 3 H$, where now $ H^2_{S_2} = n_2 -1$. It follows that $C^2_{S_2} =((a +1) F + 3 H)_{S_2}^2 = 6 (a + 1) + 9 (n_2-1) = 3 g + 3$, which is a nontrivial check that this geometry is consistent with the phase structure of the $G_2$ theory. On the other hand, in the case of a special blowup, the difference is that the proper transform of $H \subset S_2$ is $H$, so that $C_{S_2} = H + (a - 2) F$, where now $H_{S_2}^2 = n_2 +1$. We again confirm that $C^2_{S_2} = ((a -2) F + 3 H)_{S_2}^2 = 6 (a - 2) + 9 (n_2 +1) = 3 g+ 3$.
In order to reach the fourth and final phase, the matter curve with volume $\phi_1$ which we blow down is $F - Y_2 \subset S_1$. The geometry of the fourth phase is $\text{Bl}_{2g-2} \mathbb F^{(g-1)}_{9- g} \cup \mathbb F_{n_2 \pm{} 1}$. Keeping in mind the previous identity $n_1 = 8- g$ along with the fact that we blow down the curve $F - Y_2 \subset S_1$, we compute the canonical class:
\begin{align}
\begin{split}
K_{S_1} &= -2 H + (n_1 -2) F + 2\sum_{i=1}^{g-1} (X_{2i-1} + X_{2i} ) + Y_2\\
&= -2 H + ((n_1+1) -2) F + 2 \sum_{i=1}^{g-1} (X_{2i-1} + X_{2i} ) .
\end{split}
\end{align}
Note also that the self-intersection of $H \subset S_1$ shifts from $8-g$ to $9-g$.
\subsubsection{Hanany-Witten transitions and complex deformations}
The next type of transition we will discuss is a \emph{complex structure deformation}.
In particular, we concern ourselves with two types of complex structure deformations that preserve the rank of the 3-fold. The first type of complex structure deformation is a Hanany-Witten (HW) transition \cite{Hanany:1996ie}. This type of transition is most easily understood in the setting of $(p,q)$ 5-brane webs, and involves interchanging the relative position of a $(p,q)$ 7-brane and a $(p,q)$ 5-brane. After the transition, despite the fact that the brane webs look different, in the low-energy decoupling limit the corresponding SCFTs describe the same physics up to decoupled free sectors. The example displayed in Figure \ref{fig:HW} describes a geometric (or HW) transition from a local 3-fold $X$ with $S= \mathbb F_2$ to another 3-fold $X'$ with $S' = \mathbb F_0$. Therefore, $X$ and $X'$ are physically equivalent.
\begin{figure}
\begin{center}
\includegraphics[scale=.5]{F0-F2-HW.pdf}
\end{center}
\caption{Hanany-Witten transition from $\mathbb F_2$ to $\mathbb F_0$. The $\otimes$ symbol denotes the location of a transverse $(0,1)$ 7-brane, and the dashed line denotes the location of the 7-brane monodromy cut.}
\label{fig:HW}
\end{figure}
This example can be geometrically described as follows: ${\mathbb F}_2$ is physically equivalent to ${\mathbb F}_0$ by a (non-normalizable) complex structure deformation. One way to see this is to first contract the curve $E$ in ${\mathbb F}_2$ (with $E^2=-2$) to an $A_1$ singularity, which can be identified with the quadric cone $x^2+y^2+z^2=0$ in ${\mathbb P}^3$. A complex structure deformation takes this to a smooth quadric surface (e.g.\ $w^2+x^2+y^2+z^2 =0$), which is isomorphic to ${\mathbb P}^1\times{\mathbb P}^1={\mathbb F}_0$.
Another type of complex structure deformation involves changing special type blow ups (i.e. blow ups on top of blow ups) to generic blow ups, where the blow up points are not on top of one another, unless the blow up curve is part of the identification between $S_i$'s. We will show that in the rank 2 case this can be avoided and we can always assume general point blow ups.
\section{Classifications}
\label{sec:classification}
Let $S=\cup S_i$ be a connected union of surfaces contained in a CY 3-fold $X$. We classify all shrinkable $S$ for rank 1 and rank 2 according to the conjectures and algorithm described in Section \ref{sec:algorithm}.
We first summarize the rank 1 and rank 2 classification results and in the next two subsections we present details of the classification.
All rank 1 and rank 2 shrinkable geometries (or SCFTs) belong to one or more families of geometric RG-flows, and the geometries in each RG-flow family are related by rank-preserving mass deformations (or blowdowns of -1 curves in geometric terminology), up to physical equivalence. The ideas of geometric RG-flow and rank-preserving mass deformations will be discussed later.
Based on these ideas, we can start from a ``top'' geometry, which corresponds to a 5d CFT or a 6d CFT on a circle (equivalently, a 5d Kaluza-Klein (KK) theory), and obtain all other geometries in the same family by a finite sequence of geometric transitions or mass deformations. This UV geometry is at the top of the RG-flow in a given family and can therefore be a representative of the entire RG-flow family. We conjecture that all descendants of the top UV geometry engineer 5d SCFTs. When shrinkable, the top UV geometry itself also engineers a 5d SCFT.
For rank 1 geometries, we have only one RG-flow family corresponding to a local elliptic 3-fold defined by the del Pezzo surface $\text{dP}_9$. All other rank 1 geometries are obtained by blowing down exceptional curves. The RG-flow family of $\text{\text{dP}}_9$ involves other del Pezzo surfaces $\text{dP}_n$ with $n\le 8$ and a Hirzebruch surface $\mathbb{F}_0$; it is believed that these are the complete set of geometries leading to rank 1 5d SCFTs.
\begin{table}
\centering
\begin{tabular}{|c|c|}
\hline
$S=S_1\cup S_2$ & $G$ \\
\hline
$(\mathbb{F}_6\cup \text{dP}_4)^*$ & $Sp(2)_{\theta=0} + 3\textbf{AS}$ \\
\hline
$(\mathbb{F}_2\cup \text{dP}_7)^*$ & $SU(3)_4 + 6\textbf{F}$ \\
& $Sp(2)+ 4 \textbf{F} + 2\textbf{AS}$ \\
& $G_2 + 6 \textbf{F}$ \\
\hline
$(\text{Bl}_9\mathbb{F}_4\cup \mathbb{F}_0)^*$ & $SU(3)_{\frac{3}{2}} + 9 \textbf{F}$ \\
& $Sp(2)+8\textbf{F} + \textbf{AS}$ \\
\hline
$(\text{Bl}_{10}\mathbb{F}_6\cup \mathbb{F}_0)^*$ & $SU(3)_0+10 \textbf{F}$ \\
& $Sp(2)+ 10\textbf{F}$ \\
\hline
\end{tabular}
\caption{Rank 2 geometries with maximal $M$. In the above table, $S$ is the rank 2 K\"ahler surface, while $G$ is the corresponding gauge theory description. These geometries denoted as $(\cdot)^*$ are not shrinkable and correspond to 5d KK theories.}
\label{tb:rank2-classification}
\end{table}
Similarly, the top rank 2 geometries are summarized in Table \ref{tb:rank2-classification}. We have identified four geometric RG-flow families represented by these top geometries. These geometries are not shrinkable; rather, we expect that these geometries have 6d UV completions and thus they engineer 5d KK theories. However, their descendants, obtained by blowing down $-1$ curves, are shrinkable and therefore give rise to 5d SCFTs.
For example, the geometry $\text{Bl}_9\mathbb{F}_4\cup \mathbb{F}_0$ is ruled out from our CFT classification because its building block $\text{Bl}_9\mathbb{F}_4$ has an infinite number of Mori cone generators as explained in Appendix~\ref{sec:mori}, violating our criterion in Section \ref{sec:buildingblocks}. However, a geometric RG-flow from this geometry by blowing down an exceptional curve as well as a number of flop transitions leads to the geometry $\text{Bl}_8\mathbb{F}_3\cup \text{dP}_1$ which is now shrinkable and engineers a 5d SCFT. Similarly, other geometries in Table \ref{tb:rank2-classification} are associated to KK theories, but their descendants are shrinkable. Therefore, we find that all rank 1 and 2 smooth 3-fold geometries engineering 5d SCFTs are mass deformations of 5d KK theories. See Section \ref{sec:rank2} for further discussion.
This result confirms the existence of many new rank 2 SCFTs predicted in \cite{Jefferson:2017ahm} which are listed in Table \ref{tb:rank2-gauge-theory-clssification}.
For example, the $SU(3)_7$ gauge theory is predicted to exist in Table \ref{tb:SU3-classification}. This theory turns out to have a geometric realization as $\mathbb{F}_0\cup \mathbb{F}_8$ which is a descendant of $\mathbb{F}_2\cup \text{dP}_7$. This implies that the gauge theory approach in \cite{Jefferson:2017ahm}, which analyzes the magnetic monopole and perturbative BPS spectrum, is quite powerful and capable of predicting new interacting 5d SCFTs.
Our study also reveals that there are no smooth 3-fold geometries associated to the following gauge theories:
\begin{align}
\begin{split}
\label{eqn:ruleout}
&SU(3)_{\frac{1}{2}} + 1\bf{Sym} \ , \\
& SU(3)_{7} +2 {\bf F} \ \rightarrow \ SU(3)_{\frac{15}{2}}+1{\bf F} \ \rightarrow \ SU(3)_8 \ .
\end{split}
\end{align}
These theories are expected to have interacting CFT fixed points by the perturbative gauge theory analysis in \cite{Jefferson:2017ahm}. See Table \ref{tb:SU3-classification}.
The SCFT of the first gauge theory indeed exists---this theory is a mass deformation of the $SU(3)_0$ theory with $N_{\bf Sym}=1,N_{\bf F}=1$ whose brane construction is given in \cite{Bergman:2015dpa,Hayashi:2015vhy}. Our study of smooth 3-folds fails to capture this theory. The reason for this failure is because the corresponding geometry involves a `frozen' singularity. For example, the brane construction in \cite{Bergman:2015dpa,Hayashi:2015vhy} contains O7$^+$-planes; indeed, constructions involving O7$^+$ planes are dual to frozen singularities involving non-geometric monodromies and a fractional M-theory 3-form background as discussed in \cite{Tachikawa:2015wka}. Therefore, we do not expect that our analysis can capture this type of singularity, and hence the geometric classification in this paper is incomplete in this sense.
We nevertheless conjecture that our classification includes all 5d SCFTs coming from {\it smooth} Calabi-Yau threefolds which do not involve frozen singularities dual to brane constructions involving O7$^+$ planes.
In the following sections, we classify smooth rank 1 and rank 2 3-fold geometries engineering 5d SCFTs in their singular limits.
On the other hand, we predict that there are no SCFTs corresponding to three gauge theories belonging to the RG flow in the second line of (\ref{eqn:ruleout}). As we discuss in Section \ref{sec:rank2}, despite the fact that these gauge theories can be realized geometrically using our algorithm, they are shrinkable only when we attach a number of non-degenerate non-compact 4-cycles to the compact surface $S$. Introducing these non-compact 4-cycles entails non-normalizable K\"ahler deformations which in the field theory setting corresponds to introducing nonzero mass parameters. We find that these mass parameters cannot be set to zero in the CFT limit---at small nonzero values, the corresponding geometries develop at least one 2-cycle with negative volume and therefore their singular limits do not engineer well-defined CFT fixed points.
This computation excludes the three gauge theories in the second line of (\ref{eqn:ruleout}) as possible candidates for interacting 5d SCFTs. This is also an indication that the classification criteria described in \cite{Jefferson:2017ahm} are necessary, but not sufficient to identify 5d SCFT fixed points. The criteria of \cite{Jefferson:2017ahm} must be modified to account for non-perturbative BPS states (such as instantons in gauge theories) in order to be both necessary and sufficient.
We also remark that a single 3-fold $X$ can admit multiple gauge theory descriptions.
This is possible because some geometries admit more than one distinct choice of fiber class associated to charged gauge bosons. The existence of multiple gauge theoretic descriptions corresponding to a single geometry suggests that the gauge descriptions are dual to one another. Starting with the the ``top'' UV geometries in Table \ref{tb:rank2-classification}, we predict the following dualities:
\begin{align}
\begin{split}
\label{eqn:dual}
SU(3)_{5-\frac{N_{\bf F}}{2}} + N_{\bf F} {\bf F} ~&\cong ~Sp(2)+N_{\bf F}{\bf F} \ , \quad N_{\bf F} \le 10 \\
SU(3)_{6-\frac{N_{\bf F}}{2}}+ N_{\bf F}{\bf F} ~&\cong ~Sp(2)+1{\bf AS}+(N_{\bf F}-1){\bf F} \ , \quad 1 \le N_{\bf F} \le 9 \\
SU(3)_{7-\frac{N_{\bf F}}{2}}+N_{\bf F}{\bf F} ~&\cong ~ G_2 + N_{\bf F}{\bf F} ~ \overset{2\le N_{\bf F}}{\cong}~ Sp(2)+ 2{\bf AS}+(N_{\bf F}-2){\bf F} \ , \quad N_{\bf F} \le 6
\end{split}
\end{align}
The first and the second dualities in (\ref{eqn:dual}) were conjectured already in \cite{Gaiotto:2015una} and in \cite{Jefferson:2017ahm}, respectively. So our construction provides concrete geometric evidence for these duality conjectures. On the other hand, the third duality is a new duality discovered by an explicit geometric construction in this section.
\subsection{Rank 1 classification}
\label{sec:rank1}
We warm up by starting with rank 1, recovering the result that all rank 1 5d SCFTs are geometrically engineered by local 3-folds containing a del Pezzo surface. More precisely, our algorithm identifies del Pezzo surfaces as shrinkable, but also identifies additional shrinkable surfaces; however, each of these turns out to be physically equivalent to a del Pezzo surface.
Recall that a del Pezzo surface $S$ is defined to be a smooth algebraic surface whose anticanonical bundle $-K_S$ is ample---this means that $-K_S \cdot C > 0$ for all effective curves $ C \subset S$. The classification of del Pezzo surfaces is well known: $S$ is either $\text{dP}_n$ for $0\le n\le 8$ or ${\mathbb P}^1\times{\mathbb P}^1={\mathbb F}_0$. Such a surface satisfies (\ref{eq:shrinkability}) as well as $K_S^2>0$, so is shrinkable. We now set out to systematically classify rank 1 shrinkable surfaces up to physical equivalence.
To apply (\ref{eq:shrinkability}), we need to know $K_S$, the generators of the Mori cone of curves on $S$, and the intersection numbers of the curves in $S$. Our algorithm leads us to consider ${\mathbb P}^2$, ${\mathbb F}_n$, and their generic blowups.
${\mathbb P}^2$ is del Pezzo, but it is instructive to check shrinkability anyway. For ${\mathbb P}^2$, the Mori cone is generated by the class $\ell$ of a line, $\ell^2=1$, and
$K_{{\mathbb P}^2}=-3\ell$. So $K_{{\mathbb P}^2}^2=9>0$ and $K_{{\mathbb P}^2}\ell=-3<0$, so ${\mathbb P}^2$ is shrinkable.
Next, we consider ${\mathbb F}_0$,\ ${\mathbb F}_1$ and ${\mathbb F}_{n \geq 2}$ separately. Since ${\mathbb F}_1$ is the blowup of ${\mathbb P}^2$ at a point, ${\mathbb F}_1$ and its generic blowups are just the generic blowups of ${\mathbb P}^2$. Similarly, ${\mathbb F}_0$ is del Pezzo, and the blowup of ${\mathbb F}_0$ at a point is isomorphic to the blowup of ${\mathbb P}^2$ at two points \cite{GH}. So the possibilities for $S$ can be reduced to either generic blowups of ${\mathbb P}^2$, or ${\mathbb F}_{n \geq 2}$.
As usual, we denote by $\text{dP}_n$ the blowup of ${\mathbb P}^2$ at general points
$p_1,\ldots,p_n$. Let $X_1,\ldots, X_n$ denote the corresponding
exceptional ${\mathbb P}^1$'s,\footnote{As noted earlier, we reserve the more customary notation
$E$ for the curves on Hirzebruch surfaces described in Appendix~\ref{app:Mori}.} and we let $\ell$ denote
the class of the total transform in $\text{dP}_n$ of a line in ${\mathbb P}^2$. The intersection numbers are
\begin{equation}
\label{eq:intp2}
\ell^2=1,~~\ X_i\cdot X_j = -\delta_{ij},~~\ \ell\cdot X_i=0
\end{equation}
and $K_{\text{dP}_n}=-3\ell+\sum_{i=1}^nX_i$. Then $K_{\text{dP}_n}^2=9-n>0$ for $n\le 8$.
We first observe that $\text{dP}_n$ is not shrinkable for $n\ge9$. To see this, we simply observe that $K_{\text{dP}_n}^2\le0$ for $n\ge9$ which implies that the string tensions are not positive.
Again, we can cite known results simply say that $\text{dP}_n$ is shrinkable for $n\le8$, but it is instructive to work out details without assuming this fact.
We adopt a convenient shorthand to describe the generators of the Mori cone: Any curve
$C\subset \text{dP}_n$ other than the $X_i$ will project to a curve $D\subset {\mathbb P}^2$ of some degree $d>0$.
Let $m_i$ be the multiplicity of $D$ at $p_i$, so that $m_i=0$ if $p_i\not\in D$, $m_i=1$
if $p$ is a nonsingular point of $D$, $m_i=2$ if $p$ is a node or cusp of $D$, etc. Then
the class of $C$ is $d\ell-\sum_{i=1}^n a_i X_i$. It is customary to abbreviate this class as
$(d;m_1,\ldots,m_n)$, as well as to omit any $m_i$ which are zero. Then the
Mori cone of $\text{dP}_n$ is generated by the classes\footnote{Strictly speaking, we have only written the Mori generators for $n=8$.
For $n<8$, we modify (\ref{eq:moridp}) by removing those generators which need more than $n$ exceptional divisors to define them. In addition, for
$n=1$, we include $(1;1)$ as a generator.}
\begin{equation}
X_i,\ (1;1^2),\ (2,1^5),\ (3,2,1^6),\ (4,2^3,1^5),\ (5,2^6,1^2),\ (6;3,2^7)
\label{eq:moridp}
\end{equation}
up to permuting the order of the $p_i$. It follows from the adjunction formula (\ref{eq:adjunction}) that each of the curve
classes $C$ in (\ref{eq:moridp}) satisfies $K_{\text{dP}_n}\cdot C=-1$,\footnote{ For $n=1$, we also check that $K_{\text{dP}_1}\cdot(\ell-X_1)=-2$.} so $\text{dP}_n$ is shrinkable.
Next, consider the Hirzebruch surfaces $S={\mathbb F}_n$.
Using the notation in Appendix~\ref{app:Mori}, there are two disjoint toric sections $E,H$
and the fiber class $F$. These classes satisfy
\begin{equation}
\label{eq:intfn}
H^2=n,\ E^2=-n,\ H \cdot E=0,\ H \cdot F=E\cdot F=1,\ F^2=0,\ H=E+nF.
\end{equation}
The canonical bundle of ${\mathbb F}_n$ is $K_{{\mathbb F}_n}=-2H+(n-2)F$ and so
$K_{{\mathbb F}_n}^2=8>0$. Furthermore,
the Mori cone of effective curves is generated by $E$ and $F$. While $K_{{\mathbb F}_n} \cdot F=-2<0$,
we also have $K_{{{\mathbb F}}_n}\cdot E=n-2$, which is strictly negative for $n<2$, zero for $n=2$, but
strictly positive for $n>2$. Thus ${\mathbb F}_2$ is shrinkable. However, as discussed in section 3, this is physically equivalent to ${\mathbb F}_0$.
The same reasoning combined with the earlier observation that $\text{Bl}_1{\mathbb F}_0\simeq \text{dP}_2$ shows that $\text{Bl}_p \mathbb F_2$ is physically equivalent to $\text{dP}_{p+1}$.
In conclusion, all rank 1 shrinkable surfaces are physically equivalent to $\text{dP}_n$ for some $n$ or ${\mathbb F}_0$.
\subsection{Rank 2 classification}
\label{sec:rank2}
The main result of this paper is a full classification of shrinkable rank 2 geometries up to physical equivalence. We preface our result by arguing some further simplifying assumptions we make about the surface $S$ in order to make the classification into a manageable problem.
\subsubsection*{Three simplifications}
In this section we show that we can utilize the following three simplifying assumptions for classifying shrinkable rank 2 surfaces:
\begin{itemize}
\item $S_1 \cap S_2$ is an irreducible curve.
\item $S_1 \cap S_2$ is a rational curve.
\item The surfaces $S_i$ are equal to ${\mathbb P}^2$ or Hirzebruch surfaces and their blowups at general points.
\end{itemize}
We now discuss these three simplifications in order.
First, we argue that in the case of a rank 2 surface $S = S_1 \cup S_2$, we can assume that $S_1$ is not glued to $S_2$ along multiple curves. Namely, there exists a single edge between two nodes.
Suppose we glue two surfaces along $C_1,C_2$ with appropriate identifications. Since $S_1$ and $S_2$ should intersect transversally, we have $(C_1 \cdot C_2)_{S_1} = (C_1 \cdot C_2)_{S_2}= 0$. This means that $C_1, C_2$ do not intersect.
We claim there always exists an effective curve $D=d_1+d_2$ such that ${\rm vol}(D) \le 0$. If ${\rm vol}(D)<0$, then $S$ is not shrinkable, so it suffices to consider the situation where ${\rm vol}(D)=0$. But in that case, we will further show below that we can arrange for the curve $D$ to be elliptic (i.e. $g(D) = 1$), which would contradict our conjectures. Therefore, the full surface is not shrinkable implying that we cannot glue two surfaces along two or more curves.
In order to show this, we first prove that there always exist curves $d_i \subset S_i$ with $K_{S_i}\cdot d_i\ge-2$ that intersect both $C_1$ and $C_2$. These classes $d_1$ and $d_2$ are identified as follows.
First, if both $C_1$ and $C_2$ are not fiber classes, we can always find a curve $d_1$ satisfying these conditions among $\{F, \, F-X_i,\, H-X_i-X_j \}$\footnote{For general $n$ we choose $d_1=F-X_i$ if $C_1=X_i$ or $C_2=X_i$, otherwise $d_1=F$. When $n=2$ and $C_1=X_1,C_2=X_2$, we choose $d_1=H-X_1-X_2$.} in $\text{Bl}_p\mathbb{F}_n$, where $X_i$ are exceptional curves associated to the blowups of $\mathbb F_n$ at $p$ general points. When $n>2$, $C_1 =E$, otherwise the volume of the curve $E$ will be negative.
Next, suppose $C_1$ or $C_2$ is a fiber class. This is possible only when $S_1=\text{Bl}_p\mathbb{F}_1$ or $\text{dP}_n$, otherwise the class $E$, which has $E\cdot C_1\neq0$ or $E\cdot C_2 \neq 0$, will have negative volume thus preventing the surface $S$ from being shrinkable. In the case that $S_1=\text{Bl}_p\mathbb{F}_1$, when $C_1$ is a fiber class $F_1$, $C_2$ must be one of $X_i$'s, due to the assumption of transversal intersection. Then we can take $d_1=H-X_i$ with $H^2=1$.
With any choice of $d_1$ given here, we find that ${\rm vol}(d_1)=m\phi_1 - n\phi_2$ with $m=1,2$ and $n\ge2$ where $\phi_i \geq 0$.
We can choose $d_2 \subset S_2$ in the same manner and then show that ${\rm vol}(d_2)=m'\phi_1 - n'\phi_2$ with $m'=1,2$ and $n'\ge2$.
This proves ${\rm vol}(D) \le 0$ for an effective curve $D=d_1+d_2$. Now we will assume ${\rm vol}(C_i)\ge 0$ for all other curves $C_i$ because otherise the surface is not shrinkable and already ruled out. As already noted above, it is clear that the total surface is not shrinkable when ${\rm vol}(D)<0$. Moreover, when ${\rm vol}(D) = 0$, i.e. when $m=m'=n=n'=0$, the curves $d_1$ and $d_2$ are both fiber classes $F_i\subset S_i$. In this case, the curve $F_1$ and $F_2$ can be deformed so that $F_1\cap C_i=F_2\cap C_i$ for $i=1,2$. Then the curve $D=F_1+F_2$ is the union of two rational curves intersecting in two points, hence elliptic. By further complex structure deformation if necessary, we can arrange that all fibers $F_1$ of $S_1$ meet all fibers $F_2$ of $S_2$ in two points, or in other words, that $S=S_1\cup S_2$ is elliptically fibered.
We argue that we can deform the complex structure of $X$ if necessary so that $X$ is also elliptically fibered. To see this, let $E$ be an elliptic fiber of $S$. Since $E$ is part of an elliptic fibration of $S$, we have that $N_{E/S}\simeq\mathcal{O}_E$. Furthermore, $\det(N_{E/X})$ is trivial by the Calabi-Yau condition and the ellipticity of $E$. Then the normal bundle sequence
\begin{equation}
\label{eq:nbs}
0 \to N_{E/S} \to N_{E/X} \to N_{S/X}|_E \to 0
\end{equation}
is identified with
\begin{equation}
\label{eq:Atiyah}
0 \to \mathcal{O}_E \to N_{E/X} \to \mathcal{O}_E \to 0.
\end{equation}
However, since $H^1(\mathcal{O}_E) \ne 0$, (\ref{eq:Atiyah}) generically does not split\footnote{The non-splitting of (\ref{eq:Atiyah}) identifies $N_{E/X}$ as the Atiyah bundle on $E$.} and dim $H^0(N_{E/X})=1$. The uniqueness of a normal direction says that $E$ moves in a 1-parameter family, enough deformations to fiber $S$ but not enough to fiber $X$.
However, we can choose a complex structure deformation of $X$ so that (\ref{eq:Atiyah}) splits, and then $N_{E/X}\simeq\mathcal{O}_E^2$. In this situation, $E$ moves in two independent directions and fibers $X$.
This justifies our claim, hence $S$ is not shrinkable. The same argument holds for cases with more than two edges (i.e. gluing curves) between $S_1$ and $S_2$. Therefore rank 2 geometries formed by two surfaces glued along two or more different curves are not shrinkable.
Second, we claim that the gluing curves must be rational. Suppose $C = S_1 \cap S_2$ has $g>0$. In Appendix~\ref{app:Mori} we explain that we must have finitely many Mori cone generators in each $S_i$ (which implies a bound on the number of blowups), hence we have finitely many Mori cone generators in $X \supset S = S_1 \cup S_2$. We argue that this implies $C_{S_i}^2 \ge 0$ as follows.
We assume $C_{S_i}^2<0$ and derive a contradiction. Since $C_{S_i}^2+C \cdot K_{S_i}=2g-2\ge0$, we have $C\cdot K_{S_i}>0$.
Anticipating the next bulleted claim that the building blocks are generic blowups of Hirzebruch surfaces at a bounded number of points, we show in Appendix~\ref{app:Mori} that $C_{S_i} \cdot K_{S_i}>0$ implies $C_{S_i}=E$. This is a contradiction, since $g>0$.
Although this argument is slightly circular in its current form depending as it does on the next bulleted claim, we believe that with further care we can independently justify $C_{S_i}^2\ge0$. Furthermore, an extensive computer search has revealed no counterexamples.
Let us now return to the claim that the gluing curves are rational. Recalling equations (\ref{eq:gluingcond}) and (\ref{eq:adjunction}), we have
\begin{equation}
C^2_{S_1} + C^2_{S_2} = C_{S_i}^2 + K_{S_i}\cdot C = 2g-2 \ .
\end{equation}
These conditions tell us that $K_{S_i} \cdot C\ge0$. This implies that the volume of the intersection curve, ${\rm vol}(C)=-\phi_1 K_{S_1}\cdot C -\phi_2 K_{S_2}\cdot C$, is negative unless $C^2_{S_1}=C_{S_2}^2=0$ and $g=1$, i.e. unless $C$ is an elliptic curve. This proves that rank 2 geometries containing two surfaces meeting in a curve with genus $g>0$ are not shrinkable.
Third, we observe that many of the building blocks in our classification program are related to one another by maps (for instance, isomorphisms and complex deformations) which at the level of 5d SCFT physics constitute physical equivalences. Therefore, we observe that the full number of rank 2 surfaces that can be constructed from our list of building blocks dramatically overcounts the number of unique CFT fixed points, and hence we can reduce the complexity of the problem at the outset by restricting our attention to a minimal representative set of configurations capturing the full list of physical equivalence classes. We will argue in particular that we need only consider configurations $S = S_1 \cup S_2$ for which $S_1$ is a blowup of $\mathbb F_{n> 0}$ at $p$ generic points\footnote{By ``generic point'', we mean a point not contained in any exceptional divisors, i.e.\ rational curves with self intersection $-1$.} and $S_2$ is $\text{dP}_m$ or $\mathbb F_0$. We summarize our simplifications by stating that {\it every rank 2 shrinkable CY 3-fold can be realized locally as a neighborhood of} $S = S_1 \cup S_2${\it , for which }$S_1 = \text{Bl}_{p} \mathbb F_{n_1 > 0} $\, {\it and} $S_2 = \text{dP}_{n_2}$ {\it or }$\mathbb F_0$. {\it Moreover, the surfaces }$S_1, S_2$ {\it are glued along a single smooth rational curve} $C =S_1 \cap S_2$.
We argue the third simplification as follows. First, observe that all of the curves $C'$ with self intersection $C'{}^2 < -2$ which do not intersect the gluing curve $C$ have negative volume. Therefore, the only curves $C' \neq C$ with negative self-intersection should have $C'{}^2 \geq -2$. Suppose $C'{}^2 = -2$ and the surface $S$ is shrinkable. Then, it should follow that such a geometry is related via complex deformation to a physically-equivalent surface for which the only curves $C'$ of negative self-intersection have $C'{}^2 = -1$. The idea is essentially identical to the description of a transitions already described in Section~\ref{sec:transitions}: we perform a conifold transition. Strictly speaking, this is only true up to physical equivalence, but that is good enough for us. Hence, we may assume that the only component surfaces $S_i$ appearing in our representative classes are those for which all curves $C' \ne C$ satisfy $C'{}^2 \geq -1$. This already places a significant constraint on the possible configurations $S_1 \cup S_2$.
Next, recall that our list of possible building blocks includes $\mathbb P^2$ and $\text{Bl}_p \mathbb F_n$, where the configuration of $p$ points can be special or generic. The gluing condition (\ref{eq:gluingcond}) implies that one of the two gluing curves $C_{S_1}$ or $C_{S_2}$ must have negative self-intersection. Therefore, we are forced to fix one of the two surfaces, say $S_1 = \text{Bl}_p \mathbb F_{n_1}$. Observe that any blowup of $\mathbb F_n$ at $p$ points $F \cap E$ is always isomorphic to the blowup of $\mathbb F_{n+p}$ at $p$ generic points, so (redefining $n$) we can always assume that $S_1$ is a blowup of $\mathbb F_{n_1}$ at $p$ points away from the curve $E$ with self intersection $E^2 = -n_1$.
Assume that $n \geq 2$ and suppose we take such a surface $S_1$ and glue it to $S_2$ along some curve $C_{S_1} \ne E$. Then this violates the condition that all curves $C' \ne C_1$ satisfy $C'{}^2 \geq -1$, in particular for $C' = E$. Hence, we are forced to set $C_{S_1} = E$, and moreover we are confined to surfaces $S_1 = \text{Bl}_p \mathbb F_{n_1}$ for which the configuration of points $p$ is a generic configuration (a special configuration of points would produce curves with self-intersection less than $-1$).
Let us focus on $S_2$. If $n_1 \geq 2$, then $S_2$ must be glued to $S_1$ along a curve $C_{S_2}$ with non-negative self intersection, $C_{S_2}^2 \geq 0$. Since we may again assume that all $C' \ne C_{S_2}$ satisfy $C'{}^2 \geq -1$, it follows that $S_2 = \text{dP}_{n_2}$ or $S_2 = \mathbb F_0$. Returning to the remaining cases $n _1< 2$, we find these cases consist of gluing configurations for which $S_i = \text{dP}_{n_i}$ glued along curves $C_{S_i}$ with $C_{S_i}^2 = -1$. However, $\text{dP}_n \cong \text{Bl}_{n-1} \mathbb F_1$, and therefore in order to avoid overcounting we assume that our configuration is again of the form conjectured above.
Finally, we turn our attention to the case where one of the component surfaces $S_i$ is a ruled surface over a curve of genus $g >0$. As explained in Section \ref{sec:transitions}, a ruled surface over a curve with genus $g >0$ is physically equivalent to a blowup of $\mathbb F_n$ at $2g$ generic points with $g$ self-gluings. Notice that when $S_1$ is the $\text{Bl}_{2g}\mathbb{F}_n$ with $g$ self-gluings, the gluing curve $C_{S_1}$ should be the section $E$ (with $E^2=-n$) since otherwise $E$ has negative volume or leads to an elliptic fiber class. This implies due to the shrinkability condition that the second surface $S_2$ is again ${\rm dP}_m$ or $\mathbb{F}_0$. The self-gluing curves must always be exceptional curves, and hence we perform a flop transition in which we blow these curves down at the expense of blowing up another curve inside the surface $S_2$. Provided we always perform enough blow downs to completely eliminate the self-glued curves, we can always exchange a configuration involving a self-glued blowup of $\mathbb F_n$ with one of the configurations described in the above conjecture. This completes our argument concerning the representative configurations for rank 2 surfaces $S=S_1 \cup S_2$.
\subsubsection*{Endpoint classification: 0 and 1 mass parameters}
\label{subsec:endpoint}
In this section we show that we can first classify geometries which are blown down `as much as possible'; we refer to these as `endpoint geometries'. The general
classification then follows by classifying endpoints and subsequently classifying their possible blowups.
Suppose a SCFT admits mass deformations for its global symmetry. Then we can take a large mass limit and integrate out all the heavy degrees of freedom. This triggers an RG flow and it is expected that the SCFT below energy scales set by the masses flows to another SCFT with a lower rank global symmetry group commuting with the mass deformations of the UV SCFT. In general, such mass deformations can reduce the rank of the resulting theory. Another possibility is for the IR theory to be a trivial free theory.
We pay attention to a particular class of mass deformations which leads to interacting SCFTs while preserving the rank of the UV SCFT. Equivalently, we restrict our attention to mass deformations which do not change the dimension of the Coulomb branch. One can typically obtain a new interacting SCFT with the same rank by means of such `rank-preserving mass deformations'. We expect that RG flows of the UV SCFT triggered by such mass deformations can generate a family of SCFTs with the same rank but different global symmetries. SCFTs in the family are distinguished by their global symmetries (i.e. the number of mass parameters), as well as topological data such as the classical Chern-Simons level $k$ or ${\mathbb Z}_2$-valued $\theta$ angle.
These types of RG flows terminate in a class of interacting SCFTs which we will call `endpoint SCFTs'. An endpoint SCFT is defined to be a theory which does not admit any rank-preserving mass deformations. Thus these theories are `endpoints' of RG flows and they cannot flow to other SCFTs via rank-preserving deformations. Endpoint geometries engineer endpoint SCFTs.
Rank-preserving mass deformations and endpoint geometries are mathematically well-defined notions. We define distinct endpoint geometries to be surfaces which cannot be related to another smooth surface of the same rank via a large mass deformation. Rank-preserving mass deformations are defined as follows: suppose $S$ is shrinkable and $C\subset S_j$ is a $-1$ curve which does not intersect any $S_k$ for $k\ne j$. Then $S$ can be blown down to a surface $S'=\cup S'_i$ with $S'_j$ the blowdown of the $-1$ curve of $S_j$ and $S'_k\simeq S_k$ for $k\ne j$. This type of blowdown is the geometric realization of a rank-preserving mass deformation.
We will now show that {\it if }$S$ {\it is shrinkable, then its endpoint geometry }$S'$ {\it is also shrinkable.}
If $C'\subset S'_i$, let $C\subset S_i$ be its proper transform. We have $K_{S'_i}^2=K_{S_i}^2+1$. If $i\ne j$
we have $K_{S_j}\cdot C=K_{S'_j} \cdot C'$, so we need only consider the case $i=j$.
Let $p\in S'_j$ be the point that the $-1$ curve in $S_j$ blows down to, and suppose that $C'$ has multiplicity $m$ at $p$. Then
$K_{S'_i}\cdot C'=K_{S_i}\cdot C-m$. The desired conclusion follows immediately.
Endpoint SCFTs are interesting due to the following reasons. First, these theories are the simplest theories in their family of RG flows. Their parameter spaces are smaller, so they are comparatively easier to understand than other theories belonging to the same family.
The classification of endpoint SCFTs is therefore a much easier problem than the full classification, as we will see below.
We can thus regard the endpoint classification as a tutorial on our classification algorithm.
Second, all other SCFTs in the family of RG flows in principle can be obtained from endpoint theories by increasing the number of mass parameters. Namely, we can undo mass deformations, and retrace the RG flow to obtain an entire family of UV SCFTs. This could sound puzzling: we know that RG flow is irreversible. So it may be hard to accept the idea that we can restore UV theories starting from an IR theory. However, this turns out to be the case among 5d supersymmetric theories.
Since 5d $\mathcal{N}=1$ SCFTs are so strongly constrained by supersymmetry, one can control their RG flows by tuning discrete data such as (for theories with gauge theory descriptions) gauge algebra, matter representations, classical CS level, and discrete $\theta$ angle. We expect that this allows us to build a family of SCFTs starting from an endpoint theory.
From the geometric standpoint, these constraints can be understood as arising from the Calabi-Yau condition. Mass deformations of a 3-fold correspond to blowups or blowdowns of exceptional curves. As discussed above, a large mass deformation corresponds to blowing down a $-1$ curve which is isolated from gluing curves and is in fact a reversible geometric transition---one can just as easily blow up the same curve to recover the original 3-fold. This means that by starting from an endpoint geometry, it is possible to obtain a family of local (smooth) 3-folds by blowing up all possible exceptional curves. In this sense, the study of endpoint geometries is a good starting point for the classification of 5d SCFTs.
Let us now classify all rank 2 endpoint geometries by employing our classification algorithm. We learned above that rank 2 geometries are constructed by gluing $S_1=\text{Bl}_p\mathbb{F}_{m_1}$ and $S_2={\rm dP}_{m_2}$ or $\mathbb{F}_0$. This implies that endpoint geometries with $M=0,1$ will take the form $\mathbb{P}^2 \cup \mathbb{F}_{n}$ or $\mathbb{F}_{n_1}\cup \mathbb{F}_{n_2}$. For being an endpoint geometry with $M>1$, there must be no irreducible exceptional curve which does not intersect with the gluing curves and no flop transitions introducing such exceptional curve away from the gluing curves. This is possible only for ${\rm dP}_2 \cup {\rm dP}_2$ with $C_1 = \ell \!-\!X_1 \!-\!X_2$ and $C_2=\ell \!-\!X_1\!-\!X_2$ which is shrinkable. We thus find that ${\rm dP}_2 \cup {\rm dP}_2$ is the only endpoint geometry with $M>1$ \footnote{We thank Sung-Soo Kim for pointing out that this geometry has no rank-preserving mass deformation}. Therefore the endpoint classification reduces to a simple classification of two types of geometries, $\mathbb{P}^2 \cup \mathbb{F}_{n}$ for $M=0$ and $\mathbb{F}_{n_1}\cup \mathbb{F}_{n_2}$ for $M=1$, other than ${\rm dP}_2 \cup {\rm dP}_2$ with $M=3$.
We first classify geometries of the type $\mathbb{P}^2 \cup \mathbb{F}_{n}$. We can choose a curve class $C_{S_1}=C_1=a \ell$ in $\mathbb{P}^2$ with a positive integer $a$ and $C_{S_2}=C_2=E$ in $\mathbb{F}_n$ satisfying the gluing condition (\ref{eq:gluingcond}). Since $C$ should be rational, the integer in $C_1$ is fixed to be either $a=1$ or $a=2$. Accordingly, the second surface is fixed to be $\mathbb{F}_3$ or $\mathbb{F}_6$ respectively. Hence we find only two geometries of this type:
\begin{align}
\begin{split}
& \mathbb{P}^2 \cup \mathbb{F}_3 \quad {\rm with} \quad C_1 = \ell \ , \ C_2 = E_3 \ , \\
&\mathbb{P}^2 \cup \mathbb{F}_6 \quad {\rm with} \quad C_1 = 2\ell \ , \ C_2 = E_6 \ .
\end{split}
\end{align}
These two geometries have brane constructions as depicted in Fig \ref{fig:rank2-branes1}. These geometries have no mass parameter. Therefore we do not expect any gauge theory descriptions associated to these CFTs.
\begin{figure}
\begin{center}
\includegraphics[scale=.4]{rank2-branes-m0.pdf}
\end{center}
\caption{Brane configurations of rank 2 SCFTs with zero mass.}
\label{fig:rank2-branes1}
\end{figure}
The second type of endpoint geometry can be classified in the same manner. Due to the gluing condition (\ref{eq:gluingcond}), a gluing curve in one of two Hirzebruch surfaces should have negative self-intersection. We choose $C_2=E_2$ in the second surface $\mathbb{F}_{n_2}$. Then the gluing curve $C_1$ in the first surface $\mathbb{F}_{n_1}$ needs to be a rational irreducible curve with self-intersection $n_2-2$. The curve $C_1$ takes the form of $C_1 = aF_1+bH_1$ with $a,b\ge0$ or $C_1=E_1$, and must satisfy
\begin{equation}
C_1^2 = n_2-2 \ , \quad C_1 \cdot S_1 = -n_2 \ .
\end{equation}
We now need to check shrinkability conditions. In both irreducible components $S_i = \mathbb F_{n_i}$, the curve classes generating Mori cone are $E_i, F_i$. When these curve classes have non-negative volumes with respect to the K\"ahler class $-J=-\phi_1 S_1-\phi_2 S_2$, the local 3-fold defined by $S$ is shrinkable and thus engineers a 5d SCFT. In this case, the criteria for shrinkability are
\begin{eqnarray}
&&{\rm vol}(E_1) = (2-n_1)\phi_1-a \phi_2 \ge 0 \ , \quad {\rm vol}(F_1) = 2\phi_1-b\phi_2 \ge 0 \ , \nonumber \\
&&{\rm vol}(E_2) = (2a+2b-bn)\phi_1+(2-n)\phi_2 \ge 0 \ , \quad {\rm vol}(F_2) = -\phi_1+2\phi_2 \ge 0 \ ,
\end{eqnarray}
with $\phi_1,\phi_2>0$.
We can easily solve these conditions and the gluing condition (\ref{eq:gluingcond}). Each solution will give a shrinkable geometry and thus a SCFT. The full list of shrinkable surfaces $\mathbb{F}_{n_1}\cup \mathbb{F}_{n_2}$ (denoted by $(n_1,n_2)$) is given in Tables \ref{tb:endpoint-F-F} and \ref{tb:shirinkable-F-F}. Some of these geometries have brane constructions given in Figure \ref{fig:rank2-branes2}. We find that only the six geometries in Table \ref{tb:endpoint-F-F} are independent endpoint geometries.
\begin{table}
\centering
\begin{subtable}{.8\textwidth}
\centering
\begin{tabular}{|c|c|c|}
\hline
$S_1\cup S_2$ & $C_{S_1}$ & $C_{S_2}$\\
\hline
$\mathbb{P}^2\cup \mathbb{F}_3$ & $\ell$ & $E$ \\
\hline
$\mathbb{P}^2 \cup \mathbb{F}_6$ & $2\ell$ & $E$ \\
\hline
\end{tabular}
\caption{Endpoint geometries with $M=0$.}
\label{tb:endpoint-P-F}
\end{subtable}%
\vspace{.5cm}
\begin{subtable}{.9\textwidth}
\centering
\begin{tabular}{|c|c|c||c|c|c|}
\hline
$(n_1,n_2)$ & $C_{S_1}$ & $G$ & $(n_1,n_2)$ & $C_{S_1}$ & $G$\\
\hline
$(0,2)$ & $F$ & $SU(3)_1$ & $(0,8)$ & $F+3H$ & $SU(3)_7,G_2$ \\
\hline
$(0,4)$ & $F+H$ & $SU(3)_3$ & $(1,1)$& $E$ & $SU(3)_0$ \\
\hline
$(0,6)$ & $F+2H$ & $SU(3)_5,Sp(2)_{\pi}$ & $(1,7)$& $2F+H$ & $SU(3)_6$ \\
\hline
\end{tabular}
\caption{Endpoint geometries with $M=1$. Here $C_{S_2}=E$. These geometries have gauge theory descriptions with gauge group $G=SU(3)_k,Sp(2)_\theta,G_2$ where $k$ is the classical CS level and $\theta$ is the $\mathbb Z_2$-valued $\theta$ angle.}
\label{tb:endpoint-F-F}
\end{subtable}%
\vspace{.5cm}
\begin{subtable}{.8\textwidth}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$(n_1,n_2)$ & $C_{S_1}$ & $G$ & Endpoint \\
\hline
$(1,2)$ & $F$ & $SU(2)\hat{\times}SU(2)$ & $\mathbb{P}^2\cup \mathbb{F}_3$ \\
\hline
$(1,3)$& $H$ & $SU(3)_2$ & $\mathbb{P}^2\cup \mathbb{F}_3$\\
\hline
$(1,5)$& $F+H$ & $SU(3)_4$ & $\mathbb{P}^2\cup \mathbb{F}_6$\\
\hline
$(1,6)$ & $2H$ & $Sp(2)_{0}$ & $\mathbb{P}^2\cup \mathbb{F}_6$\\
\hline
\hline
$(2,4)$ & $H$ & $SU(3)_1$ & $\cdot$\\
\hline
$(0,10)$ & $F+4H$ & $SU(3)_9$ & $\cdot$ \\
\hline
\end{tabular}
\caption{Other geometries of $\mathbb{F}_{n_1}\cup \mathbb{F}_{n_2}$. The first four are not endpoints and flow to geometries in (a) by mass deformations. $(2,4)$ is an endpoint, but is also equivalent to $(0,4)$ by a HW transition. $(0,10)$ is an endpoint, but not shrinkable.}
\label{tb:shirinkable-F-F}
\end{subtable}
\caption{Classification of all rank 2 geometries with $M=0,1$.}\label{tb:rank2-F-F-clssification}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[scale=.35]{rank2-branes.pdf}
\end{center}
\caption{Brane configurations of rank 2 SCFTs with $M=1$.}
\label{fig:rank2-branes2}
\end{figure}
In fact, all the endpoint geometries in Table \ref{tb:endpoint-F-F} have gauge theory descriptions with simple gauge group $G$. As explained in Section \ref{sec:Mth}, a distinguished property of geometries corresponding to gauge theories is that the matrix of intersection numbers (\ref{eqn:Cartan}) of holomorphic fiber classes $f_i$ with the surfaces $S_i$ is equal to (minus) the Cartan matrix of the gauge algebra. We remark here that the Hirzebruch surface $\mathbb{F}_0$ has a base-fiber duality exchanging the base curve class $H$ and the fiber curve class $F$. Geometrically, this is an isomorphism between two geometries related by the exchange of $H$ and $F$. It is possible that the dual geometry often has different gauge theory realization from the gauge theory of the original geometry. In this case, the geometric duality leads to a duality between two different gauge theories.
Aside from studying the Cartan matrices, we can also compare the triple intersection polynomial $J^3$ to the perturbative expression for the prepotential given in (\ref{eqn:pre}). For the geometries in Table \ref{tb:endpoint-F-F} and \ref{tb:shirinkable-F-F}, the prepotentials are
\begin{equation}
6\mathcal{F} = J^3 = 8\phi_1^3 + 3\phi_1\phi_2(-n_2\phi_1+(n_2-1)\phi_2) + 8 \phi_2^3\ .
\end{equation}
We can compare these prepotentials against known gauge theory prepotentials as a means to identify the corresponding gauge theories.
Let us first select the respective fibers $H,F$ for $\mathbb{F}_{0}\cup \mathbb{F}_{n_2}$, and $F,F$ for $\mathbb{F}_{1}\cup \mathbb{F}_{n_2}$. The Cartan matrix $A_{ij}$ of the following geometries computed using these fiber classes is that of the gauge algebra $SU(3)$ as
\begin{equation}
(A_{SU(3)})_{ij} ~:~(n_1,n_2) ~=~ (0,2) \, , \ (0,4) \,, \ (0,6) \,, \ (0,8) \,, \ (1,1) \,, \ (1,7) \ ,
\end{equation}
for the choices of degrees $(n_1,n_2)$ of $\mathbb F_{n_1} \cup \mathbb F_{n_2}$.
Moreover, their triple intersections agree with gauge theory prepotentials of $SU(3)_k$ listed in Table \ref{tb:endpoint-F-F}. Therefore, we expect that these endpoint geometries have $SU(3)_k$ gauge theory realizations.
The geometries $(0,6)$ and $(0,8)$ are particularly interesting, as they have two different gauge theory descriptions related by the base-fiber exchange of $\mathbb{F}_0$.
When we consider the fibers classes to be $F,F$, the two geometries $(0,6),(0,8)$ exhibit (respectively) $Sp(2),G_2$ Cartan matrices. On the other hand, if we choose fiber classes $H,F$, the geometries exhibit the $SU(3)$ Cartan matrix in both cases.
Studying triple intersection numbers gives us a means to narrow down the precise gauge theory that corresponds to these geometries. The triple intersection polynomial $J^3$ of the geometry $(0,6)$ is identical to the prepotentials of both pure $SU(3)_5$ gauge theory and also pure $Sp(2)_\theta$ theory, which can have either $\theta=0$ or $\theta=\pi$. However, the prepotential cannot distinguish two $Sp(2)$ cases. We can instead determine the $\theta$ angle using the known duality between $SU(3)$ and $Sp(2)$. In \cite{Gaiotto:2015una}, it was conjectured that $SU(3)_5$ is dual to $Sp(2)_\pi$. This suggests that the geometry $(0,6)$ corresponds to $Sp(2)_\pi$ while $(1,6)$ corresponds to $Sp(2)_0$. Thus, the geometric construction provides yet additional evidence supporting the duality between the $SU(3)_5$ and $Sp(2)_{\pi}$ gauge theories.
As another example of a duality between gauge theories, the triple intersections of $(0,8)$ agree with the prepotentials of $SU(3)_7$ and $G_2$ gauge theories. We thus conjecture that $SU(3)_7$ and $G_2$ theories are dual and describe the low energy physics of the SCFT corresponding to $\mathbb{F}_0\cup \mathbb{F}_8$.
Additional (not necessarily endpoint) geometries of type $\mathbb{F}_{n_1}\cup \mathbb{F}_{n_2}$ are displayed in Table \ref{tb:shirinkable-F-F}. The first five geometries in Table \ref{tb:shirinkable-F-F} are shrinkable. However, the first four geometries of these are not endpoints. They all can be obtained from other endpoint geometries, $\mathbb{P}^2\cup \mathbb{F}_3$ or $\mathbb{P}^2\cup \mathbb{F}_6$, by blowing up a point and performing flop transitions; see Figure \ref{fig:P2-F-transition} for more details. We find that these geometries but $(1,2)$ have gauge theory descriptions as listed in Table \ref{tb:shirinkable-F-F}. The geometry $(1,2)$ has gauge algebra $SU(2)\hat{\times}SU(2)$ where $\hat{\times}$ denotes that we gauge the $SU(2)$ global symmetry of another $SU(2)$ gauge theory which arises from the $U(1)_I$ instanton symmetry in the IR gauge theory.
\begin{figure}
\centering
\includegraphics[scale=.45]{P2-F-transitions.pdf}
\caption{Geometric transitions from $\mathbb{P}^2\cup \mathbb{F}_3$ and $\mathbb{P}^2\cup \mathbb{F}_6$ to $\mathbb{F}_1\cup \mathbb{F}_n$'s with $n=2,3,5,6$.}
\label{fig:P2-F-transition}
\end{figure}
The geometry $(2,4)$ in Table \ref{tb:shirinkable-F-F} is an endpoint geometry admitting no additional rank preserving mass deformations. However, this geometry is equivalent to another endpoint geometry $(0,4)$ by a complex structure deformation, or a Hanany-Witten transition. Thus these two geometries belong to the same physical equivalence class.
Lastly, the geometry $(0,10)$ is not shrinkable. This geometry satisfies all other shrinkablity conditions, but we find that no 4-cycles have nonzero volume at any point in the K\"ahler cone. Thus $(0,10)$ is not shrinkable unless we make a non-normalizable K\"ahler deformation. This means the corresponding field theory possesses an intrinsic energy scale set by the K\"ahler parameter of the non-compact 4-cycle. Therefore, we do not expect that this geometry corresponds to a 5d SCFT. Indeed, in Section \ref{sec:rank2}, we will argue that this geometry gives a 5d KK theory.
We have finished the full classification of rank 2 endpoint geometries (thus rank 2 endpoint SCFTs), which have $M=0,1$. The result is rather surprising---we observe that all rank 2 SCFTs are actually realized by gauge theories and their mass deformations. Note that geometries $\mathbb{P}^2\cup\mathbb{F}_3$ and $\mathbb{P}^2\cup\mathbb{F}_6$ corresponding to non-Lagrangian theories can also viewed as deformations of geometries which admit gauge theory descriptions, for example (respectively) $\mathbb{F}_1\cup\mathbb{F}_2$ and $\mathbb{F}_1\cup\mathbb{F}_5$. This seems to suggest that gauge theory descriptions are generally quite useful, even for 5d SCFTs of higher rank.
Furthermore, all geometries in Table \ref{tb:rank2-F-F-clssification} except for $(1,2)$ were already predicted in \cite{Jefferson:2017ahm} using perturbative gauge theory analysis. In fact these geometric constructions confirm all predictions with $r=2$ and $M=1$ in \cite{Jefferson:2017ahm} except for $SU(3)_8$.
It was conjectured in \cite{Jefferson:2017ahm} that the $SU(3)_8$ theory exists and has an interacting UV fixed point. However, the existence of this theory appears to be ruled out by our geometric classification.
Let us briefly discuss the geometry of the $SU(3)_8$ gauge theory.
This theory in fact has a geometric realization as the local 3-fold with K\"ahler surface $\mathbb{F}_1\cup \mathbb{F}_9$, where we identify the 2-cycles $C_{S_1}=3F_1+H_1$ and $C_{S_2} = E_2$. However, this geometry is not shrinkable because at least one 2-cycle contained in $S$ has negative volume. For example, the volumes
\begin{equation}
\text{vol}(E_1) = \phi_1 - 3\phi_2 \ , \quad \text{vol}(F_2) = 2\phi_2 - \phi_1 \
\end{equation}
with $\phi_1,\phi_2>0$ cannot be both non-negative. Therefore the Coulomb branch of this geometry is trivial and this geometry is not shrinkable.
In order to make the geometry shrinkable we need to attach a non-compact 4-cycle with non-zero K\"ahler parameter corresponding to bare gauge coupling constant $1/g^2$. This K\"ahler parameter cannot be tuned to zero while maintaining positivity of the K\"ahler metric. So even though the IR gauge description with $1/g^2\not=0$ makes sense geometrically, we cannot take the $1/g^2=0$ limit without taking the Coulomb branch parameter to $0$. This means that if the point $1/g^2=0$ is a CFT point, then it has no Coulomb branch deformation, and thus in conflict with a SCFT from this gauge theory based on our assumptions. Thus we do not expect that this geometry has a CFT limit. The gauge theory analysis in \cite{Jefferson:2017ahm} uses only the perturbative spectrum and monopole tensions and thus cannot capture the spectrum of M2-branes wrapping the curve $E_1 \subset \mathbb F_1$ (which correspond to instantons in the gauge theory). Missing non-perturbative states such as these are crucial for assessing whether or not a geometry is shrinkable. This again shows that the perturbative constraints used in \cite{Jefferson:2017ahm} are necessary but not sufficient to guarantee the existence of CFT fixed points.
\subsubsection*{Full rank 2 classification}
We showed in the previous section that our classification program can be reduced to a classification of the following types of geometric configurations: $\text{Bl}_{p_1} \mathbb F_n \cup \text{dP}_{p_2}$ and $ \text{Bl}_{p_1} \mathbb F_n \cup \mathbb F_0$. As already discussed $p_2$ and $p_1$ are bounded above by $p_{\text{max}}(n)$, which we note depends upon both the degree $n$ and the type of gluing configuration. However, we are still faced with the problem of restricting the range of (non-negative) integer $n$ for which there exist shrinkable configurations. It turns out that some necessary conditions of shrinkability allows us to derive a crude bound on $n$. From a physical perspective, the existence of such a bound is not surprising as it is closely tied to the existence of only a finite number of 5d interacting fixed CFT points for a fixed rank.
Appropriate bounds on $n$ can be determined in the two separate cases of $S_2 = \text{dP}_{p_2}$ or $S_2= \mathbb F_0$. For both cases, we need only consider $n \geq 2$, since setting $n=0,1$ produces a geometric configuration isomorphic to $\text{dP}_{p_1+1} \cup \text{dP}_{p_2}$. In the case of $S_2 = \text{dP}_2$, we find that $n \leq 7$, while in the case of $S_2 = \mathbb F_0$, we find that $n \leq 8$. See Appendix \ref{app:bound} for proofs of these bounds.
We present our classification of rank 2 K\"ahler surfaces associated to 5d UV interacting fixed points in Figures \ref{fig:11}-\ref{fig:0}. These results are organized by the number of mass parameters $M$, with $0 \leq M \leq 11$. Given $M >0$ mass parameters, a shrinkable geometry with $M-1$ mass parameters may be obtained by performing a blowdown of an exceptional divisor (possibly after a sequence of flops) in the surface $S$; in the associated field theory, blowing down an exceptional curve corresponds to integrating out a massive matter hypermultiplet.
In each figure, we list the K\"ahler surface $S= S_1 \overset{C_{S_2}}{\cup} S_2$, where $C_{S_2}=( S_1 \cap S_2)_{S_2}$ is the curve along which the two surfaces are glued, restricted to the \emph{second} surface $S_2$. Geometries marked with $( \cdot )^*$ correspond to 5d KK theories. Beneath each geometry, we also list the associated gauge theory; geometries with no associated gauge system indicated do not admit a known description as a gauge theory.
Our method for identifying gauge theoretic descriptions involves comparing the triple intersection $J^3$ with the gauge-theoretic prepotential $6\mathcal F$ in (\ref{eqn:pre}) for given gauge group and matter content in the K\"ahler cone, as well as identifying a geometric realization of the Cartan matrix of associated to the gauge algebra.
The Cartan matrices are determined up to sign by a choice of fibers\footnote{In the present discussion, a \emph{fiber} is a rational curve $f$ with self intersection $f^2= 0$.} $f_1\subset S_1, f_2 \subset S_2$ satisfying
\begin{align}
(f_i \cdot S_j)_{S_i} = - (A_G)_{ij}.
\end{align}
Geometrically, these fibers are rational curves over which M2-branes may be wrapped to give rise to charged BPS vectors in the 5d spectrum. In Figures \ref{fig:11}-\ref{fig:0}, we indicate to the left of each gauge description a possible choice of fibers giving rise to stated gauge algebra. We merely list all possible gauge theory descriptions and do not attempt to list all possible configurations of fibers. When there is more than one choice of fiber leading to different Cartan matrices (and hence different gauge symmetries), there are dualities between the associated gauge theory descriptions. For $\text{dP}_{p_2 <8}$, the possible fibers are (using the same notation as in \ref{eq:moridp})
\begin{align}
(1;1)~,~(2;1^4)~,~(3;2,1^6)~,~(4;2^3,1^4)~,~(5;2^6,1).
\end{align}
The list of possible fibers in $\text{Bl}_{p_1} \mathbb F_n$ is significantly more complicated; see Appendix \ref{app:fiber}.
We also note that the double arrows connecting pairs of different geometries $S$ indicate flop transitions mapping the geometries into one another. Each figure contains several clusters of geometries connected by arrows, with each cluster belonging to the same birational, and thus physical, equivalence class. Arrows decorated with the symbol $\phi_1 \leftrightarrow \phi_2$ indicate that the flop transition requires us to reverse our identifications $S_1 \leftrightarrow S_2$, and flip the sign of the Chern-Simons level, $k \to - k$.
Finally, we remark that the gluing curves $C_{S_2} \in \text{dP}_{p_2 \geq 3}$ are only listed up to the action of the Weyl group $W(E_{p_2})$. Said differently, each choice of gluing curve displayed in the figures is a single element in the Weyl orbit. We now briefly describe the Weyl group action in $\text{dP}_{p_2}$ and explain why in most cases we only need to distinguish geometric configurations whose gluing curves belong to the same Weyl orbit in a given surface. Given a simple root $\alpha_i = X_i - X_{i+1}, i = 1, \dots, p_2-1$, and an effective curve
\begin{align}
C= d \ell - m_i X_i,
\end{align}
the Weyl reflections $w_{\alpha_i}$ act by transposing exceptional divisors, $X_i \leftrightarrow X_{i+1}$, while the reflection $w_{\alpha_{p_2}}$ associated to the root $\alpha_{p_2} = \ell - \sum_{i=1}^3 X_i$ acts on $C$ as follows:
\begin{align}
\begin{split}
w_{\alpha_{p_2}}(C) &= (2 d - m_1 - m_2 -m_3) \ell - ( d - m_2 - m_3 ) X_1 - ( d - m_1 - m_3) X_2 \\
&- (d - m_1 - m_2) X_3 - \sum_{i > 3} m_i X_i.
\end{split}
\end{align}
As was shown in \cite{Iqbal:2001ye}, the action of $W(E_{p_2})$ on a rational curve $C \in \text{dP}_{p_2}$ for $p_2 \geq 4$ and degree $d_C \equiv - K \cdot C = C^2 + 2 = n$ in all cases studied in this paper is transitive. Therefore, since the Weyl action $w_{\alpha}: C \mapsto C + (C \cdot \alpha) C$ preserves intersection products,
\begin{align}
C \cdot C' = ( C + (C \cdot \alpha) \alpha) \cdot ( C' + (C'\cdot \alpha) \alpha),
\end{align}
it is sufficient to set the gluing curve $C_{S_2}$ equal to a single element of the Weyl orbit in order to understand the full intersection structure, as the intersection numbers are identical up to permutation for any two elements belonging to the same Weyl orbit. For $p_2 <3$, the Weyl group either has multiple orbits (as in the case of $p_2 =3$) or is otherwise undefined (as in the case of $p_2 <3$), and so for $p_2<4$ we only list gluing curves $C_{S_2}$ up to cyclic permutations of the exceptional divisors $X_i$.
Upon mass deforming these SCFTs and flowing to the IR we get a tree of relations between these conformal theories which is summarized in the RG flow tree diagram in Figure \ref{tree}. The top theories of the RG families are related to 5d KK theories which are discussed in the next section.
\begin{figure}
\begin{center}
\noindent\makebox[\textwidth]{
$
\begin{array}{c}
\begin{tikzpicture}[]
\node[yscale=1.2,xscale=1.1] at (0,0) {\includegraphics[scale=.5]{rank2.pdf}};
\end{tikzpicture}
\end{array}
$
}
\end{center}
\caption[]{The diagram above shows the RG flow among rank 1 and rank 2 SCFTs obtained by mass deformations. The first and the second rows in each box correspond to the geometric and the gauge theoretic descriptions respectively of a 5d theory \footnotemark. The parent theory in each branch is a 5d KK theory related to a 6d theory on $S^1$.}
\label{tree}
\end{figure}
\footnotetext{\label{foot:GZ}
We note that while $\text{Bl}_8\mathbb{F}_3\cup \mathbb{P}^2$ has no gauge theory description, it is nonetheless related to $[SU(2)+5{\bf F}]\times SU(2)_0$ by a flop transition: a flop of $\text{Bl}_8\mathbb{F}_3\cup \mathbb{P}^2$ leads to the geometry $\text{Bl}_7 \mathbb{F}_2 \overset{\ell-X_1}{\cup} \text{dP}_1$, which has gauge theory description $[SU(2)+5{\bf F}]\times SU(2)_0$. However, $\text{Bl}_7 \mathbb{F}_2 \overset{\ell-X_1}{\cup} \text{dP}_1$ is not shrinkable, which implies that the BPS spectrum of the gauge theory will develop a negative mass before reaching a CFT fixed point. Nevertheless, this gauge theory theory makes sense as an effective description of the CFT from $\text{Bl}_8\mathbb{F}_3\cup \mathbb{P}^2$ through a flop transition to $\text{Bl}_7 \mathbb{F}_2 \overset{\ell-X_1}{\cup} \text{dP}_1$ when mass parameters are turned on. We are greatful to Gabi Zafrir for pointing out that the CFT related to the $[SU(2)+5{\bf F}]\times SU(2)_0$ gauge theory should exist since an associated $(p,q)$ 5-brane system exists.}
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{tikzpicture}[yscale=1.4]
\node[](a) at (-4,2) {$ \begin{array}{c} (\text{Bl}_{10} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1)^* \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 10 \textbf{F} \\\hline H + 2 F - \sum X_i , \ell - X_1 &\hat A_1\\\hline \end{array}$}\end{array}$};
\node[](b1) at (0,0) {$\begin{array}{c} (\text{Bl}_{9} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2)^* \\\scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_1 & SU(3)_0 + 10 \textbf{F} \\\hline F, \ell - X_2 & Sp(2) + 10\textbf{F} \\\hline H + 2 F - \sum X_i , \ell - X_2& \hat A_1 \\\hline \end{array} $}\end{array}$};
\node[](b2) at (4,2) {$\begin{array}{c} (\text{Bl}_{10} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0)^* \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,F & Sp(2) + 10\textbf{F} \\\hline F, E & SU(3)_0 + 10 \textbf{F} \\\hline H + 2 F - \sum X_i ,F& \hat A_1 \\\hline \end{array}$} \end{array}$};
\node[](c) at (0,-2.5) {$\begin{array}{c} (\text{Bl}_{8} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3)^* \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_0 + 10 \textbf{F} \\\hline F,\ell - X_3 & Sp(2) + 10\textbf{F}\\\hline H + 2 F - \sum X_i, \ell - X_3 & \hat A_1 \\\hline \end{array}$} \end{array}$};
\node[](d) at (0,-5) {$\begin{array}{c} (\text{Bl}_{7} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4)^*\\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_0 + 10 \textbf{F} \\\hline F,\ell - X_4 & Sp(2) + 10\textbf{F}\\\hline H + 2 F - \sum X_i, \ell -X_4 & \hat A_1 \\\hline \end{array}$} \end{array}$};
\node[](e) at (0,-7.5) {$\begin{array}{c} (\text{Bl}_{6} \mathbb F_2 \overset{2 \ell-\sum_{i=1}^4 X_i}{\cup} \text{dP}_5)^* \\\scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_0 + 10 \textbf{F} \\\hline F,\ell - X_5 & Sp(2) + 10\textbf{F}\\\hline H- X_1 - X_2, 2 \ell-\sum_{i=1}^4 X_i & [SU(2) + 4 \textbf{F}] \times [SU(2) + 4 \textbf{F}] \\\hline H + 2 F - \sum X_i,\ell-X_5 & \hat A_1 \\\hline \end{array} $} \end{array}$};
\node[](f) at (0,-10) {$\begin{array}{c} (\text{Bl}_{5} \mathbb F_1 \overset{2 \ell-\sum_{i=1}^5 X_i }{\cup} \text{dP}_6)^*\\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_0 + 10 \textbf{F} \\\hline F, \ell - X_6 & Sp(2) + 10\textbf{F} \\\hline f_1 \cdot E = 0, 2\ell - \sum_{i=2}^5 X_i & [SU(2) + 4 \textbf{F}] \times [SU(2) + 4 \textbf{F}] \\\hline f_1 \cdot E= 2, \ell- X_6 & \hat A_1 \\\hline \end{array} $}\end{array}$};
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={yshift=-.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
$
\end{center}
\caption{$M=11$ geometries.}
\label{fig:11}
\end{figure}
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1.3]
\node[](a) at (7.5,10) {$ \begin{array}{c} \text{Bl}_{9} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 9 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](b1) at (5.5,8) {$\begin{array}{c} \text{Bl}_{8} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_1 & SU(3)_{\frac{1}{2}} + 9 \textbf{F} \\\hline F, \ell - X_2 & Sp(2) + 9\textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](b2) at (3.5,10) {$\begin{array}{c} \text{Bl}_{9} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,F & Sp(2) + 9\textbf{F} \\\hline F, E & SU(3)_{\frac{1}{2}} + 9 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](c) at (5.5,6) {$\begin{array}{c} \text{Bl}_{7} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_{\frac{1}{2}} + 9 \textbf{F} \\\hline F, \ell - X_3 & Sp(2) + 9\textbf{F}\\\hline \end{array} $}\end{array}$};
\node[](d) at (5.5,4) {$\begin{array}{c} \text{Bl}_{6} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4\\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{1}{2}} + 9 \textbf{F} \\\hline F, \ell - X_4 & Sp(2) + 9\textbf{F}\\\hline \end{array}$} \end{array}$};
\node[](e) at (5.5,1.5) {$\begin{array}{c} \text{Bl}_{5} \mathbb F_2 \overset{2 \ell-\sum_{i=1}^4 X_i}{\cup} \text{dP}_5 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_{\frac{1}{2}} + 9 \textbf{F} \\\hline F,\ell - X_5 & Sp(2) + 9\textbf{F}\\\hline H - X_1-X_2, 2 \ell- \sum_{i=1}^4 X_i & [SU(2) + 3 \textbf{F} ] \times [SU(2) + 4 \textbf{F}] \\\hline \end{array} $}\end{array}$};
\node[](f) at (-2,1.5) {$\begin{array}{c} \text{Bl}_{4} \mathbb F_1 \overset{2 \ell-\sum_{i=1}^5 X_i }{\cup} \text{dP}_6\\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F, \ell- X_5 &SU(3)_{\frac{1}{2}} + 9 \textbf{F} \\\hline F, \ell - X_6 & Sp(2) + 9\textbf{F}\\\hline f_1 \cdot E = 0, \ell-X_6 & [SU(2) + 3 \textbf{F}] \times [ SU(2) + 4 \textbf{F}]\\\hline \end{array}$} \end{array}$};
\node[](a3) at (-2,11.5) {$ \begin{array}{c} \text{Bl}_{10} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[] at (5.5,12) {$ \begin{array}{c} (\text{Bl}_{9} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0)^* \\\scalebox{.7}{$ \begin{array}{|c|c|} \hline F,E& SU(3)_{-\frac{3}{2}} + 9 \textbf{F} \\\hline H + 2 F - \sum_{i=1}^8 X_i ,E & Sp(2) + 8 \textbf{F}+ 1\textbf{AS} \\\hline \end{array}$} \end{array} $};
\node[](b3) at (-2,10) {$ \begin{array}{c} \text{Bl}_{9} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_1 & SU(3)_{-\frac{1}{2}}+ 9 \textbf{F} \\\hline H + 2 F- \sum X_i , \ell- X_1 & Sp(2) + 9 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](c3) at (-2,8) {$ \begin{array}{c} \text{Bl}_{8} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_{-\frac{1}{2}}+ 9 \textbf{F} \\\hline H + 2 F- \sum X_i , \ell- X_1 & Sp(2) + 9 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](d3) at (-2,6) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{-\frac{1}{2}}+ 9 \textbf{F} \\\hline H + 2 F - \sum X_i , \ell - X_{1} & Sp(2) + 9 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](e3) at (-2,4) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_2 \overset{2 \ell - \sum_{i=1}^4 X_i}{\cup} \text{dP}_4 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{-\frac{1}{2}}+ 9 \textbf{F} \\\hline H + 2 F - \sum X_i, \ell - X_{1} & Sp(2) + 9 \textbf{F} \\\hline H - X_1 -X_2 , 2 \ell - \sum X_i & [SU(2) + 4 \textbf{F}] \times [ SU(2) + 3 \textbf{F}] \\\hline \end{array}$}\end{array}$};
\draw[big arrow] (a3) -- (b3);
\draw[big arrow] (e3) -- (f);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- node[right,midway]{$\phi_1 \leftrightarrow \phi_2$} (e3);
\draw[big arrow] (b3) -- (c3);
\draw[big arrow] (c3) -- (d3);
\draw[big arrow] (d3) -- (e3);
\draw[big arrow,transform canvas={xshift=.5em}] (b3) -- (a3);
\draw[big arrow,transform canvas={xshift=.5em}] (c3) -- (b3);
\draw[big arrow,transform canvas={xshift=.5em}] (d3) -- (c3);
\draw[big arrow,transform canvas={xshift=.5em}] (e3) -- (d3);
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={yshift=-.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=10$ geometries.}
\label{fig:10}
\end{figure}
\begin{figure}
\begin{center}
\noindent\makebox[\textwidth]{ $
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\node[](a) at (9,-4) {$ \begin{array}{c} \text{Bl}_{9} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (5,-4) {$ \begin{array}{c} \text{Bl}_{8} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{0}+ 8 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](c) at (0,-4) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{0}+ 8 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](d) at (0,-6) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{0}+ 8 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](e) at (6,-6) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_2 \overset{2 \ell - \sum_{i=1}^4 X_i}{\cup} \text{dP}_4 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{0}+ 8 \textbf{F} \\\hline H- X_1 - X_2 , 2\ell - \sum X_i & [SU(2) + 3 \textbf{F}] \times [SU(2) + 3 \textbf{F}]\\\hline \end{array}$}\end{array}$};
\node[](f) at (6,-8.5) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_1 \overset{2 \ell - \sum_{i=1}^5 X_i}{\cup} \text{dP}_5 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell-X_1 & SU(3)_0 + 8 \textbf{F} \\\hline f_1 \cdot E = 0, 2l- \sum_{i=1}^4 X_i & [SU(2) + 3 \textbf{F} ] \times [SU(2) + 3 \textbf{F}]\\\hline \end{array}$}\end{array}$};
\node[] at (4,-1.5) {$ \begin{array}{c} \text{Bl}_{8} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{-2} + 8 \textbf{F} \\\hline H + 2 F - \sum_{i=1}^7 X_i , \ell - X_1 & Sp(2) + 7 \textbf{F} + 1 \textbf{AS} \\\hline \end{array}$}\end{array} $};
\draw[big arrow] (a) -- (b);
\draw[big arrow] (b) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={yshift=-.5em}] (e) -- (d);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}[yscale=1.2]
\node[](a) at (5,4.5) {$ \begin{array}{c} \text{Bl}_{8} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_1 & Sp(2) + 8 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](b1) at (5,2) {$\begin{array}{c} \text{Bl}_{7} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_1 & SU(3)_{1} + 8 \textbf{F} \\\hline F,\ell - X_2 & Sp(2) + 8 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](b2) at (5,-.5) {$\begin{array}{c} \text{Bl}_{8} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,F & Sp(2) + 8 \textbf{F} \\\hline F,E & SU(3)_{1} + 8 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](c) at (.5,2) {$\begin{array}{c} \text{Bl}_{6} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{1} + 8 \textbf{F} \\\hline F, \ell - X_3 & Sp(2) + 8 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](d) at (.5,-.5) {$\begin{array}{c} \text{Bl}_{5} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4\\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{1} + 8 \textbf{F} \\\hline F,\ell - X_4 & Sp(2) + 8 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](e) at (-6.5,-.5) {$\begin{array}{c} \text{Bl}_{4} \mathbb F_2 \overset{2 \ell-\sum_{i=1}^4 X_i}{\cup} \text{dP}_5 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{1} + 8 \textbf{F} \\\hline F,\ell - X_5 & Sp(2) + 8 \textbf{F} \\\hline H- X_1 - X_2 , 2\ell - \sum_{i=1}^4 X_i & [SU(2) + 2 \textbf{F} ] \times [ SU(2) + 4 \textbf{F}] \\\hline \end{array}$} \end{array}$};
\node[](f) at (-6.5,2) {$\begin{array}{c} \text{Bl}_{3} \mathbb F_1 \overset{2 \ell-\sum_{i=1}^5 X_i }{\cup} \text{dP}_6\\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{1} + 8 \textbf{F} \\\hline F, \ell - X_6 & Sp(2) + 8 \textbf{F} \\\hline f_1 \cdot E = 0, 2\ell - \sum_{i=1}^4 X_i & [SU(2)+ 2 \textbf{F} ] \times [SU(2) + 4 \textbf{F}]\\\hline \end{array}$} \end{array}$};
\node(z) at (.5,6.5) {$ \begin{array}{c} \text{Bl}_{8} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F, F & SU(3)_{-1} + 8 \textbf{F} \\\hline H + 2 F - \sum X_i , F & Sp(2) + 8 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node(y) at (.5,4.5) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{-1} + 8 \textbf{F} \\\hline H + 2 F - \sum X_i , \ell- X_1 & Sp(2) + 8 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node(x) at (-6.5,4.5) {$\begin{array}{c} \text{Bl}_{6} \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_3 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{-1} + 8 \textbf{F} \\\hline H + 2 F - \sum X_i , \ell - X_{2} & Sp(2) + 8 \textbf{F} \\\hline H-X_1 -X_2 , \ell-X_1 & [SU(2) +4 \textbf{F} ] \times [ SU(2) + 2 \textbf{F}] \\\hline \end{array} $}\end{array}$};
\draw[big arrow,transform canvas={xshift=.7em}] (f) --node[right,midway]{$\phi_1 \leftrightarrow \phi_2$} (x);
\draw[big arrow] (x)-- (f);
\draw[big arrow] (y) -- (x);
\draw[big arrow,transform canvas={yshift=-.5em}] (x) -- (y);
\draw[big arrow] (y) -- (x);
\draw[big arrow,transform canvas={yshift=-.5em}] (x) -- (y);
\draw[big arrow] (y) -- (x);
\draw[big arrow,transform canvas={xshift=.5em}] (y) -- (z);
\draw[big arrow] (z) -- (y);
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={xshift=.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={yshift=-.5em}] (e) -- (d);
\draw[big arrow,transform canvas={yshift=-.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\begin{array}{c}
\begin{tikzpicture}
\end{tikzpicture}
\end{array}
\end{array}
$}
\end{center}
\caption{$M=9$ geometries.}
\label{fig:9}
\end{figure}
\begin{figure}
\begin{center}
\noindent\makebox[\textwidth]{ $
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1.4]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \scalebox{.65}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 7 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](b1) at (0,-2) {$\begin{array}{c} \text{Bl}_{6} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\\scalebox{.65}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{3}{2}} + 7 \textbf{F} \\\hline F, \ell - X_2 & Sp(2) + 7 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](b2) at (-5,-2) {$\begin{array}{c} \text{Bl}_{7} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\\scalebox{.65}{$ \begin{array}{|c|c|} \hline
F,F & Sp(2) + 7 \textbf{F} \\\hline F, E & SU(3)_{\frac{3}{2}} + 7 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](c) at (5.5,-2) {$\begin{array}{c} \text{Bl}_{5} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \scalebox{.65}{$ \begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_{\frac{3}{2}} + 7 \textbf{F} \\\hline F,\ell - X_3 & Sp(2) + 7 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](d) at (5.5,0) {$\begin{array}{c} \text{Bl}_{4} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4\\ \scalebox{.65}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{3}{2}} + 7 \textbf{F} \\\hline F,\ell - X_4 & Sp(2) + 7 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](e) at (5.5,2) {$\begin{array}{c} \text{Bl}_{3} \mathbb F_2 \overset{2 \ell-\sum_{i=1}^4 X_i}{\cup} \text{dP}_5 \\ \scalebox{.65}{$\begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_{\frac{3}{2}} + 7 \textbf{F} \\\hline F, \ell - X_5 & Sp(2) + 7 \textbf{F} \\\hline H - X_1 - X_2 ,2\ell - \sum_{i=1}^4 X_i & [SU(2) + 1\textbf{F}] \times [ SU(2) + 4 \textbf{F}]\\\hline \end{array}$} \end{array}$};
\node[](f) at (0,4) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_1 \overset{2 \ell-\sum_{i=1}^5 X_i }{\cup} \text{dP}_6\\ \scalebox{.65}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{3}{2}} + 7 \textbf{F} \\\hline F, \ell - X_6 & Sp(2) + 7 \textbf{F} \\\hline f_1 \cdot E = 0, 2l - \sum_{i=1}^4 X_i & [SU(2) + 1 \textbf{F}] \times [ SU(2) + 4 \textbf{F}] \\\hline \end{array}$} \end{array}$};
\node[](a4) at (-5,0) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\ \scalebox{.65}{$\begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{-\frac{3}{2}} + 7 \textbf{F} \\\hline H + 2 F - \sum X_i , \ell - X_1 & Sp(2) + 7\textbf{F}\\\hline \end{array}$}\end{array} $};
\node[](a1) at (-5,2) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_2 \overset{ \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{.65}{$\begin{array}{|c|c|}\hline F,\ell - X_2 & SU(3)_{-\frac{3}{2}} + 7 \textbf{F} \\\hline H+ 2F - \sum X_i , \ell-X_2 & Sp(2) + 7 \textbf{F} \\\hline H -X_1 - X_2 , \ell- X_1 & [SU(2) + 4\textbf{F}]\times [SU(2) + 1 \textbf{F}] \\\hline \end{array}$}\end{array} $};
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow] (a1)-- node[left,midway]{$\phi_1 \leftrightarrow \phi_2$}(f);
\draw[big arrow] (a1)-- (a4);
\draw[big arrow,transform canvas={xshift=.5em}] (a4) -- (a1);
\draw[big arrow,transform canvas={yshift=-.5em}] (f) -- (a1);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={yshift=-.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\node[](a) at (-2,0) {$ \begin{array}{c} \text{Bl}_{8} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (-2,-2) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \scalebox{.65}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{1}{2}}+ 7 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](c) at (-2,-4) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \scalebox{.65}{$ \begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_{\frac{1}{2}}+ 7 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](d) at (-2,-6) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\\scalebox{.65}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{1}{2}}+ 7 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](e) at (-2,-8) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_2 \overset{2 \ell - \sum_{i=1}^4 X_i}{\cup} \text{dP}_4 \\ \scalebox{.65}{$ \begin{array}{|c|c|} \hline
F, \ell - X_{1} &SU(3)_{\frac{1}{2}}+ 7 \textbf{F} \\\hline H-X_1 -X_2 , 2\ell - \sum X_i & [SU(2) + 2 \textbf{F}] \times [SU(2) + 3 \textbf{F}] \\\hline \end{array}$}\end{array}$};
\node[](f) at (7,-8) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_1 \overset{2 \ell - \sum_{i=1}^5 X_i}{\cup} \text{dP}_5 \\ \scalebox{.65}{$ \begin{array}{|c|c|} \hline
F,\ell- X_1 & SU(3)_{\frac{1}{2}} + 7 \textbf{F} \\\hline f_1 \cdot E = 0, 2\ell - \sum_{i=1}^4 X_i & [SU(2) + 2 \textbf{F}] \times [ SU(2) + 3 \textbf{F}] \\\hline \end{array}$}\end{array}$};
\node[](a2) at (2.5,0) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_2 \overset{ E}{\cup} \mathbb F_0 \\\scalebox{.65}{$\begin{array}{|c|c|}\hline F,F & SU(3)_{-\frac{5}{2}} + 7 \textbf{F} \\\hline H-X_1 -X_2 , E & [SU(2) + 5 \textbf{F}] \times SU(2)_\pi \\\hline H + 2F - \sum_{i=1}^6 X_i , F & Sp(2) + 6 \textbf{F} + 1 \textbf{AS} \\\hline \end{array}$}\end{array} $};
\node at (2.5,-3) {$ \begin{array}{c} \text{Bl}_{8} \mathbb F_3 \overset{ \ell}{\cup} \mathbb P^2 \end{array} $};
\node (b3) at (7,-1) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \scalebox{.65}{$\begin{array}{|c|c|}\hline F,F & SU(3)_{-\frac{1}{2}} + 7 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (c3) at (7,-3) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\ \scalebox{.65}{$ \begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{-\frac{1}{2}} + 7 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (d3) at (7,-5) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_3 \\ \scalebox{.65}{$\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{-\frac{1}{2}} + 7 \textbf{F} \\\hline H-X_1 - X_2 ,\ell- X_1 & [SU(2) + 3 \textbf{F} ] \times [SU(2) + 2 \textbf{F}] \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (b3) -- (c3);
\draw[big arrow] (c3) -- (b3);
\draw[big arrow,transform canvas={xshift=.5em}] (c3) -- (d3);
\draw[big arrow] (d3) -- (c3);
\draw[big arrow] (a) -- (b);
\draw[big arrow] (b) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (f) -- node[left,midway]{$\phi_1 \leftrightarrow \phi_2$} (d3);
\draw[big arrow,transform canvas={xshift=.5em}] (d3) -- (f);
\draw[big arrow,transform canvas={xshift=.5em}] (b) -- (a);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={yshift=-.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\begin{array}{c}
\begin{tikzpicture}
\end{tikzpicture}
\end{array}
\end{array}
$}
\end{center}
\caption{$M=8$ geometries. (See Footnote \ref{foot:GZ} for a comment about $\text{Bl}_{8} \mathbb F_3 \cup \mathbb P^2$.)}
\label{fig:8}
\end{figure}
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1.3]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 6 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](b1) at (0,-2) {$\begin{array}{c} \text{Bl}_{5} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{2} + 6 \textbf{F} \\\hline F, \ell - X_2 & Sp(2) + 6 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](b2) at (-5,-2) {$\begin{array}{c} \text{Bl}_{6} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,F & Sp(2) + 6 \textbf{F} \\\hline F, E & SU(3)_{2} + 6 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](c) at (5.5,-2) {$\begin{array}{c} \text{Bl}_{4} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{2} + 6 \textbf{F} \\\hline F,\ell - X_3 & Sp(2) + 6 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](d) at (5.5,0) {$\begin{array}{c} \text{Bl}_{3} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4\\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{2} + 6 \textbf{F} \\\hline F, \ell - X_4 & Sp(2) + 6 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](e) at (5.5,2) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_2 \overset{2 \ell-\sum_{i=1}^4 X_i}{\cup} \text{dP}_5 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{2} + 6 \textbf{F} \\\hline F,\ell - X_5 & Sp(2) + 6 \textbf{F} \\\hline H- X_1 - X_2 , 2 \ell - \sum_{i=1}^4 X_i & SU(2)_\pi \times [SU(2) + 4 \textbf{F}] \\\hline \end{array} $}\end{array}$};
\node[](f) at (5.5,4.5) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_1 \overset{2 \ell-\sum_{i=1}^5 X_i }{\cup} \text{dP}_6\\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F, \ell - X_{1} &SU(3)_{2} + 6 \textbf{F} \\\hline F, \ell - X_6 & Sp(2) + 6 \textbf{F} \\\hline f_1 \cdot E = 0 , 2\ell - \sum_{i=1}^4 X_i & SU(2)_\pi \times [SU(2) + 4 \textbf{F} ] \\\hline \end{array} $}\end{array}$};
\node(g) at (-2,4.5) {$\begin{array}{c} \text{Bl}_6 \mathbb F_2 \overset{E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F, F & SU(3)_{-2} + 6 \textbf{F} \\\hline H+2F - \sum X_i ,F & Sp(2) + 6 \textbf{F} \\\hline H-X_1-X_2, E & [SU(2) + 4 \textbf{F}] \times SU(2)_{\pi} \\\hline \end{array}$} \end{array}$};
\node(A) at (7.5,6) {$\begin{array}{c} \text{Bl}_7 \mathbb F_3 \overset{\ell}{\cup} \mathbb P^2\end{array}$};
\node(B) at (3,6) {$\begin{array}{c} \text{Bl}_6 \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline H-X_1-X_2 , \ell - X_1 & [SU(2) + 4 \textbf{F} ] \times SU(2)_0 \\\hline \end{array}$} \end{array}$};
\node(C) at (-3,6) {$ \begin{array}{c} \text{Bl}_5 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_2 \\\scalebox{.7}{$\begin{array}{|c|c|}\hline F,\ell-X_{1} & [SU(2) + 4 \textbf{F}]\times SU(2)_0 \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={yshift=-.5em}] (A) -- (B);
\draw[big arrow] (B) -- (A);
\draw[big arrow](C) -- (B);
\draw[big arrow,transform canvas={yshift=-.5em}](B) -- (C);
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow] (f) -- node[above,pos=.5]{$\phi_1 \leftrightarrow \phi_2$} (g);
\draw[big arrow,transform canvas={yshift=-.5em}] (g) -- (f);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\end{array}
\begin{array}{c}
\begin{tikzpicture}
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=7$ geometries.}
\label{fig:7a}
\end{figure}
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}
\node (e) at (2.5,-1) {$\begin{array}{c} (\mathbb F_2 \overset{\ell - X_1}{\cup} \text{dP}_7 )^*\\ \scalebox{.7}{$\begin{array}{|c|c|} \hline F, \ell- X_2 & SU(3)_4 + 6 \textbf{F}\\\hline F, 2\ell- \sum_{i=2}^5 X_i & Sp(2) + 4 \textbf{F} + 2 \textbf{AS} \\\hline F,4\ell - \sum_{i=1}^4 X_i - 2 \sum_{j=5}^7 X_j& G_2 + 6 \textbf{F} \\\hline F,5\ell - X_1 - 2 \sum_{i=2}^7 X_i &A^{(2)}_2 \\\hline \end{array}$} \end{array}$};
\end{tikzpicture}
\end{array}
\\\\
\begin{array}{c}
\begin{tikzpicture}[yscale=1,xscale=1.2]
\node[](a) at (3.2,0) {$ \begin{array}{c} \text{Bl}_{7} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (0,0) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{1}+ 6 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](c) at (0,-2) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} & SU(3)_{1}+ 6 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](d) at (0,-4) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{1}+ 6 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](e) at (0,-6) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_2 \overset{2 \ell - \sum_{i=1}^4 X_i}{\cup} \text{dP}_4 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{1}+ 6 \textbf{F} \\\hline H-X_1-X_2 , 2\ell - \sum X_i & [SU(2)+ 1 \textbf{F}] \times [SU(2) + 3 \textbf{F}] \\\hline \end{array}$}\end{array}$};
\node[](f) at (0,-8.5) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_1 \overset{2 \ell - \sum_{i=1}^5 X_i}{\cup} \text{dP}_5 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell-X_1 & SU(3)_{1} + 6 \textbf{F} \\\hline f_1 \cdot E = 0, 2 \ell - \sum_{i=1}^4 X_i & [SU(2) + 1 \textbf{F} ] \times [ SU(2)\times 3 \textbf{F}] \\\hline \end{array}$}\end{array}$};
\node[](a3) at (0,-11) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_2 \overset{ \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F,\ell - X_2 & SU(3)_{-1} + 6 \textbf{F} \\\hline H-X_1-X_2, \ell-X_1 & [SU(2) + 3 \textbf{F}] \times [ SU(2)+ 1 \textbf{F}] \\\hline \end{array}$}\end{array} $};
\node[](a4) at (0,-13) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\\scalebox{.7}{$\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{-1} + 6 \textbf{F} \\\hline \end{array}$}\end{array} $};
\node[] at (7.5,0) {$ \begin{array}{c} \mathbb F_1 \overset{X_1}{\cup} \text{dP}_7 \\\scalebox{.7}{$\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{3} + 6 \textbf{F} \\\hline F,\ell-X_2 & Sp(2) + 5 \textbf{F} + 1 \textbf{AS} \\\hline \end{array}$}\end{array} $};
\node (b3) at (7.5,-3) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F,E & SU(3)_{0} + 6 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (c3) at (7.5,-5) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\\scalebox{.7}{$ \begin{array}{|c|c|}\hline F, \ell - X_{1} & SU(3)_{0} + 6 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (d3) at (7.5,-7) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_2 \overset{\ell- X_1}{\cup} \text{dP}_3 \\ \scalebox{.7}{$ \begin{array}{|c|c|}\hline F,\ell - X_{2} & SU(3)_{0} + 6 \textbf{F} \\\hline H-X_1-X_2 , \ell -X_1 & [SU(2) + 2 \textbf{F} ] \times [SU(2) \times 2\textbf{F}] \\\hline\end{array} $}\end{array}$};
\node(e3) at (7.5,-9) {$\begin{array}{c} \text{Bl}_3 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_4 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F, \ell - X_{3} & SU(3)_0 + 6 \textbf{F} \\\hline f_1 \cdot E = 0, \ell- X_{1} & [SU(2) + 2 \textbf{F} ] \times [SU(2) \times 2\textbf{F}] \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (b3) -- (c3);
\draw[big arrow] (c3) -- (b3);
\draw[big arrow,transform canvas={xshift=.5em}] (c3) -- (d3);
\draw[big arrow] (d3) -- (c3);
\draw[big arrow,transform canvas={xshift=.5em}] (d3) -- (e3);
\draw[big arrow] (e3) -- (d3);
\draw[big arrow] (a) -- (b);
\draw[big arrow] (a3) -- (f);
\draw[big arrow] (b) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (a3) -- (a4);
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (a);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- (e);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- node[right,midway]{$\phi_1 \leftrightarrow \phi_2$} (a3);
\draw[big arrow,transform canvas={xshift=.5em}] (a4) -- (a3);
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=7$ geometries, cont.}
\label{fig:7b}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1.3]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 5 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](b1) at (0,-2) {$\begin{array}{c} \text{Bl}_{4} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{5}{2}} + 5 \textbf{F} \\\hline F, \ell - X_2 & Sp(2) + 5 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](b2) at (-5,-2) {$\begin{array}{c} \text{Bl}_{5} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,F & Sp(2) + 5 \textbf{F} \\\hline F, E & SU(3)_{\frac{5}{2}} + 5 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](c) at (5.5,-2) {$\begin{array}{c} \text{Bl}_{3} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F, \ell - X_{1} &SU(3)_{\frac{5}{2}} + 5 \textbf{F} \\\hline F, \ell - X_3 & Sp(2) + 5 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](d) at (5.5,0) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4\\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{5}{2}} + 5 \textbf{F} \\\hline F,\ell - X_4 & Sp(2) + 5 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](e) at (5.5,2) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_2 \overset{2 \ell-\sum_{i=1}^4 X_i}{\cup} \text{dP}_5 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{5}{2}} + 5 \textbf{F} \\\hline F,\ell - X_5 & Sp(2) + 5 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](f) at (-1,2) {$\begin{array}{c} \mathbb F_1 \overset{2 \ell-\sum_{i=1}^5 X_i }{\cup} \text{dP}_6\\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{5}{2}} + 5 \textbf{F} \\\hline F, \ell - X_6 & Sp(2) + 5 \textbf{F} \\\hline \end{array}$} \end{array}$};
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={yshift=-.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\node[](a) at (4,-2) {$ \begin{array}{c} \text{Bl}_{6} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (0,-2) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{3}{2}}+ 5 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](c) at (0,-4) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline
F, \ell - X_{1} & SU(3)_{\frac{3}{2}}+ 5 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](d) at (0,-6) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{3}{2}}+ 5 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](e) at (0,-8.5) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_2 \overset{2 \ell - \sum_{i=1}^4 X_i}{\cup} \text{dP}_4 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{3}{2}}+ 5 \textbf{F} \\\hline H - X_1 - X_2 , 2\ell - \sum X_i & SU(2)_\pi \times [SU(2) + 3 \textbf{F} ] \\\hline \end{array}$}\end{array}$};
\node[](f) at (0,-10.8) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_1 \overset{2 \ell - \sum_{i=1}^5 X_i}{\cup} \text{dP}_5 \\ \scalebox{.7}{$ \begin{array}{|c|c|} \hline
F,\ell - X_1 &SU(3)_{\frac{3}{2}}+ 5 \textbf{F} \\\hline f_1 \cdot E = 0, 2\ell - \sum_{i=1}^4 X_i & SU(2)_\pi \times [ SU(2) + 3 \textbf{F}]\\\hline \end{array}$}\end{array}$};
\node(g) at (0,-13) {$\begin{array}{c} \text{Bl}_5 \mathbb F_2 \overset{E}{\cup} \mathbb F_0 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F, F & SU(3)_{-\frac{3}{2}} + 5 \textbf{F} \\\hline H-X_1 - X_2 ,E& [SU(2) + 3\textbf{F}] \times SU(2)_\pi \\\hline \end{array}$} \end{array}$};
\draw[big arrow] (a) -- (b);
\draw[big arrow] (b) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow] (f) -- (g);
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (a);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- (e);
\draw[big arrow,transform canvas={xshift=.5em}] (g) -- node[right,midway]{$\phi_1 \leftrightarrow \phi_2$} (f);
\end{tikzpicture}
\end{array}
\begin{array}{c}
\begin{tikzpicture}
\node(A) at (8,-8) {$\begin{array}{c} \text{Bl}_6 \mathbb F_3 \overset{\ell}{\cup} \mathbb P^2\end{array}$};
\node(B) at (8,-10) {$\begin{array}{c} \text{Bl}_5 \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_1 \\ \scalebox{.7}{$\begin{array}{|c|c|} \hline H-X_1-X_2 , \ell - X_1 & [SU(2) + 3 \textbf{F} ] \times SU(2)_0 \\\hline \end{array}$} \end{array}$};
\node(C) at (8,-12) {$\begin{array}{c} \text{Bl}_4 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_2 \\ \scalebox{.7}{$\begin{array}{|c|c|}\hline F,\ell-X_{1} & [SU(2) + 3 \textbf{F}] \times SU(2)_{0} \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (A) -- (B);
\draw[big arrow,transform canvas={xshift=.5em}] (C) -- (B);
\draw[big arrow] (B) -- (A);
\draw[big arrow] (B) -- (C);
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=6$ geometries.}
\label{fig:6a}
\end{figure}
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}
\node (e) at (1.5,-2.5) {$\begin{array}{c} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_6 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline F,\ell - X_6 & SU(3)_{\frac{9}{2}} + 5 \textbf{F}\\\hline F, 2\ell - \sum_{i=3}^6 X_{i} & Sp(2) + 3 \textbf{F} + 2 \textbf{AS} \\\hline F, 3 \ell - \sum_{i=1}^5 X_i - 2 X_6 & G_2 + 5 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (f) at (9,-2.5) {$\begin{array}{c} \mathbb F_2 \overset{\ell - X_1}{\cup} \text{dP}_6 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline F,\ell-X_2 & SU(3)_{\frac{7}{2}} + 5 \textbf{F}\\\hline F,2 \ell- \sum_{i=3}^6 X_i & Sp(2) + 4 \textbf{F} + 1 \textbf{AS} \\\hline \end{array}$} \end{array}$};
\end{tikzpicture}
\end{array}\\
\begin{array}{c}
\begin{tikzpicture}
\node (bn) at (8.1,-1) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \scalebox{1}{$ \begin{array}{|c|c|}\hline F,F & SU(3)_{\frac{1}{2}} + 5 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (cn) at (8.1,-3) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{\frac{1}{2}} + 5 \textbf{F}\\\hline \end{array}$} \end{array}$};
\node (dn) at (8.1,-5.5) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_2 \overset{\ell- X_1}{\cup} \text{dP}_3 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{\frac{1}{2}} + 5 \textbf{F} \\\hline H- X_1 - X_2 , \ell-X_1 & [SU(2) + 1 \textbf{F} ] \times [SU(2) +2 \textbf{F}] \\\hline \end{array}$} \end{array}$};
\node(en) at (8.1,-8) {$\begin{array}{c} \text{Bl}_2 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_4 \\ \scalebox{1}{$ \begin{array}{|c|c|}\hline F,\ell -X_4 & SU(3)_{\frac{1}{2}} + 5 \textbf{F} \\\hline f_1 \cdot E =0 , \ell-X_1 & [SU(2) + 1 \textbf{F} ] \times [SU(2) +2 \textbf{F}] \\\hline \end{array}$} \end{array}$};
\node[](a1n) at (8.1,-13) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\\scalebox{1}{$\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{-\frac{1}{2}} + 5 \textbf{F} \\\hline \end{array}$}\end{array} $};
\node[](a2n) at (8.1,-10.5) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_2 \overset{ \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{-\frac{1}{2}} + 5 \textbf{F} \\\hline H- X_1 -X_2 ,\ell-X_1 & [SU(2) + 2 \textbf{F}] \times [ SU(2) + 1 \textbf{F}]\\\hline \end{array}$}\end{array} $};
\draw[big arrow,transform canvas={xshift=.5em}] (bn) -- (cn);
\draw[big arrow] (cn) -- (bn);
\draw[big arrow,transform canvas={xshift=.5em}] (cn) -- (dn);
\draw[big arrow] (dn) -- (cn);
\draw[big arrow,transform canvas={xshift=.5em}] (dn) -- (en);
\draw[big arrow] (en) -- (dn);
\draw[big arrow,transform canvas={xshift=.5em}] (a1n) -- (a2n);
\draw[big arrow] (a2n) -- (a1n);
\draw[big arrow,transform canvas={xshift=.5em}] (a2n) -- node[right,midway]{$\phi_1 \leftrightarrow \phi_2$} (en);
\draw[big arrow] (en) -- (a2n);
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=6$ geometries, cont.}
\label{fig:6b}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1.3]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 4 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](b1) at (0,-2) {$\begin{array}{c} \text{Bl}_{3} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F, \ell - X_1 & SU(3)_{3} + 4 \textbf{F} \\\hline F,\ell - X_2 & Sp(2) + 4 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](b2) at (-5,-2) {$\begin{array}{c} \text{Bl}_{4} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,F & Sp(2) + 4 \textbf{F} \\\hline F, E & SU(3)_{3} + 4 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](c) at (5.5,-2) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{3} + 4 \textbf{F} \\\hline F,\ell - X_3 & Sp(2) + 4 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](d) at (5.5,0) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4\\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{3} + 4 \textbf{F} \\\hline \ell - X_4 & Sp(2) + 4 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node[](e) at (5.5,2) {$\begin{array}{c} \mathbb F_2 \overset{2 \ell-\sum_{i=1}^4 X_i}{\cup} \text{dP}_5 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{3} + 4 \textbf{F} \\\hline F, \ell - X_5 & Sp(2) + 4 \textbf{F} \\\hline \end{array} $}\end{array}$};
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{5} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (0,-2) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{2}+ 4 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](c) at (0,-4) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{j=1,2} &SU(3)_{2}+ 4 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](d) at (0,-6) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_{j=1,\dots,3} &SU(3)_{2}+ 4 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](e) at (0,-8) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_2 \overset{2 \ell - \sum_{i=1}^4 X_i}{\cup} \text{dP}_4 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_{j=1,\dots,4} &SU(3)_{2}+ 4 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](f) at (0,-10) {$ \begin{array}{c} \mathbb F_1 \overset{2 \ell - \sum_{i=1}^5 X_i}{\cup} \text{dP}_5 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 &SU(3)_{2}+ 4 \textbf{F} \\\hline \end{array}$}\end{array}$};
\draw[big arrow] (a) -- (b);
\draw[big arrow] (b) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow] (e) -- (f);
\draw[big arrow,transform canvas={xshift=.5em}] (b) -- (a);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- (e);
\end{tikzpicture}
\end{array}
\begin{array}{c}
\begin{tikzpicture}
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=5$ geometries.}
\label{fig:5a}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
\noindent\makebox[\textwidth]{ $
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}
\node (b) at (0,-3) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \scalebox{1}{$ \begin{array}{|c|c|}\hline F,F & SU(3)_{1} + 4 \textbf{F} \\\hline \end{array} $}\end{array}$};
\node (c) at (0,-5) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\ \scalebox{1}{$ \begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{1} + 4 \textbf{F}\\\hline \end{array}$} \end{array}$};
\node (d) at (0,-7) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_2 \overset{\ell- X_1}{\cup} \text{dP}_3 \\\scalebox{1}{$ \begin{array}{|c|c|}\hline F, \ell - X_2 & SU(3)_{1} + 4 \textbf{F} \\\hline H-X_1-X_2 , \ell - X_1 & SU(2)_\pi \times [SU(2) + 2 \textbf{F}]\\\hline\end{array}$} \end{array}$};
\node(e) at (0,-9.5) {$\begin{array}{c} \text{Bl}_1 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_4 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F,\ell- X_3 & SU(3)_{1} + 4 \textbf{F} \\\hline f_1 \cdot E =0, \ell - X_1 & SU(2)_\pi \times [ SU(2) + 2 \textbf{F}] \\\hline \end{array} $}\end{array}$};
\node(f) at (0,-12.5) {$\begin{array}{c} \text{Bl}_4 \mathbb F_2 \overset{E}{\cup} \mathbb F_0 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F,F & SU(3)_{-1} + 4 \textbf{F} \\\hline H-X_1-X_2 ,E&[SU(2) + 2 \textbf{F}] \times SU(2)_\pi \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (b) -- (c);
\draw[big arrow] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (d);
\draw[big arrow] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (e);
\draw[big arrow] (e) -- (d);
\draw[big arrow] (e) --(f);
\draw[big arrow,transform canvas={xshift=.5em}] (f) -- node[right,midway]{$\phi_1 \leftrightarrow \phi_2$}(e);
\end{tikzpicture}
\end{array} ~~~~~~~ \begin{array}{c}\begin{tikzpicture} \node(A) at (6,-7) {$\begin{array}{c} \text{Bl}_5 \mathbb F_3 \overset{\ell}{\cup} \mathbb P^2\end{array}$};
\node(B) at (6,-9) {$\begin{array}{c} \text{Bl}_4 \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_1 \\\begin{array}{|c|c|} \hline H-X_1-X_2,\ell-X_1 &[SU(2) + 2 \textbf{F}] \times SU(2)_0 \\\hline \end{array} \end{array}$};
\node(C) at (6,-11) {$\begin{array}{c} \text{Bl}_3 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_2 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F, \ell-X_{1} & [SU(2) + 2 \textbf{F}] \times SU(2)_{0} \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (A) -- (B);
\draw[big arrow] (B) -- (A);
\draw[big arrow] (B) -- (C);
\draw[big arrow,transform canvas={xshift=.5em}] (C) -- (B);
\end{tikzpicture}\end{array}\\ \\
\begin{array}{c}
\begin{tikzpicture}
\node (e) at (-2,-7.5) {$\begin{array}{c} \mathbb F_4 \overset{2\ell-X_1-X_2}{\cup} \text{dP}_5 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline F,\ell- X_1 & SU(3)_{5} + 4 \textbf{F}\\\hline F, \ell - X_5 & Sp(2) + 2 \textbf{F} + 2 \textbf{AS} \\\hline F, 2\ell - \sum_{i=2}^5 X_i & G_2 + 4 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (f) at (5,-7.5) {$\begin{array}{c} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_5 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline F, \ell - X_1 & SU(3)_{4} + 4 \textbf{F}\\\hline F, 2\ell - \sum X_i & Sp(2) + 3 \textbf{F} + 1 \textbf{AS} \\\hline \end{array} $}\end{array}$};
\node[](k) at (-4,-4.5) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{0} + 4 \textbf{F} \\\hline \end{array}$}\end{array} $};
\node (l) at (3.5,-4.5) {$\begin{array}{c} \text{Bl}_{3} \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline F,\ell-X_2 & SU(3)_{0} + 4 \textbf{F} \\\hline H-X_1-X_2 , \ell-X_1 & [SU(2) +1 \textbf{F}] \times[ SU(2) +1 \textbf{F}] \\\hline \end{array} $}\end{array}$};
\node (m) at (3.5,-2) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_1 \overset{\ell-X_1-X_2}{\cup} \text{dP}_3 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline F,\ell - X_3 & SU(3)_{0} + 4 \textbf{F} \\\hline f_1 \cdot E = 0,\ell- X_1 & [SU(2) +1 \textbf{F}] \times[ SU(2) +1 \textbf{F}] \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={yshift=-.5em}] (k) -- (l);
\draw[big arrow] (l) -- (k);
\draw[big arrow,transform canvas={xshift=.5em}] (l) -- (m);
\draw[big arrow] (m) -- (l);
\end{tikzpicture}
\end{array}
\end{array}
$}
\end{center}
\caption{$M=5$ geometries, cont.}
\label{fig:5b}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1.3]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F, \ell - X_1 & Sp(2) + 3 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](b1) at (0,-2) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{7}{2}} + 3 \textbf{F} \\\hline F, \ell - X_2 & Sp(2) + 3 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](b2) at (-5,-2) {$\begin{array}{c} \text{Bl}_{3} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\\scalebox{1}{$ \begin{array}{|c|c|} \hline
F,F & Sp(2) + 3 \textbf{F} \\\hline F,E &SU(3)_{\frac{7}{2}} + 3 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](c) at (5.5,-2) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{7}{2}} + 3 \textbf{F} \\\hline F, \ell - X_3 & Sp(2) + 3 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node[](d) at (5.5,0) {$\begin{array}{c} \mathbb F_3 \overset{2 \ell-\sum_{i=1}^3 X_i}{\cup} \text{dP}_4\\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{7}{2}} + 3 \textbf{F} \\\hline F,\ell - X_4 & Sp(2) + 3 \textbf{F} \\\hline \end{array}$} \end{array}$};
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{4} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (0,-2) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{5}{2}}+ 3 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](c) at (0,-4) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{5}{2}}+ 3 \textbf{F} \\\hline \end{array}$}\end{array}$};
\node[](d) at (0,-6) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\ \scalebox{1}{$\begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{5}{2}}+ 3 \textbf{F}\\\hline \end{array}$}\end{array}$};
\node[](e) at (0,-8) {$ \begin{array}{c} \mathbb F_2 \overset{2 \ell - \sum_{i=1}^4 X_i}{\cup} \text{dP}_4 \\ \scalebox{1}{$ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{\frac{5}{2}}+ 3 \textbf{F} \\\hline \end{array}$}\end{array}$};
\draw[big arrow] (a) -- (b);
\draw[big arrow] (b) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow] (d) -- (e);
\draw[big arrow,transform canvas={xshift=.5em}] (b) -- (a);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (e) -- (d);
\end{tikzpicture}
\end{array}
~~~~~~~~~~~~~~~~~~~
\begin{array}{c}
\begin{tikzpicture}
\node (b) at (1,-3) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F,F & SU(3)_{\frac{3}{2}} + 3 \textbf{F} \\\hline \end{array}$} \end{array}$};
\node (c) at (1,-5) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F,\ell - X_1 &SU(3)_{\frac{3}{2}} + 3 \textbf{F}\\\hline \end{array}$} \end{array}$};
\node (d) at (1,-7) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_2 \overset{\ell- X_1}{\cup} \text{dP}_3 \\ \scalebox{1}{$ \begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{\frac{3}{2}} + 3 \textbf{F}\\\hline \end{array}$} \end{array}$};
\node(e) at (1,-9) {$\begin{array}{c} \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_4 \\ \scalebox{1}{$\begin{array}{|c|c|}\hline F, \ell - X_3 &SU(3)_{\frac{3}{2}} + 3 \textbf{F} \\\hline \end{array}$} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (b) -- (c);
\draw[big arrow] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (c) -- (d);
\draw[big arrow] (d) -- (c);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (e);
\draw[big arrow] (e) -- (d);
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=4$ geometries.}
\label{fig:4a}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
\noindent\makebox[\textwidth]{ $
\begin{array}{c}
\begin{tikzpicture}
\node(a) at (-2,2) {$\begin{array}{c} \text{Bl}_4 \mathbb F_3 \overset{\ell}{\cup} \mathbb P^2\end{array}$};
\node(b) at (-2,0) {$\begin{array}{c} \text{Bl}_3 \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline H-X_1 -X_2 , \ell - X_1 &[SU(2) + 1\textbf{F}] \times SU(2)_0\end{array} \\\hline \end{array}$};
\node(C) at (7,0) {$\begin{array}{c} \text{Bl}_2 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_2 \\\begin{array}{|c|c|}\hline F,\ell-X_{j=1,2} & [SU(2) + \textbf{F}] \times SU(2)_0 \\\hline \end{array} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (a) -- (b);
\draw[big arrow] (b) -- (a);
\draw[big arrow] (b) -- (C);
\draw[big arrow,transform canvas={yshift=-.5em}] (C) -- (b);
\node (e) at (6,-7.5) {$\begin{array}{c} \mathbb F_4 \overset{2\ell-X_1-X_2}{\cup} \text{dP}_4 \\ \begin{array}{|c|c|} \hline F, \ell -X_1 & SU(3)_{\frac{9}{2}} + 3 \textbf{F}\\\hline F,\ell - X_3 & Sp(2) + 2 \textbf{F} + 1\textbf{AS} \\\hline \end{array} \end{array}$};
\node (f) at (-2,-7.5) {$\begin{array}{c} \mathbb F_5 \overset{2\ell - X_1}{\cup} \text{dP}_4 \\ \begin{array}{|c|c|} \hline F , \ell - X_1 & SU(3)_{\frac{11}{2}} + 3 \textbf{F}\\\hline F, \ell - X_1 & Sp(2) + \textbf{F} + 2 \textbf{AS} \\\hline F , 2\ell - \sum X_i & G_2 + 3 \textbf{F} \\\hline \end{array} \end{array}$};
\node (f) at (3,2.5) {$\begin{array}{c} (\mathbb F_6 \overset{2\ell}{\cup} \text{dP}_4)^* \\ \begin{array}{|c|c|} \hline F, \ell - X_1 & Sp(2)_{ 0} + 3 \textbf{AS} \\\hline F, 2\ell - \sum X_i & A^{(2)}_2 \\\hline \end{array} \end{array}$};
\node[](k) at (-2.5,-4.5) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\\begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{\frac{1}{2}} + 3 \textbf{F} \\\hline \end{array}\end{array} $};
\node (l) at (-2.5,-2) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|} \hline F,\ell -X_2 &SU(3)_{\frac{1}{2}} + 3 \textbf{F} \\\hline H-X_1 - X_2 , \ell - X_1 & SU(2)_\pi \times [ SU(2) + 1 \textbf{F}] \\\hline\end{array} \end{array}$};
\node (m) at (6.5,-2) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_1 \overset{\ell-X_1-X_2}{\cup} \text{dP}_3 \\ \begin{array}{|c|c|} \hline F, \ell - X_3 &SU(3)_{\frac{1}{2}} + 3 \textbf{F} \\\hline f_1 \cdot E =0, \ell - X_1 & SU(2)_\pi \times [ SU(2) + 1 \textbf{F}] \\\hline \end{array} \end{array}$};
\node(n) at (6.5,-4.5) {$\begin{array}{c} \text{Bl}_3 \mathbb F_2 \overset{E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|}\hline F, F & SU(3)_{-\frac{1}{2}} + 3 \textbf{F} \\\hline H - X_1 - X_2, E & [SU(2) + 1 \textbf{F}] \times SU(2)_\pi \\\hline \end{array} \end{array}$};
\draw[big arrow,transform canvas={xshift=.5em}] (k) -- (l);
\draw[big arrow] (l) -- (k);
\draw[big arrow,transform canvas={yshift=-.5em}] (l) -- (m);
\draw[big arrow] (m) -- (l);
\draw[big arrow] (m) -- (n);
\draw[big arrow,transform canvas={xshift=.5em}] (n) -- node[right,midway]{$\phi_1\leftrightarrow \phi_2$} (m);
\end{tikzpicture}
\end{array}
$}
\end{center}
\caption{$M=4$ geometries, cont.}
\label{fig:4b}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\node[](a) at (-4.5,-2) {$ \begin{array}{c} \text{Bl}_{3} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (0,-2) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline
F, \ell - X_1 & SU(3)_{3}+ 2 \textbf{F} \\\hline \end{array}\end{array}$};
\node[](c) at (5.5,-2) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|} \hline
F,\ell - X_{j=1,2} &SU(3)_{3}+ 2 \textbf{F} \\\hline \end{array}\end{array}$};
\node[](d) at (5.5,-4.5) {$ \begin{array}{c} \mathbb F_3 \overset{2 \ell - \sum_{i=1}^3 X_i}{\cup} \text{dP}_3 \\ \begin{array}{|c|c|} \hline
F, \ell - X_{1} &SU(3)_{3}+ 2 \textbf{F} \\\hline \end{array}\end{array}$};
\draw[big arrow] (a) -- (b);
\draw[big arrow] (b) -- (c);
\draw[big arrow] (c) -- (d);
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (d) -- (c);
\end{tikzpicture}
\end{array}
\\
\begin{tikzpicture}[yscale=1.2]
\node[](a) at (0,0) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 2 \textbf{F} \\\hline \end{array}\end{array}$};
\node[](b1) at (0,-2) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{4} + 2 \textbf{F} \\\hline F,\ell - X_2 & Sp(2) + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\node[](b2) at (-5,-2) {$\begin{array}{c} \text{Bl}_{2} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|} \hline
F, F & Sp(2) + 2 \textbf{F} \\\hline F, E &SU(3)_{4} + 2 \textbf{F}\\\hline \end{array} \end{array}$};
\node[](c) at (5.5,-2) {$\begin{array}{c} \mathbb F_4 \overset{2 \ell-\sum_{i=1}^2 X_i}{\cup} \text{dP}_3 \\ \begin{array}{|c|c|} \hline
F,\ell - X_{1} &SU(3)_{4} + 2 \textbf{F} \\\hline F, \ell - X_3 & Sp(2) + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (c);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={xshift=.5em}] (b1) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b1);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}
\node (b) at (-5,-2.5) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|}\hline F,F & SU(3)_{2} + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\node (c) at (-.5,-2.5) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|}\hline F, \ell - X_1 &SU(3)_{2} + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\node (d) at (4.5,-2.5) {$ \begin{array}{c} \mathbb F_2 \overset{\ell- X_1}{\cup} \text{dP}_3 \\ \begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{2} + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (c);
\draw[big arrow] (c) -- (b);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (d);
\draw[big arrow] (d) -- (c);
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=3$ geometries.}
\label{fig:3b}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{tikzpicture}
\node(C2) at (4,5) {$\begin{array}{c} \text{dP}_2 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_2 \\\begin{array}{|c|c|}\hline \ell- X_1, \ell-X_{1} & SU(2)_{0} \times SU(2)_0 \\\hline \end{array} \end{array}$};
\node(a) at (-2,5) {$\begin{array}{c} \text{Bl}_3 \mathbb F_3 \overset{\ell}{\cup} \mathbb P^2\end{array}$};
\node(b) at (-2,3) {$\begin{array}{c} \text{Bl}_2 \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline H-X_1-X_2, \ell - X_1 & SU(2)_\pi \times SU(2)_0 \\\hline \end{array} \end{array}$};
\node(C1) at (6,3) {$\begin{array}{c} \text{Bl}_1 \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_2 \\\begin{array}{|c|c|} \hline f_1 \cdot E = 0, \ell-X_{1} & SU(2)_{\pi} \times SU(2)_0 \\\hline \end{array} \end{array}$};
\draw[big arrow] (b) -- (C1);
\draw[big arrow,transform canvas={yshift=-.5em}] (C1) -- (b);
\draw[big arrow,transform canvas={xshift=.5em}] (a) -- (b);
\draw[big arrow] (b) -- (a);
\node(d1) at (-2.5,1) {$\begin{array}{c} \text{Bl}_2 \mathbb F_2 \overset{E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|}\hline F,F & SU(3)_{0} + 2 \textbf{F} \\\hline H - X_1 - X_2 , E & SU(2)_\pi \times SU(2)_\pi \\\hline \end{array} \end{array}$};
\node(d2) at (5.5,1) {$\begin{array}{c} \text{Bl}_1 \mathbb F_1 \overset{X_1}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{0} + 2 \textbf{F} \\\hline f_1 \cdot E = 0, \ell - X_2 & SU(2)_\pi \times SU(2)_\pi \\\hline \end{array} \end{array}$};
\node (f) at (-3.5,-3) {$\begin{array}{c} \mathbb F_5 \overset{2\ell - X_1}{\cup} \text{dP}_3 \\ \begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{5} + 2 \textbf{F}\\\hline F,\ell - X_{2} & Sp(2) + 1\textbf{F} + 1\textbf{AS} \\\hline \end{array} \end{array}$};
\node (f) at (2,-3) {$\begin{array}{c} \mathbb F_6 \overset{2\ell}{\cup} \text{dP}_3\\ \begin{array}{|c|c|} \hline F,\ell -X_{1} & Sp(2)_{0}+ 2 \textbf{AS} \\\hline \end{array} \end{array}$};
\node[](k) at (-3,-1) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\\begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{1} + 2 \textbf{F} \\\hline \end{array}\end{array} $};
\node (l) at (2,-1) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|} \hline F,\ell-X_2 &SU(3)_{1} + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\node (m) at (7,-1) {$\begin{array}{c} \mathbb F_1 \overset{\ell-X_1-X_2}{\cup} \text{dP}_3 \\ \begin{array}{|c|c|} \hline F, \ell - X_3 &SU(3)_{1} + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\node (g) at (7,-3.5) {$\begin{array}{c} \mathbb F_6 \overset{3 \ell - 2 X_1 - X_2 }{\cup} \text{dP}_3\\ \begin{array}{|c|c|} \hline F, \ell-X_2 &Sp(2)_{\pi} + 2 \textbf{AS} \\\hline F,\ell - X_1 & SU(3)_6 + 2 \textbf{F} \\\hline F, \ell- X_3 &G_2 + 2 \textbf{F} \\\hline \end{array} \end{array}$};
\draw[big arrow,transform canvas={yshift=-.5em}] (k) -- (l);
\draw[big arrow] (l) -- (k);
\draw[big arrow,transform canvas={yshift=-.5em}] (l) -- (m);
\draw[big arrow] (m) -- (l);
\draw[big arrow,transform canvas={yshift=-.5em}] (d1) -- (d2);
\draw[big arrow] (d2) -- (d1);
\end{tikzpicture}
\end{array}
$
\end{center}
\caption{$M=3$ geometries, cont. Note that for the geometry $\text{dP}_2 \cup \text{dP}_2$ at the top, the gluing curves in \emph{both} surfaces are $C = \ell - X_1 - X_2$, in contrast to the other geometries.}
\label{fig:3b}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1]
\node[](a) at (-5,-2) {$ \begin{array}{c} \text{Bl}_{2} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b) at (0,-2) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{7}{2}}+1 \textbf{F} \\\hline \end{array}\end{array}$};
\node[](c) at (5,-2) {$ \begin{array}{c} \mathbb F_4 \overset{2 \ell - \sum_{i=1}^2 X_i}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|} \hline
F,\ell - X_{j=1,2} &SU(3)_{\frac{7}{2}}+ 1 \textbf{F} \\\hline \end{array}\end{array}$};
\draw[big arrow] (a) -- (b);
\draw[big arrow] (b) -- (c);
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (a);
\draw[big arrow,transform canvas={yshift=-.5em}] (c) -- (b);
\end{tikzpicture}
\end{array}
\\ \\
\begin{tikzpicture}[yscale=1.2]
\node[](a) at (5,0) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline
F,\ell - X_1 & Sp(2) + 1\textbf{F} \\\hline \end{array}\end{array}$};
\node[](b1) at (0,0) {$\begin{array}{c} \mathbb F_5 \overset{2 \ell-X_1}{\cup} \text{dP}_2 \\\begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{\frac{9}{2}} + 1 \textbf{F} \\\hline F,\ell - X_2 & Sp(2) + 1 \textbf{F} \\\hline \end{array} \end{array}$};
\node[](b2) at (-5,0) {$\begin{array}{c} \text{Bl}_{1} \mathbb F_6 \overset{F + 2 E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|} \hline
F, F & Sp(2) + 1\textbf{F} \\\hline F,E &SU(3)_{\frac{9}{2}} + 1\textbf{F}\\\hline \end{array} \end{array}$};
\draw[big arrow] (a) -- (b1);
\draw[big arrow] (b1) -- (b2);
\draw[big arrow,transform canvas={yshift=-.5em}] (b2) -- (b1);
\draw[big arrow,transform canvas={yshift=-.5em}] (b1) -- (a);
\end{tikzpicture}
\end{array}
\\ \\
\begin{array}{c}
\begin{tikzpicture}
\node (b) at (-2.5,-2.5) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_4 \overset{F+E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|}\hline F,F & SU(3)_{\frac{5}{2}} + 1\textbf{F} \\\hline \end{array} \end{array}$};
\node (c) at (2.5,-2.5) {$ \begin{array}{c} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|}\hline F, \ell - X_1 &SU(3)_{\frac{5}{2}} + 1\textbf{F} \\\hline \end{array} \end{array}$};
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (c);
\draw[big arrow] (c) -- (b);
\end{tikzpicture}
\end{array}
\\
\begin{tikzpicture}
\node(a) at (-1,3) {$\begin{array}{c} \text{Bl}_2 \mathbb F_3 \overset{\ell}{\cup} \mathbb P^2\end{array}$};
\node(b) at (3,3) {$\begin{array}{c} \text{Bl}_1 \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_1 \end{array}$};
\draw[big arrow,transform canvas={yshift=-.5em}] (a) -- (b);
\draw[big arrow] (b) -- (a);
\node(d1) at (.5,1.5) {$\begin{array}{c} \text{Bl}_1 \mathbb F_2 \overset{E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|}\hline F, F & SU(3)_{\frac{1}{2}} + 1\textbf{F} \\\hline \end{array} \end{array}$};
\node(d2) at (5.5,1.5) {$\begin{array}{c} \mathbb F_1 \overset{X_1}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{\frac{1}{2}} + 1\textbf{F} \\\hline \end{array} \end{array}$};
\node(d3) at (7,3) {$\begin{array}{c} \mathbb F_1 \overset{\ell - X_1 - X_2}{\cup} \text{dP}_2 \end{array}$};
\draw[big arrow] (d3) -- (b);
\draw[big arrow,transform canvas={yshift=-.5em}] (b) -- (d3);
\node (f) at (-2,-2.5) {$\begin{array}{c} \mathbb F_7 \overset{3\ell -2 X_1}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|}\hline F,\ell - X_1 & SU(3)_{\frac{13}{2}} + 1\textbf{F}\\\hline F, \ell - X_2 & G_2 + 1\textbf{F} \\\hline \end{array} \end{array}$};
\node (f) at (3,-2.5) {$\begin{array}{c} \mathbb F_6 \overset{2\ell}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|} \hline F, \ell - X_1 &Sp(2)_0 + 1\textbf{AS} \\\hline \end{array} \end{array}$};
\node (f1) at (8,-2.5) {$\begin{array}{c} \mathbb F_6 \overset{3\ell - 2 X_1 - X_2}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|} \hline F, \ell - X_1& SU(3)_{\frac{11}{2}} + 1\textbf{F}\\\hline F,\ell - X_2 & Sp(2)_\pi +1 \textbf{AS} \\\hline \end{array} \end{array}$};
\node[](k) at (0,-.5) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_3 \overset{ \ell}{\cup} \text{dP}_1 \\\begin{array}{|c|c|}\hline F, \ell - X_1 & SU(3)_{\frac{3}{2}} + 1\textbf{F} \\\hline \end{array}\end{array} $};
\node (l) at (5,-.5) {$\begin{array}{c} \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_2 \\ \begin{array}{|c|c|} \hline F, \ell -X_2 &SU(3)_{\frac{3}{2}} + 1\textbf{F} \\\hline \end{array} \end{array}$};
\draw[big arrow,transform canvas={yshift=-.5em}] (k) -- (l);
\draw[big arrow] (l) -- (k);
\draw[big arrow,transform canvas={yshift=-.5em}] (d1) -- (d2);
\draw[big arrow] (d2) -- (d1);
\end{tikzpicture}
\end{array}
$
\end{center}
\caption{$M=2$ geometries.}
\label{fig:2a}
\end{figure}
\clearpage
\begin{figure}
\begin{center}
$
\begin{array}{c}
\begin{array}{c}
\begin{tikzpicture}[yscale=1.2]
\node[](a1) at (-4,-3) {$ \begin{array}{c} \text{Bl}_{1} \mathbb F_6 \overset{2 \ell}{\cup} \mathbb P^2 \end{array} $};
\node[](b1) at (-.5,-3) {$ \begin{array}{c} \mathbb F_5 \overset{2 \ell - X_1}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline
F,\ell - X_1 & SU(3)_{4} \\\hline \end{array}\end{array}$};
\draw[big arrow] (a1) -- (b1);
\draw[big arrow,transform canvas={yshift=-.5em}] (b1) -- (a1);
\node[](a) at (3,-3) {$ \begin{array}{c} \text{Bl}_1 \mathbb F_3 \overset{\ell}{\cup} \mathbb P^2 \end{array}$};
\node(ee) at (6.5,-3) {$\begin{array}{c} \mathbb F_2 \overset{\ell-X_1}{\cup} \text{dP}_1 \end{array}$};
\node[](b2) at (1.5,.5) {$\begin{array}{c} \mathbb F_{2b} \overset{F + (b-1) E}{\cup} \mathbb F_0 \\ \begin{array}{|c|c|c|} \hline
b=1 & F,E & SU(3)_1 \\\hline b=2 &F,F & SU(3)_3 \\\hline b=3 &F, E & SU(3)_5 \\\hline b = 3 &F,F & Sp(2)_{\pi} \\\hline b = 4 & F,E & SU(3)_7 \\\hline b = 4 &F,F &G_2 \\\hline b = 5 & F,E & SU(3)_9\\\hline b = 5 & F, F & A^{(2)}_2 \\\hline \end{array} \end{array}$};
\draw[big arrow] (a) -- (ee);
\draw[big arrow,transform canvas={yshift=-.5em}] (ee) -- (a);
\end{tikzpicture}
\end{array}
\\
\begin{array}{c}
\begin{tikzpicture}
\node at (-2.5,0) {$\begin{array}{c} \mathbb F_3 \overset{\ell}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline F,\ell - X_1 & SU(3)_2 \\\hline\end{array} \end{array}$};
\node[](a) at (1,0) {$ \begin{array}{c} \mathbb F_6 \overset{2 \ell}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline
F,\ell - X_1 &Sp(2)_0\\\hline \end{array}\end{array}$};
\node at (4.5,0) {$\begin{array}{c} \mathbb F_1 \overset{X_1}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline F,\ell - X_1 & SU(3)_0 \\\hline\end{array} \end{array}$};
\node at (8,0) {$\begin{array}{c} \mathbb F_7 \overset{3\ell - 2X_1}{\cup} \text{dP}_1 \\ \begin{array}{|c|c|} \hline F, \ell - X_1 & SU(3)_6 \\\hline\end{array} \end{array}$};
\end{tikzpicture}
\end{array}
\begin{array}{c}
\begin{tikzpicture}
\end{tikzpicture}
\end{array}
\end{array}
$
\end{center}
\caption{$M=1$ geometries.}
\label{fig:1}
\end{figure}
\begin{figure}
\centering
$
\begin{array}{ccc}
\mathbb F_3 \overset{\ell}{\cup} \mathbb P^2
& ~~~~&
\mathbb F_6 \overset{2\ell}{\cup} \mathbb P^2
\\ \\
\end{array}
$
\caption{$M=0$ geometries.}
\label{fig:0}
\centering
\end{figure}
\subsubsection*{6d Theories on a Circle}
In this section we show that the complicated web of theories we have uncovered are actually unified from the perspective of 5d Kaluza-Klein (KK) theories arising from 6d SCFTs compactified on a circle (up to possible automorphism twists and holonomies).
As discussed in Section \ref{sec:rank1}, shrinkable rank 1 geometries are classified by del Pezzo surfaces $\text{dP}_{n\leq 8}$ and $\mathbb{F}_0$ up to physical equivalence. Interestingly, all of them can be obtained via geometric RG flows from $\text{dP}_9$ (equivalently, $\frac{1}{2}$K3). The local $\text{dP}_9$ model is an elliptic 3-fold engineering the 6d SCFT called the `E-string theory'. Therefore all rank 1 5d SCFTs are descendants (i.e. related by rank preserving mass deformations) of the 6d E-string theory compactified on a circle.
We also find that all rank 2 5d SCFTs have 6d origin, but the rank 2 case is significantly more elaborate than the rank 1 case. Geometric constructions produce 5d SCFTs belonging to the four distinct families displayed in Table \ref{tb:rank2-classification}. The geometries of type $(\cdot)^*$ are not shrinkable but rather 5d KK theories~\footnote{These theories are also called \emph{marginal} theories \cite{Jefferson:2017ahm}.}. We expect that these geometries correspond to 6d SCFTs compactified on a circle, possibly with automorphism twists.
One distinguished property of geometries corresponding to 5d KK theories is that there must exist an elliptic curve class whose volume is not controlled by normalizable K\"ahler moduli. The M2-branes wrapping this elliptic class correspond to KK momentum states. For example, the canonical class $-K_{\text{dP}_9} \subset\text{dP}_9$ is an elliptic class with zero volume associated to the KK momenta of the E-string theory compactified on a circle. Another important property is that some KK geometries contain fiber classes forming an affine gauge algebra. Namely, we can find fiber classes $f_i$ such that
\begin{equation}
-f_i \cdot S_j = (A_{\hat{G}})_{ij},
\end{equation}
where $\hat{G}$ denotes an affine gauge algebra. This signals that the corresponding geometry is an elliptic geometry realizing a 5d KK theory. We will now identify 6d origins of the geometries in Table \ref{tb:rank2-classification} using these properties.
We begin with $\text{Bl}_{10}\mathbb{F}_6\cup \mathbb{F}_0$. This geometry has two gauge theory descriptions, namely $SU(3)_0 + 10 \textbf{F}$ and $Sp(2)+ 10 \textbf{F}$. The 6d origin of these gauge theories is discussed in \cite{Yonekura:2015ksa,Hayashi:2015fsa,Gaiotto:2015una,Hayashi:2016abm}. These theories are a circle reduction of the 6d $(D_5,D_5)$ conformal matter theory introduced in \cite{Heckman:2013pva,DelZotto:2014hpa}. The geometry $\text{Bl}_{10}\mathbb{F}_6\cup \mathbb{F}_0$ realizes the circle compactification of this 6d theory. This theory has another duality frame in which an affine gauge algebra is manifest. To see this, choose the fiber classes $f_1=H+2F-\sum_{i=1}^{10}X_i $ and $f_2= F$. These fiber classes indeed form the affine $\hat{A}_1$ Cartan matrix:
\begin{equation}
-(f_i \cdot S_j) = \begin{pmatrix} 2 & -2 \\ -2 & 2 \end{pmatrix} .
\end{equation}
Another geometry $\mathbb{F}_2\cup \text{dP}_7$ is interesting for similar reasons. This geometry admits three different gauge theory descriptions corresponding to the following choices of fiber classes:
\begin{align}
\begin{split}
f_1&= F,~~ f_2= \ell - X_2 \ \ \rightarrow \ \ SU(3)_4+6{\bf F} \ , \\
f_1&= F,~~ f_2= 2\ell - \sum_{i=2}^5X_i \ \ \rightarrow \ \ Sp(2)+2{\bf AS}+4{\bf F} \ , \\
f_1&= F,~~ f_2= 3\ell - \sum_{i=2}^6X_i -2X_7 \ \ \rightarrow \ \ G_2+6{\bf F} \ .
\end{split}
\end{align}
Here, the two surfaces are glued along the curves $C_{S_1}=E$ and $C_{S_2}=\ell-X_1$. This implies new dualities between these three gauge theories and their descendants obtained by RG-flows induced by relevant mass deformations. In addition, we find another distinct duality frame:
\begin{equation}
f_1 = F \,, \ f_2 = 5\ell -X_1-2\sum_{i=2}^7X_i \ .
\end{equation}
The fiber classes in this last frame form the affine Cartan matrix $ A^{(2)}_2$:
\begin{equation}
-(f_i\cdot S_j) = \left(\begin{array}{cc} 2 & -1 \\ -4 & 2 \end{array}\right) \ .
\end{equation}
This algebra $ A^{(2)}_2$ is obtained by an outer automorphism twist of the affine $A^{(1)}_2=\hat{A}_2$ algebra which identifies ${\bf 3}$ and $\bar{\bf 3}$ representations in $A_2\subset \hat{A}_2$. Therefore, one can expect that this geometry is also a KK geometry corresponding to a 6d $SU(3)$ gauge theory compactified on a circle with an outer automorphism twist. The unique 6d theory satisfying these properties is the 6d $\mathcal{N}=(1,0)$ SCFT with $SU(3)$ gauge group and $N_\textbf{F}=12$ fundamental hypermultiplets. Circle compactification of this 6d theory with an outer automorphism twist of the $SU(3)$ gauge algebra leads to a 5d theory with affine $A^{(2)}_2$ gauge algebra and 6 flavors. This interpretation agrees with the geometric model $\mathbb{F}_2\cup \text{dP}_7$. Therefore, we conclude that $\mathbb{F}_2\cup \text{dP}_7$ is a `KK geometry' engineering the circle compactification of the 6d $SU(3)$ theory with $N_\textbf{F} = 12$.
$\mathbb{F}_6\cup \text{dP}_4$ is also a KK geometry. When one chooses the fiber classes $f_1=F_1,f_2=\ell-X_1$ (with the gluing curve $C_{S_2}=2\ell$), this geometry has a gauge theory description as $Sp(2)_{0}+3{\bf AS}$. However, if we choose the fiber classes $f_1=F,f_2=2\ell-\sum_{i=1}^4X_i$, their intersections with the irreducible components $S_i$ form the affine $A^{(2)}_2$ Cartan matrix, up to sign. This suggests that $\mathbb{F}_6\cup \text{dP}_4$ is a KK geometry. Indeed we find that the 6d $SU(3)$ gauge theory with $N_\textbf{F}=6$ can give rise to the 5d KK theory associated to this geometry upon circle reduction with an outer automorphism twist.
$\mathbb{F}_{10}\cup \mathbb{F}_0$ is yet another KK geometry constructed by our building blocks. This geometry admits two dual descriptions related to the base-fiber exchange symmetry of $\mathbb{F}_0$. One description is $SU(3)_9$, while the other is the $ A^{(2)}_2$ gauge theory description without matter hypermultiplets. We anticipate that this affine $ A^{(2)}_2$ gauge theory is the 5d KK theory coming from the 6d theory $\mathcal{O}(-3)$ minimal SCFT with $SU(3)$ gauge group compactified on a circle with an outer automorphism twist of the $SU(3)$ gauge algebra.
Lastly, $\text{Bl}_9\mathbb{F}_4\cup \mathbb{F}_0$ is a KK geometry. This geometry is formed by gluing two surfaces along $C_{S_1}=E$ in $\text{Bl}_9\mathbb{F}_4$ and $C_{S_2}=F+H$ in $\mathbb{F}_0$. We find that this geometry involves an elliptic fiber class given by $E+2X$ (with $E^2=-4,X^2=-1,E\cdot X=2$) in $\text{Bl}_9\mathbb{F}_4$ which signals that this geometry is an elliptic CY 3-fold.
In the 5d reduction, this geometry has two gauge theory descriptions as predicted in \cite{Jefferson:2017ahm}: $SU(3)_{\frac{3}{2}}$ with $N_{\bf F}=9$ and $Sp(2)$ with $N_{\bf AS}=1,N_{\bf F}=8$. This geometry is associated to the 6d rank 2 E-string theory on a circle. This becomes clearer after a flop transition with respect to the exceptional curve $X$. The flop transition described in Section \ref{sec:transitions} leads to $\text{dP}_9\cup \mathbb{F}_0^{g=1}$ geometry where we glue the anticanonical class in $\text{dP}_9$ to the elliptic class $E$ (with $E^2=0$) in $\mathbb{F}_0^{g=1}$. This is the rank 2 generalization of $\text{dP}_9$ (or the 6d rank 2 E-string theory).
All top geometries in Table \ref{tb:rank2-classification} come from 6d SCFTs.
We also claim that all smooth rank 2 3-folds engineering 5d SCFTs belong to one of the RG-flow families exhibited in Table \ref{tb:rank2-classification}.
Therefore, we deduce the following conclusion: \emph{All rank 2 5d SCFTs realized by smooth non-compact 3-folds have 6d SCFT origins.}
This is one of the most important lessons from our classification of rank 2 5d SCFTs.
The same conclusion may hold also for singular geometries involving $\text{O7}^+$-planes. As mentioned earlier, the classification of smooth 3-folds misses a single geometry corresponding to the theory $SU(3)_{\frac{1}{2}} + 1\textbf{Sym}$, despite the fact that this theory is known to have a brane construction involving $\text{O7}^+$-planes \cite{Hayashi:2015vhy}. This theory may be the only rank 2 SCFT which cannot be engineered by a smooth 3-fold. But, we also know that this theory can be obtained from a KK theory with 6d origin, so we have found no counterexamples to the notion that all rank 2 5d SCFTs come from 6d SCFTs.
The above discussion motivates classifying automorphisms of 6d SCFTs which lead to 5d KK theories, as in \cite{Apruzzi:2017iqe}. Given the fact that 6d SCFTs are already classified (not counting frozen singularities involving $\text{O7}^+$ planes), the possible automorphisms can be deduced from symmetries of the tensor branch diagrams of 6d SCFTs dressed by gauge symmetries which respect the automorphisms.
\section*{Acknowledgements}
We would like to thank Ron Donagi, Hirotaka Hayashi, Sung-Soo Kim, Kimyeong Lee, Dave Morrison, Kantaro Ohmori and Gabi Zafrir for useful comments and discussions. We also like to thank SCGP summer workshop 2017 for hospitality during part of this work. The research of P.J. and H.K. and C.V. is supported in part by NSF grant PHY-1067976. S.K. is supported by NSF grant DMS-1502170.
|
\section{Introduction}
\label{sec:intro_mgiifeii}
Understanding the surface star-formation rate (\ssfr) and its redshift
dependence associated with quasar absorbers is of utmost importance
for measuring the star-formation rate density (SFRD) as a function of
cosmic time in a luminosity independent way
\citep{Wolfe2003ApJ...593..235W,Srianand2005MNRAS.362..549S,Rahmani2010MNRAS.409L..59R,Menard2011MNRAS.417..801M}
and to probe the Kennicutt-Schmidt law in low metallicity gas at high
redshifts \citep{Rahmani2010MNRAS.409L..59R}. Evidence for large scale
outflows are seen in nearly all star-forming galaxies in the local
Universe and out to z $\sim$ 6
\citep{Pettini2001ApJ...554..981P,Shapley2003ApJ...588...65S,Steidel2010ApJ...717..289S,Martin2012ApJ...760..127M,
Newman2012ApJ...761...43N,Lundgren2012ApJ...760...49L,Zhu2015ApJ...815...48Z}.
Wind detection rates in galaxies at intermediate and low redshifts are
found to be depending on the galaxy orientation with the outflow
geometry being consistent with a bi-conical flow
\citep{Martin2012ApJ...760..127M}. These winds are thought to be
responsible for enriching the intergalactic medium (IGM) and
circumgalactic medium (CGM) around galaxies
\citep{Schaye2001ApJ...559L...1S,Simcoe2012Natur.492...79S}. At high
redshifts, for isolated galaxies, gas inflowing rate is found to be
roughly comparable to the sum of the star formation rate and the
outflowing rate
\citep{Erb2008ApJ...674..151E,Seko2016ApJ...833...53S}. \par
At present, the best way to probe the low density outflowing gas is
to study the metal absorption lines they imprint in the spectra of
background luminous sources (at small impact parameters; $\rho < 10$
kpc) which trace the dynamic environment, i.e., gas inflows and
outflows in the outskirt of galaxies over cosmic time-line
\citep{Weisheit1978ApJ...219..829W,
Lanzetta1992ApJ...391...48L,Mo1996ApJ...469..589M,
Tinker2008ApJ...679.1218T,Chelouche2010ApJ...722.1821C,Bouche2012MNRAS.426..801B}.
In the local Universe winds are ubiquitous in galaxies having
\ssfr\ $ \ge 0.1\ \rm {M_\odot\ yr^{-1}\ kpc^{-2}}$
\citep[][]{Heckman2001ASPC..240..345H,Heckman2002ASPC..254..292H}. If
high \mgii equivalent width systems are associated with outflows, as
suggested by \citet{Nestor2011MNRAS.412.1559N} and
\citet{Bouche2012MNRAS.426..801B}, then the associated galaxy is
expected to have high \ssfr. Therefore, understanding the connection
between absorber properties and their associated galaxies (e.g.,
equivalent width vs \ssfr) is vital for understanding various
feedback processes that shape up the galaxy evolution.
Interestingly, the average \ssfr\ per absorber can be obtained using
spectral stacking exercise
\citep{Wild2007MNRAS.374..292W,Noterdaeme2010MNRAS.403..906N,
Menard2011MNRAS.417..801M} by making suitable assumptions related to
fibre losses. Note that, in spectroscopic surveys using fibres (e.g.,
3 and 2 arcsec fibres used in SDSS-DR7 and DR12 surveys) one not only
integrate light from the background quasar but also from all
foreground galaxies that happen to fall within the fibre. However,
because of various practical reasons it is highly probable that only a
part of the line emitting region of foreground galaxy may come inside
the fibre and this leads to the fiber loss. This loss will also have
redshift dependence. At low redshifts $0.4 < z < 1.3$, in a stacking
experiment of 3461 \mgii and 345 \caii\ absorbers,
\citet{Wild2007MNRAS.374..292W} have detected the \oii\ nebular
emission and measured an average star formation rate (SFR) of
$0.11-0.14~ {\rm M_\odot\ yr^{-1}}$ and $0.11-0.48~ {\rm
M_\odot\ yr^{-1}}$, respectively. Using the enlarged sample of \mgii
absorbers from SDSS-DR7, \citet{Noterdaeme2010MNRAS.403..906N} have
detected the average \oii\ luminosity of $\sim$ $1.4-5.1 \times
10^{40} \rm ~erg\ s^{-1}$ for strong \mgii\ absorbers (i.e., \ew\ $\ge
1$~\AA) at $0.5 < z < 0.7$. \citet{Menard2011MNRAS.417..801M} proposed
that \mgii absorbers recover the overall star formation history of the
universe and can be used as a new tool to probe the redshift evolution
of SFR in a luminosity independent manner.
\citet{Rahmani2010MNRAS.409L..59R} have stacked the damped-\lya
absorbers (DLA) with log \nhi $\ge 20.3$ to detect the average \lya
emission and set an upper limit on the contribution of DLA galaxies to
the cosmic SFRD of $\sim$ 0.13~ ${\rm M_\odot\ yr^{-1}\ Mpc^{-3}}$ at
$z \sim 3$ \citep[see
also,][]{Noterdaeme2014A&A...566A..24N,Joshi2017MNRAS.465..701J}.
However, as mentioned above physical parameters derived from fibre
spectra suffer from effect of fibre losses and one has to be mindful
of this while interpreting these results
\citep{Lopez2012MNRAS.419.3553L, Joshi2017MNRAS.471.1910J}.
Fibre losses can in principle be considered as negligible for systems
with large \ew\ because of the well known anticorrelation between
impact parameter ($\rho$) and \ew\ \citep{Bergeron1986A&A...155L...8B,
Steidel1995qal..conf..139S, Chen2010ApJ...714.1521C,
Nielsen2013ApJ...776..114N}. \citet{Lopez2012MNRAS.419.3553L} have
shown that for the fibre size used in SDSS-II (3 arcsec), galaxies
associated with more than 90\% of the systems will fall inside the
fibre for \ew $> 3$~\AA\ (see their figure 4). However, large scatter
in \ew\ vs $\rho$ relationship could increase the fibre loss. A strong
correlation is seen between the \mgii equivalent width (at low
spectral resolution is a good proxy to the velocity spread along the
line of sight) and galaxy color where a stronger \mgii absorbers tend
to be present in the vicinity of star-forming galaxies and most likely
to be associated with the outflows \citep{Zibetti2007ApJ...658..161Z,
Noterdaeme2010MNRAS.403..906N, Bordoloi2011ApJ...743...10B,
Lan2014ApJ...795...31L,Nielsen2016ApJ...818..171N}. The absorbing
gas traced by \mgii absorbers appear to have a bimodality in azimuthal
angle distribution where the cool dense gas is preferred to lie near
major and minor axes of galaxies \citep{Bouche2012MNRAS.426..801B,
Kacprzak2012ApJ...760L...7K}.
The \mgii absorbers with strong \feii\ absorption are likely to arise
from either very high metallicity sub-DLAs or high \nhi\ DLAs
\citep{Srianand1996ApJ...462..643S, Rao2006ApJ...636..610R}.
\citet{Rao2006ApJ...636..610R} have found that the detection rate of
DLAs in \mgii systems increases if one puts additional constrains
based on equivalent width ratios of Mg~{\sc ii}, \mgi and \feii
absorption. They detected DLAs with a success rate of $\sim$ 42 per
cent by selecting \mgii absorbers with strong \feii absorption (i.e.,
\rratio $\ge 0.5$) and \ewmgi\ $> 0.1$~\AA. In recent efforts to
detect cold gas in strong \mgii systems
\citet{Dutta2017MNRAS.465.4249D} have found a factor four times higher
detection rate of \hi 21-cm absorption in systems with \ewfeii\ $\ge
1$~\AA\ \citep[see also,][]{Gupta2012A&A...544A..21G}. Recently, we
have detected 198 \oii\ emitting galaxies associated with strong \mgii
absorbers in the spectra of SDSS \citep[][hereinafter refer as Paper
1]{Joshi2017MNRAS.471.1910J}. We have found that the \mgii absorbers
detected in \oii\ nebular emission (with \loii\ $> 2 \times 10^{40}
\rm ~erg\ s^{-1}$) typically have \ew\ $\ge$ 1~\AA, \mgii doublet
ratio ($DR = W_{\rm Mg~II \lambda2796}/W_{\rm Mg~II \lambda2803}$)
close to unity and \rrr $\ge$ 0.5. Therefore, given the above facts,
naively one would expect a strong dependence of average
\oii\ luminosity of \mgii absorbers in the stacked spectra of systems
with different \rrr\ parameter ranges. \par
There are two main motivations behind this work: (i) to study the
dependence of average \loii\ on \rrr\ and (ii) to understand the
effect of fibre size on different observed correlations between
\loii\ and other parameters. While the former allows us to probe the
nature of SFR in potential DLA candidates the latter will allow us to
probe the gas distribution at different scale around the star forming
regions probed by different samples of \mgii\ absorbers.
\begin{figure}
\centering
\epsfig{figure=dr7_vs_dr4.ps,height=8.5cm,width=8.5cm,angle=0}
\caption{\emph{Top left panel:} The signal-to-noise ratio
distribution of quasar spectra around the expected position of
\oii\ emission line in SDSS-DR7 and DR12 spectra. The cumulative
distribution of the \zabs\ (\emph{top right}), \ew (\emph{bottom
left}) and \rratio (\emph{bottom right}) of our \mgii systems
from SDSS-DR7 and SDSS-DR4 \mgii systems used in
\citet{Menard2011MNRAS.417..801M}.}
\label{fig:sample_comp}
\end{figure}
\begin{figure*}
\centering
\epsfig{figure=ew_vs_r.ps,height=12.5cm,width=13.5cm,angle=0}
\caption{\emph{Top panel:} The fraction of \mgii systems with $R \ge
0.5$ versus $z$ for two different \ew\ ranges. Vertical dotted
lines mark the redshift range of interest for this study.
\emph{Bottom right panel:} The dependence of \rrr\ parameter on
\ew\ for the \mgii systems. The corresponding fraction is shown as
\emph{stars} in the right side ordinate. For this plot we consider
only \mgii systems in the redshift range of our interest.
\emph{Bottom left panel:} \rrr\ as a function of redshift for
systems with \ew\ $\ge 1$~\AA. The solid blue circles show the
median value of \rrr\ in different redshift bins. The results are
for \mgii\ systems from SDSS-DR7 (see Fig~\ref{fig:fraction_dr12}
in Appendix~\ref{sec:apd1} for the results based on SDSS-DR12).}
\label{fig:fraction}
\end{figure*}
This article is organized as follows. Section 2 describes our sample
of absorbers used in this study. In Section 3, we present the spectral
staking analysis. In Section 4, we present various correlations seen
in the stacked spectra of quasars from SDSS-DR7 and DR12. Here, we
also discuss the dependence of \oii\ on \rrr\ parameter. Finally,
discussions and conclusions are presented in Section 5. Throughout, we
have assumed a flat cosmology with $H_0 =$ 70~\kms\ $\rm Mpc^{-1}$,
$\Omega_{\rm m} = 0.3$ and $\Omega_{\rm \Lambda} = 0.7$.
\section{Sample}
\label{sec:sample}
We have constructed a sample of strong \mgii absorbers, defined as
the ones with rest equivalent width \ew\ $\ge 1$~\AA\ (detected at
$\ge 4 \sigma$), by using the compilation of \mgii systems from the
\emph{expanded-version} of JHU-SDSS Metal Absorption Line
Catalog{\footnote{\href{http://www.pha.jhu.edu/$\sim$gz323/Site/Download\_Absorber\_Catalog\_files/fits/}
{http://www.pha.jhu.edu/$\sim$gz323/Site/}}
\citep[][]{Zhu2013ApJ...770..130Z}, compiled from the SDSS$-$DR7
and SDSS$-$DR12. For a fully saturated line this equivalent width
threshold corresponds to a velocity spread of $\sim 107$~\kms. We
select only systems with velocity offset of $>$ 5000~\kms\ with
respect to the quasar emission redshift and avoid sightlines
having broad absorption lines produced by quasar outflows (i.e.,
BALQSOs). To investigate the dependence of the average
\oii\ luminosity in the stacked spectra on \rrr\ parameter we
ensured that the \feiia line falls in the wavelength range of
higher completeness at $\lambda >$ 4000~\AA, i.e., $z \ge 0.55$
and redward of \lya emission of the quasar. In addition, we
restrict ourselves to $z \le 1.3$ (i.e. $\lambda \lesssim
8500$~\AA\ for \oiiab\ line) to avoid the \oii\ emission being
contaminated by most crowded telluric lines. For the above
redshift range, i.e., $ 0.55 \le z \le 1.3$, the fibres of 3 and 2
arcsec diameter used for the SDSS-DR7 and DR12 observations
project an angular size of $\sim 9.6-12.5$ kpc and $\sim 6.4-8.4$
kpc, respectively, in the sky. Our final sample consists of 10,083
and 12,116 \mgii systems from SDSS-DR7 and DR12, respectively.
\par
We note that our sample of \mgii systems based on SDSS-DR7 is
similar to the SDSS-DR4 sample used by
\citet{Menard2011MNRAS.417..801M} in terms of \zabs, \ew\ and
\rratio (see also, Fig.~\ref{fig:sample_comp}). A two-sided
Kolmogorov-Smirnov test (KS-test) finds no difference between the
two sub-samples based on \zabs, \ew\ and \rrr\ parameters with a
null probability of being drawn from the same parent distribution
to be $P_{KS} =$ 0.8, 0.8 and 0.7, respectively. The median
redshift probed in SDSS-DR7 and DR4 sample is \zabs\ $\sim$ 0.9.
However, SDSS-DR7 has twice the number of \mgii systems in the
sample of \citet{Menard2011MNRAS.417..801M}. The \mgii systems
detected in SDSS-DR12 also show a similar \ew\ distribution as
seen in SDSS-DR7 and DR4. However, the absorption redshifts
(\zabs) of \mgii systems in SDSS-DR12 are found to be slightly
lower with a median \zabs\ $\sim$ 0.8. One can note that the
average $SNR$, measured around the expected \oii\ nebular
emission, for the SDSS-DR7 spectra is slightly higher than the
SDSS-DR12 spectra (top left panel in Fig.~\ref{fig:sample_comp}).
The mean (respectively, median) \rrr\ parameter of \mgii systems
from SDSS-DR4, DR7 and DR12 is found to be $\sim$ 0.57 (0.56),
$\sim$ 0.56 (0.56), $\sim$ 0.55 (0.54), respectively.
In panel (a) of Fig.~\ref{fig:fraction} we show the evolution of
fraction of strong \mgii absorbers with \rrr\ $\ge 0.5$ as a
function of redshift. It is clear from the figure that over the
redshift range of our interest (i.e. $0.55 \le z \le 1.3$
identified with vertical dotted lines in Fig.~\ref{fig:fraction})
the fraction of absorbers with \rrr $\ge 0.5$ in the range $\rm
1~\AA\ \le$ \ew\ $< 2$~\AA\ decreases from 53\% to 44\%. However,
this fraction remains nearly constant at $\sim$ 70\% for the
systems with $\rm 2~\AA\ \le$ \ew\ $\le 6$~\AA\ over the same
redshift range (i.e., region between the two dotted lines in
Fig.~\ref{fig:fraction}). Even these strong systems show
decreasing trend when we consider the full observed $z$ range.
\par
\citet{Dey2015MNRAS.451.1806D} have found the median \rrr\ values
to decrease with increasing $z$ for \mgii systems detected in
SDSS-DR7. We see the same trend with \mgii\ systems detected in
SDSS-DR12 for the full sample (see Fig.~\ref{fig:fraction_dr12} in
Appendix~\ref{sec:apd1}). Based on their fit (see their figure 2)
we except {\rm {\ensuremath{\mathcal{R}}}} to decrease by 0.04
between redshift 0.5 and 1.3. For systems in the restricted
redshift range of our interest (i.e., $0.55 \le z \le 1.3$) we do
see a similar evolution of {\rm {\ensuremath{\mathcal{R}}}} with
redshift where {\rm {\ensuremath{\mathcal{R}}}} decreases by
$\sim$0.03 (see panel b of Fig.~\ref{fig:fraction}). They
interpreted this evolution to be due to evolution in the
metallicity ratio $\rm [Fe/Mg]$, most probably caused by the
cosmic evolution in the SNIa rates. In panel (c) of
Fig.~\ref{fig:fraction} we plot \rrr\ as a function of \ew\ for
all the systems in the $z$ range of our interest. The dependence
of \rrr\ on \ew\ is apparent as the fraction of systems with
\rrr\ $\ge 0.5$ increases from $\sim$ 50\% to $\sim$ 80\% for the
\ew\ range from 1~\AA\ to 3~\AA.
\begin{figure*}
\centering
\epsfig{figure=loii_profile_sdss_boss_all.ps,height=12.0cm,width=12.0cm,angle=0}
\caption{The \oii\ emission line profile seen in SDSS-DR7
(\emph{left}) and DR12 (\emph{right}) stacked spectra for various
\ew\ bins. The \emph{solid} line shows the best fit double
Gaussian to the data. The continuum-subtracted spectra are shifted
by a constant offset in luminosity for display purpose.}
\label{fig:profile}
\end{figure*}
\begin{figure*}
\centering
\epsfig{figure=ew_loii.ps,height=8.0cm,width=8.0cm,angle=0}
\epsfig{figure=ew_vs_loii_comp_menard.ps,height=8cm,width=8cm,angle=0}
\caption{\emph{Left panel:} The \oii\ luminosity of \mgii absorbers
for various \ew\ bins for SDSS-DR7 (\emph{red circles}) and DR12
(\emph{blue triangles}). \emph{Right panel:} The \oii\ luminosity
surface density (\sloii) as a function of \ew\ for SDSS-DR7
(\emph{red circles}) and DR12 (\emph{blue triangles}). The
\emph{dotted line } shows the best fit for \sloii\ versus
\ew\ from \citet{Menard2011MNRAS.417..801M} while shaded region
show the 1$\sigma$ uncertainty (see Appendix~\ref{sec:apd2} for
more discussion on this difference). The \emph{dashed} and
\emph{dot-dashed} line are the best-fitting power law for our
measurements based on SDSS-DR7 and DR12 data set.}
\label{fig:loiivsew}
\end{figure*}
\begin{figure}
\centering
\epsfig{figure=delvsigma_loii_sdss_boss.ps,height=8.5cm,width=8.5cm,angle=0,bbllx=20bp,bblly=145bp,bburx=589bp,bbury=712bp,clip=true}
\caption{\emph{Lower panel:} The distribution of $\sigma_{\rm [O
II]}$ as a function of \ew\ for SDSS-DR7 (\emph{diamonds}) and
SDSS-DR12 (\emph{stars}). \emph{Upper panel:} shows the \oii\ line
profile for various \ew\ bins for SDSS-DR7 and DR12.}
\label{fig:sigma_delv}
\end{figure}
\section{Analysis}
\label{sec:analysis}
To detect the \oii\ nebular emission from \mgii absorbers we have
constructed composite spectra, using continuum subtracted spectra and
median statistics. For this, we have shifted the individual spectrum
to the rest-frame of the \mgii absorber by conserving the flux and
rebinning on to a uniform rest wavelength grid as the original data
\citep{Bolton2012AJ....144..144B}. We modeled the local continuum by a
low order (typically a third order) polynomial fit, within the
proximity of \oiiab\ line (i.e., rest wavelength range of 3700 $-$
3750~\AA). While stacking, we have masked the absorption line features
originating from systems at other redshifts as well as the sky
emission lines in the spectrum. The 1$\sigma$ flux uncertainty in each
pixel in the stacked spectrum is estimated from the central interval
encompassing 68\% of the flux distribution of the corresponding pixel
as in \citet{Joshi2017MNRAS.465..701J}.
Recall that the physical area corresponding to a given angular
aperture varies with redshift. As suggested by
\citet{Menard2011MNRAS.417..801M}, to account for the fibre effects,
we also generate the stacked spectrum of \oii\ luminosity surface
density (\sloii). To estimate the \sloii, we convert each spectrum
into luminosity units, at the redshift of the absorber and divide by
the projected surface area of the fibre at the absorber redshift
before co-adding them to get the composite spectrum that we will call
\sloii.
As expected we detect \oii\ emission in most of our composite spectra.
In each composite spectrum we model the observed \oii\ emission line
with a double Gaussian profile with a tied linewidth but freely
varying the line ratio in a range of 3.4-1.5{\footnote{The {\sc [O
II]$\lambda3729/\lambda3727$} intensity ratio in the range
3.4-1.5 is predicted in photoionization models for the electron
density in the range ${\it n_e = \rm 10^1-10^5 cm^{-3}}$ for the
kinetic temperature $T = 10,000 \rm K$
\citep{Osterbrock2006agna.book.....O}.}}, allowing for the typical
range in the electron density of the gas under photoionization
equilibrium. Gaussian fitting is mainly used to verify any possible
dependences of intensity ratio of \oii\ doublet and its FWHM with
\mgii absorption line properties. However, we simply integrate the
stacked spectrum over the central 12 pixles (i.e., $\sim$800~\kms) for
measuring the \oii\ line luminosity (or \sloii) and the respective
error is computed by propagating the flux uncertainty in each pixel.
We have also generated geometric mean composite spectra to study the
effect of \rrr\ parameter on the average reddening induced by \mgii
absorbers
\citep[][]{York2006MNRAS.367..945Y,Khare2012MNRAS.419.1028K}. For
this, we have shifted each spectrum to the absorber rest frame and
computed their geometric mean. Here, we do not use the normalized or
continuum subtracted spectrum to preserve the average continuum shape
which is important for determining the characteristic extinction law
\citep[see also,][]{York2006MNRAS.367..945Y}. To compute the relative
extinction we have generated the stacked spectra for a control sample
of quasars, within $\Delta z = \pm 0.05$ of \zem and $\Delta r_{\rm
mag} = \pm 0.5$ of $r_{\rm mag}$, without absorbers in their
spectra. We will discuss the results of this analysis in
Section~\ref{sub:red}.
\section{Results}
We generate several composite spectra in various \ew\ and redshift
bins. Since, we are interested in measuring the average \oii\ nebular
emission line luminosity (or surface brightness) associated with the
\mgii absorbers, the direct detection of \oii\ nebular emission
reported in Paper 1 are also included in most of our analysis. We also
present the results when these systems are excluded from the analysis.
\subsection{Average \oii\ emission and fibre effects}
\label{sub:loiivsew}
To quantify the fibre loss effect we have constructed composite
spectra by dividing our sample into five \ew\ bins of $1-1.5,
1.5-2.0,2.0-2.5, 2.5-3.0$ and $\ge 3$~\AA\ for SDSS-DR7 and DR12. The
stacked profiles for various \ew\ bins and the number of systems used
to construct the stacked spectra are shown in Fig.~\ref{fig:profile}.
It is apparent that for a given equivalent width bin the \oii\ line
luminosity found for SDSS-DR7 is higher than that of SDSS-DR12. In
the \emph{left panel} of Fig.~\ref{fig:loiivsew}, we show the median
\oii\ line luminosity as a function of \ew\ for the \mgii systems in
SDSS-DR7 (\emph{circles}) and DR12 (\emph{triangles}). We find a
clear increasing trend of \loii\ with increasing \ew. This is
consistent with the trend found by
\citet{Noterdaeme2010MNRAS.403..906N} in a stacking analysis of \mgii
systems they have found from SDSS-DR7 quasars. We note that median
\loii\ we detect in SDSS-DR7 sample is similar to that obtained by
\citet[][see also their table 5]{Noterdaeme2010MNRAS.403..906N}.
In \emph{right panel} of Fig.~\ref{fig:loiivsew}, we show \sloii\ as a
function of \ew. We confirm the strong correlation between \ew\ and
\sloii\ in both the SDSS-DR7 (\emph{circles}) and DR12
(\emph{triangles}) datasets. The \sloii\ measurement for SDSS-DR7 and
DR12 datasets for each \ew\ bin are listed in column 4 and 7 of
Table~\ref{tab:sigloii}, respectively. We first compare our results
based on SDSS-DR7 stacked spectra with that of
\citet{Menard2011MNRAS.417..801M}, based on SDSS-DR4, as spectra were
obtained using 3 arcsec fibre in both cases. Note that our redshift
range of $0.55 \le z \le 1.3 $ is slightly different from that of
\citet{Menard2011MNRAS.417..801M} of $0.36 \le z \le 1.3$. As
suggested by \citet{Menard2011MNRAS.417..801M} we model this
relationship using a powerlaw of the form $\langle$\sloii$\rangle$ =
$A W_0^{\alpha}$ and compute the best fit parameters of $\rm \alpha =
1.69 \pm 0.11$ and $A=$ $ \rm (2.55 \pm 0.27) \times 10^{37} \rm
erg\ s^{-1}\ kpc^{-2}$ for SDSS-DR7. While comparing our best fit
parameters with \citet[][]{Menard2011MNRAS.417..801M}, i.e., $\rm
\alpha = 1.75 \pm 0.11$ and $A=$ $ \rm (1.48 \pm 0.18) \times 10^{37}
\rm erg\ s^{-1}\ kpc^{-2}$, the $\alpha$ is found to be same whereas
normalization factor is found to be different (i.e. $\sim$ 1.7 times
higher) at a significance level of $\sim 3.3\sigma$. We note that this
difference in the normalization factor is mainly due to difference in
the way quasar continuum is modelled.
\citet[][]{Menard2011MNRAS.417..801M} have modelled the continuum with
an iterative running median of sizes ranging from 500 to 15 pixels.
This basically smooths all small scale fluctuations. In fact, we get
the similar values of \sloii\ if we use the continuum fitting
procedure adopted by \citet[][]{Menard2011MNRAS.417..801M} [see
Fig.~\ref{fig:loiivsew_menard} and related discussions in the
Appendix]. However, for rest of the paper we will present result
from data using our continuum fitting procedure.
\par
\begin{table*}
\centering
\begin{minipage}{150mm}
{\small
\caption{\oii\ luminosity surface density traced by \mgii absorbers
as a function of equivalent width (\ew).}
\label{tab:sigloii}
\begin{tabular}{@{} c ccccc c @{}}
\hline \hline
\multicolumn{1}{c}{\ew} &\multicolumn{3}{c}{\sloii\ for SDSS-DR7} &\multicolumn{3}{c}{\sloii\ for SDSS-DR12} \\
\multicolumn{1}{c}{interval} &\multicolumn{1}{c}{$\langle$\ew $\rangle$} &\multicolumn{1}{c}{$\langle$ $z$ $\rangle$} &\multicolumn{1}{c}{\sloii} &\multicolumn{1}{c}{$\langle$\ew $\rangle$} &\multicolumn{1}{c}{$\langle$ $z$ $\rangle$} &\multicolumn{1}{c}{\sloii} \\
\multicolumn{1}{c}{ (\AA) } &\multicolumn{1}{c}{ (\AA) } &\multicolumn{1}{c}{} &\multicolumn{1}{c}{($\rm \times 10^{38} erg\ s^{-1}\ kcp^{-2}$)} &\multicolumn{1}{c}{(\AA)} &\multicolumn{1}{c}{} &\multicolumn{1}{c}{($\rm \times 10^{38} erg\ s^{-1}\ kcp^{-2}$)} \\
\hline
1.0~\AA$ \le $ \ew $ < 1.5$~\AA\ & 1.23 & 0.93 & $0.28 \pm 0.04 $ & 1.23 & 0.89 & $0.37 \pm0.04 $ \\
1.5~\AA$ \le $ \ew $ < 2.0$~\AA\ & 1.71 & 0.94 & $0.68 \pm 0.06 $ & 1.73 & 0.90 & $0.49 \pm0.05 $ \\
2.0~\AA$ \le $ \ew $ < 2.5$~\AA\ & 2.22 & 0.95 & $1.07 \pm 0.09 $ & 2.23 & 0.91 & $0.99 \pm0.08 $ \\
3.0~\AA$ \le $ \ew $ < 3.5$~\AA\ & 2.73 & 0.95 & $1.71 \pm 0.13 $ & 2.72 & 0.92 & $1.36 \pm0.11 $ \\
3.5~\AA$ \le $ \ew $ < 6.0$~\AA\ & 3.62 & 0.97 & $1.99 \pm 0.15 $ & 3.66 & 0.91 & $1.33 \pm0.12 $ \\
\hline
\end{tabular}
}
\end{minipage}
\end{table*}
It is clear from the figure that \sloii\ measured from the SDSS-DR7
data are higher than those from SDSS-DR12 data (apart from the first
equivalent width bin). The difference in 3 out of the 5 equivalent
width bin considered here is more than 2$\sigma$ (see column 4 and 7
of Table~\ref{tab:sigloii}). The best fitted relationship between
\sloii\ and \ew\ for the SDSS-DR12 data are shown in \emph{dot-dashed}
line (i.e., $\langle$\sloii$\rangle$ = $A W_0^{\alpha}$ with $\rm
\alpha = 1.28 \pm 0.11$ and $A=$ $ \rm (2.95 \pm 0.29) \times 10^{37}
\rm erg\ s^{-1}\ kpc^{-2}$) in the figure. While this fit is not as
good as that for DR7 data, what is evident is that the fit to DR12
data is consistently lower than that for the DR7 data in all but one
equivalent width bin. This clearly confirms that the observed
\oii\ surface brightness is affected by fibre size effects even after
normalizing the flux with the projected fibre size. In particular our
results are consistent with the fact that (1) there is a large scatter
in the \ew\ vs impact parameter relationship even at high equivalent
widths (see figure 1 of
\citealt{Nielsen2013ApJ...776..114N,Lan2014ApJ...795...31L,Huang2016MNRAS.455.1713H})
and (2) gradients are known to be present in the surface brightness
profiles of galaxies which results in lower measured
\oii\ luminosities when smaller fibres probe predominately the outer
parts of galaxies.
\begin{figure}
\centering
\epsfig{figure=scatter.ps,height=8.5cm,width=8.5cm,angle=0}
\caption{The \mgii doublet ratio versus \rrr\ parameter of \mgii
absorbers with nebular emission line detected at $\ge 4\sigma$
(\emph{stars} from Paper 1), $3-4\sigma$ (\emph{circles}) and
without (i.e., $< 3\sigma$) nebular emission (\emph{squares}) line
detection for \ew\ $> 1$~\AA\ in SDSS-DR7. The upper and
right-hand panels show, respectively, the $DR$ and
\rrr\ distributions for the \mgii systems with nebular emission
line detected at $\ge 4\sigma$ (shaded with blue slanted lines at
$-45^{\circ}$), $3-4\sigma$ (shaded with red slanted lines at
$45^{\circ}$) and without (i.e., $< 3\sigma$) nebular emission
(unfilled histogram). }
\label{fig:scatter}
\end{figure}
In the lower panel of Fig.~\ref{fig:sigma_delv} we also show the
dependence of velocity width (deconvolved for the instrumental
broadening) of \oii\ line ($\sigma_{\rm[O II]}$) on the \ew\ of
\mgii\ systems. A clear increasing trend of $\sigma_{\rm[O II]}$ with
\ew\ is apparent from the figure. The $\sigma_{\rm[O II]}$ increases
from $\sim$ 90~\kms\ to $\sim$ 200~\kms\ when the average
\ew\ increases from 1~\AA\ to 3~\AA\ in both SDSS-DR7 and DR12. The
\oii\ line profiles for each \ew\ bin are shown in the upper panel of
Fig.~\ref{fig:sigma_delv}. One can see that emission line peaks for
both the \oiiab\ doublet components are clearly visible in the stacked
profile of lower \ew\ bins, i.e., $< 2.5$~\AA. However, \oiiab\ line
is blended for the strong \mgii systems with \ew\ $> 2.5$~\AA.
Typically $\sigma$ of any emission line from a galaxy can be a good
probe of the underlying mass. However, in the stacked spectrum it can
also be a reflection of spread in the difference between emission and
absorption redshifts. Therefore, a correlation between \ew\ and
$\sigma_{\rm[O II]}$ could either reflects velocity offset ($\Delta
v$) between emission and absorption increasing with \ew\ and/or high
\ew\ systems originating from massive halos. However, in Paper 1 when
we considered the direct \oii\ detections we do not find any
correlation between \ew\ vs $\sigma_{\rm[O II]}$ as well as \ew\ vs
$\Delta v$ (relative velocity between absorption and emission
redshift). It is also clear from Fig.~\ref{fig:sigma_delv}, that
$\sigma_{\rm[O II]}$ measured for a given \ew\ bin matches very well
between SDSS-DR7 and SDSS-DR12. In addition, we do not find any
dependence of \oiiab\ doublet ratio with \ew.
It is clear from the above discussions that \loii\ and
\sloii\ measured in the SDSS-DR12 composite are under estimated.
Therefore, to minimize the fibre size effects, in most discussions
that follows we concentrate on results based on DR7.
\subsection{Dependence of \oii\ emission on \rratio}
\label{sec:rpara}
In Paper 1, we have shown that for a given luminosity threshold
(i.e. \loii), direct \oii\ nebular line detection fraction
increases with increasing \ew. We have considered \oii\ detections
with more than 4$\sigma$ significance level in that study. Here,
before doing the stacking analysis, we also identify systems with
tentative emission feature (hereinafter, candidate \oii\ emitters)
at the expected position of \oii\ at $ 3 < \sigma < 4$
level{\footnote{Please refer to paper 1 for how we compute the
significance level for the \oii\ feature.}}. Note that by
lowering the significance level to confirm a detection, we might
have enhanced the number of false positives. In
Fig.~\ref{fig:scatter}, we compare the distribution of these \mgii
systems with (\emph{circles}) and without (\emph{squares}) emission
feature in the {\rm {\ensuremath{\mathcal{R}}}} vs $DR$ plane
(systems discussed in Paper 1 marked with star symbol).
Interestingly, like the trend shown by firm detections, most of the
\mgii systems with consistent features (at $3-4 \sigma$ level) at
the location of \oii\ nebular line also have {\rm
{\ensuremath{\mathcal{R}}}} $\ge 0.5$ and $DR$ $\sim$ 1 (see
also, figure 8 in Paper 1). In Fig.~\ref{fig:scatter}, we also plot
the histogram of {\rm {\ensuremath{\mathcal{R}}}} and $DR$
distribution in the right and upper panel respectively. The
sub-samples of ``candidate \oii\ emitters'' and systems without
nebular emission are found to be drawn from different distribution
of {\rm {\ensuremath{\mathcal{R}}}} with $KS-test$ null probability
of $p_{null} = 0.004$. It again indicates that the luminosity of
\oii\ emission in the stacked spectrum will depend on {\rm
{\ensuremath{\mathcal{R}}}} parameter.
\begin{figure*}
\centering
\epsfig{figure=sigoii.ps,height=10cm,width=12.0cm,bbllx=37bp,bblly=151bp,bburx=545bp,bbury=517bp,clip=true}
\caption{\oii\ luminosity (\loii) in the stacked
spectra obtained for three bins of \rratio\ for strong (\ew $>
1$~\AA) \mgii absorbers detected in SDSS-DR7 spectra. \emph{Right
panel:} The same for the \mgii absorber detected in SDSS-DR12
spectra. Total number of \mgii\ systems involved in each bin is
also given in the figure.}
\label{fig:stackfeiiall}
\end{figure*}
\begin{table*}
\centering
\begin{minipage}{110mm}
{\scriptsize
\caption{The \loii\ for the subset based on \rrr\ parameter.}
\label{tab:stat}
\begin{tabular}{@{} c ccccc c @{}}
\hline \hline
\multicolumn{1}{c}{Criteria} &\multicolumn{5}{c}{SDSS-DR7} \\
&\multicolumn{1}{c}{No} &\multicolumn{1}{c}{$\langle$\ew $\rangle$} &\multicolumn{1}{c}{$\langle z \rangle$} &\multicolumn{1}{c}{\loii}&\multicolumn{1}{c}{\loii$^P$\textcolor{blue}{$^a$}} \\
&\multicolumn{1}{c}{ } &\multicolumn{1}{c}{(\AA)} &\multicolumn{1}{c}{ } &\multicolumn{1}{c}{($\rm \times 10^{40}~erg~s^{-1}$)} &\multicolumn{1}{c}{($\rm \times 10^{40}~erg~s^{-1}$)}\\
\hline
1~\AA$ \le $ \ew $ < 2$~\AA\ & & & & & \\
\rrr $\le 0.44$ & 2871 & 1.36 & 0.95 & $0.99 \pm 0.26$ & $1.66 $\\
$0.44 < $ \rrr $\le 0.64$ & 2169 & 1.42 & 0.93 & $1.48 \pm 0.28$ & $1.79 $\\
\rrr $> 0.64$ & 2291 & 1.43 & 0.92 & $2.76 \pm 0.26$ & $1.82 $\\
& & & & & \\
\hline
2~\AA$ \le $ \ew $ < 3$~\AA\ & & & & & \\
\rrr $\le 0.44$ &450 & 2.35 & 0.95 & $1.52 \pm 0.69$ & $4.40 $ \\
$0.44 < $ \rrr $\le 0.64$ &675 & 2.41 & 0.95 & $4.29 \pm 0.59$ & $4.60 $ \\
\rrr $> 0.64$ &801 & 2.41 & 0.94 & $7.94 \pm 0.48$ & $4.60 $ \\
\hline
\end{tabular}
}
\\
\textcolor{blue}{$^a$}{ expected \oii\ luminosity measured from the \ew\ vs \loii\ \\ correlation for the average \ew\ per bin.}
\end{minipage}
\end{table*}
\begin{figure*}
\centering
\epsfig{figure=ew_loii_Rbased.ps,height=8.3cm,width=8.3cm,angle=0}
\epsfig{figure=ew_sigloii_Rbased.ps,height=8cm,width=8cm,angle=0}
\caption{\emph{Left panel:} The \oii\ luminosity of \mgii systems
detected in SDSS-DR7 with $R \ge 0.5$ (\emph{circles}) and $R < 0.5$
(\emph{triangles}) for various \ew\ bins. The number of systems
used for the stack and the stacked profiles are shown. The
\loii\ for all the systems for different \ew\ bins are shown as
\emph{squares}. \emph{Right panel:} The \oii\ luminosity surface
density (\sloii) as a function of \ew. The symbols are as in left
panel. (\emph{circles}). }
\label{fig:loiivsew_r}
\end{figure*}
\begin{figure}
\centering
\epsfig{figure=frac_gt3sig_vs_ew_new.ps,height=7.8cm,width=7.8cm,angle=0}
\caption{The fraction of \mgii systems with \oii\ nebular emission
detected at $\ge 3\sigma$ in SDSS-DR7 (\emph{circles}) and DR12
(\emph{triangles}). The \emph{dotted} and \emph{dashed} lines show
the expected fraction of \mgii\ systems detected within a
projected fibre radius of 10 and 7 kpc at a redshift of 0.6, from
\ew\ vs $\rho$ distribution of confirmed \mgii galaxies by
\citet{Nielsen2013ApJ...776..114N}, given in the right side
ordinates. }
\label{fig:loiivsfrac}
\end{figure}
To explore the dependence of average \loii\ and \sloii\ emission on
{\rm {\ensuremath{\mathcal{R}}}} parameter we generate the composite
spectra based on {\rm {\ensuremath{\mathcal{R}}}} by dividing our
sample into three bins of {\rm {\ensuremath{\mathcal{R}}}} $\le
0.44$, $0.44 < $ {\rm {\ensuremath{\mathcal{R}}}} $\le 0.64$ and {\rm
{\ensuremath{\mathcal{R}}}} $> 0.64$, having almost equal number of
systems in each bin. In Fig.~\ref{fig:stackfeiiall}, we show the
median \oii\ emission line profile in sub-samples based on
\rrr\ parameter for SDSS-DR7. For comparison, we also show the
stacked profiles from SDSS-DR12. It is apparent that the strength of
\oii\ emission is higher for the systems with higher \rrr\ parameter.
This trend is consistent with the fact that when \oii\ nebular
emission is detected (with $\ge 3 \sigma$ level of significance) in
individual cases one finds \rrr\ values to be higher (see
Fig.~\ref{fig:scatter}).
In order to decipher the effect of \rrr\ (without being affected by
\ew\ vs \loii\ correlations) on the strength of \oii\ luminosity (or
\sloii) we restrict ourselves to two narrow \ew\ range of $\rm 1~\AA
\le $\ew$ < 2$~\AA\ and $\rm 2~\AA \le $\ew$ < 3$~\AA, having
sufficient number of \mgii systems with \rrr $< 0.5$ (see
Fig~\ref{fig:fraction}). Further, we divide each of them in to three
bins of \rrr\ $\le 0.44$, $0.44 < $ \rrr\ $\le 0.64$ and \rrr\ $>
0.64$. The details of these sub-samples are given in
Table~\ref{tab:stat}. The number of systems in each sub-sample,
average \ew\ and \zabs\ are given in column 2, 3 and 4 of this table,
respectively. It is clear that these quantities do not differ by a
wide margin between the sub-samples. However, we clearly see an
increasing trend in \loii\ with \rrr.
A two-sided Kolmogorov-Smirnov test (KS − test) shows that the
\ew\ distribution of two sub-samples with higher {\rm
{\ensuremath{\mathcal{R}}}} values, i.e., $0.44 < $ {\rm
{\ensuremath{\mathcal{R}}}} $\le 0.64$ and {\rm
{\ensuremath{\mathcal{R}}}} $> 0.64$, are drawn from a parent
distribution with a null probability of being drawn from same parent
distribution of $P_{KS} =$ 0.79 and $P_{KS} =$ 0.87 for both
\ew\ bins. The systems with {\rm {\ensuremath{\mathcal{R}}}} $\le
0.44$ show slightly lower \ew. The average \ew\ is lower by 4\% in
this case (i.e., {\rm {\ensuremath{\mathcal{R}}}} $\le 0.44$) compare
to other two cases. We find that the probability for sub-samples with
{\rm {\ensuremath{\mathcal{R}}}} $\le 0.44$ and {\rm
{\ensuremath{\mathcal{R}}}} $> 0.64$ are drawn from two different
populations is $P_{KS} = 1.7 \times 10^{-11}$ (respectively, 0.03) for
the systems having $\rm 1~\AA \le $\ew$ < 2$~\AA\ (respectively, $\rm
2~\AA \le $\ew$ < 3$~\AA).
Next, we ask whether difference in \loii\ between \rrr\ $\le 0.44$ and
\rrr\ $> 0.64$ comes from the previously discussed correlation between
\ew\ and \loii. Based on our \loii\ vs \ew\ best fit parameters (i.e.,
$A W_0^{\alpha}$ with $\rm \alpha = 0.97 \pm 0.07$ and $A=$ $ \rm
(1.78 \pm 0.08) \times 10^{40} \rm erg\ s^{-1}$) we compute the
expected \loii\ for the mean \ew\ probed by the above sub-samples and
is given in column 6 of Table~\ref{tab:stat}. It is clear from this
table that purely based on the \ew\ vs \loii\ we do not expect
\loii\ to be very different between different sub-samples.
Interestingly, for a fixed \ew\ range of $\rm 1~\AA \le $\ew$ <
2$~\AA\ a clear difference in \loii\ of factor 2.8 (significant at 3.6
$\sigma$) is apparent for the subset having different \rrr\ parameter
of \rrr\ $\le 0.44$ and \rrr\ $> 0.64$. The difference in \loii\ is
found to be even higher of about factor 5.7 (significant at
4.8$\sigma$) if we consider the systems with $\rm 2~\AA \le $\ew$ <
3$~\AA.
\emph{The discussions presented here confirm that \loii\ obtained in
the stacked spectra depends strongly on \rrr} \emph {parameter. }
\subsection{Dependence of \loii\ versus \ew\ based on \rrr}
\label{sub:loiivsr}
In this section we explore the effect of \rrr\ on \loii\ vs
\ew\ relation using SDSS-DR7 data. For this, we generate composite
spectra for two subsets with \rrr\ $\ge 0.5$ and \rrr\ $< 0.5$ for
various \ew\ bins. In the left panel Fig.~\ref{fig:loiivsew_r}, we
plot the \loii\ vs \ew. We also show the \oii\ profile for each
\ew\ bin for two ranges in \rrr. It is clear from the figure that
the \oii\ emission has systematically higher luminosity in the case
of \rrr\ $\ge 0.5$ in each \ew\ bin which is consistent with our
findings discussed above. For the subset with \rrr\ $< 0.5$ we note
that even for the lower \ew\ bins, where there are a good number of
systems available for the stacking, either the emission is
significantly weaker or is not detected. \par
In right panel of Fig.~\ref{fig:loiivsew_r}, we show that the median
surface luminosity density for the systems with \rrr\ $\ge 0.5$
(\emph{open circles}) is higher than those of all \mgii systems
(\emph{solid circles}). It is clear from this figure that the
correlation, we as well \citet[][]{Menard2011MNRAS.417..801M}, found
between \loii\ and \ew\ in the whole \mgii sample is mainly dominated
by systems showing strong Fe{\sc~ii}. For systems with \rrr\ $< 0.5$
no clear trend between \ew\ and \loii\ is visible and we could measure
only upper limits in several \ew\ bins.
Note, contrary to the correlation seen between \ew\ and \loii\ (or
\sloii) in the stacked spectra individual detections do not follow
this correlation. In Paper 1, we suggested that the \ew\ vs
\loii\ correlation may come from increase in \oii\ detection fraction
with increasing \ew. We explore this point further here.
In Fig.~\ref{fig:loiivsfrac}, we show the fraction of \mgii systems
with an emission feature detected at $\rm \ge 3\sigma$ at the expected
location of \oii\ doublet, having \ew\ threshold of $\ge 1$~\AA\ and
\loii\ $\ge 0.6$~\lsoii. A similar plot for \mgii systems with
\oii\ nebular emission detected at $\ge 4 \sigma$ ($\sim$ 198 systems)
are presented in Paper 1. Interestingly, we find a clear increasing
trend between the fraction of \mgii systems with \oii\ nebular
emission and \ew\ even among the tentative detections. A similar trend
in also seen when we consider the fraction of systems with detection
threshold of $\ge 2\sigma$ as well. This can be naturally explained
with the known anti-correlation between \ew\ and $\rho$ where the
galaxies at lower projected distances (i.e., impact parameters)
produce on an average stronger \mgii absorption and higher probability
of the associated \oii\ emission falling inside the fibre. The
fraction further increases if we put an additional constraint of
\rrr\ $\ge 0.5$. The reason being, as discussed in
Fig.~\ref{fig:scatter} most of the direct detections have \rrr\ $\ge
0.5$. \par
Using the impact parameter distribution of 183 spectroscopically
confirmed \mgii galaxies from the compilation of
\citet{Nielsen2013ApJ...776..114N} we compute the detection fraction
of \mgii\ systems within the impact parameter of $\sim 10$ and $\sim
7$ kpc for SDSS-DR7 and DR12, i.e., the projected radius of 3 and 2
arcsec fibres at $z=0.6$. For this, we measure the average $\rho$
and its standard deviation from the $\rho$ distribution of \mgii
systems for various \ew\ bins. For each \ew\ bin we compute the
probability of galaxy to come inside the fibre by randomly
generating 10,000 values of $\rho$ by assuming a Gaussian
distribution and considering $1 \sigma$ error over each measurement.
In Fig.~\ref{fig:loiivsfrac}, we show the expected fraction of \mgii
systems in SDSS-DR7 (\emph{dotted line}) and DR12 (\emph{dashed
line}) which roughly follow the observed trend in our systems.
This once again reiterates the importance of anticorrelation between
\ew\ vs $\rho$ and the fibre losses in deriving the correlations
seen in the stacked spectra. However, it is important to note that
the sample of \citet{Nielsen2013ApJ...776..114N} does not contain
enough systems at $\rho\le10$ kpc. As discussed before at such
impact parameters the relationship between \ew\ and $\rho$ may not
be similar to what we see at high $\rho$ values. Thus it is very
important to quantify the extent and nature of star forming regions
associated with strong Mg~{\sc ii} systems at low impact parameter
through direct observations.
\par
Our results also suggest that most of the \mgii absorbers with
\rrr\ $\ge 0.5$ should have systematically lower impact parameter (or
higher \loii\ and hence higher star formation rate) compared to those
with \rrr\ $< 0.5$. We could not check this with the existing data of
\citet{Nielsen2013ApJ...776..114N} as \feiia measurements are not
available for most of the systems.
\subsection{ Dependence of \loii\ versus $z$ based on \rrr}
\label{sub:loiivsz}
The strong correlation between \sloii\ and \ew\ is also found to be
evolving with redshift in the sense that a system with a given
\ew\ seems to be associated with larger \loii\ at high redshift
compared to that at low redshift \citep{Menard2011MNRAS.417..801M}.
Such a redshift evolution of \loii\ is also seen among direct
detections discussed in Paper 1. Here, to explore the dependence of
\loii\ as a function of redshift based on \rrr\ parameter we make two
subsets with \ew\ bins of 1~\AA\ $\le$ \ew $<2$~\AA\ and 2~\AA\ $\le$
\ew $<6$~\AA. We further divide each subset in to three redshift bins
of $0.55 \le z < 0.75$, $0.75 \le z < 0.95$ and $0.95 \le z < 1.3$,
respectively. For these redshifts ranges, the fibre of 3 arcsec
diameter used in SDSS-DR7 projects an angular size of $6.4-7.3$ kpc,
$7.3-7.9$ kpc and $7.9-8.4$ kpc, respectively, in the sky. \par
In Fig.~\ref{fig:loiivsz}, we show the \loii\ versus redshift as a
function of \rrr\ for the subset with 1~\AA\ $\le$ \ew\ $<
2$~\AA\ (\emph{top left panel}) and 2~\AA\ $\le$ \ew\ $<
6$~\AA\ (\emph{top right panel}). It is clear from the figure that the
median \loii\ of \mgii systems is higher at high redshifts. Here also,
we note that the systems with \rrr\ $< 0.5$ show very less emission at
each redshift bin, albeit having similar number of systems as in the
sub-sample of $\rrr$ $\ge 0.5$. The average \loii\ probed in the
stacked spectra corresponds to sub-\lsoii\ [with log \lsoii $\rm
(erg\ s^{-1}) = 41.60$ at the median $z$ of 0.65] galaxies with
\oii\ luminosity of $\sim 0.01$\lsoii\ and $\sim 0.1$\lsoii, for the
systems with \ew\ ranging from 1~\AA\ $\le$ \ew $<2$~\AA\ and
2~\AA\ $\le$ \ew $<6$~\AA, respectively. The \emph{dashed curve} in
Fig.~\ref{fig:loiivsz} shows the expected luminosity of $0.1$ and
$0.03$ \lsoii\ galaxy as a function of redshift using the redshift
evolution of field galaxies luminosity function by \citet[][see their
Table 7]{Comparat2016MNRAS.461.1076C}. Note that, these average
luminosities are smaller than the direct detection of \oii\ emission
associated to individual \mgii systems (see figure 9, 10 of Paper 1)}.
It is clear from Fig.~\ref{fig:loiivsz} that the increase in
\loii\ associated with \mgii systems roughly follows the luminosity
evolution of field galaxies and is mostly due to the systems with
\rrr\ $\ge 0.5$.
\begin{table}
\centering
{\scriptsize
\caption{The best-fit parameters for the \sloii\ as a function of redshift, $A (1+z)^{\alpha}$.}
\label{tab:sample}
\begin{tabular}{@{} c c c c @{}}
\hline \hline
\multicolumn{1}{c}{\ew} &\multicolumn{2}{c}{SDSS-DR7 with $R \ge 0.5$} \\
\multicolumn{1}{c}{(\AA)} &\multicolumn{1}{c}{A} & \multicolumn{1}{c}{$\alpha$} \\
\hline
1~\AA\ $\le$ \ew\ $<$ 2~\AA\ & 0.005$\pm$0.003 & 0.01$\pm$0.84 \\
2~\AA\ $\le$ \ew\ $<$ 6~\AA\ & 0.010$\pm$0.002 & 0.98$\pm$0.39 \\
\hline
\end{tabular}
}
\end{table}
Next, to account for the effect of increasing fibre size with redshift
we also plot \sloii\ as a function of redshift for two equivalent
width bins of 1~\AA\ $\le$ \ew $<2$~\AA\ (\emph{lower left panel}) and
2~\AA\ $\le$ \ew $<6$~\AA\ (\emph{lower right panel}). For
1~\AA\ $\le$ \ew $<2$~\AA\ the \sloii\ seem to be constant with
redshift. However, a mild increase in \sloii\ with redshift is seen
for the subset of 2~\AA\ $\le$ \ew $<6$~\AA. We model \sloii\ vs $z$
with a power-law of the form $A (1+z)^{\alpha}$. The best fit
parameters, i.e., normalization and slope for above two \ew\ bins are
given in column 2 and 3 of Table~\ref{tab:sample}, respectively. We
note that the evolution of \loii\ with redshift is similar (within
$1\sigma$) for both the subsets (see column 3 of
Table~\ref{tab:sample}).
\begin{figure*}
\centering
\epsfig{figure=z_loii_Rbased_profile_EW1_2.ps,height=7.0cm,width=7.0cm,angle=0}
\epsfig{figure=z_loii_Rbased_profile_EW2_6.ps,height=7cm,width=7cm,angle=0}
\epsfig{figure=z_sigloii_Rbased_profile_EW1_2.ps,height=7.0cm,width=7.0cm,angle=0}
\epsfig{figure=z_sigloii_Rbased_profile_EW2_6.ps,height=7cm,width=7cm,angle=0}
\caption{\emph{Top left panel:} The \oii\ luminosity as a function
of \zabs\ for a subset of \mgii systems from SDSS-DR7 with
1~\AA\ $\le$ \ew $<2$~\AA\ and $R \ge 0.5$ (\emph{circles}) and $R
< 0.5$ (\emph{triangles}), respectively. The \loii\ for all the
systems is shown as \emph{squares}. The number of systems used for
the stack and the stacked profiles are shown. The \emph{dashed
line} show the luminosity evolution of 0.03\ls (\emph{left
panel}) and 0.1\ls (\emph{right panel}) galaxy as a function of
redshift. \emph{Top right panel:} The same for the subset of \mgii
systems with 2~\AA\ $\le$ \ew $<6$~\AA. \emph{Bottom panel:} show
the \sloii\ as a function of redshift for two \ew\ bins of
1~\AA\ $\le$ \ew $<2$~\AA\ (\emph{left panel}) and 2~\AA\ $\le$
\ew $<6$~\AA\ (\emph{right panel}), respectively. The \emph{dashed
line} show the luminosity surface density evolution of 0.03\ls
(\emph{left panel}) and 0.1\ls (\emph{right panel}) galaxy as a
function of redshift.}
\label{fig:loiivsz}
\end{figure*}
Furthermore, we try to explore the contribution of direct detections
to the stacked spectra. For this, we have selected systems with
1~\AA\ $\le$ \ew $<2$~\AA\ and $0.55 \le z < 0.75$ where we have
sufficient number of systems and the \oii\ emission falls in the
wavelength range free from most crowded telluric emission line region.
Here, if one considers systems with {\rm {\ensuremath{\mathcal{R}}}}
$\ge 0.5$ where the \oii\ emission is clearly detected, the \loii\ is
found to be $(1.88 \pm 0.18) \times 10^{40} \rm ~erg\ s^{-1}$ (i.e,
$\sim 0.047$ \lsoii; where, log~\lsoii\ $=41.6$ at average $z$ of
$\sim 0.65$). However, if we exclude the candidate \oii\ emitters
(i.e., systems with \oii\ emission detected at $\ge 3 \sigma$ level),
which accounts for $\sim$5\% of the systems, we still detect the
\oii\ emission with slightly lower \loii\ of (1.45+0.19) $\times
10^{40} \rm ~erg\ s^{-1}$ (i.e., 0.036\lsoii), albeit consistent
within 1.6$\sigma$ level. Here, it is interesting to ask that what
kind of galaxies do contribute to this \loii. For this, we first
compute the average \loii\ of galaxies by using the \oii\ luminosity
function by \citet{Comparat2016MNRAS.461.1076C} at average $z$ of
0.65, for different lower limits on \loii\ ranging between $ \rm
L_{min} =$ 0.001 - 0.01\lsoii, as:
\begin{equation}
\rm \langle L \rangle = \frac {\int_{L_{min}}^{L_{max}} L \phi(L)
dL} {{\int_{L_{min}}^{L_{max}} \phi(L)dL}} .
\label{eq:ewvsd}
\end{equation}
The average \loii\ expected for different values of $\rm L_{min}$ are
listed in column 2 of Table~\ref{tab:lf}. It varies from
0.10\lsoii\ to 0.03\lsoii\ for the $ \rm L_{min}$ ranging from
$0.01$\lsoii\ and $0.001$~\lsoii, respectively. Assuming that the
average \loii\ detected in the stacked spectra after removal of
candidate \oii\ emitters is mainly due to galaxies with \loii\ smaller
than that seen in case of direct detections, we compute the average
\loii\ by restricting ourselves to $\rm L_{max} = 0.23$\lsoii. This
upper limit is set to the average luminosity of lower 5\% systems from
the cumulative distribution of \oii\ luminosity of candidate
\oii\ emitters while considering it as the lowest \loii\ from direct
detections. It is clear from Table ~\ref{tab:lf}, that we recover the
observed \loii\ (i.e., $\sim$ 0.05\lsoii) seen in the stacked spectra
of all the \mgii\ systems when we integrate down to 0.003\lsoii.
Interestingly, the average \loii\ obtained by integrating the
luminosity function over the luminosity range of 0.003\lsoii\ to
0.23\lsoii\ is similar to what we find in the stacked spectra without
direct detections (i.e., $\sim 0.03$\lsoii). This clearly suggest that
strong \mgii absorbers also originate from low luminosity galaxies at
small impact parameters. Note the above estimate is based on the
assumption that galaxies are like point sources and without
considering fibre losses. This should be considered more as an
indicative result. The actual calculations should include size of
galaxies, their orientations and fibre losses into account. We
postpone this for a future work. In addition, the importance of low
mass galaxies contributing to high \ew\ absorbers can be probed
through clustering analysis.
\begin{table}
\centering
{\scriptsize
\caption{The average \oii\ luminosity in the stacked spectra.}
\label{tab:lf}
\begin{tabular}{@{} r c c c @{}}
\hline \hline
\multicolumn{1}{c}{$\rm L_{min}$} &\multicolumn{2}{c}{Average \loii\textcolor{blue}{$^a$}} \\
\multicolumn{1}{c}{($\times$ \lsoii)} &\multicolumn{1}{c}{$\rm \langle L \rangle$ } & \multicolumn{1}{c}{$\rm \langle L^r \rangle$ = $ \frac {\rm \int_{L_{min}}^{0.23L*} L
\phi(L) dL}{{\rm \int_{L_{min}}^{\infty} \phi(L)dL}}$} \\
\hline
0.01 & 0.10& 0.042 \\
0.005 & 0.07& 0.031 \\
0.003 & 0.05& 0.024 \\
0.001 & 0.03& 0.015 \\
\hline
\end{tabular}
\\
} \textcolor{blue}{$^a$}{The upper limits is infinity for the second
column and 0.23\lsoii\ in the numerator for the third column.}
\end{table}
\begin{figure*}
\centering
\epsfig{figure=reddening_new.ps,height=14.0cm,width=14.5cm,bbllx=19bp,bblly=147bp,bburx=592bp,bbury=716bp,clip=true}
\caption{\emph{Left panel:} The geometric mean composite spectra for
the sub-samples with 2~\AA\ $\le$ \ew\ $<$ 6~\AA\ and three
redshift bins of $0.55 \le z < 0.75$ (bottom panel), $0.75 \le z <
0.95$ (middle panel) and $0.95 \le z < 1.3$ (top panel),
respectively. The lower part of each panel shows the flux
ratio of the two spectra for the systems with {\rm
{\ensuremath{\mathcal{R}}}} $\ge 0.5$ and control sample
(black) along with the best fit SMC extinction curve
(\emph{solid red line}). The flux ratio for the systems with
{\rm {\ensuremath{\mathcal{R}}}} $< 0.5$ and control sample are
shown in gray along with the best fit SMC extinction curve
(\emph{dashed blue line}). \emph{Right panel:} The same as
left for the subset with 1~\AA\ $\le$ \ew\ $<$ 2~\AA.}
\label{fig:stack_red}
\end{figure*}
\emph{Therefore, based on the above discussions we conclude that
within the impact parameters probed by the SDSS fibre(s) \mgii
absorbers having higher {\rm {\ensuremath{\mathcal{R}}}} are
associated with regions having higher \loii. This means, either
systems with high {\rm {\ensuremath{\mathcal{R}}}} originate from
regions having high SFR (for a given impact parameter) or have
smaller impact parameter to the star forming region (for a given
\ew). Obtaining spatially resolved spectroscopy as well as image
stacking of these systems \citep{Zibetti2007ApJ...658..161Z} could
help in discriminating between these two alternatives}.
\begin{table*}
\centering
\begin{minipage}{100mm}
{\scriptsize
\caption{The E($B-V$) for various sub-samples.}
\label{tab:ebmv}
\begin{tabular}{@{} c c r r r r r r @{}}
\hline \hline
\multicolumn{1}{c}{Redshift} &\multicolumn{4}{c}{E($B-V$)} \\
\multicolumn{1}{c}{($z$)} &\multicolumn{2}{c}{2~\AA\ $\le$ \ew\ $<$6~\AA} &\multicolumn{2}{c}{1~\AA\ $\le$ \ew\ $<$2~\AA} \\
\multicolumn{1}{c}{} &\multicolumn{1}{c}{\rrr $\ge 0.5$} & \multicolumn{1}{c}{\rrr $< 0.5$} &\multicolumn{1}{c}{\rrr $\ge 0.5$} & \multicolumn{1}{c}{\rrr $< 0.5$}\\
\hline
$0.55 \le z < 0.75$ & $0.011\pm0.002$ & $0.013\pm0.004$ & $-0.001\pm0.002$ &$-0.003\pm0.001$ \\
$0.75 \le z < 0.95$ & $0.027\pm0.001$ & $0.017\pm0.004$ & $ 0.004\pm0.001$ &$0.002\pm 0.001$ \\
$0.95 \le z < 1.30$ & $0.026\pm0.001$ & $0.018\pm0.002$ & $ 0.005\pm0.001$ &$0.004\pm0.001$ \\
\hline
\end{tabular}
}
\end{minipage}
\end{table*}
\subsection{Average dust content}
\label{sub:red}
Evidences for the presence of dust in the intervening absorbers are
commonly seen in the form of continuum reddening. This makes quasar
absorption systems to be a good tracers of dust content within
gaseous haloes surrounding galaxies
\citep{York2006MNRAS.367..945Y,Menard2008MNRAS.385.1053M,Khare2012MNRAS.419.1028K,Menard2012ApJ...754..116M,
Fukugita2015ApJ...799..195F, Sardane2015MNRAS.452.3192S,
Murphy2016MNRAS.455.1043M}. Here, we study the dependence of
average dust content in \mgii\ systems on the \rrr\ parameter, i.e.,
\rrr $\ge 0.5$ and \rrr $< 0.5$. For this, we have made two subsets
based on \ew\ with 1~\AA\ $\le$ \ew $<2$~\AA\ and 2~\AA\ $\le$ \ew
$<6$~\AA. We further divide each subset in to three redshift bins of
$0.55 \le z < 0.75$, $0.75 \le z < 0.95$ and $0.95 \le z < 1.3$. We
have generated geometric mean spectra for various sub-samples as well
as for the control samples of quasars, within $\Delta z = \pm 0.05$
of \zem and $\Delta r_{\rm mag} = \pm 0.5$ of $r_{\rm mag}$, without
absorption in their spectra. \par
The stacked spectra for various subsets are shown in
Fig.~\ref{fig:stack_red}. We estimate the reddening, E($B-V$), by
fitting the spectral energy distribution (SED) of the control sample,
reddened by the Small Magellanic Cloud (SMC) extinction curves
\citep{Gordon2003ApJ...594..279G}. The flux ratio of the composite
spectra for the systems with {\rm {\ensuremath{\mathcal{R}}}} $\ge
0.5$ (\emph{black}) and {\rm {\ensuremath{\mathcal{R}}}} $< 0.5$
(\emph{gray}) to the control sample are shown in the lower panel of
Fig.~\ref{fig:stack_red}. The best fit SMC extinction curve for {\rm
{\ensuremath{\mathcal{R}}}} $\ge 0.5$ and {\rm
{\ensuremath{\mathcal{R}}}} $< 0.5$ sample are overlaid in
\emph{solid red} line and \emph{dotted} line, respectively. We perform
a bootstrap analysis to measure the uncertainties over each
measurement. For this, we make stacked spectra for 1000 sub-samples by
randomly selecting 70 per cent of the sample and measure the E($B-V$)
by fitting the SED of control sample, reddened by SMC extinction
curves. We consider the standard deviation of E($B-V$) distribution as
1$\sigma$ uncertainty.
In Table~\ref{tab:ebmv}, we have summarized the colour excess,
E($B-V$), for each subset. At first, we confirm that the E($B-V$) is
more towards high \ew\ systems
\citep{Budzynski2011MNRAS.416.1871B,Jiang2011ApJ...732..110J} at any
redshift bin. We also find that the E($B-V$) is higher for the systems
at high redshift \citep[see
also,][]{Budzynski2011MNRAS.416.1871B,Menard2012ApJ...754..116M}.
For the strong \mgii\ systems (i.e., \ew\ $\ge 1$\AA) in our sample
the E($B-V$) is found to be in the range of -0.001 to 0.027. Using 809
\mgii absorption systems with $1.0 \le$ \zabs\ $\le 1.86$
\citet{York2006MNRAS.367..945Y} have shown that the typical colour
excess, E($B-V$), introduced by these systems ranges from $< 0.001$ to
0.085 \citep[see also,][]{Wild2007MNRAS.374..292W}. Interestingly, for
the subset with 2~\AA\ $\le$ \ew\ $<$6~\AA\ and $0.95 \le z < 1.3$ we
find that \mgii systems with {\rm {\ensuremath{\mathcal{R}}}} $\ge
0.5$ are redder (significant at 3.6$\sigma$ level) than the systems
with {\rm {\ensuremath{\mathcal{R}}}} $< 0.5$. A similar trend is seen
for the subset with 2~\AA\ $\le$ \ew\ $<$6~\AA\ and $0.75 \le z <
0.95$ albeit with a lower significance at 2.4$\sigma$ level. However,
we do not see this trend for any other subsets (see also,
Table~\ref{tab:ebmv}). Note that the typical E($B-V$) of $< 0.02$ has
been inferred using the DLAs \citep{Vladilo2008A&A...478..701V} and
$\sim 0.046$ from the \caii\ absorbers with equivalent width of $\ge
0.7$\AA\ \citep{Sardane2015MNRAS.452.3192S}. Interestingly, the
average E($B-V$) for the sub-sample with 2~\AA\ $\le$ \ew
$<6$~\AA\ and redshift bins of $0.75 \le z < 0.95$ and $0.95 \le z <
1.3$ show the reddening of $\sim$ 0.03 as seen in the extreme-DLAs,
i.e., log~\nhi $\ge 21.7$, by \citet{Noterdaeme2014A&A...566A..24N}.
Moreover, the \hi 21-cm absorbers that also show \rrr\ $\ge 0.5$ tend
to produce more significant reddening in the spectrum of the
background quasars than \mgii\ systems without 21-cm absorbers
\citep{Dutta2017MNRAS.465.4249D}.
\section{Discussion and conclusions}
\label{lab:dicuss}
We have investigated the effect of fibre size as well as the metal
absorption line ratio (\rratio) on the average luminosity of
\oiiab\ nebular emission from \mgii absorbers (at $0.55 \le z \le
1.3$) in the composite spectra by utilizing quasar spectra obtained
with 3 and 2 arcsec fibres in Sloan Digital Sky Survey. We have found
the following interesting results:
1. We confirm the presence of a strong correlation between
\oii\ luminosity and \ew, in both the data sets. The \sloii\ measured
for SDSS-DR7 is found to be higher than those measured with SDSS-DR12.
This suggests that the fibre effects are not fully taken care of even
when we normalize the luminosity by the projected area. Interestingly,
the difference is found to be largest for the highest \ew\ bin of
$3-6$~\AA. This might be due to the observed large scatter between
\ew\ vs $\rho$ relation which implies that even for the Ultra strong
(\ew $\ge 3$~\AA) \mgii\ systems there is a non negligible probability
of galaxy being outside the fibre (i.e., at large impact parameters).
While discussing the difference between SDSS-DR7 and SDSS-DR12
observations we need to also remember some differences in the
observational strategy adopted. In SDSS-DR7 the fibres are centered on
red whereas in SDSS-DR12 a centering offset is introduced to the
fibres, taking into account atmospheric dispersion, to improve the
flux on blue part. At this stage it is not clear what is the
contribution of this effect to the differences we discuss in this
work. Therefore, it is important to systematically study a sample of
strong Mg~{\sc ii} systems with integral field spectroscopy to map the
extent and nature of star formation associated with Mg~{\sc ii}
systems of different equivalent widths.
2. We also explore the dependence of \loii\ on \rratio. We have found
that the \mgii absorbers with \rrr\ $\ge 0.5$ tend to show higher
\loii\ and \sloii\ (see Fig.~\ref{fig:loiivsew_r}). In fact, the
strong correlation seen between \loii\ vs \ew\ and \loii\ vs $z$ is
mainly driven by systems with strong \feii (i.e. \rrr\ $\ge$ 0.5). For
systems with \rrr\ $< 0.5$ no such trend between \ew\ and \loii\ is
visible. We also show that the fraction of systems with direct
detection increases as a function of \ew\ which gives a hint that the
strong correlation seen in the stacked spectra is possibly a combined
result of the \ew versus $\rho$ anti-correlation and the redshift
dependent fibre losses.
3. Strong dependence of \oii\ luminosity on \rrr\ could mean the
correlation between impact parameter and \rrr\ could be stronger than
that between \ew\ and impact parameter. It will be important to check
this before ascribing any physical connection between \rrr\ and star
formation rate associated with the absorbing galaxy. Unfortunately
\rrr\ values are not available for systems used to define the
correlation between \ew\ and $\rho$
\citep{Nielsen2013ApJ...776..114N}. Therefore, it will be an important
step to explore the correlation between \ew\ and \rrr\ for building a
clear connection between \ew\ and associated star formation rate.
4. We clearly detect the \oii\ emission in the stacked spectra even if
we exclude the candidate \oii\ emitters (systems with nebular emission
detected at $\ge 3 \sigma$ level of significance). This could either
means appreciable contribution from low luminosity galaxies at small
impact parameters or galaxies at larger impact parameters with only
light from outer regions of galaxies contributing to the emission in
the stacked spectra.
5. We confirm the trend of increasing $E(B-V)$ of \mgii absorbers with
increasing \ew\ as well as redshift (see also
Fig.~\ref{fig:stack_red}). Interestingly, for the subset with
2~\AA\ $\le$ \ew\ $<$6~\AA\ and $0.95 \le z < 1.3$ the $E(B-V)$ of
\mgii systems with {\rm {\ensuremath{\mathcal{R}}}} $\ge 0.5$ is found
to be higher than that for the systems with {\rm
{\ensuremath{\mathcal{R}}}} $< 0.5$ (at 3.6 $\sigma$ level). The
$E(B-V)$ found in {\rm {\ensuremath{\mathcal{R}}}} $> 0.5$ is similar
to the $E(B-V)$ inferred for DLAs. Using the \mgii\ systems searched
for the \hi absorption in STIS survey by
\citet{Rao2006ApJ...636..610R} and \citet{Rao2017MNRAS.471.3428R} we
compute the fraction of strong \mgii systems (\ew $\ge 1$~\AA) having
\rrr\ $\ge 0.5$ being DLAs, i.e., log~\nhi $\ge 20.3$, is $\sim$38\%
which is only $\sim$6 \% in case of \rrr\ $< 0.5$. This fraction
increase to $\sim$60\% if we consider the systems with \ew $\ge
2$~\AA\ and \rrr\ $\ge 0.5$. Therefore, the \mgii absorbers with
\rrr\ $\ge 0.5$ systems may be related to high probability of them
being sub-DLA and DLAs.
\label{lastpage}
\section*{Acknowledgments}
RS, PN, and PPJ acknowledge the support from Indo-French Centre for
the Promotion of Advance Research (IFCPAR) under project number
5504$-$2.\par
Funding for the Sloan Digital Sky Survey IV has been provided by the
Alfred P. Sloan Foundation, the U.S. Department of Energy Office of
Science, and the Participating Institutions. SDSS-IV acknowledges
support and resources from the Center for High-Performance Computing
at the University of Utah. The SDSS web site is www.sdss.org.
SDSS-IV is managed by the Astrophysical Research Consortium for the
Participating Institutions of the SDSS Collaboration including the
Brazilian Participation Group, the Carnegie Institution for Science,
Carnegie Mellon University, the Chilean Participation Group, the
French Participation Group, Harvard-Smithsonian Center for
Astrophysics, Instituto de Astrof\'isica de Canarias, The Johns
Hopkins University, Kavli Institute for the Physics and Mathematics of
the Universe (IPMU) / University of Tokyo, Lawrence Berkeley National
Laboratory, Leibniz Institut f\"ur Astrophysik Potsdam (AIP),
Max-Planck-Institut f\"ur Astronomie (MPIA Heidelberg),
Max-Planck-Institut f\"ur Astrophysik (MPA Garching),
Max-Planck-Institut f\"ur Extraterrestrische Physik (MPE), National
Astronomical Observatories of China, New Mexico State University, New
York University, University of Notre Dame, Observat\'ario Nacional /
MCTI, The Ohio State University, Pennsylvania State University,
Shanghai Astronomical Observatory, United Kingdom Participation Group,
Universidad Nacional Aut\'onoma de M\'exico, University of Arizona,
University of Colorado Boulder, University of Oxford, University of
Portsmouth, University of Utah, University of Virginia, University of
Washington, University of Wisconsin, Vanderbilt University, and Yale
University. |
\section{INTRODUCTION}
Because of the complexity of robotics, its research depends substantially on experimental evaluations to reveal consequences that cannot be seen from the start. Unfortunately, this complexity even makes the execution and reproduction of experiments very time-consuming and difficult. For example, the execution of an experiment with several robots, external sensors and actuators as well as additional monitoring systems needs the setup of various complex software artifacts. In many cases, these artifacts are exchanged by researchers across the globe, which makes it even more complicated to execute and, above all, reproduce an experiment. Some research networks have established special interest groups, e.g. EURON Good Experimental Methodology (GEM) or the IEEE Technical Committee on Performance Evaluation and Benchmarking of Robotic and Automation Systems (TC-PEBRAS), to develop methodologies and tools that improve the situation. One outcome is the demand by Bonsignorio and del Pobil~\cite{bonsignorio2015}, who call for a possibility to practically replicate results for validation. The authors suggest a new community-wide agreement on the content of a research paper in robotics. They propose the additional accompaniment of \textit{data sets}, \textit{code identifiers} and \textit{hardware identifiers} to provide a level of transparency where the experiment itself can be replicated in another environment by an objective researcher. Guglielmelli~\cite{guglielmelli2015} points out that reproducibility is one of the critical factors for dependable robots and supports the trust of the society in such systems. To increase the maturity of robotics research, the assurance of reproducibility and the possibility of objective disconfirmation is indispensable. This paper proposes a novel workflow that addresses the requirements stated by Bonsignorio and del Pobil~\cite{bonsignorio2015} and introduces a framework that abstracts the complexity of experimental setups by using containerization technologies. The proposed approach makes it significantly easier to share the experimental artifacts and reproduce the environment of the inital experiment.
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{figure1}
\caption{The Architecture of \textit{SwarmRob} - The green cubes represent the worker nodes and the red cubes represent the master nodes. Every bounding box illustrates a swarm.The outer box illustrates the local network of the laboratory and the grey boxes illustrates the repositories where the worker can obtain the definition files.}
\label{fig:sr_architecture}
\end{figure}
\section{RELATED WORK}
The following section discusses the related work that is focused on systems consisting of multiple robots, namely \textit{Multi-Robot Testbeds} and \textit{Application Containerization}.
\subsection{Multi-Robot Testbeds}
\par In some branches of robotics, e.g. evolutionary robotics, it is common to evaluate approaches by simulating real-world experiments. Unfortunately, this approach leads oftentimes to differing results, compared to results of the system in the real world~\cite{koos2013}. To synergize the evaluation capabilities of simulation evironments and equally prevent the ``reality gap``, testbeds for multiple robots have been developed. Michael et al.~\cite{michael2008} introduced a testbed for large multi-robot systems with a strong focus on classical swarm robotics and homogeneous compositions. The system is limited to two different robot platforms and supports the integration of a three-dimensional simulation environment. Johnson et al.~\cite{Johnson2006} proposed one of the first remotely accessible testbeds for mobile robots combined with sensor networks. The testbed uses a client-server architecture with a web-based front-end for creating and managing experiments, including logging and remote code execution. Additionally, it supports robot localization using external cameras and remote robot control. A very similar approach called \textit{HoTDeC} were introduced by Stubbs et al.~\cite{Stubbs2006}. While \textit{HoTDeC} also uses external cameras, it is designed for homogeneous composition of autonomous hovercrafts. Another variant, which is designed for cheaper and smaller robots, were proposed by Pickem et al.~\cite{Pickem2017}. The system relies on very small robots called \textit{GRITSbots}~\cite{Pickem2015} and should also be available to external users for remotely carrying out their experiments. The disadvantages of testbeds are that they are limited to a homogeneous set of components and are usually inflexible to technical expansions. Additionally, the possibility to reproduce experiments is limited to a very small subset of researchers which are in possession of the complete testbed. A more inclusive and flexible alternative referred to in the literature is the use of containerization technologies.
\subsection{Containerization in Multi-Robot Systems}
A more integral approach to the problem of reproducibility is the use of container technologies to encapsulate the complexity of software and hardware. Boettiger~\cite{boettiger2015} and Cito et al.~\cite{cito2016} discussed the relevance of containers for software engineering research and mention four technical challenges that prevent reproducibility in software engineering reserach that are also problems in robotics research: (\textit{i.}) the \textit{``Dependency Hell''}, which is the problem of reproducing computational environments to run the software, (\textit{ii.}) \textit{Imprecise Documentations}, which multiplies the problem \textit{(i.)} and is therefore another barrier to install and run the software, (\textit{iii.}) \textit{Code erosion}, which is the problem of running outdated or updated code in current environments and (\textit{iiii.}) \textit{Barriers to Adoption with Existing Solutions}, which is the problem that existing technological solutions that would solve some of the problems need a high level of expertise and are therefore neglected by the researchers. Both propose containers as an approach to face these challenges. Currently, the use of containers in robotics research is not very popular because a practical, generic approach is missing. One of the first approaches that addressed this deficiency and proposed an experimental workflow in robotics research inspired by containerization is the \textit{Cognitive Interaction Toolkit (CITk)} by Lier et al.~\cite{lier2014}. The approach uses containerization in an automated build and deployment process for simulation environments. An extension to the \textit{CITk} is the \textit{RoboBench} project presented by Weisz et al.~\cite{weisz2016} that extends the \textit{CITk} by a benchmarking suite that allows the distribution of system-wide benchmarking containers via public repositories. Unfortunately, the initial approach as well as the extension do not support physical robots.
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{figure2}
\caption{Workflow of SwarmRob - The figure illustrates the research phase (left timeline) and the review phase (right timeline) of the workflow with their related subphases.}
\label{fig:workflow}
\end{figure*}
\section{CONTRIBUTION TO THE STATE OF THE ART}
In order to address the previously discussed deficiency, this paper presents a novel framework that enables the distribution of experimental artifacts in robotics research based on a container technology, called \textit{Docker}~\cite{docker2017}. It supports the dynamic assignment of containers to nodes based on predefined composition definitions, which can be easily provided to other researchers. Unlike other orchestration solutions, the system is tailored to robotics and incorporates the underlying hardware as a parameter within the service allocation process. In the previous section, some of the technical challenges that prevent reproducibility have been discussed. Below, the particular challenges are revisited and solution concepts of the approach will be discussed:
\begin{enumerate}[(i)]
\item \textbf{Imprecise Documentations}: The approach tackles the problem of \textit{Imprecise Documentations} by detailed definition files. They are the important basis for the bootstrapping of the experiment.
\item \textbf{``Dependency Hell``}: Because of the packaging of software and dependencies in containers, the approach addresses the several forms of the \textit{Dependency Hell} appropriately .
\item \textbf{Code erosion}: The \textit{Code erosion} in robotics applications is addressed by the use of software containers, which make it possible to use outdated software via virtualized operating system environments.
\item \textbf{Barriers to Adopting Existing Solutions}: Because the use of container technology in robotics needs a high level of expertise and therefore raises \textit{Barriers to Adopt the Solution}, the system automates most of the necessary steps, e.g. container distribution, virtualized networking, and thus reduces obstacles to use the technology.
\end{enumerate}
The previous section has highlighted the contribution to the state of the art as well as the advantages of the proposed solution comparing to other approaches. The following section describes the system in detail.
\section{SWARMROB}
The \textit{SwarmRob} system simplifies the re-execution and reproducibility of experiments in robotics research. Therefore, it uses containers for the deployment of robotics applications and bootstrap as well as orchestration mechanisms to abstract the complexity of the technology. The remainder of this section discusses the several partial aspects in detail.
\subsection{Definitions}
For clarification, several notions need to be defined: An experiment is a composition of services $z \in Z$, where $Z$ is finite set of services. A service $z$ is a detailed description of the assembly of a specific artifact of the experiment. A \textit{swarm} is an element $s \in S$, where $S$ is a finite set of swarms and $s$ itself is also a finite set which means that $S$ is a set of sets. The atomic elements of $s$ are agents $a \in A$, which implies that $s \subset A$. Every agent in a swarm has either the role of a \textit{worker} $w \in W$, where $W \subset A$ or the role of a \textit{master} $m \in M$, where $M \subset A$ with the restriction that $M$ is a singleton set. As a consequence, one has $A \setminus (M\cup W) = \emptyset$. Furthermore, the amount of workers is the amount of robots in the swarm less the \textit{master} $W \approx (A\setminus M)$ and no participant can be worker and master at the same time $(M\cap W) = \emptyset$.
\subsection{The SwarmRob Workflow}
Along with the framework, \textit{SwarmRob} proposes a novel workflow for the documentation, distribution and execution of experiments (Fig. \ref{fig:workflow}). The workflow of the system can be subdivided in two meta-phases: the \textit{(i.) Research Phase}, where the initial experiment is constructed and evaluated and the \textit{(ii.) Review Phase}, where the results are reproduced and reviewed by the community. The research phase is subdivided in four consecutive phases: The first phase is concerned with the \textit{Specification of Individual Services}, where the hardware and software configurations of services are documented via \textit{CDFs (Container Definition Files)}. The \textit{CDF} can be interpreted as a basic description of an artifact of the experiment. The \textit{CDF} describes e.g. the operating system of the service, the required software packages or the code repositories that should be cloned within the initialization. The deliverable of the specification phase is a service that is ready for execution. A service can be reused in another experiment, by forking its \textit{CDF}. The second phase of the workflow is the \textit{Specification of the Experiment}. In this phase, the composition of the services, which is the basis of the experiment, is defined. The specification is deposited in a so called \textit{EDF (Experiment Definition File)}. The \textit{EDF} refers to the \textit{CDFs} and adds additional information like network configurations, relationships between services (e.g. two services should be executed on the same machine) and hardware requirements. The deliverable of this phase is a complete specification of an experiment composition that can be re-executed in various environments. The third phase of the workflow is the actual experiment. The experiment is performed by the help of \textit{SwarmRob} but is completely independent of the framework in terms of evaluation and monitoring. The deliverable of this phase is the result of the experiment that is planned to be shared, e.g. in a scientific publication. In the subsequent phase, the researcher can make the composition of the experiment (CDFs and EDF) and the results of the experiment available to other researchers using public repositories and academic publications. In the \textit{Review Phase}, the community can easily obtain the results and definitions and is able to review the performed experimental evaluations. \textit{SwarmRob} automates the distribution of the containers, the initialization of the robots as well as the configuration of the inter-robot network. The deliverable of this phase is a bootstrapped system that resembles the situation of the inital experiment as close as possible. Based on this, the experiment can be reproduced and a much more qualified feedback for the author is possible.
\subsection{The Container Ecosystem}
The packaging of software using containers is founded on the idea of abstracting and isolating hardware and encapsulating software. It guarantees that any process inside a container cannot see a process outside of the container. Every container can be easily distributed and executed using a \textit{CDF} that defines among other things, the operating system, required software packages and dependencies as well as mounted volumes. The behaviour of each container is managed by a \textit{container engine (CE)} that handles its life cycle. In \textit{SwarmRob}, the container technology is used to encapsulate the services running on each robot. Using the example of the \textit{Robot Operating System (ROS)}, a high amount of software dependencies and configuration is required to replicate a system. The container technology of \textit{SwarmRob} simplifies this since it manages all of the dependencies and configurations in one single \textit{CDF} that can be easily distributed and executed on every robot with the same hardware configuration. Because the manual setup of large-scale containerized systems is time-consuming and needs expertise, \textit{SwarmRob} takes care of the necessary steps. For this purpose, every node of the system runs a \textit{SwarmRob} daemon. The daemon automates all aspects of the container workflow, e.g. obtain and start containers, setup networks or provide system information of the worker to the master. Because of the highly distributed architecture of system, the daemon also manages the distributed access of information used for inter-service networking and controls the service allocation process participation. The technology stack is illustrated by the left robot shown in Fig. \ref{fig:network_architecture}.
\begin{figure}
\centering
\includegraphics[width=0.35\textwidth]{figure3}
\caption{System and Inter-Robot Network Architecture using Overlay Networks - The \textit{Underlay Network} represents the physical network connection between the hosts, the \textit{Intra-Swarm Communication} represents the commands and information exchanged between the participants of a swarm and the \textit{Overlay Network} is the communication channel used for the communication between containerized applications.}
\label{fig:network_architecture}
\end{figure}
\subsection{Network Architecture}
Information exchange between robots is one of the originating aspects of multi-robot systems. \textit{SwarmRob} supports the communication between services using \textit{VXLAN}~\cite{rfc7348}. It encapsulates the actual communication via a tunneling mechanism on top of the actual physical infrastructure. \textit{VXLAN} uses the existing underlay network and encapsulates the traffic in UDP packets of the underlay network and adds an additional \textit{VXLAN} header. In a \textit{VXLAN}-based overlay network, only the \textit{VTEPs} (\textit{VXLAN Tunnel Endpoints}) can encapsulate, respectively decapsulate the traffic. In the context of \textit{SwarmRob}, the \textit{VTEPs} of an overlay network are the services that assemble an experiment. For that purpose, additional information, e.g. network configuration and the members of the communication, need to be shared between the \textit{VTEPs}. In the toolkit, this is implemented using a \textit{distributed key-value store (D-KV)} that guarantees the conflict-free and highly topical access to this information. The isolated experiment-specific network traffic has several benefits: First of all, due to the isolation, the monitoring of the communication traffic is much simpler, because potential noise (e.g. broadcast, web traffic) of the host system is filtered. This results in a better reproducibility of communication experiments and comparability of metrics. In addition, the risk of potential network conflicts (e.g. already used resp. blocked ports or IPs, misconfigured networks or communication channels) can be reduced, because the network traffic as well as the configuration (IPs, Ports, Subnets) of the overlay network tunnel is completely independent of the underlay network and unique for each instance of the overlay network. The communication between the master and its workers within a swarm is implemented using remote procedure calls. The network architecture of the system is illustrated in Fig.~\ref{fig:network_architecture}.
\subsection{The SwarmRob Architecture}
The architecture of \textit{SwarmRob} considers the special requirements of multi-robot experiments by using a master-worker architecture, where every participant in the system is either a worker that can run one or more services or a master that controls the swarm. Every swarm of the system is a unique instance initialized by its master node. The master node is an indespensable component of the swarm. It manages the lifecycle of the workers, the service allocation process as well as the overall logging of the system. Every instance of a swarm can be joined as a worker node by its unique identifier and the network address of the master. If the swarm should execute an experiment, the master loads the \textit{EDF} and allocates the services dynamically to the workers based on their hardware capabilities as well as their workload. The worker that is used to run a specific service obtains the \textit{CDF} of the assigned service either by a publically or locally accessible repository. The whole architecture of the system is illustrated in Fig.~\ref{fig:sr_architecture}.
\begin{figure}
\centering
\includegraphics[width=0.42\textwidth]{figure4}
\caption{Model of the Service Allocation Problem as a Flow Network - The nodes on the left side represent the set of worker, the node on the right side represent the set of services and pooled services. The edge labels represent the costs to run a service \textit{j+k} on a worker \textit{i}.}
\label{fig:container_allocation_graph}
\end{figure}
\subsection{Dynamic Service Allocation}
The experiment bootstrap process is implemented using a dynamic allocation of services to workers. The problem is an instance of the \textit{assignment problem} and can be modelled as a flow network (Fig.~\ref{fig:container_allocation_graph}) and formulated as a maximum-flow with minimum-cost problem. A flow network is a directed graph $G=(V,E)$ with a source vertex $s \in V$ and a sink vertex $t \in V$. Each edge has a capacity $c:E \rightarrow \mathbb{R}^{+}$, denoted by $c_{uv}$ that represents the maximum amount of flow that can pass through an edge. The costs $a:E \rightarrow \mathbb{R}^{+}$ to run a container are denoted by $a_{uv}$. A flow is a mapping $f:E \rightarrow \mathbb{R}^{+}$, denoted by $f_{uv}$ subjected to a capacity constraint (Equation~\ref{eq:capacity_constraint}) that guarantees that an arc's flow cannot exceeds its capacity and a conservation constraint that equals Kirchhoffs current law (Equation~\ref{eq:conservation_constraint}).
\begin{subequations}
\begin{alignat}{2}
&\forall (u,v) \in E: f_{uv} \leq c_{uv}\label{eq:capacity_constraint}\\
&\forall v\in V \setminus \{s,t\}: \sum_{u:(u,v)\in E}{f_{uv}} =\sum_{u:(v,u)\in E}{f_{vu}}. \label{eq:conservation_constraint}
\end{alignat}
\end{subequations}
The value of the flow (amount of flow passing from source to sink) is defined by
\begin{flalign}
|f|=\sum_{v:(s,v)\in E}{f_{sv}} - \sum_{v:(v,s)\in E}{f_{vs}}\text{.}
\end{flalign}
The total cost over all edges is given by
\begin{flalign}
|a| = \sum_{(u,v) \in E}{a_{uv} \cdot f_{uv}}\text{.}
\end{flalign}
The maximum-flow with minimum-cost problem is to maximize $|f|$ while minimizing $|a|$.
Each worker $w \in W$ and each service $z \in Z$ is a vertex $\omega : (W\cup Z) \rightarrow V$ subjected to the constraint that an edge is only allowed between a worker and a service.
Whether two vertices are connected by an edge is given by the hardware capabilities of that worker, that is, if the worker is able to run the specified service. This restriction on the edges is specified via a binary matrix $H$, where $R$ is a binary relation $R \subseteq W\times Z$ that holds iff. a worker $w_i \in W$ has the hardware capabilties to run a service $z_j \in Z$. As a consequence, the entries of $H$ are defined by
\begin{equation}
H_{i,j}=\begin{cases}
1 \text{~~}(w_i,z_j) \in R \\
0 \text{~~}(w_i,z_j) \notin R\text{.}
\end{cases}
\end{equation}
The corresponding transformation function is a mapping $\kappa: Z \times W \times H \rightarrow E$. The overall costs $a_{uv}$ for running a service on a worker is given by the workload of the worker and the pre-defined costs for a single container. The overall workload of the worker is composed of the CPU load denoted by $\epsilon$, the VRAM load denoted by $\eta$, the SWAP load denoted by $\zeta$ and the bandwith of the worker denoted by $\theta$, all multiplied by a weighting factor $\delta$:
\begin{flalign}
a_{uv} = \epsilon_{uv} \cdot \delta_{\epsilon} + \eta_{uv} \cdot \delta_{\eta} + \zeta_{uv} \cdot \delta_{\zeta} +\theta_{uv} \cdot \delta_{\theta},
\end{flalign}
such that the sum of all weights $\delta$ equal 1. The single cost of each load is given by their related cost functions:
\begin{subequations}
\begin{flalign}
\epsilon_{uv} &= \alpha \cdot \beta^4 \textit{~(CPU load)}\\
\eta_{uv} &= \alpha \cdot \beta^4 \textit{~(VRAM load)}\\
\zeta_{uv} &= \alpha \cdot \beta \textit{~(SWAP load)}\\
\theta_{uv} &= \alpha \cdot (1-\beta)^{4} \textit{~(Bandwidth),}
\end{flalign}
\end{subequations}
where $\alpha \in \{x \in \mathbb{R}|0\leq x \leq 100\}$ is the predefined costs for the service and $\beta \in \{x \in \mathbb{R}|0\leq x \leq 1\}$ is the currently measured relative workload. Fig.~\ref{fig:contour_plots} illustrates the cost functions as contour plots. It shows that $\epsilon$ and $\eta$ are progressive cost functions, where the exponent is chosen such that the cost increase with a strong degree of progressivity in the second half of the workload scale to relieve high-loaded workers. Because of the assumption that a high swap indicates a very high workload of some worker, the exponent of the cost function $\zeta$ is chosen such that the cost rise earlier than $\epsilon$ and $\eta$. In the case of $\theta$, the cost function is chosen such that the cost decrease along with the available bandwith of the worker. If the bandwith tends to zero, the cost increases. The cost calculation algorithm is designed as a distributed algorithm that allows uses parallelization of the cost calculation. We assume that $m$ is the total number of agents, $n$ is the total number of services, $i$ lies in the interval $[1,m]$ and $j$ lies in the interval $[1,n]$. The cost per worker and per service are represented as a cost matrix $A$ of size $m \times n$, where the entries of $A$ are defined by $A_{i,j}=a_{uv}$. Each worker can run exactly one container, which implies that each capacity is given by $c_{uv} = 1$. The overall capacities are represented as a matrix $C$ of size $m \times n$, where the entries of $C$ are defined by $C_{i,j}=c_{uv}$.
\begin{figure}[!t]
\centering
\includegraphics[width=0.49\textwidth]{figure5}
\caption{Contour Plots of the Workload Cost Functions - The several x-axes represent the relative workload, the y-axes represent the pre-defined costs per image and the color of the contour plot indicates the resulting cost.}
\label{fig:contour_plots}
\end{figure}
In order to enable that two or more services can be executed on one worker, the system supports \textit{pooled services} $\Phi$, where $ \Phi \subset \mathcal{P}(Z)$. To take into account that in some cases (e.g. when all machines have a high workload) the execution of a pooled service on different machines is more favorable than the execution on a single machine, the set of service vertices is the set of the single services united with the set of the pooled services such that $\tau: (W\cup Z \cup \Phi) \rightarrow V$. We assume that $k$ lies in the interval $[1,n]$. The dependencies of services are represented via a binary matrix $\Xi$ of size $n \times n$, where $D$ is a binary relation $D \subseteq Z \times Z$ that holds iff. a service $z_k \in Z$ depends on a another service $z_j \in Z$. The entries of $\Xi$ are given by
\begin{equation}
\Xi_{k,j}=\begin{cases}
1 \text{~~}(z_k,z_j) \in D \\
0 \text{~~}(z_k,z_j) \notin D\text{.}
\end{cases}
\end{equation}
The cost of a pooled service $\phi \in \Phi$ for a worker $w \in W$ is given by the sum of cost of their respective containers, multiplied by a discount factor $\delta$.
The corresponding cell of the hardware capability matrix $H$ of a pooled container $\phi \in \Phi$ for a worker $w \in W$ is the logical conjuction of all entries of $H$ for that service.
The entries of the capacity matrix $C$ stays the same as for single services.
The problem is solved using the \textit{cost-scaling push-relabel (CSPR)} algorithm by Goldberg and Kennedy~\cite{goldberg1997} implemented in the Google Optimization Toolbox~\cite{google2017}.
\begin{figure}[b]
\centering
\includegraphics[width=0.35\textwidth]{figure6}
\caption{Boxplots of the Elapsed Time in Seconds of Container and Native \textit{SLAMBench}, each for x64 and ARM.}
\label{fig:slambench}
\end{figure}
\section{EVALUATION}
The following section discusses the results of the evaluation. To reveal possible differences between system architectures, the experiments were performed using an x64-based system consisting of an \textit{Intel Core i5-6200U} with \textit{2.30 GHz}, \textit{4096 MB LPDDR3} and \textit{IEEE 802.11ac-WLAN} connection and an ARM-based system consisting of an \textit{ARM Cortex-A7} with \textit{900 MHz}, \textit{1024MB DDR2-SDRAM} and a \textit{IEEE 802.3-10/100-Mbit/s} connection. The evaluation of the service allocation algorithm as well as the overall system evaluation were performed using the previously described ARM-based system composed as a 12-node cluster.
\begin{figure*}[t]
\centering
\includegraphics[width=1\textwidth]{figure7}
\caption{Boxplots of Network Performance Measures - The left figure shows the TCP/UDP Request/Response Performance, the figure in the middle shows the TCP Stream Performance and the right figure shows the UDP Stream Performance, each for the containerized and the non-containerized version}
\label{fig:netperf}
\end{figure*}
\begin{table}
\setlength\extrarowheight{2.5pt}
\caption{Statistical Dispersion of the Container Performance Experiment}
\label{tab:msd_container_performance}
\centering
\begin{tabular}{|p{4cm}|l|l|}
\hline
\textit{Metric} & \textit{Std. Deviation} & \textit{CV}\\ \hline \hline
\textit{x64} & 0.0908 & 0.2800 \\
\textit{x64\_Cont} & 0.0908 & 0.2800 \\
\textit{ARM} & 1.2258 & 0.2429 \\
\textit{ARM\_Cont} & 1.1820 & 0.2426 \\ \hline
\end{tabular}
\end{table}
\subsection{Container Performance}
An important aspect is the question whether the containerization of robotics applications affect the computational performance of the system. To consider the special requirements of robotics applications and to ensure reproducibility, the evaluation was performed using \textit{SLAMBench}~\cite{nardi2015}, which is a framework for quantifying \textit{SLAM} algorithms. Additionally, the framework contains implementations of \textit{SLAM} algorithms and it is able to run on pre-recorded datasets. For the performance evaluation of the containerized case compared to the native case, the experiment was performed using the \textit{KinectFusion} algorithm~\cite{newcombe2011} executed on the \textit{ICL-NUIM} dataset~\cite{handa2014}, each for x64-based systems and ARM-based systems. The results presented in Fig.~\ref{fig:slambench} show a comparison of boxplots of the total time elapsed for the processing of one frame. The processing of one frame includes pre-processing, tracking, integration and raycast as well as the acquisition and rendering of a single frame of the dataset. The dataset consists of \textit{881} frames and the experiment was repeated \textit{10} times. The \textit{standard deviations} and the \textit{coefficient of variations (CVs) }are shown in Tab.~\ref{tab:msd_container_performance}. For the x64-based architecture (x64/x64-Cont) as well as for the ARM-based architecture (ARM/ARM-Cont), the experiment indicates that there is no performance loss when using containers instead of native applications, for this particular benchmark. The slightly better performance of the container can be attributed to normal fluctuations of the system utilization. Earlier studies suggest that the same observation applies to the GPU passthrough performance of LXC containers~\cite{walters2014}. Both implies that the use of virtualized robot applications has no impact on the comparability and transferability of the experiments as well as on the computational performance of the underlying hardware, which is a necessary property for the success and a further dissemination of the approach.
\begin{table}
\setlength\extrarowheight{2.5pt}
\footnotesize
\caption{Measures of Statistical Dispersion of the Network Performance Experiment}
\label{tab:msd_network_performance}
\centering
\begin{tabular}{|p{4.3cm}|l|l|}
\hline
\textit{Metric} & \textit{Std. Deviation} & \textit{CV}\\ \hline\hline
\textit{TCP\_RR\_Cont} & 51.5990 & 0.1237 \\
\textit{TCP\_RR} & 57.1165 & 0.1280 \\
\textit{UDP\_RR\_Cont} & 83.4594 & 0.2029 \\
\textit{UDP\_RR} & 78.3588 & 0.1806 \\
\textit{TCP\_Stream\_Cont} & 12.3056 & 0.1279 \\
\textit{TCP\_Stream} & 12.4611 & 0.1280 \\
\textit{UDP\_Steam\_Cont} & 14.2138 & 0.1469 \\
\textit{UDP\_Stream} & 14.5790 & 0.1425 \\ \hline
\end{tabular}
\end{table}
\subsection{Network Performace}
The following section discusses the impact on the network performace when using virtualized networks on top of physical infrastructures. In order to assure meaningful results, several different aspects of the network performance were measured, each for the native and the virtualized case. For simulating the message exchange between robots, the request/response performance of TCP (\textit{TCP\_RR}) respectively UDP (\textit{UDP\_RR}) were measured. For simulating streaming data, the streaming performance of TCP (\textit{TCP\_Stream}) and UDP (\textit{UDP\_Stream}) were measured. To reflect the real-world use case, the experiment was performed on two hosts, which are connected wirelessly. The measurements were conducted using \textit{netperf}, which is an open-source network benchmark tool. The experiment was repeated a \textit{100} times. The statistical dispersions are shown in Tab.~\ref{tab:msd_network_performance}. The results reveal that the network performance in the native case is marginal better than the performance in the container case. For \textit{TCP\_RR}, the average deviation is \textit{5.64} percent and \textit{5.45} percent for \textit{UDP\_RR}. The average streaming performance of \textit{TCP} is \textit{1.12} percent better in the native case than the streaming performance in the container case. In the case of UDP, the streaming performance in the native case is \textit{5.76} percent better than the performance in the container case. It is important to note that, because UDP is connectionless, the streaming performance was determined using the received throughput of the receiver. It can be concluded that the performance loss of the virtualized connection is so small that it can be neglected in practice and is thus no limiting factor for the approach.
\begin{figure}[b]
\centering
\includegraphics[width=0.43\textwidth]{figure8}
\caption{Time Series of the Jain's Index of Allocated Costs - The x-axis represents the time in iterations and the y-axis represents the Jain's Index}
\label{fig:jain_index}
\end{figure}
\subsection{Optimality and Fairness}
The experiment regarding the optimality and the fairness of the approach were performed a \textit{100} times on a \textit{12}-node ARM-based cluster with a balanced workload. The optimality of the approach is shown in Fig.~\ref{fig:alloc_costs}. The figure illustrates the allocation of \textit{6} services to \textit{12} workers over \textit{100} iterations. The colorbar indicates the number of services a worker got allocated. The surface represents the average calculated costs over 100 iterations per worker, per service. The standard deviation over all costs is \textit{0.0156} and the coefficient of variation over all costs is \textit{0.0066}. A comparison of the allocations and the average costs shows that the services are allocated to the workers with the lowest costs and accordingly the lowest workload, which implies an optimal allocation of services to workers. The figure also reveals that, for this particular case, \textit{7} workers carry out the whole workload of the experiment, which seems not to be optimal with respect to the overall fairness of the approach. The \textit{Jain's Index}~\cite{jain1998} is a fairness measure that is typically used to reveal fair share of system resources. In this case, a large variance of allocated costs between the workers would result in a \textit{Jain's Index} that tends to 0. A small variance of costs and thus a balanced and fair allocation tends to a \textit{Jain's Index} of 1. The assumption that the approach not seems to be fair is supported by the results shown in Fig.~\ref{fig:jain_index}. It reveals that approx. 58\% of the workers consume 100\% of the costs. For future iterations of the system, a much better trade-off between optimality and fairness is preferable.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figure9}
\caption{Surface Plot of Average Cost and Bar Plot of Allocations - The surface of the plot indicates the average cost per worker, per service and the color and height of the bars indicate the number of allocations per worker, per service}
\label{fig:alloc_costs}
\end{figure}
\subsection{Runtime}
The following section considers the runtime of the approach. Because the time that is needed to join the swarm is very individual with regard to the environment (e.g. SSH-Access, Physical Access etc.), the comparable measure is the elapsed time from the start of the swarm to the start of the containers. The evaluation was performed for \textit{1} to \textit{12} workers and \textit{1} to \textit{12} services with the previously described cluster configuration. The result of the evaluation is shown in Fig.~\ref{fig:overall_eval}. The evaluation reveals that the time from the start of the swarm to the start of the container depends much more on the number of services that should be allocated than on the number of workers that are part of the swarm. To be precise, because of the sequential determination of costs for each service, the runtime of the approach increases linearly with the number of services. The costs for each service are calculated in a distributed and parallelized fashion, which results in a nearly constant runtime also in case of an increasing number of workers.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figure10}
\caption{Surface Plot of the Elapsed Time in Seconds - The figure illustrates the elapsed time in seconds from the start of the swarm to the start of the services, per service and per worker}
\label{fig:overall_eval}
\end{figure}
\section{CONCLUSIONS AND FUTURE WORK}
The present study discusses the increasing importance of reproducibility in robotics research, the approaches of the community to improve the situation and the challenges that still remains. The presented toolkit, called \textit{SwarmRob}, uses virtualization of robotics applications embedded in a novel experimental workflow to manage some of the challenges. Unlike other approaches discussed in the related work, \textit{SwarmRob} is designed to be open, flexible and extensible and is therefore easily adoptable for other researchers. The toolkit assists the researcher with the implementation of the workflow by abstracting the complexity of the used technologies in terms of orchestration, networking and initialization. The performed evaluations show that the use of containers in robotics has only a little impact on the performance of the underlying hardware, which is a critical requirement to ensure, on the one hand, the transferability of results and, on the other hand, the universal application capabilities of the toolkit. The evaluations also demonstrate the substantial time saving of the system, compared to manual experiment setups. The evaluation of the orchestration mechanism shows an optimal allocation of services to workers, which is to the detirement of fairness. Regarding this point, the measurements reveal opportunities for improvement. While currently, the initial definition of services and experiments is a cumbersome task, in future iterations, the system should be extented by a mechanism that allows to ``snapshot'' an entire robot and experiment automatically. This feature would follow up on the goal of an independent an easily accessible toolkit for reproducibility in experimental robotics research. While the system is currently under practical evaluation in a multi-robot search scenario~\cite{Sprute2017}, it should also be evaluated and evolved by the help and guidance of the community. Therefore, the entire toolkit will be publicly available at \url{https://iot-lab-minden.github.io/SwarmRob/}.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The shape of a galaxy's globally integrated 21-cm line profile encodes information about the kinematics of neutral hydrogen (HI) in its interstellar medium (ISM). Because the 21-cm transition is usually optically thin, an observed line profile can be directly translated into a line-of-sight velocity distribution (LOSVD) of HI gas in the galaxy. The shape of the 21-cm line thus depends on a galaxy's internal kinematics, and it will in general be different for disky, rotation-supported galaxies than for irregular galaxies in which gas is primarily supported by dispersion.
Most large spiral galaxies produce characteristic ``double-horned'' unresolved 21-cm line profiles, which arise naturally from differential Doppler effects in rotating disks \citep{Epstein_1964, Shostak_1977}. In fact, double-horned line profiles are not unique to 21-cm emission from HI in galaxies but are a generic consequence of rotating disk geometries and are produced in a wide range of astrophysical settings, including lines formed in accretion disks in some broad-line AGN \citep{Stella_1990, Storchi_2017}, protostellar disks \citep{Yamada_2009}, and accretion disks around massive stars \citep{Doazan_1965} and white dwarfs \citep{Tylenda_1981}. In many cases, the precise shape of such double-peaked emission lines can be used to constrain the rotation curve and/or spatial distribution of the emitting material.
Unlike large spiral galaxies, many low-mass galaxies do not produce double-horned profiles, instead exhibiting single-peaked, Gaussian-like profiles with sloping wings \citep[e.g.][]{Geha_2006, Begum_2008, Haynes_2011, Bradford_2015}. Single-peaked, Gaussian-like profiles are typically interpreted as evidence of lower rotation velocities and more disordered gas kinematics than is typical for more massive disk galaxies \citep{Singhal_2008, Stewart_2014, Papastergis_2016}. This suggests that in the absence of spatially resolved observations, the shape of a galaxy's unresolved HI profile can be used as a rough proxy for $V/\sigma$ (i.e., the ratio of ordered rotation to dispersion) of the galaxy's gas.
In contrast to resolved HI kinematic maps, which are expensive to obtain and have only been produced for of order one hundred galaxies in the local Universe \citep[e.g.][]{Walter_2008, Hunter_2012, Ott_2012}, unresolved global 21-cm line profiles can be efficiently obtained with short, single-dish observations. Unresolved HI profiles are already cataloged for a few $\times\, 10^4$ low-redshift galaxies \citep[e.g.][]{Paturel_2003, Meyer_2004, Begum_2008, Haynes_2011, Giovanelli_2015}, and next-generation radio surveys are expected to increase the size of unresolved HI samples by a factor of 10 within a decade \citep{Dewdney_2009, Duffy_2012, Li_2017}.
Besides their availability, another advantage of using unresolved 21-cm line profiles as probes of HI kinematics is that most unresolved surveys are morphologically blind, providing a homogeneous and unbiased sample of the local gas-rich galaxy population against which to test the predictions of theoretical models. This is in contrast to spatially resolved studies, which often preferentially target ``well-behaved'' disk galaxies. Mock unresolved HI observations are also straightforward to produce from simulations with minimal post-processing and can be compared directly to observations, without the need to fit kinematic models.
Both observational \citep{McGaugh_2000, Haynes_2011, Bradford_2015, Bradford_2016} and theoretical \citep{Obreschkow_2009, Obreschkow_2009b, Brook_2016, Maccio_2016, Brooks_2017} works have previously interpreted HI profiles from unresolved data. However, these works have typically not used the \textit{shape} of line profiles to extract kinematic information, but have instead quantified observed 21-cm lines with only two numbers: the integrated flux, which can be translated to total HI mass, and the line width, which to first order traces the depth of the gravitational potential. These parameters are not sensitive to whether the emitting gas is supported primarily by rotation or dispersion.
In this paper, we compare unresolved 21-cm line profiles from simulations to profiles of real galaxies, with the goal of comparing the simulated and observed galaxies' HI kinematics. We first show how the information encoded in the shapes of unresolved 21-cm line profiles can be leveraged to statistically constrain the gas kinematics of the low-redshift galaxy population. We use mock HI profiles produced from simulated galaxies, for which the internal gas kinematics are known a priori, as a guide for interpreting the line profiles of real galaxies. We then compare the line profiles of real galaxies to the mock profiles in order to compare the degree of rotation vs. dispersion support in the simulated galaxies to that in the observed population.
The remainder of this paper is organized as follows. In Section~\ref{sec:mock_obs}, we describe our procedure for mock-observing simulated galaxies and show how unresolved HI line shape is related to internal rotation vs. dispersion support. In Section~\ref{sec:obs_comp}, we quantify the shapes of observed HI line profiles and compare to our simulated galaxies. We summarize our results and discuss their implications in Section~\ref{sec:conclusion}. In Appendix~\ref{sec:toy_model}, we describe a simple analytic model for predicting the shape of unresolved HI line profiles. We discuss the relation between rotation/dispersion support and other galaxy properties in Appendix~\ref{sec:sfhs}.
\section{Mock observations}
\label{sec:mock_obs}
\begin{figure*}
\centering
\subfigure[simulated galaxies]{\label{fig:schematic}\includegraphics[width=\columnwidth]{schematic_unresolved_HI.pdf}}
\subfigure[real galaxies]{\label{fig:data_schematic}\includegraphics[width=\columnwidth]{bradford_and_sdss_image_and_hi_profile.pdf}}
\caption{\textbf{Left}: Effect of internal gas kinematics on the shape of galaxies' unresolved 21-cm lines. First (second) row shows the resolved zeroth (first) HI moment map for three low-mass galaxies in our sample, with increasing rotational support from left to right. Each map is produced with inclination $i=60$ deg. Third row shows the global HI mass-weighted line-of-sight velocity distribution. Bottom row shows the mock 21-cm spectrum (gray) and identified emission region (black; see Section~\ref{sec:mock_obs}). Unresolved spectra encode information about the galaxies' gas kinematics: rotation-supported galaxies produce double-horned line profiles with steep wings, while dispersion-supported galaxies produce single-peaked profiles with shallower wings. \textbf{Right}: Bottom panels show unresolved HI data for two low-mass galaxies ($M_{\rm star}\sim 10^8 M_{\odot}$), one with a Gaussian-like profile and one with a double-horned profile. Top panels show corresponding SDSS images.}
\end{figure*}
We analyze the same suite of cosmological hydrodynamic zoom-in simulations from the FIRE project \citep{Hopkins_2014} that was studied by \citet[][hereafter E18]{ElBadry_2018}. The simulations were run with the \texttt{GIZMO} hydrodynamics code \citep{Hopkins_2015} in the Lagrangian ``meshless finite mass'' (MFM) mode, using the FIRE-2 model for galaxy formation and feedback\footnote{See the FIRE project website at \href{http://fire.northwestern.edu. A public version of the \texttt{GIZMO} code is available at \href{http://www.tapir.caltech.edu/~phopkins/Site/GIZMO.html}{\url{http://www.tapir.caltech.edu/~phopkins/Site/GIZMO.html}}.} \citep{Hopkins_2017}. For details regarding the physical processes modeled in these simulations and their numerical implementation, we refer the reader to \citet{Hopkins_2017} and \citet{Hopkins_2017b}.
This simulation suite contains 24 isolated galaxies, with $z=0$ stellar masses spanning $M_{\rm star}=10^{6.3-11.1} M_{\odot}$, residing in halos with $z=0$ masses $M_{\rm 200m} = 10^{9.9-12.2}M_{\odot}$. Here $M_{\rm 200m}$ is the total mass within $R_{\rm 200m}$, which is the radius within which the matter density is 200 $\times$ the mean matter density at $z=0$. \citetalias{ElBadry_2018} studied the gas properties of these galaxies in detail, showing that they exhibit a diverse range of gas kinematics and morphologies: thin, rotation-supported disk galaxies dominate at $M_{\rm star} \gtrsim 10^{10} M_{\odot}$, while puffy, dispersion-supported irregular galaxies dominate at lower masses. The properties of the individual simulated galaxies studied in this work are summarized in \citetalias{ElBadry_2018} (their Table 1).
We mock-observe these simulated galaxies in order to produce synthetic unresolved HI profiles that can be compared to observations. The observational sample to which we compare is the homogeneously reduced sample of unresolved HI observations produced by \citet[][hereafter B15]{Bradford_2015}; our mock observations are designed to emulate these data. Most of the lower-mass galaxies ($M_{\rm star} \lesssim 10^{8.5} M_{\odot}$) in the observational sample were observed with a typical velocity resolution of $\Delta v = 0.65\,{\rm km\,s^{-1}}$ as part of the low-mass HI program introduced in \citet{Geha_2006}; see Figure 3 of \citetalias{Bradford_2015}. Most of the higher-mass galaxies were originally observed with a velocity resolution of $\Delta v = 5.5\,{\rm km\,s^{-1}}$ through the ALFALFA survey \citep{Haynes_2011} and subsequently re-reduced by \citetalias{Bradford_2015}. Spectra collected through the low-mass HI program were Hanning smoothed to a resolution of $5\,{\rm km\,s^{-1}}$ by \citetalias{Bradford_2015}; those from the ALFALFA survey, to a resolution of $10\, {\rm km\,s^{-1}}$. In this work, we only analyze the subsample of the \citetalias{Bradford_2015} data with a smoothed signal-to-noise ratio ${\rm SNR} > 15$ (see Equation~\ref{eqn:snr_eqn}); this subsample contains 2002 galaxies with a median SNR of 25.
For our main analysis, we mock-observe all simulated galaxies at $z=0$ from 100 random viewing angles distributed uniformly on the unit sphere. For a given viewing angle and desired signal-to-noise ratio, we apply the following procedure to create a mock-HI profile:
\begin{enumerate}
\item We consider all HI gas within a circular aperture of diameter 80 kpc centered on the galaxy. This is comparable to the 3.5 arcmin Arecibo beam diameter for an object at a distance of 70 Mpc, which is typical for our sample of observations. Our results are not sensitive to this choice, because the aperture is larger than the region containing significant HI for all the simulated galaxies in our sample. We calculate the HI mass of each gas element in the simulation as in \citetalias{ElBadry_2018}, excluding ionized gas and cold gas likely to be molecular.\footnote{Gas with $T<300\,\rm K$ and $n_{\rm H} > 10\,{\rm cm^{-3}}$ is treated as molecular. Our results are not sensitive to excluding molecular gas, as only a small fraction ($\sim$10\% by mass in MW-mass galaxies, and less at lower masses) of the ISM passes these molecular criteria at $z=0$. We note that these molecular fractions are likely somewhat lower than typical values for observed galaxies.} The neutral hydrogen fraction of each gas element is computed by \texttt{GIZMO}, which incorporates ionization from a spatially uniform, redshift-dependent UV background \citep{FG_2009} and approximates the radiation field due to local sources (see \citealt[][]{Hopkins_2017}).
\item We construct the HI mass-weighted LOSVD of all HI gas in this aperture, binned at a velocity resolution $\Delta v$. To account for thermal broadening, we treat each gas element as a Gaussian in velocity space with $\sigma = c_s=\sqrt{kT/(\mu m_p)}$, where $c_s$ is the isothermal sound speed.
We choose $\Delta v$ to match the observations against which we compare: for galaxies with $M_{\rm star} < 10^{8.5} M_{\odot}$, we use $\Delta v = 0.65\,\rm km\,s^{-1}$; for more massive galaxies, we take $\Delta v = 5.5\,\rm km\,s^{-1}$.
\item For mock-observations of a target at a distance $D$, we compute the 21-cm flux density in the $i$-th velocity channel using the formula
\begin{equation}
\label{eqn:S_nu}
\left(\frac{S_{{21,i}}}{{\rm Jy}}\right)=\frac{1}{2.36\times10^{5}}\left(\frac{m_{{\rm HI},i}}{M_{\odot}}\right)\left(\frac{\Delta v}{{\rm km\,s^{-1}}}\right)^{-1}\left(\frac{D}{{\rm Mpc}}\right)^{-2},
\end{equation}
implicitly assuming the 21-cm transition is optically thin \citep[e.g.][]{Haynes_1984}. Here $m_{{\rm HI},i}$ is the HI mass in the $i$-th velocity channel. In practice, we solve for $D$ for each observation in order to yield the desired signal-to-noise ratio, as quantified in Equation~\ref{eqn:snr_eqn}.
\item We add Gaussian noise to the spectrum with a fixed dispersion of 0.004 Jy, which is typical for the observations against which we compare. We then Hanning smooth the spectrum to a resolution consistent with the observations: for galaxies with $M_{\rm star} < 10^{8.5} M_{\odot}$, we smooth to a resolution of $5\,\rm km\,s^{-1}$; for higher-mass galaxies, to a resolution of $10\,\rm km\,s^{-1}$.
\item We identify the HI emission region of the spectrum in velocity space as follows. We construct a copy of the noisy spectrum that is Hanning smoothed to twice the resolution adopted in (iv). We define the edges of the emission region as the first point on either side of the emission peak where this doubly-smoothed spectrum drops below zero flux; we set the emission to zero outside this region.
Denoting the peak flux density $f_{\rm peak}$, we fit a quadratic polynomial to the wings on either side of the line profile between $0.5f_{\rm peak}$ and the edges of the emission region, replacing the flux in this region with the polynomial interpolation. This procedure was found by \citetalias{Bradford_2015} to minimize spurious signal due to noise in the wings of the profile; we adopt it here for consistency.
\item We calculate the total integrated flux of the HI line as $S_{21,{\rm tot}}=\int_{-\infty}^{\infty}S_{{\rm 21}}\left(v\right)\,{\rm d}v$, and $W_{\rm 50}$, the velocity width at 50\% of the maximum flux, as the difference between the velocities above and below the peak flux at which the interpolated emission profile reaches $0.5 f_{\rm peak}$. We then calculate the signal-to-noise ratio of the integrated emission using the same definition used by \citetalias{Bradford_2015}:
\begin{equation}
\label{eqn:snr_eqn}
{\rm SNR}=\left(\frac{S_{21,{\rm tot}}}{W_{50}}\right)\frac{w_{{\rm smo}}^{1/2}}{\sigma_{{\rm rms}}}.
\end{equation}
Here $w_{\rm smo}=W_{50}/(10\rm\,km\,s^{-1})$, and $\sigma_{\rm rms}$ is the standard deviation of the spectrum outside the emission region, after the smoothing in (iv) is applied. Note that this definition represents the total signal-to-noise ratio of the HI line integrated over all channels; the SNR in any one channel will be lower than the value yielded by Equation~\ref{eqn:snr_eqn}.
\end{enumerate}
\subsection{Quantifying Unresolved Line Shapes}
\label{sec:kurt}
We use two distinct but related statistics to quantify the shape of galaxies' integrated HI lines.
\subsubsection{Kurtosis}
The kurtosis, which we denote $\kappa$, is a measure of the ``peakedness'' of the line; i.e., how centrally concentrated the emission is in velocity space. Single-peaked, Gaussian-like HI lines have relatively high kurtosis, while line profiles with more weight in the edges, such as the characteristic double-horned profile, have lower kurtosis. The kurtosis of galaxies' integrated HI lines was previously used as a proxy for $V/\sigma$ by \citet{Papastergis_2016}, who used a cut of $\kappa \leq -1.2$ to select rotation-supported galaxies in a study of the local baryonic Tully-Fisher relation (BTFR). For an observed line profile $S_{21}(v)$, we define the kurtosis as
\begin{equation}
\label{eqn:kurt}
\kappa=\frac{\mu_{4}}{\mu_{2}^{2}}-3.
\end{equation}
Here $\mu_n$ is the $n$th central moment:
\begin{equation}
\label{eqn:cen_mom}
\mu_{n}=\frac{\int_{-\infty}^{\infty}S_{{\rm 21}}\left(v\right)\left(v-\overline{v}\right)^{n}\,{\rm d}v}{\int_{-\infty}^{\infty}S_{{\rm 21}}\left(v\right)\,{\rm d}v},
\end{equation}
where $\overline{v}$ is the flux-weighted mean velocity:
\begin{equation}
\label{eqn:mean_v}
\overline{v}=\frac{\int_{-\infty}^{\infty}v S_{{\rm 21}}\left(v\right)\,{\rm d}v}{\int_{-\infty}^{\infty}S_{{\rm 21}}\left(v\right)\,{\rm d}v}.
\end{equation}
The constant $-3$ in Equation~\ref{eqn:kurt} sets the kurtosis scale so that a Gaussian has $\kappa=0$. We find that for both our mock observations and real galaxies, the majority of HI lines have $-1.5 \lesssim \kappa \lesssim 0$.
\subsubsection{Steepness of profile wings}
To quantify the steepness of a profile's wings, we first calculate $W_{20}$ and $W_{50}$, the width of the line profile at 20\% and 50\% of its peak value. We then measure their scaled difference, $\Delta W \equiv (W_{\rm 20} - W_{\rm 50})/W_{50}$. This quantity is a measure of the slope of the line profile wings: for profiles with steep wings, $W_{20}$ and $W_{50}$ are very similar, and $\Delta W$ is small. For profiles with sloping wings, $W_{\rm 20}$ is significantly larger than $W_{50}$, so $\Delta W$ is larger; for a Gaussian, $\Delta W = 0.52$. Most of the line profiles in both our observed and simulated samples have $0 \lesssim \Delta W \lesssim 1$.
\subsection{Relation between line shape and kinematics}
In Figure~\ref{fig:schematic}, we show how a galaxy's gas kinematics change the shape of its unresolved 21-cm profile. We show column density and velocity maps of three simulated galaxies spanning the range of gas kinematics and morphology in our sample, ranging from irregular and dispersion supported (\texttt{m11a}, with $M_{\rm star} = 10^{8}M_{\odot}$) to rotationally supported but with significant disordered motion (\texttt{m11h}, with $M_{\rm star} = 4\times 10^{9}M_{\odot}$), to rotationally supported and morphologically disky (\texttt{m11b}, with $M_{\rm star} = 10^{8}M_{\odot}$). Details regarding the construction of these maps, as well as comparable maps for all the galaxies in our sample, can be found in \citetalias{ElBadry_2018}. Figure~\ref{fig:schematic} shows that the simulated galaxies' unresolved HI profiles become single-peaked, with increasingly sloping wings and higher kurtosis, as they become more dispersion supported.
In Figure~\ref{fig:data_schematic}, we show the unresolved HI profiles of two galaxies in our observed sample, both with $M_{\rm star}\sim 10^8 M_{\odot}$. One of these galaxies exhibits a double-horned profile with low kurtosis, (qualitatively similar to \texttt{m11b}), while the other produces a single-peaked profile with higher kurtosis (qualitatively similar to \texttt{m11a}). Spatially resolved HI kinematics are not available for our observed sample, so we instead show the galaxies' optical images from the NASA-Sloan Atlas, a reprocessing of SDSS DR8 \citep{Blanton_2011}. The optical morphology of the target with the double-peaked 21-cm profile is clearly disky, while that of the target with the Gaussian-like profile appears clumpy and irregular. Stellar morphology from optical images is an imperfect proxy for HI morphology, and we do not carry out any quantitative analysis on optical images. However, we find that on average, galaxies with double-peaked HI profiles exhibit diskier optical morphologies upon visual inspection than those with single-peaked profiles.
\begin{figure}
\includegraphics[width=\columnwidth]{kurtosis_vs_vsig_w20_w50.pdf}
\caption{Kurtosis (top) and steepness of the wings (bottom) of unresolved 21-cm line profiles of simulated galaxies vs. $V/\sigma$ (from mock-IFU measurements; Equation~\ref{eqn:v_sig}). Error bars show 68\% scatter across 100 different viewing angles. Disky, rotationally supported galaxies with higher $V/\sigma$ have systematically lower kurtosis ($\kappa \lesssim -1$) and steeper profile wings ($(W_{20}-W_{50})/W_{50}\lesssim 0.2$) than dispersion supported galaxies. Rotation- and dispersion-supported galaxies can thus be separated statistically based only on their unresolved line profile shape.}
\label{fig:vsig}
\end{figure}
To determine quantitatively how the shape of the unresolved HI line is related to a galaxy's \textit{resolved} gas kinematics, we mock-observe all our simulations from 100 random viewing angles and calculate the kurtosis, $\kappa$, wing steepness, $\Delta W$, and the degree of rotation vs. dispersion support, $V/\sigma$, for each viewing angle. To calculate $V/\sigma$ for a given viewing angle, we first construct resolved mock HI velocity maps with spatial resolution 0.1 kpc (see \citetalias{ElBadry_2018} for details) and then define \begin{equation}
\label{eqn:v_sig}
V/\sigma=\frac{\sum_{j}m_{j}\left|V_{{\rm los},j}\right|}{\sum_{j}m_{j}\sigma_{{\rm los},j}}.
\end{equation}
Here $m_{j}$ is the HI mass in a given spatial pixel, $V_{{\rm los},j}$ and $\sigma_{{\rm los},j}$ are the HI mass-weighted mean line-of-sight velocity and velocity dispersion in a pixel, and the sum is over all spatial pixels in the data cube. We take this definition from \citet{Obreja_2016}. Although $V/\sigma$ as calculated in Equation~\ref{eqn:v_sig} is dependent on the size of spatial pixels; i.e., the beam size in resolved observations, it has the advantage of being defined in terms of quantities measured directly from the simulation, without requiring any kinematic model fitting. Like $\kappa$ and $\Delta W$, $V/\sigma$ is a line-of-sight dependent quantity. Projection effects can cause rotational velocities to contribute significantly to $\sigma$, so $V/\sigma$ values calculated from Equation~\ref{eqn:v_sig} are typically lower than values calculated from cylindrically-averaged 3D quantities \citepalias[see][]{ElBadry_2018}.
In Figure~\ref{fig:vsig}, we plot $\kappa$ and $\Delta W$ for our simulated galaxies as a function of $V/\sigma$. For all quantities, we show the median and middle 68\% values across 100 random viewing angles. $\kappa$ and $\Delta W$ are measured from mock unresolved HI profiles with a SNR of 25, the median value for our observed sample. In agreement with the qualitative trend apparent in Figure~\ref{fig:schematic}, we find that $\kappa$ and $\Delta W$ decrease nearly monotonically with increasing $V/\sigma$, indicating that these quantities can be used as a proxy for $V/\sigma$. We find that $\kappa$ and $\Delta W$ are also both anticorrelated with the intrinsic, three-dimensional $V_{\rm rot}/\sigma$, as defined in \citetalias{ElBadry_2018}. However, the correlation is somewhat weaker than the correlation with $V/\sigma$ as defined in Equation~\ref{eqn:v_sig} because the unambiguously rotation-supported galaxies all have $\kappa \sim -1.4$ and $\Delta W \sim 0.1$ but have a wider range of $V_{\rm rot}/\sigma$, with $6 \lesssim V_{\rm rot}/\sigma \lesssim 12$.
Figure~\ref{fig:vsig} also shows that, for the simulated galaxies, $V/\sigma$, $\kappa$, and $\Delta W$ are all strong functions of mass, with higher-mass galaxies having higher $V/\sigma$ and lower $\kappa$ and $\Delta W$. However, the trend of decreasing kurtosis and increasingly steep profile wings at higher $V/\sigma$ is \textit{not} simply a result of the joint dependence on mass: the lowest-mass galaxy in our sample that is strongly rotation-supported (\texttt{m11b}, with $M_{\rm star}=10^8 M_{\odot}$) still has low $\kappa$ and $\Delta W$. For more detailed discussion of trends in gas kinematics with mass for this sample of simulated galaxies, we refer to \citetalias{ElBadry_2018}.
\begin{figure}
\includegraphics[width=\columnwidth]{kurtosis_vs_inclination_snr_alfalfa.pdf}
\caption{\textbf{Left}: Effect of signal-to-noise ratio on 21-cm line profile kurtosis measurements for a simulated disky, rotationally supported galaxy (blue) and an irregular, dispersion supported galaxy (red). Dashed line shows $\rm SNR=25$, the median value for our observed sample. At each SNR, we mock-observe each galaxy from 100 random viewing angles and generate 100 different noise realizations for each viewing angle. We plot the median and middle 68\% scatter for the resulting $10^4$ profiles generated at each SNR. \textbf{Right}: Here we vary inclination while fixing $\rm SNR=25$, the median for the observed sample in Figures~\ref{fig:kurt_vs_mstar} and~\ref{fig:W20_W50}. $\kappa$ is robust to changes in SNR down to ${\rm SNR}\sim 15$. $\kappa$ does vary with inclination for disky systems, but the effects are large only for rare, nearly face-on viewing angles.}
\label{fig:snr}
\end{figure}
Although the non-negligible 68\% scatter in Figure~\ref{fig:vsig} shows that $\kappa$ and $\Delta W$ do vary across viewing angles, rotation- and dispersion-supported systems are still separated on average. In Figure~\ref{fig:snr}, we show explicitly how $\kappa$ varies with SNR (left) and inclination (right). For each inclination and SNR value, we compute the kurtosis of $10^4$ mock HI profiles for \texttt{m11a} and \texttt{m11b}, two of the galaxies shown in Figure~\ref{fig:schematic}. These galaxies both have $M_{\rm star} \approx 10^8 M_{\odot}$ and $M_{\rm 200m}\approx 10^{10.7}M_{\odot}$, but \texttt{m11b} is rotation-supported while \texttt{m11a} is dispersion-supported. The left panel of Figure~\ref{fig:snr} shows that, at ${\rm SNR} > 15$ (the minimum SNR of our observed sample), kurtosis measurements for a given system are robust: the typical $\kappa$ values for the two galaxies remain well-separated, and the spread of $\kappa$ values across different SNR is less than that across different viewing angles at fixed SNR.
The right panel of Figure~\ref{fig:snr} shows how kurtosis varies with inclination angle, which is defined in terms of the net angular momentum vector of HI. For the disky, rotation-supported galaxy, kurtosis varies significantly with inclination: it has $\kappa \sim -1.2$ when viewed edge on, but $\kappa \sim 0$ when viewed face on. This is unsurprising, since the face-on disk has no rotation projected along the line of sight; generically, rotation-supported disk galaxies will produce single-peaked, Gaussian-like profiles if viewed completely face-on.
Fortunately, most rotation-supported galaxies will nevertheless produce 21-cm profiles with low kurtosis, because face-on viewing angles are intrinsically rare: for random disk orientations, the probability distribution of inclinations is $p\left(i\right)\,{\rm d}i=\sin i\,{\rm d}i$, meaning that far more galaxies are observed nearly edge-on than nearly face-on. Because inclination effects on kurtosis measurements only become significant at $i \lesssim 40$ degrees (representing 23\% of viewing angles), rotationally supported galaxies will produce unresolved HI profiles with low kurtosis for most viewing angles.
For the dispersion-supported galaxy, there is no significant trend in $\kappa$ with inclination, though there is still significant scatter across different viewing angles at fixed inclination. This is unsurprising, since dispersion-supported systems have no unambiguously preferred kinematic axis. Most lines of sight for such systems produce single-peaked line profiles roughly resembling Gaussians, but large-scale inflows and outflows of HI, as well as irregular HI distributions, can produce a variety of irregular and asymmetric profiles.
Although it is not shown in Figure~\ref{fig:snr}, we find that $\Delta W$ has a similar scaling with SNR and inclination to $\kappa$. Due to viewing angle effects, the mapping from line shape as quantified by $\kappa$ or $\Delta W$ to $V/\sigma$ is not unique. However, since real galaxies should have random orientations, the observed distribution of $\kappa$ and $\Delta W$ values places a statistical constraint on the observed population's distribution of $V/\sigma$ values, even though it is not always possible to unambiguously measure the kinematics of a particular galaxy from its unresolved line profile.
\subsubsection{Effect of rotation curve shape}
The detailed shape of a galaxy's unresolved 21-cm line profile depends not only on rotation vs. dispersion support, but also on the shape of its rotation curve. We investigate this dependence in detail in Appendix~\ref{sec:toy_model}. Our primary conclusion is that while a double-peaked 21-cm profile is almost always indicative of rotational support, a single-peaked, Gaussian-like profile can arise either due to dispersion support \textit{or} due to a linearly rising rotation curve: if a galaxy's circular velocity curve rises sufficiently slowly and does not begin to flatten in the region occupied by the galaxy, a single-peaked profile can be produced even in the limit where HI is completely rotationally supported and follows the circular velocity curve, $v_c(r)=\sqrt{r\times{\rm d\Phi}/{\rm d}r}$. For the simulated galaxies, we can measure the HI rotation curve and $v_c(r)$ directly, and we find that Gaussian-like profiles are due to dispersion support, not linearly rising rotation curves. However, for observed galaxies with Gaussian-like HI profiles, it is generally not possible to distinguish between a dispersion-supported system and a rotation-supported systems with a linearly rising rotation curve.
\section{Comparison to observations}
\label{sec:obs_comp}
\begin{figure}
\includegraphics[width=\columnwidth]{kurt_and_fgas_vs_mstar.pdf}
\caption{\textbf{Top}: 21-cm line kurtosis measurements for simulated galaxies compared to $\sim$2000 observed galaxies from \citetalias{Bradford_2015}. Error bars show 68\% scatter across 100 random viewing angles; dashed lines show median and 68\% scatter for observed galaxies in each mass bin. At $M_{\rm star} > 10^{10} M_{\odot}$, where all the simulated galaxies have rotation-supported HI disks, the simulated galaxies produce line profiles with kurtosis values very similar to the observed galaxy population. At $M_{\rm star} = 10^{8-10} M_{\odot}$, most of the simulated profiles have higher $\kappa$ than the observed population, corresponding to lower $V/\sigma$ (see Figure~\ref{fig:vsig}).
\textbf{Bottom}: Gas fractions. Most low-mass simulated galaxies with higher $\kappa$ than the median of the observed sample also have somewhat lower gas fractions than typical observed galaxies. }
\label{fig:kurt_vs_mstar}
\end{figure}
\subsection{Kurtosis}
We now compare the unresolved HI profiles generated from our simulations to those of real galaxies. Our observed sample consists of 2002 21-cm HI profiles with ${\rm SNR} > 15$. Almost all of these profiles were presented in \citetalias{Bradford_2015}; we also include in our sample observations of few dozen low-mass galaxies that were obtained since the publication of that work and were reduced following the same procedure described in \citetalias{Bradford_2015}. The majority of this observational sample is completely morphologically blind. 210 of the profiles first presented by \citetalias{Bradford_2015} were selected with a weak selection criterion against galaxies with inclinations near 0. However, we find the full observational sample has a distribution of inclinations (computed from optical axis ratios by \citetalias{Bradford_2015}) consistent with random orientations.
We compute the kurtosis of the smoothed emission regions of these profiles, as defined in Section~\ref{sec:mock_obs}, using Equation~\ref{eqn:kurt}. Stellar masses were taken from the NASA-Sloan atlas \citep{Blanton_2011}. The median SNR of these data is 25, so we adopt ${\rm SNR=25}$ in generating mock HI profiles for the simulated galaxies. In this observational sample, there is no significant correlation between SNR and $\kappa$ or $M_{\rm star}$.
In Figure~\ref{fig:kurt_vs_mstar}, we show the distribution of kurtosis values for these galaxies as a function of mass, overplotting measurements for our simulated galaxies. At $M_{\rm star} \gtrsim 10^{10} M_{\odot}$, most observed galaxies have $\kappa \lesssim -1$, indicating that they have rotation-supported kinematics. This is also true for the simulated galaxies, which fall on top of the median observed relation. At this mass scale, the scatter in the kurtosis measurements of different profiles for individual galaxies, which is due primarily to different viewing angles, is almost as large as all the scatter in the observed sample. This suggests that gas-rich galaxies at $M_{\rm star} \gtrsim 10^{10} M_{\odot}$ likely all have similar rotation-supported HI kinematics.
At lower masses, particularly $M_{\rm star} = 10^{8-10} M_{\odot}$, most of the simulated galaxies have significantly higher kurtosis than at Milky Way (MW) masses, with $\kappa \sim -0.5$. A similar trend toward higher $\kappa$ at lower masses is evident in the observations, but the trend is less pronounced than for the simulated galaxies: more than half of the observed galaxies have $\kappa < -1$ over this mass range, indicating rotation-supported kinematics. Of the 8 simulated galaxies with $M_{\rm star} = 10^{8-10} M_{\odot}$, 7 have median $\kappa$ values above the median observed $\kappa - M_{\rm star}$ relation. The only exception is \texttt{m11b}, which has a rotation-supported disk (Figure~\ref{fig:schematic}) and falls below the median observed relation.
\subsection{Gas Fractions}
In the bottom panel of Figure~\ref{fig:kurt_vs_mstar}, we compare the gas fractions of the simulated galaxies to observations. For both the observed and simulated galaxies, we define $M_{{\rm gas}} \equiv 1.4M_{{\rm HI}}$, where the factor of 1.4 accounts for helium, and then plot $f_{{\rm gas}}=M_{{\rm gas}}/(M_{{\rm gas}}+M_{{\rm star}})$. For both the observed and simulated samples, $f_{\rm gas}$ decreases with increasing $M_{\rm star}$. The simulations with $M_{\rm star} > 10^{10} M_{\odot}$ have gas fractions similar to the median of the observed population, but at lower masses, most of the simulated galaxies have somewhat lower $f_{\rm gas}$ than the median observed relation.
In particular, most of the low-mass simulated galaxies with higher-than-average $\kappa$ (i.e., enhanced dispersion support) also have lower gas fractions than the median of the observed population. The lowest-mass simulated galaxy with a clear rotation-supported disk (\texttt{m11b}, with $M_{\rm star}=10^{8} M_{\odot}$) also has a higher-than-average gas fraction ($f_{\rm gas} = 0.94$). This suggests that for the simulated galaxies, increased dispersion supported is likely connected to reduced gas content. Intriguingly, we find no correlation between $f_{\rm gas}$ and $\kappa$ at fixed mass in the observed sample: at all masses, observed galaxies with lower-than-average gas fractions appear to have similar gas kinematics to those with higher-than-average gas fractions. We return to this issue in Section~\ref{sec:biases}.
\begin{figure}
\includegraphics[width=\columnwidth]{obs_w20_and_w50_vs_sim.pdf}
\caption{Spatially-unresolved baryonic Tully-Fisher relation as probed by $W_{20}$ (top) and $W_{50}$ (middle), without inclination corrections. Error bars show 68\% scatter across 100 viewing angles. Bottom panel shows $\Delta W = (W_{20}-W_{50})/W_{50}$, an estimate of the slope of the wings of the HI profile. The simulations reproduce the observed BTFR as probed by $W_{20}$ well. At $\sim$MW masses ($M_{\rm bar} > 10^{10.5}M_{\odot}$), the simulated galaxies also have similar $W_{50}$ and $\Delta W$ to the observed galaxies. At lower masses, most simulated galaxies have $W_{20}$ similar to the observed population but lower $W_{50}$ and thus higher $\Delta W$, indicating less-steep profile wings and higher dispersion than average for the observed galaxies.}
\label{fig:W20_W50}
\end{figure}
\subsection{Line Widths and BTFR}
The width of a galaxy's unresolved 21-cm line, as quantified by $W_{20}$ or $W_{50}$, is typically used as a diagnostic of its maximum circular velocity, and thus, the depth of its gravitational potential. Velocity widths are used to construct the BTFR for gas-rich galaxies \citep[e.g.][]{McGaugh_2000}. In this context, individual line widths are to first order agnostic to rotation vs. dispersion support, since the range of line-of-sight velocities in a virialized system depends primarily on the depth of the gravitational potential. Different BTFR slopes can be constructed depending on whether $W_{20}$ or $W_{50}$ is used; see \citet{Bradford_2016} for a review.
To compare the BTFR for our simulated galaxies with observations, we follow \citetalias{Bradford_2015} in defining baryonic masses for the simulated galaxies as $M_{\rm bar} = M_{\rm star}+1.4M_{\rm HI}$. We correct the observed $W_{20}$ and $W_{50}$ values for redshift broadening by dividing them by $(1+z)$; this correction is always small because $z\ll 1$. Most previous works on the BTFR have also attempted to correct for inclination effects by dividing observed line widths by $\sin i$, where inclination is calculated from galaxies' projected optical axis ratios with the assumption of an intrinsic thin-disk morphology. Such corrections can be problematic for low-mass galaxies, both because optical axis ratios are an imperfect tracer of HI morphology, and because there is no guarantee that galaxies' HI morphologies are intrinsically disky. We therefore do not apply any inclination correction to the observed or simulated line widths, instead accounting for the random orientations of the observed galaxies by mock-observing the simulated galaxies along many random lines of sight. As a result, the intrinsic scatter in the observed line width -- mass relation is larger than that found when an inclination correction is applied (see \citetalias{Bradford_2015}).
In Figure~\ref{fig:W20_W50}, we compare the BTFR predicted by our simulations as quantified by $W_{20}$ (top) and $W_{50}$ (middle) to that of the observational sample from \citetalias{Bradford_2015}. At $M_{\rm bar} > 10^{10.5} M_{\odot}$, the simulated HI profiles have both $W_{\rm 20}$ and $W_{50}$ in good agreement with the observed sample. At lower masses, the simulated $W_{20}$ values remain in reasonable agreement with the observed values, while the simulated $W_{50}$ values are on average lower than the median of the observed values. Because our simulated galaxies -- and the observed galaxies, to a lesser extent -- preferentially produce 21-cm profiles with sloping wings at lower masses, the slope of the BTFR measured by $W_{50}$ is steeper than that measured by $W_{20}$.
The bottom panel of Figure~\ref{fig:W20_W50} shows $\Delta W$. At $M_{\rm bar} \gtrsim 10^{10.5} M_{\odot}$, values for the simulated and observed galaxies agree well. At $M_{\rm bar}=10^{8.5-10} M_{\odot}$, most of the simulated galaxies have higher $\Delta W$, corresponding to shallower HI profile wings, than typical observed galaxies. Consistent with Figure~\ref{fig:kurt_vs_mstar}, the higher $\Delta W$ for low-mass simulated galaxies suggests that they have enhanced dispersion support compared to the observed galaxies.
Measured from unresolved line widths, the BTFR is relatively insensitive to rotation vs. dispersion support. This is in contrast to measurements from spatially resolved gas kinematics \citep[e.g.][]{Simons_2015, Bloom_2017}, which can distinguish between rotation and dispersion. For unresolved observations, simulated galaxies can be expected to match the observed BTFR as long as they fall on a realistic $M_{\rm bar}-M_{\rm halo}$ relation and are approximately virialized, irrespective of the mode of kinematic support. The simulated galaxies reproduce the observed BTFR as measured by $W_{20}$ well, despite the apparent excess dispersion support in low-mass galaxies. On the other hand, the dispersion-supported simulated galaxies fall slightly below the observed BTFR as measured by $W_{50}$. This occurs because $W_{50}$ is significantly less than $W_{20}$ for a dispersion-supported system with a Gaussian LOSVD, but the two quantities are similar for a rotation-supported systems with steep wings. This suggests that the scatter in the BTFR can be reduced by measuring $W_{20}$ as opposed to $W_{50}$, or by inflating $W_{50}$ values for targets with large $\Delta W$ and $\kappa$. Previous studies with observational data \citep{Bradford_2016, Papastergis_2016} have indeed found this to be the case. Incorporating measurements of line shape as quantified by $\kappa$ and $\Delta W$ in the observed BTFR will likely prove useful in minimizing the relation's scatter for the purpose of using it as a redshift-independent distance estimator \citep[e.g.][]{Obreschkow_2013}.
\section{Summary and Discussion}
\label{sec:conclusion}
We have investigated the connection between the shape of galaxies' spatially unresolved 21-cm line profiles and the kinematics of their gas. Rotation-supported galaxies typically produce double-peaked line profiles with steep wings, while dispersion-supported galaxies produce single-peaked Gaussian-like profiles with more sloping wings (see Figure~\ref{fig:schematic}). Using a suite of simulations from the FIRE project, we showed that the kurtosis and wing steepness of galaxies' unresolved HI lines can thus be used as a proxy for $V/\sigma$ (Figure~\ref{fig:vsig}). Although the mapping between line shape and $V/\sigma$ also depends on viewing angle (Figure~\ref{fig:snr}) and on the shape of galaxies' rotation curves (Appendix~\ref{sec:toy_model}), marginalization over these parameters with a large observational sample provides statistical constraints on the degree of rotational support in the local galaxy population, without requiring spatially resolved data.
Unresolved 21-cm line profiles provide a homogeneous, unbiased observational sample with which to test the predictions of simulations. Because mock observations of unresolved HI are straightforward to carry out with minimal assumptions and no model fitting, they allow for robust comparison of simulations and observational data on equal terms. We compared mock 21-cm line profiles of our simulated galaxies to those of a sample of $\sim$2000 low-redshift observed galaxies, quantifying their shape through the measured kurtosis (Figure~\ref{fig:kurt_vs_mstar}) and the difference between their measured 20\% and 50\% velocity widths (Figure~\ref{fig:W20_W50}). The primary results of these comparisons are as follows:
\begin{enumerate}
\item \textit{Kinematics in Milky Way-mass galaxies}: At $M_{\rm star} = 10^{10-11}M_{\odot}$, the simulated galaxies all have rotation-supported gas disks and produce HI profiles with shapes in excellent agreement with the observed population. At this mass scale, the scatter in observed kurtosis values is small enough to be explained almost entirely by inclination effects, indicating that most HI-rich galaxies with $M_{\rm star} = 10^{10-11} M_{\odot}$ have similar, rotation-supported gas kinematics.
\item \textit{Kinematics in lower-mass galaxies}: At $M_{\rm star}=10^{8-10} M_{\odot}$, most of the simulated galaxies have higher kurtosis values and less-steep wings than more massive galaxies. This is true for the observed sample as well. However, the simulated lower-mass galaxies appear to have somewhat more dispersion support than the observed systems (Figures~\ref{fig:kurt_vs_mstar} and ~\ref{fig:W20_W50}). This is not universally true: one low-mass galaxy (\texttt{m11b}, with $M_{\rm star} = 10^{8} M_{\odot}$) does have a kinematically cold, rotationally supported gas disk and produces a characteristic double-horned HI profile. However, 7 of the 8 simulated galaxies in our sample with $M_{\rm star}=10^{8-10}M_{\odot}$ produce HI profiles with higher kurtosis and shallower wings than the median of the observed sample.\footnote{We also note that the \texttt{m11b} halo has a relatively high spin parameter, with $\lambda = 0.077$, which is $2\sigma$ above the mean for all dark matter halos \citep{Bullock_2001}; this may be related to its diskier-than-average morphology and kinematics.}
\item \textit{Gas fractions}: The simulated MW-way mass galaxies have gas fractions similar to the observed galaxies. At lower masses ($M_{\rm star}=10^{8-10}M_{\odot}$), the gas fractions of most simulated galaxies are somewhat lower than the median of the observed population (Figure~\ref{fig:kurt_vs_mstar}). For the simulated low-mass galaxies, $f_{\rm gas}$ and $V/\sigma$ are correlated at fixed mass: galaxies with lower-than-average gas fractions have higher-than-average $\kappa$ and $\Delta W$, indicating enhanced dispersion support. We do not find such a correlation for the observed galaxies.
\item \textit{Baryonic Tully-Fisher relation}: At all masses, the simulated galaxies have 20\% line-widths in good agreement with the observed baryonic Tully-Fisher relation, irrespective of whether they are supported by rotation or dispersion (Figure~\ref{fig:W20_W50}). Global line widths are primarily probes of galaxies' virial masses and are agnostic to the mode of kinematic support.
\end{enumerate}
At the lowest mass scales probed by our simulations ($M_{\rm star} \simeq 10^{6-7.5} M_{\odot}$), the observed sample does not contain enough observed galaxies for a robust comparison between the simulations and observations. Like the simulated galaxies, the lowest-mass observed galaxies preferentially have single-peaked profiles with high $\kappa$ and $\Delta W$, but this could be explained by either increased dispersion support or linearly rising circular velocity curves (see Appendix~\ref{sec:toy_model}).
\subsection{Possible selection effects}
\label{sec:biases}
We note that although no explicit morphological selection criteria were used in constructing our observed sample, some selection effects could be present. In particular, our observed galaxies are all detected in both HI and optical light, which could select against (a) low surface brightness or (b) HI-poor galaxies. Many of our low-mass galaxies do have low surface brightness \citep{Chan_2017}. We do not, however, expect a selection effect against low surface brightness galaxies in our observed sample to significantly bias our observationally inferred kinematics, as observed low surface brightness galaxies typically have ordinary levels of rotational support in their gas kinematics \citep{deBlok_2001, Oh_2015}.
Because most of the dispersion-supported galaxies in our simulation sample also have somewhat lower gas fractions than typical observed galaxies with similar mass, one might expect selection effects against HI-poor galaxies to introduce biases against dispersion-supported systems in the observed sample. However, the lack of any correlation between gas fraction and kinematics in the observed sample suggests that the excess dispersion found in the simulated galaxies is \textit{not} primarily due to poor observational completeness at low $f_{\rm gas}$: if it were, observed galaxies with lower-than-average $f_{\rm gas}$ would show enhanced dispersion support.
We also consider the possibility of biases in the sample of simulated galaxies. The primary selection effect in our simulation sample is that we target halos that are isolated at $z=0$, meaning that they have no more massive neighbors within at least $3\times R_{\rm 200m}$. We have verified that the sample spans a normal range of halo spin, concentration and formation time \citepalias{ElBadry_2018}, and previous work has also found that targeting isolated halos does not introduce significant biases in most halo properties \citep{Onorbe_2014}.
To test explicitly whether our selection of isolated halos introduces biases in the galaxy properties studied here, we repeated our analysis with only the 923 galaxies in our observational sample (out of 2002 total) that were found to be isolated by \citetalias{Bradford_2015}. We found this isolated subsample to have similar $f_{\rm gas}$, $\kappa$, and $\Delta W$ to the full observational sample.\footnote{This is true only because the observed galaxies all retain enough cold gas to be detected in HI. Stripped galaxies in dense cluster environments likely do have different gas properties than isolated galaxies, but such targets do no enter our observational sample.} We therefore conclude that simulating halos that are isolated at late times does not introduce significant biases for the galaxy properties studied here.
\subsection{Excess dispersion in low-mass galaxies}
The \textit{resolved} gas properties of the simulated galaxies studied here were previously investigated and compared to observations by \citetalias{ElBadry_2018}. In qualitative agreement with our results, that work found most galaxies with $M_{\rm star}=10^{8-10}M_{\odot}$ to be somewhat less rotationally supported than typical observed galaxies at the same mass scale (e.g., their Figure 15). Due to imperfectly understood selection effects in the observed samples to which they compared, \citetalias{ElBadry_2018} concluded that there were hints of enhanced dispersion support in the simulations at low masses, but they could not verify the disagreement to be robust. Our results also find evidence of excess dispersion support at low masses, but with comparison to a larger, morphologically unbiased observational sample. We emphasize that although our simulated low-mass galaxies are \textit{on average} more dispersion-supported than the observed population, they still fall in regions of parameter space populated with real galaxies; i.e., the simulations do not produce galaxies in observationally ``forbidden'' regions of parameter space.
The mass scale at which we find the most enhanced gas dispersion in the simulated galaxies compared to observations is precisely the scale at which feedback-driven outflows have been shown to most strongly alter galaxies' gravitational potentials \citep{DiCintio_2014, Chan_2015, ElBadry_2016, Bullock_2017}. In the FIRE simulations, dynamical heating through a feedback-driven, time-variable potential at this mass scale has been shown to produce dark matter cores \citep{Chan_2015}, low central stellar surface densities \citep{Chan_2017}, and dispersion-supported stellar dynamics with primarily radial orbits \citep{ElBadry_2017}.
The same large-scale outflows that lead to core formation (a) continually inject energy into low-mass galaxies' ISM, making it difficult for gas to settle into a rotationally supported disk \citepalias{ElBadry_2018} and (b) produce very bursty star formation histories at low masses. This suggests that the excess gas dispersion we find at this mass scale is likely connected to tensions between the simulations and observations identified by previous work. \citet{Sparre_2017} showed that FIRE galaxies with $M_{\rm star}\lesssim 10^{9.5} M_{\odot}$ likely have star formation histories that at $z=0$ are on average burstier than those of typical low-mass field galaxies (though observations also imply nontrivial SFH burstiness at this mass scale; see \citealt{Weisz_2012}), and \citet{Chan_2017} showed that galaxies from the FIRE simulations with $M_{\rm star}=10^{7-9} M_{\odot}$ have, on average, lower surface brightnesses than typical observed galaxies at the same stellar mass.
One might therefore expect that low-mass galaxies with rotation-supported gas kinematics would have calmer star formation histories and smaller cores than similar-mass galaxies with dispersion-supported gas; we investigate this further in Appendix~\ref{sec:sfhs}. We find that the lowest-mass galaxy in our simulation sample with a clear rotation-supported disk has a star formation history (SFH) that is substantially less bursty at late times than the SFHs of the dispersion-supported simulated galaxies at the same mass scale. It nevertheless has a dark matter core at $z=0$, likely due to burstier star formation at early times, but the core is smaller than in dispersion-supported systems at the same mass scale.
The apparent excess of dispersion support in the low-mass simulated galaxies can serve as a useful constraint on feedback models and may motivate modifications to the FIRE model or the incorporation of additional physical processes in the future. We first consider the possibility that the FIRE model produces feedback that is in some sense ``too strong'' at low masses. Because feedback energetics are taken directly from stellar evolution calculations, there is little room to arbitrarily reduce or increase the overall feedback energy budget.\footnote{In \citet{Hopkins_2017} and Hopkins et al. (in prep) we have nevertheless carried out tests of the effects of modified feedback recipes on the burstiness of star formation in dwarfs. While reducing the feedback energy budget can make galaxies somewhat less bursty and morphologically diskier at fixed halo mass, doing so increases galaxy stellar masses, such that at fixed stellar mass, burstiness is not significantly changed. Significantly changing the feedback energy budget also produces a $M_{\rm star} - M_{\rm halo}$ relation in worse agreement with constraints from abundance matching than the default FIRE model.} It is possible, however, that changing details in the implementation of physical processes that are currently modeled approximately (e.g., adding on-the-fly radiative transport or more explicit treatment of HII regions) could change the burstiness of star formation. For example, if feedback can create hot ``chimneys'' surrounding star forming regions such that localized outflows can vent energy out of the disk \citep[e.g.][]{Ceverino_2009, Hopkins_2012}, the effects of feedback on the bulk of a galaxy's gas may be less catastrophic than when energy and momentum are primarily deposited in cold gas. Alternatively, if radiative feedback can suppress star formation locally on timescales shorter than the free-fall time in bound gas clouds \citep[e.g.][]{FG_2018}, star formation can become less spatially and temporally clustered, leading to weaker outflows and less dispersion-supported gas.
Simply turning off some feedback sources (e.g. radiative feedback or stellar winds) does not seem to reduce SFH burstiness, as it allows more stars to form before collapse is regulated by supernovae \citep{Hopkins_2017}. As long as star formation is restricted to self-gravitating gas clouds, varying the local star formation efficiency or the critical density above which star formation is allowed to occur also does not significantly change galaxy SFHs in the default FIRE model, since star formation proceeds until feedback produces enough energy to regulate it \citep{Hopkins_2017}. It may thus if fact be easier to facilitate disk formation by incorporating additional feedback processes, albeit processes that act less violently than supernovae, than by reducing the types of feedback sources or their energy budget. For example, \citet{Chen_2016} found that including cosmic ray diffusion in galaxy formation simulations caused their low-mass galaxies to transition from being dispersion-supported, with bursty SFHs, to rotation-supported, with smoother SFHs. Further investigation of the effects of changes in the details of feedback implementations on SFH burstiness and galaxy morphology represents a promising avenue for future work. We note, for example, that feedback resulting in weaker large-scale outflows in low-mass galaxies could also alleviate the difficulty identified by \citet{AnglesAlcazar_2017} of growing massive black holes in the FIRE simulations at early times.
The fact that our simulations do produce some low-mass galaxies with rotation-supported disks also suggests that at fixed mass, morphology is driven by a galaxy's formation history and/or environment. This warrants further investigation of the assembly histories of our low-mass galaxies and the primary factors that drive them to become rotation- or dispersion-supported, as well as simulation of low-mass galaxies in a wider range of environments.
\section*{Acknowledgements}
We thank the referee, Danail Obreschkow, for a constructive report and Alyson Brooks and Pascal Elahi for helpful discussions.
KE acknowledges support from a Berkeley graduate fellowship, a Hellman award for graduate study, and an NSF Graduate Research Fellowship.
EQ and KE are supported by a Simons Investigator Award from the Simons Foundation and by NSF grant AST-1715070.
MBK acknowledges support from NSF grant AST-1517226 and from NASA grants NNX17AG29G and HST-AR-13888, HST-AR-13896, HST-AR-14282, HST-AR-14554, HST-GO-12914, and HST-GO-14191 from STScI.
DRW is supported by a fellowship from the Alfred P. Sloan Foundation.
AW was supported by NASA through grants HST-GO-14734 and HST-AR-15057 from STScI.
Support for PFH was provided by an Alfred P. Sloan Research Fellowship, NASA ATP Grant NNX14AH35G, and NSF Collaborative Research Grant \#1411920 and CAREER grant \#1455342.
DK and TKC were supported by NSF grants AST-1412153 and AST-1715101 and the Cottrell Scholar Award from the Research Corporation for Science Advancement.
CAFG was supported by NSF through grants AST-1412836, AST-1517491, AST-1715216, and CAREER award AST-1652522, and by NASA through grant NNX15AB22G.
We ran numerical calculations on the Caltech compute cluster ``Wheeler,'' allocations TG-AST130039 \& TG-AST150080 granted by the Extreme Science and Engineering Discovery Environment (XSEDE) supported by the NSF, and the NASA HEC Program through the NAS Division at Ames Research Center and the NCCS at Goddard Space Flight Center.
The analysis in this paper relied on the python packages \texttt{NumPy} \citep{vanderwalt_2011}, \texttt{Matplotlib} \citep{Hunter_2007}, and \texttt{AstroPy} \citep{Astropy_2013}.
\bibliographystyle{mnras}
|
\section{Introduction}
Tremendous progress in the implementation of quantum simulators~\cite{feynman82,houck12,ma14,omalley16,hensgens17,loredo16,monroe17,lukin17} has led us to an enviable scenario in which predicting the general dynamics of a quantum simulator exceeds available classical computation power.
How then are we to measure the performance of such devices?
In benchmarking, one technique is so-called homomorphic encryption in which the desired computation is hidden from, e.g., the server upon which it is implemented and then decrypted post facto.
In the quantum domain, such blind computing has been suggested using the circuit model and the measurement-based approaches~\cite{Childs05,Broadbent09}.
These protocols rely upon a high bandwidth quantum communication channel between the server and the client.
Further developments along these lines have improved security, blindness, and provided a connection to quantum interactive proofs~\cite{Dunjko12,Morimae12,Dunjko14,Morimae15,Aharonov10,Aharonov17,Fitzsimons17,Gheorghiu17}.
Inspired by the central spin Hamiltonian~\cite{Gaudin76,Taylor03,Yuzbashyan05}, here we build from the ground up a blind computation scheme using a quantum simulator. While we require two-way quantum communication, we show that passing a single qubit back and forth suffices to ensure security.
This approach should be accessible not only in natural central spin implementation, such as quantum dots~\cite{Merkulov2002,Khaetskii2002,deSousa2003-dots,Schliemann03,Erlingsson04,Chekhovich2013}, NV centers~\cite{Wrachtrup06,Childress06,Balasubramanian09,Shin13,Wang13,Hall14} and NMR molecules~\cite{Pastawski95,Cory98,Laflamme2002}, but also in simulators that can implement Heisenberg-like interactions such as ion traps~\cite{Porras04,Lanyon11,Arrazola16,Bermudez17} and circuit QED systems~\cite{Cho08,houck12,Heras14,Salathe15,Lamata17}.
The paper is structured as follows.
We first review the central spin Hamiltonian where a central spin is coupled to a number of bath spins.
We show how different states of the central spin effectively leads to different dynamics of the bath spins.
With this observation, we present a delegated simulation scheme in which the dynamics of the bath spins are controlled by a single central spin communicated back and forth between the client and the server.
However, the simulation is blind only if the server is trusted not to measure the central spin.
Such a vulnerability is later removed in an improved scheme with ``honeypots" added to detect measurement attempts during the computation.
Finally, we show that our blind simulation scheme is capable of simulating a universal quantum gate set, and therefore allows the client to perform universal, blind computation on the server.
\section{Central spin model}
We consider a system in which the central spin Hamiltonian is either natural (e.g. in quantum dots, NV centers) or can be simulated (e.g. using ion traps and circuit QED).
Such a Hamiltonian describes interaction between a spin-$\frac12$ central spin $\vec S_0$ with $n$ spin-$\frac{1}{2}$ bath spins $\vec S_j$ for $j=1,\dots,n$,
\begin{align}
H_c = \sum_{j=1}^n \gamma_j \vec S_0\cdot \vec S_j - h_0 S_0^z - \sum_{j=1}^n h_j S_j^z, \label{EQ_HC}
\end{align}
where $\vec S_j=\left(S_j^x,S_j^y,S_j^z\right)$ is the spin operator vector of the $j$th spin and $\gamma_j$ denotes the interaction strength between $\vec S_0$ and $\vec S_j$.
The last two terms are the result of the interaction between the spins and an external magnetic field.
Without loss of generality, we assume the field to be along the $z$ axis.
For our blind computation protocol, we also assume that $h_0$ and $h_j,\gamma_j$ for $j=1,\dots,n$ are tunable parameters of the system.
In the following discussion, we further assume that the magnetic field on the central spin is much larger than the interaction between the spins, i.e. $h_0\gg n\gamma = \eta$ with $\gamma = \frac{1}{n}\sum_j\abs{\gamma_j}$ being the average interaction strength.
Without loss of generality and for simplicity, we further set $h_0 = 1$ in our calculation.
In this $\eta\ll 1$ limit, the Hilbert space is well separated into two subspaces, each corresponds to an eigenstate of $S_0^z$ of the central spin.
Although there is no interaction term between the bath spins in Eq.~\eqref{EQ_HC}, they can still interact with each other via an interaction mediated by the central spin~\cite{YaoLS06,LiuYS07}.
Using the Schrieffer-Wolff approximation~\cite{Bravyi11}, we find the Hamiltonians describing such effective interactions among the bath spins in the two subspaces (Supplemental Material).
For example, in the subspace that corresponds to the central spin being in $\ket{0}$, i.e. the ``up" state, the effective Hamiltonian is
\begin{align}
H_{\uparrow} = &-\frac{1}{2} \sum_{j<k}\gamma_j\gamma_k\left(S_j^xS_k^x+S_j^yS_k^y\right)\nonumber\\
&-\sum_{j=1}^n\left(h_j-\frac{\gamma_j}{2}-\frac{\gamma_j^2}{4}\right)S_j^z+\O{\eta^3}, \label{EQ_Hup_hb}
\end{align}
where the first sum is over all $1\leq j<k\leq n$.
Note that the first sum is also an all-to-all interaction using which we shall engineer two-qubit gates between any two spins.
Similarly, in the ``down" subspace, i.e. the central spin is in $\ket{1}$, the effective Hamiltonian between the bath spins is
\begin{align}
H_{\downarrow} =& \frac{1}{2} \sum_{j<k}\gamma_j\gamma_k\left(S_j^xS_k^x+S_j^yS_k^y\right)\nonumber\\
&-\sum_{j=1}^n\left(h_j+\frac{\gamma_j}{2}-\frac{\gamma_j^2}{4}\right)S_j^z+\O{\eta^3}. \label{EQ_Hdowno_hb}
\end{align}
In our discussion below, we shall choose $h_j = -\gamma^2_j/4$ in the system.
With this choice, the two effective Hamiltonians are different by only a sign, i.e.
\begin{align}
H_\uparrow \approx -H_\downarrow \approx &\sum_{j<k}\frac{\gamma_j\gamma_k}{2}\left(S_j^xS_k^x+S_j^yS_k^y\right)-\sum_{j=1}^n\frac{\gamma_j}{2} S_j^z. \label{EQ_Hup}
\end{align}
This is a key ingredient in achieving blindness later in our protocols.
Knowing the effective Hamiltonians, we may approximate the time evolution under $H_c$ as
\begin{align}
e^{-iH_ct} &\approx \ket{0}\bra{0}\otimes e^{-iH_{\uparrow}t} + \ket{1}\bra{1}\otimes e^{-iH_{\downarrow}t}\\
&= e^{-i Z_0\otimes H_\uparrow t},\label{EQ_UApprox}
\end{align}
with $\ket{\alpha}\bra{\alpha}$ $(\alpha=0,1)$ being the projection operator onto the state $\ket{\alpha}$ of the central spin and $Z_0$ is the Pauli-$Z$ acting on the central spin.
Our protocol leverages the emergence of a three-spin interaction that arises naturally in the central spin settings.
This is in contrast to many quantum gadgets simulation approaches~\cite{Kempe06,Jordan08}, where 3-local terms require substantial additional resources for implementations.
This interaction allows the central spin to control the dynamics of the bath spins.
We now show how such a feature allows Alice and Bob to perform blind quantum simulation and, in particular, universal blind quantum computation.
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{delegate-comp-demo.pdf}
\caption{
A demonstration of our blind simulation and blind computation schemes.
The server Bob can simulate dynamics of the central spin Hamiltonian in Eq.~\eqref{EQ_HC} where a central spin (orange) interacts with a number of bath spins (blue).
The bath spins can only interact with each other via the central spin.
The effective dynamics of the bath spins is therefore encrypted in the state of the central spin, which is communicated back and forth between the server and the client.
}
\label{FIG_Delegate_demo}
\end{figure}
\section{Blind quantum simulation}
The three-body interaction in Eq.~\eqref{EQ_UApprox} allows a client (Alice) to perform quantum simulation on a server (Bob) such that details of the simulation are hidden from Bob.
Here we present a protocol for simulating dynamics of $n$ bath spins on the server using a single spin communicated back and forth with the client.
In our scenario, Alice has a series of $m$ Hamiltonians $H^{(k)}_{\uparrow}$ $(k=1,\dots,m)$, each of which is of the form \eqref{EQ_Hup} and is characterized by $n$ real parameters, namely the $\gamma_{j}^{(k)}$ for $j=1,\dots,n$.
Note that for each $k$, these parameters also characterize a corresponding central spin Hamiltonian $H^{(k)}_c$ as in Eq.~\eqref{EQ_HC}.
In addition to the Hamiltonians, Alice also chooses a set of $m$ constants $\{t_1,\dots,t_m\}$ and a binary string $\alpha$ of length $m$, i.e. $\alpha\in\{0,1\}^m$.
The former is to play the role of desired evolution times under the Hamiltonians of the same index and the latter is a secret key that encrypts the simulation.
The blind simulation protocol consists of $m$ iterations. In the $k$th iteration, Alice will communicate a central spin to Bob along with the classical parameters $\gamma_j^{(k)}$.
Bob then simulates evolution of the $n+1$ spins under the $H_c^{(k)}$ specified by $\gamma_j^{(k)}$ for a time $t_k$.
Although Bob knows the evolution of the whole system, the effective time evolution of the bath spins under $H_{\uparrow}^{(k)}$ is either forward or backward in time and is encoded by the secret key $\alpha_k$ known only to Alice.
Such a protocol for blind quantum simulation can be summarized in Protocol~\ref{PRO_SIM}.
\SetNlSkip{0.8em}
\SetInd{0.5em}{1em}
\begin{algorithm}[t]
\caption{Blind quantum simulation} \label{PRO_SIM}
\SetInd{1em}{1em}
\setstretch{0}
\DontPrintSemicolon
\KwIn{the $n$ bath spins in a state $\ket{\psi^{(0)}}$.}
\For{$k\in\{1,\dots,m\}$}{
\jus{1} Alice sends Bob classical parameters $t_k$ and $\gamma_j^{(k)}$ for all $j = 1,\dots,n$.\;
\jus{2} Alice sets the central spin to the state $\ket{\alpha_k}$ and sends to Bob.\;
\jus{3} Bob simulates evolution of the $n+1$ spins under $H^{(k)}_c$ for time $t_k$.\;
\jus{4} Bob sends the central spin back to Alice.\;
}
\KwRet the $n$ bath spins in the final state $\ket{\psi^{(m)}}$
\end{algorithm}
Let us examine the final state of the $n$ bath spins at the end of this protocol.
Denote by $\ket{\psi^{(k)}}$ the state of the bath spins after the $k$th iteration.
By the end of line 3 in the $k$th iteration, Bob has in his possession $n+1$ spins, including the central spin sent by Alice, in the state $\ket{\alpha_{k}}\otimes\ket{\psi^{(k-1)}}$.
Bob then simulates the evolution under $H^{(k)}_c$ for the time $t_{k}$ and brings the $n+1$ spins to
\begin{align}
&\exp{\left(-iH^{(k)}_c t_{k}\right)}\ket{\alpha_{k}}\otimes \ket{\psi^{(k-1)}}\nonumber\\
\approx& \ket{\alpha_{k}}\otimes \exp{\left(-i(-1)^{\alpha_{k}}H^{(k)}_\uparrow t_{k}\right)}\ket{\psi^{(0)}},
\end{align}
where we have applied Eq.~\eqref{EQ_UApprox} to approximate the time evolution of the $n+1$ spins by an effective time evolution of the bath spins only.
By induction from $k=1$ to $k=m$, the final state of the $n$ spins at the end of the protocol is
\begin{align}
\ket{\psi^{(m)}} &= \prod_{k=1}^m \exp{\left(-i(-1)^{\alpha_{k}}H^{(k)}_\uparrow t_{k}\right)}\ket{\psi^{(0)}} \nonumber\\
&\equiv U_m\ket{\psi^{(0)}}.\label{EQ_Unet}
\end{align}
Here the information about the effective unitary $U_m$ is partially encrypted by the key $\alpha$ known only to Alice.
For a generic set of $H^{(k)}_c$ and $t_k$, there are $2^m$ possibilities of $U_m$ only one of which is actually implemented on the bath spins.
Therefore with a long enough key $\alpha$, Alice can be confident that Bob has almost no information about the unitary performed.
\begin{figure}[t]
\includegraphics[width=0.5\textwidth]{hide-circuit.pdf}
\caption{
The security of Protocol~\ref{PRO_SIM_W_TRAP} and Protocol~\ref{PRO_COMP} is guaranteed by embedding the desired circuit (blue blocks) into a much longer circuit.
The longer circuit consists of mostly honeypots (orange blocks) that do not contribute to the overall computation.
Instead, they serve to detect measurement attempts from the server.
The probability for the server to measure without being detected decays exponentially with the number of honeypots.
}
\label{FIG_hide_circuit}
\end{figure}
Note, however, that this simple protocol does not protect Alice's secret from a malicious Bob who tries to determine the key $\alpha$ by measuring the central spin every time Alice sends it over.
Indeed, if the protocol is followed, Bob knows the central spin can only be in one of the two orthogonal states and hence can be deterministically identified by an appropriate projective measurement.
Although Alice may not have the power to stop Bob from measuring, she can set up honeypots in the middle of the simulation to trap and abort the simulation as soon as such a malicious attempt is detected.
Using the same idea as in the BB84 quantum key distribution scheme \cite{BB84}, the set of available states of the central spin can be extended to $\left\{\ket{0},\ket{1},\ket{+},\ket{-}\right\}$ where $\ket{\pm}=\ket{0}\pm\ket{1}$ up to a normalization constant.
Since the four states form a nonorthogonal set, it is impossible for Bob to determine with certainty which of the states is prepared by Alice.
In particular, attempts to measure the central spin will collapse its state and therefore can be detected by Alice when the central spin is returned to her at the end of each iteration.
We present below Protocol~\ref{PRO_SIM_W_TRAP} with such honeypots.
\begin{algorithm}[t]
\caption{Secured blind simulation} \label{PRO_SIM_W_TRAP}
\SetInd{1em}{1em}
\setstretch{0}
\DontPrintSemicolon
\KwIn{the $n$ bath spins in a state $\ket{\psi^{(0)}}$.}
\For{$k\in\{1,\dots,m\}$}{
\jus{1} Alice sends Bob the classical parameters $t_k$ and $\gamma_j^{(k)}$.\;
\jus{2} Alice sets the central spin to $\ket{\alpha_k}$ and sends to Bob.\;
\jus{3} Bob simulates evolution of the $n+1$ spins under $H^{(k)}_c$ for time $\frac{t_k}{2}$.\;
\jus{4} Bob returns the central spin back to Alice.\;
\jus{5} Alice applies either a $\pi$ pulse, a $\pi/2$ pulse or identity to rotate the central spin to $\ket{\beta_k}$ and sends to Bob.\;
\jus{6} Bob simulates evolution of the $n+1$ spins under $H^{(k)}_c$ for time $\frac{t_k}{2}$.\;
\jus{7} Bob sends the central spin back to Alice.\;
\jus{8} Alice aborts if the returned central spin is not $\ket{\beta_k}$.\;
}
\KwRet the bath spins in the final state $\ket{\psi^{(m)}}$.
\end{algorithm}
In addition to the key $\alpha\in\{0,1,\pm 1\}^m$, Protocol~\ref{PRO_SIM_W_TRAP} requires Alice to choose another key $\beta\in\{0,1,\pm\}^m$ of the same length $m$, with one restriction that $\beta_k = \pm$ if $\alpha_k=\mp$ for all $k$.
With this restriction, it is straightforward to verify that whenever $\alpha_k=\pm$, the net unitary applied on the bath spins in the $k$th iteration is equivalent to identity.
Such iterations therefore only play the role of flagging malicious measurement attempts and do not contribute to the overall simulation.
Note that when $\beta=\alpha\in\{0,1\}^m$, Protocol~\ref{PRO_SIM_W_TRAP} reduces to Protocol~\ref{PRO_SIM}.
\section{Universal blind computation}
So far we have shown that Alice can request a general simulation $U_m$ given by Eq.~\eqref{EQ_Unet} on $n$ bath spins without revealing her data.
But how general is $U_m$?
In other words, what type of quantum computation Alice can achieve by simulating $U_m$?
We now show that by choosing the right parameters $\gamma_j^{(k)},t_k$ in each iteration, Alice can simulate any gate in a universal gate set and therefore is able to perform a blind simulation of an arbitrary quantum circuit.
Indeed, by turning off $\gamma_j$ for all except $j = 1$, the time evolution unitary is a local rotation of the first spin about the $z$ axis,
\begin{align}
e^{-i \alpha Z_1 t} =e^{-i \frac{\hbar\gamma}{2}\gamma_1 t} \left(
\begin{array}{cc}
1 & 0 \\
0 & e^{i\hbar\gamma_1 t} \\
\end{array}
\right),
\end{align}
where $Z_1$ is the Pauli-$Z$ matrix on the first qubit.
Using this Hamiltonian, we can obtain phase-shift gates such as the $T$ gate by choosing the right evolution time $t$.
Similarly, by changing the magnetic field in Eq.~\eqref{EQ_HC} to the $x$ axis, Alice can simulate the Hadamard gate $H$.
To form a universal gate set, we still need a two-qubit gate such as the following $U_{XY}$ gate~\cite{ImamogluABDLSS99}:
\begin{align}
U_{XY}\equiv \left(
\begin{array}{cccc}
1 & 0 & 0 & 0 \\
0 & \frac{1}{\sqrt{2}} & \frac{i}{\sqrt{2}} & 0 \\
0 & \frac{i}{\sqrt{2}} & \frac{1}{\sqrt{2}} & 0 \\
0 & 0 & 0 & 1 \\
\end{array}
\right).
\end{align}
To engineer a $U_{XY}$ gate, for example, between the first and the second qubits, we turn all $\gamma_j$ off except for $\gamma_1=\gamma_2=\gamma$.
The effective time evolution of the bath spins is
\begin{align}
\exp\left\{-i\left(\frac{\gamma^2}{8}\left(X_1X_2+Y_1Y_2\right)+\frac{\gamma}{2}(Z_1+Z_2)\right)t\right\},\label{EQ_UHc}
\end{align}
where $X,Y$ are respective Pauli gates.
With $\gamma = \frac{7}{2c}$ for some large integer $c$ and $t = \frac{7\pi}{\gamma^2}$, the above unitary reduces to the $U_{XY}$ gate.
Since the $U_{XY}$ gate and single-qubit gates form a universal gate set~\cite{ImamogluABDLSS99}, Alice can effectively perform universal quantum computation on the bath spins.
In the following discussion, we refer to this universal gate set as $\mathcal{U}$.
Protocol~\ref{PRO_SIM_W_TRAP} can be further adapted to this situation to guarantee the security of the computation by using honeypots to detect measurement attacks.
A desired circuit $U_{m_0}$ of length $m_0$ on $n$ qubits can be embedded into a much larger circuit $U_m$ with $m\gg m_0$ that consists mostly of honeypots (Fig.~\ref{FIG_hide_circuit}).
These honeypots perform trivial operations on the qubits and serve only as detectors of malicious behaviors.
Mean while in each non-honeypot iteration of the protocol, i.e. $\alpha_k\in\{0,1\}$, Alice can choose a Hamiltonian $H^{(k)}_\uparrow$ and a time $t_k$ such that $G_k = \exp\left\{-iH^{(k)}_\uparrow t_k\right\}$ is a gate in the universal gate set $\mathcal{U}$.
Note that such a gate (or its inverse) is only implemented on the bath spins at the end of the iteration if $\beta_k \oplus \alpha_k = 0$.
On the other hand, if $\beta_k \oplus \alpha_k =1$, the two evolutions in the $k$th iteration cancel each other out and therefore only a trivial gate is implemented.
Denote by $\omega(\alpha_k,\beta_k)$ a function of the characters $\alpha_k,\beta_k$ such that $\omega(\alpha_k,\beta_k)=1$ if $\alpha_k\oplus\beta_k=0$ and $\omega(\alpha_k,\beta_k)=0$ otherwise.
The gate sequence generated by the protocol can then be summarized by the following equation:
\begin{align}
U_m = \prod_{k=1}^m G_{k}^{\omega(\alpha_k,\beta_k)}.\label{EQ_Ugate}
\end{align}
Therefore the keys $\alpha,\beta$ effectively encrypt the circuit Alice implements and make the quantum computation blind.
Our protocol for universal blind quantum computation is be summarized in Protocol~\ref{PRO_COMP} below.
\begin{algorithm}[h!]
\caption{Universal blind computation} \label{PRO_COMP}
\SetInd{1em}{1em}
\setstretch{0}
\DontPrintSemicolon
\vspace{0.1in}
\jus{0} Alice has a circuit $U_{m_0}$ of length $m_0$ to be performed on $n$ qubits.\;
\jus{1} Alice embeds $U_{m_0}$ into a much larger circuit $U_m$ by choosing two keys $\alpha,\beta\in\{0,1,\pm\}^m$, each of length $m\gg m_0$ such that $U_m$ in Eq.~\eqref{EQ_Ugate} reduces to $U_{m_0}$.\;
\jus{2} Alice and Bob perform Protocol~\ref{PRO_SIM_W_TRAP} to implement $U_m$ on $\ket{\psi}$.\;
\end{algorithm}
\section{Quantum verification}
Blindness
allows the client Alice to not only hide the computation from the server Bob but also to verify if Bob performs the correct computation.
Indeed, Alice can verify Bob by simply requesting quantum circuits that have outcomes that can be classically \emph{verified}.
By the definition of blind computation used in this work, Bob has no information about what circuits are being implemented, and the only way he can return the correct output to Alice is to perform the exact simulation sequence as instructed.
For example, Alice can ask Bob to initialize the $n$ bath spins in a product state such that some of the spins are ``up" and some are ``down", and use Protocol~\ref{PRO_COMP} to simulate a sequence of SWAP gates between the bath spins known only to Alice.
The final state is a permutation of the initial state and is known only to Alice.
Bob has to find the final state and the only way he can pass with certainty is to correctly perform the computation.
Since both the initial state and the final state are fully separable, the permutation circuit is essentially classical.
Stabilizer circuits~\cite{Bennett96,Gottesman96}, on the other hand, can perform nontrivial quantum operations, such as quantum teleportation~\cite{Bennett93} and preparation of highly entangled states.
They consist of only Clifford gates and can be simulated efficiently on a classical computer~\cite{Aaronson04}.
Therefore by requesting simulation of an arbitrary stabilizer circuits, Alice can also efficiently verify quantumness of the server.
Alice can also take a step further to verify even the quantum computing power of the server by requesting a quantum circuit that is known to solve a problem faster than classical algorithms.
For example, in the Simon's problem \cite{Simon97}, a function $f:\{0,1\}^n\rightarrow\{0,1\}^n$ is promised to satisfy that $f(x)=f(y)$ if and only if $x=y$ or $x\oplus y = s$ for all $x,y\in\{0,1\}^n$ and a fixed string $s\in\{0,1\}^n$.
To find $s$, classical algorithms require at least $\Omega(2^{n/2})$ queries to the function $f$ while the quantum Simon's algorithm can solve the problem using only $\O{n}$ queries.
Using Protocol~\ref{PRO_COMP}, Alice can simulate a quantum circuit corresponding to a secret string $s$.
She then asks Bob to measure the output and announce the measured string.
If Bob is able to answer correctly what $s$ is for large enough $n$, Alice can be confident that the server Bob has access to at least a BQP machine.
\section{Outlook}
Here we have shown how to implement an arbitrary circuit controlled by a single spin.
This enables us to define several blind computing protocols that can be a powerful test of computing power in quantum simulators.
However, we have not yet developed natural observables whose, e.g., distribution function is distinctly different given classical versus quantum computational power.
We consider this an intriguing direction for future research.
\begin{acknowledgments}
We thank S. -H Hung, B. Lackey, R. Matthew, and Y. Wang for helpful discussions. This research was supported in part by the NSF funded Physics Frontier Center at the Joint Quantum Institute and the Army Research Laboratory's CDQI.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction and notations}
Let $G$ be a connected graph. The \emph{open neighbourhood} of a vertex $u$ is $N(u)$ the set of vertices adjacent to $u$. The \emph{closed neighbourhood} of $u$ is $N[u]=N(u)\cup\{u\}$. The \emph{distance} between two vertices noted $d_G(x,y)$, or $d(x,y)$ when the graph is unambiguous, is the length of the shortest path between $x$ and $y$. We have thus $N[u]=\{v \in V(G);d(u,v)\leq1\}$.
A \emph{dominating set} $D$ of $G$ is a set of vertices such that every vertex of $G$ belongs to the closed neighbourhood of at least one vertex of $D$.
In \cite{Biggs}, Biggs initiated the study of perfect codes in graphs a generalization of classical 1-error perfect correcting codes. A \emph{code} $C$ in $G$ is a set of vertices $C$ such that for all pair of distinct vertices $c,c'$ of $C$ we have $N[c]\cap N[c']=\emptyset$ or equivalently such that $d_G(c,c')\geq3$.
A \emph{perfect code} of a graph $G$ is both a dominating set and a code. It is thus a set of vertices $C$ such that every vertex of $G$ belongs to the closed neighbourhood of exactly one vertex of $C$. A perfect code is some time called an efficient dominating set. The existence or non-existence of perfect codes have been considered for many graphs. See the introduction of \cite{aabfk-2016} for some references.
The vertex set of the \emph{$n$-cube} $Q_n$ is the set $\mathbb{B}_n$ of binary strings of length $n$, two vertices being adjacent if they differ in precisely one position. Classical 1-error correcting codes and perfect codes are codes and perfect codes in the graph $Q_n$.
The concatenation of strings $\bm{x}$ and $\bm{y}$ is noted $\bm{x}||\bm{y}$ or just $\bm{x}\bm{y}$ when there is no ambiguity. A string $\bm{f}$ is a \emph{substring} of a string $\bm{s}$ if there exist strings $\bm{x}$ and $\bm{y}$, may be empty, such that $\bm{s}=\bm{x}\bm{f}\bm{y}$.
A {\em Fibonacci string} of length $n$ is a binary string $\bm{b}=b_1\ldots b_n$ with $b_i\cdot b_{i+1}=0$ for $1\leq i<n$. In other words a Fibonacci string is a binary string without $11$ as substring.
The {\em Fibonacci cube} $\Gamma_n$ ($n\geq 1$) is the subgraph of $Q_n$ induced by the Fibonacci strings of length $n$.
Fibonacci cubes were introduced as a model for interconnection networks~\cite{hsu-93} and received a lot of attention afterwards.
These graphs also found an application in theoretical chemistry. See the survey \cite{survey} for more results and applications about Fibonacci cubes.
The sets $\{00\}$ and $\{010,101\}$ are perfect codes in respectively $\Gamma_2$ and $\Gamma_3$. In a recent paper \cite{aabfk-2016} Ashrafi and his co-authors proved the non-existence of perfect codes in $\Gamma_n$ for $n\geq 4$. As an open problem the authors suggest to consider the existence of perfect codes in generalization of Fibonacci cubes.
The most complete generalization proposed in \cite{ikr} is, for a given string $\bm{f}$, to consider $\Gamma_n(\bm{f})$ the subgraph of $Q_n$ induced by strings that do not contain $\bm{f}$ as substring. Since Fibonacci cubes are $\Gamma_n(11)$ the most immediate generalization \cite{hsuliu,Zag} is to consider $\Gamma_n({1^s})$ for a given integer $s$. We will prove the existence of perfect codes in $\Gamma_n({1^s})$ for an infinite family of parameters $(n,s)$.
It will be convenient to consider the binary strings of length $n$ as vectors of $\mathbb{F}^n$ the vector space of dimension $n$ over the field $F=\mathbb{Z}_2$ thus to associate to a string $x_1 x_2 \dots x_n$ the vector $\theta(x_1 x_2 \dots x_n)=(x_1,x_2,\ldots,x_n)$.
The \emph{Hamming distance} between two vectors $\bm{x},\bm{y} \in \mathbb{F}^n$, $d(\bm{x},\bm{y})$ is the number of coordinates in which they differ.
The $\emph{parity function}$ is the function from $\mathbb{F}^n$ to $\mathbb{Z}_2$ defined by $\pi(\bm{x})=\pi(x_1,x_2,\ldots,x_n)= x_1+x_2+\ldots+x_n$.
By the correspondence $\theta$ we can define the sum $\bm{x}+\bm{y}$, the Hamming distance $d(\bm{x},\bm{y})$ and the parity $\pi(\bm{x})$ of strings in $\mathbb{B}_n$. Note that Hamming distance is the usual graph distance in $Q_n$.
The complement of a string $\bm{x}\in \mathbb{B}_n$ is the string $\overline{\bm{x}}=\bm{x}+1^n$.
We will first recall some basic results about perfect codes in $Q_n$.
Since $Q_n$ is a regular graph of degree $n$ the existence of a perfect code of cardinality $|C|$ implies $|C|(n+1)=2^n$ thus a necessary condition of existence is that $n+1$ is a power of 2 thus that $n=2^p-1$ for some integer $p$.
For any integer $p$ Hamming \cite{Ha1950} constructed, a linear subspace of $\mathbb{F}^{2^p-1}$ which is a perfect code. It is easy to prove that all linear perfect codes are Hamming codes.\\
In 1961 Vasilev \cite{{Va1962}}, and later many authors, see \cite{Co1,Sol2008} for a survey, constructed perfect codes which are not linear codes.
Let us recall Vasilev's construction of perfect codes.
\begin{theorem}\label{thvas}\cite{Va1962}
Let $C_r$ be a perfect code of $Q_r$. Let $f$ be a function from $C_r$ to $\mathbb{Z}_2$ and $\pi$ be the $\emph{parity function}$.
Then the set $C_{2r+1}= \left\{\bm{x}||\pi(\bm{x})+f(\bm{c})||\bm{x}+\bm{c};\bm{x}\in \mathbb{B}_r,\bm{c}\in C_r\right\}$ is a perfect code of $Q_{2r+1}$
\end{theorem}
We recall also the proof of Theorem \ref{thvas} in such a way our article will be self contained.
\begin{proof}
Fist notice that $|C_{2r+1}|=2^r|C_r|=2^r\frac{2^{r}}{r+1}=\frac{2^{2r+1}}{2r+2}$. Thus if is sufficient to prove that the distance between to different elements of $C_{2r+1}$ is at least 3.\\
Consider $d(\bm{x}||\pi(\bm{x})+f(\bm{c})||\bm{x}+\bm{c}, \bm{x'}||\pi(\bm{x'})+f(\bm{c'})||\bm{x'}+\bm{c'})=d_1+d_2+d_3$ where $d_1=d(\bm{x},\bm{x'})$, $d_2=d(\pi(\bm{x})+f(\bm{c}),\pi(\bm{x'})+f(\bm{c'}))$ and $d_3=d(\bm{x}+\bm{c},\bm{x'}+\bm{c'})$.\\
If $d_1=0$ then $\bm{x}= \bm{x'}$ thus $d_3=d(\bm{c},\bm{c'})\geq 3$.\\
If $d_1=1$ and $\bm{c}= \bm{c'}$ then $d_2=d_3=1$\\
If $d_1=1$ and $\bm{c} \neq \bm{c'}$ then $d_3\geq 2$ otherwise $d(\bm{c},\bm{c'})\leq2$\\
If $d_1=2$ then $d_3\neq 0$ otherwise $d(\bm{c},\bm{c'})=2$\\
Thus $d=d_1+d_2+d_3\geq3$.
\end {proof}
\qed
If $f(\bm{c})=0$ for any $\bm{c}\in C_r$ we obtain the classical inductive construction of Hamming codes with $C_1=\{0\}$ as basis.
In the next section we will use this construction starting from the Hamming code in $Q_r$ as $C_r$ and a function $f$ chosen in such way that the strings of the constructed code $C_{2r+1}$ has not a too big number of consecutive 1's.
\section{Main Result}
\begin{lemma}
Let $m$ be an integer. Let $A_0$ be the set of strings $A_0=\{0^{m+1}\bm{y};\bm{y}\in \mathbb{B}_{m}\}$. For $i\in \{1,\dots,m\}$ let $A_i=\{\bm{z}10^{m+1}\bm{y}; \bm{z}\in \mathbb{B}_{i-1}, \bm{y}\in \mathbb{B}_{m-i}\}$.
Then the sets $A_i$ are disjoint and any string of ${B}_{2m+1}$ containing $0^{m+1}$ as substring belongs to a $A_i$.
\end{lemma}
\begin{proof}
Let $\bm{x}$ be a string of ${B}_{2m+1}$ containing $0^{m+1}$ as substring and $i$ be the minimum integer such that $x_{i+1}x_{i+2}\dots x_{i+m+1}=0^{m+1}$. Then $i=0$, and $\bm{x}$ belongs to $A_0$, or $m \geq i\geq 1$. In this case $x_{i}=1$ thus $\bm{x}\in A_i$.
Assume $\bm{x}\in A_i\cap A_j$ with $m\geq j>i\geq 0$ then $x_{j}=1$ thus $j\geq i+m+2>m$ a contradiction.
\end{proof}
\begin{theorem}\label{thmain}
Let $n=2^p-1$ where $p\geq 2$ and let $s =3.2^{p-2}$. There exists a perfect code $C$ in $Q_n$ such that no elements of $C$ contains $1^s$ as substring.
\end{theorem}
\begin{proof}
Let $m=2^{p-2}-1$ thus $2m+1=2^{p-1}-1$ and $s=3m+3$. Let $C_{2m+1}$ be a perfect code in $Q_{2m+1}$. Let $f$ be the function from $\mathbb{B}_{2m+1}$ to $\mathbb{Z}_2$ defined by
\begin{itemize}
\item $f(0^{m+1}\bm{y})=1$ for $\bm{y}\in \mathbb{B}_{m}$
\item $f(10^{m+1}\bm{y})=0$ for $\bm{y}\in \mathbb{B}_{m-1}$
\item $f(\bm{z}10^{m+1}\bm{y})=\pi(\bm{z})$ for $\bm{z}\in \mathbb{B}_{i-1}$ and $\bm{y}\in \mathbb{B}_{m-i}$ for $i=2$ to $m$.
\item $f=0$ otherwise.
\end{itemize}
Note that from the previous lemma the function is well defined.
Let $C$ be the perfect code obtained from Vasilev's construction from $C_{2m+1}$ and $f$.
Assume there exists a string $\bm{d}$ in $C$ with $1^{3m+3}$ as substring. Therefore $\bm{d}$ is obtained from $\bm{x}=d_1d_2\dots d_{2m+1}$ and $\bm{c}\in C_{2m+1}$. Since $n= 4m+3$ note first that $d_{m+1}d_{m+2}\dots d_{3m+3}=1^{2m+2}$.
Let $i$ be the minimum integer such that $d_{i}d_{i+1}\dots d_{3m+i+2}=1^{3m+3}$. We consider 3 cases
\begin{itemize}
\item $i=1$ then $x=d_1d_2\dots d_{2m+1}=1^{2m+1}$ and $d_{2m+2}d_{2m+3}\dots d_{3m+3}=1^{m+2}$. Since $\bm{c}+\bm{x}=1^{m+1}d_{3m+4}d_{3m+5}\dots d_{4m+3}$ we have $\bm{c}=0^{m+1}\bm{y} $ for some $\bm{y}\in \mathbb{B}_{m}$. Thus $f(\bm{c})=1$ and since $\pi(\bm{x})=1$ we obtain $d_{2m+2}= f(\bm{c})+\pi(\bm{x})=0$ a contradiction.
\item $i=2$ then $\bm{x}=01^{2m}$ and $d_{2m+2}d_{2m+3}\dots d_{3m+4}=1^{m+3}$. Since $\bm{c}+\bm{x}=1^{m+2}d_{3m+5}d_{3m+6}\dots d_{4m+3}$ we have $\bm{c}=10^{m+1}\bm{y} $ for some $\bm{y}\in \mathbb{B}_{m-1}$. Thus $f(\bm{c})=0$ and since $\pi(\bm{x})=0$ we obtain $d_{2m+2}=f(\bm{c})+\pi(\bm{x})=0$ a contradiction.
\item $i\geq3$ then $\bm{x}=\bm{z}01^{2m-i+2}$ for $\bm{z}\in \mathbb{B}_{i-2}$ and $d_{2m+2}d_{2m+3}\dots d_{3m+2+i}=1^{m+i+1}$. Since $\bm{c}+\bm{x}=1^{m+i}d_{3m+i+3}d_{3m+i+4}\dots d_{4m+3}$ we have $\bm{c}=\overline{\bm{z}}10^{m+1}\bm{y} $ for some $\bm{y}\in \mathbb{B}_{m-i+1}$. Thus $f(\bm{c})=\pi(\overline{\bm{z}})$. Since $\pi(\bm{x})=\pi(\bm{{z}})+\pi({1^{2m-i+2}})$ and $\pi(\overline{\bm{z}})+\pi(\bm{z})=\pi({1^{i-2}})$ we obtain $d_{2m+2}= f(\bm{c})+\pi(\bm{x})=\pi({1^{2m}})=0$ a contradiction.
\end{itemize}
Therefore there exists no string $d$ in $C$ with $1^{3m+3}$ as substring.
\end{proof}
\qed
\begin{corollary}\label{cormain}
Let $n=2^p-1$ where $p\geq 2$ and let $s \geq 3.2^{p-2}$. There exists a perfect code in $\Gamma_n(1^s)$.
\end{corollary}
\begin{proof}
Indeed let $C$ be a perfect code in $Q_n$ such that no element of $C$ contains $1^{3.2^{p-2}}$ as substring.
The strings of $C$ are in $V(\Gamma_n(1^s))$. Let $x$ be a vertex of $V(\Gamma_n(1^s))$. If $x\notin C$ then $x$ is adjacent in $Q_n$ to a vertex $c$ in $C$. Note that $x$ and $c$ are also adjacent in $\Gamma_n(1^s)$ thus $C$ is a dominating set of $\Gamma_n(1^s)$. If $c$ and $c'$ are two strings of $C$ then $d_{\Gamma_n(1^s)}(c,c')\geq d_{Q_n}(c,c')\geq 3$. Therefore $C$ is a perfect code in $\Gamma_n(1^s)$.
\end{proof}
\section{Concluding remark and open problems}
Whenever $n=2^p-1$ it will be interesting to determine
the minimum $s$ such that there exists a perfect code in $\Gamma_n(1^s)$.
Corollary \ref{cormain} is not always the best result possible. For example for $n=7$ the code $C_7$ obtained in Vasilev's construction starting from $C_3=\{000,111\}$ with $f(000)=f(111)=1$ is a perfect code in $\Gamma_n(1^5)$. Indeed
\begin{itemize}
\item $11111ab$ or $0011111$ cannot be in $C_7$ since the $P(111)+1=P(001)+1=0$
\item $011111a$ cannot be in $C_7$ since the possible codewords begining with $011$ are $0111011$ and $0111100$.
\end{itemize}
Note that that all strings of this code are obtained from strings in the Hamming code of length 7 by a translation of $0001000$. This simple idea can be generalized but is less efficient than our result in the general case.
We propose also the following conjecture:
\begin{conjecture}
For $n\geq 3$ and $s\geq 1$ if $C$ is a perfect code in $\Gamma_n(1^s)$ then $n=2^p-1$ for some integer $p$ and furthermore $C$ is a perfect code in $Q_n$.
\end{conjecture}.
|
\section{Conclusions and Outlook}
\label{sec:c+o}
Future IoT networking is one of the most challenging use cases of the Internet today and a potential deployment regime of ICN. In this work, we revisited Information-Centric Networking in the IoT from a variety of perspectives and concluded that (a) publish-subscribe with named topics largely facilitates to manage the complexity of naming data, and (b) NDN without a push option for data has striking advantages for security and resilience in constrained environments. We propose HoPP, a lightweight publish-subscribe system that was implemented on RIOT and CCN-lite and experimentally evaluated on large, realistic testbeds. Our findings confirmed that constrained lossy networks can admit largely unforeseeable behaviour. Nevertheless, our approach turned out robust and resilient while performing well in the majority of experiments.
In future work, we will enhance our implementation and work towards prototypic deployment in more intricate use cases. Prior to that, we will study mobility and disruption tolerance in closer detail using multi-proxy set-ups and content redundancy. Adding an analytic model that complements our understanding of the different protocol control loops will be valuable for optimizing parameters and the overall performance.
\section{Implementation and Evaluation}
\label{sec:eval}
\subsection{Implementation for CCN-lite on RIOT}
\begin{figure}[h]
\centering
\resizebox{0.90\columnwidth}{!}{\import{figs/}{arch}}
\caption{IoT Publish-Subscribe Architecture}
\label{fig:arch}
\end{figure}
We implemented the HoPP extensions on the CCN-lite version ported to RIOT and deploy NDN. It is noteworthy that this software stack supports both, the NDN core protocol as well as CCNx. On RIOT, CCN-lite implements the {\tt netdev} interface and runs as a dedicated single-threaded network stack.
The architecture of the extended CCN-lite is depicted in Figure
\ref{fig:arch}. It mainly adds a new control protocol block that
handles exchange and processing of the two new packet types (\texttt{PAM, NAM}) on the control plane. This extends the
\texttt{forwarder} module of CCN-lite. The \texttt{forwarder} allows extensions for the
packet parsing by the use of user-defined callback functions on a suite
basis. Considering this loose coupling, the actual topology maintenance was
implemented separately from the CCN-lite core. The \texttt{topology manager}
handles \texttt{PAM} scheduling and parent selection to form and maintain the
routing topology (DODAG). Resulting forwarding states are reflected in the
FIB with the help of the CCN-lite API. The Name Advertisement
Daemon (\texttt{NAD}) module handles parsing and scheduling of \texttt{NAM}
messages. A \texttt{NAM} Cache (\texttt{NC}) is used to intermittently track
the hop-wise propagation and to reschedule \texttt{NAM} transmissions in case
of network disruptions. For each entry in the \texttt{NC}, the \texttt{NAD}
triggers the \texttt{replicator} to invoke a hop-wise content replication
on the data plane via pull-driven Interest-Data.
To ensure hop-wise replication of published content, a caching strategy was added to CCN-lite that
hinders replicated content to be cached out during publishing.
After a successful Interest-Data exchange, the \texttt{replicator} notifies the \texttt{NAD} module and the appropriate \texttt{NC} entry is freed for removal.
\begin{figure}
\includegraphics[width=0.94\columnwidth]{success_rate}
\caption{Success rate of content delivery to one consumer as a function of hop count}
\label{fig:reliability}
\end{figure}
\subsection{Basic Testbed Setup}
All experiments are conducted in the FIT~IoT-LAB testbed \cite{ivlso-14} to reflect common IoT properties.
The testbed consists of several hundreds of class~2 devices equipped with an ARM Cortex-M3 MCU,
64~kB of RAM and 512~kB of ROM, and an IEEE~802.15.4 radio (Atmel AT86RF231).
The radio card provides basic MAC layer functions implemented in hardware, such as ACK handling,
retransmissions, and CSMA/CA (Carrier sense multiple access/Collision avoidance).
The software platform is based on RIOT \cite{bhgws-rotoi-13} and the CCN-lite network stack \cite{ccn-lite}, including the protocol extensions described above.
The performance of the HoPP publish-subscribe IoT system is evaluated on the three different topologies:
\begin{LaTeXdescription}
\item[Paris] is a densely connected topology of 69 nodes all within radio reach.
\item[Grenoble (ring)] is formed of a closed rectangle with two double-stacked edges. 178 nodes form a heterogeneously meshed network with a maximal hop distance of four.
\item[Grenoble] consists of about 350 nodes, where half of them is situated on the rectangle, the other half forms linear extensions leading outwards. This network organizes in complex, fluctuating topologies with a node distance up to 9 hops.
\end{LaTeXdescription}
\subsection{Performance evaluation}
The first evaluation inspects the reliability of HoPP as compared to plain Interest notification.
We investigate the content reception rate on a given consumer in the Grenoble ring multi-hop topology using a converge cast traffic pattern, where each device generates sensor readings every $30 \pm 15$~seconds.
While HoPP is able to build and maintain the topology, static forwarding states were installed on the devices for the {Interest Notification} approach using the same routing information as HoPP.
Figure \ref{fig:reliability} compares the reliability of HoPP with the common {Interest Notification} approach in relation to the hop distance of the consumer.
For HoPP, we observe a steady high content delivery rate above $96~\%$ for all hop distances in the topology.
NDN {Interest Notification} admits significantly lower reliability and shows a decline in transmission with increasing hop distance. While a hop count of $1$ yields $~70~\%$ packet arrivals, success ratio decreases to $~41~\%$ for hop distances of $5$ and larger.
Next, we investigate performance metrics that relate to the temporal behaviour of the protocol. Since deficits of the core protocol, but also different failures of networked elements (radio/link layer, CCN-layer, pub-sub, and node layer) translate into delays due to retransmissions and re-arrangements, times to completion are a key performance indicators. In detail, we study (i) routing convergence, (ii) times to publish content items, (iii) times to publish under network partitioning, and (iv) times to issue alerts (from publisher to the subscribers).
\begin{figure}
\includegraphics[width=0.94\columnwidth]{convtime}
\caption{Routing convergence time for the testbed topologies}
\label{fig:routing_convergence}
\end{figure}
\begin{figure*}
\subfigure[Paris topology]{\includegraphics[width=0.66\columnwidth]{time_to_publish_e2e_paris_full}
\label{fig:pubtime_paris}}
\subfigure[Grenoble (ring) topology]{\includegraphics[width=0.66\columnwidth]{time_to_publish_e2e_grenoble_ring}
\label{fig:pubtime_grenoble-ring}}
\subfigure[Grenoble topology]{\includegraphics[width=0.66\columnwidth]{time_to_publish_e2e_grenoble_full}
\label{fig:pubtime_grenoble}}
\caption{Time to content publishing}
\end{figure*}
Routing convergence times in the three testbeds are displayed in Figure \ref{fig:routing_convergence}. Clearly visible is the dependence on hop counts, each counting for an average delay of $\approx 100$ ms---the PAM timer. While Paris is single-hop network and exhibits a single step in distribution, multiple steps represent hop count multiplicities in the multi-hop cases. No exceptional delays become visible. This is due to the moderate timing of the routing protocol which causes a low network utilization.
For the evaluation of the times needed to publish a content item, we iterate the following scenario. For each topology, a Content Proxy is positioned in the center of the network, while randomly chosen nodes publish a single, individually named chunk to the network. Publication is initiated every second and depending on the nodes position in the tree, one to several data packets might traverse the same sub-paths within this time frame.
Results for the single-hop network (Paris) are displayed in Figure \ref{fig:pubtime_paris}. Observing round-trip ping values of $\approx 10$~ms, the NAM timer ($\mathtt{nam_t}$) of $125 \pm 25$~ms, and the CCN-lite processing, a mean time to publish of about $135$ ms would be expected. Small fluctuations at $\approx 2 \times \mathtt{nam_t}$ indicate additional delays that result from network disturbances and node congestion leading to paths of hop count two.
\iffalse
\begin{figure}
\includegraphics[width=0.94\columnwidth]{time_to_publish_e2e_paris_full}
\caption{Time to content publishing in the Paris topology}
\label{fig:pubtime_paris}
\end{figure}
\begin{figure}
\includegraphics[width=0.94\columnwidth]{time_to_publish_e2e_grenoble_ring}
\caption{Time to content publishing in the Grenoble (ring) topology}
\label{fig:pubtime_grenoble-ring}
\end{figure}
\begin{figure}
\includegraphics[width=0.94\columnwidth]{time_to_publish_e2e_grenoble_full}
\caption{Time to content publishing in the Grenoble topology}
\label{fig:pubtime_grenoble}
\end{figure}
\fi
Similar results become visible from the Grenoble experiments in Figs. \ref{fig:pubtime_grenoble-ring} and \ref{fig:pubtime_grenoble}. Clearly pronounced are the first four routing hops, higher hop counts in Fig. \ref{fig:pubtime_grenoble} blur according to increasing fluctuations. These results clearly show the fragility of the lossy wireless regime, but also confirm a majority of these challenging transmissions did complete on the expected time scale.
We analyzed a scenario of network partitioning on the Grenoble ring topology. To quantify the effects of a major network disruption, we disabled all nodes of rank two every 60 s for an off-time interval of 60 s. This isolated the Content Proxy periodically. Content publishing proceeded randomly with a frequency of one per second.
\begin{figure}
\includegraphics[width=0.94\columnwidth]{partitioned_grenoble_ring}
\caption{Time to content publishing at network partitioning}
\label{fig:partitioning_grenoble-ring}
\end{figure}
Results in Figure \ref{fig:partitioning_grenoble-ring} highlight a smooth content transition to the CP with a timing almost linearly stretched over the 60 s off-period. No unexpected content delays become visible, which indicates the protocol robustness on this macroscopic time scale.
Finally, the end-to-end delay from the publisher to the subscriber was examined. This corresponds to the use case of issuing alerts between nodes from the local IoT network. The scenarios correspond to the previous measurements of the publishing time, i.e., publishing and subscription requests are issued randomly scattered within the topology at intervals of one second.
\iffalse
\begin{figure}
\includegraphics[width=0.94\columnwidth]{time_to_content_delivery_paris_full}
\caption{Time to issue alerts in the Paris topology}
\label{fig:alerttime_paris}
\end{figure}
\begin{figure}
\includegraphics[width=0.94\columnwidth]{time_to_content_delivery_grenoble_ring}
\caption{Time to issue alerts in the Grenoble (ring) topology}
\label{fig:alerttime_grenoble-ring}
\end{figure}
\begin{figure}
\includegraphics[width=0.94\columnwidth]{time_to_content_delivery_grenoble_full}
\caption{Time to issue alerts in the Grenoble topology}
\label{fig:alerttime_grenoble}
\end{figure}
\fi
\begin{figure*}
\subfigure[Paris topology]{\includegraphics[width=0.66\columnwidth]{time_to_content_delivery_paris_full}
\label{fig:alerttime_paris}}
\subfigure[Grenoble (ring) topology]{\includegraphics[width=0.66\columnwidth]{time_to_content_delivery_grenoble_ring}
\label{fig:alerttime_grenoble-ring}}
\subfigure[Grenoble topology]{\includegraphics[width=0.66\columnwidth]{time_to_content_delivery_grenoble_full}
\label{fig:alerttime_grenoble}}
\caption{Time to issue alerts}
\end{figure*}
The experimental output for the three topologies are displayed in Figs. \ref{fig:alerttime_paris}, \ref{fig:alerttime_grenoble-ring}, and \ref{fig:alerttime_grenoble} respectively. As we might expect, blurring fluctuations have enhanced with only a few pronounced signatures of hops and the means increased slightly by the extended paths towards the subscribers. Notably, the single-hop testbed from Paris performed best under the extended communication load, whereas the full Grenoble testbed clearly runs at its limit. The latter can be easily explained by the many hop transitions required at Grenoble, each of which requires an additional packet exchange which potentially impacts on neighbors within radio range.
Low power lossy networks that connect heavily constrained IoT nodes are known to be infeasible for such heavy load. We consider it therefore a success that a notable fraction of the content arrived at its receivers on within about 500 ms -- a timescale which is considered normal in multi-hop WPANs. To a certain degree, we account this for the robustness of our hopwise content publishing and replication protocol.
\section{Introduction}
\label{sec:intro}
The Internet of Things (IoT) is emerging, and billions of new networked devices are forecasted. However, no common networking technology for the IoT has been agreed upon. Despite of a maturing IETF protocol suite, dozens of incompatible industry solutions are rolled out to meet device and network constraints, as well as application specific needs.
Facing this huge world of mainly constrained devices, it seems worth rethinking its networking paradigm. A very loose coupling appears most appropriate between nodes that often run on battery with long sleep cycles and connect via lossy wireless links. Information-Centric Networking (ICN) \cite{adiko-sind-12,xvsft-sinr-14} decouples content provisioning from data producers in space which makes it a promising candidate. Additional decoupling in time and synchronization is desirable and attainable by a publish-subscribe layer.
Information-centric publish-subscribe networks have been proposed. PSIRP/PURSUIT \cite{lvt-psipp-10} is an early, prominent candidate. However, its central control architecture seems more suitable for an SDN-type deployment in LANs. Publish-subscribe schemes based on NDN like Content-based pub/sub \cite{cpw-cpsni-11} and COPSS \cite{cajfr-cecop-11} violate the loose coupling principle in their use of name-based routing or forwarding. Facing the current state of the art, we explore the problem of information-centric publish-subscribe for IoT networking open.
In this paper, we take up the challenge and seek for an information-centric IoT networking solution that satisfies all challenges of real-world sensor-actuator networks and allows for an easy deployment. We base our work on NDN \cite{jstp-nnc-09} not only because of its widespread availability and implementations on IoT operating systems, but in particular because of its clean request-response scheme that prevents unwanted traffic at the constrained end nodes. We design and evaluate HoP-and-Pull (HoPP), a lean, adaptive publish-subscribe layer that strictly adheres to the NDN communication pattern. Our experimental findings on large IoT testbeds indicate that our system complies indeed to the challenging requirements of IoT use case with promising performance. In particular, reliability and resilience of HoPP largely outperforms previously advised push notifications.
The structure of this paper continues as follows. In Section~\ref{sec:use-cases}, we outline distinctive use cases that motivate the following contributions. Section~\ref{sec:related_work} explores the problem space and discusses related concepts and work. In Section~\ref{sec:pub-sub}, we dive into the design details of our publish-subscribe scheme, including the key aspects of network partitioning and publisher mobility. Implementation and evaluations of our system are described in Section~\ref{sec:eval}. Finally, we conclude with an outlook in Section~\ref{sec:c+o}.
\section{HoP and Pull: A Publish-Subscribe Approach to Lightweight Routing on Names}
\label{sec:pub-sub}
\subsection{Overview}
We are now ready to describe HoP-and-Pull (HoPP), our pub-sub system for lightweight IoT deployment in detail. For a confined IoT environment, we make the common assumption that nodes form a stub network that may be connected to the outside by one or several gateways.
Some global prefix is given to a gateway, but (wireless) IoT nodes can reach a gateway without global prefix changes in one or several hops unless they are temporarily disconnected. Internally, nodes may be grouped according to one or several sub-network prefixes (e.g., {\tt /lighting}).
We select one or several distinguished nodes to serve as Content Proxies (CPs). CPs are typically more stable and more powerful such as gateways or other infrastructural entities. These Content Proxies take the role of data caches and persistent access points. They will be reachable throughout the network by default routes, unless temporary partitioning occurs. Note that one CP can serve several local prefixes, but a local prefix may also belong to several CPs. The latter scenario will lead to replicated caching with higher and faster data availability.
Our publish-subscribe protocol for the IoT is then composed of three core primitives:
\begin{enumerate}
\item Establishing and maintaining the routing system
\item Publishing content to the CPs
\item Subscribing content from the CPs
\end{enumerate}
Our following protocol definition strictly complies with the design principles: (a) minimal FIBs that only contain default routes, (b) no push primitive or polling, (c) no broadcast or flooding on the data plane.
\subsection{Prefix-specific default routing}
\label{sec:default_routing}
Content Proxies advertise the prefix(es) they own on the control plane to all neighbors in a Prefix Advertisement Message (PAM). Observing nodes will adopt a CP as their parent and re-broadcast the PAM message with an increased distance value. Much like in the core RPL \cite{RFC-6550}, all nodes will be members of a Destination-Oriented Directed Acyclic Graph (DODAG) after routing convergence. Nodes will include the selected best uplink in their FIB as default route to the announced prefix, but may add additional uplinks with lower priority.
\begin{figure}[h]
\centering
\import{figs/}{build_dodag}
\caption{Building a routing DODAG by prefix advertisements}
\label{fig:build_dodag}
\end{figure}
Figure \ref{fig:build_dodag} visualizes the PAM prefix distribution and the corresponding FIB entry for the sample prefix /$\rho$/. All nodes establish a default on shortest paths upstream. In addition, node 4 learns a backup path of equal hop distance, but lower radio quality.
\subsection{Publishing content}
\label{sec:pub}
An IoT node (sensor) that has new data to publish will first select a name. It may choose either from a predefined scheme accessible by local controllers, some common standard set, or decide individually. It will advertise this content name to its upstream neighbor via a (unicast) Name Advertisement Message (NAM). It will also associate the content with one or several topic names and adds these to the content metadata.
\begin{figure*}[t]
\centering
\import{figs/}{publish}
\caption{Publishing new content: Advertising names and pulling content hop-by-hop}
\label{fig:publish}
\end{figure*}
Under regular network conditions, the upstream neighbor is expected to retrieve the advertised content via the incoming interface of the NAM. It proceeds according to the standard NDN scheme: An Interest requests the name, the data is returned in response. Concurrently, the upstream issues a corresponding NAM to its parent, which in turn pulls the content one hop closer to the CP. This hop-wise content replication proceeds until the data arrives at the Content Proxy.
It is worth noting that the NAM content alerting is situated on the control plane using {\em link-local unicast} signaling. Neither a data path is established in the PIT, nor are FIBs modified. Hop-wise content retrieval is also more robust to changing network conditions, while experiencing little temporal overhead when executed in parallel.
The publishing mechanism is depicted in Figure~\ref{fig:publish}. Publisher~3 issues a NAM to its parent 1, which requests the content and republishes the NAM to the CP in parallel. After arrival of the data, node 1 can satisfy the Interest which was received by the CP.
Under irregular network conditions, a node may not receive an Interest that matches its previous name advertisements. This may be due to broken links, failing or deep-sleeping nodes, or enduring overload. After a deployment-specific timeout, the content owner will adapt and try to publish the content on an alternate path by sending a NAM up on a backup link. In case of a complete failure, the content node can follow two strategies: Either it waits and re-advertises according to an exponential back-off, or it solicits a refresh of router advertisements for learning new, operational routes.
\subsection{Subscribing to content}
A subscriber in HoPP behaves almost like any content requester in NDN. It issues a regular Interest request up the default route to the CP and awaits the response. There are two deviations from plain NDN, though. First, the subscriber cannot extract content names from its FIB, since FIBs only contain prefixes. Second, it does not expect an immediate reply, but issues Interests with extended lifetimes.
\begin{figure*}[t]
\centering
\import{figs/}{subscribe}
\caption{Content subscription: Requesting name by topic with asynchronous delivery}
\label{fig:subscribe}
\end{figure*}
Names are expected to follow an application-specific logic. Following up the discussion in Sections \ref{sec:naming+context}, we argue that names (of content or topics) in machine-to-machine communication must be processable in the context of the endpoint and thus known. Names of individual content items can be learned by issuing Interests on topics. The corresponding CP will then answer the request with an empty data chunk that carries available content name(s) as metadata.
Figure \ref{fig:subscribe} displays the operations of a subscriber. An Interest for named content is sent up to the proper prefix owner (CP) and remains for a predefined lifetime, if the Content Proxy cannot supply the data. In case content is arriving from a publisher to the CP, data is transferred automatically down the reverse Interest path---as a regular NDN operation. We anticipate that in common sensor-actuator networks of the IoT, the application semantic will define meaningful Interest lifetimes. Otherwise, in regimes of largely fluctuating temporal behaviours or long-lasting subscriptions (e.g., alerts), the subscriber may refresh and maintain the request at its discretion.
Note that in contrast to {\em long-lived Interests} or the COPSS {\em subscription tables} (s. Sec. \ref{sec:related_work}), such Interests of extended lifetime are consumed by arriving content and do not open a persistent, uncontrolled data path. Subscribers continue to apply flow control and may discontinue subscriptions to unwanted content.
\subsection{Publisher mobility and network partitioning}
A publishing node that moves from one point of attachment to another within the IoT domain, will experience stable routing conditions in the sense that default routes to active prefixes should exist everywhere in a connected network. Correspondingly, the mobile node (MN) can re-configure its upstream route either by wait for the next prefix advertisement (PAM), or may actively solicit an additional PAM. Note that these link-local route configurations closely resemble the autoconfiguration of IPv6 default gateways. However, in contrast to mobile IPv6, the MN in our publish-subscribe system can continue publication immediately after a link-local route is established.
\begin{figure}[h]
\centering
\import{figs/}{publisher_mobility}
\caption{Publisher Mobility: Switching DODAGs}
\label{fig:publisher_mobility}
\end{figure}
Figure \ref{fig:publisher_mobility} illustrates provider mobility. Node 3 removes from the network while trying to publish a content item and enters the radio range of node 4. It may now actively learn about network re-attachment (e.g., from link triggers), or learn from a newly arriving PAM. After the local upstream is configured, the mobile publisher can successfully complete its publishing handshake.
\begin{figure}[h]
\centering
\import{figs/}{partitioning}
\caption{Temporary network partitioning: Interim Content Proxies (ICPs) buffer publishing}
\label{fig:partitioning}
\end{figure}
Temporary network partitioning proceeds very similar to mobility. An intermediate node that looses upstream connectivity will explore alternate paths (cf. Sec.~\ref{sec:pub}), but has to await a re-attachment in case of a complete failure. Such node will continue to receive publishing demands (NAMs) from the downstream, which it will satisfy in accordance with its resources. On overload, it will terminate to retrieve content from its children. Proceeding this way will establish a classic backpressure mechanism of flow control.
Operations under network partitioning are shown in Figure~\ref{fig:partitioning}. Following an outage of the CP, nodes 1 and 2 experience a disconnect. They continue to handle publications (as well as subscriptions) until connectivity to the CP is reestablished.
\section{The Problem of Information Centric IoT Networking and Related Work}
\label{sec:related_work}
\subsection{Deployment in the constrained IoT}
Things in the IoT are often represented by small embedded controllers which possess orders of magnitude less resources (kBytes of memory, MHz CPU speed, mW of power) than regular Internet nodes, but still need to communicate using protocols that interoperate in a shared infrastructure.
These things are commonly sensors or actuators that speak with a remote 'cloud' or talk with each other locally. The predominant communication for edge devices happens on wireless channels of low power lossy networks (LLNs) in the battery-powered world. Following the IEEE 802.15.4, BLE, or LWPAN standard, these nodes can exchange only small packets at very low rates and sleep frequently. Violating constraints quickly leads to successive overload, extreme packet losses, and may strongly degrade network operation or node availability. Repeated incidents have told that the mass of IoT nodes can be both highly threatened and a threat to the global Internet.
\subsubsection{ICN in the IoT}
It became apparent \cite{olg-ccnte-10,bmhsw-icnie-14,RFC-7476} that ICN/NDN exhibit great potentials for the IoT. Not only allows the access of named content instead of distant nodes a much leaner and more robust implementation of a network layer, but in particular prevents the request-response pattern of NDN any overloading with data at the receiver. For a few years, it was the believe that NDN can be DoS resistant by design, until Interest- and state-based attacks were discovered \cite{wsv-bipmc-12}. Subsequent work \cite{gtuz-ddndn-13,wsv-bdpts-13} elaborated the threats of Interest flooding and overloading FIB and PIT tables by user-generated names and content requests. This has proven difficult to mitigate \cite{sws-rcani-15} and is a particular threat to memory-constrained nodes.
In the subsequent Section \ref{sec:pub-sub}, we will show how a FIB with simple default routes can serve the IoT, and how PITs remain minimal by hop-wise content replication between nodes.
ICN deployment in the IoT has been studied with increasing intensity, touching protocol design aspects \cite{bmhsw-icnie-14,pf-britu-15,sqvcg-simab-16,abcmr-inmcd-16,mwt-tucin-16}, architecture work \cite{g-ainai-17,szsmb-avdir-17}, and practical use cases \cite{bgnt-sieoc-13,acim-icnis-15,srs-sndnt-15,gkslp-inii-17}. Emerging link-layer extensions for the wireless like TSCH turned out to be beneficial for the interaction of NDN communication patterns and channel management \cite{habsw-itpla-16}. Several implementations have become available. CCN-Lite \cite{ccn-lite} runs on RIOT \cite{bhgws-rotoi-13,bmhsw-icnie-14} and on Contiki \cite{dgv-clfos-04,alw-defsc-16}, NDN has been ported to RIOT \cite{saz-dinps-16}. Thus, grounds seem to be prepared for opening the floor to real-world IoT applications with NDN.
Many deployments in the IoT, though, follow the communication patterns {\em on demand}, {\em scheduled}, and {\em unscheduled}. Actuators in particular rely on unscheduled control messages. Since NDN is built on the request-response scheme of data-follows-Interest, unscheduled push message are not natively supported. For the IoT, this has been identified as a major research challenge \cite{RFC-7927}.
\subsubsection{Push communication}
Several extensions have been proposed to enable an unsolicited push of data, among them {\em Interest-follows-Interest} \cite{bgnt-sieoc-13}, {\em Interest notification} \cite{acim-ndnia-14}, and a dedicated {\em push packet} \cite{draft-ravi-icnrg-ccn-notification}. All these push messages are sent immediately to a prospective consumer node, which not only conflicts with the ICN paradigm of naming content instead of hosts, but has no forwarding supported on the network layer. No push packet will reach its destination unless potential receivers are announced to the routing using a node-centric name. Unidirectional data push to named nodes, however, lacks flow as well as congestion control, and opens an attack surface to DoS. In the IoT with its constrained nodes, this must be rated a particularly severe disadvantage.
Carzaniga et al. \cite{cpw-cpsni-11} with a proposal of {\em long-lived Interest} seem to be the first in addressing the push challenge in a natural NDN fashion. Subscribers issue a persistent Interest that is not consumed at content arrival, and thereby establish a (static) data path from the producer. Unfortunately, long-lived Interests open an unrestricted data path to the recipient and thereby inherit the threats of overload as other push primitives. In addition, persistent forwarding states in PITs lead to self-reinforcing broadcast storms whenever L2 broadcasts are used \cite{kgshw-nnmam-17}.
Finally, frequent topology changes as characteristic for the IoT will routinely break paths. In the following, we will show how regular Interests with appropriate lifetime can serve this purpose equally well, without suffering from its drawbacks.
\subsubsection{The role of a control plane}
Lessons from Internet decades have told that the networking layer should be composed of well defined and clearly separated control and data planes. NDN has primarily focused on a stateful forwarding plane.
We argue that the ICN community has payed too little attention on clearly separating a control plane \cite{wsv-lpwds-13}.
Current proposals of routing protocols that fill NDN FIBs mainly rely on brodcasts, and often misuse Interest messages of the forwarding plane to disseminate control information. The distance vector content routing protocol DCR \cite{g-ncric-14} and the link state content routing protocol LSCR \cite{hg-nanlr-15} use broadcast pushs to
distribute control traffic over multiple hops. This flooding is controlled by
utilizing sequence numbers and anchor nodes that store copies of the content.
An approach to reducing traffic overhead by scoped-flooding is outlined in Pro-Diluvian \cite{wboks-puscd-15}.
The set synchronization protocols ChronoSync \cite{za-lcdds-13}, iSync
\cite{fbc-snibf-15} and PartialSync \cite{zlw-pespn-16} rely on a broadcast-pull
pattern, where an Interest message containing name information is distributed
into a broadcast domain and served by the first node that maintains
conflicting name information. NLSR \cite{haazz-nnlsr-13} is a link state routing
protocol that uses ChronoSync to distribute link information in the same
manner.
Panini \cite{swbw-lcnhp-16} explicitly defines a unicast name advertisement message (NAM) on the control plane which we will re-use when designing the publish-subscribe scheme in Section \ref{sec:pub-sub}.
\subsection{Naming and routing}
Naming content on an information-centric network layer promises a simplified access to information. Routing on names directly designs a lean network without further address mapping. It obsoletes infrastructure like the DNS and eliminates the attack surface inherent to the mapping. Both aspects are of great advantage in a constrained IoT network.
However, name-based routing encounters the problems of (a) exploding routing tables, as the number of names largely exceeds common routing resources, and (b) limited aggregation potentials, as names are specific to appliances and applications, but independent of content locations. More severely and in contrast to IP, a local router cannot decide on aggregating names since the symbol space of names is not enumerable in practice \cite{swbw-lcnhp-16}. Limiting the complexity of name-based routing and FIB table state is one of the major challenges in IoT networks \cite{RFC-7927}.
\subsubsection{Naming in context}
\label{sec:naming+context}
In a typical IoT scenario, there are sensor readings that are reported to a (remote) cloud, or to a controller that operates actuators. In some cases (s. Section \ref{sec:use-cases}), sensors are co-located with a controller that generates control information for immediate actuation---a safety alert for example after a sensor threshold was exceeded.
Names need to be shared between the sending and the receiving side so that requests can be issued. Advertising all names throughout the routing system is infeasible and will quickly explode the FIBs. However, there are ways to mitigate this. An application-specific common knowledge, or standard naming schemes for sensor data \cite{draft-ietf-core-senml} and alerts may obsolete the need to distribute every name to the FIBs. More generally, named {\em topics} serve as the common link in publish-subscribe systems.
In a sense, this natural approach relates to an old discussion about accessing named information in Hypermedia. Before the invention of the Worl Wide Web, Landow \cite{l-rhm-89} already pointed out that information exchange always carries two contexts, the context of departure and that of arrival. Departure and arrival translate to publish and subscribe in our discussion.
\subsubsection{Name-based routing, forwarding, and caching}
Routing normally proceeds according to location information from the FIB. Names in FIBs only aggregate well if naming follows the topological hierarchy of the network. This rarely holds, since naming is application-specific, and cannot be detected without distributed knowledge. To overcome FIB explosion, several authors refer to the NDN capabilities of stateful forwarding, using the option of distributing requests to several interfaces simultaneously \cite{yamwz-csfp-12,yaawz-rrndn-14}. Such Interest multicasting will lead to duplicate content deliveries if the network is densely connected. In 'Pro Diluvian' \cite{wboks-puscd-15}, the effects of such scoped flooding are analyzed, and authors find a utility limited over very few ($\approx$\ 2--3) hops. Such opportunistic forwarding can also lead to loops, as was pointed out by Garcia-Luna-Aceves \cite{gm-lfpcn-16}. In any case, the excessive traffic, as well as redundant PIT states make this approach infeasible for the IoT.
COPSS \cite{cajfr-cecop-11}, an earlier publish-subscribe approach inspired by PIM \cite{RFC-4601} multicast routing, selects a rendezvous point to interconnect publishers and subscribers. Such dedicated routing point naturally allows for name aggregation. Like PIM-SM (Phase 2), COPSS further establishes a dedicated forwarding infrastructure (subscription table) that establishes persistent forwarding paths from the publisher via the rendezvous point to the receivers. PANINI \cite{swbw-panii-15,swbw-lcnhp-16} re-uses the idea of an aggregation point called Name Collector, but does not establish a (persistent) forwarding plane like COPSS. Instead, PANINI uses selective broadcasts to discover unpopular routes towards the network edge. For the IoT, we want to minimize control traffic and avoid flooding. We restrict our solution to a lean default routing, instead.
The ICN support of data replication and caching is of particular interest for the IoT, where wireless channels are lossy and nodes are often asleep. Hop-wise data transport with intermediate storage of chunks is a built-in feature of NDN which we extend to account for node heterogeneity. IoT deployments often consist of very constrained nodes at the edge with more powerful border routers, gateways, or other node infrastructure---many of them equipped with larger hardware, electrical connectivity, and network uplinks. In the following, we will make use of Content Proxy nodes, which are meant to be chosen from this kind.
\subsubsection{Mobility and network partitioning}
Mobile nodes are part of many IoT deployments. While mobility is natively supported at the receiver side of NDN, publisher mobility is considered difficult to solve in a generic way \cite{tscrm-smin-13}. Translated to IoT use cases, this means mobile sensors are hard to integrate---a particular problem for surveillance and safety sensing applications. These use cases may also experience temporary network partitioning (see Section \ref{sec:use-cases}), which can be treated with correspondence to network mobility.
Several solutions have been built for specific applications \cite{wakvw-rtidu-12,gpwpv-hpvin-13}, but the complexity of the name-based routing system often withstands a generic mobility management. We will show in the following how prevalent default routes can naturally accommodate publisher mobility, as well as network partitioning.
\section{Deployment Considerations for IoT Use Cases}
\label{sec:use-cases}
In this section, we focus on two use cases for the deployment of an information-centric IoT---the simple, well-known Lighting Control \cite{bgnt-sieoc-13}, and the more challenging application of an industrial Internet for Safety Control in harsh environments.
\subsection{Lighting control}
Smart lighting control is essentially the task of setting the state of various lights according to preconfigured scenarios in response to triggering events. The latter may be generated by plain switches, complex controllers, or by other machinery like an elevator that is transporting people to a currently unilluminated floor. Configuring the proper light consists of turning various fixtures into selective settings.
We revisit this basic use case, because it raises two interesting aspects of networking. First, lighting control foremost follows an actuation pattern, i.e., different signals request for immediate state changes at specific groups of fixtures. The information of turning a switch must somehow propagate to distributed ensembles of lights under soft real-time constraints. Burke et al. \cite{bgnt-sieoc-13} define authenticated Interests to push signalling to the actuator, thereby inverting the NDN request-response pattern. We argue for preserving the NDN communication paradigm in Section \ref{sec:related_work}.
Second the deployment of names is closely related to the application logic and often more involved than accessing data directly. Lighting control may switch individually located fixtures (e.g., corridor light 5), or fixture groups (e.g., room 5, front), activate functions (e.g., fading), or integrate aspects into schemes (e.g., background illumination). Smart systems most likely combine lighting control features with further sensor readings (e.g., user presence, brightness detection) to apply adaptive functions to varying device groups etc..
While authors in \cite{bgnt-sieoc-13} chose to combine locations and applications within names that are preconfigured by a control manager, we argue that preconfigured application groups at the device level are too static and violate the device context: IoT devices have an identity, capabilities, and sometimes a known location. Their role in varying application contexts, though, is extrinsic and requires a coordinating function on the application level. This cannot be hard-coded in data names.
\subsection{Industrial safety networks}
Industrial safety and control systems are increasingly interconnected and often operate under harsh conditions. In this use case, we consider industrial environments with a threat of hazardous contaminant (e.g., explosive gas) that need continuous monitoring by stationary, as well as mobile sensors. In case of an emergency, immediate actions are required such as issuing local alarms, activating protective shut-downs (e.g., closing valves, halting pumps), initiating a remote recording for first responders and forensic purposes.
Typical industrial plants are widespread with sparse network coverage, so that mobile workers or machines face intermittent connectivity at scattered gateways. Some sensors and actuators are infrastructure bound, others are independent, battery-powered embedded devices (e.g., body equipment). The latter aspects resemble the challenges faced in previous DTN-work such as in mines \cite{gkrao-dcm-10}.
Like the previous, this use case relies on a fast sensor-actuator network including embedded IoT nodes. In addition, the harsh industrial environment raises the challenges of mobile, intermittently connected end nodes, and network partitioning. Still, enhanced reliability is required in the safety context. We will show in the following, how configurable data replication with dynamically generated content proxies can meet these challenges and how they combine in a lightweight system suitable for real-world deployment \cite{gkslp-inii-17}.
|
\section{Introduction}
Hartree--Fock theory is ubiquitous in quantum chemistry.
Representing the many-electron wave function as a single Slater determinant, the Hartree--Fock approximation provides a mean-field description of molecular electronic structure.\cite{SzaboOstlund}
Through the long established self-consistent field method (SCF), the Hartree--Fock energy is minimised with respect to variations of a set of orbitals expressed in a given finite basis set.
This optimal set of orbitals therefore exists as a stationary point of the energy.\cite{Hall1951, Roothaan1951}
However, it is less widely appreciated that the nonlinear form of the SCF equations can lead to convergence onto a range of different solutions.\cite{Lions1987, Thom2008}
These solutions represent additional local minima, maxima or saddle points of the energy.
Through methods including SCF metadynamics\cite{Thom2008} and the maximum overlap method (MOM)\cite{Gilbert2008}, locating higher energy stationary points has become relatively routine, and several authors have sought to interpret these as physical excited states.\cite{Gilbert2008, Besley2009, Barca2014, Peng2013, Glushkov2015}
However, for many systems there exist multiple low energy solutions that may cross as the geometry changes, presenting a dilemma when correlation techniques require a single reference determinant to be chosen.\cite{Thom2009}
Recently there has been increasing interest in using multiple Hartree--Fock states as a basis for non-orthogonal configuration interaction (NOCI) calculations, providing a more egalitarian treatment of individual low energy SCF solutions.\cite{Malmqvist1986, Ayala1998, Thom2009, Mayhall2014, Sundstrom2014, Jake2017}
Since each Hartree--Fock state is itself an mean-field optimised solution, excited configurations can be accurately represented in the NOCI basis set\cite{Sundstrom2014}.
Consequently, NOCI also provides an alternative to the multiconfigurational Complete Active Space SCF (CASSCF)\cite{Helgaker} approach
using an ``active space'' of relevant Hartree--Fock determinants.
This results in wave functions that reproduce avoided crossings and conical intersections at a similar scaling to the SCF method itself.\cite{Thom2009, Jake2017}
Furthermore, the inherent multireference nature of NOCI enables strong static correlation to be captured, whilst additional dynamic correlation can subsequently be computed using the pertubative NOCI-MP2 approach.\cite{Yost2013, Yost2016}
Unless it is strictly enforced, there is no guarantee that SCF solutions possess the same symmetries as the exact wave function.\cite{Fukutome1981, Jimenez-Hoyos2012}
Restricted Hartree--Fock (RHF) wave functions, for example, are eigenfunctions of the spin operator $\hat{S}^2$ but may break the molecular point group symmetry at singlet instabilities\cite{Mestechkin1978,Mestechkin1979,Mestechkin1988}.
In constrast, the unrestricted Hartree--Fock (UHF) approach allows the wave function to break both spatial and $\hat{S}^2$ symmetry, leading to spin contaminated states containing a mixture of singlet and triplet components.\cite{Fukutome1981,Fukutome1974, Fukutome1975}
Alongside capturing static correlation, including symmetry broken SCF states in a NOCI calculation allows spatial symmetry to be restored and reduces spin contamination in a similar style to the Projected\cite{Lowdin1955, Scuseria2011, Ellis2013, Jimenez-Hoyos2012} and Half-Projected\cite{Smeyers1973,Smeyers1974,Cox1976,Smeyers1976} Hartree--Fock approaches.
However, as a projection after variation approach, NOCI retains the size--consistency of the SCF determinants to provide size--consistent approximations for singlet and triplet wave functions.
Crucially, NOCI requires the existence of multiple Hartree--Fock solutions across all molecular geometries of interest to ensure the basis set size is consistent and prevent discontinuous NOCI energies.\cite{Thom2009, Mayhall2014}
There is, however, no guarantee that Hartree--Fock states must exist everywhere, and in fact they often vanish as the geometry is varied.
This is demonstrated by the coalescence of the low energy UHF states with the ground state RHF solution at the Coulson--Fischer point in \ce{H2}\cite{Coulson1949}, although further examples are observed in a wide array of molecular systems.\cite{Dunietz2003, Cui2013, Mori-Sanchez2014}
To construct a continuous basis set of SCF solutions for NOCI, Thom and Head--Gordon proposed that Hartree--Fock states may need to be followed into the complex plane.\cite{Thom2009}
We have recently reported a holomorphic Hartree--Fock theory as a method for constructing a continuous basis of SCF determinants in this manner.\cite{Hiscock2014, Burton2016}
In holomorphic Hartree--Fock theory, the complex conjugation of orbital coefficients is removed from the standard Hartree--Fock energy to yield a complex differentiable function which we believe has a constant number of stationary points across all geometries.\cite{Hiscock2014}
Using a revised SCF method, we have demonstrated the existence of holomorphic UHF (h-UHF) solutions for \ce{H2}, \ce{H4^2+} and \ce{H4}\cite{Burton2016}.
The h-UHF states exist across all geometries, corresponding to real Hartree--Fock solutions when these are present and extending into the complex plane when the real states disappear.
Despite the promise of holomorphic Hartree--Fock theory, there is currently limited understanding about the nature of holomorphic solutions.
For example, underpinning this theory we believe that the number of stationary points of the holomorphic energy function is constant (including solutions with multiplicity greater than one), and thus states must exist for all geometries.
In the current work, we attempt to understand the simplest application of holomorphic Hartree--Fock theory by investigating the holomorphic RHF (h-RHF) solutions to two--electron problems.
First, we outline the key concepts of the theory before providing a derivation for the exact number of h-RHF states for two electrons in $n$ basis functions.
In doing so we demonstrate that this number is constant for all geometries.
We then study the full set of h-RHF states for \ce{HZ}, \ce{HHeH^2+} and \ce{HHeH} (STO-3G), investigating the behaviour of these states as molecular geometry or atomic charges are varied, and discussing the isomorphism between systems with two electrons and two electron holes.
Finally, we investigate the h-RHF states of ethene (STO-3G) and demonstrate the application of NOCI to its internal rotation by considering the $\uppi$ and $\uppi^*$ orbitals with frozen core and virtual orbitals as a two--electron SCF problem.
\section{Holomorphic Hartree--Fock Theory}
We begin with an orthonormal set of $n$ real single-particle basis functions, denoted $\{ \chi_{\mu} \}$, from which the closed-shell molecular orbitals can be constructed as
\begin{align}
\phi_{i} = \sum_{\mu}^{n} \chi_{\mu} c_{\mu i}.
\label{eq:MolecularOrbitals}
\end{align}
In standard RHF theory, to ensure orthogonality of molecular orbitals, the orbital coefficients $\{ c_{\mu i} \}$ are elements of a unitary matrix
\begin{align}
\sum_{\mu}^{n} c_{\mu i}^* c_{\mu j}^{\vphantom{*}} = \delta_{i j}.
\label{eq:UnitaryConstraint}
\end{align}
The density matrix is then constructed as $P_{\mu \nu} = \sum_{i}^{N} c_{\mu i}^{\vphantom{*}} c_{\nu i}^*$, where $N$ is the number of occupied spatial orbitals, and the Hartree--Fock energy is given by
\begin{align}
E &= h_{0} + 2 \sum_{\mu \nu}^n P_{\mu \nu} h_{\mu \nu} \nonumber \\
&+ \sum_{\mu \nu \sigma \tau}^n P_{\mu \nu} \left[ 2( \mu \nu | \sigma \tau) - ( \mu \tau | \sigma \nu) \right] P_{\sigma \tau}.
\label{eq:ConvetionalHartreeFockE}
\end{align}
where $h_{0}$ is the nuclear repulsion, $h_{\mu \nu}$ are the one-electron integrals and $( \mu \nu | \sigma \tau)$ are the two-electron integrals.
RHF solutions then exist as stationary points of Equation \ref{eq:ConvetionalHartreeFockE} constrained by Equation \ref{eq:UnitaryConstraint}.
Conventionally, the Hartree--Fock energy function is considered only over the domain of real orbital coefficients.
Extending this domain to the complex plane turns Equation \ref{eq:ConvetionalHartreeFockE} into a function of several complex variables $\{ c_{\mu k} \}$ and their complex conjugates $\{ c_{\mu k}^* \}$.
However, since the dependence on $\{ c_{\mu k}^* \}$ violates the Cauch--Riemann conditions,\cite{Fischer} the Hartree--Fock energy is usually interpreted as a function of the real variables $\{ \Re \left[ c_{\mu k} \right] \}$ and $\{ \Im \left[ c_{\mu k} \right] \}$ to ensure gradients are well-defined.
Consequently, $E$ remains a polynomial of only real variables and we cannot expect the number of solutions to be constant for all geometries, as previously demonstrated in the single variable case.\cite{Hiscock2014}
In holomorphic Hartree--Fock, states that disappear can be followed into the complex plane by defining a revised complex-analytic energy as a function of the holomorphic density matrix
$\widetilde{P}_{\mu \nu} = \sum_{i}^{N} c_{\mu i} c_{\nu i}$,
where now the complex conjugation of orbital coefficients has been removed.
Since $\widetilde{P}$ is a complex symmetric matrix, its eigenvectors --- which form the holomorphic one-electron orbitals --- must be complex orthogonal\cite{Craven1969, Gantmacher} and the orbital coefficients are elements of a complex orthogonal matrix such that
\begin{align}
\sum_{\mu}^{n} c_{\mu i} c_{\mu j} = \delta_{i j}.
\label{eq:ComplexOrthogonalConstraint}
\end{align}
The h-RHF energy is then defined in terms of $\widetilde{P}$ as
\begin{align}
\label{eq:HolomorphicHartreeFockE}
\widetilde{E} &= h_{0} + 2 \sum_{\mu \nu}^n h_{\mu \nu} \widetilde{P}_{\mu \nu} \nonumber \\
&+ \sum_{\mu \nu \sigma \tau}^n \widetilde{P}_{\mu \nu} \left[2(\mu \nu | \sigma \tau) - (\mu \tau | \sigma \nu) \right] \widetilde{P}_{\sigma \tau}.
\end{align}
With no dependence on the complex conjugate of orbital coefficients, this function is a complex analytic polynomial which, by taking inspiration from the fundamental theorem of algebra, we believe must have a constant number of solutions at all geometries.\cite{Hiscock2014, Burton2016}
\section{Enumerating the h-RHF states}
The closed-shell h-RHF approach with two-electrons, described by a single molecular orbital $\phi$ and $n$ orbital coefficients $\{ c_{\mu} \}$, provides the simplest system in which we can consider proving the number of holomorphic Hartree--Fock states is constant for all geometries.
In this case, $\phi$ is constructed from a linear combination of $n$ real orthogonal basis functions as
\begin{equation}
\phi\left( \mathbf{r} \right) = \sum_{\mu=1}^n c_{\mu} \chi_{\mu} \left( \mathbf{r} \right),
\label{eq:basis_expansion}
\end{equation}
with the requirement for complex orthonormalization introducing the constraint
\begin{align}
\sum_{\mu=1}^n c_{\mu}^2 = 1 .
\label{eq:normalisation}
\end{align}
The holomorphic restricted Hartree--Fock energy is given by the polynomial
\begin{align}
\label{eq:HoloEnergy}
\widetilde{E} \left( c_1,\dots, c_n \right) &= h_{0} + 2 \sum_{\mu, \nu=1}^n h_{\mu \nu} c_{\mu} c_{\nu} \nonumber \\
&+ \sum_{\mu, \nu, \sigma, \tau=1}^n h_{\mu \nu \sigma \tau} c_{\mu} c_{\nu} c_{\sigma} c_{\tau} ,
\end{align}
where $ h_{\mu \nu \sigma \tau} = 2(\mu \nu | \sigma \tau) - (\mu \tau | \sigma \nu)$, and the h-RHF states exist as stationary points constrained by Equation \ref{eq:normalisation}.
Identifying the number of these stationary points can be achieved through the mathematical framework of algebraic geometry.\cite{Hartshorne}
Algebraic geometry forms a vast and complex field, encompassing the study of solutions to systems of polynomial equations in an affine or projective space.
Affine spaces provide a generalisation to Euclidean space independent of a specific coordinate system.
An $n$-dimensional affine space $\mathbb{A}^n = \mathbb{C}^n$ is described by the $n$-tuples $(a_1, \dots, a_n)$, where $a_i \in \mathbb{C}$ are coordinates of the space.
Alternatively, a projective $n$-space $\mathbb{P}^n = \mathbb{C}^{n+1}$ is described by the $(n+1)$-tuples $(a_0, \dots , a_n)$ under the scaling relation $(a_0 , \dots , a_n) \sim ( \lambda a_0 , \dots , \lambda a_n)$, where $\lambda$ is a non-zero scalar and the point $(a_0 , \dots , a_n) = 0$ is excluded.\cite{Hartshorne}
An affine space can be viewed as the subset of a projective space where $a_0 \neq 0$.
In contrast, points where $a_0=0$ are referred to as ``points at infinity'' and allow geometric intersection results to be consistently defined without exceptions.
For example, in $\mathbb{A}^2$ two lines must always intersect exactly once unless they are parallel, whilst in $\mathbb{P}^2$ parallel lines intersect at a point at infinity.
Therefore, in the projective space $\mathbb{P}^2$, the intersection rule is generalised without exceptions.
\begin{figure*}[tbh!]
\input{figures/TikZ_circle_projection.tex}
\hspace{1em}\includegraphics[scale=1.25, trim={3.7cm 2cm 1.2cm 2.cm}, clip]{figures/HH_RHF_surf}
\caption{
Constrained projective h-RHF energy $F_{\bar{X}} $ for the $n=2$ system \ce{H2} (STO-3G) at a bond length of $2.5$\r{A}.
The coordinate $c_0$ defines the normalisation constant, representing the distance of a point from the origin, and $\frac{\partial F_{\bar{X}}}{\partial c_0} = 0$ for all $c_0 \neq 0$.
Exploiting this invariance, every stationary point constrained to the circle $c_1^2 + c_2^2 = 1$ (black curve) with $c_2 \neq 0$ can also be located as a stationary point constrained to the line $c_2 = 1$ (blue line), where rescaling recovers the normalised h-RHF state (dashed line).
Due to the overall sign symmetry, only half the stationary points need to be considered (filled circles vs. open circles).
}
\label{fig:n_2_solutions}
\end{figure*}
Using this terminology, the spatial orbital $\phi$ with $n$ basis functions is represented by a point $\left( c_1,\dots, c_n \right)$ in the affine space $\mathbb{A}^{n}=\mathbb{C}^{n}$.
The holomorphic Hartree--Fock energy (Equation \ref{eq:HoloEnergy}) is a function $\widetilde{E} : \mathbb{A}^{n} \rightarrow \mathbb{C}$ given by a polynomial of degree 4.
Satisfying the normalisation constraint (Equation \ref{eq:normalisation}) involves restricting solutions to the hypersurface $X \subseteq \mathbb{A}^n$ defined as
\begin{align}
\label{eq:AffineHypersurface}
X = \Big\lbrace (c_1, \dots, c_n) \in \mathbb{A}^{n} \mid \sum_{\mu=1}^n c_{\mu}^2 =1 \Big\rbrace .
\end{align}
Points corresponding to h-RHF states are then the vanishing points of the differential $\mathrm{d} \widetilde{E}$ restricted to $X$.
To enable a complete enumeration of these points, we must first convert to the projective space $\mathbb{P}^{n}$ represented by the points $\left( c_0,\dots,c_n \right)$.
This is achieved through the mapping $(c_1, \dots, c_n) \mapsto (\frac{c_1}{c_0}, \dots, \frac{c_n}{c_0})$, converting all polynomials in the affine coordinates $\{ c_1, \dots, c_n \}$ to homogeneous polynomials in the projective coordinates $\{ c_0, \dots, c_n \}$.
Following this transformation, the constraint becomes
\begin{align}
\label{eq:ProjectiveConstraint}
\sum_{\mu=1}^n c_{\mu}^2 = c_0^2
\end{align}
and solutions are therefore restricted to the hypersurface $\widebar{X} \subseteq \mathbb{P}^n$ defined as
\begin{align}
\label{eq:ProjectiveHypersurface}
\widebar{X} = \Big\lbrace (c_0, \dots , c_n) \in \mathbb{P}^{n} | \sum_{\mu=1}^n c_{\mu}^2 = c_0^2 \Big\rbrace .
\end{align}
The holomorphic energy can then be written as a rational function on $\mathbb{P}^n$
\begin{align}
\label{eq:ProjectiveHoloEnergy}
F\left(c_0,\dots, c_n \right) &= \widetilde{E} \left( \frac{c_1}{c_0}, \dots, \frac{c_n}{c_0} \right) \nonumber \\
&= \frac{\widebar{E}\left(c_0,\dots, c_n \right)}{c_0^{4}},
\end{align}
where $\widebar{E}$ is the homogeneous version of $\widetilde{E}$ given by
\begin{align}
\label{eq:HomogeneousHoloEnergy}
\widebar{E}\left(c_0,\dots, c_n \right) &= h_{0}^{\vphantom{4}} c_0^4 + 2 \sum_{\mu, \nu=1}^n h_{\mu \nu}^{\vphantom{4}} c_{\mu}^{\vphantom{4}} c_{\nu}^{\vphantom{4}} c_0^2 \nonumber \\
&+ \sum_{\mu, \nu, \sigma, \tau=1}^n h_{\mu \nu \sigma \tau} c_{\mu} c_{\nu} c_{\sigma} c_{\tau} .
\end{align}
Consequently, h-RHF states exist as vanishing points of the differential
\begin{align}
\label{eq:ProjectiveDifferential}
\mathrm{d} F = \frac{\partial F}{\partial c_0} \mathrm{d} c_0 + \sum_{\mu = 1}^{n} \frac{\partial F}{\partial c_{\mu}} \mathrm{d} c_{\mu}
\end{align}
restricted to the hypersurface $\widebar{X}$.
From here, it can be shown that, including multiplicities, the number of such vanishing points (and thus the exact number of h-RHF solutions) is given by
\begin{align}
\label{eq:NumberSolutions}
N_{\mathrm{solutions}} = \frac{1}{2} \left( 3^n - 1 \right).
\end{align}
A rigorous proof of this relationship is mathematically involved and beyond the scope of the current communication, but will form the focus of a future publication.
Instead, here we present a more intuitive derivation.
Consider the case of one basis function, $n=1$; clearly there are two trivial solutions at $(-1)$ and $(1)$ in the affine space $\mathbb{A}^1$.
Both points give the same density matrix and therefore describe equivalent h-RHF states.
This overall sign symmetry arises for all h-RHF states and is henceforth implicit.
\begin{figure*}[htb!]
\input{figures/OrbitalPlots.tex}
\caption{
Constrained projective h-RHF energy $F_{\bar{X}} $ plotted on the sphere $c_1^2 + c_2^2 + c_3^2 = 1$ and the plane $c_3=1$ for linear \ce{H3^+} (STO-3G) at a bond length of $2.5$\r{A}.
Nine stationary points can be located on the plane $c_3 = 1$, as shown alongside their corresponding orbital plots.
The remaining four h-RHF states exist at the infinities of this plane, and can be located by finding the stationary points constrained to the plane $c_3=0$.
The bond length has been chosen such that all solutions and their energies are real, although the results extend to geometries where complex h-RHF states are present.
}
\label{fig:n_3_solutions}
\end{figure*}
Now consider the $n=2$ case, represented by a point $(c_0 , c_1 , c_2)$ in projective space $\mathbb{P}^2$.
The projective h-RHF energy $F$ is then given by
\begin{align}
F \left( c_0, c_1, c_2 \right) &= h_{0} + 2 \sum_{\mu, \nu=1}^2 h_{\mu \nu} \frac{c_{\mu} c_{\nu}}{c_0^2} \nonumber \\
&+ \sum_{\mu, \nu, \sigma, \tau =1}^2 h_{\mu \nu \sigma \tau} \frac{c_{\mu} c_{\nu} c_{\sigma} c_{\tau}}{c_0^4}.
\end{align}
Restriction to the hypersurface $\widebar{X}$, in this case given by $c_0^2 = c_1^2 + c_2^2$, causes $F$ to become equivalent to the normalised h-RHF energy where $c_0$ provides the normalisation factor.
Consequently, the constrained function $F_{\widebar{X}}$ is invariant to a global rescaling of the orbital coefficients $c_1$ and $c_2$ and the partial derivative $\frac{\partial F_{\widebar{X}}}{\partial c_0}$ is zero for all $c_0 \neq 0$, as illustrated for \ce{H2} (STO-3G) in Figure \ref{fig:n_2_solutions}.
Although it is possible for solutions to exist with $c_0 = 0$, we find these arise only when electron-electron interactions vanish completely.
We do not expect this to occur in real molecular systems, and therefore continue our intuitive derivation under the assumption that $c_0 \neq 0$ for all stationary points.
First consider the case $c_2 \neq 0$.
Exploiting the invariance of $F_{\widebar{X}}$ to $c_0$ allows the h-RHF solutions to be located as stationary points along either the circle $c_1^2 + c_2^2 = 1$ (black curve in Figure \ref{fig:n_2_solutions}) or the line $c_2 = 1$ (blue line in Figure \ref{fig:n_2_solutions}).
Taking the latter approach enforces $\mathrm{d} c_2 = 0$ and, when combined with $ \frac{\partial F_{\widebar{X}}}{\partial c_0} = 0$, the constrained differential becomes
\begin{align}
\label{eq:ProjetiveDifferential_n=2}
\mathrm{d} F_{\widebar{X}} =\left. \frac{\partial F_{\widebar{X}}}{\partial c_1}\right|_{c_2 = 1} \mathrm{d} c_1.
\end{align}
Since $F$ is a fourth degree polynomial in $c_1$, the partial derivative $\left. \frac{\partial F_{\widebar{X}}}{\partial c_1}\right|_{c_2 = 1}$ is third degree in $c_1$ and has three roots, each defining an h-RHF state.
Next we consider the case $c_2 = 0$, recovering the $n=1$ system and yielding one further solution in the affine space $\mathbb{A}^2$ at $(1,0)$.
The total number of solutions for two basis functions is therefore $ 3 + 1 = 4 $.
We continue by adding a third basis function, represented in $\mathbb{P}^3$ by the point $(c_0 , c_1 , c_2 , c_3)$, and rotate the orbital coefficients such that $\frac{\partial F_{\widebar{X}}}{\partial c_3} = 0$ wherever $c_3=0$.
First consider $c_3 \neq 0$.
Similarly to $n=2$, the h-RHF states can be located as stationary points on either the sphere $c_1^2 + c_2^2 + c_3^2 = 1$ or the plane $c_3 = 1$, as shown for \ce{H3^+} (STO-3G) in Figure \ref{fig:n_3_solutions}.
By considering the stationary points on the plane $c_3 = 1$, the constrained differential $\mathrm{d} F_{\widebar{X}}$ reduces to
\begin{align}
\label{eq:ProjetiveDifferential_n=3}
\mathrm{d} F_{\widebar{X}} = \left. \frac{\partial F_{\widebar{X}}}{\partial c_1}\right|_{c_3 = 1} \mathrm{d} c_1 + \left. \frac{\partial F_{\widebar{X}}}{\partial c_2}\right|_{c_3 = 1} \mathrm{d} c_2.
\end{align}
The required solutions are now located by finding the common intersections of the third degree homogeneous polynomials
\begin{align}
\left. \frac{\partial F_{\widebar{X}}}{\partial c_1}\right|_{c_3 = 1} = 0 \hspace{1em} \mathrm{and} \hspace{1em}
\left. \frac{\partial F_{\widebar{X}}}{\partial c_2}\right|_{c_3 = 1} = 0.
\label{eq:Intersections_n=3}
\end{align}
B\'{e}zout's Theorem states that the number of common intersections of $n$ homogeneous polynomials in the projective space $\mathbb{P}^{n}$ is given by the product of the degrees of each polynomial.\cite{Hartshorne}
Consequently, the number of solutions to (\ref{eq:Intersections_n=3}) is given by $3 \times 3 = 9$, yielding nine h-RHF states with $c_3 \neq 0$.
We continue by considering the case where $c_3 = 0$ and recover a system of two basis functions analogous to Figure \ref{fig:n_2_solutions}.
This regime yields a further $3 + 1 = 4$ solutions, and thus the total number of h-RHF states for $n=3$ is $9 + 3 + 1 = 13$.
We can iteratively extend this argument to a general two--electron system with $n$ basis functions and find the number of solutions is given by
\begin{align}
\label{eq:NumberSolutionsIntuitive}
N_{\mathrm{solutions}} &= \sum_{i=0}^{n-1} 3^{i}.
\end{align}
Expressing this geometric series in a closed form then recovers Equation \ref{eq:NumberSolutions}.
Crucially, both this intuitive derivation and the more rigorous proof are independent of the nuclear repulsion, one- and two-electron integrals.
Therefore, the number of h-RHF solutions depends only on the number of basis functions and every solution must be conserved as the geometry or atomic charges of a system are varied.
It is important to note that Equation \ref{eq:NumberSolutions} may include solutions with a multiplicity greater than one, for example exactly when states coalesce at the Coulson--Fischer point.
Alternatively, it is possible for continuous lines or planes of solutions in the orbital coefficient space to exist.
We believe this will occur for systems with degenerate basis functions, for example molecules with cylindrical symmetry, however an infinite number of solutions can be avoided by forcing the single-particle orbitals to transform as an irreducible representation of the molecular point group.
We also note that Equation \ref{eq:NumberSolutions} has previously been identified by Stanton as an upper bound on the number of real closed-shell Hartree--Fock solutions for two-electron systems.\cite{Stanton1968}
Stanton arrived at this result geometrically for the $n=2$ case, but was restricted to considering the $n\geq 3$ case in the zero differential overlap limit, where
\begin{align}
\label{eq:ZeroDiffOvA}
h_{\mu \nu} = \delta_{\mu \nu} h_{\mu \mu}
\end{align}
and
\begin{align}
\label{eq:ZeroDiffOvB}
(\mu \nu | \sigma \tau ) = \delta_{\mu \nu} \delta_{\sigma \tau} (\mu \mu | \sigma \sigma ).
\end{align}
In contrast, employing the algebraic geometry approach presented above yields an entirely generalised geometric proof.
Furthermore, our approach proves that Equation \ref{eq:NumberSolutions} provides not only an upper bound on the number of real RHF solutions, but also the \textit{exact} number of h-RHF states for two--electron systems.
We believe that using algebraic geometry will subsequently enable the number of holomorphic solutions to be computed for both unrestricted or many electron systems, however there may be challenges in obtaining a general closed formula for these cases.
Finally, we note that the number of h-RHF states predicted by Equation \ref{eq:NumberSolutions} is much larger than the dimension of the full configuration interaction (FCI) space for two-electrons, given by $n^2$.
However, the non-orthogonality of different SCF solutions allows each state to span multiple FCI determinants, enabling a more compact description of the Hilbert space through a small number of relevant h-RHF states.
\section{Closed-shell states of \ce{HZ} in STO-3G}
\label{Sect:HZ}
\begin{figure*}[!htb]
\includegraphics[scale=1.0, trim={0.7cm 0.1cm 0.8cm 0.2cm}, clip]{figures/HZ_r_vary}
\caption{Four h-RHF states for \ce{HZ} (STO--3G) are located for all bond lengths and charges $Q_{\mathrm{Z}}$.
Each h-RHF state corresponds to a real RHF state (red/solid) where such states exist.
h-RHF states with complex orbital coefficients (magenta/dashed) form in complex--conjugate pairs with degenerate standard Hartree--Fock energies.
In the case of $Q_{\mathrm{Z}}=2.00 \ \mathrm{a.u.}$, corresponding to \ce{HHe^+} a pair of degenerate states exist with complex coefficients across all geometries.}
\label{fig:HZ_plotVsR}
\end{figure*}
Since the number of h-RHF states for two-electron systems depends on only the number of basis functions, any pair of distinct two-electron systems with the same number of basis functions can be smoothly interconverted by either moving the basis function centres (eg. changing structure) or adjusting the atomic charges (eg. changing atoms).
This concept can be demonstrated by considering the h-RHF solutions of the \ce{HZ} molecule using the STO-3G basis set.
Varying the nuclear charge of the hydrogenic \ce{Z} atom, $Q_{\mathrm{Z}}$, between $0$ and $2$, enables the smooth interconversion along the isoelectronic sequence \ce{H^-} $\rightarrow$ \ce{H2} $\rightarrow$ \ce{HHe^+}.\cite{King1969}
This simple system is of particular interest as an archetypal model for the qualitative nature of the h-RHF states in symmetric and asymmetric diatomics.
The sole occupied spatial orbital is expressed in terms of the RHF rotation angle $\theta$ describing the degree of mixing between the $1\mathrm{s}$ atomic orbitals on H and Z,
\begin{align}
\phi \left( \mathbf{r} \right) &= \sin \left( \theta - \frac{\pi}{4} \right) \chi_{\mathrm{1s,H}} \left( \mathbf{r} \right) \nonumber \\
&+ \cos \left( \theta - \frac{\pi}{4} \right) \chi_{\mathrm{1s,Z}} \left( \mathbf{r} \right).
\label{eq:RotatingBasisFunctions}
\end{align}
With two basis functions, Equation \ref{eq:NumberSolutions} dictates that four h-RHF states exist for all bond lengths $R_{\mathrm{HZ}}$ and values of $Q_{\mathrm{Z}}$.
We begin by considering the case where $Q_{\mathrm{Z}} = 1.00 \ \mathrm{a.u.}$, corresponding to \ce{H2}, and plot the conventional Hartree--Fock energy of each h-RHF solution from $R_{\mathrm{HZ}}=0.5$ \r{A} to $R_{\mathrm{HZ}}=4.0$ \r{A} in Figure \ref{fig:HZ_plotVsR} (left panel).
In the dissociation limit, each solution has real orbital coefficients and corresponds to a real Hartree--Fock state (red/solid), representing the $\upsigma_\mathrm{g}^2$, $\upsigma_{\mathrm{u}}^2$ and degenerate ionic \ce{H+-Z-} and \ce{H^{-}-Z+} states in order of ascending energy.
As the internuclear distance is reduced, the ionic states coalesce with the $\upsigma_{\mathrm{u}}^2$ state at a Hartree--Fock instability threshold and disappear at shorter bond lengths.
In contrast, the corresponding h-RHF solutions continue to exist with complex orbital coefficients (magenta/dashed), forming a degenerate pair related by complex conjugation.
Significantly, although the conventional Hartree--Fock energy of these states appears kinked at the coalescence point, their path through orbital coefficient space is both smooth and continuous, and it is this property that is essential for NOCI.
Using the classification of Hartree--Fock singlet instability thresholds developed by Mestechkin\cite{Mestechkin1978,Mestechkin1979,Mestechkin1988}, the coalescence point for $Q_{\mathrm{z}} = 1.00\ \mathrm{a.u.}$ can be identified as a ``confluence'' point, where two maxima converge onto a minimum, as shown in Figure \ref{fig:HZ_surface_z1-00}.
\begin{figure*}[!bt]
\centering
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=1, trim={0.7cm 0.cm 1.5cm 1.1cm}, clip]{figures/HZ_surface_z1-00}
\caption{}
\label{fig:HZ_surface_z1-00}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\centering
\includegraphics[scale=1., trim={0.7cm 0.cm 1.5cm 1.1cm}, clip]{figures/HZ_surface_z1-15}
\caption{}
\label{fig:HZ_surface_z1-15}
\end{subfigure}
\caption{Conventional Hartree--Fock energy plotted as a function of the RHF rotation angle, $\theta$ (Equation \ref{eq:RotatingBasisFunctions}) for $Q_{\mathrm{Z}} = 1.00\ \mathrm{a.u.}$ and $Q_{\mathrm{Z}} = 1.15\ \mathrm{a.u.}$.
\subref{fig:HZ_surface_z1-00} When $Q_{\mathrm{Z}} = 1.00\ \mathrm{a.u.}$ the molecule possesses $\mathcal{D}_{\infty h}$ symmetry and the ionic solutions simultaneously converge with the $\upsigma_{\mathrm{u}}^2$ state, disappearing at a triply degenerate $\mathrm{A_{3}}$ cusp catastrophe in a pitchfork bifurcation.
\subref{fig:HZ_surface_z1-15} For $Q_{\mathrm{Z}} = 1.15\ \mathrm{a.u.}$ the molecular symmetry becomes $\mathcal{C}_{\infty v}$, decomposing the pitchfork bifurcation into a primary branch and two secondary modes that coalesce and disappear at a doubly degenerate $\mathrm{A_{2}}$ fold catastrophe.}
\label{fig:HZ_surfaces}
\end{figure*}
In contrast, the molecular symmetry is broken by moving to $Q_{\mathrm{Z}} = 1.15\ \mathrm{a.u.}$ (middle panel of Figure \ref{fig:HZ_plotVsR}), lifting the degeneracy of the ionic states and leading to the coalescence of only the $\upsigma_{\mathrm{u}}^2$ and \ce{H+ - Z^-} states at a ``pair annihilation'' point.
Beyond this point, both real RHF solutions disappear whilst, again, their h-RHF counterparts continue as a complex degenerate pair.
The existence of complex h-RHF states arising at this pair annhilation point indicates the applicability of holomorphic Hartree--Fock for vanishing states in asymmetric diatomics including \ce{LiF}\cite{Thom2009}.
As $Q_{\mathrm{Z}}$ is increased further, the coalescence point occurs at increasing bond lengths until eventually only two real Hartree--Fock solutions exist across all geometries, as demonstrated for $Q_{\mathrm{Z}} = 2.00\ \mathrm{a.u.}$ (right panel of Figure \ref{fig:HZ_plotVsR}).
The remaining two h-RHF solutions form a degenerate pair with complex orbital coefficients across all geometries.
Although no electronic state appears to correspond to these complex solutions, they can be smoothly evolved into real states with physical significance by varying $Q_{\mathrm{Z}}$, as shown in Figure \ref{fig:HZ_charge_250}.
Consequently, we believe these states should be considered as `dormant' analytic continuations of real states.
\begin{figure}[!hb]
\center
\includegraphics[scale=0.95, trim={0.5cm 0.2cm 0.8cm 0.5cm},clip]{figures/HZ_z_vary}
\caption{
Four h-RHF solutions of \ce{HZ} (STO--3G) are plotted against the nuclear charge $Q_{\mathrm{Z}}$ for a bond length of $2.50$\r{A}, showing the smooth interconversion between the complex h-RHF states of \ce{HHe^+} (at $Q_{\mathrm{Z}}=2.00\ \mathrm{a.u.}$) and \ce{H2} (at $Q_{\mathrm{Z}}=1.00\ \mathrm{a.u.}$).
Two h-RHF states of \ce{HHe^+} that exist with complex coefficients for all bond lengths are seen to smoothly interconvert with real h-RHF states of \ce{H2} as $Q_{\mathrm{Z}}$ is varied.
}
\label{fig:HZ_charge_250}
\end{figure}
The field of catastrophe theory allows the nature of real RHF coalescence points in \ce{HZ} to be comprehensively understood as both $Q_{\mathrm{Z}}$ and $R_{\mathrm{HZ}}$ are varied.
Catastrophe theory provides a framework for qualitatively investigating stationary points for potentials that depend on a certain set of system control parameters.\cite{Gilmore}
Generally, applications focus on degenerate equilibrium points where one or more higher derivatives of the potential function are zero, referred to as non-Morse critical points.
Expanding the potential at these points as a Taylor series in small perturbations of the parameters allows the degeneracy to be lifted in a process referred to as ``unfolding''.
For one-dimensional potentials, this allows any non-Morse critical point to be classified as one of only seven ``elementary catastrophes''\cite{Thom}.
In \ce{HZ}, the number of stationary points of the conventional Hartree--Fock energy is controlled by two physical parameters $R_{\mathrm{HZ}}$ and $Q_{\mathrm{Z}}$, and we consider the behaviour of stationary points around $R_{\mathrm{HZ}} = 1.19$ \r{A}, $Q_{\mathrm{Z}} = 1.00\ \mathrm{a.u.}$ and $\theta = 0$, corresponding to the RHF confluence point of \ce{H2}.
When $Q_{\mathrm{Z}} = 1.00\ \mathrm{a.u.}$, the confluence point is a triply degenerate non-Morse critical point and the RHF solutions disappear in a pitchfork bifurcation as shown in Figure \ref{fig:HZ_surface_z1-00}.
In contrast, for $Q_{\mathrm{Z}} \neq 1.00\ \mathrm{a.u.}$ the pair annihilation point corresponds to a doubly degenerate non-Morse critical point and the pitchfork bifurcation is broken into a primary branch, existing across all geometries, and two secondary solutions which coalesce and disappear, as shown in Figure \ref{fig:HZ_surface_z1-15}.
Simultaneously considering the stationary points as both $R_{\mathrm{HZ}}$ and $Q_{\mathrm{Z}}$ are varied reveals the related elementary catastrophe to be a triply degenerate cusp or $\mathrm{A}_3$ catastrophe.\cite{Gilmore}
In contrast, pair annihilation points correspond to doubly degenerate fold or $\mathrm{A}_2$ catastrophe.
This identification indicates fold catastrophes are significantly more widespread than cusp catastrophes in molecular systems, with the simultaneous convergence of three RHF states in \ce{H2} arising directly from the additional plane of symmetry.
Despite this, the existence of complex h-RHF solutions for each case in Figure \ref{fig:HZ_plotVsR} demonstrates that holomorphic Hartree--Fock states will always exist regardless of the molecular symmetry or the nature of the singlet instability.
\section{Isomorphism of \ce{HHeH^2+} and \ce{HHeH}}
\label{Sect:Isomorphism}
\begin{figure*}[htb!]
\centering
\begin{subfigure}{.49\textwidth}
\flushleft
\includegraphics[scale=1.0, trim={0.0cm 0.0cm 0.6cm 0.5cm},clip]{figures/H2He_comparison_A}
\caption{\ce{HHeH^{2+}}}
\label{fig:H2He_2p}
\end{subfigure}
\hspace{-1.4cm}
\begin{subfigure}{.49\textwidth}
\flushright
\includegraphics[scale=1.0, trim={1.0cm 0.0cm 0.6cm 0.5cm},clip]{figures/H2He_comparison_B}
\caption{\ce{HHeH}}
\label{fig:H2He}
\end{subfigure}
\caption{The 13 h-RHF states for \subref{fig:H2He_2p} \ce{HHeH^{2+}} and \subref{fig:H2He} \ce{HHeH} in STO-3G for a range of symmetric \ce{H-He} bond lengths.
Six solutions have complex coefficients across all bond lengths, arising in degenerate complex-conjugate pairs.
At around $0.7$ \r{A} in \ce{HHeH^2+} and around $0.9$ \r{A} in \ce{HHeH}, two pairs of complex solutions coalesce to form a four-fold degenerate set of complex h-RHF states.}
\label{fig:H2He_comparison}
\end{figure*}
We now consider the two--electron linear \ce{HHeH^2+} molecule using the STO-3G basis set.
As a system with three basis functions, Equation \ref{eq:NumberSolutions} predicts 13 h-RHF states, plotted across a range of symmetric bond lengths in Figure \ref{fig:H2He_2p}.
Similarly to \ce{H2}, \ce{HHeH^2+} possesses $\mathcal{D}_{\infty \mathrm{h}}$ symmetry and thus the disappearance of the high energy real RHF states occurs at a triply degenerate cusp $\mathrm{A}_3$ catastrophe.
Beyond this point, the corresponding h-RHF states become complex, forming a degenerate pair related by complex conjugation.
Eight further dormant solutions similar to those seen in \ce{HHe^+} exist with complex coefficients across all geometries.
Furthermore, at $R = 0.5$ \r{A} we observe the convergence of two pairs of degenerate complex h-RHF states to form a set of four degenerate complex solutions.
Mathematically, systems with two electrons or two electron holes in $n$ basis functions are isomorphic and have the same number of h-RHF states.
The \ce{HHeH^2+} and \ce{HHeH} systems in STO-3G provide one of the simplest example, as shown in Figure \ref{fig:H2He_comparison}.
In both cases there are 13 h-RHF states across the all molecular geometries including eight dormant states.
Again, in \ce{HHeH} the coalescence of two pairs of degenerate complex h-RHF states to form a set of four degenerate complex solutions can be observed at $R~=~0.7$~\r{A}.
Although the relative standard Hartree--Fock energies of the states in \ce{HHeH} are in the reverse order to those in \ce{HHeH^2+}, the qualitative behaviour of solutions at coalescence points is equivalent and arises between the same pairs of h-RHF states.
Exploiting this isomorphism allows Equation \ref{eq:NumberSolutions} to be extended to systems with $2n~-~2$ electrons.
To our knowledge, only Fukutome has previously attempted to enumerate the Hartree--Fock states for a general multiple electron system.\cite{Fukutome1971}
Fukutome expressed the Hartree--Fock problem as a density matrix equation to obtain lower and upper bounds on the number of complex Hartree--Fock solutions as $2^K$ and $2^{\left( n^2- K \right)}$, where $K = \min \left(N, n - N \right)$.
To represent closed-shell systems with two-electron holes we take $N=n-1$ and $K=1$, and thus Fukutome's result predicts lower and upper bounds of $2$ and $2^{(n^2-1)} = \frac{1}{2} \times 4^n$ respectively.
Since all real Hartree--Fock solutions are simultaneously also complex and holomorphic Hartree--Fock solutions, both Fukutome's expression and Equation \ref{eq:NumberSolutions} provide independent upper bounds on the number of real RHF states.
Consequently, our result of $\frac{1}{2} \times (3^n - 1)$ provides a significantly reduced upper bound on the number of real RHF states in these systems.
\section{Rotation of ethene}
\begin{figure*}[htb!]
\centering
\begin{subfigure}{.49\textwidth}
\flushleft
\includegraphics[scale=1.0, trim={0.2cm 0.0cm 0.0cm 0.0cm},clip]{figures/C2H4_aTwist_sto-3g}
\caption{$Q_{\mathrm{X}}=6.0$\ a.u. (symmetric ethene)}
\label{fig:C2H4}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\flushright
\includegraphics[scale=1.0, trim={1.4cm 0.0cm 0.0cm 0.0cm},clip]{figures/CZH4_aTwist_sto-3g}
\caption{$Q_{\mathrm{X}}=6.1$\ a.u. (asymmetric ethene)}
\label{fig:CZH4}
\end{subfigure}
\caption{The four h-RHF states for two electrons in the space of $\uppi$ and $\uppi^*$ orbitals of \ce{CH2XH2} (STO-3G) freezing the remaining core electrons and vitual orbitals at a \ce{C-X} bond length of $R_{\mathrm{CC}} = 1.256$ \AA.
\ce{X} is a carbon-like atom with nuclear charge $Q_{\mathrm{X}}$.
At $\phi_{\mathrm{T}} = 90^{\circ}$, every h-RHF solution corresponds to a real RHF state (red/solid).
As $\phi_{\mathrm{T}}$ moves towards $0^{\circ}$ or $180^{\circ}$, real RHF states coalesce whilst h-RHF solutions continue with complex orbital coefficients (magenta/dashed).
Breaking the molecular symmetry splits the degeneracy of the high energy ionic states, changing the coalescence points from triply degenerate $\mathrm{A_{3}}$ cusp catastrophes to doubly degenerate $\mathrm{A_{2}}$ fold catastrophes.}
\label{fig:C2H4_r125_aTwist}
\end{figure*}
Although the examples presented in Sections \ref{Sect:HZ} and \ref{Sect:Isomorphism} provide insightful models for understanding the emergence of h-RHF solutions, we are not restricted to molecular systems containing only two electrons.
The properties and reactivity of many molecules are dominated by a subset of only two electrons which, by freezing the remaining core electrons, can also be considered as two-electron problems.
The electronic energy levels in the rotation of ethene, for example, depend strongly on the two-electron, two-centre $\uppi$ bond.
Starting with an orthogonal basis set composed of the STO-3G ground state RHF molecular orbitals at the optimised planar $\mathcal{D}_{\mathrm{2h}}$ geometry, we select the $\mathrm{b_{3u}}$ ($\uppi$) and $\mathrm{b_{2g}}$ ($\uppi^*$) orbitals as an active pair and freeze the remaining core electrons and virtual orbitals.
An h-RHF calculation using the $\uppi$ electrons in this active space reduces the system to a two-electron problem in two basis functions, yielding 4 solutions through Equation \ref{eq:NumberSolutions}.
Due to the symmetry equivalence of the carbon centres, the h-RHF states resemble those of \ce{H2}, corresponding at dissociation to the $\left( \uppi \right)^2$ and $\left( \uppi^* \right)^2$ configurations and the degenerate symmetry broken \ce{H2C^{+}-C^{-}H2} and \ce{H2C^{-}-C^{+}H2} ionic states.
As the carbon-carbon bond length $R_{\mathrm{CC}}$ is shortened, the ionic states coalesce with the $\left( \uppi^* \right)^2$ state at around $R_\mathrm{CC} = 1.29$ \AA\ in a triply degenerate $\mathrm{A_3}$ cusp catastrophe analogous to Figure \ref{fig:HZ_surface_z1-00}.
\begin{figure*}[!tb]
\input{figures/LipsBifurcation.tex}
\caption{Sketch of the critical manifold (left) showing the types of coalescence points between real RHF solutions in ethene and their dependence on the molecular control parameters $R_{\mathrm{CC}}$, $\phi_{\mathrm{T}}$ and $Q_{\mathrm{X}}$.
Sections through the critical manifold (right) demonstrate the dependence of these coalescence points on $\phi_{\mathrm{T}}$ and $Q_{\mathrm{X}}$ at various values of $R_{\mathrm{CC}}$.
Doubly and triply degenerate coalescence points correspond to $\mathrm{A_2}$ fold and $\mathrm{A_3}$ cusp catastrophes respectively.
Two cusp catastrophes emerge from an $\mathrm{A_3^+}$ cusp creation catastrophe and recombine at an $\mathrm{A_3^-}$ cusp annihilation catastrophe.
Within the conoidal structure (shaded) there exist four real RHF states whilst outside there are only two.}
\label{fig:C2H4_critical_manifold}
\end{figure*}
The evolution of the h-RHF states as the torsion angle $\phi_{\mathrm{T}}$ varies for $R_{\mathrm{CC}}=1.256$~\r{A} is shown in Figure \ref{fig:C2H4}.
As $\phi_{\mathrm{T}}$ increases or decreases from the $90^{\circ}$ perpendicular structure ($\mathcal{D}_{\mathrm{2d}}$) towards the planar geometry, the ionic states simultaneously coalesce with the anti-bonding $\left( \uppi^* \right)^2$ state at two $\mathrm{A_3}$ cusp catastrophes located at around $\phi_{\mathrm{T}}=45^{\circ}$ and $135^{\circ}$.
This mirrors a previous analysis by Fukutome.\cite{Fukutome1973}
Beyond these singlet instability points, the related h-RHF states continue to exist with complex orbital coefficients.
Similarly to \ce{HZ}, breaking of the molecular symmetry can be modelled by replacing one carbon with a carbon-like nucleus \ce{X} containing six electrons and a variable nuclear charge $Q_{\mathrm{X}}$.
Increasing $Q_{\mathrm{X}}$ from $6.0$ a.u. splits the degeneracy of the ionic states, leading to the coalescence of only the $\left( \uppi^* \right)^2$ and \ce{H2X^{-}-C^{+}H2} states at two doubly degenerate $\mathrm{A_2}$ fold catastrophes that shift towards $\phi_{\mathrm{T}}=90^{\circ}$, as shown for $Q_{\mathrm{X}}=6.1$~a.u. in Figure \ref{fig:CZH4}.
The critical manifold, sketched in Figure \ref{fig:C2H4_critical_manifold}, demonstrates the evolution of these coalescence points over all possible variations of $R_{\mathrm{CC}}$, $\phi_{\mathrm{T}}$ and $Q_{\mathrm{X}}$.
For very short $R_{\mathrm{CC}}$ there exist only two real RHF states for all $\phi_{\mathrm{T}}$ and $Q_{\mathrm{X}}$.
As the bond length increases, two $\mathrm{A_3}$ cusp catstrophes emerge in the plane $Q_{\mathrm{X}}=6$ a.u. from an $\mathrm{A_3^+}$ cusp creation point.\cite{Hidding2014}
These symmetry related $\mathrm{A}_3$ catastrophes are connected by two $\mathrm{A}_2$ fold catastrophes when $Q_{\mathrm{X}} \neq 6$ a.u.
Further increasing $R_{\mathrm{CC}}$ causes the $\mathrm{A_3}$ catastrophes to move away from $\phi_{\mathrm{T}}=90^{\circ}$ until they recombine at $\phi_{\mathrm{T}}=0^{\circ}$ (or the symmetry related point $\phi_{\mathrm{T}}=180^{\circ}$) at an $\mathrm{A_{3}^{-}}$ cusp annhilation point\cite{Hidding2014}.
At larger bond lengths there are no coalescence points for $Q_{\mathrm{X}}=6$ a.u. whilst the two $\mathrm{A}_2$ fold catastrophes remain when $Q_{\mathrm{X}} \neq 6$ a.u.
\begin{figure*}[!htb]
\includegraphics[scale=1.0, trim={0.2cm 0.5cm 0.0cm 1.0cm},clip]{figures/C2H4_sto-3g_r130_NOCI}
\caption{Lowest four h-UHF (blue/cyan) and four h-RHF (red/magenta) states located using two electrons in the $\uppi$ and $\uppi^*$ orbitals of ethene (STO-3G) with a frozen core and virtual approximation.
At this bond length, the h-UHF states have real orbital coefficients (blue) across all torsion angles $\phi_{\mathrm{T}}$.
The \ce{C-H} and \ce{C-C} bond lengths and angles are fixed at their optimal values and the virtual orbitals relaxed as $\phi_{\mathrm{T}}$ varies between $0^{\circ}$ and $180^{\circ}$.
Using these solutions as a basis for NOCI recovers the $\mathrm{^{1}B_{1}}$, $\mathrm{^{3}A_{2}}$ and $\mathrm{^{1}B_{2}}$ states (green), predicting the crossing of the lowest energy singlet and triplet surfaces in agreement with the exact FCIQMC results for the STO-3G basis set.}
\label{fig:C2H4_r130_NOCI}
\end{figure*}
We next use these states as a basis for NOCI to investigate the multireference energy levels as $\phi_{\mathrm{T}}$ varies, whilst retaining the STO-3G optimised bond lengths and angles.
Increasing $\phi_{\mathrm{T}}$ from $0^{\circ}$ to $90^{\circ}$ causes the energies of the bonding $\uppi$ and anti-bonding $\uppi^*$ orbitals to converge, forming a degenerate $\mathrm{e}$ molecular orbital pair.
Consequently, the perpendicular $\mathcal{D}_{\mathrm{2d}}$ structure consists of the nearly degenerate $\mathrm{^{1}B_{1}}$ and $\mathrm{^{3}A_{2}}$ states accompanied by the higher energy $\mathrm{^{1}B_2}$ and $\mathrm{^{1}A_1}$ states.
These correlate respectively to the $\mathrm{^{1}A_{g}}$, $\mathrm{^{3}B_{1u}}$, $\mathrm{^{1}B_{1u}}$, and $\mathrm{^{1}A_{g}}$ states at the planar $\mathcal{D}_{\mathrm{2h}}$ geometry.
The correct ordering of the $\mathrm{^{1}B_{1}}$ and $\mathrm{^{3}A_{2}}$ states has long been a subject of particular interest, with general valence bond\cite{Voter1985} and multiconfigurational SCF\cite{Brooks1979, Schmidt1987, Benassi2000, Oyedepo2010} calculations both indicating the $\mathrm{^{1}B_{1}}$ state lies below the $\mathrm{^{3}A_{2}}$ state in a rare violation of Hund's rules.\cite{Walsh1953, Merer1969}
To capture the triplet states using NOCI, spin contaminated h-UHF solutions must be included in the basis set.
For this particular case where $n=2$, our formal mathematical treatment indicates a further four h-UHF states exist across all geometries.
Retaining the frozen core and virtual orbital approximations, these additional h-UHF solutions have real orbital coefficients for all $\phi_{\mathrm{T}}$, corresponding to the diradical states and the $\left( \uppi \right)^1 \left( \uppi^* \right)^1$ configurations.
Unfreezing the virtual orbitals and relaxing the SCF states then allows the inclusion of hyperconjugation with the C-H $\upsigma^*$ orbitals.
Using these eight solutions, the three lowest NOCI energy levels are computed as shown in Figure \ref{fig:C2H4_r130_NOCI}.
With this basis set and carbon-carbon bond length only the ionic h-RHF states become complex, however including these states is essential to prevent discontinuities in the singlet NOCI energy levels.
The NOCI results presented in Figure \ref{fig:C2H4_r130_NOCI} indicate that the $\mathrm{^{3}A_{2}}$ state lies below the $\mathrm{^{1}B_{1}}$ state at the $90^{\circ}$ transition structure, as predicted by Hund's rules.\cite{Walsh1953, Merer1969}
However, Schmidt \textit{et al.} note that such results can arise when only the two $\uppi$ electrons are correlated\cite{Schmidt1987} --- for example in the two-configuration SCF calculations of Yamaguchi \textit{et al.}\cite{Yamaguchi1983} --- whilst the correct ordering requires correlation with the core electrons to be included.
To verify our NOCI results, we compute the exact, fully correlated energies of the $\mathrm{^{1}B_{1}}$ and $\mathrm{^{3}A_{2}}$ states within the STO-3G basis set using Full Configuration Interaction Quantum Monte--Carlo (FCIQMC)\cite{Booth2009} and obtain energies of $-76.98253(5)$ $\mathrm{E_h}$ and $-76.98833(5)$ $\mathrm{E_h}$ respectively, confirming the ordering predicted by NOCI.
Further comparison with the FCIQMC results indicates that, despite only including 8 out of $1.1 \times 10^7$ determinants from the full Hilbert space, NOCI captures 93\% and 92\% of the $\mathrm{^{1}B_{1}}$ and $\mathrm{^{3}A_{2}}$ correlation energies
Although these NOCI energy levels suggest the $\mathrm{^{1}B_{1}}$ and $\mathrm{^{3}A_{2}}$ states do cross in the rotation of ethene, it is important to remember that this is a minimal basis set calculation ignoring any geometrical relaxation for the triplet $\mathrm{^{3}A_{2}}$ state or the non-planar structures.
Regardless, it is reassuring to observe the qualitative accuracy of NOCI within the STO-3G basis set approximation using a minimal number of determinants and a frozen core approximation.
Furthermore, the occurrence of complex h-RHF solutions as the molecular control parameters vary highlights the important role of holomorphic Hartree--Fock theory if NOCI is to be applied over all ranges of molecular geometries and compositions.
\section{Computational details}
Calculations to locate h-RHF and h-UHF solutions were performed using a holomorphic analogue to the Geometric Direct Minimisation\cite{VanVoorhis2002} method implemented with processing from \texttt{SciPy}.\cite{SciPy}
FCIQMC energies were obtained using the \texttt{HANDE 1.1}\cite{Spencer2015} stochastic quantum chemistry package.
All one- and two-electron integrals were computed in \texttt{Q-Chem 4.3}\cite{QChem4-0} whilst all figures were plotted using \texttt{Matplotlib}\cite{Matplotlib}.
\section{Conclusions}
In this work we have highlighted the properties and behaviour of h-RHF solutions for two-electron problems.
By formulating the h-RHF problem in the framework of algebraic geometry, the exact number of h-RHF states (counted with multiplicity) has been identified as $\frac{1}{2} \left(3^n - 1\right)$, where $n$ is the number of basis functions.
Consequently, h-RHF states exist for all geometries and atomic charges, and always provide a continuous basis for NOCI.
Furthermore, this expression provides an upper bound on the number of real RHF states, rigorously proving the result obtained by Stanton.\cite{Stanton1968}
We believe that algebraic geometry will also yield a generalised result for unrestricted or multiple electron systems, although it may be challenging to obtain a closed formula for these cases.
Through an in-depth study of \ce{HZ}, \ce{HHeH^2+}, \ce{HHeH} and ethene we have demonstrated the behaviour of h-RHF states as molecular geometry or atomic charges are changed.
For \ce{HZ} and ethene, the presence of molecular symmetry determines whether real RHF states coalesce at a triply degenerate confluence or a doubly degenerate pair annihilation point, although complex holomorphic states emerge in both cases.
By applying the generalised framework of catastrophe theory, we have illustrated the influence of molecular control parameters including geometry and atomic compositions on the type of these coalescence points.
Moreover, we have identified dormant h-RHF states with complex orbital coefficients across all geometries.
These states are not observed in standard Hartree--Fock but can be smoothly evolved into real RHF states by changing geometry or atomic charges and represent analytic continuations of the corresponding real RHF states.
Further investigating the h-RHF states of \ce{HHeH^2+} and \ce{HHeH} in STO-3G demonstrates the isomorphism between systems with two electrons and systems with two electron holes.
Exploiting this isomorphism allows the number of h-RHF states to be identified for both types of system.
Comparing to the upper bound of $\frac{1}{2} \times 4^n$ real RHF states obtained by Fukutome\cite{Fukutome1971} indicates that the number of h-RHF states provides a new reduced upper bound for systems with two electron holes.
Finally, by considering the $\uppi$ electrons in ethene as a two-electron problem, we have used the four h-RHF states and four h-UHF as a basis for NOCI to identify a crossing of the lowest energy singlet and triplet states at a torsion angle of $90^{\circ}$.
Comparing with the exact STO-3G energies computed using FCIQMC then verifies this result within the basis set approximation, demonstrating the potential of combining holomorphic Hartree--Fock theory and NOCI.
Ultimately, the understanding on the nature of h-RHF solutions developed in this study provides a stronger platform for exploiting holomorphic states as a basis for NOCI, whilst also providing insight into the nature of Hartree--Fock states in general.
H.G.A.B. thanks the Cambridge and Commonwealth Trust for a Vice--Chancellor's Award Scholarship and A.J.W.T. thanks the Royal Society for a University Research Fellowship (UF110161).
We also acknowledge Dr. James Farrell for insightful discussions and assistance.
|
\section{Introduction
Given a definite integral depending on several parameters, a
\emph{Landen transformation} is a map on these parameters that
leaves invariant the integral. In \cite{BM0,BM}, G.~Boros and
V.~Moll introduced the dynamical system given by
$$
\left\{ \begin{array}{l}
a_{n+1}=\dfrac{5a_n+5b_n+a_nb_n+9}{(a_n+b_n+2)^{4/3}},\quad
b_{n+1}=\dfrac{a_n+b_n+6}{(a_n+b_n+2)^{2/3}},\\
\ \\
c_{n+1}=\dfrac{d_n+e_n+c_n}{(a_n+b_n+2)^{2/3}},\quad
d_{n+1}=\dfrac{(b_n+3)c_n+(a_n+3)e_n+2d_n}{a_n+b_n+2},\quad
e_{n+1}=\dfrac{c_n+e_n}{(a_n+b_n+2)^{1/3}},\\
\end{array} \right.
$$
as a Landen transformation associated to the integral
\begin{equation}\label{eq6}
I(a,b,c,d,e)=\int_{0}^{\infty} {\frac{cx^4+dx^2+e}
{x^6+ax^4+bx^2+1}\mathrm{d}x},
\end{equation}
that is,
$I(a_{n+1},b_{n+1},c_{n+1},d_{n+1},e_{n+1})=I(a_{n},b_{n},c_{n},d_{n},e_{n})$.
This dynamical system contains a 2-dimensional uncoupled subsystem. M.~Chamberland and
V.~Moll in \cite{chamb}, related the convergence of the integral~\eqref{eq6} with the dynamics
given by the iteration of the planar, non invertible map
associated to it:
$$
G(a,b):=\left(\frac{5a+5b+ab+9}{(a+b+2)^{4/3}},
\frac{a+b+6}{(a+b+2)^{2/3}}\right).
$$
In particular they proved that the map $G$ has only three fixed points,
characterizing their nature, and they also proved that the region of the $(a,b)$-plane where the
integral \eqref{eq6} converges is the basin of attraction of the fixed point $(3,3)$.
In Section \ref{S-Landen} we will give a brief description of the known results about
the dynamics of the map $G$. In Section
\ref{S-fixos-i-periodics} we prove our main result,
\begin{teo}\label{T-main} Consider the map $G$.
\begin{enumerate}
\item[(a)] It has exactly three fixed points. A super-attracting point in
$(3,3)$, an oscillatory saddle in the boundary of
the basin of attraction of $(3,3)$ and an unstable focus.
\item[(b)] It has not periodic points with minimal period $2$.
\item[(c)] It has exactly twelve periodic points of minimal period\, $3$, that correspond with four $3$-periodic orbits.
\end{enumerate}
\end{teo}
For completeness we include in its statement the results about fixed
points already proved in~\cite{chamb}. In fact, in that paper it
is also proved that there are no periodic points with minimal
period~$2$ above the line $a+b+2=0$. Our statement $(b)$ extends
their result to the whole plane.
As we will see, item $(c)$
disproves a conjecture about the dynamics of this map, see \cite[Conj. 15.6.3]{llibre-Moll} or Section
\ref{S-Landen}. We will also determine analytically the location of the $3$-periodic orbits.
Although it is easy to find $3$-periodic points numerically, when
trying to prove their existence there appear important computational
obstacles. Thus, to prove the existence of $3$-periodic points of
$G,$ as well as the non-existence of $2$-periodic points, we have
developed a procedure to determine analytically the number of
isolated periodic points of discrete dynamical systems of algebraic
nature and locate them with a prescribed precision. This method
consists in the following four steps:
\begin{itemize}
\item Convert the problem into an algebraic one, characterizing
the periodic points as the solutions of a system of polynomial
equations.
\item Include these solutions into the ones of an uncoupled system
of equations given by one-variable polynomials.
\item Combine an algorithm based on the \emph{Sturm's method} for
isolating the
real roots of a one-variable polynomial with a \emph{discard procedure} for
systems of polynomial equations in order to efficiently remove those solutions
of the later system that do not correspond with the periodic
points.
\item The
application of the \emph{Poincar\'e-Miranda theorem} to prove that the non discarded
solutions are actual solutions of the first system of polynomial equations
and, in consequence, give rise to periodic points.
\end{itemize}
This procedure is explained in detail in
next section. Recall that the Poincar\'{e}-Miranda
theorem is essentially the extension of Bolzano theorem to higher
dimensions. It was stated by H.~Poincar\'e in 1883 and 1884, and proved by
himself in 1886, \cite{Poinc1,Poinc3}. In 1940, C.~Miranda
re-obtained the result as an equivalent formulation of Brouwer fixed
point theorem, \cite{Miranda}. Recent proofs are presented in
\cite{K,V}. We also recall this theorem in Section
\ref{Ss-metode-general}.
As a complement, in Section \ref{S-Num-An} we characterize the
stable set associated to the fixed point of $G$ of saddle type, and
we provide an analytic-numeric study that gives evidences of the
existence of homoclinic trajectories associated to it, as well as of
the existence of some points in the intersection of the unstable set
of this fixed point and the non-definition set of the map, which
recall that it is formed by all the preimages of the straight line
$a+b+2=0.$
\section{Determination of periodic points of discrete dynamical systems}\label{Ss-metode-general}
We consider a discrete dynamical system defined by a map
$F:\mathcal{U}\subseteq{\mathbb R}^k\rightarrow\mathcal{U}$ where
$\mathcal{U}$ is an open set. Fix $p\in{\mathbb N}$ and assume that it has
finitely many $p$-periodic points. These points are characterized by
the real solutions of the system of $k$ equations given by
$F^p=\mathrm{Id}$. Let us suppose that the solutions of the above
system are in correspondence with the ones of a new system of
$n\geq k$ non-trivial \emph{polynomial} equations given by
\begin{equation}\label{desc1}
\left\{ \begin{array}{c}
f_1(\mathbf{x})=0,\, f_2(\mathbf{x})=0,\,
\cdots,\,
f_n(\mathbf{x})=0,
\end{array} \right.
\end{equation}
where $\mathbf{x}=(x_1,\ldots,x_n)$ are not necessarily the $k$-independent variables of $F.$
Suppose also
that using some algebraic transformations, like for instance successive resultants between the given equations, we reach an
uncoupled polynomial system whose set of solutions \emph{contains} all the solutions
of system~\eqref{desc1}:
\begin{equation}\label{desc2}
\left\{ \begin{array}{c}
q_1(x_1)=0,\,
q_2(x_2)=0,\,
\cdots,\,
q_n(x_n)=0.
\end{array} \right.
\end{equation}
To clarify with an example the above situation we sketch here the
systems involved in the computation of the
$3$-periodic points of the map $G.$ The $k=2$ equations corresponding to $G^3(a,b)=(a,b)$
can be transformed into a new system of $n=3$ polynomial equations
(see system \eqref{E-d10d11d12}) in the new variables $m,n,r$ given
by \eqref{E-abcdef}. This new system plays the role of
system \eqref{desc1}, and its solutions are in correspondence with
the periodic points, by forthcoming Lemma~\ref{O-1}.
Lemma~\ref{L-lem2} will show that the solutions of system
\eqref{E-d10d11d12} are included in the set of solutions of the
uncoupled system $\{d_{17}(m)=d_{17}(n)=d_{17}(r)=0\}$, where
$d_{17}$ is a polynomial of degree 371 introduced in~\eqref{eq:d17}.
This system plays the role of system~\eqref{desc2}.
In this setting, the proposed methodology applies in the cases where
we do not know how to obtain explicitly the solutions of systems
\eqref{desc1} or \eqref{desc2} and follows the next steps:
\smallskip
\noindent \textbf{Step 1:} By using an algorithm based on the
Sturm's method (\cite[Chap. 5.6]{StB}) and for each polynomial
$q_j$, it is possible to isolate and count all its real roots by
finding intervals with preset maximum length and rational ends, each
one of them containing only one isolated root. For each
$j=1,2,\ldots,n,$ let $k_j$ be the number of real roots of $q_j,$
without counting their multiplicities, and denote by
$I_{j,m}:=[u_{j,m},v_{j,m}],$ $m=1,2,\ldots,k_j$ the found
intervals, such that each one of them contains exactly one of these
roots. Proceeding in this way we obtain that the set of solutions of
system \eqref{desc1} \emph{is contained} in the set formed by
$\prod _{j=1}^n k_j$ boxes ($n$-dimensional orthohedrons), of the form
$$\mathcal{I}_{m_1,\ldots,m_n}:=I_{1,m_1}\times I_{2,m_2}\times \cdots \times I_{n,m_n},$$ where each $m_j\in\{1,\ldots,k_j\}$, for $j=1,\ldots,n$.
\smallskip
\noindent \textbf{Step 2:} In order to detect those boxes that do
not contain any solution of system \eqref{desc1} we apply a
\emph{discard procedure} to each box $\mathcal{I}_{m_1,\ldots,m_n}.$ This procedure is
inspired in a technique used in~\cite{JD2}. To prove that a certain
polynomial
$P(\mathbf{x})$ has no zeros in a given box $\mathcal{I}_{m_1,\ldots,m_n},$ that for the sake of simplicity
we denote as $\mathcal{I},$ we proceed as follows:
\begin{itemize}
\item We numerically evaluate $P$ at the center of $\mathcal{I}$. If,
compared with the working precision, this value is far from zero, we
suspect that $P$ restricted to $\mathcal{I}$ has a given sign. According
whether this value is positive or negative we continue with one of
next two steps.
\item For trying to prove that $P(\mathbf{x})>0$ for all
$\mathbf{x}\in \mathcal{I}$, we search a $L$ such that
$0<L<P(\mathbf{x})$ on $\mathcal{I}.$ Write
$P(\mathbf{x})=\sum_\ell M_\ell(\mathbf{x})$ where $M_\ell(\mathbf{x})=a_\ell x_1^{\ell_1} x_2^{\ell_2}\cdots x_n^{\ell_n} $,
we find $\underline{M}_{\,\ell}\in \mathbb{R}$ such that
$\underline{M}_{\,\ell}<M_\ell(\mathbf{x})$ for all $\mathbf{x}\in \mathcal{I}$ (this can be done using
the formulas in forthcoming Lemma \ref{L-minmaj-nou-n}). If the following condition is satisfied:
$0<L:=\sum_\ell \underline{M}_{\,\ell}<\sum_\ell
M_\ell(\mathbf{x})=P(\mathbf{x}),$
then we can discard the box $\mathcal{I}$.
\item For trying to prove that $P(\mathbf{x})<0$ for all
$\mathbf{x}\in \mathcal{I}$, we look for $U\in{\mathbb R}$ such that
$P(\mathbf{x})<U<0$ on~$\mathcal{I}.$ To do this, similarly than in the previous situation,
we find $\overline{M}_{\,\ell}\in \mathbb{R}$ such that
$M_\ell(\mathbf{x})<\overline{M}_{\,\ell}$ for all $\mathbf{x}\in \mathcal{I}$. If
it holds that
$P(\mathbf{x})=\sum_\ell M_\ell(\mathbf{x})<\sum_\ell
\overline{M}_{\,\ell}=:U<0,$
then we can discard the box $\mathcal{I}$.
\end{itemize}
To compute the bounds $\underline{M}_{\,\ell}$ and
$\overline{M}_{\,\ell}$, we use the following straightforward result, which can be easily implemented in any computer algebra software.
\begin{lem}\label{L-minmaj-nou-n} Consider
$P(\mathbf{x})=\sum_\ell M_\ell(\mathbf{x})$ where
$M_\ell(\mathbf{x})=a_\ell x_1^{\ell_1} x_2^{\ell_2}\cdots
x_n^{\ell_n} $, and a box $ \mathcal{I}=[u_1,v_1]\times
[u_2,v_2]\times\cdots\times [u_n,v_n].$ Set
$O^{+}=\{(x_1,\ldots,x_n),$ such that $x_i>0$ for all
$i=1,\ldots,n\}$. Then,
\begin{enumerate}[(i)]
\item If $ \mathcal{I}\subset O^{+}\subset{\mathbb R}^n,$ then for all $\mathbf{x}\in
\mathcal{I},$ $\sum _\ell \underline{M}_{\,\ell}\le P(\mathbf{x}) \le \sum
_\ell
\overline{M}_{\ell},$ where
\begin{enumerate}
\item[(a)] $\underline{M}_{\,\ell}=a_\ell\,u_1^{\ell_1} u_2^{\ell_2}\cdots u_n^{\ell_n}$
and $\overline{M}_\ell=a_\ell\,v_1^{\ell_1} v_2^{\ell_2}\cdots v_n^{\ell_n}$ if $a_\ell>0$.
\item[(b)] $\underline{M}_{\,\ell}=a_\ell\,v_1^{\ell_1} v_2^{\ell_2}\cdots v_n^{\ell_n}$
and $\overline{M}_{\ell}=a_\ell\,u_1^{\ell_1} u_2^{\ell_2}\cdots u_n^{\ell_n}$ if $a_\ell<0$.
\end{enumerate}
\item If $\mathcal{I}\not\subset O^{+}$ we can always take a number $\xi>0$,
$\xi\in\mathbb{Q}$ such that the new box
$\widetilde{\mathcal{I}}=[u_1+\xi,v_1+\xi]\times
[u_2+\xi,v_2+\xi]\times\cdots\times [u_n+\xi,v_n+\xi]\subset O^{+},
$ and then find bounds for $P$ on $\mathcal{I},$ using the bounds given in
item $(i)$ for $\widetilde P_\xi(x_1,x_2,\ldots,x_n):=
P(x_1-\xi,x_2-\xi,\ldots,x_n-\xi)$ on $\widetilde{\mathcal{I}}.$
\end{enumerate}
\end{lem}
We try to apply the discard procedure until the number of remaining
boxes coincides with our hopes. These hopes usually came from a
previous numerical study of the problem. We start trying to prove
that the first function $f_1$ does not vanish in the given box. It
may happen that it is easier to try to prove the same with another
$f_j.$ Notice also that sometimes to discard a box we must go to the
Step 1 and start with smaller boxes.
\smallskip
\noindent \textbf{Step 3:} Once it is achieved an optimized list of
non-discarded boxes, we identify those boxes that correspond to either fixed points or periodic points
with a period being a divisor of $p$, which we assume that we already know, and we also discard them.
\smallskip
\noindent \textbf{Step 4:} From the non-discarded boxes list obtained in the previous step,
we try to show that each box actually contains a solution by applying the Poincar\'e-Miranda
theorem. For completeness, we recall it. As usual, $\overline B$ and
$\partial B$ denote, respectively, the closure and the boundary of a
set $B\subset{\mathbb R}^n.$
\begin{teo}[Poincar\'e-Miranda]\label{T-PM-n}
Set $\mathcal{I}=\{\mathbf{x}=(x_1,\ldots,x_n)\in{\mathbb R}^n\,:\,L_i<x_i<U_i, 1\leq i\leq n\}$. Suppose that
$f=(f_1,f_2,\ldots,f_n):\overline{\mathcal{I}}\rightarrow R^n$ is
continuous, $f(\mathbf{x})\neq\mathbf{0}$
for all $\mathbf{x}\in\partial \mathcal{I}$, and for $1\leq i\leq
n,$
$$f_i(x_1,\ldots,x_{i-1},L_i,x_{i+1},\ldots,x_n)\cdot f_i(x_1,\ldots,x_{i-1},U_i,x_{i+1},\ldots,x_n)\leq 0,$$
Then, there exists $\mathbf{s}\in\mathcal{I}$ such that $f(\mathbf{s})= \mathbf{0}$.
\end{teo}
It is clear that when we define $f$ to try to apply Poincar\'{e}-Miranda
theorem, the order of the components matters. So, sometimes to be
under the hypotheses of the theorem it is better to consider
$f=(f_{\sigma_1},f_{\sigma_2},\ldots,f_{\sigma_n})$ for some
permutation $\sigma.$ In fact, more in general, it is convenient to
apply the theorem to $A (f(\mathbf{x}))^t,$ where $A$ is a suitable
$n\times n$ invertible matrix. When $f$ is differentiable, as we
will see it is useful to chose $A=(\mathrm{D}f(\widehat
{\mathbf{s}}))^{-1},$ where $\widehat{\mathbf{s}}\in{\mathbb Q}^n$ is a
numerical approximation of a zero of $f$ in $\mathcal{I}.$
If we succeed in proving that there is at least a solution in each box,
its uniqueness is given by the fact that each of the intervals $ I_{j,m} $ contains
only a single solution of each polynomial $q_j$. Otherwise we can refine
boxes, taking them with smaller size, and then repeating the computations in Step 1.
\section{An overview of the dynamics of $G$}\label{S-Landen}
In this section we briefly summarize the known results on the
dynamics of the map $G$ and we characterize their invariant sets. We
mainly follow the steps in \cite{chamb}. The rational integral
\eqref{eq6} is well-defined and convergent if $P(x)=x^3+ax^2+bx+1$
has not real positive roots. To study the number of real roots of
$P$ when $a$ and $b$ vary, we consider
$$
R(a,b):=\mathrm{Res}(P,P';x)=-\Delta_x(P)=-{a}^{2}{b}^{2}+4\,{a}^{3}+4\,
{b}^{3}-18\,ab+27,
$$
\noindent where $\Delta_x$ is the discriminant. The curve $R(a,b)=0$
is known as the \emph{resolvent} one, and after removing the point
$(-1,-1)$ it is invariant by $G$ because
\begin{equation}\label{E-Rinv}
R(G(a,b))=\frac{(a-b)^2}{(a+b+2)^4}R(a,b).
\end{equation}
The curve has two connected components
$L_1$ i $L_2$ (see Figure 1 (a)). Note that
the fixed point $(3,3)$ is the cusp of $L_1$.
\begin{center}
\includegraphics[scale=0.3]{regionsresultant-v2.pdf} \hspace{1cm} \includegraphics[scale=0.55]{fase-v6.pdf}
(a) $\phantom{xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx} $ (b)
\smallskip
Figure 1. (a) Connected components $L_1$ and $L_2$ of the curve $R(a,b)=0$,
and regions $C_1$, $C_2$ and $C_3$ of the plane. (b) 10000 iterates of an orbit with initial condition in $C_3$.
\end{center}
The resolvent curve defines three open unbounded sets $C_1$, $C_2$
and $C_3$ depicted in Figure 1~(a). By studying the sign of the
discriminant of $P$, and by using the Descartes rule of signs, it is
straightforward to obtain that on $L_1\cup C_2\cup C_3$, all the
real roots of $P(x)$ are negative, so the integral \eqref{eq6} is
convergent; and on $C_1\cup L_2$ there exists at least one positive
real root, so the integral diverges.
In \cite{chamb}, the authors proved that $G$ has only three fixed
points, namely $P_i$ for $i=1,2,3,$ which are described in Theorem
\ref{T-main} and given in Equation \eqref{E-PF}. One of them, the
point $P_1=(3,3)$, is a super-attracting one, i.e. both eigenvalues
of the jacobian matrix are $0$. Their main result states that
\emph{the basin of attraction of the fixed point $P_1$ for the map
$G$, is the region of the $(a,b)$-plane where the
integral \eqref{eq6} converge}. As a consequence, the basin of attraction of $P_1$ is $L_1\cup C_2\cup C_3$.
On the other hand, the connected component
$L_2':=L_2\setminus\{(-1,-1)\}$ is \emph{positively} invariant (as
we will see, there are points on $C_1$ which are mapped into $L_2$).
On $L_2$ there is one fixed point, $P_2$, which is a saddle. In
Proposition \ref{P-atraccio-g} we prove that any orbit with initial
condition on $L_2'$ converges to $P_2$.
In summary, the dynamics of $G$ on the invariant sets
$\mathcal{A}:=L_1\cup C_2\cup C_3$ and $L_2$ is known and simple. However, there is a poor knowledge of
the dynamics of $G$ in the set
$\mathcal{B}:=C_1\setminus \mathcal{F},$ where $
\mathcal{F}=\{(a,b)\in{\mathbb R}^2:\,\exists\, n\geq 0\,:\, G^n(a,b)\in
\{a+b+2=0\}\}$, is the forbidden set of $G$.
In \cite[Conj. 15.6.3]{llibre-Moll}, V.~Moll established the
following conjecture about the dynamics of~$G$ in~$\mathcal{B}$:
\emph{``The orbit of any point below the resolvent curve is dense in
the open region below this curve.''}
Of course one has to exclude from this conjecture
the third fixed point $P_3$ which is in the set $C_1$, and the
points in the forbidden set $\mathcal{F}$. In Theorem
\ref{T-main}, we prove the existence of $3$-periodic points in
$C_1,$ result that disproves the conjecture. In fact it is not
difficult to find \emph{numerically} these orbits as well as other
periodic points, however to prove the existence of $3$-periodic
points is far from being trivial, and it is the main objective of
this paper.
In Section \ref{S-Num-An} we present an analytic-numeric study that evidences the existence of points in the
unstable manifold of $P_2$ which belongs to its stable
set, i.e. \emph{homoclinic points}. In the case of
diffeomorphisms, by the Smale-Birkhoff homoclinic theorem, the
existence of such points implies the existence of a hyperbolic
invariant set on which the dynamics is equivalent to a subshift of
finite type, see \cite{GuH}. Similar results are developed in \cite{Gard,GarSus11} in the non-invertible setting.
We also give evidences of the existence of points in the
unstable manifold that also belong to the forbidden set~$\mathcal{F}$.
\section{Proof of Theorem \ref{T-main}}\label{S-fixos-i-periodics}
In this section we prove Theorem \ref{T-main}. We split the proof in
two subsections. The first one dedicated to the fixed and 2-periodic
points, and the second one to study the 3-periodic points.
\subsection{Fixed and $2$-periodic points}
\begin{proof}[Proof of statements (a) and (b)] (a) Following \cite{chamb}
we consider the equations given by $G(a,b)=(a,b)$, and we introduce the auxiliary variable $m^3=a+b+2$, obtaining
\begin{equation}\label{eq:pfixos}
\begin{cases}
\begin{array}{ll}
d_1(a,b,m)&:=m^3-a-b-2=0,\\
d_2(a,b,m)&:=-am^4+ab+5a+5b+9=0,\\
d_3(a,b,m)&:=-bm^2+a+b+6=0.
\end{array}
\end{cases}
\end{equation}
Isolating $a$ and $b$ from the first and third equations, and
substituting the obtained expressions in the second one we get:
\begin{align}\label{eq:fixos}
d_4(m):=- ( m-2 ) ( {m}^{2}-m+1 ) ( {m}^{2}+m+2
) ( {m}^{3}+{m}^{2}-m-2 ) ( {m}^{3}+{m}^{2}+m +2 )=0.
\end{align}
The only real roots of the above equation
are
{\scriptsize
\begin{align}\label{eq:fixos2}
m_1=2,\quad m_2=\frac{1}{6}\,\sqrt [3]{A}+\frac{8}{3}\,{\dfrac
{1}{\sqrt [3]
{A}}}-\frac{1}{3}\simeq 1.20557,\quad
m_3=-\frac{1}{6}\,\sqrt [3]{B}+\frac{4}{3}\,{\dfrac {1}
{\sqrt [3]{B}}}-\frac{1}{3}\simeq -1.35321,
\end{align}}
where $A=172+12\,\sqrt {177}$ and $B=188+12\,\sqrt {249}$. From
these values and \eqref{eq:pfixos} we obtain
{\scriptsize \begin{align}\label{E-PF}
P_1&=(3,3), \nonumber\\
P_2&=\Bigg( {\frac
{-43+3\sqrt {177}}{384}} A^{2/3}-\frac{1}{6} A^{1/3}-\frac{8}{3},
{\frac{13-\sqrt {177}}{48}} A^{2/3}+\frac{7+\sqrt {177}}{48} A^{1/3}
+\frac{4}{3} \Bigg)\simeq(-4.20557, 3.95774), \\
P_3&=\Bigg( {\frac{-21+\sqrt {249}}{96}} B^{2/3}+\frac{15-\sqrt {249}}{12} B^{1/3} -2,{\frac{17-\sqrt {249}}{48}} B ^{2/3}
+{\frac{-13+\sqrt {249}}{24}} B^{1/3}-\frac{4}{3} \Bigg)\simeq(-5.30914,
0.83118)\nonumber.
\end{align}}
A straightforward computation of the differential matrix at these
points give that the points are, respectively, a super-attractor
(null eigenvalues), an oscillatory saddle, and an unstable focus.
Moreover $P_2$ is in $L_2$ and $P_3$ is in $C_1.$
\medskip
(b) Again, following \cite{chamb}, we consider $c$ and $d$ such that
$G(a,b)=(c,d)$ and
$G(c,d)=(a,b)$. By introducing the two auxiliary variables
$m$ and $n$ such that
$m^3=a+b+2$ and $n^3=c+d+2$, we get:
$$\begin{cases}
\begin{array}{ll}
d_1&:=m^3-a-b-2=0,\quad
d_2:=n^3-c-d-2=0,\\
d_3&:=-cm^4+ab+5a+5b+9=0,\quad
d_4:=-dm^2+a+b+6=0,\\
d_5&:=-an^4+cd+5c+5d+9=0,\quad
d_6:=-bn^2+c+d+6=0.
\end{array}
\end{cases}
$$
Solving
$\{d_1=0,d_2=0, d_4=0, d_6=0\}$ we obtain
$$
a={\dfrac {{m}^{3}{n}^{2}-{n}^{3}-2\,{n}^{2}-4}{{n}^{2}}},\quad
b=\dfrac {{n}^{3}+4}{{n}^{2}},\quad
c=\dfrac {{m}^{2}{n}^{3}-{m}^{3}-2\,{m}^{2}-4}{{m}^{2}},\quad
d=\dfrac {{m}^{3}+4}{{m}^{2}}.
$$
By substituting the above result in $d_3$ and
$d_5$ we reach the following system, which plays the role of system \eqref{desc1} in our methodology:
\begin{equation}\label{E-2pd7d8}
\begin{cases}\begin{array}{rl}
d_7(m,n):=&-{m}^{4}{n}^{7}+{m}^{5}{n}^{4}+2 {m}^{4}{n}^{4}
+{m}^{3}{n}^{5}+5 {m}^{3}{n}^{4}+4 {m}^{2}{n}^{4}-{n}^{6}\\
&+4 {m}^{3}{n}^{2}-2 {n}^{5}-{n}^{4}-8 {n}^{3}-8 {n}^{2}
-16=0,\\
d_8(m,n):=& d_7(n,m)=0.
\end{array}\end{cases}
\end{equation}
Now we consider the polynomial
\begin{align*}
d_9(m):=&\mathrm{Res}(d_{7}(m,n),d_{8}(m,n);n)\\
=&{m}^{4} \left( m-2 \right) \left( m+1 \right) ^{2}\left( {m}^{3}+{m}^{2}+m+2 \right)
\left( {m}^{3}+{m}^ {2}-m-2 \right)P(m),
\end{align*}
where $P$ is a polynomial of degree 56, without real roots. This is
proved by using the Sturm's method and can also be done, for
instance, by using the command \texttt{realroot} of the computer
algebra system Maple. Similarly,
$d_{10}(n):=\mathrm{Res}(d_{7}(m,n),d_{8}(m,n);m)$. As a consequence
of the symmetry we get
that $d_{10}(n)=-d_{9}(n)$ and
system $\{d_9(m)=0,\, d_9(n)=0\}$ plays the role of system~\eqref{desc2}
in our methodology. Hence, the only
non-zero
reals roots of $d_9$ are $-1,2, m_1$ and $m_2,$ where these values
correspond to the ones associated with the fixed points, because
the two degree~3 factors coincide with the ones given in
\eqref{eq:fixos}.
Hence the 2-periodic points are included in the set with 16 elements
$\{-1,2,m_1,m_2\}^2.$ In this particular case, because the real
solutions of the uncoupled system are explicit, in Steps 3 and 4 of
our approach we have not boxes but points, and the problem is much
easier. It is not difficult to check that in this set of points the
only solutions of \eqref{E-2pd7d8} are $(1,1), (m_1,m_1)$ and
$(m_2,m_2)$ which correspond to the fixed points of $G.$ In
consequence there are not points of minimal period $2$
for~$G$~\end{proof}
\subsection{Proof of Theorem \ref{T-main} (c): $3$-periodic points}
We
need some preliminary results. Proceeding as in the previous
cases we look for $a$, $b$, $c$, $d$, $e$, and $f\in{\mathbb R}$, such that $G(a,b)=(c,d)$, $G(c,d)=(e,f)$ and $G(e,f)=(a,b)$,
that is
\begin{align*}
&\dfrac{5a+5b+ab+9}{(a+b+2)^{4/3}}=c,
&&\dfrac{a+b+6}{(a+b+2)^{2/3}}=d,
&&\dfrac{5c+5d+cd+9}{(c+d+2)^{4/3}}=e,\\[0.1cm]
&\dfrac{c+d+6}{(c+d+2)^{2/3}}=f,
&&\dfrac{5e+5f+ef+9}{(e+f+2)^{4/3}}=a,
&&\dfrac{e+f+6}{(e+f+2)^{2/3}}=b.
\end{align*}
We introduce the auxiliary variables
$m$, $n$ and $r$, such that
$m^3=a+b+2$, $n^3=c+d+2$ and $r^3=e+f+2$.
Using this notation we get
$$\begin{cases}
\begin{array}{lll}
d_1:=m^3-a-b-2=0,& d_4:=-cm^4+ab+5a+5b+9=0,& d_5:=-dm^2+a+b+6=0,\\
d_2:=n^3-c-d-2=0,& d_6:=-en^4+cd+5c+5d+9=0,& d_7:=-fn^2+c+d+6=0,\\
d_3:=r^3-e-f-2=0,& d_8:=-ar^4+ef+5e+5f+9=0,&d_9:=-br^2+e+f+6=0.
\end{array}
\end{cases}
$$
First we solve the system
$\{d_1=0,d_2=0,d_3=0,d_5=0,d_7=0,d_9=0\}$ obtaining:
\begin{align}\label{E-abcdef}
a&=\dfrac{{m}^{3}{r}^{2}-{r}^{3}-2\,{r}^{2}-4}{{r}^{2}},\quad
b=\dfrac{{r}^{3}+4}{{r}^{2}},\quad
c=\dfrac{{m}^{2}{n}^{3}-{m}^{3}-2\,{m}^{2}-4}{{m}^{2}},\nonumber \\
d&=\dfrac{{m}^{3}+4}{{m}^{2}}, \quad
e=\dfrac{{n}^{2}{r}^{3}-{n}^{3}-2\,{n}^{2}-4}{{n}^{2}}, \quad
f=\dfrac{{n}^{3}+4}{{n}^{2}}.
\end{align}
Substituting this result in the expressions $d_4,$ $d_6,$ and $d_8$,
we obtain the equations
\begin{equation}\label{E-d10d11d12}
\begin{cases}
\begin{array}{ll}
d_{10}(m,n,r):=&-{m}^{4}{n}^{3}{r}^{4}+{m}^{5}{r}^{4}
+2\,{m}^{4}{r}^{4}+{m}^{3}{r}^{5}+5\,{m}^{3}{r}^{4}
+4\,{m}^{2}{r}^{4}-{r}^{6}+4\,{m}^{3}{r}^{2}\\
&-2\,{r}^{5}-{r}^{4}-8\,{r}^{3}-8\,{r}^{2}-16=0,\\
d_{11}(m,n,r):=&d_{10}(n,r,m)=0,\\
d_{12}(m,n,r):=&d_{10}(r,m,n)=0.
\end{array} \end{cases}
\end{equation}
From the equations \eqref{E-abcdef}, if $(m,n,r)$ is a real solution
of \eqref{E-d10d11d12} such that $m \cdot n\cdot r\neq 0$,
there exists either an orbit with minimal period
$3$ given by \eqref{E-abcdef} or a fixed point or $G$. Moreover,
if $(m_0,n_0,r_0)$ is a solution of system \eqref{E-d10d11d12}, then so are
$(n_0,r_0,m_0)$ and $(r_0,m_0,n_0)$. As a consequence, we obtain
\begin{lem}\label{O-1}
Any $3$-periodic orbit of $G$,
$\{(a,b); (c,d); (e,f)\}$ with associated parameters $m,n$ and $r$,
is in correspondence, via \eqref{E-abcdef}, with the solutions $(m,n,r)$, $(n,r,m)$ and $(r,m,n)$ of the system~\eqref{E-d10d11d12}.
\end{lem}
The forthcoming Lemma \ref{L-lem2} gives a first characterization
of the locus where the solutions of system \eqref{E-d10d11d12} are
located. Prior to state this result we introduce the following
auxiliary polynomials
\begin{align*}
d_{13}(n,r)&:=\mathrm{Res}(d_{10}(m,n,r),d_{12}(m,n,r);m),
\, \mbox{with degree $37$ in $n$ and degree $37$ in $r$,}\\
d_{14}(n,r)&:=\mathrm{Res}(d_{11}(m,n,r),d_{12}(m,n,r);m),
\, \mbox{with degree $47$ in $n$ and degree $37$ in $r$.}\
\end{align*}
We apply the resultant once again to obtain the polynomials $
d_{15}(n):=\mathrm{Res}(d_{13}(n,r),d_{14}(n,r);r),$ and
$ d_{16}(r):=\mathrm{Res}(d_{13}(n,r),d_{14}(n,r);n),
$
where $\mathrm{deg}_n(d_{15}(n))=2521$ and $\mathrm{deg}_r(d_{16}(r))=1985$.
Finally, we introduce the polynomial
\begin{equation}\label{eq:d17}
d_{17}(n):=\gcd\left(d_{15}(n),d_{16}(n)\right)/n^{716}.
\end{equation}
This polynomial has degree $371$, and using once more Sturm's method we get that
it has exactly $16$ different
real non-zero roots.
\begin{lem}\label{L-lem2}
Let $I_i,$ with $i=1,\ldots, 16,$ be disjoint intervals, each one
of them containing a unique real root of $d_{17}$. Then, any real
solution $(m,n,r)$ of system \eqref{E-d10d11d12} is contained in
one of the $16^3$ sets
\begin{equation}\label{E-caixesR3}
\mathcal{I}_{i,j,k}:=I_i\times I_j\times I_k,\quad i,j,k\in\{1,\ldots,16\}.
\end{equation}
\end{lem}
\begin{proof}
Let $(m_0,n_0,r_0)$ be a real solution of \eqref{E-d10d11d12}. We
want to show that it is also a solution of $\{d_{17}(m)=0,
d_{17}(n)=0, d_{17}(r)=0\}$. Observe that by construction, $n_0$
must be a root of $d_{15}$. From Lemma \ref{O-1}, we have that
$(m_0,n_0)$ must be also a zero of $d_{13}$ and $d_{14}$. Hence
$n_0$ must be also a zero of $d_{16}$, and therefore of
$\gcd\left(d_{15}(n),d_{16}(n)\right)$ which is a polynomial of
degree $1087$ with the factor $n^{716}$. Since we are interested in
its non-zero roots, we remove this factor, obtaining that $n_0$ must
be a root of $d_{17}$. By using an analogous argument and Lemma
\ref{O-1} again, we can see that $m_0$ and $r_0$ are also roots of
$d_{17}$.
Since $d_{17}$ has $16$ different real roots, any solution
$(m,n,r)$ of system \eqref{E-d10d11d12} must be contained in a box
of the form \eqref{E-caixesR3}, and each box contains at most one
solution.
\end{proof}
Now we can prove statement (c) of Theorem \ref{T-main}. We follow
the steps explained in Section \ref{Ss-metode-general}:
\smallskip
\noindent \textbf{Step 1:} Recall that $d_{17}$ has 16 non-zero real roots. Two of them are $n=-1$, $n=2.$
Although two more explicit roots are
$n_i=m_i, i=1,2$ given in \eqref{eq:fixos2}, we prefer to take 14 intervals with rational ends and length smaller than $10^{-20},$
$I_i, i=1,2,\ldots,14,$ each one of them containing a unique root.
We consider:
{\tiny
\begin{align*}
&I_1= \left[
-{\frac{4308988841618670568853}{147573952589676412928}},-{\frac{
34471910732949364550823}{1180591620717411303424}} \right], && I_2=
\left[ -{\frac{
34411805733101949308435}{1180591620717411303424}},-{\frac{
17205902866550974654217}{590295810358705651712}} \right],\\[0.1cm]
&I_3= \left[ -{\frac{
9138398550509024508051}{1180591620717411303424}},-{\frac{
4569199275254512254025}{590295810358705651712}} \right], &&
I_4= \left[ -{\frac{
4416518740855918762195}{590295810358705651712}},-{\frac{
8833037481711837524389}{1180591620717411303424}} \right],\\[0.1cm]
&I_5= \left[ -{\frac{
994661336537171251825}{295147905179352825856}},-{\frac{
3978645346148685007299}{1180591620717411303424}} \right], && I_6=
\left[ -{\frac{
3977374161031280580629}{1180591620717411303424}},-{\frac{
994343540257820145157}{295147905179352825856}} \right],\\[0.1cm]
&I_7= \left[-{\frac{
197879469664271669175}{73786976294838206464}},-{\frac{
3166071514628346706799}{1180591620717411303424}} \right], &&
I_8= \left[ -{\frac{
3144313156826151948503}{1180591620717411303424}},-{\frac{
1572156578413075974251}{590295810358705651712}} \right],\\[0.1cm]
&I_9= \left[ -{\frac{
399397086201257638833}{295147905179352825856}},-{\frac{
1597588344805030555331}{1180591620717411303424}} \right], &&
I_{10}= \left[-{\frac{
1053526769518098399097}{4722366482869645213696}},-{\frac{
131690846189762299887}{590295810358705651712}} \right],\\[0.1cm]
&I_{11}= \left[ -{\frac{
1064910654630154190265}{9444732965739290427392}},-{\frac{
133113831828769273783}{1180591620717411303424}} \right],&&
I_{12}= \left[ {\frac{
1065572958580542810237}{9444732965739290427392}},{\frac{
532786479290271405119}{4722366482869645213696}} \right],\\[0.1cm]
&I_{13}= \left[ {\frac{
128535594827653577343}{590295810358705651712}},{\frac{
1028284758621228618745}{4722366482869645213696}} \right],
&&I_{14}= \left[ {\frac{
177910645965499912685}{147573952589676412928}},{\frac{
1423285167723999301481}{1180591620717411303424}} \right].\\
\end{align*}
}
We also introduce the degenerate intervals $I_{15}=[-1,1]$ and
$I_{16}=[2,2]$ containing the exact roots $n=-1$ and $n=2$. By Lemma
\ref{L-lem2}, all the real solutions of system \eqref{E-d10d11d12}
are contained in one of the $16^3$ boxes \eqref{E-caixesR3}, where
we also call boxes the ones with some degenerate interval. Recall
that if a box $\mathcal{I}_{i,j,k}$ contains a solution of system
\eqref{E-d10d11d12}, then this solution is unique.
\smallskip
\noindent \textbf{Step 2:} We apply the discard procedure to
$d_{10}$ and the $4096$ boxes of the form \eqref{E-caixesR3} given
by the intervals computed before. In Lemma \ref{L-minmaj-nou-n} we
use the value $\xi=30$ and consider the polynomial
$P(m,n,r)=d_{10}(m-\xi,n-\xi,r-\xi),$ which has $224$ monomials. The
procedure implemented in Maple v.17 took 5.61s of real time in an
Intel i7-3770-3.4GHz CPU to discard $4080$ boxes. The code is given
in \cite[Chap. 5]{Llor}. In short, we obtain that each solution of
system \eqref{E-d10d11d12} must be contained in one of the following
$16$ non-discarded boxes
$$
\begin{array}{| l | l | l | l | l | l | l | l | }
\hline
\mathcal{I}_{1,5,11} & \mathcal{I}_{2,6,12} & \mathcal{I}_{3,7,13} & \mathcal{I}_{4,8,10} &
\mathcal{I}_{5,11,1} & \mathcal{I}_{6,12,2} & \mathcal{I}_{7,13,3} & \mathcal{I}_{8,10,4} \\
\hline
\mathcal{I}_{9,9,9} & \mathcal{I}_{10,4,8} & \mathcal{I}_{11,1,5} & \mathcal{I}_{12,2,6} &
\mathcal{I}_{13,3,7} & \mathcal{I}_{14,14,14} & \mathcal{I}_{16,16,15} & \mathcal{I}_{16,16,16}\\
\hline
\end{array}
$$
Observe that the degenerated box $\mathcal{I}_{16,16,15}$, which
corresponds with $(m,n,r)=(2,2,-1)$, must also be discarded because
$d_{10}(2,2,-1)=0$, but $d_{11}(2,2,-1)=2304$.
\smallskip
\noindent \textbf{Step 3:} Following similar arguments that in the
proof of statement (b) we can discard boxes
$\mathcal{I}_{16,16,16},$ $\mathcal{I}_{14,14,14}$ and
$\mathcal{I}_{9,9,9}$ because they correspond to the fixed points
$P_1,P_2$ and $P_3,$ respectively.
\smallskip
\noindent \textbf{Step 4:} We have obtained
$12$ non-discarded boxes that, from Lemma \ref{O-1}, if they correspond to periodic points of minimum period $3$,
they would contain the parameters $(m,n,r)$ corresponding to the periodic points
according to the following groupings:
\begin{equation}\label{E-supercaixes}
\begin{array}{cc}
\mathcal{O}_1\subset
\mathcal{I}_{1,5,11}\cup \mathcal{I}_{5,11,1}\cup \mathcal{I}_{11,1,5},&
\mathcal{O}_2\subset
\mathcal{I}_{2,6,12}\cup \mathcal{I}_{6,12,2}\cup \mathcal{I}_{12,2,6},\\
{}\\
\mathcal{O}_3\subset
\mathcal{I}_{3,7,13}\cup \mathcal{I}_{7,13,3}\cup \mathcal{I}_{13,3,7}, &
\mathcal{O}_4\subset \mathcal{I}_{4,8,14}\cup \mathcal{I}_{8,14,4}\cup \mathcal{I}_{14,4,8}.
\end{array}
\end{equation}
We will prove that the above $12$ boxes indeed contain a solution
of system \eqref{E-d10d11d12}, which will be unique as reasoned
above. To do this, we will apply the Poincar\'e-Miranda theorem
(Theorem \ref{T-PM-n}). Again by Lemma \ref{O-1} we only need to
prove that there is a solution of the system \eqref{E-d10d11d12} in
the boxes:
$\mathcal{I}_{1,5,11}$, $\mathcal{I}_{2,6,12}$, $\mathcal{I}_{3,7,13}$, and $\mathcal{I}_{4,8,14}$.
For reasons of space we only give details for the first box.
We consider the polynomial map
$f(m,n,r):=\left(d_{10}(m,n,r),d_{11}(m,n,r),d_{12}(m,n,r)\right).$
We denote the ends of the intervals $I_1$, $I_5$ and $I_{11}$
respectively: $ [\underline m,\overline m]:=I_1,\,[\underline n
,\overline n]:=I_5,\,[\underline r,\overline r]:=I_{11}.$ Consider
also
the middle point of $\mathcal{I}_{1,5,11},$ $ \widehat
{p}=(\widehat {m},\widehat {n},\widehat {r})=\left(\big(\underline m
+\overline m\big)/2,\big(\underline n+\overline
n\big)/2,\big(\underline r+\overline r\big)/2\right). $
\begin{center}
\begin{minipage}{0.5\textwidth}
\includegraphics[scale=0.50]{pm2.pdf}
\end{minipage}
\begin{minipage}{0.49\textwidth}
Figure 2. The surfaces $g_1(m,n,r)=0$, $g_2(m,n,r)=0$ and $g_3(m,n,r)=0$ in blue, green and red, respectively, in the box
$[\widehat {m}-\varepsilon,\widehat {m}+\varepsilon]
\times[\widehat {n}-\varepsilon,\widehat {n}+\varepsilon]\times[\widehat {r}-\varepsilon,\widehat {r}+\varepsilon],$ where $\varepsilon=10^{-10}$.
\end{minipage}
\end{center}
The hypothesis of Poincar\'e-Miranda theorem for $f$ using the box
$\mathcal{I}_{1,5,11}$ are not satisfied: for instance, at the
points $(\underline m,\widehat {n},\widehat {r})$ and $(\overline
m,\widehat {n},\widehat {r})$ none of the functions $d_{10}, d_{11}$
and $d_{12}$ changes sign. So in order to rectify the level 0
surfaces of the components of $f$, we consider the new function
$$
g(m,n,r)=\left(g_1(m,n,r),g_2(m,n,r),g_3(m,n,r)\right):=(\mathrm{D}f(\widehat
{p}))^{-1} (f(m,n,r))^t.
$$
We omit here the expressions of
$(\mathrm{D}f(\widehat {p}))^{-1}$ and $g$ since they involve huge
rational numbers with numerators and denominators with hundreds of
digits. Notice that since
$\operatorname{det}\left(\mathrm{D}f(\widehat {p})\right)\ne0$
the point $(m_0,n_0,r_0)$ is a zero
of $g$ if and only if it is a zero of $f$.
Observe that $ g(m,n,r)=g(\widehat {p})+(m-\widehat {m},n-\widehat
{n},r-\widehat {r})+O(||(m-\widehat {m},n-\widehat {n},r-\widehat
{r})||^2)$. Since $g(\widehat {p})\simeq0,$ near $\widehat p$ it
holds that $ g(m,n,r)\simeq (m-\widehat {m},n-\widehat
{n},r-\widehat {r})$ and so, a small enough box centered at
$\widehat p$ should be under the hypotheses of Poincar\'{e}-Miranda
theorem, see Figure 2. Now we will check that, indeed, this is the
situation for the function $g$ in the box $\mathcal{I}_{1,5,11}$.
In order to prove that the components of the function $g$ have
no roots, and alternate signs at the faces of $\mathcal{I}_{1,5,11}$
we will apply repeatedly the following technical result, that is a simplified
version adapted to our interests
of a result given in
\cite{JD2}:
\begin{lem}\label{zeros2}
Let $G_\alpha(x)=g_n(\alpha) x^n+g_{n-1}(\alpha) x^{n-1}+\cdots+g_1(\alpha)
x+g_0(\alpha)$ be
a family of real polynomials that depend continuously on a real parameter
$\alpha\in\Lambda=[\alpha_1,\alpha_2]\subset\mathbb{R}$. Fix $J=[a,b]\subset{\mathbb R}$ and assume that:
\begin{enumerate}[(i)]
\item There exists $\alpha_0\in \Lambda$ such that $G_{\alpha_0}(x)$ has no real roots in $J$.
\item For all $\alpha\in \Lambda$, $G_\alpha(a)\cdot G_\alpha(b)\cdot \Delta_x(G_\alpha)\neq 0,$ where
$\Delta_x(G_\alpha)$
is the discriminant of $G_\alpha$ with respect to $x.$
\end{enumerate}
Then for all $\alpha\in \Lambda$, $G_\alpha(x)$ has no real roots in $ J$.
\end{lem}
We will prove that the first component of $g$ has no roots, and
alternates signs at the faces $m=\underline m$ and $m=\overline m$
of the box $\mathcal{I}_{1,5,11}$. Consider the function $
G_n(r)=g_1(\underline m,n,r)\cdot g_1(\overline m,n,r). $ We will
prove that $G_n(r)<0$ for all $(n,r)\in I_5\times I_{11}$ using
Lemma~\ref{zeros2} with $\Lambda=I_5$ and $J=I_{11}.$
By the Sturm's method it can be seen that the polynomial
$G_{\widehat {n}}(r)$ has only 6 different real roots and that none
of them is in the interval $I_{11}.$ Hence the hypothesis $(i)$ is
satisfied. Moreover $G_{\widehat {n}}(r)$ restricted to $I_{11}$ is
negative.
Proceeding in an analogous way, we obtain that $G_n(\underline
r)\cdot G_n(\overline r)$ has only 4 different real roots and none
of them belongs to $I_{5}$. Hence $G_n(\underline r)\cdot
G_n(\overline r)\neq 0$ for all $n\in I_5$. We also check that the
discriminant $\Delta_{r}(G_n(r))$, which is a polynomial of degree
192 in $n$, has $37$ different real roots. Again, we prove that
they are not in $I_5$ and so, we are under the hypothesis $(ii)$ of
Lemma \ref{zeros2}. Hence by this lemma we get that $G_n(r)<0$
for all $(n,r)\in I_5\times I_{11},$ as we wanted to prove.
Doing similar arguments and computations, we obtain that the second
and third component of $g$ do not vanish, and alternate signs on the
faces $n=\underline n$ and $n=\overline n$, and $r=\underline r$
and $r=\overline r$ of $\mathcal{I}_{1,5,11}$, respectively. See
\cite[Chap. 5]{Llor} for more details. Thus $g(m,n,r)$ verifies the
hypothesis of the Poincar\'e-Miranda theorem in
$\mathcal{I}_{1,5,11}.$ Hence the function $g$, and therefore the
function $f$, have at least one zero in this box, which is unique
by construction.
\subsection{Analytic location of the $3$-periodic points}\label{ss:location}
In this section we use that the parameters $m$, $n$ and $r$
associated to each periodic point are located in the $12$ boxes
given in \eqref{E-supercaixes}, to obtain an analytic location of
them in the $(a,b)$-plane.
\begin{lem}\label{L-acotacions-final}
Let $m\in[\underline m ,\overline m ]$ and $r\in[\underline r
,\overline r ]$, and the functions $a(m,r)$ and $b(r)$ given by
\eqref{E-abcdef}, then:
\begin{enumerate}
\item[(i)] If $0<\underline r \leq \overline r $ then
$\underline a :=\underline m ^3-\overline r -2-{4}/{\underline
r ^2}\leq a(m,r)\leq \overline m ^3-\underline r -2-{4}/{\overline
r ^2}=:\overline a $ and $\underline b :=\underline r
+{4}/{\overline r ^2}\leq b(r)\leq \overline r +{4}/{\underline r
^2}=:\overline b $
\item[(ii)] If $\underline r \leq \overline r <0$ then
$ \underline a :=\underline m ^3-\overline r -2-{4}/{\overline
r ^2}\leq a(m,r)\leq \overline m ^3-\underline r
-2-{4}/{\underline r ^2}=:\overline a $ and $\underline b
:=\underline r +{4}/{\underline r ^2}\leq b(r)\leq \overline r
+{4}/{\overline r ^2}=:\overline b $.
\end{enumerate}
\end{lem}
\begin{proof}
(i) From \eqref{E-abcdef} we get that $a(m,r)=m^3-r-2-{4}/{r^2}$ and
$ b(r)=r+{4}/{r^2}.$ Notice that if $0<\underline r \leq r \leq
\overline r $ then $-\overline r \leq -r \leq -\underline r <0$ and
$ -\frac{4}{\underline r ^2}\leq -\frac{4}{r^2}\leq
-\frac{4}{\overline r ^2}<0 .$ Moreover, $\underline m ^3\leq m^3
\leq \overline m ^3.$ By combining all these chains of inequalities
we get statement (i). The statement (ii) follows similarly.~\end{proof}
By using the inequalities in $(ii)$ of Lemma
\ref{L-acotacions-final} we obtain, for example, that the
$3$-periodic point $(a,b)$ of $G$, associated to the parameters
$(m,n,r)\in \mathcal{I}_{1,5,11}$, satisfies $a\in[\underline a
,\overline a ]$ and $b\in[\underline b ,\overline b ]$ where
{\tiny \begin{align*}\underline a &= -\frac {
1435686715756812113129131753291751212473714621389705932746390847605145815709035232062993533718832495489341
}{
56947609584619278435915236206283183709714097978506070511694763452312581699417401160811385506316156928
},\\
\overline a &=-\frac{
47044582301919219323098597682011430719430330620984084471100755414697990442772197382375529104298060913286874119
}{
1866059270868804515791575090678155019012542140400238364480469557193740712623709418953041434224970065510400
}
\end{align*}}
and {\tiny \begin{align*} \underline b &={\frac {
3368785687756582636246263551756811406295236320753178521304454421527}{
10710654937528498667637446691242283113536911386660380934878003200}},\\
\overline b & ={\frac {
6579659546399575461418490144259606329620802274396204966850496409}{
20919247924860348960190099800217926294342327605140376818548736}}.
\end{align*}}
By using the decimal approximation we get, {\scriptsize
\begin{align*}
&a\in[\underline a ,\overline a ]\simeq
[-25210.658115921519312682, -25210.658115921519312679],\\
&b\in[\underline b ,\overline b ]\simeq
[314.5265819322469464743, 314.5265819322469464749],
\end{align*}}
where we observe that $\max(\overline a -\underline a
,\overline b -\underline b )\simeq 2.8\times 10^{-18}.$
Applying Lemma \ref{L-acotacions-final} to each of the $12$ boxes \eqref{E-supercaixes}, we obtain
rational bounds for the components of each periodic point of minimal
period $3$, that are summarized in the following tables, where only
the decimal expression of some significative digits is given. In all
the cases the maximum length of the interval localizing the
3-periodic points is smaller than $10^{-17},$ so the given expression of both ends of the intervals coincide.
{\scriptsize
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
$\mathcal{O}_1$ & $a$ & $b$ \\
\hline
& -25210.658115921519313 & 314.52658193224694647 \\
\hline
& -11.080089229288244821& -29.194152462502174029 \\
\hline
& 1.0164106270635353803& -3.0178440371837045505 \\
\hline
\end{tabular}
\begin{tabular}{|c|c|c|}
\hline
$\mathcal{O}_2$ & $a$ & $b$ \\
\hline
& -25080.503857555317449& 314.36115078061939834\\
\hline
& -11.094342178650567807& -29.143225143670723223 \\
\hline
& 1.0179782228602330827& -3.0165421366176918413 \\
\hline
\end{tabular}
\end{center}}
{\scriptsize
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
$\mathcal{O}_3$ & $a$ & $b$ \\
\hline
& -550.35997876621370288& 84.580855473468510676 \\
\hline
&-13.613164340185764400 & -7.6737642167841728949 \\
\hline
& 0.13590789992610542444& -2.1255835876361107899 \\
\hline
\end{tabular}
\begin{tabular}{|c|c|c|}
\hline
$\mathcal{O}_4$ & $a$ & $b$ \\
\hline
&-500.96942815695686889 & 80.145842594816842809 \\
\hline
&-13.481597649423988848 & -7.4104176831057891201 \\
\hline
& 0.088325991394389446424& -2.0994294342645985249\\
\hline
\end{tabular}
\end{center}}
\section{Dynamics associated to the saddle point $P_2$}
\label{S-Num-An}
In this section we study the invariant sets of the saddle point
$P_2$. First, Proposition \ref{P-atraccio-g} characterizes the
stable set of $P_2.$ We will also give numerical evidences of the
existence of homoclinic orbits, that is, initial conditions on the
local unstable manifold, whose orbit converges to $P_2$. Finally, we
will provide numerical evidences of the existence of points on the
local unstable manifold, whose orbits end in the non-definition set.
\subsection{The stable set of $P_2$}\label{SS-stab-set}
We denote the \emph{stable set} of the fixed point $P_2$ as
$W^s(P_2)=\{(a,b)\in {\mathbb R}^2:\, \lim_{n\to\infty} G^n(a,b)=P_2\}.$ This set is
not necessarily a manifold. For hyperbolic points, like $P_2$ there
is also the so called \emph{local stable manifold}
$W^s_{\textrm{loc}}(P_2)\subset W^s(P_2),$ that is an actual
manifold and is only defined in a small neighborhood of the fixed
point. Our first result characterizes totally the stable set of
$P_2$ and its local stable manifold.
\begin{propo}\label{P-atraccio-g} It holds that
\[
W^s(P_2)=\big(L_2\cup\{(a,b)\in{\mathbb R}^2:\, \exists\, n\ge0: G^n(a,b)\in
R_1\cup C_1\}\big)\setminus\{(-1,-1)\},
\]
where $R_1=\{a-b=0\}$. Moreover, $W^s_{\textrm{loc}}(P_2)$ is
contained in $L_2.$
\end{propo}
\begin{proof} Observe that Equation (\ref{E-Rinv}) implies that the only initial conditions
mapped by $G$ to the resolvent curve are the points of the straight line
$R_1,$ except $(-1,-1).$
Hence, to prove the proposition it suffices to show that $L_2'=L_2\setminus\{(-1,-1)\}\subset W^s(P_2).$
Let us prove this inclusion.
Recall that $L_2\subset\{R(a,b)=0\}.$ The
resolvent curve $R(a,b)=0$ is algebraic and has genus 0, so it
admits \emph{rational parametrizations}. For instance, if we define
$P(t)=(P_1(t),P_2(t))=\left(\frac{t^3+4}{t^2},\frac{t^3+16}{4t}\right)$ it holds that
$R(P_1(t),P_2(t))\equiv0.$ This parametrization has been already was also used in \cite[Thms 3 and 4]{chamb}.
The component $L_2$ corresponds with
$t\in(-\infty,0)$, and $L_1$ with
$t\in(0,\infty)$. Some computations give
$P^{-1}(a,b)={\frac{4\,({a}^{2}-3\,b)}{{a}^{2}b-4\,{b}^{2}+3\,a}}.$
Hence, to study the dynamics of $G$ on the component
$L_2$ we need to study the one-dimensional map
$$
g(t)=P^{-1}\circ G\circ P(t)=\sqrt[3]{4}\,\frac{t}{(t+2)^2}\,
\left(\frac{(t^2+4)(t+2)^2}{t^2}\right)^{2/3}, \mbox{ for }t\in
\mathcal{I}:=(-\infty,0)\setminus\{-2\},
$$ see also \cite[Thm 4]{chamb}. Observe that $t=-2$
corresponds with $(a,b)=(-1,-1)$ which belongs to the non-definition line $\{a+b+2=0\}$ and is excluded in our statement. The map $g(t)$ has a
unique fixed point in $\mathcal{I}$
$$
p=-\frac{1}{3}\,{\frac{4\,C+\sqrt [3]{2}{C}^{2}+8\,{2}^{2/3}}{C}}\simeq -4.4111,\,\mbox{ where } C=\sqrt [3]{86+6\,\sqrt {177}}.
$$
Our objective is to prove that this fixed point is a global attractor of $g(t)$ in $\mathcal{I}$.
First we summarize some features of $g(t)$ in $\mathcal{I}$ that
we will need (see Figure 3): (i) It has only two relative extremes
(maximum) in $\mathcal{I}$ given by
$t=-4\mp 2\sqrt{3}$ ($t\simeq -7.4641$ and $t\simeq -0.5359$ respectively), and such that $g(-4\mp 2\sqrt{3})=-4$.
We denote $m:=-4- 2\sqrt{3}$. (ii) It holds that
$\lim\limits_{t\to-2^{\pm}}g(t)=\lim\limits_{t\to0^{-}}g(t)=-\infty$.
(iii) It also holds that $\lim\limits_{t\to-\infty}g(t)=-\infty$.
(iv) For all $t\in(-\infty,p)$, we have $g(t)>t$.
(v) The map $g$ has not $2$-periodic points as a
consequence of Theorem \ref{T-main} (b).
\begin{center}
\includegraphics[scale=0.42]{g3.pdf}
Figure 3. Graph of the function $g(t)$ in $\mathcal{I}$.
\end{center}
The proof has three steps, namely (A)--(C):
(A) From the properties (i) and (ii), we conclude $g\left((-2,0)\right) = (-\infty,-4]$.
\noindent (B) Using (i) and (ii) again, we conclude that
$g\left((-\infty,-2)\right) = (-\infty,-4]$,
hence the interval $(-\infty,-2)$ is invariant by $g$.
We will study the dynamics in this interval.
Let $\ell\in(p,-2)$ be the unique value in this interval such that $g(\ell)=m$ ($\ell\simeq -2.6675$).
By using the monotony of $g$, the interval $[m,\ell]$ is invariant. Indeed, $g([m,\ell])=[g(\ell),g(m)]=[m,-4]\subset[m,\ell],$
see again Figure 3. Now we claim that \emph{for all $t\in (-\infty,m)\cup(\ell,-2)$
there exists $n>0$ such that $t_n=g^n(t)\in [m,\ell]$.}
Indeed, by the monotonicity of $g$ in $(\ell,-2)$ we have that
$g((\ell,-2))=(-\infty,m)$. Since for all $t\in(-\infty,m)$, we have
$g(t)<-4<\ell$, then $g(t)\notin (\ell,-2)$, hence
$g(t)\in(-\infty,\ell)$. We only need, therefore to prove the claim
in $t\in (-\infty,m)$. We proceed by contradiction. Consider
$t_0\in(-\infty,m)$ and suppose that none iterate $t_n\in[m,\ell]$,
so that for all $n>0$ we have $t_n\in(-\infty,m)$. From (iv), the
sequence $\{t_n\}$ is increasing, and as we are assuming that it is
bounded from above by $m$, the sequence must have a limit that, by
continuity, must be a fixed point, which is a contradiction because
there is no fixed point in $(-\infty,m]$. Hence the claim is
proved, and we only have to study the dynamics of $g$ in
$[m,\ell]$.
\noindent (C) We study now the dynamics on the interval $[m,\ell]$.
We denote $m_0:=m$ and $\ell_0:=\ell$, and consider the sequences
\begin{equation}\label{E-g2k}
\begin{array}{l}
\ell_k=g(m_{k-1})=g^2(\ell_{k-1}) \,\mbox{ and }
m_k=g(\ell_k)=g^2(m_{k-1}).
\end{array}
\end{equation}
Observe that since $g$ is strictly decreasing in $[m,\ell]$, for $k\geq 1$ we obtain
$$
\left[m_{k-1},{\ell}_k\right]:= g^{2k-1}\left([m ,\ell]\right) \,\mbox{ and }\,
\left[m_{k},{\ell}_k\right]:= g^{2k}\left([m ,\ell]\right).
$$
We will prove that $\{m_k\}$ and $\{\ell_k\}$ are increasing and
decreasing sequences, respectively, that converge to the fixed point
$p$, thus proving the result.
Some computations show that $\ell_1=g(m)=-4<\ell=\ell_0$, and that $m_0=m<m_1=g(\ell_1)=g(-4)$. We proceed by induction,
assuming that $m_{k-1}<m_k$ and $\ell_k<\ell_{k-1}$.
Using \eqref{E-g2k}, as
$g$ is decreasing and $m_{k-1}<m_k$, we have $
\ell_k=g(m_{k-1})>g(m_k)=\ell_{k+1}.$
Likewise, since $\ell_{k+1}<\ell_k$ we have
$
m_{k+1}=g(\ell_{k+1})>g(\ell_k)=m_k$. Therefore, the sequences
$\{m_k\}$ and $\{\ell_k\}$ are monotonous increasing and
decreasing, respectively. Since both sequences are bounded,
and using the expressions in \eqref{E-g2k}, we have that both converge to a fixed point of
$g^2$. But since there are not $2$-periodic points, except
the fixed point $p$, we have $\lim\limits_{k\to \infty} m_k=\lim\limits_{k\to \infty} \ell_k=p.$~
\end{proof}
\subsection{Local expression of the unstable manifold
In order to search numerically the homoclinic points associated to
$P_2$, we compute an approximation of the local unstable manifold of
the saddle point $P_2=(a_2,b_2)$. We consider the change
$u=a-a_2$ and $v=b-b_2$, which brings $P_2$ to the origin $(0,0)$. We also consider the map
$\widetilde{G}(u,v)=G(u+a_2,v+b_2)-(a_2,b_2)$ which is conjugate with $G$, and
the linear map given by
$H(r,s)=L \cdot (r,s)^t$,
where $L$ is the matrix formed by the eigenvectors of
$\mathrm{D}G(P_2)$. Hence
$$L^{-1} \cdot \mathrm{D}G(P_2) \cdot L=
\left( \begin{array}{cc}
\lambda_1 & 0 \\ 0 & \lambda_2
\end{array} \right),$$
where $\lambda_1$ and $\lambda_2$ are the eigenvalues of $\mathrm{D}G(P_2)$, given by
{\footnotesize \begin{align*}
\lambda_1&=\frac{-1}{384}\left(\left(7\,
\sqrt {177}-111\right) \,A^{2/3} +\left(8\,\sqrt {177}-264\right)\,A^{1/3}-768\right)\simeq 7.0701,\\
\lambda_2&={\frac{-1}{1152}}
\left(\left(\sqrt {177}
-25\right)\, A^{2/3}+\left(8\,\sqrt {177}
-136\right)\,A^{1/3}+1280\right)\simeq -0.4470,
\end{align*}}
where $A:=172+12\,\sqrt {177}$.
We compute the Taylor development of the unstable manifold
associated to the origin of the map $ F(r,s)=H^{-1} \circ
\widetilde{G} \circ H(r,s)=\left(\lambda_1
r+O(||(r,s)||^2),\lambda_2 s+O(||(r,s)||^2\right). $ \emph{The
expression of the local unstable manifold $W^u_{\textrm{loc}}(0,0)$
of $F$ is $s=w(r)=w_2 r^2+w_3 r^3+w_4 r^4+w_5 r^5+O(r^6),$
where}
{\footnotesize \begin{align*}
&w_2\simeq- 0.00259107002218996975513519324145,\,
&&w_3\simeq- 0.00013220529650666650558465802906, \\
&w_4\simeq- 0.00000889870356674847560384348601, \,
&&w_5\simeq- 0.00000069374812274441343473691330.
\end{align*}}
These coefficients have been computed using the formulas in Lemma
\ref{var_inest_expr} of the Appendix, by using floating-point
arithmetic with $60$ digits in the mantissa.
Observe that we can parametrize $W^u_{\textrm{loc}}(P_2)$
using the function $s=w(r)$, by considering
\begin{equation}\label{E-paramWu}
r\longrightarrow H(r,w(r))+P_2,\, \mbox{ for } r\simeq 0.
\end{equation}
We use this parametrization to obtain Figures 4 and 5.
Finally, from the expression of the local unstable manifold of the
origin for the map $F$, we obtain that
the points $(a,b)\in W^u_{\mathrm{loc}}(P_2)$ satisfy
$
w(H_1^{-1}(a-a_2,b-b_2))-H_2^{-1}(a-a_2,b-b_2)=0,
$
that can be approximated by
$$
D_1(a,b):=\sum\limits_{i=2}^5 w_i \big(H_1^{-1}(a-a_2,b-b_2)\big)^i -H_2^{-1}(a-a_2,b-b_2)=0,
$$
where $D_1(a,b)$ is a polynomial of degree 5 that we do not
explicite for the sake of shortness, see \cite[Chap. 5]{Llor} for
more details.
\subsection{Computation of the homoclinic point}
Previous to find a homoclinic point we remember that, by Proposition
\ref{P-atraccio-g}, any point $(a,b)$ such that there exists
$k\in{\mathbb N}$ verifying $G^k(a,b)\in R_1=\{a-b=0\} \cap
C_1\setminus\{(-1,-1)\}$ belongs to the stable set of $P_2$, since
$G^{k+1}(a,b)\in L_2$. In this sense, we have graphically observed
that, except for the point $P_2$, there is no intersection of
$W^u_{\mathrm{loc}}(P_2)$ with the curve $L_2$. Also we have
observed neither intersections of $W^u_{\mathrm {loc}}(P_2)$ with
$R_1=\{a-b=0\}$ at the region $C_1$, nor points $(a,b)\in
W^u_{\textrm{loc}}(P_2)$ such that $G(a,b)\in R_1\cup C_1$, but we
have seen the existence of at least one point such that $G^2(a,b)\in
R_1\cup C_1$. See Figure 4.
Imposing $G_1(a,b)-G_2(a,b)=0$, we find that the points $(a,b)$ such that $G(a,b)\in R_1$ satisfy:
$${D_2}(a,b):=\left( ab+5\,a+5\,b+9 \right) ^{3}-
\left( a+b+6 \right) ^{3} \left(a+b+2 \right) ^{2}=0.$$
Hence, the points such that $G^2(a,b)\in R_1$ are those satisfying ${D_2}
(G(a,b))=0$, or equivalently $D_3(a,b):=\mathrm{numer}({D_2}
(G(a,b)))=0,$
where $D_3(a,b)$ is a polynomial of degree $10$ in the variable
$m=(a+b+2)^{2/3}$ with $22$ terms, that we omit here.
Therefore, the homoclinic point $P$ must verify the system
$\{D_1(a,b)=0, D_3(a,b)=0\}$.
We solve it numerically, using floating-point arithmetic with $60$
digits in the mantissa, and we get a solution in
$[-6,-5]\times[3.5,5]$, given by $P=(p_1,p_2)$ where
{\footnotesize
\begin{align*}
p_1&\simeq-5.67750144031789435343891174392876990152177028290023619512062, \\
p_2&\simeq 4.10574868714920935493626045239900450809925741194290963919902 .
\end{align*}}
By using the parametrization of $W^u_{\mathrm{loc}}(P_2)$ given by
\eqref{E-paramWu}, we find that the point
$P$ corresponds with the parameter
$r\simeq -1.48202 15208 77494 33523.$
By construction, $G^3(P)$ which must lie on $L_2$. A computation
shows that the absolute error when we evaluate $R(a,b)$ on this
point, is $\big|R\big(G^3(p)\big)\big|\simeq 10^{-58}$. Accordingly,
the point $P$ exhibits, numerically, a homoclinic behavior.
As can be seen in Figure 4, there exists another solution of
$\{D_1(a,b)=0, D_3(a,b)=0\}$ in $[-8,-6]\times[3.5,5]$, given by
$\tilde{P}=(\tilde{p}_1,\tilde{p}_2)$ where
{\footnotesize
\begin{align*}
\tilde{p}_1&\simeq-7.32664831286596004531700787733138125161658087249633041273728, \\
\tilde{p}_2&\simeq 4.26205920129322448141657538934356322617112224124511704493689.
\end{align*}
} The point $\tilde{P}$ corresponds to the parameter value $r\simeq
-3.14702 44917 79071 04545$.
\begin{center}
\begin{minipage}{0.5\textwidth}
\includegraphics[scale=0.37]{figura6sellaretallada.pdf}
\end{minipage}
\begin{minipage}{0.47\textwidth}
Figure 4. Location of the points $P, G^2(P)$ and
$G^3(P)$ on $W^u_{\textrm{loc}}(P_2)$ (brown), the curve $D_3(a,b)=0$ (green), the diagonal $R_1$ (blue) and
the resolvent curve (red), respectively. The point $G(P)$ is outside the image.
\end{minipage}
\end{center}
\subsection{Computation of points in $W^u_{\mathrm{loc}}(P_2)\cap\mathcal{F}$}
To find a point in $W^u_{\mathrm{loc}}(P_2)\cap\mathcal{F}$, we solve numerically the system $
\left\{D_1(a,b)=0,a+b+2=0\right\},$
obtaining the point $Q=(q_1,q_2)$, where
{\footnotesize
\[
q_1\simeq-6.15163017029193114270539883292276699558057876233980350720282,
\quad q_2=-2-q_1,
\]}
This point corresponds with the parameter $r\simeq -1.96025 81538
61616 87597 .$
To find another point with a parameter value closer to zero (hence giving a better evidence of really being in
$W^u_{\mathrm{loc}}(P_2)$), we find a point $Q_{-1}$ such that $G(Q_{-1})=Q$. The points $(a,b)$ such that
$G(a,b)\in\{a+b+2=0\}$,
verify
$$D_4(a,b):=5\,a+5\,b+ab+9+ \left( a+b+6 \right) \sqrt [3]{ \left( a+b+2 \right) ^
{2}}+2\,\sqrt [3]{ \left( a+b+2 \right) ^{4}}=0.$$
By solving numerically the system
$\left\{D_1(a,b)=0,D_4(a,b)=0\right\},$
we find $Q_{-1}:=(z_1,z_2)$ where
{\footnotesize
\begin{align*}
z_1&\simeq-4.43931733951927306713914976146761550810750048579478327758904, \\
z_2&\simeq 3.98185284365899589972467095578564600569428848801825836848384.
\end{align*}}
The point $Q_{-1}$ has an associated parameter value
$r\simeq -0.23505 95678 85428 61108.$ The location of the above points is shown in Figure 5.
Observe that the parameters of the points $Q$, $P$ and $Q_{-1}$ are
interspersed, so the points are also interspersed in
$W^u_{\mathrm{loc}}(P_2)$. An analytic proof of this fact would show
that arbitrarily near of
$P_2$ there are homoclinic points and points in
$\mathcal{F}$.
\begin{center}
\begin{minipage}{0.5\textwidth}
\includegraphics[scale=0.37]{figura7retallada.pdf}
\end{minipage}
\begin{minipage}{0.47\textwidth}
Figure 5. Location of the points $Q$ and $Q_{-1}$
in $W^u_{\mathrm{loc}}(P_2)\cap\mathcal{F}$; the curve $D_4(a,b)=0$ (green);
the line $a+b+2=0$ (blue); and the resolvent curve
(red), respectively.
\end{minipage}
\end{center}
\section*{Appendix: Local unstable manifold near a hyperbolic saddle point}
\begin{lem}\label{var_inest_expr}
Consider the smooth map, defined in a neighborhood of the origin \
$\mathcal{U}$:
$$
F(x,y)=\left(\lambda x+ \sum_{{i+j}={2}}^5{f_{i,j}
x^i y^j}+O(||(x,y)||^6),\mu y+ \sum_{{i+j}={2}}^5{g_{i,j} x^i y^j}+O(||(x,y)||^6)\right)
$$
where $\arrowvert \lambda \arrowvert > 1 >
\arrowvert\mu\arrowvert$, so that the origin is a hyperbolic saddle. Let
$y=w(x)=\sum_{k=2}^5{w_k x^k}+O(x^6)$ be the expression of the local
unstable manifold in a neighborhood of the origin. Then:
$$ w_2=\frac{g_{2,0}}{\lambda^2-\mu},\quad w_3={\frac
{{\lambda}^{2}g_{{3,0}}-2\,\lambda\,f_{{2,0}}g_{{2,0}}-\mu\,g_{
{3,0}}+g_{{1,1}}g_{{2,0}}}{ \left( {\lambda}^{2}-\mu \right) \left(
{ \lambda}^{3}-\mu \right) }} ,\quad
w_4=\frac{W_4}{ \left( {\lambda}^{2}-\mu \right) ^{2} \left( {\lambda}^{3}-\mu
\right) \left( {\lambda}^{4}-\mu \right)} ,
$$
where
\noindent\begin{align*} W_4&= g_{{4,0}}{\lambda}^{7}+ \left(
-3\,f_{{2,0}}g_{{3,0}}-2\,f_{{3,0}}g_{{ 2,0}} \right) {\lambda}^{6}+
\left( 5\,f_{2,0}^2g_{{2,0}}-2\,g_{
{4,0}}\mu+g_{{2,0}}g_{{2,1}} \right) {\lambda}^{5}\\
&+ \left( \left( 6\,
f_{{2,0}}g_{{3,0}}+2\,f_{{3,0}}g_{{2,0}}-g_{{4,0}} \right) \mu-2\,f_{{
1,1}}g_{2,0}^2-3\,f_{{2,0}}g_{{1,1}}g_{{2,0}}+g_{{1,1}}g_{{3,0}}
\right) {\lambda}^{4}\\
&+ \left( g_{{4,0}}{\mu}^{2}+ \left( -5\,
f_{2,0}^{2}g_{{2,0}}+2\,f_{{3,0}}g_{{2,0}}-g_{{2,0}}g_{{2,1}} \right) \mu-
2\,f_{{2,0}}g_{{1,1}}g_{{2,0}}+g_{{0,2}}g_{2,0}^2 \right) {
\lambda}^{3}\\
&+ \left( \left( -3\,f_{{2,0}}g_{{3,0}}+2\,g_{{4,0}}
\right) {\mu}^{2}+ \left( f_{2,0}^2g_{{2,0}}+3\,f_{{2,0}}g_{{1,
1}}g_{{2,0}}-2\,g_{{1,1}}g_{{3,0}}-g_{{2,0}}g_{{2,1}} \right) \mu\right.\\
&\left.+g_{1,1}^{2}g_{{2,0}} \right) {\lambda}^{2}+ \left( -2\,f_{{3,0}}g_{{2,0
}}{\mu}^{2}+ \left( 2\,f_{{1,1}}g_{2,0}^2+2\,f_{{2,0}}g_{{1,1}}g
_{{2,0}} \right) \mu \right) \lambda\\
&-g_{{4,0}}{\mu}^{3}+ \left( -f_{{2,0}}^{2}g_{{2,0}}+g_{{1,1}}g_{{3,0}}+g_{{2,0}}g_{{2,1}} \right) {\mu
}^{2}+ \left( -g_{{0,2}}g_{2,0}^2-g_{1,1}^2g_{{2,0}}
\right) \mu
\end{align*}
and
$$
w_5=\frac{\sum_{i=0}^{13} p_i\,\lambda^i}{\left( {\lambda}^{2}-\mu \right) ^{3} \left( {\lambda}^{3}-\mu
\right) \left( {\lambda}^{4}-\mu \right) \left( {\lambda}^{5}-\mu
\right)},
$$
where $p_i, i=1,2,\ldots,13$ are polynomials in the other variables
of $F$ that we skip, although we have used, for the sake of
shortness (they are given in \cite[Chapter 5]{Llor}).
\end{lem}
\begin{proof} Due to the particular form of the linear part of $F$, the local unstable manifold $W^u_{\mathrm{loc}}(0,0)$ is given by
a smooth function of the form
$y=w(x)=w_2x^2+w_3x^3+w_4x^4+w_5x^5+O(x^6),$ that is, a point is on
the local stable manifold if it is of the form
$(x,w(x))$. Imposing that $F(x,w(x))=(F_1(x,w(x)),F_2(x,w(x)))$ is also on this curve we get that
the points on the local unstable manifold must satisfy
$
F_2(x,w(x))=w\big(F_1(x,w(x))\big).
$
The result follows by comparing the terms in the Taylor development
of both members of the last equation.~\end{proof}
\subsection*{Acknowledgements} The authors are supported by
Ministry of Economy, Industry and Competitiveness of the Spanish
Government through grants MINECO/FEDER MTM2016-77278-P (first
author) and DPI2016-77407-P
(AEI/FEDER, UE, second and third author). The first author is also supported by the grant 2014-SGR-568 from
AGAUR, Generalitat de Catalunya. The third author is supported by
the grant 2014-SGR-859 from AGAUR, Generalitat de Catalunya.
|
\section{Adaptive Bit Allocation Algorithm} \label{sec:algo}
This section presents the adaptive bit allocation algorithm to maximize the OFDM CU throughput, i.e. the transmitted number of bits per OFDM symbol, while maintaining the average BER across all OFDM subcarriers below a target BER, in the presence of a NB interference and imperfect channel estimation. This can be formally expressed as
\begin{eqnarray}
\underset{m_k}{max} & & \sum_{k=1}^{N}m_k, \nonumber \\
\textup{subject to} & & \overline{BER} = \frac{\sum_{k=1}^{N}m_k (BER_k)}{\sum_{k=1}^{N}m_k}\leq BER_T,
\end{eqnarray}
where $m_k$ is the number of bits allocated to subcarrier $k$, $\overline{BER}$ is the mean BER, $BER_T$ is the target BER, and $BER_k$ is the BER for subcarrier $k$.
The idea behind the algorithm is to load all subcarriers with the highest possible constellation size and uniform power, and then calculate the BER per subcarrier depending on the channel state condition and the NB interference per subcarrier. The average BER is finally calculated and checked against the target BER. If the average BER meets the target BER, then the final bit allocation is reached; otherwise, the signal constellation on the worst performance subcarrier is decreased and the process repeats.
The modulation schemes considered in this work are BPSK, QPSK, 16-QAM, and 64-QAM, i.e. each subcarrier can be loaded with a symbol drawn from one of the previously mentioned modulation schemes. The closed-form expression for the $BER_k$ of B/QPSK PSK is given by \cite{proakisdigital}
\begin{eqnarray}
BER_k &{} = &{} Q\left ( \sqrt{2\frac{T_u}{T_o}\gamma_k} \right ),
\label{eq:PSK}
\end{eqnarray}
where $T_u$ and $T_o$ are the useful OFDM symbol duration and the OFDM symbol duration including the cyclic prefix, respectively, $\frac{T_u}{T_o}$ is the loss due to cyclic prefix, and $Q(.)$ represents the Q-function. The closed-form expression for the $BER_k$ of 16-QAM and 64-QAM is given by \cite{proakisdigital}
\begin{eqnarray}
BER_k & = &\frac{4}{m_k}\left ( 1-\frac{1}{\sqrt{M_k}} \right ) Q\left ( \sqrt{\frac{3}{M_k-1}\frac{T_u}{T_o}\gamma_k} \right ) \nonumber \\
& &\left ( 1- ( 1-\frac{1}{\sqrt{M_k}} ) Q\left ( \sqrt{\frac{3}{M_k-1}\frac{T_u}{T_o}\gamma_k} \right ) \right ), \nonumber \\
\label{eq:QAM}
\end{eqnarray}
where $M_k$ is the constellation size per subcarrier $k$. Note that (\ref{eq:PSK}) and (\ref{eq:QAM}) are written under the assumption of Gaussian interference at the OFDM receiver FFT output. As the interference per subcarrier after the OFDM receiver FFT represents the contribution of the interference samples at the FFT input, this tends to have a normal distribution according to the central limit theorem \cite{papoulis1965probability}.
The algorithm can be formally described as follows
\floatname{algorithm}{}
\begin {algorithm}
\renewcommand{\thealgorithm}{}
\caption{\textbf{Adaptive Bit Allocation Algorithm}}
\begin{algorithmic}[1]
\STATE Initialization: set the modulation scheme of all the subcarriers to 64-QAM.
\STATE Determine $BER_k$, $k$=1,...,$N$, given the SINR $\gamma_k$ values, using (\ref{eq:PSK}) or (\ref{eq:QAM}).
\STATE Compare $\overline{BER}$ with $BER_T$. If $\overline{BER}$ is less than $BER_T$, the current configuration is kept and the algorithm ends.
\STATE Search for the subcarrier with the worst $BER_k$ and reduce the constellation size. If $m_k$ = 1, null the subcarrier (i.e., set $m_k$ = 0).
\STATE Recompute $BER_k$ of all subcarriers with changed allocations and return to step 3.
\STATE If $BER_T$ cannot be met, the transmission stops.
\end{algorithmic}
\end{algorithm}
We should mention that \cite{wyglinski2005bit} proposes a similar bit allocation algorithm without considering the interference effect, which is crucial in the CR environment.
\section*{Appendix: \small{Derivation of NB EU variance per subcarrier ($\sigma^2_{I,k}$) (\ref{eq:var})}}
\addcontentsline{toc}{chapter}{Appendix}
\markboth{Appendix}{Appendix}
Here we derive the analytical expression for the NB EU variance per subcarrier ($\sigma^2_{I,k}$) after the OFDM receiver FFT. The NB EU signal is given by:
\begin{eqnarray}
I(t) & = &\sum_{l =-\infty }^{\infty}b_{l}\:p(t-lT-\xi) \:e^{j2\pi f_{c}t},
\end{eqnarray}
where $b_{l}$ is the $lth$ data symbol, $p(t)$ is the impulse response of the transmit filter, $T$ is the symbol period, $\xi$ is the time delay, and $f_{c}$ is the NB EU frequency deviation from the OFDM carrier frequency. As we are interested in the variance of this NB EU signal after the OFDM receiver FFT, this signal is sampled by the OFDM sampling time $T_s = \frac{1}{BW}$ resulting in
\begin{eqnarray}
I(vT_s) & = & \sum_{l =-\infty }^{\infty}b_{l}\:p(vT_s-lT-\xi)\: e^{j2\pi f_{c}vT_s}.
\label{eq:int_sam}
\end{eqnarray}
By replacing $v$ by $n+rN$, with $n$ = $0,..., N-1$ and $r$ integer, (\ref{eq:int_sam}) becomes
\begin{eqnarray}
I^r(n) & = & \sum_{l =-\infty }^{\infty}b_{l}\:p((n + rN)T_s-lT-\xi)\: e^{j2\pi f_{c}(n+rN)T_s}.
\end{eqnarray}
The resulting signal after the OFDM receiver FFT is
\begin{eqnarray}
X^r(k) & = & \frac{1}{\sqrt{N}}\sum_{n=0}^{N-1} I^r(n) \:e^{-j\frac{2\pi kn}{N}}, \\
& = & \frac{1}{\sqrt{N}}\sum_{n=0}^{N-1} \sum_{l =-\infty }^{\infty}b_{l}\: p((n + rN)T_s-lT-\xi)\: e^{j2\pi f_{c}(n+rN)T_s} \: e^{-j\frac{2\pi kn}{N}}.
\end{eqnarray}
Finally, the variance of NB EU signal after the OFDM receiver FFT can be written as
\begin{eqnarray}
(\sigma^{r}_{I,k})^2 & = & \mathbb E\left \{ X^r(k) X^{*r}(k) \right \} \nonumber \\
& = & \frac{\sigma_b^2}{N} \sum_{n=0}^{N-1}\sum_{n'=0}^{N-1} \sum_{l =-\infty }^{\infty} p((n + rN)T_s-lT-\xi) \: p((n + rN)T_s-lT-\xi) \nonumber \\
& & e^{j2\pi f_{c}(n-n')T_s} \: e^{-j\frac{2\pi k}{N}(n-n')}, \nonumber
\end{eqnarray}
where $\mathbb E\{.\}$ denotes the statistical expectation operator, and $\sigma_b^2$ = $\mathbb E\{b_l b^*_l\}$. By averaging over the number of OFDM symbols $R$, the NB EU variance per subcarrier can be finally expressed as
\begin{eqnarray}
\sigma^2_{I,k} & = & \lim_{R \rightarrow \infty} \frac{1}{R} \frac{\sigma_b^2}{N} \sum_{r = 0}^{R-1} \sum_{n=0}^{N-1}\sum_{n'=0}^{N-1} \sum_{l =-\infty }^{\infty} p((n + rN)T_s-lT-\xi) \: p((n + rN)T_s-lT-\xi) \nonumber \\
& & e^{j2\pi f_{c}(n-n')T_s} \: e^{-j\frac{2\pi k}{N}(n-n')}.
\end{eqnarray}
\section{Conclusion} \label{sec:future}
In this paper, we presented an adaptive bit allocation algorithm to maximize the OFDM CU throughput in the presence of NB PU for CR systems with imperfect channel estimation. The average throughput of the OFDM CU is maximized under the constraint of a BER below a target value. As expected, the average throughput increases as the CU-PU frequency separation and SIR increase, respectively. Moreover, increasing the channel estimation error variance reduces the average throughput. In future work, we will extend the proposed algorithm to include power loading, and impose a constraint on the maximum transmit power to reduce the OFDM CU spectrum leakage to PUs, in addition to considering spectrum sculpting techniques.
\section{Introduction}
Cognitive Radio (CR), first introduced by Mitola in \cite{Mitola1999}, is a promising wireless communication paradigm that is aware of its radio surroundings and adapts intelligently to improve spectrum efficiency. Unlicensed or cognitive users (CUs) seek to \textit{underlay}, \textit{overlay}, or \textit{interweave} their signals with licensed or primary users (PUs) \cite{kolodzy2005cognitive, srinivasa2007cognitive, goldsmith2009breaking}. The underlay approach allows concurrent transmission of PUs and CUs as in ultrawide band (UWB) systems. CUs spread their transmission over a wide bandwidth, hence their interference is below an acceptable noise floor to PUs. The overlay approach also allows concurrent transmission of PUs and CUs with a premise that CUs can use part of their power to assist/relay PUs transmission. The interweave approach allows CUs to opportunistically access voids in PUs frequency bands/time slots under the condition that no harmful interference occurs to PUs. In this paper, we focus on the interweave CR systems.
Orthogonal frequency division multiplexing (OFDM) is recognized as an attractive modulation technique for CR due to its flexibility and adaptivity in allocating vacant radio resources among CUs \cite{weiss2004spectrum}. In conventional OFDM-based systems, a fixed bit allocation is used on all subcarriers. Thus, the total BER is dominated by the subcarriers that have the worst performance. To improve the system's BER, adaptive bit allocation can be employed such that the information is redistributed across subcarriers according to the channel state information. Consequently, adaptive bit allocation requires accurate channel estimation at the receiver and a reliable feedback channel between the receiver and the transmitter.
Different adaptive bit allocation algorithms are presented in the literature \cite{hughes1988ensemble, de1998optimal, levin2001complete, wyglinski2005bit, fox1966discrete, papandreou2005new, goldsmith1997variable, fischer1996new, kalet1989multitone}. These algorithms can be categorized according to their operation as follows: greedy algorithms \cite{hughes1988ensemble, de1998optimal, levin2001complete, wyglinski2005bit, fox1966discrete, papandreou2005new}, and water-filling based algorithms \cite{goldsmith1997variable, fischer1996new, kalet1989multitone}.
Greedy algorithms provide optimal performance by incrementally allocating an integer number of bits at the cost of high complexity. This was first suggested by Hughes-Hartog in \cite{hughes1988ensemble}, where one bit is added at a time to the subcarrier requiring the smallest incremental power to maximize the throughput. Unfortunately, the algorithm is complex and it converges very slowly. Campello de Souza \cite{de1998optimal} and Levin \cite{levin2001complete} developed a complete and mathematically verifiable algorithm, known as ``Levin-Campello,'' that offers significant improvement to the work of Hughes-Hartog.
On the other hand, water-filling based algorithms formulate the adaptive bit allocation problem as a constrained optimization problem that can be solved by classical optimization methods \cite{cover2004elements}. These algorithms allocate non-integer number of bits to subcarriers in a non-iterative manner. Hence, it compromises performance for lower complexity, as it is generally followed by a rounding-off step to allocate an integer number of bits to the transmitted symbols across all subcarriers, thus lowering the overall data rate \cite{papandreou2005new}.
The work in \cite{hughes1988ensemble, de1998optimal, levin2001complete, wyglinski2005bit, fox1966discrete, papandreou2005new, goldsmith1997variable, fischer1996new, kalet1989multitone} assumes that the OFDM system is interference-free. In this paper, we consider the coexistence between an OFDM CU and a narrowband (NB) PU, and present an adaptive bit allocation algorithm to maximize the average throughput of the CU under an average BER constraint in the presence of the NB interference. The level of NB interference depends on how close the OFDM CU transmits in frequency when compared to the NB PU, for a certain NB PU power. Moreover, the effect of imperfect channel estimation on the algorithm performance is investigated. On the other hand, the OFDM CU spectrum leakage to the NB PU can be reduced by straightforwardly combining the proposed algorithm with spectrum sculpting techniques \cite{budiarjo2006combined,bedeer2011partial}; this is beyond the scope of this paper.
The remainder of the paper is organized as follows. Section \ref{sec:model} presents the system models. Section \ref{sec:algo} delineates the adaptive bit allocation algorithm. Simulation results are presented in Section \ref{sec:sim}. Finally, conclusions are drawn in Section \ref{sec:future}.
\section*{Acknowledgment}
This work has been supported in part by the Communications Research Centre, Canada.
\bibliographystyle{IEEEtran}
\section{System Models} \label{sec:model}
\subsection{OFDM system model}
We consider an OFDM system with $N$ subcarriers that are orthogonal to each other and intersymbol interference (ISI) free. Hence, each subcarrier can be detected independently using a simple maximum likelihood detector, as shown in Fig.~\ref{fig:ch}. Let $X_k$, $Y_k$, $H_k$, $\widehat{H}_k$, and $W_k$ represent the transmitted symbol, the received symbol, the channel gain, the estimated channel gain, and the additive white Gaussian noise (AWGN), respectively, for subcarrier $k$. The received symbol $Y_k$ can be written as
\begin{eqnarray}
Y_k & = &\frac{1}{\widehat{H}_k}\Big ( H_k X_k + W_k \Big ) = X_k \Big ( 1 + \frac{H_k - \widehat{H}_k}{\widehat{H}_k} \Big ) + W'_k,\nonumber \\
& = & X_k + \frac{H_k - \widehat{H}_k}{\widehat{H}_k} X_k + W'_k,
\end{eqnarray}
where $\frac{H_k - \widehat{H}_k}{\widehat{H}_k} X_k$ represents the error added to the transmitted symbol on subcarrier $k$ due to imperfect channel estimation. Leke \textit{et al.} in \cite{leke1998impact,wyglinski2005bit} show that this error has a Gaussian distribution with zero mean and variance $\sigma_{h}^2$ given by
\begin{eqnarray}
\sigma_{h}^2 = \left | \frac{H_k - \widehat{H}_k}{\widehat{H}_k} \right |^2 \sigma_{s,k}^2,
\end{eqnarray}
where $\sigma_{s,k}^2$ is the transmitted symbol power per subcarrier $k$.
\begin{figure}[!t]
\centering
\includegraphics[width=0.50\textwidth]{Ch_est}
\caption{OFDM subcarrier model with imperfect channel estimation.}
\label{fig:ch}
\end{figure}
\subsection{NB system model}
We consider a NB PU, coexisting with the OFDM CU, whose signal is given by
\begin{eqnarray}
I(t) & = &\sum_{l =-\infty }^{\infty}b_{l}\:p(t-lT-\xi) \:e^{j2\pi f_{c}t},
\end{eqnarray}
where $b_{l}$ is the data symbol transmitted in the $l$th period, $p(t)$ is the impulse response of the transmit filter, $T$ is the symbol period, $\xi$ is the time delay, and $f_{c}$ is the NB PU frequency deviation from the OFDM carrier frequency.
The instantaneous signal-to-interference-plus-noise ratio (SINR) on subcarrier $k$, $\gamma_k$, in case of imperfect channel estimation and NB interference can be written as
\begin{eqnarray}
\gamma_k = \frac{\sigma_{s,k}^2}{\sigma_{n}^2 + \sigma_{h}^2 + \sigma_{I,k}^2}|H_k|^2,
\label{eq:SINR}
\end{eqnarray}
where $\sigma_n^2$ is the noise variance and $\sigma_{I,k}^2$ is the interference variance per subcarrier $k$. The value of $\sigma_{I,k}^2$ is calculated by evaluating the NB PU signal power per subcarrier at the OFDM receiver FFT output. The resultant NB signal after the OFDM receiver FFT in the $r$th OFDM symbol can be written as
\begin{eqnarray}
Z^r_k = \frac{1}{\sqrt{N}}\sum_{n=0}^{N-1} \sum_{l =-\infty }^{\infty}b_{l}\: p\Big((n + rN)T_s-lT-\xi\Big) \nonumber \\ \: e^{j2\pi f_{c}(n+rN)T_s} \:e^{-j\frac{2\pi kn}{N}}.
\end{eqnarray}
Hence, the variance of NB signal after the OFDM receiver FFT for the $r$th OFDM block can be written as
\begin{eqnarray}
(\sigma^{r}_{I,k})^2 & = & \mathbb E\left \{ Z^r_k Z^{*r}_k \right \} \nonumber \\
& = & \frac{\sigma_b^2}{N} \sum_{n=0}^{N-1}\sum_{n'=0}^{N-1} \sum_{l =-\infty }^{\infty} p((n + rN)T_s-lT-\xi) \nonumber \\ & & p((n' + rN)T_s-lT-\xi) \: e^{j2\pi f_{c}(n-n')T_s} \nonumber \\
& & \: e^{-j\frac{2\pi k}{N}(n-n')},
\end{eqnarray}
where $\mathbb E\{.\}$ denotes the statistical expectation operator, and $\sigma_b^2$ = $\mathbb E\{b_l b^*_l\}$. By averaging over the number of OFDM symbols $R$, the NB PU variance per subcarrier can be finally expressed as
\begin{eqnarray}
\sigma^2_{I,k} & = & \lim_{R \rightarrow \infty} \frac{1}{R} \frac{\sigma_b^2}{N} \sum_{r = 0}^{R-1} \sum_{n=0}^{N-1}\sum_{n'=0}^{N-1} \sum_{l =-\infty }^{\infty} \nonumber \\
& & p((n + rN)T_s-lT-\xi) \: p((n' + rN)T_s-lT-\xi) \nonumber \\
& & e^{j2\pi f_{c}(n-n')T_s} \: e^{-j\frac{2\pi k}{N}(n-n')}.
\label{eq:sigma}
\end{eqnarray}
\section{Simulation Setup and Results} \label{sec:sim}
\subsection{Simulation setup}
The parameters of the systems considered in this study are provided in Table \ref{tab:SimPar}. A frequency selective fading channel is used for the OFDM CU. The channel impulse response $h(n)$ has a length of $N_{ch}$ = 5 taps, where their components vary independently and are modeled as complex valued Gaussian random variables with zero mean and an exponential power delay profile \cite{morelli2004timing}
\setlength{\arraycolsep}{0.0em}
\begin{eqnarray}
\mathbb{E}\{\left | h(n) \right |^2\} = \sigma_c^2 \: e^{-n\Xi}, \qquad n = 0, 1, ..., N_{ch}-1,
\end{eqnarray}
where $\sigma_c^2$ is a constant chosen such that the average energy per subcarrier is normalized to unity, i.e. $\mathbb{E}\{\left | H_k \right |^2\}$ = 1, and $\Xi$ represents the decay factor, $\Xi = \frac{1}{5}$. Also, we assume that the frequency selective channel is fixed over a number of OFDM symbols, and a total of $10^4$ channel realization are considered. A root-raised cosine transmit filter is considered for the adjacent NB system, and the time delay $\xi$ is taken as a random variable uniformly distributed between 0 and the NB symbol duration, $T$. The target BER, $BER_T$, is chosen to be $10^{-4}$. The normalized frequency $F_n$, which represents the spectral distance between the OFDM CU and NB PU, is defined as $F_n = \frac{f_c}{BW}$, where $BW$ is the OFDM bandwidth.
\begin{table}[!t]
\centering
\caption{Simulation parameters.}
\begin{tabular}{rl}
\hline
{\bf } & {\bf OFDM-based system} \\ \hline
{ Bandwidth, $BW$} & 1.25 MHz \\
{ Window roll-off factor, $\beta$} & 0 \\
{ Number of subcarriers, $N$} & 128 \\
{ Subcarrier spacing, $\Delta F$} & 9.7656 kHz \\
{ Useful symbol duration, $T_u$} & 102.4 $\mu$sec \\
{ CP duration, $T_{cp}$} & 0.25$T_u$ = 25.6 $\mu$sec \\
{ Postfix duration, $T_{p}$} & 0 \\
{ Modulation type} & BPSK, QPSK, 16-QAM, 64-QAM \\ \hline
{\bf } & {\bf NB system} \\ \hline
{ Bandwidth, $BW_N$} & 15 kHz \\
{ Roll-off factor, $\alpha_N$} & 0.35 \\
{ Modulation type} & QPSK \\ \hline
\end{tabular}
\label{tab:SimPar}
\end{table}
\subsection{Simulation results}
Fig. \ref{fig:Th_Fn} shows the average throughput as a function of $F_n$ for different values of SIRs\footnote[1]{The SIR is defined after the OFDM receiver FFT as the ratio between the OFDM and NB interference average powers, respectively, as in \cite{bedeer2011partial}.} at average SNR\footnote[2]{The average SNR is calculated by averaging the instantaneous SNR values per subcarrier over the total number of subcarriers and the total number of channel realizations, respectively.} = 20 dB, assuming perfect channel estimation, i.e. $\sigma_h^2$ = 0. As one can notice, as $F_n$ increases, the NB PU effect on the OFDM CU decreases; hence, more bits can be transmitted on average per each OFDM symbol while achieving a target BER of $10^{-4}$. Furthermore, the average throughput increases as the SIR increases (i.e. the effect of the NB PU interference on the OFDM subcarriers decreases).
\begin{figure}[!t]
\centering
\includegraphics[width=0.50\textwidth]{Throughput_SNR_20}
\caption{Average throughput as a function of $F_n$ for different SIRs at average SNR = 20 dB in the case of perfect channel estimation.}
\label{fig:Th_Fn}
\end{figure}
Fig. \ref{fig:Th_SNR} depicts the average throughput as a function of average SNR for different values of SIRs at $F_n$ = 0.52, assuming perfect channel estimation. As one can observe, as the average SNR or SIR increases, more bits can be loaded on OFDM subcarriers while achieving the target BER, which translates into an increase in the average throughput. However, the throughput saturates beyond a certain average SNR at a given SIR. This can be explained as, at a certain average SNR value, the OFDM CU subcarriers are loaded with the maximum constellation given a certain SIR value and a further increase in the average SNR will not improve the average throughput. Moreover, as the SIR increases, the effect of the NB PU on the OFDM CU reduces, which translates into an increase in the average throughput. It is worth pointing out that for SIR values of $- 10$ and $- 20$ dB, the OFDM CU subcarriers adjacent to the NB PU are nulled, which in turn reduces the OFDM CU spectrum leakage to the NB PU.
\begin{figure}[!t]
\centering
\includegraphics[width=0.50\textwidth]{Throughput_Fn_052}
\caption{Average throughput as a function of average SNR for different SIRs at $F_n$ = 0.52 in case of perfect channel estimation.}
\label{fig:Th_SNR}
\end{figure}
The effect of imperfect channel estimation on the average throughput is shown in Fig. \ref{fig:Th_ch}. The average throughput is plotted as a function of average SNR for different values of $\sigma_h^2$ for $F_n$ = 0.52 and SIR = 0 dB. As expected, increasing the channel estimation error variance, $\sigma_h^2$, reduces the average throughput. According to (\ref{eq:SINR}), the channel estimation error has the same effect as the interference which can be noticed from Figs. \ref{fig:Th_SNR} and \ref{fig:Th_ch}.
\begin{figure}[!t]
\centering
\includegraphics[width=0.50\textwidth]{Throughput_ch}
\caption{Average throughput as a function of average SNR for different values of $\sigma_h^2$ at $F_n$ = 0.52 and SIR = 0 dB.}
\label{fig:Th_ch}
\end{figure} |
\section{Historical Background}
For sets $A,B \subseteq \mathbb{R}$ we define the algebraic sum $A+B=\{a+b|\, a\in A, \, b \in B\}$. Study of algebraic sums of this kind has been
around for almost a century. The first result in this topic seems to be due to Sierpi\'nski, who proved in 1920 that there exists two sets of measure zero whose sum is non-measurable \cite{Sie}. Rubel \cite{Rub} showed that these two sets can be chosen to be equal. This result was later generalized in many directions. For example, related results for other $\sigma$-ideals were obtained by Kharazishvili \cite{Kha} and by Cicho\'n and Jasi\'nski \cite{C-J}. In another direction, Ciesielski, Fejzi\'c and Freiling \cite{C-F} proved among others, that for every set $C\subseteq \mathbb{R}$, there exists a set $A \subset C$ such that $\lambda_*(A+A)=0$ and $\lambda^*(A+A)=\lambda^*(C+C)$, where $\lambda_*$ and $\lambda^*$ denote the inner and the outer Lebesgue measure respectively (but for simpler proof see the work by Marcin Kysiak \cite{kys}). It is also worth to mention the famous Erd{\"o}s-Kunen-Mauldin theorem \cite{EKM}.
\par It is easy to see that the sum of compact (open) sets is compact (open), the sum of $F_\sigma$ sets is $F_\sigma$, but for higher Borel classes this is not the case \cite{C-J}. Even the sum of a compact set with $G_\delta$ doesn't have to be a Borel set, and this was shown by Sodnomow in 1954 \cite{Sod} and independently by Erd\"os and Stone in 1970 \cite{S-E}.
\par The study of algebraic sums of subsets of real line is closely related to the study of additive subgroups of $(\mathbb{R},+)$. Erd\"os proved, that under CH there exists a non-meagre, null additive subgroup of reals, as well as a non-measurable, meagre additive subgroup of reals \cite{Erd}. The same can be proved under MA, but somehow surprisingly, while non-meagre subgroups of measure zero always exists, some additional set-theoretic assumption turns out to be necessary to prove the existence of a subgroup which is non-measurable and meagre. This was proved recently by Ros{\l}anowski and Shelah \cite{r-s}.
\par Results of this paper were obtained during the work on the author's Master's Thesis. Author would like to thank his advisor, dr. Marcin Kysiak, as well as prof. Witold Marciszewski and prof. Piotr Zakrzewski for many helpful remarks and suggestions.
\section{Preliminaries}
In private communication Sergei Akbarov posed the following problem, sometimes referred to as the Akbarov Problem.
\begin{prob}
Let $A\subseteq \mathbb{R}$ be a nonempty null set. Does there exist a set $B\subseteq \mathbb{R}$ with the property that $A+B$ is Lebesgue non-measurable?
\end{prob}
One of the natural ways to approaching such problem is to try to find a non-measurable dense subgroup $G \le \mathbb{R}$ disjoint with some translation of $A$.
Indeed, assume we have a non-measurable dense subgroup $G \le \mathbb{R}$, and $(A+v) \cap G =\emptyset$ holds. Then also
$$(A+v)\cap (G-G)=\emptyset,$$
$$(A+v+G)\cap G=\emptyset.$$
It is well-known (see for example \cite{buk}, Thm. 7.36) that every dense subgroup of $\mathbb{R}$ is either null or has full outer measure. Both $G$ and $A+v+G$ have full outer measure, hence
both have inner measure zero, and so are non-measurable.
\par This approach, however sufficient to solve the problem with certain additional assumptions like Martin's Axiom, won't work in ZFC alone. In 2016 Andrzej Ros{\l}anowski and Saharon Shelah proved
the following theorem \cite{r-s}.
\begin{thm}
It is relatively consistent with ZFC that any meagre subgroup of reals is null.
\end{thm}
Indeed, consider a dense $G_\delta$ null subset of reals. Any subgroup disjoint with its translation must be meagre. So, consistently, also null.
\par In the case of Baire category however, situation is different. The following is the main result of this paper.
\begin{mainthm}
Let $X=(\mathbb{R},+)$ or $X=(2^\omega, +_2)$. For any meagre set $F \subseteq X$, there exists $x\in X$, and a dense subgroup $H \le X$ without the~Baire property such that $(F+x)\cap H=\emptyset$.
\end{mainthm}
From this follows the affirmative answer to the category version of Problem 1.
\begin{cor}
For any meagre set $A\subseteq X$, there exist a set $B \subseteq X$ such that $A+B$ doesn't have the Baire property.
\end{cor}
\begin{proof}
Just take as $B$ a dense subgroup without the Baire property, which is disjoint with a translation of $A$.
\end{proof}
Another consequence is
\begin{cor}
There exists a null subgroup of $X$ which is not meagre.
\end{cor}
This was firstly proved by Talagrand \cite{tal}, and more recently by Ros{\l}anowski and Shelah \cite{r-s}.
\begin{proof}
Just take $F$ in Theorem 2 of full measure.
\end{proof}
In the whole paper $X$ denote the group of reals $(\mathbb{R},+)$ or the Cantor space $(2^\omega, +_2)$ with coordinate-wise addition modulo 2. Instead of $A+\{x\}$, we write $A+x$. $\mathcal{M}$ and $\mathcal{N}$ denote $\sigma$-ideals of meagre and null sets respectively.
A partition of $\omega$ is always a partition on finite intervals.
\par The following quantifiers are commonly used in the infinite combinatorics:
\begin{itemize}
\item $\forall^\infty_{n<\omega}\; \psi(n)$, denoting ``$\psi$ holds for sufficiently large $n$'';
\item $\exists^\infty_{n<\omega}\; \psi(n)$, denoting ``$\psi$ holds for infinitely many $n$''.
\end{itemize}
Here, one more similar notation will prove useful. Let $\textswab{U}$ be a fixed non-principial ultrafilter on $\omega$. Expression
$$\textswab{U}_{n<\omega} \; \psi(n)$$ will mean
$$\{n| \; \psi(n)\} \in \textswab{U}.$$
$\textswab{U}$ can be seen as something between $\forall^\infty$, and $\exists^\infty$. If $\forall^\infty_{n<\omega}\; \psi(n)$ and $\forall^\infty_{n<\omega}\; \phi(n)$, then clearly
$\forall^\infty_{n<\omega}\; \psi(n) \wedge\phi(n)$, and this is not the case for $\exists^\infty$. On the other hand, for any $\psi$ either $\exists^\infty_{n<\omega} \psi(n)$ or $\exists^\infty_{n<\omega} \neg \psi(n)$, and this is not the case for $\forall^\infty$.
It is straightforward from the definition of a non-principial ultrafilter that for $\textswab{U}$ both mentioned conditions holds.
\par The following combinatorial characterization of meagre sets in $2^\omega$, due to Bartoszy\'nski (for the proof see \cite{b-j}, Thm. 2.2.4), will be crucial in our considerations.
\begin{thm} Every meagre subset of $2^\omega$ is contained in a meagre set \\of the form
$$F=\{x \in 2^\omega |\, \forall^\infty_{n<\omega}\, x \restriction{I_n} \neq v \restriction{I_n}\},$$
where $\{I_n\}_{n<\omega}$ is a partition of $\omega$, and $v\in 2^\omega$.
\end{thm}
\begin{xrem} It is not hard to see, that once we have the partition $\{I_n\}_{n<\omega}$, we can replace it with a ``thicker'' partition, i.e. one in which the end of every interval lies in some fixed infinite subset of $\omega$.
\end{xrem}
\begin{hist} H. Friedman and S. Shelah independently proved, that if a model $V$ results from adding $\omega_2$ Cohen reals to a model of $CH$, then in $V$ the following holds: if $E$ is an $F_{\sigma}$ subset of $X \times X$, which contains a rectangle of positive outer measure, then it contains a rectangle of positive measure. From this, M. Burke \cite{burke} concludes Theorem 1 as a corollary. For another reference, see \cite{pawl}.
\end{hist}
\section{Main theorem}
We turn to the proof of our main theorem. Firstly, we prove it for $X=2^\omega$, and then for $X=\mathbb{R}$, which will turn out to be more complicated. The following lemma was implicitly used in \cite{r-s} to obtain null, non-meagre subgroup of the Cantor Space, but in fact, there's more we can get from it.
\begin{lem} Let $\{I_n\}_{n<\omega}$ be a partition. Then $G=\{x \in 2^\omega | \, \textswab{U}_{n<\omega} \, x \restriction{I_n} \equiv 0 \}$ is a non-meagre dense subgroup of $2^\omega$.\end{lem}
\begin{proof} The fact that $G$ is a group is straightforward from properties of the ultrafilter. It is dense, since every sequence eventually equal 0 is in $G$. The non-trivial part is to show that it doesn't have the Baire property. It is well-known (see for example \cite{buk}, Thm. 7.38) that dense, proper subgroups of $2^\omega$ which have the Baire property are meagre, so it's enough to show that group $G$ is not meagre. Consider any $M \in \mathcal{M}(2^\omega)$. By virtue of Theorem 2 we can assume, possibly enlarging $M$, that
$$M=\{x \in 2^\omega |\, \forall_{k<\omega}^{\infty}\, x \restriction{J_k} \neq v \restriction{J_k} \},$$
for some $v\in 2^\omega$ and a partition $\{J_n\}_{n<\omega}$. Moreover, applying Remark after Theorem 2, we can choose intervals $J_k$ in such a way that each of them is a finite sum of consecutive intervals of the form $I_r$, like below.
$$\overbrace{
\underbrace{0010}_{\text{$I_0$}}\underbrace{010}_{\text{$I_1$}}\underbrace{000111111}_{\text{$I_2$}}
}^{\text{$J_0$}}
\overbrace{
\underbrace{01101010100010}_{\text{$I_3$}}\underbrace{0010000101101}_{\text{$I_4$}}
}^{\text{$J_1$}}
\overbrace{
\underbrace{0001100001000}_{\text{$I_5$}}
\underbrace{11010011}_{\text{$I_6$}}
}^{\text{$J_2$}} \ldots
$$
\par Let now $$A_0=\{n <\omega | \, \exists_{r < \omega}\, I_n \subseteq J_{2r}\},$$ and
$$A_1=\{n <\omega | \, \exists_{r < \omega}\, I_n \subseteq J_{2r+1}\}.$$ One of these sets belongs to the ultrafilter $\textswab{U}$. Suppose it's $A_0$.
Then we put
\begin{equation*}
x \restriction{I_k}=
\begin{cases}
0, & \text{if}\ k \in A_0 \\
v \restriction{I_k}, & \text{if}\ k \in A_1.
\end{cases}
\end{equation*} Similarily, if $A_1 \in \textswab{U}$, we put
\begin{equation*}
x \restriction{I_k}=
\begin{cases}
0, & \text{if}\ k \in A_1 \\
v \restriction{I_k}, & \text{if}\ k \in A_0.
\end{cases}
\end{equation*} In any case, $x$ is constructed in such a way that $\textswab{U}_{n<\omega} \, x \restriction{I_n} \equiv 0$, but also $\exists^\infty_{n<\omega}\; x \restriction{J_n}=v \restriction{J_n}$. In fact, $\{n < \omega | \, x \restriction{J_n}=v\restriction{J_n}\}$ is either the set of even or the set of odd non-negative integers. This means that $x \in G \setminus M$, and since $M$ was arbitrary, this shows that $G \notin \mathcal{M}.$
\end{proof}
\begin{thm}
For any set $F \in \mathcal{M}$, there exist a $x\in X$ and a dense subgroup $H \le X$ without the Baire property such that $(F+x)\cap H=\emptyset$.
\end{thm}
\subsection{Version for $2^\omega$}
Take any $F \in \mathcal{M}(2^\omega)$. Using Theorem 2, we can assume that $$F = \{x \in 2^\omega |\, \forall_{k< \omega}^{\infty}\, x \restriction{I_k} \neq x_F \restriction{I_k} \},$$ where $x_F \in 2^\omega$, and $\{I_n\}_{n<\omega}$ is a partition of $\omega$.
Then, we only have to notice that $\{x \in 2^\omega | \, \textswab{U}_{n<\omega} \, x \restriction{I_n} \equiv 0 \} \cap (F+_2x_F)=\emptyset$ and use Lemma 1. To this end, see that
$$\{x \in 2^\omega | \, \textswab{U}_{n<\omega} \, x \restriction{I_n} \equiv 0 \} \subseteq \{x \in 2^\omega | \, \exists^\infty_{n<\omega} \, x \restriction{I_n} \equiv 0\},$$
and
$$F+_2x_F \subseteq \{x \in 2^\omega | \, \forall^\infty_{n<\omega} \, x \restriction{I_n} \neq 0 \}.$$
\subsection{Version for $\mathbb{R}$}
For any (finite or infinite) binary sequence $x$, we denote by $x^{op}$ the sequence obtained from $x$ by changing every $0$ to $1$, and vice versa. Let $D^{\infty}=\{x \in~ 2^\omega |\, \exists^{\infty}_{n<\omega} \, x(n)=0\}$. It is known that there exists the continuous bijection preserving measure and category, $\phi : D^{\infty} \rightarrow [0,1)$, given by the formula
$$\phi(x)=\sum_{i=0}^\infty{\frac{x(i)}{2^{i+1}}}.$$
\par Consider any $F \in \mathcal{M(\mathbb{R})}$. Replacing if needed $F$ by $F+\mathbb{Z}$, we can assume that $F+\mathbb{Z}=F$. Let $\widetilde{F}=\phi^{-1}[F \cap [0,1)]$. There exists a $w \in 2^{\omega}$ and a partition of $\omega, \; \{I_n\}_{n<\omega}$, with the property that
\begin{equation} \widetilde{F} \subseteq \{x \in 2^{\omega}|\, \forall^{\infty}_{n<\omega}\, x \restriction{I_n} \neq w \restriction{I_n} \}. \end{equation}
In this case, we'll need a combinatorial characterization with some stronger properties. We choose another partition of $\omega$, $\{J_n\}_{n<\omega}$, satisfying the following conditions:
\begin{itemize}
\item each interval $J_n$ is a sum of at least three consecutive intervals $I_k$;
\item if $J_n=I_{r_0} \cup \ldots \cup I_{r_{k}}$, and indices are increasing, then \\
$|I_{r_2} \cup \ldots \cup I_{r_k}| > |I_{r_0} \cup I_{r_1}|;$
\item sequence $\{|J_n|\}_{n<\omega}$ is increasing.
\end{itemize}
We now construct the sequence $v \in 2^{\omega}$. For every $n < \omega$, we write $$J_n=I_{r_0} \cup \ldots \cup I_{r_{k}}$$ as a sum of consecutive intervals of the form $I_k$, with increasing indices, and put:
$$v \restriction{I_{r_0}} = w\restriction{I_{r_0}},$$
$$v\restriction{I_{r_1}}=w^{op}\restriction{I_{r_1}},$$
$$v\restriction{I_{r_2} \cup \ldots \cup I_{r_k}}=010101...01(0).$$
Given (3.1), it is evident that
\begin{equation}\widetilde{F}\subseteq \{x \in 2^{\omega}| \, \forall^{\infty}_{n<\omega}\, x \restriction{J_n} \neq v \restriction{J_n} \}. \end{equation}
Let $J_n=[a_n,b_n]$ for every $n$.
\begin{lem}Assume we have sequences $x,y,z \in D^\infty$ such that $x\restriction{[a,b]} \equiv 0$, $y(b)=0$, and $\phi(z)+\rho=\phi(x) + \phi(y)$, where $\rho \in \{0,1\}$. Then $y\restriction{[a,b-1]}=z\restriction{[a,b-1]}$.
\end{lem}
\begin{proof}
\begin{align*}
\phi(x)+\phi(y) = \sum_{i=0}^{a-1}{\frac{x(i)}{2^{i+1}}}+
\sum_{i=a}^{b}{\frac{x(i)}{2^{i+1}}}+\sum_{i=b+1}^{\infty}{\frac{x(i)}{2^{i+1}}}&+\\
\sum_{i=0}^{a-1}{\frac{y(i)}{2^{i+1}}}+\sum_{i=a}^{b}{\frac{y(i)}{2^{i+1}}}+\sum_{i=b+1}^{\infty}{\frac{y(i)}{2^{i+1}}} &=\\
\sum_{i=0}^{a-1}{\frac{x(i)+y(i)}{2^{i+1}}}+\sum_{i=a}^{b-1}{\frac{y(i)}{2^{i+1}}}+
\sum_{i=b+1}^{\infty}{\frac{x(i)+y(i)}{2^{i+1}}} &=\\ \sum_{i=0}^{\infty}{\frac{z(i)}{2^{i+1}}}+\rho.
\end{align*}
$\displaystyle{\sum_{i=b+1}^{\infty}{\frac{x(i)+y(i)}{2^{i+1}}} < \sum_{i=b+1}^{\infty}{\frac{1}{2^i}}=2^{-b}}$, so this part of the sum does not affect positions of $z$ with indices less than $b$.
\end{proof}
Let us now define
$$\widetilde{G}=\{x \in 2^{\omega} |\, \exists_{m<\omega} \, \textswab{U}_{n<\omega} \, x\restriction{[a_n,b_n-m]}\equiv 0\}.$$
In fact, we didn't rule out the possibility that $m>b_n-a_n$, but recall that $|a_n-~b_n|\rightarrow~ \infty$, so this situation can only occur on finitely many positions. This is clearly a~ subgroup of $2^{\omega}$, and $\widetilde{G}
\supseteq \{x \in 2^\omega | \, \textswab{U}_{n<\omega} \, x \restriction{J_n} \equiv 0 \}$,
which, by Lemma~ 1, is not meagre, so we obtain that $\widetilde{G}$ is not meagre.
\par Group $\widetilde{G}$ defined this way can be ``transferred'' to $\mathbb{R}$, and in fact authors of \cite{r-s} apply this kind of idea.
\begin{lem} $\phi[\widetilde{G}] +\mathbb{Z}$ is a subsemigroup of $\mathbb{R}$.
\end{lem}
\begin{proof}
Assume we have an equality $$\phi(g)+k_1+\phi(h)+k_2=\phi(f)+l,$$ where $g,h \in \widetilde{G}, \, k_1,k_2,l \in \mathbb{Z}$. We shall show that $f \in \widetilde{G}$.
\par There exist integers $m_1,m_2$ such that $$\textswab{U}_n\, g\restriction{[a_n,b_n-m_1]\equiv 0}, \; \textswab{U}_n \, h\restriction{[a_n,b_n-m_2]\equiv 0},$$
so that if we set $\overline{m}$=max$\{m_1,m_2\}$, we obtain
$$\textswab{U}_n\, g\restriction{[a_n,b_n-\overline{m}]} \equiv h\restriction{[a_n,b_n-\overline{m}]} \equiv 0.$$
Then, from Lemma 2 we conclude that $\textswab{U}_n \, f \restriction[a_n,b_n-(\overline{m}+1)]\equiv 0$, which shows that $f \in \widetilde{G}$.
\end{proof}
\begin{lem} $\widetilde{G}+_2v, \, \widetilde{G}+_2v^{op}\subseteq D^\infty$, \\
moreover $\phi[\widetilde{G}+_2v]+\mathbb{Z} \supseteq \phi[\widetilde{G}] + \phi[\widetilde{G}+_2v] + \mathbb{Z}$, \\and $\phi[\widetilde{G}+_2v^{op}]+\mathbb{Z} \supseteq \phi[\widetilde{G}] + \phi[\widetilde{G}+_2v^{op}] + \mathbb{Z}$.
\end{lem}
\begin{proof}
The first part easily follows from the definitions. For the proof of the~~second part, assume we have a $z \in D^\infty$ with the property that $\phi(z)+\rho=\phi(g_1)+\phi(g_2+_2v)$, for some $g_1, g_2 \in \widetilde{G}$ and $\rho \in \{0,1\}.$ We'll show that $z \in \widetilde{G}+_2v$.
\par Let $m_1,m_2 < \omega$ satisfy
$$\textswab{U}_{n<\omega} \, g_i\restriction{[a_n,b_n-m_i]} \equiv 0,$$
for $i=0,1$, and $\overline{m}=\text{max}\{m_1,m_2\}$. Then
$$\textswab{U}_{n<\omega}\, (g_2+_2v)\restriction{[a_n,b_n-\overline{m}]}=v\restriction{[a_n,b_n-\overline{m}]}$$ and
$$\textswab{U}_{n<\omega}\, g_1\restriction{[a_n,b_n-\overline{m}]}\equiv 0.$$
Because $|a_n-b_n| \rightarrow \infty$, for sufficiently large $n$, $\frac{a_n+b_n}{2}\le b_n-\overline{m},$ which means that $b_n-\overline{m}$ lies in the right half of the interval $J_n$. Recall that $v$ was defined in such a way that in the right half of every interval $J_n$, it is of the~form $0101010 \ldots1(0)$. Thus for sufficiently large $n\text{, either} \; v(b_n-\overline{m})=0$ or $v(b_n-(\overline{m}+1))=0$. From Lemma 2 (putting $z=z, x=g_1, y= v$), we infer that
$$\textswab{U}_{n<\omega}\, z\restriction{[a_n,b_n - (\overline{m}+2)]}= v\restriction{[a_n,b_n -(\overline{m}+2)]}.$$
In conclusion $z+_2v\in \widetilde{G}$, which is exactly $z \in \widetilde{G}+_2v$. For $v^{op}$ proof is the same, except that we replace every instance of $v$ with $v^{op}$.
\end{proof}
Notice now, that
\begin{equation}(\widetilde{G}+v) \cap \widetilde{F} = \emptyset. \end{equation}
This is true, because of (3.1) and the inclusion
$$(\widetilde{G}+v) \subseteq \{x \in 2^\omega | \, \exists^\infty_{n<\omega} \, x \restriction{I_n}=w\restriction{I_n}\}.$$
Set $H=\phi[\widetilde{G}]-\phi[\widetilde{G}]+\mathbb{Z}$. From Lemma 3 directly follows that it is a subgroup of $\mathbb{R}$. What's left, is to verify that
$$(\phi[\widetilde{G}]-\phi[\widetilde{G}]+\mathbb{Z}) \cap (F-\phi(v))=\emptyset.$$
Suppose to the contrary, that for some $g,h \in \widetilde{G}, \; k \in \mathbb{Z} \text{ and } f\in F$, an~equality
$$\phi(g)-\phi(h)+\phi(v)+k=f$$
holds.
From Lemma 4 we know, that $\phi(g)+\phi(v)+k=\phi(g'+_2v)+k'; \\ g'\in \widetilde{G},\; k' \in \mathbb{Z},$ so we obtain
\begin{equation}\phi(g'+_2v)-\phi(h)+k'=f.\end{equation}
One can verify that for any sequence $x \in D^\infty$, for which also $x^{op} \in D^\infty$, identity $\phi(x)+\phi(x^{op})=1$ holds. This allows us to write
\begin{equation}\phi(g'+_2v)=1-\phi((g'+_2v)^{op})=1-\phi(g'+_2v^{op}),\end{equation}
because $^{op}$ operation is just the addition of a constant sequence, and so plugging (3.5) into (3.4) yields
\begin{align*}f=1-\phi(g'+_2v^{op})-\phi(h)+k' &=\\
1-[\phi(g'+_2v^{op})+\phi(h)]+k' &=\\
k''+1-\phi(\overline{g}+_2v^{op}) &=\\
k''+\phi(\overline{g}+_2v).
\end{align*}
We used Lemma 4 in the third equality, and $\phi(\overline{g}+_2v)+\phi(\overline{g}+_2v^{op})=1$ in the last one.
Since $F+\mathbb{Z}=F$, we conclude that
$$\phi(\overline{g}+_2v) \in F\cap[0,1)=\phi[\widetilde{F}],$$
$$\overline{g}+_2v \in \widetilde{F},$$
which contradicts (3.3). Finally, let us notice that during the whole calculation arguments of $\phi$ were always within its domain, $D^\infty$.
\section{Open problems}
The proof of Theorem 2 uses the characterization of sets from the $\mathcal{M}(2^\omega)$ ideal, which turns out to match very well with the properties of ultrafilters. There's another class of sets with elegant combinatorial characterization, namely $\mathcal{E}$ -- the $\sigma$-ideal generated by closed sets of measure zero. The following theorem is due to Bartoszy\'nski and Shelah and is proved in \cite{b-j}.
\begin{thm}
$E \in \mathcal{E}(2^\omega)$ if and only if
$$E \subseteq \{x \in 2^\omega | \; \forall^\infty_{n<\omega} \; x \restriction{I_n} \in K_n \},$$
where $\{I_n\}_{n<\omega}$ is a partition of $\omega$, $K_n \subseteq 2^{I_n}$ and $\forall_{n<\omega}\, \frac{|K_n|}{2^{|I_n|}} \le 2^{-n}$.
\end{thm}
The following seems to be a reasonable question.
\begin{prob}
Let $E \in \mathcal{E}(2^\omega)$. Does there necessarily exists a dense non-measurable subgroup $G\le 2^\omega$, disjoint with some translation of $E$?
\end{prob}
Another related question was asked by Taras Banakh on the Mathoverflow.
\begin{prob}
Does there exist (in ZFC) a subgroup $G \le X$, such that \\$G \in \mathcal{N}\cap\mathcal{M}$, but $G \notin \mathcal{E}$?
\end{prob}
|
\section{Introduction}
Finding observational consequences of quantum gravity is obviously a major challenge. In the last decade most attempts have focused on the early Universe, evaporating black holes or Lorentz invariance violation (see \cite{Barrau:2017tcd} for a recent overview). In the last years, the idea that quantum gravity effects could be seen in higher-mass black holes has attracted a lot of interest \cite{Rovelli:2014cta,Haggard:2014rza,Barcelo:2015uff,Barcelo:2017lnx,Haggard:2016ibp}. In particular it was suggested that the quite mysterious Fast Radio Bursts (FRBs) \cite{Katz:2016dti} could be explained by bouncing black holes \cite{Barrau:2014yka}. There are unquestionably simpler astrophysical explanations that we consider to be more probable but this hypothesis is worth a deeper look. At the heuristic and intuitive level, this bounce can be understood as a phenomenon quite similar to what is expected to happen to the Universe in loop quantum cosmology \cite{lqc9,Barrau:2017rwl}. In the cosmological framework, the classically contracting branch is linked to the classically expanding one by a quantum tunneling, whereas in the black hole sector, the classically collapsing solution is glued to the classically exploding one (on the double cover of the Kruskal map \cite{Haggard:2014rza}). The usual event horizon is replaced by a trapping horizon \cite{Ashtekar:2004cn}. In this brief article we revisit this hypothesis by taking into account the fundamental randomness of the tunneling process that was previously ignored. In the first section we assume a peaked mass spectrum for the bouncing black holes and show that the 3 orders of magnitude in energy thought to be missing to explain FRBs can easily be accounted for. In the second section we consider a wide mass spectrum and investigate the sensitivity of the signal to the spectral index. We show that the expected emission remains compatible with measurements and make clear predictions.
\section{Peaked mass spectrum}
The heuristic arguments given by Rovelli, Haggard and Vidotto in the previously mentioned articles suggested that the black hole lifetime could be of the order of $M^2$ in Planck units (those units are used throughout the rest of the article except otherwise stated). As this is shorter that the Hawking evaporation time (of the order of $M^3$) this means that black holes might bounce before they evaporate: the Hawking effect would just be a dissipative correction. An exact calculation of this lifetime is in principle possible in loop quantum gravity (see, {\it e.g.}, \cite{Rovelli:2011eq}), but it is still hard to perform accurately at this stage \cite{Christodoulou:2016vny}. The previous phenomenological works around this hypothesis have focused on gamma-ray bursts \cite{Barrau:2014hda}, FRBs \cite{Barrau:2014yka}, the space-integrated signal \cite{Barrau:2015uca}, and trying to explain the Fermi excess \cite{Barrau:2016fcg}. In all of them the lifetime was taken, as a first approximation, to be deterministic, fixed at the value $\tau=kM^2$ where $k$ was chosen to be of the order of 0.05 (however in one of the studies \cite{Barrau:2015uca} its value was varied). We also assume this value in the present article as it the most phenomenologically interesting one (and the smallest one theoretically allowed). However, as the black-to-white hole transformation is to be understood as a tunneling process, the lifetime of a black hole should be considered as a random variable.\\
The probability that a black hole has not yet bounced after a time $t$ is given by
\begin{equation}
P(t)=\frac{1}{\tau}e^{-\frac{t}{\tau}}.
\end{equation}
This is the usual ``nuclear decay" behavior which comes directly from the fact that the number of bouncing black holes during a time interval $dt$ is proportional to the full number of black holes and to $dt$.
We focus in this study on local effects and neglect the redshift integration as this will play only a minor role in the analysis carried out. The black holes we are interested in can be considered to have been produced in the early universe as the range of masses -- much below a Solar mass - leading to bounces occurring in the contemporary universe can only be associated with primordial black holes (PBHs, see \cite{Carr:2009jm} for a quite recent review on the limits on the PBH abundance and references therein for possible formation mechanisms. In general, the number of black holes of a given type bouncing after a time $t_H$ (taken to be the Hubble time as we are considering phenomena taking place nowadays) in a time interval $dt$ is:
\begin{equation}
dN=\frac{N_0}{kM^2}e^{-\frac{t_H}{kM^2}}dt,
\end{equation}
where $N_0$ is the initial abundance. The exponential function entering this calculation directly comes the random nature of the bounce, as in the previous formula.
Let us assume that the initial differential mass spectrum of the considered PBHs is given by $dN/dM$.\\
In this study, we focus on the so-called bouncing black hole {\it low-energy} component as this is the one that is relevant for a possible link with FRBs. This specific component is based on a simple dimensional analysis : photons are assumed to be emitted with a characteristic wavelength that is of the order of the size of the black hole, which is the only length scale of the problem. As in \cite{Barrau:2016fcg}, we model the shape of the signal emitted by a single black hole by a simple Gaussian function:
\begin{equation}
\frac{dN_{\gamma}^{BH}}{dE}=Ae^{-\frac{(E-E_0)^2}{2\sigma_E^2}},
\end{equation}
where $E_0=1/(2R_S)=1/(4M)$, $R_S$ is the Schwarzschild radius and $M$ is the mass of the considered black hole. This choice is arbitrary and simply taken as an example. The width is typically fixed to be $\sigma_E=0.1E_0$ but the results do not critically depend on this value or the detailed shape of the distribution.\\
The full signal due to a local distribution of bouncing black holes is given by
\begin{equation}
\frac{dN_{\gamma}}{dE}=\int_{M_{Pl}}^{\infty}Ae^{-\frac{(E-E_0)^2}{2\sigma_E^2}}\cdot \frac{dN}{dM}(M)\cdot\frac{1}{kM^2}e^{-\frac{t_H}{kM^2}}.
\end{equation}
The point we want to raise in this study is that the mean energy of the detected signal might {\it not} be the naively expected one, that is may {\it not} be $E\sim1/(4M_{t_H})$ where $M_{t_H}$ is the mass satisfying $t_H=kM_{t_H}^2$ (this would correspond to black holes having a characteristic lifetime equal to the age of the Universe). The naive expectation, $E\sim1/(4M_{t_H})$ is not in the radio band, but rather 3 orders of magnitude higher in energy, in the infrared band. If the initial mass spectrum is peaked around a value $M_0$, {\it e.g.}, according to
\begin{equation}
\frac{dN}{dM}\propto e^{-\frac{(M-M_0)^2}{2\sigma_{\footnotesize{M}}^2}},
\end{equation}
which can in principle be different than $\sqrt{t_H/k}$, the energy will however be peaked around $1/(4M_0)$ which can differ from $1/(4M_{t_H})$. This is possible precisely because of the distributional nature of the actual bouncing time.\\
Considering a peaked mass spectrum is not arbitrary and can be justified if PBHs are created, for example, because of a phase transition in the early Universe (see, {\it e.g.}, \cite{Sobrinho:2016fay}). As the primordial cosmological power spectrum is now clearly known not to be blue \cite{Ade:2015xua} (at least at large scales), the naturally expected density contrast is not high enough to produce PBHs \cite{Carr:1975qj} and specific post-inflationary phenomena are generically required (see, {\it e.g.}, \cite{Jedamzik:1999am}).
\begin{figure}[H]
\includegraphics[width=85mm,center]{Gauss_Norm2.pdf}
\caption{Differential electromagnetic flux emitted by bouncing PBHs for a central mass $M_0$ equal (from right to left) to $M_{t_H}$, $10M_{t_H}$, $100M_{t_H}$, and $1000M_{t_H}$. The normalization is such that the total mass going into PBHs is the same in all cases.}
\label{f2}
\end{figure}
In Fig \ref{f2}, the expected emitted flux is shown for different values of the central mass $M_0$ of the initial mass spectrum: $M_{t_H}$, $10M_{t_H}$, $100M_{t_H}$, and $1000M_{t_H}$. As expected, this shows that the energy of the signal depends on the mass spectrum even if the parameters of the model are fixed. Naturally, when the mass spectrum is peaked at masses well above $M_{t_H}$, the amplitude of the expected signal decreases as BHs that are exploding today constitute an increasingly smaller fraction of the full population. However, the key point we stress here is that a given mean lifetime $\tau=kM^2$ does not imply a fixed expected energy.\\
In particular, it was previously emphasized that the expected mean wavelength (obtained by fixing $\tau=t_H$) of the electromagnetic emission associated with bouncing black holes was basically one thousand times smaller than required to explain the FRBs. If the mass spectrum is peaked at masses higher than $M_{t_H}$, it is however perfectly possible to precisely account for the expected wavelength. The curve on the left in Fig \ref{f2} is peaked around 1.5 GHz, which corresponds to the typical wavelength of FRBs. At this stage, there is no obvious motivation for choosing a specific value for the peak mass. Interesting proposals were recently suggested, for example in the framework of critical Higgs inflation \cite{Ezquiaga:2017fvi}, but as pointed out in the mentioned reference, the actual peak value could differ from the naively calculated one by several orders of magnitude due to accretion and merging and many other models do exist that suggest other mass values.
In Fig \ref{f2} the normalization between the different curves is such that the total mass going into black holes is the same:
\begin{equation}
\int_{M_{Pl}}^{\infty}M\frac{dN}{dM}=cte.
\end{equation}
This is somehow justified if ones tries to account for dark matter with PBHs. The point we want to stress with this remark is simply that the decrease in flux when one moves below the ``natural" mass $M_{T_H}$ is not drastic. Accounting for the observed events by shifting the peaked mass to higher values requires a higher density of PBHs. This cannot be done up to arbitrary values as the upper bounds on the density of PBHs would then be violated. However, orders of magnitude show that the density of PBHs required to account for observed events is very far below the known bounds and this does not limit the present proposal as the rate of FRBs is actually very small \cite{Fialkov:2017qoz}. There is no point is performing a detailed normalization of the expected spectrum at this stage as the initial mass spectrum normalization is totally unknown and the calculation of any observable would directly depend on it.
We have also considered a second normalization, such that the total number of black holes is the same,
\begin{equation}
\int_{M_{Pl}}^{\infty}\frac{dN}{dM}=cte,
\end{equation}
and this basically leads to the very same results.
\\
Beyond FRBs -- which can be explained by astrophysical phenomena -- the point raised here is simply the fact that when the probabilistic nature of the bouncing time is accounted for, the mean energy of the emitted signal is also determined by the mass spectrum and not only by the lifetime of the black holes.
\section{Wide mass spectrum}
It is also possible that the mass spectrum of PBHs is quite wide. As a toy model, if it is directly produced by scale-invariant density perturbations in a perfect fluid with equation of state $w=p/\rho$, the mass spectrum can be approximated by \cite{Carr:1975qj}
\begin{equation}
\frac{dN}{dM} \propto M^{-1-\frac{1+3w}{1+w}}.
\label{spec}
\end{equation}
In this study, we just consider -- as a first approximation -- a spectrum
\begin{equation}
\frac{dN}{dM} \propto M^{\alpha},
\label{spec}
\end{equation}
where $\alpha$ is an unknown parameter. In Fig \ref{f3} we present the expected signal for $\alpha=\{-3,-2,-1,0\}$ (a spectrum rising with an increasing mass on a wide interval would be rather unphysical). Once again, the shape of the mass spectrum does influence the expected signal as the probabilistic nature of the lifetime is now taken into account: black holes with masses smaller or larger than $M_{t_H}$ do also contribute to the emitted radiation and changing their relative weights does change the result.
\begin{figure}[H]
\includegraphics[width=85mm,center]{Phi.pdf}
\caption{Signal expected from a wide mass spectrum, with $\alpha=\{-3,-2,-1,0\}$ from the lower curve to the upper curve at $10^{-6}$ eV.}
\label{f3}
\end{figure}
This leads to another way of addressing the discrepancy between the ``natural" wavelength (around 0.02 cm $\sim2\times10^{-6}$ eV) of bouncing black holes and the observed wavelength (around 20 cm $\sim2\times10^{-3}$ eV) of FRBs. It could indeed be that most bouncing black holes do lead to a signal of wavelength $\sim$ 0.02 cm and that only the tail (which exists because of the probabilistic nature of the lifetime) of the distribution is observed in the radio band. If the peak is in the infrared -- which should occur if the mass spectrum is wide -- it might be that it is simply unobserved today. Detectors in the infrared band have proper time constants that are way to high to allow for the measurement of such fast transient phenomena and there are no deep surveys being carried out.\\
In this case, as shown in Fig \ref{f3}, a clear prediction of this model for future observations is that one should expect a higher flux as the energy increases (up to the infrared band). The slope of this increase reflects that of the mass spectrum. This is qualitatively quite independent of the details of the mass spectrum.
\section{Conclusion}
The possible existence of a black-to-white hole transition through a kind of tunneling process has recently received a lot of attention in quantum gravity. In this brief article we have taken into account the fundamentally random nature of the black hole lifetime in those models. We showed that this can induce a substantial shift with respect to previous studies in which the characteristic lifetime $\tau$ (either derived from the full theory -- first attempts can be found in \cite{Christodoulou:2016vny} -- or inferred by heuristic arguments) was taken as an actual bouncing time.
In a Poisson process, the distribution of time intervals is wide and exponentially decreasing. A bounce can occur after a time which is very different from its characteristic timescale, with the smallest time being always the most probable one. This should be taken into account (and this was indeed accounted for in \cite{Raccanelli:2017xee}).\\
Beyond this quite trivial statement, we have shown that, because of this stochastic process, the mean energy of the emitted signal can be different that previously considered. In particular, if the mass spectrum of PBHs is peaked, it is perfectly possible to match the observed FRBs.
In addition if the mass spectrum of PBHs is wide and continuous it is still possible to explain the data and a prediction was suggested for future observations.\\
The main point of this study was not to revive at any price the hypothesis that FRBs are due to bouncing black holes. Our point was to show that the randomness of the lifetime of black holes in quantum gravity can drastically change the spectral characteristic of the expected signal when the mass spectrum is highly peaked and can lead to interesting predictions in any case.
\section{Acknowledgments}
K.M is supported by a grant from the CFM foundation.
|
\section{Introduction}
\label{sec:intro}
The determination of the exact mass scale of the Standard Model neutrinos is one of the most exciting questions for current research in both particle physics and cosmology. Neutrino oscillation experiments have established the mass square differences between the mass eigenstates \cite{SuperK98,SNO2001,K2K2003,DayaBay}, but are not suited for determining either the absolute mass scale of each eigenstate, or the total mass of the three mass eigenstates combined. The high number density of relic thermal neutrinos from the early Universe (e.g. \cite{Dodelson2003book}), and consequently the non-negligible energy fraction in the neutrinos means that various cosmological observables can be used to constrain the masses of neutrinos.
The most immediate effect of the presence of massive neutrinos is to damp the power spectrum on scales below their free streaming scale, relative to the massless case \cite{Hu1997,Lesgourgues2006}. The amount of damping of the power spectrum is proportional to the total energy density in neutrinos, and therefore the total mass of all the neutrino species. However, the exact shape and the scale at which the damping sets in, are, in principle, sensitive to the individual mass of each eigenstate. Various authors have set bounds on the sum of neutrino masses by looking at, for example, the lensing of the CMB \cite{Planck2013,Planck2015,Sherwin2016,deHaan2016}, the Lyman alpha forest power spectrum \cite{Palanque-Delabrouille:2015pga,Palanque-Delabrouille:2014jca}, combining the autocorrelation of galaxies with galaxy lensing and cosmic shear \cite{Abbott:2017wau,Troxel:2017xyo}, and measurements of Baryon Acoustic Oscillations \cite{Giusarma2016,Vagnozzi:2017ovm,Beutler:2014yhv}. Almost all of these results are based on measuring the small scale damping of the power spectrum. While the effect of neutrinos on the power spectrum is well-understood on scales which are linear, i.e. large scales at low redshifts and progressively smaller scales at higher redshifts, fully describing the effects on small scales requires running cosmological simulations which correctly include the effect of massive neutrinos. Since the useful information in various cosmological surveys is proportional to the number of independent modes, and pushing to smaller scales dramatically increases the number of independent modes in the survey, such studies are extremely useful in tightening the bounds on various cosmological parameters, including the neutrino mass. However, these attempts are complicated by the fact that neutrinos, especially at early times, have very high thermal velocities (see e.g., \cite{Lesgourgues2006}), unlike Cold Dark Matter (CDM).
Various attempts have been made to incorporate the effect of massive neutrinos into the standard N-body simulations used to study structure formation in the $\Lambda$-CDM cosmologies. These can be broadly classified into two approaches. The first approach is to use a linear, or perturbative approach for the neutrinos, but coupling to the non-linear gravitational potential sourced by the Cold Dark Matter component \cite{Brandbyge:2008js,Archidiacono:2015ota,Upadhye:2015lia,AliHaimoud:2012vj,Shoji:2010hm,Inman:2016qmg,Senatore:2017hyk,Saito:2008bp,Dakin:2017idt,Wright:2017dkw}. These approaches are useful at intermediate scales and relatively high redshifts, but break down at late times and small scales, depending on the mass of the individual neutrino species being simulated.
The second approach has been to include neutrinos as an extra set of N-body particles in cosmological simulations \cite{Viel:2010bn,Bird:2011rb,Brandbyge:2009ce,Villaescusa-Navarro:2013pva,Costanzi:2013bha,Castorina:2013wga,Castorina2015,Carbone:2016nzj,Yu:2016yfe,Emberson:2016ecv,RSD2017,Adamek:2017uiq}. In addition to a bulk velocity determined by the power spectrum, each neutrino particle is also given a random thermal velocity by sampling the Fermi-Dirac distribution. This method is fully non-linear, but suffers from Poissonian shot-noise.
The reason is that neutrinos have very large thermal velocities at high-redshifts, which allows them to cross the simulation box multiple times. Because of this, neutrino particles quickly lose memory about their initial conditions and distribute themselves randomly in the box, giving rise to shot-noise in their power spectrum. N-body methods rely on the assumption that the number of particles in a given volume is a faithful representation of the actual physical density in that volume. However, for neutrinos free-streaming with large thermal velocities, this assumption is no longer valid, and the number of neutrino particles in a given volume is a just described by Poissonian statistics. This also means that error on the neutrino power spectrum only improves as $1/N$ with the number of particles used in these simulations. To completely remove shot noise from typical simulations with massive neutrinos, just by increasing the number of particles, one would need roughly need a factor of $10^7$ more particles than used in the largest simulations today! While some methods have been suggested to get rid of the noise in the neutrino power spectrum by sub-sampling the particles to get two independent realizations of the density field (see e.g. \cite{Adamek:2017uiq,Inman:2015pfa}), it does not actually get rid of the inherent noise in field that is used to source the Poisson equation. Therefore new methods are required to test the accuracy of various quantities measured from simulations run using this technique. This is especially true in the context of upcoming cosmological surveys which can potentially pin down the matter power spectrum to the accuracy of $1\%$ \cite{DESI,EUCLID,WFIRST,LSST} at a range of scales.
While \cite{Banerjee2016} has proposed a method to remove the shot noise issue while fully capturing the non-linearities, using a combination of N-body and hydrodynamic techniques, in this paper we describe a completely new approach to remove shot noise from simulations of massive neutrinos, run using the standard N-body technique, by just changing the initial conditions used to initialize these simulations.
The plan of the paper is as follows: we describe our new method for generating initial conditions in \S \ref{sec:method}. In \S \ref{sec:sims}, we describe the details of the cosmological simulations that are used in the paper. In \S \ref{sec:results}, we present results at high redshifts and at $z=0$ to explore the convergence properties and accuracy of our new scheme. We also compare our results to those obtained from simulations with shot noise, as well as various fitting functions which try to capture the effect of massive neutrinos on the matter power spectrum. In \S \ref{sec:sheets}, we explore how this method also allows us to trace shells of different velocities in the neutrino phase space, and allow for a better understanding of the structure of the distribution function of neutrinos in different environments, such as halos and voids. Finally in \S \ref{sec:disc}, we summarize our results, and discuss various future directions of work.
\section{Method}
\label{sec:method}
Before decoupling, neutrinos follow a thermal Fermi-Dirac distribution. Once the neutrinos decouple from the rest of the universe, the distribution gets frozen in, but the momenta of the neutrinos redshift as $1/a$, where $a$ is the cosmological scale factor. As long as the neutrinos were relativistic at the time of decoupling, this is equivalent to having a Fermi-Dirac distribution with the temperature falling as $T\propto 1/a$. Therefore the neutrino distribution at a redshift $z$ is given by
\begin{equation}
\label{eq:fd}
f(\mathbf p) = \frac{4\pi g_\nu}{(2\pi \hbar)^3} \frac 1 {e ^{\frac{pc}{k_{\footnotesize B} T_\nu (1+z)}}+1} \, ,
\end{equation}
where $p = \sqrt{\mathbf p . \mathbf p}$, and $g_\nu$ is the degeneracy (three neutrino species and three anti-neutrino species). $T_\nu(z) \simeq 1.95\,$K is the neutrino temperature today. Our aim is to sample the velocity or momentum space in a regular manner, which is replicated at every point in configuration space - i.e. at every point on the grid used to generate the initial conditions. This sort of approach has been adopted previously in a different context in \cite{1970PhFl...13.1819B}. An inspection of Eq. \ref{eq:fd} shows that choosing to tile the velocity space using a Cartesian grid structure is difficult as the Fermi-Dirac distribution does not factorize along the three directions. On the other hand, the distribution looks much simpler if one considers the magnitude of the momentum, the distribution for which is given simply by $p^2\, f(p)$. Therefore, we choose to first decompose the distribution into intervals of the magnitude of velocity, and then choose angular directions for each velocity magnitude. We discuss the method for each step in detail below.
The reason this method is effective at removing shot noise is that it fixes the magnitudes and directions of the thermal velocities of the neutrino particles. This means that the number of neutrino particles moving into and out of some given volume of the simulation box is not a random number, as is the case when the distribution is sampled randomly. Instead, it is completely determined by the physical power spectrum of the neutrinos, which sets the initial displacements off grid points, and the gravitational evolution of the trajectories. For example, if we decided to initialize the neutrino particles exactly off grid points, i.e. without a physical power spectrum or a completely uniform initial density field, the regular sampling of magnitudes and directions would mean that, in the absence of gravity, at any later time, the density field would still be exactly uniform. Obviously, were we to choose velocity magnitudes and directions randomly, this would not be true, even if we initialized the neutrino particles exactly off the grid points, and gravity is turned off. Note that the ``pairing'' scheme (see for e.g. \cite{Ma:1993xs}), where after the velocity magnitude and direction are chosen at random, two neutrino particles are generated in opposite directions to conserve momentum, would also suffer from the same problem. This is because the neutrino particles generated at adjacent grid points would still have different magnitudes and directions for the velocity.
We now discuss the recipe for dividing the Fermi-Dirac distribution into shells of equal ``mass" or phase space volume. In order to divide the distribution into $N_{\rm shell}$ shells, we construct each shell $i$ such that
\begin{equation}
\label{shell_division}
\frac{\int_{p_{min}^i}^{p_{max}^i} p^2 f(p) \, \dd{p}}{\int_{0}^{p_{max}} p^2 f(p) \, \dd{p}} = \frac 1 {N_{\rm shell}} \, ,
\end{equation}
where $p_{min}^0 = 0$, $p_{min}^i$ and $p_{max}^i$ are the minimum and maximum value of momentum for each bin, and $p_{max}$ is some reasonable value where we can truncate the distribution. In practice, we work in the dimensionless units of $p/T$, which has the additional advantage of making $p_{max}$ independent of the neutrino temperature. We set the magnitude of the velocity associated with each shell in the following manner:
\begin{equation}
\label{eq:shell_vel}
\langle p_i \rangle = \sqrt{\frac{\int_{p_{min}^i}^{p_{max}^i} p^4 f(p) \, \dd{p}}{\int_{p_{min}^i}^{p_{max}^i} p^2 f(p) \, \dd{p}}} \, .
\end{equation}
Eq. \ref{eq:shell_vel} ensures that the velocity dispersion of the neutrinos is set to the correct analytic value.
As we will discuss in detail in \S3, we also need a recipe for dividing up the Fermi-Dirac distribution in a way that allows us to better resolve the slower moving neutrino shells. We do this in the following manner. We choose a weighting function $g(p)$ such that $g(p)$ has higher weight at low $p$, and lower weight at high $p$, compared to $p^2\, f(p)$. For example, in our calculations, we have used $g(p) = p \, f(p)$, and $g(p) = \ln(1+p) f(p)$. Once the function $g(p)$ has been chosen, the different shells are defined in terms of their $p_{\rm min}^i $ and $p_{\rm max}^i$ similar to Eq.\ref{shell_division},
\begin{equation}
\label{shell_division_unequal}
\frac{\int_{p_{min}^i}^{p_{max}^i} g(p) \, \dd{p}}{\int_{0}^{p_{max}} g(p) \, \dd{p}} = \frac 1 {N_{\rm shell}} \, .
\end{equation}
The fractional mass in each shell is calculated in the following manner:
\begin{equation}
\label{shell_mass}
m^i_{\rm shell} = \frac{\int_{p_{min}^i}^{p_{max}^i} p^2 f(p) \, \dd{p}}{\int_{0}^{p_{max}} p^2 f(p) \, \dd{p}} \, .
\end{equation}
So, for this method neutrino particles from different shells have different masses. We can still use Eq. \ref{eq:shell_vel} to determine the magnitude of the velocity corresponding to each shell.
\begin{figure}[tbp]
\centering
\includegraphics[width=.7\textwidth]{fermidirac_new}
\caption{\label{fig:l} Illustration of our method for dividing up the initial Fermi-Dirac distribution for neutrinos. The solid lines represent the boundaries of different momentum bins, while the dashed lines represent the velocity associated with each shell ($\langle p_i\rangle$). On the top, we show the case when all the subdivisions have the same mass or phase-space volume. This has more resolution near the peak of the distribution. At the bottom, we show the subdivisions for the case where $g(p)$ as defined in the text is taken to be $g(p) = p \, f(p)$. This puts more resolution on the slower end of the distribution function. Each sub-division of the distribution corresponds to a ``shell'' in our terminology.}
\end{figure}
Once we have divided up the distribution function into shells in velocity space, we subdivide each of these spherical shells into equal area elements using the \texttt{HEALPIX} algorithm ~\cite{healpix}. At the lowest refinement level, \texttt{HEALPIX} tiles the surface of the sphere with 12 elements. Higher refinement levels are labeled by the variable $N_{\rm side}$, where for a given value of $N_{\rm side}$, the sphere is divided into $12 N_{\rm side}^2$ elements. \texttt{HEALPIX} elements are always symmetric, irrespective of the value of $N_{\rm side}$. However, for low $N_{\rm side}$, especially $N_{\rm side} = 1$ and $N_{\rm side} =2$, there are some residual anisotropies. This comes from the fact that the equatorial tiling and polar tilings are very different, causing the $\hat z$ direction to be different from the $\hat x$ and $\hat y$ directions. For low $N_{\rm side}$, we try to minimize the effects of this anisotropy by rescaling the velocities to ensure that the dispersion along each of the three directions is the same. In summary, therefore, the magnitude of the velocity is chosen by dividing up the Fermi-Dirac distribution; the direction of the velocity is given by the \texttt{HEALPIX} algorithm, and this procedure is repeated for every point on the grid on which the initial conditions are generated. While we have used \texttt{HEALPIX} to generate the velocity directions, other schemes of uniformly dividing the unit sphere \cite{2016arXiv160704590H} would also produce similar effects as those discussed in this work.
\section{Simulations}
\label{sec:sims}
To generate cosmological initial conditions with this method, we modify a version of N-GenIC, that computes displacements and peculiar velocities accountting for the fact that in cosmologies with massive neutrinos the growth factor and the growth rate are scale-dependent \cite[e.g.][]{RSD2017, Raccanelli_2017, Roncarelli_2017, Zennaro:2016nqo, Massara_2015}. The simulations are then run with the Gadget-3 cosmological N-body code, which is a modified version of the publicly available Gadget-2 code \cite{Springel:2005mi}. We note that we turn off the short-range force for neutrinos at early times - turning it on only at redshift $z=9$. The reason we do this at early times is that our method for generating initial conditions produces multiple neutrino particles at the same position. While constructing the tree in Gadget, particles at very close positions are randomized in order to complete the tree construction, leading to artifacts in the simulation. We check that our results are not sensitive to the exact redshift at which we turn on the tree construction for the neutrinos.
We run all our simulations for a comoving box size of $1$ Gpc/h. The cosmological parameters we use are the following: $\Omega_m = 0.3175$, $\Omega_b = 0.049$, $\Omega_\Lambda = 0.6825$, $n_s = 0.9624$, and $h = 0.6711$. Note that in our parameterization $\Omega_\nu$ is always included in $\Omega_m$, i.e. $\Omega_{\rm m}=\Omega_{\rm cdm}+\Omega_{\rm b}+\Omega_\nu$.
In order to generate the initial conditions for high redshift studies in \S \ref{sec:conv}, each point on the initial grid starts $12\times N_{\rm side}^2\times N_{\rm shell}$ neutrino particles, along with one CDM particle. For these studies, all initial conditions were generated with equal mass neutrinos. The particles are displaced off the grids using first order Lagrangian perturbation theory (Zeldovich approximation). The masses of the neutrino particles are adjusted so that they add up to give the correct $\Omega_\nu$ for the simulation box. Of course, this implementation is very expensive in terms of particle numbers, and can only be used to run convergence tests with small numbers of CDM particles. These runs correspond to the first five simulations shown in Table \ref{tab:a}.
For simulations which run to $z=0$, in \S \ref{sec:full_runs}, we follow a different strategy. We generate the initial conditions for the CDM particles from a fine grid, and for the neutrinos from a coarser grid. Therefore, we will have $N_{\rm fine}^3$ CDM particles, and $12\times N_{\rm side}^2 \times N_{\rm shell} \times N_{\rm coarse}^3$ neutrino particles. By playing around with the sizes of the two grids, we can ensure that the total number of CDM particles and neutrino particles are of the same magnitude, as is the case in most simulations run with CDM and neutrino particles. We list the choices of parameters for various runs in Table \ref{tab:a}. For the simulations with massive neutrinos represented in Table \ref{tab:a}, we consider the degenerate neutrino mass scenario, where all the individual mass eigenstates are equal in mass. We do consider one specific case where there is only one massive neutrino species - this case is discussed in \S \ref{sec:results}.
\begin{table}
\begin{center}
\begin{tabular}{| c | c | c | c | c | c | c | c |}
\hline
Name & $\sum m_\nu$ & $N_{\rm CDM}^{1/3}$ & RV & $N_{\rm coarse}$ & $N_{\rm shell}$ & $N_{\rm side}$ & EM \\
\hline
\texttt{E\_Rand\_15} & 0.15eV & 128 & Y & - & - & - & Y \\
\hline
\texttt{E\_Rand2\_15} & 0.15eV & 128 & Y & - & - & - & Y \\
\hline
\texttt{E\_Rand3\_15} & 0.15eV & 128 & Y & - & - & - & Y \\
\hline
\texttt{E\_SH10\_NS1} & 0.15eV & 128 & N & 128 & 10 & 1 & Y \\
\hline
\texttt{E\_SH40\_NS1} & 0.15eV & 128 & N & 128 & 40 & 1 & Y \\
\hline
\texttt{E\_SH10\_NS2} & 0.15eV & 128 & N & 128 & 10 & 2 & Y \\
\hline
\texttt{E\_SH10\_NS4} & 0.15eV & 128 & N & 128 & 10 & 4 & Y \\
\hline
\texttt{E\_SH5\_NS4} & 0.15eV & 128 & N & 128 & 5 & 4 & Y \\
\hline
\texttt{L\_CDM} & 0 eV & 512 & - & - & - & - & - \\
\hline
\texttt{L\_Rand\_15} & 0.15eV & 512 & Y & - & - & - & Y \\
\hline
\texttt{L\_Rand\_30} & 0.30eV & 512 & Y & - & - & - & Y \\
\hline
\texttt{L\_SH5\_NS1} & 0.15eV & 512 & N & 256 & 5 & 1 & Y \\
\hline
\texttt{L\_SH5\_NS2} & 0.15eV & 512 & N & 128 & 5 & 2 & Y \\
\hline
\texttt{L\_SH5\_NS4} & 0.15eV & 512 & N & 128 & 5 & 4 & Y \\
\hline
\texttt{L\_SH10\_NS2} & 0.15eV & 512 & N & 128 & 10 & 2 & Y \\
\hline
\texttt{LU\_SH10\_NS2} & 0.15eV & 512 & N & 128 & 10 & 2 & N (scheme 1) \\
\hline
\texttt{LU2\_SH10\_NS2} & 0.15eV & 512 & N & 128 & 10 & 2 & N (scheme 2) \\
\hline
\texttt{LU\_SH20\_NS2} & 0.15eV & 512 & N & 128 & 20 & 2 & N (scheme 1)\\
\hline
\texttt{LU3\_SH10\_NS2} & 0.30eV & 512 & N & 128 & 10 & 2 & N (scheme 1)\\
\hline
\end{tabular}
\caption{\label{tab:a} A summary of the parameters of various runs. The column labeled ``RV" corresponds to whether the neutrino particles were given random thermal velocities. The column labeled ``EM'' indicates if all the neutrino particles in the simulation had the same masses (Y) or not (N). Scheme 1 corresponds to setting the function $g(p) = p\, f(p)$, while scheme 2 corresponds to setting $g(p) = \ln(1+p) \, f(p)$. Note that the run \texttt{E\_Rand\_15} used the same number of neutrino particles as \texttt{E\_SH10\_NS1}, while \texttt{E\_Rand2\_15} had the same number of neutrino particles as \texttt{E\_SH10\_NS2}, and \texttt{E\_Rand2\_15} used the same of neutrino particles as \texttt{E\_SH10\_NS4}.}
\end{center}
\end{table}
\section{Results}
\label{sec:results}
\subsection{High redshift}
\label{sec:conv}
The shot noise arising from the thermal motion of neutrino particles in simulations is more dominant at early redshifts. This is because at early redshifts, the amplitude of the neutrino power spectrum is small compared to the $1/N$ shot noise power spectrum on almost all scales in the simulation volume. As the thermal velocities of the neutrinos goes down, due to the Universe's expansion, and the physical perturbations grow, the scale at which the noise power spectrum dominates over the physical power spectrum moves to smaller values. Therefore, to display the effectiveness of our initial conditions to get rid of the shot noise in the neutrino power spectrum is best demonstrated by comparing the results of simulations initialized by our method to those from simulations initialized using earlier methods of sampling the thermal velocities of neutrinos. Also, at high redshifts, linear perturbation theory holds at all scales of interest, and therefore we can safely compare the results of our simulations to linear theory predictions from Boltzmann codes like \texttt{CAMB} \cite{CAMB} and \texttt{CLASS} \cite{CLASS1} without worrying about non-linear effects.
\begin{figure}[tbp]
\centering
\includegraphics[width=.6\textwidth]{z49}
\caption{\label{fig:c} Comparison of the neutrino power spectrum at $z=49$ for various choices of $N_{\rm shell}$ and $N_{\rm side}$. We compare the results from our simulations (solid curves) to linear theory (black), as well as a simulations where the Fermi Dirac distribution was sampled randomly for the thermal velocities (dashed curves). For each color, the solid curve and dashed curve used the same number of neutrino particles in the simulations. The curves for \texttt{E\_SH10\_NS1} and \texttt{E\_SH40\_NS1} have been displaced slightly from each other to make them visible on the plot. We find that at early times, the isotropy of the neutrino distribution (higher $N_{\rm side}$) is more important for matching the linear theory prediction for the neutrino power spectrum compared to the number of radial shells ($N_{\rm shell}$).}
\end{figure}
To perform the comparisons, we initialize our simulations at $z=99$, and run up to $z=49$. We explore the case where the sum of the neutrino masses $\sum m_\nu = 0.15$eV, and each individual mass eigenstate has mass $m_i = 0.05$eV. For these light neutrinos, the free streaming scale is larger than the simulation box size of $1$ Gpc/h, implying that the neutrino transfer function is damped compared to the CDM transfer function at all scales.
All of the runs considered in this section were performed with $128^3$ CDM particles, and the same grid size was used to generate the neutrino initial conditions. For a fixed size of the initial conditions grid, there are two parameters which can be changed in our method - the number of radial velocity shells, $N_{\rm shell}$, and the number of \texttt{HEALPIX} pixels, controlled by $N_{\rm side}$. As we increase $N_{\rm shell}$, we sample the radial velocity distribution more finely. Increasing $N_{\rm side}$ results in more more directions in which the neutrinos can move, and results in increasing the isotropy of the distribution.
As pointed out previously, the method in this paper generates multiple neutrino particles per point of the initial conditions grid. To make a fair comparison to the results of simulations initialized using the recipes followed in previous studies, we modify the procedure so that we use the same number of neutrino particles to randomly sample the Fermi-Dirac distribution as are used in runs with initial conditions generated using the new method.
We plot the results of our study in Fig. \ref{fig:c}, where the solid curves are from the ``noiseless" simulations and the dashed curves are from simulations with shot noise. Curves of the same color used the same numbers of neutrino particles. We find that even with increased number of particles, the neutrino power spectrum from the randomly sampled runs (dashed curves) is shot-noise dominated at all scales in the simulation box. This is because the physical power spectrum of the neutrinos at this redshift is smaller than the shot noise floor determined by particle number. We also find that for a fixed $N_{\rm side}$, changing $N_{\rm shell}$ does not have a very big effect, as can be seem by comparing the results from runs \texttt{E\_SH10\_NS1} and \texttt{E\_SH40\_NS1}. The first run uses 10 radial shells, while the second uses 40 radial shells, with $N_{\rm side} =1$ for both. On the other hand, if we fix $N_{\rm shell}$ and increase $N_{\rm side}$, we move closer to the linear theory prediction for the neutrino power spectrum, and for $N_{\rm side} = 4$, we obtain a pretty good match with theory. While there is still some noise in the neutrino power spectrum, we are able to beat down shot noise in the power spectrum by a factor of nearly $10^8$.
These results imply that at early redshifts, ensuring the isotropy of the neutrino distribution (higher $N_{\rm side}$) is more important than sampling the Fermi-Dirac distribution finely (higher $N_{\rm shell}$). This is not very surprising, because at early times, all the radial velocity shells redshift in the same way. It is therefore possible to accurately reconstruct the Fermi-Dirac distribution with a small number of points. This would no longer be true if each velocity shell behaved differently, as will happen at low redshifts, and we explore this behavior in detail in the next section.
\subsection{Low redshift}
\label{sec:full_runs}
We looked at the neutrino power spectrum at high redshifts to demonstrate the ability of our method to eliminate Poisson noise from the simulations. However, this is currently not directly observable, so at low redshifts, we shift our attention to the total matter auto power spectrum, and the neutrino-CDM cross power spectrum. In this section, we perform a study of the two spectra at $z=0$ for various parameter choices. We compare the results from these simulations to simulations which have shot noise in them, as well as to various fitting functions from literature.
To compute the matter power spectrum, we can first define the matter overdensity field in terms of the overdensity fields of the CDM component ($\delta_c$) and the neutrinos ($\delta_\nu$) as
\begin{equation}
\delta_m = f_c \delta_c + f_\nu \delta_\nu \,
\end{equation}
where $f_c$ and $f_\nu$ are the fractions of the total matter density that is in each component. That is,
\begin{equation}
f_c = \frac{\Omega_c}{\Omega_c + \Omega_\nu} \qquad ; \qquad f_\nu = \frac{\Omega_\nu}{\Omega_c + \Omega_\nu} \, .
\end{equation}
The matter power spectrum is then given by
\begin{equation}
P_{mm} = f_c^2 P_{cc} + 2 f_c f_\nu P_{c\nu} + f_\nu^2 P_{\nu\nu} \,,
\label{eq:Pkmm}
\end{equation}
where $P_{cc}$ and $P_{\nu\nu}$ are the auto-spectra of the CDM and neutrino fields, and $P_{c\nu}$ is the cross spectrum.
For the simulations considered here, we generate the initial conditions for the CDM particles from a $512^3$ grid. As in the previous section, all simulations were initialized at $z=99$. For the neutrinos, we explore a range of parameters allowed by computational resources. We vary both $N_{\rm shell}$ and $N_{\rm side}$, as well as the size of the grid on which the initial conditions for the neutrinos are generated. The last is to study the effects of the Nyquist frequency of the neutrino particles on the different observables. We concentrate on $\sum m_\nu=0.15$eV, with individual masses $m_i = 0.05$eV. We also present some results for the case $\sum m_\nu = 0.3$eV, with $m_i = 0.1$eV.
The results of the various runs are summarized in Fig. \ref{fig:f}. We choose the run \texttt{LU\_SH10\_NS2} as our fiducial run. This run has $N_{\rm shell}=10$, $N_{\rm side}=2$, and a $128^3$ grid was used to generate the neutrino initial conditions. This run also uses unequal neutrino particle masses to allow better resolution of the slow-moving shells from the Fermi-Dirac distribution. If we look at the ratio of the matter power spectrum for various runs compared to the fiducial case (the left panel of Fig. \ref{fig:f}), we find that it is remarkably well converged. The run with $N_{\rm side} = 1$, \texttt{L\_SH5\_NS1} is somewhat noisy on large and intermediate scales, but agrees well with the fiducial run on small scales. Since this run (\texttt{L\_SH5\_NS1}) used a $256^3$ grid to generate the neutrino initial conditions, the agreement at larger values of $k$ implies that, at least on the scales we are interested in, increasing the neutrino grid size beyond the fiducial value, does not affect the computation of the matter power spectrum. Increasing $N_{\rm side}$ from $2$ to $4$ produces very little change, with the noise level at about $0.2\%$.
Similarly, we find that as we increase $N_{\rm shell}$ while keeping fixed the other parameters, the ratio of the matter power spectrum barely changes. In terms of the matter power spectrum alone, therefore, as long as $N_{\rm side}$ is larger than $1$, the result is well-converged on all scales. This suggests that, just like at high redshifts, it is important that the neutrino distribution is isotropic to a certain degree. We also point out that for all the runs, the mean of the ratio is very close to $1$, implying that the matter power spectrum is insensitive to whether we use equal mass neutrino or neutrino particles with unequal masses, as is the case with the fiducial run \texttt{LU\_SH10\_NS2}.
\begin{figure}[tbp]
\centering
\includegraphics[width=.475\textwidth]{ratio_matter_convergence}
\includegraphics[width=.475\textwidth]{ratio_cross_convergence}
\caption{\label{fig:f} We look at the convergence of the low redshift results with different number of radial shells into which the Fermi Dirac distribution is divided, and different number of angular directions in which the neutrinos can move initially. The matter power spectrum (left panel) is well converged for most choices of parameters. For the CDM-$\nu$ cross-power spectrum, we find that $N_{\rm shell}$ is important to resolve the full clustering of neutrinos. Having higher resolution on the lower velocity shells helps with the convergence tests. The fiducial simulation used here is \texttt{LU\_SH10\_NS2}.}
\end{figure}
While the matter power spectrum is, of course, the most important observable that is affected by the presence of massive neutrinos, and that can be calculated from these simulations, we also look at the cross-power spectrum between the CDM and the neutrino fields. While this quantity may not be directly observable, it is an important piece to model the total matter distribution and the non-linear coupling between CDM and neutrinos.
The results of this investigation is presented in the right panel of Fig. \ref{fig:f}. As with the matter power spectrum, we present the ratio of the cross spectrum normalized by the cross spectrum from the fiducial run \texttt{LU\_SH10\_NS2}. We find that as $N_{\rm side}$ is increased from $1$ to $2$ to $4$, while keeping $N_{\rm shell}$ fixed at 5, the mean behavior does not change. There is some amount of noise about the mean as we change $N_{\rm side}$.
There is, however, a trend in the mean behavior as we increase $N_{\rm shell}$ for a fixed $N_{\rm side}$. We find that for $k \geq 2 \times 10^{-2} \, h/$Mpc, the magnitude of the cross spectrum increases as we increase the number of radial shells that are used to split up the Fermi-Dirac distribution from $5$ to $10$ - going from \texttt{L\_SH5\_NS2} to \texttt{L\_SH10\_NS2}. While we have not plotted it in the figure to reduce the number of curves, this trend continues as we further increase $N_{\rm shell}$ from $10$ to $20$. This indicates that at late times, the clustering of the neutrinos about the CDM field, as measured by the cross spectrum, is dominated by the slowest moving neutrinos, as expected from theoretical calculations \cite{Massara_2014}, and disproportionate to the phase space ``mass'' in these particles. To clarify, we find this trend in the simulations run with equal neutrino particle masses. For these runs, we divided up the initial Fermi-Dirac distribution such that the different shells covered equal volumes in phase space - i.e. there is more resolution in near the peak of the initial Fermi-Dirac distribution than either at the high-velocity tail or the low-velocity end of the distribution. However, at low redshifts, the slowest moving neutrinos will be the ones whose motion is most affected by the gravitational potential sourced by the CDM field, and they contribute the most to the cross correlation power spectrum that we measure from the simulations.
This motivates us to consider situations where there is more resolution on the low velocity end of the Fermi-Dirac distribution, while also indicating that we do not require much resolution for the higher velocity parts of the distribution. This is exactly what we do for the runs with unequal neutrino particles mass - we use the function $g(p)$ as outlined in \S \ref{sec:method} to ensure that the lower velocity parts of the distribution are populated by larger numbers of particles in the simulations. When we do this, we find that we get convergence in the cross correlation spectrum by using $N_{\rm shell} = 10$ (fiducial run), as the result does not change much when we increase $N_{\rm shell}$ to $20$ (\texttt{LU\_SH20\_NS2}). We also find that the result is not very sensitive to our choice of the function $g(p)$ between $g(p) = p \, f(p)$ and $g(p) = \ln (1+p) \, f(p)$, as seen by inspecting the curve from the simulation \texttt{LU2\_SH10\_NS2}, where the latter choice was used to generate the neutrino particle masses and velocities. The fiducial run used the first choice for $g(p)$ to generate masses and velocities. We further explore the clustering of different parts of the initial Fermi-Dirac distribution of velocities in \S \ref{sec:sheets}.
We have thus demonstrated the convergence of the results of our simulations, in terms of the matter power spectrum and the CDM-neutrino cross-power spectrum, as a function of the free parameters of our method for generating initial conditions. As the next step, we compare the results from these simulations to those produced by the same N-body method, but with previous methods of generating initial conditions. As discussed previously, the older method samples the thermal velocity distribution randomly, which leads directly to the shot noise in the neutrino density field. Comparing the results from the two methods, is therefore a direct way of testing the effects of the shot noise on different observables. To clarify, we will compare the results of the fiducial simulations \texttt{LU\_SH10\_NS2}, and \texttt{LU3\_SH10\_NS2}, to the runs with random velocities. We also point out that the simulations with random velocities will, in general, have lower numbers of neutrino particles compared to the runs mentioned above, since these runs are usually initialized with $1$ neutrino particle per grid point on the initial conditions grid. We also compare two different scenarios - one where $\sum m_\nu = 0.15\,$eV, and the other where $\sum m_\nu = 0.3\,$eV. This helps us understand how the simulations compare as $f_\nu$ is varied. In both cases, we focus on the degenerate scenario, where each mass eigenstate has the same mass.
\begin{figure}[tbp]
\centering
\includegraphics[width=.45\textwidth]{ratio_matter}
\includegraphics[width=.45\textwidth]{ratio_cross}
\caption{\label{fig:a} The left panel shows the ratio of the matter power spectrum from simulations with shot noise to the same quantity from simulations without shot noise. The right hand panel shows the same for the cross-spectrum between CDM and neutrinos. The two cases considered are $\sum m_\nu = 0.15\,$eV and $\sum m_\nu = 0.3\,$eV. We find that simulations with shot noise introduce extra noise in the matter power spectrum on large scales, while the small scales are not very affected. The mean behavior of the cross correlation spectrum is also consistent between the simulations with shot noise and those without.}
\end{figure}
We first compare the ratio of the total matter power spectrum at $z=0$ from the two simulations. We find that the largest differences on the matter power spectrum from the two sets of simulations is on the larger scales, or small wavenumber. The power spectrum of the simulations with shot noise is noisy on large scales, and its magnitude is larger than that expected from cosmic variance.
This is because at high redshifts the neutrino density is completely shot-noise dominated, on all scales in the box. On the other hand, in our method, the large scales do not suffer from any discreteness effects. The magnitude of the noise in the ratio of the matter power spectrum is relatively small, of the order of $1\%$, and increases with $f_\nu$, as illustrated in Fig \ref{fig:a}. We note that achieving sub-percent precision in simulations is mandatory in order to pin point percent-level effects like the ones in \cite{Villaescusa-Navarro:2013pva, Castorina:2013wga, 2017arXiv171001310C}. Interestingly, on small scales, the matter power spectrum from the two sets of simulations agree extremely well, even though these are exactly the scales on which shot noise has the most effect on the neutrino density field. This can be understood from Eq. \ref{eq:Pkmm}. On small scales, neutrinos barely cluster so the dominant term is the CDM auto-power spectrum. The standard method induces a shot-noise term only in the neutrino-power spectrum, as the cross-power spectrum does not suffer from this. The contribution of the neutrino shot-noise to the matter power spectrum is however weighted by $f_\nu^2$, so for realistic neutrino masses this contribution to the matter power spectrum on small scales should be sub-dominant. It is however reassuring that simulations with two different setups, and very different discreteness effects in terms of sampling the neutrino distribution function agree so well for this important quantity.
Next we compare the cross-power spectrum between CDM and neutrinos. We find that for both mass scenarios considered, the fiducial run using the new initial conditions agree well with the cross-spectra from runs with shot noise on all scales of cosmological interest. As can be seen in the right panel of Fig. \ref{fig:a}, the mean behavior matches well, but there is some scatter between the two. Since our fiducial run was chosen such that the cross-correlation spectrum was converged with respect to various parameters of our initial conditions scheme, this implies that the random sampling of the initial distribution is able to capture the correct cross-correlation, which is dominated by the lower velocity end of the initial distribution function. It should also be pointed out that the random sampling also requires fewer particles to capture the correct cross-correlation.
Overall, we conclude that the main effect of shot noise in the neutrino density field is that the matter power spectrum becomes somewhat noisy on large scales. Using the initial condition generation scheme outlined in this paper, therefore, can help reduce the sample variance of simulations with massive neutrinos. On small scales, the effects of shot noise is hardly measurable. As long as the contribution of the neutrinos to the gravitational potential is sub-dominant to CDM, shot noise in the neutrino density field should not be correlated with fluctuations in the CDM field.
This is exactly what we find - the cross-power spectrum of the CDM and neutrinos do not show the effects of shot noise that is clearly seen in the neutrino auto power spectrum.
We now explore how the results from our simulations compare to various fitting functions and related approaches which seek to capture the effect of massive neutrinos on cosmological observables like the nonlinear matter power spectrum. The effects of massive neutrinos on the matter power spectrum was incorporated into the HALOFIT framework (\cite{Smith:2002dz,Takahashi:2012em}) in \cite{Bird:2011rb}. This method modifies the HALOFIT prescription for $\Lambda$CDM cosmologies by including terms proportional to $f_\nu$. It also takes the total matter power spectrum from linear theory as the input, rather than the cold dark matter part only. \cite{Castorina2015} suggested an alternative way to describe the effects of massive neutrinos on the total matter power spectrum. Instead of providing the linear matter power spectrum as an input to HALOFIT, the authors suggested that only the linear CDM power spectrum be used as the input to the original HALOFIT (i.e. the one without a correction for $f_\nu$). The resultant non-linear CDM power spectrum would then be combined with the linear theory predictions for the cross power spectrum of CDM and neutrinos, and the neutrino power spectrum, to yield the final matter power spectrum (see e.g. \cite{Massara_2014}). In particular,
\begin{equation}
P_{mm}^{HF} = f_c^2 P_{cc}^{HF} + 2 f_c f_\nu P_{c\nu}^{L} + f_\nu^2 P_{\nu\nu}^{L} \,,
\end{equation}
Comparing the matter power spectrum at $z=0$ from our simulations to the fitting functions above, we find excellent agreement on large scales, seen in Fig. \ref{fig:d} for both $\sum m_\nu = 0.15$eV and $\sum m_\nu = 0.3$eV. We note that there is very little scatter on large scales, in contrast to simulations with shot noise. There are differences of a few percent on smaller scales, and we find that the prescription in \cite{Castorina2015} provides a slightly better fit out to higher values of $k$.
\begin{figure}[tbp]
\centering
\includegraphics[width=.45\textwidth]{ratio_15}
\includegraphics[width=.45\textwidth]{ratio_3}
\caption{\label{fig:d} Ratio of the fully nonlinear power spectrum at $z=0$, for cosmologies with massive neutrinos to the cosmology with no massive neutrinos. The simulation results are compared to the prediction from HALOFIT modified for massive neutrinos \cite{Bird:2011rb}, and the recipe in \cite{Castorina2015}. In the $\sum m_\nu = 0.15\,$eV case, we find very good agreement on large scales with both fitting formulas, with the latter doing somewhat better down to smaller scales. For $\sum m_\nu =0.3\,$eV, there are slightly larger differences between the results from the simulations and the two fitting formulas on smaller scales.}
\end{figure}
We also compare the results of our simulations to the predictions from the Cosmic Emulator from \cite{Lawrence2017}. For $\sum m_\nu = 0.15$eV and for degenerate neutrino masses, we find that there is a significant difference between the two even on large scales, as shown in Fig. \ref{fig:e}. We find this difference arises from the fact that the emulator only takes $\omega_\nu$ as the input to parameterize neutrinos. $\omega_\nu$, of course, depends on the sum of the neutrino masses, $M_\nu = \sum_i m_\nu^i$, but is insensitive to the individual masses $m_i$. Our first set of simulations were run with three degenerate neutrinos, each with mass $0.05$eV. However, when we change the parameters of our simulations to run for a single massive neutrino species of mass $0.15$eV, we find good agreement between the predictions of the emulator and our simulations, as illustrated in Fig. \ref{fig:e}.
\begin{figure}[tbp]
\centering
\includegraphics[width=.55\textwidth]{ratio_emu}
\caption{\label{fig:e} Comparison of simulation results to the predictions of the Cosmic Emulator in \cite{Lawrence2017}. The emulator takes in only $\omega_\nu$ as an input, and assumes that this is due a single massive species, and is able to match simulations (green) with the same assumptions. However, for the same total mass, the Emulator cannot match the degenerate neutrinos scenario from the simulations (red) even on large scales.}
\end{figure}
This result is interesting, since it tells us that, in principle, we could be sensitive to the individual masses of the neutrino species, along with the total mass. This idea has been explored in literature in the context of determining the neutrino mass hierarchy \cite{DeBernardis:2009di,Jimenez:2010ev}. By comparing the different scenarios in Fig. \ref{fig:e}, we can see that the total amount of damping does depend on the sum of the masses of the neutrinos, though there are small differences in the exact amplitude. However, the scale at which the damping begins, as well as the shape of the damping on intermediate scale depends on the individual masses. This is because the mass of an individual species determines properties like the free streaming scale and transfer functions. For future surveys, where the matter power spectrum will be measured at the percent level, it will be important to explore this systematic carefully both in forecasts, as well as in the analysis. To get to the accuracy levels required for these surveys, the effect of neutrinos cannot then be captured effectively by the variation of one single parameter which only depends on the total mass of the three neutrino species, across the whole range of allowed masses and splittings. For example, if the sum of neutrino masses is close to the minimum value allowed by oscillation experiments ($\sim 0.06\,$eV), the single species is a better approximation than the degenerate mass approximation. On the other hand, if the sum of masses is much higher than the individual mass splittings, then the degenerate mass approximation is accurate, and the single species approximation will be incorrect. Since we do not want to assume, a priori, which regime we are in, the analysis needs to take into account the effects of different mass schemes for the same total mass of neutrinos.
\section{Evolution of velocity shells}
\label{sec:sheets}
We now discuss the evolution of individual neutrino shells in velocity space. A ``shell'' consists of all neutrino particles
from a given radial bin in velocity-space, or all particles with the same velocity magnitude in the initial conditions, as described in \S \ref{sec:method}. As long as the initial particle IDs for the neutrino particles are assigned carefully, it is simple to reconstruct individual shells at any later time. By studying individual neutrino shells, we can separate out how different parts of the initial Fermi-Dirac distribution are distorted by the gravitational potential chiefly determined by the Cold Dark Matter part of the matter distribution.
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\textwidth]{pic_vs_sic.png}
\caption{\label{fig:picvssic} A visual comparison of the total neutrino density at $z=0$, simulated using traditional noisy initial conditions and cloud-in-cell mass deposit (CIC, run \texttt{L\_Rand\_15}, left) versus the initial conditions conditions and simplex-in-cell (SIC, run \texttt{LU\_SH10\_NS2}) techniques presented here (\S \ref{sec:sheets}, \S \ref{sec:sims}). The color bar applies only to the SIC panel; we had to arbitrarily adjust the density range in the CIC panel in order to make the overall density field visible over the noise. The maps are projections through a $1000 \times 1000 \times 8$ Mpc/h slab, normalized to the mean density. We immediately see two major features: First is the much lower noise present in the SIC map, with physical features easily visible. Second is the pattern of ``streaks'' visible at discrete angles in the SIC map. This is a result of the finite number of bulk velocity directions available in this method for IC generation (see \S \ref{sec:method}). Neutrino streams tend to fall into massive structures preferentially along these directions. See Figure \ref{fig:timeevolution} for a map of a single one of these directional bins. We reiterate that the orderly nature of the IC generation described here allows us to uniformly discretize Lagrangian space and apply the SIC technique for interpolating mass between tracer particles.}
\end{figure}
We generate density maps for each neutrino sheet (each velocity direction within a shell, of which there are $12\times N_{\rm side}^2\times N_{\rm shell}$ total) using the simplex-in-cell technique (SIC). By analogy with cloud-in-cell (CIC), this is a method of mapping mass from N-body particles to a Cartesian grid. It has been shown in \cite{Abel2011} that new physical insights can be gained by treating Cold Dark Matter as a 3-dimensional manifold in a 6-dimensional phase space, the ``phase-space sheet'' (PSS) approach. Practically speaking, this scheme demotes N-body particles to Lagrangian flow tracers and interpolates mass between the tracer particles, rather than concentrating mass at particle locations. This gives a smooth density field that is a more realistic representation of the true underlying phase-space structure.
In this application, we connect cubes of 8 N-body particles (in the initial conditions and Lagrangian space) as the mass elements in our post-processing discretization. Each neutrino sheet in the velocity discretization (as discussed in \S \ref{sec:method}) thus consists of a periodic Lagrangian grid of mass-carrying cubes. As described above, the same 8 particle IDs remain associated with the same cube for all time, so that the interpolated mass follows the bulk flow for its neutrino sheet. During the mass remapping step, each cube is refined recursively into tetrahedra using standard trilinear interpolation until each is smaller than $0.1\times\Delta x$, where $\Delta x$ is the grid resolution. Each tetrahedron is then passed to the remapping kernel, which uses the geometric intersection between a homogeneous tetrahedron and the cubical cells of a Cartesian grid to apportion mass into the final density map. This gives a smooth density field which is continuous everywhere and conserves total mass. This allows us to recover continuous density maps on a $1024^3$ grid, even after simulating only $128^3$ particles per neutrino sheet.
The kernel to this algorithm is the geometric remapping technique of \cite{powell2015a}.
The remapping algorithm is designed to map mass from a tetrahedral mass element, in this case with constant mass and density inversely proportional to the volume, to a grid of cubical cells. This is done using the exactly reconstructed polyhedral intersections between each tetrahedral mass element and each cell of the grid. The resulting density field is geometrically precise, mass-conserving, and naturally anti-aliased. The algorithmic details are complex for a robust floating-point implementation; the most recent description is \cite{powell2015b}.
Prior to this work our software package, the \texttt{Phase Sheet Intersector} (PSI), has been applied successfully in several contexts, including analysis of N-body simulations \citep{powell2015a,wojtak2016}, evolution of the Vlasov-Poisson system \citep{sousbie2015, hahn2016}, and the study of cold plasmas \citep{kh2016}.
The PSS approach as previously applied takes advantage of the coldness of the fluid in question, which can be modeled as a 3-manifold embedded in the 6D phase space.
For neutrinos, there is the additional complication that the initial phase space does not form a single 3-dimensional manifold in 6-dimensional phase space. This arises straightforwardly from the fact that the temperatures of the neutrinos is high compared to the bulk motion. However, since neutrinos are essentially collisionless in the context of cosmology, we could decompose the initial 6-dimensional phase space into a set of 3-dimensional manifolds (the ``sheets''), which then interact with each other and to Cold Dark Matter only through gravity. As a final step, the density maps for the individual neutrino sheets are stacked into shells, with each shell containing all sheets of the same initial velocity magnitude.
We plot the density map of neutrinos from a slice through the simulation volume on the right panel in Fig. \ref{fig:picvssic}. On the left panel, we plot the density field from the same volume, but from a simulation in which the neutrinos were initialized with random thermal velocities (\texttt{L\_Rand\_15}). Since the SIC method is not applicable to this sort of initialization, we use the traditional CIC deposition scheme to generate the density maps. The right hand panel exhibits much lower noise compared to the left, and the physical structures are clearly visible. However, since we have used a finite number of velocity directions while initializing the neutrinos, these directions remain visible in the final density map. We note that this is because we have implemented the analog of the ``horizontal streams'' discussed in \cite{kh2016}, where each velocity shell produces its own density field, and the total density field is just a sum over the density fields of all shells. Instead of using horizontal streams, one could use ``vertical streams'', also defined in \cite{kh2016}. In this case, the phase space sheets are connected along the velocity direction - i.e. all particles that started off at the same position but with different velocities now form a connected sheet. To produce the density maps of each sheet, one has to therefore interpolate in velocity space. This should lead to a smoother stacked density map despite having finite number of velocity magnitudes and directions. We plan to investigate this is detail in future work.
In Fig. \ref{fig:timeevolution}, we present the time evolution of density field, but this time using only one of the velocity directions chosen using \texttt{HEALPIX}. To generate this density field, we have summed over the different velocity magnitudes. The velocity direction is shown with the black arrow in the upper left frame. As expected, the clustering in the neutrinos increases with time, as is correlated with the clustering of the CDM component shown in black. Since the neutrinos fall in along one direction, the visible structures at $z=0$ are all elongated along this direction.
\begin{figure}[tbp]
\centering
\includegraphics[width=1.0\textwidth]{time_evolution.png}
\caption{\label{fig:timeevolution} Time-evolution of the neutrino density for a single initial velocity direction; this is the sum of all neutrino shells (velocity magnitude bins) for \texttt{HEALPIX} pixel 9 from run \texttt{LU\_SH10\_NS2} (see \S \ref{sec:method}, \S \ref{sec:sims}). The bulk velocity is shown by the arrow in the upper-left frame. The CDM density is given by the black contours. Thin, thick contours correspond to $\log_{10}(1+\delta_\mathrm{CDM})=0,1$ respectively.
All maps are projections through a $250 \times 250 \times 8$ Mpc/h slab, normalized to the mean density. We use the simplex-in-cell (SIC) method (\S \ref{sec:sheets}) to generate smooth and accurate density maps from Gadget particle data.}
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=.45\textwidth]{dimlesspower_49}
\includegraphics[width=.45\textwidth]{dimlesspower_0}
\caption{\label{fig:b} Dimensionless power of different shells of neutrinos at $z=49$ (left panel) and at $z=0$ (right panel) from the simulation \texttt{LU\_SH10\_NS2}. The slower neutrino shells always have higher power, and their power at small scales (large $k$) is enhanced at late times. The feature in the faster shells on the right panel near $k = 0.8\,$h/Mpc is an artifact of the original grid size. For the slower shells, the physical power is larger than the grid artifact, and hence does not appear for them. Blue corresponds to neutrino shells with slow bulk velocity, and red corresponds to fast-moving shells.}
\end{figure}
Using the density maps, we study the evolution of the dimensionless power ($k^3\,P(k)/(2\pi^3)$) for each shell at different redshifts - we concentrate on $z=49$ and $z=0$ as being representative of high redshift and low redshift behavior, respectively. Even though all the sheets started off with the same initial conditions at $z=99$, the left panel of Fig. \ref{fig:b} shows that, even at high redshifts, the power in each shell is different. As expected, the power in the slower shells is higher at a given redshift when compared to the power in faster shells. This is a direct consequence of the fact that the shells have different thermal velocities, and therefore, a different effective ``sound speed''. However, at such high redshifts, the sound speed of each shell is given roughly by the unperturbed Fermi-Dirac distribution. The sound speed of the shells therefore redshift in the same manner, and the offsets between the different sheets seen in Fig. \ref{fig:b} (left panel) do not grow significantly at high redshifts. This fits well with our finding in \S \ref{sec:conv} that we do not need many neutrino shells at early times to follow the evolution correctly.
At low redshifts, however, we expect more complicated behavior in terms of the evolution of the neutrino shells. The largest scales still evolve in the same way for the different shells. But on small scales, the slower shells are expected to respond more strongly to the gravitational potential which is mainly sourced by CDM, whereas the faster shells are still relatively unaffected. We see this in the right panel of Fig. \ref{fig:b}, where the shape of the power spectrum of the slowest shell is quite different on scales of $k \sim 0.1\,$h/Mpc to $k\sim 1\,$h/Mpc compared to the faster shells.
The effect of the gravitational potential on different shells at low redshift becomes clear when we examine the velocity distribution function (VDF) of the shells from the simulation \texttt{LU\_SH10\_NS2}. When the initial conditions are generated, the distribution function of each shell is a Dirac delta function at the velocities given by Eq. \ref{eq:shell_vel}, as can be seen by the dashed curves in Fig. \ref{fig:vdf}. At $z=0$, the distribution functions are modified, as illustrated by the solid curves. In particular, the distribution functions of all the shells broaden, with the slower shells broadening much more than the faster ones, in agreement with previous works \citep{2013JCAP...03..019V}. It is interesting to note that for the fastest shell, the total amount of broadening about the initial velocity is roughly $10^3\,$km/s, roughly the velocity dispersion of the largest galaxy clusters. The slow shells have a significant fraction of particles which get sped up by many multiples of the initial velocity with which they started.
\begin{figure}[tbp]
\centering
\includegraphics[width=.8\textwidth]{vdfs.pdf}
\caption{\label{fig:vdf} The velocity distribution function of each neutrino shell, at $z=0$ (solid) and the initial conditions (dashed). The initial velocity-space distribution reflects spherical ``shells'' in velocity-space that each sample a different energy range of the initial Fermi-Dirac distribution (see Figure \ref{fig:l}). Blue corresponds to neutrino shells with slow bulk velocity, and red corresponds to fast-moving shells. }
\end{figure}
To understand these high velocity tails, we compute the quantity $\left(\mathbf{v} - \mathbf{v_0}\right).\mathbf{v_0}/\left|\mathbf{v_0}\right|^2$ for particles in every shell, and plot the mass-weighted distribution of this quantity in Fig. \ref{fig:deltavdf}. Here, $\mathbf{v_0}$ represents the original velocity of the neutrino particle, but whose magnitude has been scaled to account for the decay of peculiar velocities with redshift. For neutrino particles which are not captured by the captured by gravitational potentials, but are instead just sped up or slowed down by a small amount along the original direction of motion of the neutrino particles, we expect $\left(\mathbf{v} - \mathbf{v_0}\right).\mathbf{v_0}/\left|\mathbf{v_0}\right|^2$ to be relatively small. This is clearly illustrated by the fastest neutrino shell, which shows the smallest deviation from $0$ in Fig. \ref{fig:deltavdf}. For particles which are actually captured by the gravitational potentials, there are two effects. The first is that the magnitude of the total velocity of the particles can change by an amount much larger than the magnitude of the original velocity of the particle. The second effect is that the direction of the velocity of the particle at $z=0$ is not correlated with the original velocity direction of the particle when the initial conditions were generated. The first effect is illustrated by the large broadening of the distribution for the slower neutrino shells. The second effect shows up as the asymmetry of the distribution function about $0$ - i.e. particles which have been captured have, at some point in their trajectory, changed their direction of motion with respect to the original velocity.
\begin{figure}[tbp]
\centering
\includegraphics[width=.8\textwidth]{deltavdot.pdf}
\caption{\label{fig:deltavdf} The distribution of changes in velocity for each neutrino shell. Note that the low-velocity shells are much more susceptible to broadening of the VDF due to interactions with massive galaxy clusters. }
\end{figure}
\begin{figure}[tbp]
\centering
\includegraphics[width=.8\textwidth]{vdf_vs_rho.pdf}
\caption{\label{fig:vdf_vs_rho} The mass-weighted distribution of neutrinos in the space of neutrino velocities and CDM density at $z=0$. Contours are drawn for the total mass fraction in each sheet \emph{outside} of each contour. The grayscale image is logarithmic in the PDF, and is meant to guide the eye through the tails of the total distribution. Note that the low-velocity shells tend to spread towards regions with higher CDM density.}
\end{figure}
To check that the large change in the velocities of neutrino particles occurs in regions where the CDM density, and therefore the potential, is much larger than the background value, we plot the local CDM density of each neutrino particle vs. its own velocity in Fig. \ref{fig:vdf_vs_rho}. Once again, for the high velocity shells, the local CDM density does not produce a significant change in the equal mass contours plotted in Fig. \ref{fig:vdf_vs_rho}. For the slower shells on the other hand, we find that the shapes of the equal mass contours are very different in underdense vs. overdense regions of CDM. For a given shell, many more of its fastest particles can be found in regions of the simulation volume where the CDM density is high compared to those regions of low density.
This technique therefore enables us to follow the evolution of the distribution function of the neutrinos in different environments in a systematic manner. Since the original thermal velocities for all neutrino particles in the simulation come from a finite number of pre-determined values, it is easy to follow how different parts of the original Fermi-Dirac distribution are modified as result of gravitational interactions with CDM. It is interesting to note that even for neutrino masses as low as $0.05\,$eV for individual species, there is significant clustering of the slower shells in our simulations. We found this in the cross-correlation power spectrum discussed in \S \ref{sec:results}, as well as in the broadening of the distribution function shown in Figs. \ref{fig:vdf}, \ref{fig:deltavdf}, and \ref{fig:vdf_vs_rho}. Studying this clustering of neutrinos can be useful in the context of experiments like PTOLEMY \cite{Betts:2013uya}, which attempt to make direct detections of the cosmic neutrino background.
\section{Summary and Discussion}
\label{sec:disc}
We have presented a new method for generating initial conditions for cosmological simulations with massive neutrinos as an extra set of N-body particles. Previous methods for generating such initial conditions were plagued by Poissonian shot noise arising from the streaming of neutrino particles. The method presented in this paper is able to eliminate this shot noise by choosing appropriate initial conditions. We achieve this by sampling the Fermi-Dirac distribution, from which the thermal velocities of the neutrinos are drawn, in a regular manner, as opposed to randomly sampling it, as was done previously. We divide the Fermi-Dirac into intervals to choose the magnitude of the velocity of neutrinos, and the directions of the neutrino particles are chosen following the \texttt{HEALPIX} algorithm. The crucial point here is that the same values are repeated over all points on the grid on which the initial conditions are generated. This ensures that the number of neutrino particles moving in and out of various regions of the simulation volume is no longer random - this being the source of the Poisson noise seen in previous simulations.
Of course, this method has its own discreteness effects - namely the fact that we choose finite number of points to sample the initial distribution function and a finite number of directions in which we add the thermal velocities to the neutrino particles. We have investigated how the different choices of parameters in our current method affect the high-redshift neutrino power spectrum, and the matter power spectrum and CDM-neutrino cross correlation spectrum at $z=0$. At high redshifts, as long as the initial distribution is roughly isotropic, the simulations can match linear theory predictions for the neutrino power spectrum. The number of shells into which we divide the Fermi-Dirac distribution is not very important in this regime. At low redshifts, while considering the matter power spectrum, the results from the simulations are well-converged with respect to various parameters of our method. However, when we consider the cross spectrum, which measures how correlated the clustering of the neutrinos is with respect to CDM, we find that we need to sample the lower end of the initial Fermi-Dirac distribution well in order to achieve convergence. We find that the slow-moving neutrinos contribute most to the cross correlation, disproportionate to the initial phase space volume that they occupy.
When we compare the results of the simulations run with these initial conditions to simulations with shot noise in the neutrinos, we can essentially isolate the effects of shot noise on various observables. As mentioned earlier, this provides an extremely valuable check on the results of various neutrino simulations run using the N-body method. We note that while the new method does use more neutrino particles than is used currently, the new method is far more effective at removing shot noise than by just increasing the number of particles while using existing methods to assign thermal velocities to neutrino particles. For example, using $\sim 1000^3$ neutrino particles, this method is able to better the shot noise levels of the largest current neutrino simulations \cite{Yu:2016yfe} ($\sim 13000^3$ neutrino particles) by at least three orders of magnitude. For the matter power spectrum, we find that there is virtually no difference between the two sets of simulations on small scales, even though this is exactly where the shot noise issue is most dominant for neutrinos. This confirms that the neutrino power spectrum is so damped on these scales that even large amounts of noise in it will be sub-dominant, and not show up in observables like the matter power spectrum. However, we find that using the ``noiseless'' initial conditions reduces the noise on large scales. This is because in the ``noisy'' simulations, even the largest scales in the simulation box are affected by shot noise at early times, and the effect of this is imprinted onto the late time matter power spectrum. Looking ahead, this reduction in the noise will useful in beating down the sample variance of massive neutrino simulations. In terms of the cross-correlation, we find that the random sampling of the initial Fermi-Dirac distribution is quite effective in being able to follow the clustering of the low end of the velocity distribution.
Given that our results mostly match those from previous simulations, it is not surprising that we also find a close match between our results and various fitting formulas and emulators in literature trying to capture the effects of massive neutrinos on the nonlinear matter power spectrum \cite{Bird:2011rb,Castorina2015,Lawrence2017}. This is especially true at large scales. We also point out that the damping of the power spectrum can be sensitive to what the individual masses of the neutrino mass eigenstates are, in addition to the total mass. Individual neutrino masses set the free streaming mass, which determines the scale at which the damping due to neutrinos starts becoming effective. While this has been known, we point out that in certain situations, the difference between a single massive neutrino species, and the degenerate scenario, for the same total mass, can produce differences which could be measurable in future surveys \cite{DESI,EUCLID,WFIRST,LSST}. This means that using a single parameter, such as the total mass, to capture the effects of neutrino mass in the forecasts and analysis for these surveys may no longer be sufficient.
By suitably ordering the particle IDs in the initial conditions, we can also follow the evolution of the individual ``shells'' into which we divide the Fermi-Dirac distribution. This is done by modifying the ``phase-space sheet'' outlined in \cite{Abel2011} to account for the initial velocity distribution function of neutrinos. We find that even for light neutrino species, there can be significant clustering for the slow shells. This is especially true in regions where the density of CDM itself is large - i.e. inside dark matter halos. This method, therefore, can be extremely useful in understanding the modification of the distribution function of neutrinos as a function of the local environment. It is also useful to study the enhancement of the local neutrino density in dense environments, in the context of experiments which seek to detect the cosmic neutrino background \cite{Betts:2013uya}. In this context, we point out that in this paper, we have used a scheme analogous to the ``horizontal streams'' outlined in \cite{kh2016}. While this scheme already outperforms the existing methods for the generation of initial conditions for neutrino particles, some of the discreteness effects associated with it (such as the discrete initial velocity magnitudes and directions) can be alleviated by implementing ``vertical streams'' scheme outlined in \cite{kh2016}. In that scheme, density reconstructions involve interpolations in velocity space, smoothing out the effects of having a finite number of sample points. We plan to explore these directions in more detail in future studies.
\acknowledgments
We would like to thank Matteo Viel and Volker Springel for permission to use the cosmological code Gadget-3. All simulations for this paper were run on the Sherlock cluster at Stanford University. We used \texttt{GenPK} \cite{2017ascl.soft06006B} extensively to generate different power spectra in this paper. We also used \texttt{pyGadgetReader} \cite{2014ascl.soft11001T} in our analysis pipeline to read different snapshots from simulations. This work was performed in part under DOE Contract DE-AC02-76SF00515 and benefitted from the Stanford Research Computing Center.
|
\section{Introduction}
Recent progresses in different, related fields of science and industry relies on a well-controlled decrease in dimensions during the fabrication of various nanoscaled structures. In the semiconductor industry, this results in the integration of increasingly complex 2D and 3D architectures with feature sizes in the low-nm range\cite{S.Natarajan2014,markov_limits_2014}.
A variety of other applications also benefit directly from the technological progress in the semiconductor industry. Surface-enhanced Raman scattering (SERS) \cite{A.X.Wang2015}, surface-enhanced infrared absorption spectroscopy (SEIRA) \cite{F.Le2008} and advanced light-trapping applications in solar cells \cite{Pala2013,M.L.Brongersma2014} are merely a few examples. Besides periodic lines as the simplest structure models, arrays of periodic nanostructures such as field emitter arrays \cite{Fletcher_2015} and nanorods \cite{Malerba2015} are also of great interest.
The performance of such complex nm-structured devices depends on how well controlled both the dimensional parameters and the 3D elemental compositions are controlled within the nm structures themselves, as well as across large areas of structured substrate. Therefore, two different types of metrological tools that are capable of characterizing these measurands in a non-destructive way are required. These tools must have a high spatial resolution in order to provide insight on the distribution of the atoms within single nm-sized objects or be capable of probing large areas with sufficient sensitivity to the structure geometry that their average dimensional and chemical properties can be characterized. Emerging analytical techniques such as atom probe tomography (APT) provide sufficient lateral and chemical resolution to image single structures \cite{D.J.Larson2016}, but a non-destructive characterization of large numbers of objects by means of APT in order to gain statistical information is not possible. Grazing-incidence X-ray methods such as small-angle X-ray scattering (GISAXS) \cite{levine_grazing-incidence_1989} or grazing-incidence X-ray fluorescence (GIXRF) \cite{D.K.G.DeBoer1995,M.Dialameh2017,JAAS_2012}, are non-destructive ensemble methods and can easily probe larger sample areas with sufficient sensitivity to the dimensional and analytical parameters of the structures \cite{PhysRevLett.94.145504,hofmann_grazing_2009,rueda_grazing-incidence_2012,wernecke_direct_2012-1,gollmer_fabrication_2014,V.Soltwisch2016}. As X-ray fluorescence radiation is element-specific, GIXRF can provide compositional information in addition to scattering methods. When combined with a variation of the incident photon energy, even information on chemical binding sates can be gained by means of fluorescence detected X-ray absorption spectroscopy \cite{BEA_2015}.
In GIXRF, the incident angle $\theta_i$ between the X-ray beam and sample surface is typically varied around the critical angle $\theta_c$ for total external reflection. On flat samples, the interference between the incoming beam and the reflected beam results in an X-ray standing wave (XSW) field \cite{Bedzyk_1989,Golovchenko1982}, which can strongly modulate the intensity distribution above and below the reflecting surface depending on the specific layer structure. The intensity modulation inside the XSW field is correlated with the incident angle and the wavelength and significantly impacts the X-ray fluorescence intensity of an atom depending on its position within the XSW. Performing GIXRF angular scans thus provides information about the in-depth distribution of any probed element within the sample \cite{P.Hoenicke2009}. Employing radiometrically calibrated instrumentation \cite{Beckhoff2008} for reference-free GIXRF \cite{M.Mueller2014}, also provides quantitative information on the elemental mass deposition without the need for any external reference.
To model GIXRF angular profiles, an accurate calculation of the XSW intensity is essential. For a 1D system (for example a stratified layer stack) the recursive matrix formalism developed by Parratt \cite{Parratt1954} is often used and implemented in various software packages such as IMD \cite{Windt1998} and XSWini \cite{Pollakowski_2015}. This formalism is rather fast and is an ideal candidate for layered systems. However, if 2D or even complex 3D structures are present, these software packages are no longer capable of calculating the XSW. For special cases, e.g. for nanostructures with stochastic distribution on a surface, the approach of a stratified layer with reduced density has proven \cite{F.Reinhardt2014, M.Dialameh2017} to be sufficiently precise. Thus for a GIXRF-based characterization of regularly ordered 2D or 3D nanostructures, which are more relevant to fields such as the semiconductor industry, a novel calculation scheme is required for the XSW field, or in general for the near-field distribution.
Maxwell solvers based on the finite-element method (FEM) are suited for the computation of the electric near-field distribution (or in GIXRF terminology, the local excitation condition) within periodic arrangements of surface structures. They can thus contribute to the simulation and interpretation of GIXRF measurement data of such structures in order to derive the dimensional parameters of the structures as well as information about their elemental composition. Similar studies in the optical spectral range have demonstrated the potential of the {finite-element} method \cite{barth_2017}. Expanding this approach to include the X-ray spectral range is challenging due to the fact that the finite-element discretization of the computational domain necessary for this approach depends on the wave length of the incoming plane wave in order to ensure the numerical precision of the calculated electric-field distribution. For incident radiation with wavelengths in the nm or sub-nm range and domain sizes of several 100 nm, this seems to be only possible with a high computational effort, at first glance. But the orientation of the wave vector with respect to the geometrical layout of the sample defines the accessible numerical precision within a reasonable computation time. A more detailed analysis \cite{Soltwisch2017}, however showed that for the special orientation of the incoming wave vector in the GISAXS geometry these requirements relax and the computation becomes feasible.
In this work, we demonstrate the flexibility and potential of the finite-element approach for the characterization of periodic structures using GIXRF and visualize the limitations of the Matrix method and the effective-layer approach. Experimental GIXRF results from a lithographically structured silicon nitride Si$_3$N$_4$ lamellar grating on a silicon substrate were compared to the first reconstruction results obtained with the Maxwell solver based on finite elements. The very good agreement between the measurements and the simulations, and the high sensitivity to relatively small changes in the geometrical layout, indicate the potential of the GIXRF method for the combined analytical and dimensional characterization of such nanostructured surfaces.
\section{Experimental Details}
In this work, a lithographically structured silicon nitride layer on a silicon substrate was used. A Si$_3$N$_4$ lamellar grating was manufactured by means of electron beam lithography at the Helmholtz-Zentrum Berlin. The grating has a nominal pitch of 100 nm, a line height of 90 nm and a line width of 40 nm. The grating areas measure 1 mm by 15 mm, with the grating lines oriented parallel to the long edge. To manufacture the gratings, a silicon substrate with a 90 nm-thick Si$_3$N$_4$ layer was spin coated using ZEP520A the positive resist (organic polymer). The pattern was generated using a Vistec EBPG5000+ e-beam writer, operated with an electron acceleration voltage of 100 kV. After the resist development, the grating was etched into the Si$_3$N$_4$ layer via reactive ion etching using CHF$_3$. Finally, the remaining resist was removed by means of an oxygen plasma treatment. Scanning electron microscopy (SEM) images obtained from witness samples show the high quality of the periodic structured surface (see Fig.~\ref{fig:sample}).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.48\textwidth]{sample_layout.pdf}
\caption{a) Scanning electron microscopy images (SEM) of the Si$_3$N$_4$ lamellar grating. b) SEM cross-section image obtained from a witness sample.}
\label{fig:sample}
\end{figure}
The reference-free GIXRF measurements were carried out at the plane-grating monochromator (PGM) beamline \cite{F.Senf1998} for undulator radiation in the PTB laboratory \cite{B.Beckhoff2009c} of the BESSY II synchrotron radiation facility. This beamline provides soft X-ray radiation of high spectral purity in the photon energy range of 78 eV to 1860 eV. The GIXRF experiments were conducted employing the radiometrically calibrated instrumentation \cite{Beckhoff2008} of PTB and a fundamental parameter-based reference-free quantification approach \cite{M.Mueller2014}. An ultrahigh-vacuum (UHV) chamber \cite{J.Lubeck2013} equipped with a 9-axis manipulator was used for the measurements, allowing the samples to be aligned precisely with respect to all necessary degrees of freedom. Photodiodes mounted on a 2$\theta$ axis allowed simultaneous X-ray reflectometry measurements to be performed and the samples to be aligned with respect to the incoming beam. The incident angle $\theta_i$ between the X-ray beam and the sample surface can be aligned with an uncertainty well below 0.01$^{\circ}$, which is sufficient for the GIXRF experiments. For structured surfaces, the azimuthal incidence angle $\varphi_i$ must also be taken into account. This is the angle between the lines of the grating structure and the plane of incidence defined by the incident beam and the normal to the sample surface, where $\varphi_i = 0^{\circ}$ corresponds to the position where the plane of incidence is parallel to the grating lines (conical). Under the prerequisite that at this position the grating structures be symmetrically with respect to the scattering plane, $\varphi_i$ can be directly aligned with the GIXRF signal to fit the conical mounting direction. The fluorescence radiation emitted is detected using a silicon drift detector (SDD) that has been calibrated with respect to its detector response functions and detection efficiency \cite{F.Scholze2009}. The incident photon flux is monitored by means of calibrated photodiodes.
The reference-free GIXRF experiments on the nanostructured Si$_3$N$_4$ lamellar gratings were performed using an incident photon energy of 520 eV. This photon energy is sufficiently high to excite the N-K${\alpha}$ fluorescence radiation, which serves here as a measure of the XSW intensity within the surface structure. It is also low enough to prevent the excitation of any O-K${\alpha}$ fluorescence radiation, which may complicate the spectral deconvolution of the fluorescence spectra. In addition to an angular variation of the incident angle $\theta_i$, the azimuthal angle $\varphi_i$ was also varied. The X-ray fluorescence spectra recorded for each $\theta_i$ and $\varphi_i$ combination were deconvoluted using detector response functions \cite{F.Scholze2009} for the relevant fluorescence lines.
The atomic fundamental parameters that quantitatively describe the process of absorption of the incident photons and the emission of the fluorescence photons are used together with the known instrumental parameters (e.g., the incident photon flux and the solid angle of detection) in order to also quantify the elemental mass deposition. In this work, we used an experimentally determined value for the nitrogen K-shell fluorescence yield in order to reduce the quantification uncertainty. This value was determined following the procedure described in ref.~\cite{P.Hoenicke2016a}.
\section{Simulation of fluorescence intensities}
\label{sec:fem}
\subsection{Maxwell solver based on the finite-element method}\label{sec_model}
\begin{figure}[tbp]
\centering
\includegraphics[width=0.38\textwidth]{FEM_Layout2.pdf}
\caption{a) The computational domain of a lamellar grating model demonstrating the flexibility of the adaptive finite-element meshing algorithm. b) The computation of the electric field strength in false color scale, visualizing the formation of waveguide modes inside the grating grooves. The sample surface is indicated by a white dotted line.}
\label{fig:fem_mesh}
\end{figure}
X-rays are treated as electromagnetic plane waves of wavelength $\lambda= h c_0 / E_i$ (where $h=$ Planck's constant, $c_0=$ the speed of light and $E_i=$ the photon energy), which scatter on nanostructures. The set of Maxwell's equations can be rewritten as a single, second-order curl-curl equation for the electric field\cite{pomplun_adaptive_2007}. The general idea of the finite element discretization is that the computational domain is subdivided into small patches such as triangles. On these patches, a vectorial ansatz function is usually defined by means of polynomials that have a fixed order. The approximate electric field solution is the superposition of these local ansatz functions. Several software implementations of a Maxwell solver based on the finite-element approach are available. In this study, we use the JCMsuite package~\cite{pomplun_adaptive_2007}, which implements a higher-order finite-element method.
In Fig.~\ref{fig:fem_mesh}, a computational domain for the lamellar grating model (a) and the corresponding near-field simulation (b) used in this work are shown to demonstrate the flexibility of the finite-element meshing algorithm. The electric field distribution in Fig.~\ref{fig:fem_mesh} b) was simulated for incidence angles $\theta_i = 2.4^{\circ}$ and $\varphi_i=0.6^{\circ}$ and for a photon energy of 520 eV. The electric field distribution shows a clear difference between the layer and structured surface models. The appearance of nodes and anti-nodes inside the grating grooves below the critical angle of bulk Si$_3$N$_4$ ($\theta_c\sim 4^{\circ}$) gives rise to a significant surface-shape sensitivity of the integral signal observed. The penetration of the nodes into the grating structure, which increases the measured fluorescence emission, can be adjusted by varying both angles of incidence $\theta_i$ and $\varphi_i$. The finite-element meshing algorithm allows the line shape profile to be varied to form any geometrical layout. We have choosen a model close to the scanning electron microscopy (SEM) cross-section images we obtained from witness samples (see Fig.~\ref{fig:sample} b)). The line height, line width and sidewall angle were parameterized to allow the grating line profile to be easily changed. The line width of the Si$_3$N$_4$ grating is defined at the half-height of the finite-element model in order to eliminate the correlation between the sidewall angle and the structure density. The thickness of potential residual Si$_3$N$_4$ in the grooves is also implemented. To account for possible oxidization of the Si$_3$N$_4$ surface, which is well known from the oxygen plasma cleaning \cite{Kennedy_1999} of the photo resist stripping process, we included an additional SiO$_2$ layer with a homogeneous layer thickness covering the line profile (see Fig. \ref{fig:fem_mesh} a), red area).
\subsection{Layer approach for highly periodic nanostructured surfaces}
The simulation of the electric field distribution for the lamellar grating structure (Fig.~\ref{fig:fem_mesh} b)) indicates that the electric field intensity inside the grating bars, which stimulates the N-K$\alpha$ fluorescence emission, is weak but also directly coupled with both incidence angles. In literature, the so-called effective layer approach \cite{F.Reinhardt2014, M.Dialameh2017} is often used to describe the behavior of a structured surface. The structured surface is assumed to be a stratified layer with a reduced density. This can then be calculated by means of the Matrix method, requiring less numerical effort.
However, the question arises as to whether a layer approach is capable of providing an approximate description of the fluorescence emission measured. In the first step, we thus compare the simulated electric field distribution of a grating effective layer system obtained by means of the Matrix method with that obtained by means of the Maxwell solver. This comparison is shown in Fig.~\ref{fig:XSW} (red and black dashed line) as a function of the grazing incidence angle $\theta_i$. A Si$_3$N$_4$ layer with a 90 nm thickness and half of the Si$_3$N$_4$ density ($\rho_1 = \rho_2 = \rho_{Si_{3}N_{4}}/2$) on top of a silicon substrate represents a grating effective layer of a perfect binary lamellar grating ($\rho_1 = 0, \rho_2 = \rho_{Si_{3}N_{4}}$). As expected, the results of the Maxwell solver computation are in perfect agreement with the results of the Matrix method for 1D layer systems. Both methods are able to deliver an adequate description of the fluorescence intensities measured for the layered systems. However, for the description of highly periodic structured surfaces with a simple layer approach, the effective layer approach and the Matrix method will fail. This is shown in the next step, where we simulate the field intensities of a binary lamellar grating structure (blue line). In contrast to the simulated angular electric field intensity profile of a Si$_3$N$_4$ layer with reduced density, which shows the expected kink only at the critical angle $\theta_c$ of Si$_3$N$_4$, the computation of the binary grating reveals a more complex intensity distribution, including several distinct features.
The difference becomes obvious when the electric field distributions are compared directly for both approaches (see insets of Fig. \ref{fig:XSW}). They show the electric field distribution at an incident angle $\theta_i = 5^\circ$, where large differences can be observed. In the effective layer, the electric field distribution homogeneously penetrates the Si$_3$N$_4$, whereas the electric field is mainly confined to the space between the grating lines in the lamellar grating. From these simulations, it is clear that, for highly correlated systems, the effective layer approach is not able to describe the fluorescence emission measured.
\begin{figure}[tb]
\centering
\includegraphics[width=0.4\textwidth]{layer_effective4.pdf}
\caption{Comparison between the simulation of electric field intensities in the Si$_3$N$_4$ domain from a grating effective layer system (red and black dashed line) with reduced density and the corresponding binary lamellar grating structure (blue line). The integrated field intensities were normalized to the incoming plane wave.}
\label{fig:XSW}
\end{figure}
\subsection{Conversion of simulated electric field distributions into fluorescence intensities}
The deconvolved fluorescence intensities can be used to quantify the amount of material using the Sherman equation \cite{Sherman1955} in an adopted form for GIXRF \cite{P.Hoenicke2009}. The following equation applies to the fluorescence intensity $F(\theta_i,E_i)$ measured for 1D layer systems:
\begin{widetext}
\begin{equation}
F(\theta_i,E_i) = \frac{\Omega}{4\pi\sin\theta_i}\epsilon_{E_f} W_i \rho \omega_k\tau_{E_i} N_0 \int_0^{t_{max}} I(t,\theta_i,E_i) \cdot \exp\left[-t\rho\mu_{E_i}\right] dt\textrm{.}
\label{eq:sherman}
\end{equation}
\end{widetext}
The factor before the integral consists of fundamental, experimental and instrumental parameters. As all instrumental parameters are known due to our calibrated instrumentation \cite{Beckhoff2008,J.Lubeck2013}, we can use the calculated electric near-field intensity distribution or the XSW intensity distribution in GIXRF terminology $I(t,\theta_i,E_i)$ inside the grating structure to extract a numerical approximation of the expected fluorescence intensity $F(\theta_i,E_i)$ per incident photon. For this purpose, we interpolate the square of the absolute values of the computed electric field $|E(x,y)|^2$ distribution inside a specific area to a Cartesian grid $(x,y)$ with sufficient discretization ($dx\times dy=1$ nm$^2$). To account for self-attenuation, every field intensity on this grid is damped with respect to the path length $y_{dis}=(y-y_0) / \cos(\theta_i)$ of the emitted fluorescence photons through the Si$_3$N$_4$ in the direction of the fluorescence detector. Both the solid angle $\Omega/4\pi$ and the detection efficiency for N-K$\alpha$ radiation $\epsilon_{E_f}$ are known for this detector. The density $\rho$ and the attenuation coefficient $\mu_{E_i}$ for Si$_3$N$_4$ at the photon energy of the N-K$\alpha$ fluorescence line are taken from X-raylib\cite{T.Schoonjans2011} for bulk materials.
We thus extend the Sherman equation to include 2D systems and numerically integrate $|E(x,y)|^2$ within the Si$_3$N$_4$ domain:
\begin{widetext}
\begin{equation}
\frac{4\pi\sin\theta_i}{\Omega}\frac{F(\theta_i,E_i)}{N_0\epsilon_{E_f}} =\frac{W_i \rho \tau(E_i) \omega_k}{\sum dx} \cdot \sum_{x}\sum_{y} |E(x,y)|^2 \cdot \exp\left[-\rho\mu_{E_i}y_{dis}\right]\textrm{.}
\end{equation}
\end{widetext}
This reformulation makes it possible to calculate the total emitted N-K$\alpha$ fluorescence intensity from the numerically calculated electrical field distributions and to compare it to the normalized experimental data. The mass fraction $W_i$ of nitrogen in Si$_3$N$_4$, as well as the fundamental parameters $\tau_{E_i}$ as the photo ionization cross section of the N-K shell \cite{T.Schoonjans2011} and $\omega_k$ as the fluorescence yield, are also required. $\omega_k$ was determined experimentally in a manner similar to that described for the O-K shell fluorescence yield in \cite{P.Hoenicke2016a}.
\subsection{Numerical accuracy and discretization size in the X-ray spectral range}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.4\textwidth]{numerical_error_aoi2.pdf}
\caption{Convergence study of the numerical error ($E_i = 520$ eV, $p$=4) for the finite-element discretization length $d$ and three incidence angles $\theta_i$ ($5^{\circ},10^{\circ},15^{\circ}$). The shaded areas represent the 1$\sigma$ uncertainty due different geometrical models (see text). The horizontal dashed line indicates a relative numerical error of 1\%.}
\label{fig:convergence}
\end{figure}
The numerical accuracy of the approximate electric field is a function of the discretization size $d$ of the finite elements and the degree $p$ of the polynomials. The grazing incidence conical diffraction and the invariance of the grating in the direction of the scattering plane results in a standing wave field with much larger periodicity than the wavelength of the exciting radiation. This makes it possible to significantly increase the size of the required discretization length $d$ while still being in line with the conventional rule of half the relevant wavelength for the discretization to ensure numerical accuracy.
In our experiment, the incident wavelength with $\lambda\approx 2.38$ nm (520 eV) combined with the conical scattering geometry and grazing incidence angles $\theta_i$ makes it possible to significantly increase the side lengths $d$ of the finite elements in the simulations. This enables the efficient use of a Maxwell solver based on the finite-element method to treat grazing incidence X-ray applications. Simulations based on a higher excitation energy for the investigation of other materials are also applicable.
This numerical stability is demonstrated in Fig.~\ref{fig:convergence}, which shows a convergence study of the numerical error as a function of the discretization length $d$ (with fixed polynomial order $p=4$) and of the incidence angle $\theta_i$. The absolute value of the relative numerical error $\epsilon_r$ is defined as
\begin{equation}
\epsilon_r (d,p,E,\theta_i,\varphi_i) = \frac{\left|I^{quasi}-I^{model}\right|}{I^{quasi}}\textrm{,}
\end{equation}
and is based on the computations of the integral electric field intensities within the Si$_3$N$_4$ domain $I^{model}$. The quasi-exact calculation $I^{quasi}$ is defined as the computation with the highest achievable numerical precision settings where further tuning of the precision parameters does not influence the results and a numerical stable solution is achieved. By increasing the numerical precision parameters, the $I^{model}$ calculation should converge against the quasi-exact solution. However, the incidence angles $\theta_i$ and $\varphi_i$ have a large impact on the numerical accuracy. At grazing incidence angles $\theta_i=5^{\circ}$ (see Fig.~\ref{fig:convergence}, green line), the numerical error $\epsilon_r$ is well below 10$^{-4}$, even with discretization lengths $d$ up to 10 nm, which is approximately four times the incidence wavelength. With increasing incidence angles $\theta_i=10^{\circ}$ (blue line) and $\theta_i=15^{\circ}$ (red line) in Fig.~\ref{fig:convergence}, the discretization length $d$ must be reduced to ensure a similar numerical precision. This is due to the fact that the spatial modulation frequency increases with increasing incidence angle $\theta_i$\cite{D.K.G.DeBoer1995}.
To account for changes in the numerical convergence due to different geometrical models, we simulated 1000 gratings by means of randomly distributed line shapes. The line height and line width was varied $\pm 10$ nm around the expected nominal values. The expected numerical error distribution \cite{Soltwisch2017} is shown in Fig.~\ref{fig:convergence} as shaded areas that represent the 1$\sigma$ interval. In the following reconstruction of the Si$_3$N$_4$ grating, we set the discretization length $d$ to 7 nm and the polynomial degree to 4, which is sufficient ($\epsilon_r < 1\%$) for the incidence angular range investigated in the GIXRF measurements. A dynamic incident angle-dependent adjustment of the discretization parameters during the simulation of GIXRF angular scans is possible, in principle, in order to further reduce the calculation time. However, for a simple 2D finite-element mesh with a 100 nm$^2$ domain, the computational time is well below 1 s for a single solution and sufficient for these first reconstruction attempts. Moreover, a constant remeshing of the identical structure is also computationally intensive.
\section{Reconstructing a Si$_3$N$_4$ lamellar grating by means of GIXRF}
The N-K$\alpha$ fluorescence intensity of the lamellar grating structure was determined for different combinations of $\theta_i$ and $\varphi_i$. Fig.~\ref{fig:lineplots_grating} a,b) shows the measured fluorescence intensity per incident photon with increasing incidence angle $\theta_i$ (blue dots) for two different azimuthal orientations $\varphi_i$.
\begin{figure}[htb]
\centering
\includegraphics[width=0.42\textwidth]{bestfit_p0_19_recon.pdf}
\caption{(a,b) Comparison between measured N-K$\alpha$ fluorescence intensity and the best reconstruction result for (a) conical mounting $\varphi=0^{\circ}$ and (b) $\varphi=1.9^{\circ}$. Below the critical angle $\theta_c$ of bulk Si$_3$N$_4$, the intensity distribution is strongly modulated (inset in a) and b)). (c,d) Simulation of the $\theta_i$-dependent fluorescence intensity for line widths from 35 nm to 75 nm.}
\label{fig:lineplots_grating}
\end{figure}
At grazing incidence angles ($\theta_i < 5^{\circ}$), the fluorescence intensity emitted exhibits a rather complex modulation (see insets in Fig.~\ref{fig:lineplots_grating}) that can no longer be explained with the grating effective layer model (see also Fig.~\ref{fig:XSW}). The structure-related impact on the intensity distribution is also highly visible if the azimuthal incident angle $\varphi_i$ is shifted from the perfect conical mounting of the grating (cf.~Fig.~\ref{fig:lineplots_grating}a and Fig.~\ref{fig:lineplots_grating}b). Both curves show a completely different intensity modulation at grazing incidence angles below the critical angle $\theta_c$ of total external reflection of the Si$_3$N$_4$ grating. The electric near-field calculation of the grating model (see Fig.~\ref{fig:fem_mesh}b) reveal that the electric field is strongly located inside the grooves until the critical angle $\theta_c$ is reached and the field begins to penetrate the grating bars. This feature provides a high sensitivity to the surface shape for periodically structured surfaces in GIXRF experiments.
In this study, we are not yet able to give a full reconstruction of the line shape model including uncertainties based on the Maxwell solver and the finite-element approach. A statistical evaluation of all model parameters (see Section \ref{sec_model}) is possible, in principle but requires further optimization of the simulation algorithm to reduce the computational effort. In Fig.~\ref{fig:lineplots_grating} a) and b), the red lines represent the best fit obtained with a Monte Carlo method. The starting values for the expected grating model parameters are based on the SEM cross-sections measurements from the witness samples. The 1$\sigma$ uncertainty bands (red shaded areas) in Fig.~\ref{fig:lineplots_grating} are calculated only using the uncertainties from the experimentally determined fluorescence yield $\omega_k$ and the photo ionization cross section $\tau_{E_i}$, as their contributions are expected to be dominant. Their combined uncertainty is about 11 \%, which makes it possible to neglect any other experimental or numerical uncertainty.
Even though we have not performed a full reconstruction of the line shape model including uncertainties, several conclusions can be drawn from the optimization we performed. First, the grating structure is fully etched down to the substrate, thus making the thickness of the Si$_3$N$_4$ in the grooves zero. The method is very sensitive to this parameter, as even a very thin remaining layer imposes large changes on the calculated signal. In Fig.~\ref{fig:lineplots_grating} a,b) the green solid line shows the fluorescence intensity obtained with the best model including an additional 1 nm thick etch residual. Second, an oxidized layer on the surface of the Si$_3$N$_4$ is definitely present. From the optimization performed, we cannot conclude whether it is a pure SiO$_2$ layer or something else that results in a depletion of nitrogen at the surface; However, an SiO$_2$ layer is very likely to exist due to the oxygen plasma cleaning \cite{Kennedy_1999}. The optimized thickness of this layer is 3.5 nm, assuming a box-like depth profile. A gradient profile that is more realistic could be included as well, although this would increase the numerical effort and was thus not taken into account here. However, the sensitivity of the method for this parameter is also very high (see Fig.~\ref{fig:lineplots_grating} a,b) the green dashed line where the thickness of the oxygen layer has been reduced by 1 nm). By additionally measuring the fluorescence emission from oxygen, this could even be improved in the future.
Due to the quantitative modeling enabled by the reference-free GIXRF, we can derive the mass deposition and thus the density of the Si$_3$N$_4$ layer. From the optimization performed, it is determined to be approximately $\rho = 2.8 \frac{g}{cm^2}$, which is 10 \% below literature values from Si$_3$N$_4$ thin films. It should be noted that this is influenced by the simplified model of a well-separated oxidization layer, as most of the nitrogen fluorescence signal measured is generated in close proximity to the surface of the structure (see Fig.~\ref{fig:fem_mesh}b).
The other reconstruction results (line height at $\sim 87$ nm, line width at $\sim 43$ nm and sidewall angle at $\sim 86^{\circ}$) are in good agreement with the expected nominal parameters and the SEM cross-section images obtained from witness samples (see Fig.~\ref{fig:sample}). For these dimensional parameters of the grating, the GIXRF methodology presented is also provides a good sensitivity. This is demonstrated in Fig.~\ref{fig:lineplots_grating} c) and d), where the calculated nitrogen fluorescence angular profile is shown in a false color scale as a function of the line width of the grating for two different azimuthal angles $\varphi_i$. In both geometries, the intensity distribution and the position of the different features in the GIXRF angular distributions is highly correlated with the line shape of the lamellar grating structure.\\
In Fig.~\ref{fig:cdsen}, the full angular measurement dataset (a) is shown in comparison to the simulation (b). Here, the incident angles were varied from $\theta_i = 0^{\circ}$ to $3^{\circ}$ (y-axis) and the azimuthal angle was varied between $\varphi_i = -0.1^{\circ}$ and $1.5^{\circ}$ (x-axis). Due to the symmetry at $\varphi_i = 0^{\circ}$, the experimental data is also valid for negative $\varphi_i$ angles. In Fig.~\ref{fig:cdsen} b), the simulation using the best reconstruction result obtained with the Maxwell solver is shown for the same angular ranges. Several distinct features are visible in both fluorescence maps. The formation of perfect circles around the symmetry axis corresponds to the penetration of the XSW nodes inside the Si$_3$N$_4$ structure. The overall agreement between experimental data and the calculation result is very good for the full angular ranges. The beam divergence was not included in the theoretical simulation in order not to degrade the fine details of the fluorescence map (for example the sharply curved lines that are very similar to higher-order Yoneda lines observed in GISAXS experiments \cite{V.Soltwisch2016}). These details are linked to the periodicity of the nanostructured surface.
\begin{figure*}[htb]
\centering
\includegraphics[width=0.85\textwidth]{bigmaps3.pdf}
\caption{Comparison between the measured fluorescence intensity map (a) of the Si$_3$N$_4$ grating under various incidence angles ($\theta_i,\varphi_i$) and the simulated fluorescence map (b) based on a line shape model with the best reconstruction result.}
\label{fig:cdsen}
\end{figure*}
\section{Conclusions}
In this work, we have shown that a GIXRF-based characterization of regularly ordered, nanostructured surfaces requires a finite element-based calculation scheme in order to model the experimental data. While simulations based on the conventional matrix formalism allow a GIXRF investigation of simple layered systems to take place, this approach will fail if a periodic surface structure such as a grating is present. In this case, the interference due to the periodic arrangement must be taken into account in the model. We have shown that Maxwell solvers based on finite elements are ideally suited for calculating the electric field intensities of any 2D or 3D structure. This allows a GIXRF-based characterization of regularly ordered nanoscale structured surfaces to take place, thus making it an interesting and novel approach with great potential in the different fields of nanotechnology.
We have applied the reference-free GIXRF technique of PTB to a nanoscale lamellar grating consisting of Si$_3$N$_4$ on Si. A finite element-based simulation for both incident angle-dependent intensity distributions within the nanostructures is used to model the experimental GIXRF data, and the dimensional parameters of the grating as well as the elemental distributions are derived. Even though only a rough model of the experimental data has been provided here, we have shown that this technique provides a direct access to the spatial distribution with promising sensitivity for the characterization of these parameters. For the example presented, this sensitivity could even be enhanced by changing to or adding a higher excitation photon energy in order to also gain a fluorescence signal from the oxide layer on the surface. A further improvement of the numerical accuracy is possible by implementing the self-attenuation correction directly to the integration of the finite elements. In addition, the flexibility of the finite elements provides an opportunity to gain deeper insight into the elemental distribution of the nanostructures investigated. For example, it is possible to model 2D structures including complex interdiffusion layers. The GIXRF technique can also be combined with other techniques such as grazing-incidence small-angle X-ray scattering, X-ray reflectometry or resonant X-ray scattering in future experiments to take advantage of their complementary nature \cite{Haase2016}.
In summary, reference-free GIXRF is clearly suitable as a new, non-destructive metrology tool for the dimensional and elemental characterization of nanostructured surfaces. In principle, this technique is also transferable to laboratory scale tools for GIXRF if an appropriate calibration is available and if quantitative information can be derived from the fluorescence intensities measured. As such nanostructures are of rapidly increasing relevance in many fields of applications, the technique presented here is of great interest to the field of nanotechnology.
\section{Conflicts of interest}
There are no conflicts of interest to declare.
\section{Acknowledgements}
Parts of this research were performed within the ‘3D MetChemIT’ EMPIR project. The financial support of the EMPIR program is gratefully acknowledged. Our project is jointly funded by the European Metrology Programme for Innovation and Research (EMPIR) and participating countries within the European Association of National Metrology Institutes (EURAMET) and the European Union.
|
\section{Introduction}
The problem of minimal projections has attracted the attention of approximation theorists for about half a century.
The survey of Cheney and Price \cite{ChePri} still provides a fine account on the topic.
The fact that the Fourier projection is uniquely minimal from $\mathcal{C}(\mathbb{T})$ onto the space of trigonometric polynomials of degree at most $d$,
derived from Berman--Marcinkiewicz formula,
undoubtedly stands as a highlight of the subject, see \cite{Loz,CHMSW}.
But when the focus is put on algebraic rather than trigonometric polynomials,
the situation becomes dramatically more complicated.
Besides the trivial cases of degree $d=0$ and $d=1$,
only the case $d=2$ has been resolved,
albeit at the cost of considerable efforts
deployed by Chalmers and Metcalf \cite{ChaMet}.
In fact, traditional analyses may have reached their limitation for the problem of minimal projections.
The present article is our way of advocating a change of philosophy
to promote the integration in classical Approximation Theory
of modern optimization methods,
in particular methods based on moments and positive polynomials,
see \cite{L}.
Such techniques have already been used for minimal projections \cite{NFAO} and also in the context of constrained approximation \cite{Basc}.
Before presenting the results and insights generated by our approach,
let us formalize the problem of interest precisely.
Throughout the article, we consider an $M$-dimensional subspace $\mathcal{U}$ of the space $\mathcal{C}[-1,1]$ of continuous functions defined on the interval $[-1,1]$.
Typically, the space $\mathcal{U}$ consists of algebraic polynomials.
A projection $P$ from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$ is just a linear map from $\mathcal{C}[-1,1]$ onto~$\mathcal{U}$
that reproduces $\mathcal{U}$,
i.e., that satisfies $P(u) = u$ for all $u \in \mathcal{U}$.
A minimal projection from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$ is a projection $P$ from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$ with minimal norm
\begin{equation}
\label{1}
\|P\|_{\infty \to \infty}
= \max_{ \|f\|_\infty \le 1} \|P(f)\|_\infty
= \max_{x \in [-1,1]} \max_{\|f\|_\infty \le 1} | P(f)(x)|.
\end{equation}
The projection constant $\lambda(\mathcal{U})$ of $\mathcal{U}$ (relative to $\mathcal{C}[-1,1]$) is the value of the minimum, i.e.,
\begin{equation}
\label{Ori1}
\lambda(\mathcal{U}) :=
\min_{P} \; \|P\|_{\infty \to \infty}
\quad
\mbox{s.to $P$ being a projection from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$}.
\end{equation}
Fixing a basis $(u_1,\ldots,u_M)$ for $\mathcal{U}$,
any projection $P$ from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$ can be represented
by some linear functionals $\eta_1,\ldots,\eta_M$ on $\mathcal{C}[-1,1]$
as
\begin{equation}
P(f) = \sum_{m=1}^M \eta_m(f) u_m,
\qquad
\mbox{with}
\quad
\eta_m(u_{m'}) = \delta_{m,m'}, \quad
m,m' \in \ibt{1}{M} .
\end{equation}
Moreover, any bounded linear functional $\eta$ on $\mathcal{C}[-1,1]$ can be represented
by some signed Borel measure $\mu$ on $[-1,1]$ as
\begin{equation}
\eta(f) = \int_{-1}^1 f d\mu.
\end{equation}
It follows easily that the projection constant of $\mathcal{U}$ can also be written as
\begin{equation}
\label{Original}
\lambda(\mathcal{U}) = \inf_{\mu_1,\ldots,\mu_M}
\max_{x \in [-1,1]} \int_{-1}^1 \left| \sum_{m=1}^M u_m(x) d\mu_m \right|
\quad
\mbox{s.to }
\int_{-1}^1 u_{m'} d\mu_{m} = \delta_{m,m'},
\; m,m' \in \ibt{1}{M},
\end{equation}
where the infimum (in fact, minimum) is taken over all signed Borel measures $\mu_1,\ldots,\mu_M$ on $[-1,1]$.
There is of course a difficulty originating from the infinite-dimensionality of the optimization variables $\mu_1,\ldots,\mu_M$,
so the optimization program \eqref{Original} cannot (a priori) be performed exactly.
Our strategy consists in producing computable upper bounds
(see Section \ref{SecUB})
and lower bounds (see Section~\ref{SecLB})
that are sufficiently close to determine the value of the projection constant with, say, four or five digits of accuracy
(corresponding here to three or four digits after the decimal point).
The computational burden can be lighten by exploiting some symmetry properties of minimal projections (see Section~\ref{SecSym}).
Implementing our strategy enables us, for instance, to:\vspace{-5mm}
\begin{itemize}
\item retrieve numerically the result of \cite{ChaMet} for quadratic polynomials, namely
\begin{equation}
\lambda(\mathcal{P}_2) \approx 1.2201,
\end{equation}
and allude\footnote{\label{FN}We are not being more assertive here because, strictly speaking, our computed minimal projection is not an exact minimal projection.} to the nonuniqueness of minimal projections onto the quadratics;
\item determine with five digits of accuracy the projection constants of several other three-dimensional polynomial spaces, and note in passing that
\begin{equation}
\lambda({\rm span}\{1,x^2,x^3\}) = 1;
\end{equation}
\item determine with four digits of accuracy the projection constants of the spaces of cubic, quatric, and quintic polynomials, namely
\begin{align}
\lambda(\mathcal{P}_3) &\approx 1.365,\\
\lambda(\mathcal{P}_4) & \approx 1.459,\\
\lambda(\mathcal{P}_5) & \approx 1.538,
\end{align}
and hint\footnote{Same reservation as in footnote \ref{FN}.} that minimal projections onto $\mathcal{P}_d$ do not in general preserve $d$-convexity,
thus \mbox{disproving} a conjecture from \cite{PCM};
\item provide state-of-the-art upper and lower bounds for the projection constants of the spaces of polynomials of degree at most $d$ until $d = 12$.
\end{itemize}\vspace{-5mm}
All of our results can be reproduced by downloading the {\sc matlab} code available on the authors' webpages.
The packages {\sf CVX} and {\sf Chebfun} are required to execute the code.
\section{Computable upper bound}
\label{SecUB}
We present in this section a discretization of the problem \eqref{Original} that leads to a computable upper bound for the projection constant $\lambda(\mathcal{U})$ of a polynomial subspace $\mathcal{U}$ of $\mathcal{C}[-1,1]$.
This involves in fact two discretizations:
one for the signed Borel measures $\mu_1,\ldots,\mu_M$
and one for the interval $[-1,1]$.
We start by discretizing the measures.
\begin{prop}
\label{PropUB1}
Given a basis $(u_1,\ldots,u_M)$ for a subspace $\mathcal{U}$ of $\mathcal{C}[-1,1]$,
if $v_1,\ldots,v_K$ are distinct points in $[-1,1]$,
then
\begin{equation}
\label{UB1}
\lambda(\mathcal{U}) \le \min_{A \in \mathbb{R}^{M \times K}}
\max_{x \in [-1,1]}
\sum_{k=1}^K \left| \sum_{m=1}^M A_{m,k} u_m(x) \right|
\;
\mbox{s.to }
\sum_{k=1}^K A_{m,k} u_{m'}(v_k) = \delta_{m,m'},
\, m,m' \in \ibt{1}{M} .
\end{equation}
\end{prop}
\begin{proof}
We obtain an upper bound for $\lambda(\mathcal{U})$
whenever the minimization in \eqref{Original} is carried over a subset of all signed Borel measures $\mu_1,\ldots,\mu_M$.
In particular, we choose here the subset of all linear combinations of Dirac measures at $v_1,\ldots,v_K$,
i.e., measures of the form
\begin{equation}
\mu_m = \sum_{k=1}^K A_{m,k} \delta_{v_k},
\qquad m \in \ibt{1}{M}.
\end{equation}
Under this restriction, the duality constraints in \eqref{Original} read,
for all $m,m' \in \ibt{1}{M}$,
\begin{equation}
\label{Exp1}
\delta_{m,m'} =
\int_{-1}^1 u_{m'} d\left( \sum_{k=1}^K A_{m,k} \delta_{v_k} \right)
=\sum_{k=1}^K A_{m,k} \int_{-1}^1 u_{m'} d\delta_{v_k}
= \sum_{k=1}^K A_{m,k} u_{m'}(v_k),
\end{equation}
while the integral appearing in the objective function becomes,
for a fixed $x \in [-1,1]$,
\begin{align}
\label{Exp2}
\int_{-1}^1 \left| \sum_{m=1}^M u_m(x) d \left( \sum_{k=1}^K A_{m,k} \delta_{v_k} \right) \right|
& = \int_{-1}^1 \left| \sum_{k=1}^K \left( \sum_{m=1}^M A_{m,k} u_m(x) \right) d\delta_{v_k} \right| \\
\nonumber
& = \int_{-1}^{1} \sum_{k=1}^K \left| \sum_{m=1}^M A_{m,k} u_m(x) \right| d\delta_{v_k}
= \sum_{k=1}^K \left| \sum_{m=1}^M A_{m,k} u_m(x) \right|.
\end{align}
Taking the expressions \eqref{Exp1} and \eqref{Exp2} into account yields the upper bound \eqref{UB1}.
\end{proof}
Proposition \ref{PropUB1} is not directly exploitable
due to the presence of the maximum over the interval $[-1,1]$.
We can replace it by a maximum over a discretized grid,
as long as we are able to bound the maximum over $[-1,1]$ by the maximum over this grid.
For polynomials,
the comparison between the discrete and continuous max-norms is a well-studied topic,
especially for equispaced points (see e.g.
\cite{EZ,CR,Rak}).
But equispaced points are not the most suitable,
since the two norms are comparable when the number of points scales quadratically with the degree.
In contrast,
for zeros of Chebyshev polynomials,
the number of points only needs to scale linearly with the degree.
Here is a quantitative version of this assertion,
which can be found in \cite{EZ}.
\begin{lem}
\label{LemDiscVsCont}
Let $w_1> \cdots > w_L \in [-1,1]$ be the Chebyshev zeros given by $w_\ell = \cos(\theta_\ell)$, where $\theta_\ell = \pi(\ell-1/2)/L$.
For any algebraic polynomial $p$ of degree at most $d$, one has
\begin{equation}
\label{DiscVsCont}
\max_{x \in [-1,1]}|p(x)| \le \cos\left(\dfrac{\pi}{2} \dfrac{d}{L} \right)^{-1} \max_{\ell \in \llbracket 1:L \rrbracket} |p(w_\ell)|.
\end{equation}
\end{lem}
We are now in a position to derive the awaited computable upper bound for the projection constant.
\begin{prop}
\label{PropUB}
Given a basis $(u_1,\ldots,u_M)$ for a subspace $\mathcal{U}$ of $\mathcal{C}[-1,1]$ consisting of polynomials of degree at most $d$,
let $v_1,\ldots,v_K$ be distinct points in $[-1,1]$.
With $ w_1 > \cdots > w_L$ denoting the Chebyshev zeros
$w_\ell = \cos(\pi (\ell-1/2)/L)$ and with $\rho = \cos\left( (\pi d)/(2 L) \right)^{-1} \ge 1$, one has
\begin{equation}
\label{DD}
\lambda(\mathcal{U})
\le \rho \times
\hspace{-1.5mm} \min_{A \in \mathbb{R}^{M \times K}}
\max_{\ell \in \llbracket 1 : L \rrbracket }
\sum_{k=1}^K \left| \sum_{m=1}^M \hspace{-.5mm}A_{m,k} u_m(w_\ell) \right|
\;
\mbox{s.to }
\sum_{k=1}^K \hspace{-.5mm}A_{m,k} u_{m'}(v_k) = \delta_{m,m'},
\, m,m' \hspace{-.5mm} \in \hspace{-.5mm} \ibt{1}{M}.
\end{equation}
\end{prop}
\begin{proof}
According to Proposition \ref{PropUB1}, it suffices to show that, for any $A \in \mathbb{R}^{M \times K}$,
\begin{equation}
\label{ObjUB2}
\max_{x \in [-1,1]}
\sum_{k=1}^K \left| \sum_{m=1}^M A_{m,k} u_m(x) \right|
\le \rho
\max_{\ell \in \llbracket 1: L \rrbracket}
\sum_{k=1}^K \left| \sum_{m=1}^M A_{m,k} u_m(w_\ell) \right|.
\end{equation}
For a fixed $x \in [-1,1]$,
we can find signs $\varepsilon_1,\ldots,\varepsilon_K \in \{\pm 1\}$ such that
\begin{align}
\label{CompAbs}
\sum_{k=1}^K \left| \sum_{m=1}^M A_{m,k} u_m(x) \right|
& = \left| \sum_{k=1}^K \varepsilon_k \sum_{m=1}^M A_{m,k} u_m(x) \right|
\le \rho
\max_{\ell \in \llbracket 1: L \rrbracket}
\left| \sum_{k=1}^K \varepsilon_k \sum_{m=1}^M A_{m,k} u_m(w_\ell) \right|\\
\nonumber & \le \rho
\max_{\ell \in \llbracket 1:L \rrbracket}
\sum_{k=1}^K \left| \sum_{m=1}^M A_{m,k} u_m(w_\ell) \right|,
\end{align}
where Lemma \ref{LemDiscVsCont} was used for the first inequality in \eqref{CompAbs}.
Taking the maximum over $x \in [-1,1]$
yields the desired inequality \eqref{ObjUB2}.
\end{proof}
We close this section by highlighting how the upper bound from Proposition \ref{PropUB}
is effectively computed by solving a linear program.
For this purpose, we introduce the collocation matrices $V \in \mathbb{R}^{K \times M}$
and $W \in \mathbb{R}^{L \times M}$
of the basis $(u_1,\ldots,u_M)$
at the points $v_1,\ldots,v_K$
(usually chosen as equispaced points in $[-1,1]$)
and at the Chebyshev zeros $w_1,\ldots,w_L$.
These matrices are defined by
\begin{equation}
\label{DefVW1}
V_{k,m} = u_m(v_k)
\qquad \mbox{and} \qquad
W_{\ell,m} = u_m(w_{\ell}),
\qquad
k \in \ibt{1}{K}, \;
\ell \in \ibt{1}{L}, \;
m \in \ibt{1}{M}.
\end{equation}
With this notation, the objective function and the constraints in \eqref{DD} read, respectively,
\begin{equation}
\max_{\ell \in \llbracket 1:L \rrbracket} \sum_{k=1}^K |(WA)_{\ell,k}|
\qquad \mbox{and} \qquad
(AV)_{m,m'} = \delta_{m,m'},
\quad m,m' \in \ibt{1}{M}.
\end{equation}
Then, introducing slack variables through $B \in \mathbb{R}^{L \times K}$
(such that $|(WA)_{\ell,k}| \le B_{\ell,k}$ for all $\ell,k$)
and $c \in \mathbb{R}$
(such that $\sum_{k=1}^K B_{\ell,k} \le c$ for all $\ell$),
we arrive at the following linear optimization problem,
written in an easily implementable form.
\ovalbox{
\begin{minipage}{0.97\textwidth}
\medskip
\begin{center}
\textbf{Upper bound for the projection constant}\vspace{-2mm}\\
\rule{0.96\textwidth}{.8pt}
\end{center}
Inputs: basis for a space $\mathcal{U}$ of polynomials of degree $\le d$, parameters $K$ and $L$.
\begin{equation}
\label{ProgUB}
\lambda(\mathcal{U})
\le \cos \left( \dfrac{\pi}{2}\dfrac{d}{L} \right)^{-1}\times
\min_{\substack{A \in \mathbb{R}^{M \times K}\\ B \in \mathbb{R}^{L \times K}\\ c \in \mathbb{R} }}
\; c
\qquad \mbox{s.to \;\;}
\left\{
\begin{matrix}
AV = I_M, \hfill \\
-B \le WA \le B, \hfill \\
B {\bf 1} \le c {\bf 1}, \hfill
\end{matrix}
\right.
\end{equation}
where the matrices $V$ and $W$ are defined in \eqref{DefVW1}.
\medskip
\end{minipage}
}
\section{Computable lower bound}
\label{SecLB}
We present in this section a computable lower bound for the projection constant of a \mbox{polynomial} \mbox{subspace} $\mathcal{U}$ of $\mathcal{C}[-1,1]$.
It again involves a discretization of the problem \eqref{Original},
followed by an \mbox{application} of the moment method.
This method, based on classical moment problems
(see e.g.~\cite{Schm}),
characterizes a measure by its sequence of moments,
typically via a semidefinite condition.
The \mbox{discrete} Hamburger moment problem, for instance,
states that,
for a sequence $(y_k)_{k \ge 0}$ of real \mbox{numbers},
there exists a nonnegative measure $\mu$ on $\mathbb{R}$ such that
\begin{equation}
\int_{-\infty}^\infty x^k d\mu(x) = y_k,
\qquad
k \ge 0,
\end{equation}
if and only if an infinite Hankel matrix is positive semidefinite, precisely
\begin{equation}
{\rm Hank}_\infty(y) := \begin{bmatrix}
y_0 & y_1 & y_2 & y_3 & \cdots\\
y_1 & y_2 & y_3 & \reflectbox{$\ddots$} & \\
y_2 & y_3 &\reflectbox{$\ddots$} & & \\
y_3 & \reflectbox{$\ddots$} & & & \\
\vdots & & & &
\end{bmatrix}
\succeq 0.
\end{equation}
We could directly use this characterization to substitute the measures $\mu_1,\ldots,\mu_M$ by their sequences $y_1,\ldots,y_M$ of moments as optimization variables,
but we would need to add extra semidefinite conditions ensuring that the measures $\mu_1,\ldots,\mu_M$ are localized on $[-1,1]$.
Instead, we prefer to rely on the discrete trigonometric moment problem, which states\footnote{The classical statement, found e.g. in \cite[Theorem 11.3]{Schm},
concerns sequences of complex numbers indexed by $n \in \mathbb{Z}$
and obtained as $\int_{|z|=1} z^{-n} d\mu(z)$ for some Radon measure on the unit circle.
We omit the verification that it implies the statement being made here.} that,
for a sequence $(y_k)_{k \ge 0}$ of real numbers,
there exists a nonnegative measure $\mu$ on $[0,\pi]$ such that
\begin{equation}
\int_{0}^\pi \cos( k \theta) d \mu(\theta)
= y_k,
\qquad
k \ge 0,
\end{equation}
if and only if an infinite Toeplitz matix is positive semidefinite, precisely
\begin{equation}
{\rm Toep}_\infty(y) := \begin{bmatrix}
y_0 & y_1 & y_2 & y_3 & \cdots\\
y_1 & y_0 & y_1 & y_2 & \\
y_1 & y_1 & y_0 & y_1 & \ddots \\
y_3 & y_2 & y_1 & \ddots & \ddots \\
\vdots & & \ddots & \ddots & \ddots
\end{bmatrix}
\succeq 0.
\end{equation}
There are two reasons explaining our preference:
firstly, localization conditions are not necessary;
secondly, the Toeplitz structure is numerically more favorable than the Hankel structure.
Now, in order to invoke the discrete trigonometric moment problem,
we first have to transform the expression of the projection constant given in \eqref{Original}.
This is done below.
\begin{lem}
\label{LemTransformed}
Given a basis $(u_1,\ldots,u_M)$ for a subspace $\mathcal{U}$ of $\mathcal{C}[-1,1]$,
if the functions $\widehat{u}_1, \ldots, \widehat{u}_M$ are defined on $[0,\pi]$ by
$\widehat{u}_m(\theta) = u_m( \cos(\theta))$,
then
\begin{equation}
\label{Transformed}
\lambda(\mathcal{U}) = \inf_{\mu_1,\ldots,\mu_M}
\max_{\theta \in [0,\pi]} \int_{0}^\pi \left| \sum_{m=1}^M \widehat{u}_m(\theta) d\mu_m \right|
\quad
\mbox{s.to }
\int_{0}^\pi \widehat{u}_{m'} d\mu_{m} = \delta_{m,m'},
\; m,m' \in \ibt{1}{M},
\end{equation}
where the infimum is taken over all signed Borel measures $\mu_1,\ldots,\mu_M$ on $[0,\pi]$.
\end{lem}
\begin{proof}
Let us consider the subspace $\widehat{\mathcal{U}} $ of $\mathcal{C}[0,\pi]$ defined by $\widehat{\mathcal{U}} = \{ u \circ \cos, u \in \mathcal{U} \}$,
for which $(\widehat{u}_1,\ldots,\widehat{u}_M)$ is a basis.
We readily verify that,
if $P$ is a projection from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$,
then
\begin{equation}
Q: g \in \mathcal{C}[0,\pi] \mapsto P(g \circ \arccos) \circ \cos \in \widehat{\mathcal{U}}
\end{equation}
defines a projection from $\mathcal{C}[0,\pi]$ onto $\widehat{\mathcal{U}}$ satisfying $\|Q\|_{\infty \to \infty} = \|P\|_{\infty \to \infty}$.
Conversely, we also see that if $Q$ is a projection from $\mathcal{C}[0,\pi]$ onto $\widehat{\mathcal{U}}$,
then
\begin{equation}
P: f \in \mathcal{C}[-1,1] \mapsto Q(f \circ \cos) \circ \arccos \in \mathcal{U}
\end{equation}
defines a projection from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$
satisfying $\|P\|_{\infty \to \infty} = \|Q\|_{\infty \to \infty}$.
This implies that the projection constant of $\mathcal{U}$ (relative to $\mathcal{C}[-1,1]$)
equals the projection constant of $\widehat{\mathcal{U}}$ (relative to $\mathcal{C}[0,\pi]$),
i.e.,
\begin{equation}
\label{Ori2}
\lambda(\mathcal{U}) = \lambda(\widehat{\mathcal{U}}) =
\min_{Q} \; \|Q\|_{\infty \to \infty}
\quad
\mbox{s.to $Q$ being a projection from $\mathcal{C}[0,\pi]$ onto $\widehat{\mathcal{U}}$}.
\end{equation}
The derivation of \eqref{Transformed} from \eqref{Ori2}
is similar to the derivation of \eqref{Original} from \eqref{Ori1}.
\end{proof}
Lemma \ref{LemTransformed} clearly yields a lower bound for the projection constant if we replace the maximum over the interval $[0,\pi]$ by the maximum over a discretization grid $\theta_1,\ldots,\theta_L$ of $[0,\pi]$,
i.e., we have
\begin{equation}
\label{Transformed2}
\lambda(\mathcal{U}) \ge \inf_{\mu_1,\ldots,\mu_M}
\max_{\ell \in \llbracket 1 : L \rrbracket} \int_{0}^\pi \left| \sum_{m=1}^M \widehat{u}_m(\theta_\ell) d\mu_m \right|
\quad
\mbox{s.to }
\int_{0}^\pi \widehat{u}_{m'} d\mu_{m} = \delta_{m,m'},
\; m,m' \in \ibt{1}{M},
\end{equation}
where the infimum is taken over all signed Borel measures $\mu_1,\ldots,\mu_M$ on $[0,\pi]$.
Our next step consists in recasting the latter minimization problem
so as to involve only (nonnegative) Borel measures instead of signed Borel measures.
Although the following observation may seem obvious,
we make an extra effort to verify it fully.
\begin{samepage}
\begin{lem}
Given a basis $(u_1,\ldots,u_M)$ for a subspace $\mathcal{U}$ of $\mathcal{C}[-1,1]$,
let $\widehat{u}_1,\ldots,\widehat{u}_M$ still denote the functions
$u_1 \circ \cos, \ldots, u_M \circ \cos$.
If $\theta_1,\ldots,\theta_L \in [0,\pi]$,
then
\begin{align}
\label{InfPBM}
\lambda(\mathcal{U}) \ge
\inf_{\substack{\mu_1^\pm,\ldots,\mu_M^\pm\\ \nu_1^\pm, \ldots, \nu_L^\pm}} \max_{\ell \in \llbracket 1 : L \rrbracket}
\int_{0}^\pi \left( d\nu_\ell^+ + d\nu_\ell^- \right)
& & \mbox{s.to} &
\int_0^\pi \widehat{u}_{m'} (d\mu_m^+ - d \mu_m^-) = \delta_{m,m'}, \; m,m' \in \ibt{1}{M},\\
\nonumber
& & \mbox{} &
\sum_{m = 1}^M \widehat{u}_m(\theta_\ell) (\mu_m^+ - \mu_m^-) = \nu_\ell^+ - \nu_\ell^-, \; \ell \in \ibt{1}{L},
\end{align}
where the infimum is taken over all (nonnegative) Borel measures
$\mu_1^\pm,\ldots,\mu_M^\pm, \nu_1^\pm, \ldots, \nu_L^\pm$ on $[0,\pi]$.
\end{lem}
\end{samepage}
\begin{proof}
Let $\alpha$ be the value of the infimum in \eqref{Transformed2} and let $\beta$ be the value of the infimum in \eqref{InfPBM}.
To prove that $\alpha \le \beta$,
we consider miminizers $\mu_1^\pm,\ldots,\mu_M^\pm,\nu_1^\pm,\ldots,\nu_L^\pm$ for the problem \eqref{InfPBM}.
By virtue of the first constraint in \eqref{InfPBM},
the measures $\mu_m^+ - \mu_m^-$, $m \in \ibt{1}{M}$,
are feasible for the problem \eqref{Transformed2},
so that, using the second constraint in \eqref{InfPBM},
\begin{align}
\alpha & \le \max_{\ell \in \llbracket 1: L \rrbracket} \int_0^\pi \left| \sum_{m=1}^M \widehat{u}_m(\theta_\ell) (d\mu_m^+ - d \mu_m^-) \right|
= \max_{\ell \in \llbracket 1: L \rrbracket} \int_0^\pi |d\nu_\ell^+ - d\nu_\ell^- | \\
\nonumber
& \le \max_{\ell \in \llbracket 1: L \rrbracket} \int_0^\pi (d\nu_\ell^+ + d\nu_\ell^-)
= \beta.
\end{align}
To prove that $\beta \le \alpha$,
we consider minimizers $\mu_1\ldots,\mu_M$ for the problem \eqref{Transformed2}.
Then we write the Jordan decomposition of each $\mu_m$, $m \in \ibt{1}{M}$,
as $\mu_m = \mu_m^+ - \mu_m^-$ for some (nonnegative) Borel measures $\mu_m^\pm$ satisfying $\mu_m^+ \perp \mu_m^-$.
For each $\ell \in \ibt{1}{L}$, we also write the Jordan decomposition of
$\nu_\ell := \sum_{m=1}^M \widehat{u}_m(\theta_\ell) \mu_m$ as $\nu_\ell= \nu_\ell^+ - \nu_\ell^-$
for some (nonnegative) Borel measures $\nu_\ell^\pm$ satisfying $\nu_\ell^+ \perp \nu_\ell^-$.
In particular, we have $\int_0^\pi |d\nu_\ell| = \int_0^\pi (d\nu_\ell^+ + d\nu_\ell^-)$.
Then, noticing that the Borel measures $\mu_1^\pm,\ldots,\mu_M^\pm,\nu_1^\pm, \ldots, \nu_L^\pm$
are feasible for the problem \eqref{InfPBM}, we obtain
\begin{equation}
\beta \le \max_{\ell \in \llbracket 1: L \rrbracket} \int_0^\pi (d\nu_\ell^+ + d\nu_\ell^-)
= \max_{\ell \in \llbracket 1: L \rrbracket} \int_0^\pi |d\nu_\ell|
= \max_{\ell \in \llbracket 1: L \rrbracket} \int_0^\pi \left| \sum_{m=1}^M \widehat{u}_m(\theta_\ell) d\mu_m \right|
= \alpha.
\end{equation}
The proof is now complete.
\end{proof}
The expression in \eqref{InfPBM} is still not directly exploitable due to the infinite-dimensionality of the Borel measures.
We resolve this issue by substituting these measures by their sequences of moments
and then by truncating the moment constraints.
With fewer constraints,
a smaller value for the minimum is produced.
At the same time, since the moments discarded in the constraints do not occur in the objective function either,
they can be removed altogether
to create a finite-dimensional semidefinite program.
We make all of this precise in the proof of the following result,
which presents the awaited computable lower bound.
Below, the notation ${\rm Toep}_S(y)$ stands for the $S \times S$ Toeplitz matrix constructed from the first $S$ components of a sequence $(y_k)_{k \ge 0}$,
i.e.,
\begin{equation}
{\rm Toep}_S(y) =
\begin{bmatrix}
y_0 & y_1 & \cdots & \cdots & y_{S-1}\\
y_{1} & y_{0} & y_{1} & & \vdots \\
\vdots & y_{1} & \ddots & \ddots & \vdots\\
\vdots & & \ddots & \ddots & y_{1}\\
y_{S-1} & \cdots & \cdots & y_{1} & y_{0}
\end{bmatrix} .
\end{equation}
\begin{prop}
\label{PropFinalLB}
Let $\mathcal{U}$ be a subspace of $\mathcal{C}[-1,1]$ consisting of polynomials of degree at most $ d$.
Let $(u_1,\ldots,u_M)$ be a basis for $\mathcal{U}$,
whose elements have the Chebyshev expansions
\begin{equation}
\label{CoefUinT}
u_m = \sum_{k=0}^d U_{k,m} T_k,
\qquad m \in \ibt{1}{M}.
\end{equation}
For an integer $S > d$ and for points $\theta_1,\ldots,\theta_L \in [0,\pi]$,
one has
\begin{align}
\label{FinalLB}
\lambda(\mathcal{U}) \ge \hspace{-2mm}
\inf_{\substack{y_1^\pm,\ldots,y_M^\pm\\ z_1^\pm, \ldots, z_L^\pm}} \max_{\ell \in \llbracket 1 : L \rrbracket}
(z_{\ell,0}^+ + z_{\ell,0}^- )
& & \mbox{s.to } &
\sum_{k=0}^d U_{k,m'} (y^+_{m,k} - y^-_{m,k}) = \delta_{m,m'}, \; m,m' \in \ibt{1}{M},\\
\nonumber
& & &
\sum_{m = 1}^M u_m(\cos(\theta_\ell)) (y_m^+ - y_m^-) = z_\ell^+ - z_\ell^-, \; \ell \in \ibt{1}{L},\\
\nonumber
& & & {\rm Toep}_S(y^\pm_m) \succeq 0, \; m \in \ibt{1}{M},
\\
\nonumber
& & &
{\rm Toep}_S(z^\pm_\ell) \succeq 0, \; \ell \in \ibt{1}{L},
\end{align}
where the infimum is taken over all vectors $y_1^\pm,\ldots,y_M^\pm,z_1^\pm, \ldots, z_L^\pm \in \mathbb{R}^S$ indexed from $0$ to $S-1$.
\end{prop}
\begin{proof}
In the optimization program \eqref{InfPBM}, we substitute the Borel measures $\mu_1^\pm,\ldots,\mu_M^\pm,\mu_1^\pm,\ldots,\nu_L^\pm$
by their infinite sequences $y_1^\pm,\ldots,y_M^\pm,z_1^\pm,\ldots,z_L^\pm$
of moments,
identified as
\begin{equation}
\label{TrigMom}
y^\pm_{m,k} = \int_0^\pi \cos(k \theta) d\mu^\pm_m(\theta),
\qquad
z^\pm_{\ell,k} = \int_0^\pi \cos(k \theta) d\nu_\ell^\pm(\theta),
\qquad
k \ge 0,
\end{equation}
to reach an equivalent optimization program featuring the objective function $\max_{\ell \in \llbracket 1: L \rrbracket}(z^+_{\ell,0} + z^-_{\ell,0})$.
As for the constraints (duality, consistency, moments),
the first constraint in \eqref{InfPBM} becomes the first constraint in \eqref{FinalLB} by virtue of
\begin{align}
\int_{0}^\pi \widehat{u}_{m'}(\theta) (d\mu_m^+(\theta) - d\mu_m^-(\theta))
& = \int_0^\pi \sum_{k=0}^d U_{k,m'} T_k(\cos(\theta)) (d\mu_m^+(\theta) - d\mu_m^-(\theta))\\
\nonumber
& = \sum_{k=0}^d U_{k,m'} \int_0^\pi \cos(k \theta) (d\mu_m^+(\theta) - d\mu_m^-(\theta))\\
\nonumber
& = \sum_{k=0}^d U_{k,m'} (y^+_{m,k} - y^-_{m,k});
\end{align}
the second constraint in \eqref{InfPBM} is clearly equivalent to the second constraint in \eqref{FinalLB};
while the fact that we are dealing with sequences of moments is reflected by the semidefinite conditions ${\rm Toep}_\infty(y_m^\pm) \succeq 0$, $m \in \ibt{1}{M}$, and ${\rm Toep}_\infty(z_\ell^\pm) \succeq 0$, $\ell \in \ibt{1}{L}$.
We now relax the last two sets of constraints by simply imposing,
for $\ell \in \ibt{1}{L}$,
\begin{equation}
\sum_{m = 1}^M u_m(\cos(\theta_\ell)) (y_{m,k}^+ - y_{m,k}^-) = z_{\ell,k}^+ - z_{\ell,k}^-,
\qquad
k \in \ibt{0}{S \hspace{-1mm} - \hspace{-1mm}1},
\end{equation}
as well as, for $m \in \ibt{1}{M}$ and $\ell \in \ibt{1}{L}$,
\begin{equation}
{\rm Toep}_S(y_m^\pm)
\succeq 0
\qquad \mbox{and} \qquad
{\rm Toep}_S(z_\ell^\pm) \succeq 0.
\end{equation}
This leads to a smaller minimum value for the optimization program.
And since the relaxed program only involves moments up to order $S-1$,
we can restrict the minimization to the finite sequences
$(y_{m,k}^\pm)_{k=0}^{S-1}$, $m \in \ibt{1}{M}$,
and $(z_{\ell,k}^\pm)_{k=0}^{S-1}$, $\ell \in \ibt{1}{L}$,
hence yielding the computable lower bound stated in \eqref{FinalLB}.
\end{proof}
We close this section by highlighting how the lower bound from Proposition \ref{PropFinalLB}
is expressed as a semidefinite program.
For this purpose,
besides the matrix $U \in \mathbb{R}^{(d+1) \times M}$
containing the coefficients of $u_1,\ldots,u_M$ in the Chebyshev system $(T_0,\ldots,T_d)$, as indicated in \eqref{CoefUinT},
we also consider the collocation matrix $W \in \mathbb{R}^{L \times M}$ with entries
\begin{equation}
\label{DefW2}
W_{\ell,m} = u_m(\cos(\theta_\ell)),
\qquad
\ell \in \ibt{1}{L},
\; m \in \ibt{1}{M}.
\end{equation}
By further introducing matrices $Y^{\pm} \in \mathbb{R}^{M \times S}$
with rows $y_1^\pm,\ldots,y_M^\pm \in \mathbb{R}^S$
and $Z^{\pm} \in \mathbb{R}^{L \times S}$
with rows $z_1^\pm,\ldots,z_L^\pm \in \mathbb{R}^S$,
as well as a slack variable $c \in \mathbb{R}$
(such that $z^+_{\ell,0} + z^-_{\ell,0} \le c$ for all $\ell$),
the optimization program in \eqref{FinalLB}
takes the easily implementable form below (where some convenient {\sc matlab} notation is used).\footnote{The grid points $\cos(\theta_1),\ldots,\cos(\theta_L)$ could be added as inputs --- by default, we chose them to be Chebyshev zeros.\label{FNgp}}
\ovalbox{
\begin{minipage}{0.97\textwidth}
\medskip
\begin{center}
\textbf{Lower bound for the projection constant}\vspace{-2mm}\\
\rule{0.96\textwidth}{.8pt}
\end{center}
Inputs: basis for a space $\mathcal{U}$ of polynomials of degree $\le d$, parameters $S>d$ and $L$.
\begin{equation}
\label{ProgLB}
\lambda(\mathcal{U})
\ge
\min_{\substack{Y^\pm \in \mathbb{R}^{M \times S}\\ Z^\pm \in \mathbb{R}^{L \times S}\\ c \in \mathbb{R} }}
\; \; c
\qquad \mbox{s.to \;\;}
\left\{
\begin{matrix}
(Y^+(:,1\hspace{-1mm}: \hspace{-1mm}d+1)-Y^-(:,1\hspace{-1mm}: \hspace{-1mm}d+1)) \, U= I_M, \\
W \, (Y^+ - Y^-) = Z^+ - Z^-, \hfill \\
{\rm Toep}_S(Y^\pm(m,:)) \succeq 0,
\quad m \in \ibt{1}{M}, \hfill\\
{\rm Toep}_S(Z^\pm(\ell,:)) \succeq 0,
\quad \ell \in \ibt{1}{L}, \hfill \\
Z^+(:,1) + Z^-(:,1) \le c , \hfill
\end{matrix}
\right.
\end{equation}
where the matrices $U$ and $V$ are defined in \eqref{CoefUinT} and \eqref{DefW2}.
\medskip
\end{minipage}
}
\section{Exploiting the symmetry of minimal projections}
\label{SecSym}
The programs highlighted in \eqref{ProgUB} and \eqref{ProgLB} are computationally demanding for large values of the parameters $K$, $L$, and $S$,
so any property that can reduce their complexity should be exploited.
We shall capitalize on a certain symmetry of minimal projections.
To this end, we assume from now on that the subspace $\mathcal{U}$ of $\mathcal{C}[-1,1]$ satisfies
\begin{equation}
\label{USym}
u \in \mathcal{U} \Longrightarrow u(-\cdot) \in \mathcal{U},
\end{equation}
where $u(-\cdot)$ evidently denotes the function $x \in [-1,1] \mapsto u(-x) \in \mathbb{R}$.
Under this assumption, it is known that there exists a minimal projection $P$ from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$ which is symmetric, in the sense that
\begin{equation}
\label{SymCond}
P(f(-\cdot))= (P(f))(-\cdot)
\qquad \mbox{for all $f \in \mathcal{C}[-1,1]$}.
\end{equation}
This fact has the following implication whenever assumption \eqref{USym} holds.\footnote{Under assumption \eqref{USym}, one can verify that there exists a basis $(u_1,\ldots,u_M)$ whose elements are either even or odd functions
(verify, for instance, that $\mathcal{U} = \mathcal{U}^{\rm e} \overset{\perp}{\oplus} \mathcal{U}^{\rm o}$,
where $\mathcal{U}^{\rm e/o} := \{ u \in \mathcal{U}: u \mbox{ is an even/odd function} \}$,
and concatenate a basis for $\mathcal{U}^{\rm e}$ with a basis for $\mathcal{U}^{\rm o}$).}
\begin{prop}
Let $\mathcal{U}$ be a subspace of $\mathcal{C}[-1,1]$ and let $(u_1^{\rm e},\ldots,u_{M_{\rm e}}^{\rm e},u_1^{\rm o},\ldots,u_{M_{\rm o}}^{\rm o})$ be a basis for~$\mathcal{U}$ arranged in such a way that the $u_m^{\rm e}$ are even functions and the $u_m^{\rm o}$ are odd functions.
Considering functions $\widetilde{u}_1^{\rm e},\ldots,\widetilde{u}_{M_{\rm e}}^{\rm e},\widetilde{u}_1^{\rm o},\ldots,\widetilde{u}_{M_{\rm o}}^{\rm o}$ defined on $[-1,1]$ by
\begin{equation}
\widetilde{u}^{\rm e/o}_m(t) := u^{\rm e/o}_m \left( \frac{t+1}{2} \right),
\qquad t \in [-1,1],
\end{equation}
the projection constant of $\mathcal{U}$ can be expressed as
\begin{align}
\label{NewExpr}
\lambda(\mathcal{U}) = & \inf_{\substack{\widetilde{\mu}_1^{\rm e},\ldots,\widetilde{\mu}_{M_{\rm e}}^{\rm e}\\ \widetilde{\mu}_1^{\rm o},\ldots,\widetilde{\mu}_{M_{\rm o}}^{\rm o}}}
\max_{x \in [0,1]}
\int_{-1}^1
\max \left\{
\left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}( x ) d\widetilde{\mu}_m^{\rm e} \right|,
\left| \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}( x ) d\widetilde{\mu}_m^{\rm o} \right|
\right\}\\
\nonumber
\mbox{s.to} & \int_{-1}^1 \widetilde{u}_{m'}^{\rm e} d\widetilde{\mu}_m^{\rm e} = \delta_{m,m'}, \; m,m' \in \ibt{1}{M_{\rm e}},
\qquad
\int_{-1}^1 \widetilde{u}_{m'}^{\rm o} d\widetilde{\mu}_m^{\rm o} = \delta_{m,m'}, \; m,m' \in \ibt{1}{M_{\rm o}},
\end{align}
where the infimum is taken over all signed Borel measures $\widetilde{\mu}_1^{\rm e},\ldots,\widetilde{\mu}_{M_{\rm e}}^{\rm e},\widetilde{\mu}_1^{\rm o},\ldots,\widetilde{\mu}_{M_{\rm o}}^{\rm o}$ on $[-1,1]$.
\end{prop}
\begin{proof}
Let us consider a symmetric projection from $\mathcal{C}[-1,1]$ onto $\mathcal{U}$ written as
\begin{equation}
P(f) = \sum_{m=1}^{M_{\rm e}} \eta^{\rm e}_m(f) u_m^{\rm e}
+ \sum_{m=1}^{M_{\rm o}} \eta^{\rm o}_m(f) u_m^{\rm o},
\qquad f \in \mathcal{C}[-1,1].
\end{equation}
The condition \eqref{SymCond} is readily seen to be equivalent to the conditions
\begin{equation}
\label{SymEta}
\eta_m^{\rm e}(f(-\cdot)) = \eta_m^{\rm e}(f),
\qquad
\eta_m^{\rm o}(f(-\cdot)) = -\eta_m^{\rm o}(f),
\qquad f \in \mathcal{C}[-1,1],
\end{equation}
which in turn are equivalent, in terms of measures $\mu_1^{\rm e}, \ldots, \mu_{M_{\rm e}}^{\rm e},\mu_1^{\rm o},\ldots,\mu_{M_{\rm o}}^{\rm o}$
representing the linear functionals $\eta_1^{\rm e}, \ldots, \eta_{M_{\rm e}}^{\rm e},\eta_1^{\rm o},\ldots,\eta_{M_{\rm o}}^{\rm o}$, to the conditions
\begin{equation}
\label{SymMu}
d\mu_m^{\rm e}(- \cdot) = d\mu_m^{\rm e},
\qquad
d\mu_m^{\rm o}(- \cdot) = - d\mu_m^{\rm o}.
\end{equation}
Then, the norm of the projection $P$ satisfies
\begin{align}
\|P\|_{\infty \to \infty}
& = \max_{x \in [-1,1]} \int_{-1}^1 \left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}(x) d\mu_m^{\rm e} + \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}(x) d\mu_m^{\rm o} \right|\\
\nonumber & = \max_{x \in [-1,1]} \int_{-1}^1 \left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}(x) d\mu_m^{\rm e} - \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}(x) d\mu_m^{\rm o} \right|\\
\nonumber & = \max_{x \in [-1,1]} \int_{-1}^1 \max \left\{ \left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}(x) d\mu_m^{\rm e}\right| , \left| \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}(x) d\mu_m^{\rm o} \right| \right\},
\end{align}
where we have used the identity $(|a+b|+|a-b|) / 2 = \max\{ |a|, |b| \}$.
Noticing the invariance of the above expression under the change $x \leftrightarrow -x$
and taking \eqref{SymMu} into account,
we can further write
\begin{align}
\label{ExpNormSym}
\|P\|_{\infty \to \infty}
& = \max_{x \in [0,1]} 2 \int_{0}^1 \max \left\{ \left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}(x) d\mu_m^{\rm e}\right| , \left| \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}(x) d\mu_m^{\rm o} \right| \right\}\\
\nonumber
& = \max_{x \in [0,1]} \phantom{2} \int_{-1}^1 \max \left\{ \left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}(x) d\widetilde{\mu}_m^{\rm e}\right| , \left| \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}(x) d\widetilde{\mu}_m^{\rm o} \right| \right\},
\end{align}
where the measures $\widetilde{\mu}_m^{\rm e/o}$ simply represent the restrictions to $[0,1]$ of the measures $\mu_m^{\rm e/o}$ that have been transposed to $[-1,1]$,
i.e., they are obtained through the identification
\begin{equation}
d\widetilde{\mu}_m^{\rm e/o}(t) = 2 d \mu_m^{\rm e/o}(\tau),
\qquad t \in [-1,1], \tau \in [0,1]
\mbox{ being linked via }
t = 2 \tau - 1, \,
\tau = \frac{t+1}{2}.
\end{equation}
Thanks to \eqref{SymMu},
we can take the $\widetilde{\mu}_m^{\rm e/o}$ as new optimization variables in the minimization of the norm of a symmetric projection,
whose expression \eqref{ExpNormSym} is the objective function in \eqref{NewExpr}.
We now just have to impose the appropriate duality constraints making $P$ a projection onto~$\mathcal{U}$.
Among them,
the constraints $\eta_m^{\rm e/o}(u_{m'}^{\rm o/e}) = 0$ are automatically fulfilled,
while the constraints $\eta_m^{\rm e/o}(u_{m'}^{\rm e/o}) = \delta_{m,m'}$ reduce to
\begin{equation}
\delta_{m,m'} = \int_{-1}^1 u_{m'}^{\rm e/o} d\mu_{m}^{\rm e/o}
= 2 \int_{0}^1 u_{m'}^{\rm e/o}(\tau) d\mu_{m}^{\rm e/o}(\tau)
= \int_{-1}^1 \widetilde{u}_{m'}^{\rm e/o} (t) d\widetilde{\mu}_m^{\rm e/o}(t).
\end{equation}
These are indeed the constraints in \eqref{NewExpr},
so the proposition is proved.
\end{proof}
We proceed by highlighting the computable bounds on the projection constant generated by the reformulation \eqref{NewExpr}.
Much of the ingredients for deriving these bounds are similar to the ones presented in Sections \ref{SecUB} and \ref{SecLB},
so we do expand on details at all.
\subsection{Implication for the upper bound}
As in Section \ref{SecUB}, we first derive an upper bound for the projection constant by minimizing only over linear combinations
\begin{equation}
\widetilde{\mu}_m^{\rm e/o} = \sum_{k=1}^K A^{\rm e/o}_{m,k} \delta_{v_k}.
\end{equation}
of Dirac measures at $v_1,\ldots,v_K \in [-1,1]$.
Then we again discretize by replacing the maximum over $[0,1]$ by the maximum over
a grid $w_1^+ > \cdots > w_L^+$ consisting of positive Chebyshev zeros
$w_\ell^+ = \cos(\pi (\ell-1/2)/(2L))$.
Thus,
we arrive at the computable upper bound
\begin{align}
\lambda(\mathcal{U})
\le \rho \times \hspace{-1mm}
\min_{A^{\rm e/o} \in \mathbb{R}^{M_{\rm e/o} \times K}}
\max_{\ell \in \llbracket 1 : L \rrbracket }
\sum_{k=1}^K
\max & \left\{
\left| \sum_{m=1}^{M_{\rm e}} A^{\rm e}_{m,k} u^{\rm e}_m(w_\ell^+) \right|,
\left| \sum_{m=1}^{M_{\rm o}} A^{\rm o}_{m,k} u^{\rm o}_m(w_\ell^+) \right|
\right\}\\
\nonumber
\mbox{s.to } &
\sum_{k=1}^K A^{\rm e/o}_{m,k} \widetilde{u}^{\rm e/o}_{m'}(v_k) = \delta_{m,m'},
\, m,m' \in \ibt{1}{M_{\rm e/o}},
\end{align}
where $\rho = \cos \left( (\pi d)/(4 L ) \right)^{-1}$.
To transform the latter into a linear program,
we introduce the collocation matrices $V^{\rm e/o} \in \mathbb{R}^{K \times M_{\rm e/o}}$
and $W^{\rm e/o} \in \mathbb{R}^{L \times M_{\rm e/o}}$ defined by
\begin{equation}
\label{DefVW3}
V^{\rm e/o}_{k,m} = \widetilde{u}^{\rm e/o}_m(v_k)
\qquad \mbox{and} \qquad
W^{\rm e/o}_{\ell,m} = u^{\rm e/o}_m(w_{\ell}^+),
\qquad
k \in \ibt{1}{K}, \;
\ell \in \ibt{1}{L}, \;
m \in \ibt{1}{M_{\rm e/o}},
\end{equation}
as well as slack variables through $B \in \mathbb{R}^{L \times K}$ and $c \in \mathbb{R}$.
All in all, we obtain the following implementable form of the upper bound.
\ovalbox{
\begin{minipage}{0.97\textwidth}
\medskip
\begin{center}
\textbf{Upper bound for the projection constant --- symmetry exploited}\vspace{-2mm}\\
\rule{0.96\textwidth}{.8pt}
\end{center}
Inputs: basis for a symmetric space $\mathcal{U}$ of polynomials of degree $\le d$, parameters $S>d$ and $L$.
\begin{equation}
\label{ProgUBSym}
\lambda(\mathcal{U})
\le \cos \left( \dfrac{\pi}{4}\dfrac{d}{L} \right)^{-1}
\times
\min_{\substack{ A^{\rm e/o} \in \mathbb{R}^{M_{\rm e/o} \times K}
\\
B \in \mathbb{R}^{L \times K} \\ c \in \mathbb{R}}}
\; c
\qquad
\mbox{s.to \; \;}
\left\{
\begin{matrix}
A^{\rm e/o} V^{\rm e/o} = I_{M_{\rm e/o}}, \hfill \\
-B \le W^{\rm e/o} A^{\rm e/o} \le B, \hfill \\
B {\bf 1} \le c {\bf 1}, \hfill
\end{matrix}
\right.
\end{equation}
where the matrices $V^{\rm e/o}$ and $W^{\rm e/o}$ are defined in \eqref{DefVW3}.
\medskip
\end{minipage}
}
\subsection{Implication for the lower bound}
As in Section \ref{SecLB}, the minimization program \eqref{NewExpr} is first transformed to make it amenable to the trigonometric moment problem.
The reformulation will involve a maximum over $[0,\pi/2]$,
which is lower bounded by the maximum over a grid $\theta_1^+,\ldots,\theta_L^+$.
With $w_\ell^+ := \cos(\theta_\ell^+) \in [0,1]$, we obtain
\begin{align}
\lambda(\mathcal{U}) \ge \inf_{\substack{\widetilde{\mu}_1^{\rm e},\ldots,\widetilde{\mu}_{M_{\rm e}}^{\rm e}\\ \widetilde{\mu}_1^{\rm o},\ldots,\widetilde{\mu}_{M_{\rm o}}^{\rm o}}}
\max_{\ell \in \llbracket 1: L \rrbracket }
\int_{0}^\pi
\max & \left\{
\left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}(w_\ell^+) d\widetilde{\mu}_m^{\rm e} \right|,
\left| \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}(w_\ell^+) d\widetilde{\mu}_m^{\rm o} \right|
\right\}\\
\nonumber
\mbox{s.to} & \int_{0}^\pi \widetilde{u}_{m'}^{\rm e/o} (\cos( \theta) ) d\widetilde{\mu}_m^{\rm e/o}(\theta) = \delta_{m,m'}, \quad m,m' \in \ibt{1}{M_{\rm e/o}},
\end{align}
where the infimum is taken over all signed Borel measures $\widetilde{\mu}_1^{\rm e},\ldots,\widetilde{\mu}_{M_{\rm e}}^{\rm e},\widetilde{\mu}_1^{\rm o},\ldots,\widetilde{\mu}_{M_{\rm o}}^{\rm o}$ on $[0,\pi]$.
\mbox{Deviating} slightly from the earlier strategy,
we now introduce as slack variables some \mbox{(nonnegative)} Borel measures $\nu_1,\ldots,\nu_L$
satisfying
\begin{equation}
\nu_\ell \ge
\max \left\{
\left| \sum_{m=1}^{M_{\rm e}} u_m^{\rm e}(w_\ell^+) \widetilde{\mu}_m^{\rm e} \right|,
\left| \sum_{m=1}^{M_{\rm o}} u_m^{\rm o}(w_\ell^+) \widetilde{\mu}_m^{\rm o} \right|
\right\},
\qquad \mbox{i.e.,} \qquad
\nu_\ell \ge \pm
\sum_{m=1}^{M_{\rm e/o}} u_m^{\rm e/o}(w_\ell^+) \widetilde{\mu}_m^{\rm e/o}.
\end{equation}
Writing the Chebyshev expansions of $\widetilde{u}_1^{\rm e},\ldots, \widetilde{u}_{M_{\rm e}}^{\rm e}, \widetilde{u}_1^{\rm o},\ldots, \widetilde{u}_{M_{\rm o}}^{\rm o}$
as
\begin{equation}
\label{DefU4}
\widetilde{u}_m^{\rm e/o} = \sum_{k=0}^d \widetilde{U}^{\rm e/o}_{k,m} T_k,
\qquad m \in \ibt{1}{M_{\rm e/o}},
\end{equation}
and substituting the measures $\widetilde{\mu}_m^{\rm e/o}$ and $\nu_\ell$
by their sequences $y_m^{\rm e/o}$ and $z_\ell$ of moments
yields to
\begin{align}
\lambda(\mathcal{U}) \ge \hspace{-1mm}
\inf_{\substack{y_1^{\rm e/o},\ldots,y_{M_{\rm e/o}}^{\rm e/o}\\ z_1,\ldots, z_L}}
\max_{\ell \in \llbracket 1: L \rrbracket } z_{\ell,0}
\quad \mbox{s.to \; } &
\sum_{k=0}^{d} \widetilde{U}^{\rm e/o}_{k,m'} y_{m,k}^{\rm e/o} = \delta_{m,m'},
\; m,m' \in \ibt{1}{M_{\rm e/o}},\\
\nonumber
& {\rm Toep}_\infty(z_\ell) \succeq \pm {\rm Toep}_\infty \left( \sum_{m=1}^{M_{\rm e/o}} u_m^{\rm e/o} ( w_\ell^+ ) y_m^{\rm e/o} \right),
\; \ell \in \ibt{1}{L},
\end{align}
where the infinum is taken over all infinite sequences $y_1^{\rm e/o},\ldots,y_{M_{\rm e/o}}^{\rm e/o}, z_1,\ldots, z_L \in \mathbb{R}^{\mathbb{N}}$.
The infinite semidefinite constraints are now truncated to a level $S>d$,
producing a lower bound involving only the finite sequences of moments $(y_{m,k}^{\rm e/o})_{k=0}^{S-1}$, $m \in \ibt{1}{M_{\rm e/o}}$, and $(z_{\ell,k})_{k=0}^{S-1}$, $\ell \in \ibt{1}{L}$.
Finally, in order to state the corresponding semidefinite program in an easily implementable form,
we gather these moments in matrices $Y^{\rm{e/o}} \in \mathbb{R}^{M_{\rm e/o}\times S}$ and $Z \in \mathbb{R}^{L \times S}$,
and we define collocation matrices $W^{\rm e/o} \in \mathbb{R}^{L \times M_{\rm e/o}}$ with entries
\begin{equation}
\label{DefW4}
W^{\rm e/o}_{\ell,m} = u^{\rm e/o}_m(w_\ell^+),
\qquad
\ell \in \ibt{1}{L},
\; m \in \ibt{1}{M_{\rm e/o}}.
\end{equation}
After introducing one last slack variable $c \in \mathbb{R}$,
we arrive at the following form of the computable lower bound
(where some convenient {\sc matlab} notation is again used).\footnote{In a similar spirit to footnote \ref{FNgp},
the grid points $w_1^+ = \cos(\theta_1), \ldots,w_L^+ =\cos(\theta_L)$ could be added as inputs --- by default, we chose them to be positive Chebyshev zeros.}
\ovalbox{
\begin{minipage}{0.97\textwidth}
\medskip
\begin{center}
\textbf{Lower bound for the projection constant --- symmetry exploited}\vspace{-2mm}\\
\rule{0.96\textwidth}{.8pt}
\end{center}
Inputs: basis for a symmetric space $\mathcal{U}$ of polynomials of degree $\le d$, parameters $S>d$ and $L$.
\begin{equation}
\label{ProgLBSym}
\lambda(\mathcal{U})
\ge
\min_{\substack{Y^{\rm e/o} \in \mathbb{R}^{M_{\rm e/o} \times S}\\ Z \in \mathbb{R}^{L \times S}\\ c \in \mathbb{R} }}
\; c
\qquad \mbox{s.to \;\;}
\left\{
\begin{matrix}
Y^{\rm e/o}(:,1\hspace{-1mm}: \hspace{-1mm}d+1) \, \widetilde{U}^{\rm e/o}= I_{M_{\rm e/o}}, \hfill \\
{\rm Toep}_S(Z(\ell,:)) \succeq
\pm {\rm Toep}_S(W^{\rm e/o}(\ell,:)Y^{\rm e/o}),
\; \ell \in \ibt{1}{L}, \hfill \\
Z(:,1) \le c , \hfill
\end{matrix}
\right.
\end{equation}
where the matrices $\widetilde{U}^{\rm e/o}$ and $W^{\rm e/o}$ are defined in \eqref{DefU4} and \eqref{DefW4}.
\medskip
\end{minipage}
}
\section{Computational results}
\label{SecCompRes}
This section gives an account of the experiments carried out using our method for specific spaces of univariate polynomials.
The experiments can be reproduced by downloading the {\sc matlab} file tied to this paper, available on the authors' webpages.
Note that the code relies on {\sf CVX} \cite{cvx},
a {\sc matlab} package for specifying and solving convex programs,
and on {\sf Chebfun}~\cite{Chebfun} for its convenience to deal with Chebyshev expansions.
\subsection{Validation of the code}
In order to certify the correct implementation of the codes computing the upper and lower bounds \eqref{ProgUBSym} and \eqref{ProgLBSym},
we take as a benchmark the inevitable result \cite{ChaMet} of Chalmers and Metcalf,
who managed to determine analytically the projection constant of the space of quadratic polynomials.
They obtained
\begin{equation}
\label{ValQuad}
\lambda(\mathcal{P}_2) \approx 1.220173064217988\ldots
\end{equation}
and exhibited a minimal projection given by $P(f)(x) = \sum_{m=1}^3 \eta_m(f) x^{m-1}$,
where
\begin{eqnarray}
\label{CMMeas1}\eta_1(f) = & A f(-1) + B f(0) + A f(1) & + \int_{I_1} \frac{a_1 |t| + b_1}{(1+w_1 |t|)^3} f(t)dt + \int_{I_2} \frac{a_2 |t| + b_2}{(1+w_2 |t|)^3} f(t)dt ,\\
\label{CMMeas2}\eta_2(f) = & -C f(-1) + C f(1) & + \int_{I_1} \frac{c_1t}{(1+w_1 |t|)^3} f(t) dt + \int_{I_2} \frac{c_2t}{(1+w_2 |t|)^3} f(t) dt,\\
\label{CMMeas3}\eta_3(f) = & D f(-1) - B f(0) + D f(1) & + \int_{I_1} \frac{d_1 |t| - b_1}{(1+w_1 |t|)^3} f(t)dt + \int_{I_2} \frac{d_2 |t| - b_2}{(1+w_2 |t|)^3} f(t)dt,
\end{eqnarray}
with $I_1 = [-s_{1,2},-s_{1,1}] \cup [s_{1,1},s_{1,2}]$, $I_2=[-s_{2,2},-s_{2,1}] \cup [s_{2,1} ,s_{2,2}]$, and with parameters $A,B,C,D$, $a_1,b_1,c_1,d_1$, $a_2,b_2,c_2,d_2$, $w_1,w_2$, $s_{1,1}, s_{1,2}, s_{2,1}, s_{2,2}$ determined in the body of \cite{ChaMet}.
Our code does allow us to retrieve the value \eqref{ValQuad}
up to five digits of accuracy.\footnote{in about five minutes for the upper bound and fifteen minutes for the lower bound,
with the capabilities offered by a laptop computer at the time this article was written.}
Incidentally, our experiment suggests that minimal projections onto the quadratics are not unique,
as illustrated by Figure \ref{Fig1} which superimposes the measures associated to \eqref{CMMeas1}-\eqref{CMMeas2}-\eqref{CMMeas3} and the discretized measures obtained by solving \eqref{ProgUBSym} ---
we have removed the atomic parts at $-1,0,1$,
which were similar.
It is worth noticing, nonetheless, that the supports of all the measures seem to be the same.
\begin{figure}[h]
\center
\subfigure{
\includegraphics[width=0.31\textwidth]{Mu1.eps}
}
\subfigure{
\includegraphics[width=0.31\textwidth]{Mu2.eps}
}
\subfigure{
\includegraphics[width=0.31\textwidth]{Mu3.eps}
}
\caption{The nonatomic parts of measures $\mu_1,\mu_2,\mu_3$ associated with the linear functionals $\eta_1,\eta_2,\eta_3$ found in \cite{ChaMet} (continuous lines),
together with their discrete approximations obtained as solutions of \eqref{ProgUBSym},
minus the parts at $-1$, $0$, and $1$ (circles, squares, diamonds).}
\label{Fig1}
\end{figure}
Another way validate of our code is offered by the three-dimensional space ${\rm span}\{1,x^2,x^3\}$, i.e.,
the subspace $\mathcal{C}[-1,1]$ spanned by $ x \mapsto 1$, $x \mapsto x^2$, and $x \mapsto x^3$.
Indeed, its projection constant is exactly equal to one.
More generally,
for any even integer $a>0$
and any odd integer $b>0$,
the space ${\rm span}\{ 1, x^a,x^{a+b} \} = {\rm span}\{1-x^a, x^a(1-x^b),x^a(1+x^b)\}$
admits a projection of norm one.
To see this, we notice that the interpolating projection at the points $-1$, $0$, and $1$,
which is given by
\begin{equation}
P(f)(x) = \frac{1}{2} f(-1) x^{a}(1-x^{b}) + f(0) (1-x^{a}) + \frac{1}{2} f(1) x^{a}(1+x^{b}),
\end{equation}
satisfies, for all $f \in \mathcal{C}[-1,1]$ and $x \in [-1,1]$,
\begin{equation}
|P(f)(x)| \le \max\{ |f(-1)|,|f(0)|,|f(1)| \} \left( \frac{1}{2}x^a(1-x^b) + 1-x^a + \frac{1}{2} x^a(1+x^b) \right)
\le \|f\|_\infty .
\end{equation}
Our code confirms the value $\lambda({\rm span}\{1,x^2,x^3\}) = 1$
(not with perfect accuracy, though, which is why we refrain from supplying numerical values with more than five digits in general).
\subsection{Other-three dimensional polynomial spaces}
The code distributed with the {\sc matlab} reproducible can be effortlessly applied to any univariate polynomial space,
so long as it can be executed with parameters large enough for the upper and lower bounds to match up to the desired accuracy.
Five digits of accuracy can typically be achieved for three-dimensional spaces of moderate degree.
The following table summarizes the values of projection constants obtained for several different spaces.\footnote{For the space ${\rm span}\{T_1,T_2,T_3\}$, it was more effective to compute the lower bound using a `symmetrization' of \eqref{ProgLB} more direct than \eqref{ProgLBSym}.
Its implementation is also included in the reproducible.}
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$\mathcal{V}$ spanned by &
$\{ 1,x,x^3 \}$ & $\{T_0,T_2,T_3\}$ & $\{U_0,U_2,U_3\}$ & $\{ x,x^2,x^3 \}$ & $\{T_1,T_2,T_3\}$ & $\{U_1,U_2,U_3 \}$\\
\hline
$\lambda(\mathcal{V})$ & $\approx 1.4723$ & $\approx 1.4460$ & $\approx 1.1522$ & $\approx 1.3325$ & $\approx 1.4065$ & $\approx 1.2354$ \\
\hline
\end{tabular}
\end{center}
\caption{Projection constants of three-dimensional spaces spanned by monomials and Chebyshev polynomials of the first and second kind.}
\end{table}
\subsection{Polynomial spaces of higher dimensions}
Looking at \eqref{ProgUBSym} and \eqref{ProgLBSym},
we see that the dimension $M=M_{\rm e} + M_{\rm o}$
is far from being a factor influencing the computational cost of the optimization programs,
so our code can easily be executed for polynomial spaces of dimension higher than three.
For instance,
we can deal with the spaces of cubic, quartic, and quintic polynomials
and compute their projection constants
with four digits of accuracy as
\begin{equation}
\lambda(\mathcal{P}_3) \approx 1.365,\qquad
\lambda(\mathcal{P}_4) \approx 1.459,\qquad
\lambda(\mathcal{P}_5) \approx 1.538.
\end{equation}
\begin{samepage}
Obtaining the same accuracy necessitates larger parameters $K$, $L$, $S$
when the dimension increases.
For degree $d>5$, with our modest computational investment,
we could locate the projection constants of the spaces $\mathcal{P}_d$ of polynomials of degree at most $d$ in the ranges presented in Table \ref{Table2} below.
The improvement with respect to the ranges found in \cite{HP} is particularly noticeable for the lower bounds
(recall that this is where the method of moments came into the picture).
\begin{table}[h]
\begin{center}
\begin{tabular}{|c||c||c|c||c|}
\hline
$\lambda(\mathcal{P}_d)$ & known lower bound & our lower bound & our upper bound & known upper bound \\
\hline
$d=3$ & $1.3539$ & $1.35667...$ & $1.35696...$ & $1.3577$\\
\hline
$d=4$ & $1.4524$ & $1.45902$... & $1.45951...$ & $1.4611$\\
\hline
$d = 5$ & $1.525 $ & $1.53817...$ & $1.53895...$ & $1.543 $\\
\hline
$d=6$ & $1.580$ & $1.60271...$ & $1.60383...$ & $1.613$\\
\hline
$d=7$ & $1.624$ & $1.65693...$ &$1.65859...$ & $1.669$\\
\hline
$d=8$ & $1.660$ & $1.70483...$ & $1.70731...$ & $1.721$\\
\hline
$d=9$ & $1.678$ & $1.74774...$ & $1.75107...$ & $1.775$\\
\hline
$d=10$ & $1.696$ & $1.78658...$ & $1.79076...$ & $1.814$\\
\hline
$d=11$ & NA & $1.82169...$ & $1.82701...$ & NA\\
\hline
$d=12$ & NA & $1.85380...$ & $1.86216...$ & NA\\
\hline
\end{tabular}
\end{center}
\caption{For the spaces $\mathcal{P}_d$ of polynomials of degree at most $d$,
lower and upper bounds on the projection constants obtained by our method compared to the ones stated in \cite{HP}.}
\label{Table2}
\end{table}
\end{samepage}
Let us now come to a close by examining the approximate minimal projection onto the cubics obtained by solving \eqref{ProgUBSym}.
For the measures $\mu_1,\mu_2,\mu_3,\mu_4$
associated with the functionals dual to $1,x,x^2,x^3$,
we detected atoms at $-1$ and $1$ (but none at $0$)
and Figure \ref{Fig2} indicates that their continuous parts seem to possess a common support strictly included in $[-1,1]$, as was the case for quadratics.
But a disparity with the quadratics now occurs in terms of shape preservation.
It was conjectured in \cite{PCM}, and proved for $d=2$,
that (one of the) minimal projections onto $\mathcal{P}_d$ preserve $d$-convexity.
This means that if $P$ is a minimal projection from $\mathcal{C}[-1,1]$ onto $\mathcal{P}_d$,
then,
for any $f \in \mathcal{C}^d[-1,1]$,
\begin{equation}
f^{(d)} \ge 0 \mbox{ on } [-1,1]
\overset{?}{\Longrightarrow} (P(f))^{(d)} \ge 0 \mbox{ on } [-1,1],
\end{equation}
or equivalently, writing $P(f)(x) = \sum_{m=1}^{d+1} \eta_m(f) x^{m-1}$,
\begin{equation}
f^{(d)} \ge 0 \mbox{ on } [-1,1]
\overset{?}{\Longrightarrow} \eta_{d+1}(f) \ge 0.
\end{equation}
Our computations for $d=3$ give some insight that this conjecture should be false.
Indeed,
if a minimal projection is approximated by our solution of \eqref{ProgUBSym},
which has the form
\begin{equation}
P(f)(x) = \sum_{m=1}^{4} \left( \sum_{k=-K}^K A_{m,k} f_k \right) x^{m-1},
\qquad f_k:= f\left( \frac{k}{K} \right),
\end{equation}
and if the condition $f''' \ge 0$ on $[-1,1]$ is replaced by its discrete version
\begin{equation}
\Delta_3(f)_k := f_{k+3} - 3 f_{k+2} + 3 f_{k+1} - f_k
\ge 0,
\qquad k \in \ibt{-K}{K-3},
\end{equation}
then, setting $a = A(4,:) \in \mathbb{R}^{2K-1}$, the question becomes
\begin{equation}
\min_{f \in \mathbb{R}^{2K-1}} \left\{ \langle a, f \rangle
\; : \,
\Delta_3(f) \ge 0 \right\}
\; \overset{?}{\ge} 0.
\end{equation}
This question is answered negatively
by solving a linear program.
\begin{figure}[h]
\center
\subfigure{
\includegraphics[width=0.48\textwidth]{CubMu1.eps}
}
\subfigure{
\includegraphics[width=0.48\textwidth]{CubMu2.eps}
}
\subfigure{
\includegraphics[width=0.48\textwidth]{CubMu3.eps}
}
\subfigure{
\includegraphics[width=0.48\textwidth]{CubMu4.eps}
}
\caption{The measures $\mu_1,\mu_2,\mu_3,\mu_4$
associated with the functionals dual to $1,x,x^2,x^3$
in the approximate minimal projection obtained by solving \eqref{ProgUBSym}.}
\label{Fig2}
\end{figure}
\section{Outlook}
As a concluding message,
we reiterate our belief in the usefulness of modern optimization techniques for solving problems in Approximation Theory.
Purists will argue that `solving computationally' is not really solving,
but benefits are undeniable for building intuition about the problems at hand.
This article demonstrated, for instance,
how the method of moments elucidates the problem of minimal projections onto polynomial spaces
and it strongly hinted that minimal projections are not unique and do not preserve shape.
Our technique can be extended in several directions,
as long as linear-programming upper bounds and semidefinite-programming lower bounds match up to a desired accuracy.
For example, given subspaces $\mathcal{U} \subseteq \mathcal{V}$ of $\mathcal{C}[-1,1]$,
one could find a linear map $P$ from $\mathcal{C}[-1,1]$ into $\mathcal{V}$, not $\mathcal{U}$,
with minimal norm among those satisfying $P(u) = u$ for all $u \in \mathcal{U}$,
or even satisfying $P(u) = F(u)$ for all $u \in \mathcal{U}$ with some $F \not= {\rm Id}_{\mathcal{U}}$,
and shape-preservation properties may be added, etc...
A particularly interesting situation concerns multivariate polynomial spaces,
for which nothing is known except the results of \cite{SS}.
Conceptually, the same technique applies,
but one quickly runs into numerical limitations.
One could throw in more computational power,
of course,
but it seems wiser to refine the method first,
possibly with a back-and-forth process between upper bound and lower bound.
|
\section{Background}
The classical wave equation and its time harmonic counterpart, the Helmholtz equation, provide accurate mathematical models for
acoustic wave propagation under a wide range of conditions.
Effects that are not accounted for in the linear regime are mainly related to various loss mechanisms, manifested in two different ways.
On the one hand there is bulk loss, which is a consequence of bulk shear, heat conduction and molecular exchange of energy.
Bulk losses are small and have in general an appreciable effect only for propagation over long distances~\cite[\S~6.4]{MoIn68}.
Boundary effects, on the other hand, are due to heat exchange with walls and viscous dissipation owing to the shear motion caused by the contact or non-slip boundary condition at the wall boundary.
The relative importance of the thermal and viscous losses varies with the type of medium, the wavelength, and the
characteristic size of the domain.
In air, thermal and viscous boundary effects are of the same order of magnitude at audio frequencies, since the Prandtl number for air is close to unity.
For devices such as hearing aids, microphones, and micro-loudspeakers, these effects can have a great influence on the generated or detected sound power level.
Another type of devices that often cannot be modeled with sufficient accuracy without a satisfactory damping model are musical wind instruments. Acoustic damping is also important when studying the damped vibrations of Micro Electro Mechanical Systems (MEMS) structures.
As we will see in the analysis of \S~\ref{BLT}, a main feature of systems for which visco-thermal losses need to be taken into account is that the quotient of the total solid-surface area to the air volume is significant.
All the above-mentioned damping effects may be accounted for by including the appropriate constitutive relations in the linearized, compressible Navier--Stokes equations.
There are software packages that can carry out numerical solutions of the full set of linearized Navier--Stokes equations for acoustic purposes, but the computational cost is generally very high, also for devices that are acoustically small.
The reason for the high computational cost is partly the introduction of four extra variables, three components for the acoustic velocity and one for temperature (or entropy) fluctuations, compared to the classical wave equation formulated in terms of the acoustic pressure only.
An even more serious problem is the viscous and thermal boundary layers, which typically are very thin in relation to the free-space wavelength and the characteristic dimensions of the geometry.
Thus, the computational mesh has to be extremely fine in the vicinity of solid boundaries to resolve the large gradients in the boundary layers.
A recommendation from one of the major software providers~\cite[Ch.~5]{Co17AM} is to hybridize and use the full Navier--Stokes equations only when absolutely necessary, such as in thin slits, and couple these equations to the pressure Helmholtz equation for the rest of the domain.
There is a long history of efforts to approximately account for boundary losses, going back at least to Kirchhoff~\cite{Ki68}.
One approach has been to consider particular geometries for which exact or approximate solutions to the linearized Navier--Stokes equations can be established.
For waveguides, the solutions is typically represented in a 1D analysis by a complex propagation constant $\phi$, so that the acoustic pressure at axial position $z$ satisfies $p(z,t) = (A\textrm e^{-\phi z} + B\textrm e^{\phi z})\textrm e^{\i\omega t}$, where $A$ and $B$ are amplitudes of waves propagating in the positive and negative $z$-direction, respectively.
For the case of propagation in a circular pipe, Tijdeman~\cite{Ti75} reviews and summarizes a large number of results in terms of four nondimensional parameters.
With a similar 1D-analysis, Richards~\cite{Ri86} covers the case of waves propagating between two infinite parallel plates and also, approximately, certain non-cylindrical tube geometries.
The case of arbitrary cross sections requires numerical solutions in general~\cite{Cu93}.
Another approach, also with a long history, is to use boundary-layer analysis.
Two recent expositions of the technique are by Rienstra \& Hirschberg~\cite[\S~4.5]{RiHi15} and Searby et al.~\cite{SeNiHaLa08}.
A typical use of the boundary-layer analysis has been to calculate the propagation constant for waveguides in the limit of a large radii~\cite[\S~4.5.3]{RiHi15}.
As opposed to the approach reviewed by Tijdeman~\cite{Ti75}, the results of the boundary-layer analysis will be independent of the cross-section shape, but the results are not valid for small radii in the order of the boundary-layer thickness.
Another use of the boundary-layer analysis is suggested by Searby et al.~\cite{SeNiHaLa08}, who propose a two-step procedure for calculation of the total visco--thermal losses in cavities.
The first step consists of numerically solving the Helmholtz equation for the acoustic pressure in the whole cavity.
With this isentropically calculated pressure at the boundary as input data, the total visco--thermal boundary losses are then computed in a second step using boundary-layer theory.
Bossart et al.~\cite{BoJoBr03} takes this idea one step further in a predictor--corrector manner and propose to recompute the outer problem using the boundary-layer solution in order to modify the wavenumber and an admittance boundary condition.
The aim of this article is to derive and propose a boundary condition that can be supplied to isentropic acoustics models, such as the Helmholtz equation for the acoustic pressure, in order to account for visco--thermal boundary losses in numerical simulations.
The derivation is based on a boundary-layer analysis of the linearized, compressible Navier--Stokes equations.
The basic idea is to rewrite the equation of mass conservation in the boundary layer as an equation with constant acoustic density and pressure, which would be the case under isentropic conditions, and compensate for the error, to first order, by a modified wall-normal boundary condition for the acoustic velocity.
For the pressure Helmholtz equation, the final boundary condition turns out to be a so-called Wentzell (or Venttsel') condition, a generalization of an impedance (or Robin) boundary condition, including a surface Laplacian of the pressure.
The starting point of the derivation is a flat-wall boundary-layer analysis, and it therefore assumes that the boundary-layer thickness is small compared to the isentropic wavelength as well as the maximal radius of curvature of the wall.
Based on the asymptotic analysis by Schmidt et al.~\cite{ScThJo14}, and using a different approach than ours, the viscous--but not the thermal--part of the proposed boundary conditions have previously been derived in a quite recent report by Schmidt and Thöns--Zueva~\cite{ScTh14}.
The difficulty with boundary layers whose resolution demands a very fine mesh, thus requiring significant computational resources, are present also in many turbulent flow simulations.
Common approaches to reduce the computational burden here include modelling the velocity within the boundary layer using so called wall functions~\cite{LaSp74}, or setting a modified Dirichlet~\cite{BaMiCaHu07} or Robin~\cite{JoUt15} boundary condition for the velocity a short distance away from the wall.
Although with a similar motivation as in the present study, the form of the boundary conditions, and their applications to turbulence models outside the boundary layer, are entirely different.
We show that a typical problem setup for cavity acoustics using the proposed boundary conditions leads to a well-posed mathematical problem.
Moreover, we show that our model yields the same expression for the total visco--thermal power losses in a cavity as discussed by Searby et al.~\cite{SeNiHaLa08} and, in the special case of propagation in wave guides, that our model yields the same dispersion relation as derived by Rienstra \& Hirschberg~\cite[\S~4.5]{RiHi15}.
The limits of applicability of our model is quantitatively assessed in the case of propagation in narrow wave guides with a circular cross section, a case in which exact solution of the linearized Navier--Stokes equations are available.
The proposed boundary-condition is straightforward to implement in an existing finite-element software, which we demonstrate by implementing it using Comsol Multiphysics' so-called Weak Form PDE Interface~\cite[Ch.~16]{Co17Ref}.
Our model is tested on the case of wave propagation through a compression driver, a type of transducer used to feed mid-range horns, and we show that visco--thermal losses are important to account for in this application and that the proposed model yields essentially the same results on the transmitted power spectrum as a simulation using a hybrid Navier--Stokes/Helmholtz approach at a fraction of the computation cost in terms of CPU time and computer memory usage.
\section{Visco--thermal acoustic equations}
Since our focus is on acoustics in air under atmospheric conditions, the starting point for the modeling is the compressible Navier--Stokes equations together with standard constitutive relations.
That is, air is regarded as an ideal gas and a Newtonian fluid with constant viscosity and satisfying the Stokes hypothesis, and the heat flux satisfies Fourier’s law with a constant thermal conductivity.
Linearizing the compressible Navier--Stokes equations around quiescent air at static pressure, density, and temperature $p_0$, $\rho_0$, and $T_0$, we obtain in frequency domain the system
\begin{subequations}\label{linNS}
\begin{align}
\i\omega\rho + \rho_0\nabla\cdot\bs U &= 0, \label{linMassC}
\\
\i\omega\bs U + \frac1{\rho_0}\nabla p - \nu\big(\Delta\bs U + \frac13\nabla(\nabla\cdot\bs U)\big)
&= \bs 0, \label{linMomC}
\\
\i\omega\rho_0 c_V T + p_0\nabla\cdot\bs U - \kappa\Delta T &= 0,
\label{linEnerC}
\end{align}
\end{subequations}
where $\rho$, $p$, $\bs U$, and $T$ are the complex amplitudes of the acoustic disturbances in density, pressure, velocity, and temperature, $\omega$ the angular frequency, $\nu$ the kinematic viscosity coefficient, $c_V$ the specific heat capacity at constant volume, and $\kappa$ the thermal conductivity.
We use the phase convention $\textrm e^{\i\omega t}$, so that in time domain, the acoustic pressure, for instance, will be $P(\bs x, t) = \Re p(\bs x)\textrm e^{\i\omega t}$, and we assume that system~\eqref{linNS} is driven through an inhomogeneous boundary condition or through a wave originating in the far field.
The static conditions satisfy the ideal gas law
\begin{equation}\label{idealgas}
p_0 = r \rho_0 T_0,
\end{equation}
where the gas constant $r$ is given by the difference of the specific heats at constant pressure and volume,
\begin{equation}\label{r}
r = c_p - c_V,
\end{equation}
Linearizing the ideal gas law, we find that the acoustic disturbances satisfy
\begin{equation}\label{linIGL}
\frac p{p_0} = \frac\rho{\rho_0} + \frac T{T_0}.
\end{equation}
Moreover, the speed of sound $c$ satisfies
\begin{equation}\label{c2}
c^2 = \gamma\frac{p_0}{\rho_0},
\end{equation}
where
\begin{equation}\label{gamma}
\gamma=\frac{c_p}{c_V}.
\end{equation}
Since the viscosity
and thermal conductivity coefficients are very small, visco--thermal effects may in most parts of the domain be ignored, which leads to isentropic conditions, and thus that the speed of sound in these particular regions (but not generally) will satisfy
\begin{equation}\label{c2ie}
c^2 = \frac{p}{\rho}.
\end{equation}
As a consequence, under isentropic assumptions, the variables $p$, $\rho$, and $T$ will, by equation~\eqref{linIGL}, be proportional to each other, and the system~\eqref{linMassC}--\eqref{linMomC} will then reduce to the following wave equation in first-order, frequency-domain form,
\begin{subequations}\label{wave1ord}
\begin{align}
\frac{\i\omega }{c^2}p + \nabla\cdot\rho_0\bs U &= 0,
\label{wave1ord_mc} \\
\i\omega\rho_0 \bs U + \nabla p
&= \bs 0, \label{wave1ord_momc}
\end{align}
\end{subequations}
which, after eliminating $\bs U$, can be written as a Helmholtz equation solely in the pressure,
\begin{equation}\label{Helmholtz}
-\Delta p - k_0^2 p = 0,
\end{equation}
where $k_0 = \omega/c$ is the isentropic wavenumber.
However, the isentropic assumptions break down in the vicinity of solid walls modeled with the non-slip and isothermal boundary conditions
\begin{equation}\label{NSBC}
\bs U =\bs 0,
\qquad
T = 0,
\end{equation}
which can be seen by the fact that it is only possible to set a vanishing normal component of the velocity, $\bs n\cdot\bs U=0$, as a boundary condition to system~\eqref{wave1ord}.
In the vicinity of a solid wall, normal derivatives of the tangential velocity and the temperature will be large enough so that not all visco--thermal terms can be ignored in equation~\eqref{linNS}.
In order to account for the wall effects of viscosity and thermal conductivity, the isentropic system~\eqref{wave1ord} can in a narrow region close to the wall be replaced with a system of boundary-layer equations.
Assume that there is a flat solid wall located at the plane $y=0$, on which the non-slip and isothermal boundary conditions~\eqref{NSBC} should be imposed.
In a region close to the wall measured in terms of
the boundary-layer thicknesses defined below,
the full system~\eqref{linNS} can be approximated, as derived in appendix~\ref{appBLeq}, with the acoustic boundary layer equations
\begin{subequations}\label{BLeqns}
\begin{align}
\i\omega\frac\rho{\rho_0} + \nabla_\text{T}\cdot\bs u + \pder vy1 &= 0, \label{BLeqnsMassC}
\\
\i\omega\bs u + \frac1{\rho_0}\nabla_\text{T} p
- \nu \pder{\bs u}y2 &= \bs 0, \label{BLeqnsMomCT}
\\
\pder py1 = 0,
\\
\i\omega\rho_0 c_V T + p_0\left(\nabla_\text{T}\cdot\bs u + \pder vy1\right)
- \kappa \pder Ty2 &= 0,
\label{BLeqnsEnerC}
\end{align}
\end{subequations}
where $\bs u = (u, 0, w)$ is the projection of the velocity vector $\bs U = (u, v, w)$ on the wall plane, and $\nabla_\text{T}$ the corresponding projection of the operator $\nabla$ , that is,
\begin{equation}
\begin{aligned}
\nabla_\text{T} = \left(\pder{}x1, 0, \pder{}z1\right).
\end{aligned}
\end{equation}
In appendix~\ref{appBLeq}, we construct exact solutions to system~\eqref{BLeqns} that satisfy the boundary conditions~\eqref{NSBC} and that exponentially, as $y\to+\infty$, approach the fields $\bs u^\infty(\bs r)$ and $p^\infty(\bs r)$, where $\bs r = (x, 0, z)$, which are assumed to be solutions to the isentropic equations~\eqref{wave1ord} evaluated at a point close to the wall but outside of the boundary layer.
These boundary-layer solutions can be written
\begin{subequations}\label{BLsol}
\begin{align}
\bs u &= \bs u^\infty(\bs r)\left( 1 - \textrm e^{-(1+\i)y/\delta_V}\right), \label{BLusol}
\\
\frac{\rho_\text{e}}{\rho_0} &= \frac{\gamma-1}{\gamma}\frac{p^\infty(\bs r)}{p_0}\textrm e^{-(1+\i)y/\delta_T}, \label{BLrhosol}
\end{align}
\end{subequations}
where
\begin{equation}\label{excessdens}
\frac{\rho_\text{e}}{\rho_0} = \frac1{\rho_0}\left(\rho - \rho^\infty \right)= \frac1{\rho_0}\left(\rho - \frac{p^\infty}{c^2}\right),
\end{equation}
is called the \emph{excess density}, and
\begin{equation}\label{dVdTdef}
\delta_V = \sqrt{\frac{2\nu}{\omega}},
\qquad
\delta_T =\sqrt{\frac{2\kappa}{\omega\rho_0 c_p}}
\end{equation}
the viscous and thermal \emph{acoustic boundary-layer thicknesses}.
Note that the boundary-layer solution as defined above, due to the small values of $\delta_V$ and $\delta_T$ quickly approaches the limit fields $\bs u^\infty$ and $p^\infty$, but that these values are not attained at any finite distance from the wall.
It would be possible to alter the approach and define a matched asymptotic expansion, but our purpose here is different; we will use the form of the boundary-layer solution in order to define an effective boundary condition to the isentropic equations to account for the boundary-layer effects.
\begin{remark}
The formation of boundary layers is due to the structure of the Navier--Stokes equations~\eqref{linNS} as a singularly-perturbed system.
That the thickness of the acoustic boundary layers in expressions~\eqref{dVdTdef} scales as the square root of the coefficients in the governing equations is a property that generally holds for layers associated with singularly-perturbed equations.
For instance, the thickness of the classical Prandtl-type of viscous boundary layer that develops over a flat plat subject to a steady free-stream flow parallel to the plate also scales as the square root of the viscosity.
However, in the Prandtl layer, the boundary-layer profile is not exponential like in expression~\eqref{BLusol}, and its thickness also grows as the square root of the distance from the leading edge~\cite{Schl87}.
\end{remark}
\section{Boundary layer effects modeled as a boundary condition}\label{BLasBC}
Our aim is to approximate the impact of the boundary layer with an effective wall boundary condition, which will be obtained by manipulations of the mass conservation law.
The boundary-layer analysis described above provided formulas~\eqref{BLsol} for the tangential velocity and the excess density within the boundary.
Corresponding wall-normal velocity will be an order of magnitude smaller than the tangential velocity, as can be seen from the scalings~\eqref{scalings} used to derive the boundary-layer equations.
Nevertheless, the wall-normal velocity $v$ at an arbitrary position $y = \tilde y$ within the boundary layer can be computed from the other variables by integrating equation~\eqref{BLeqnsMassC},
\begin{equation}\label{intBLMB}
\i\omega\int\limits_0^{\mathclap{\tilde y}}\frac\rho{\rho_0}\,\textit{dy}
+ \int\limits_0^{\mathclap{\tilde y}} \nabla_\text{T}\cdot\bs u\, \textit{dy} + v|_{y=\tilde y} - v|_{y=0} = 0
\end{equation}
where $v|_{y=0} = 0$ due to the non-slip boundary condition~\eqref{NSBC}, and where $\rho$ and $\bs u$ exponentially approach $\rho^\infty$ and $\bs u^\infty$ for increasing $\tilde y$.
Subtracting and adding $\rho^\infty/\rho_0$ (which is a function of wall position $\bs r$ only), using definition~\eqref{excessdens} and that $v|_{y=0} = 0$, we find that equation~\eqref{intBLMB} can be written
\begin{equation}\label{intBLMBb}
\i\omega\int\limits_0^{\mathclap{\tilde y}}\frac{\rho^e}{\rho_0}\,\textit{dy}
+\i\omega\int\limits_0^{\mathclap{\tilde y}}\frac{\rho^\infty}{\rho_0}\,\textit{dy}
+ \int\limits_0^{\mathclap{\tilde y}} \nabla_\text{T}\cdot\bs u\, \textit{dy} + v|_{y=\tilde y} = 0
\end{equation}
An integration of the first term in equation~\eqref{intBLMBb}, using formula~\eqref{BLrhosol}, yields
\begin{equation}\label{intrho}
\begin{aligned}
\i\omega\int\limits_0^{\mathclap{\tilde y}}\frac{\rho_\text{e}}{\rho_0}\,\textit{dy}
=
\delta_T\frac{\omega(\gamma-1)(1+\i)}{2\gamma p_0}p^\infty
\left( 1 - \textrm e^{-(1+\i)\tilde y/\delta_T}\right).
\end{aligned}
\end{equation}
Moreover, by expression~\eqref{BLusol}, we find that the third term in equation~\eqref{intBLMBb} can be evaluated as
\begin{equation}\label{intu}
\begin{aligned}
\int\limits_0^{\mathclap{\tilde y}} \nabla_\text{T}\cdot\bs u\, \textit{dy}
&= \int\limits_0^{\tilde y}\nabla_\text{T}\cdot\bs u^\infty\lp1 - e^{-(1+i)y/\delta_V}\right)\,\textit{dy}
\\&
=\int\limits_0^{\tilde y}\nabla_\text{T}\cdot\bs u^\infty\,\textit{dy} +
{\delta_V}\frac{1-\i}2 \nabla_\text{T}\cdot\bs u^\infty\left(\textrm e^{-(1+i)\tilde y/\delta_V }- 1\right).
\end{aligned}
\end{equation}
Substituting expressions~\eqref{intrho} and~\eqref{intu} into equation
~\eqref{intBLMBb}, we find that
\begin{equation}\label{intBLMBa}
\begin{aligned}
&\i\omega\int\limits_0^{\mathclap{\tilde y}}\frac{\rho^\infty}{\rho_0}\,\textit{dy}
+ \int\limits_0^{\tilde y}\nabla_\text{T}\cdot\bs u^\infty\,\textit{dy}
+ v|_{y=\tilde y}
+ \delta_V\frac{\i-1}{2}\nabla_\text{T}\cdot\bs u^\infty\left( 1 - \textrm e^{-(1+\i)\tilde y/\delta_V}\right)
\\&\qquad\qquad
+ \delta_T\frac{\omega(\gamma-1)(1+\i)}{2\gamma p_0}p^\infty
\left( 1 - \textrm e^{-(\i+1)\tilde y/\delta_T}\right)=0,
\end{aligned}
\end{equation}
which can be written
\begin{equation}\label{intBLMBc}
\begin{aligned}
&\i\omega\int\limits_0^{\mathclap{\tilde y}}\frac{\rho^\infty}{\rho_0}\,\textit{dy}
+ \int\limits_0^{\tilde y}\nabla_\text{T}\cdot\bs u^\infty\,\textit{dy}
+ v|_{y=\tilde y} + f(\tilde y) - v_\text{W} = 0,
\end{aligned}
\end{equation}
where
\begin{subequations}\label{BLvs}
\begin{align}
f(y) &=
-\delta_V\frac{\i-1}{2}\nabla_\text{T}\cdot\bs u^\infty \textrm e^{-(1+\i) y/\delta_V}
- \delta_T\frac{\omega(\gamma-1)(1+\i)}{2\gamma p_0}p^\infty \textrm e^{-(\i+1) y/\delta_T},
\label{BLvsy}\\
v_\text{W} = f(0)&=
- \delta_V\frac{\i-1}{2}\nabla_\text{T}\cdot\bs u^\infty
- \delta_T\frac{\omega(\gamma-1)(1+\i)}{2\gamma p_0}p^\infty.
\label{BLvsW}
\end{align}
\end{subequations}
The function $f$ is of $O(\delta_V + \delta_T)$ at the wall and decays exponentially with its argument.
We thus find that the integrated mass conservation law under boundary-layer approximation, equation~\eqref{intBLMB}, can be written as equation~\eqref{intBLMBc}, which, if the term $f(\tilde y)$ is ignored, all the effects of the boundary layer has been pushed into $v_\text{W}$.
Recall that the solution to the isentropic system~\eqref{wave1ord} will be essentially constant in the normal direction close to a solid wall due to the lack of boundary layers.
Thus, equation~\eqref{intBLMBc} is essentially an integral form of the isentropic mass conservation law~\eqref{wave1ord_mc} (recall that $p= c^2\rho$ under isentropic assumptions) in which the wall normal velocity~\eqref{BLvsW} is a perturbation of the non-penetration condition $\bs n\cdot\bs U = 0$ with coefficients of $O(\delta_V + \delta_T)$.
The form of equation~\eqref{intBLMBc} and expression~\eqref{BLvsW} suggest that boundary layer effects could be taken into account by simply solving the isentropic system~\eqref{wave1ord} and replacing the normal isentropic wall boundary condition $\bs n\cdot \bs U = 0$ with
\begin{equation}
\bs n\cdot\bs U = -v_W = \delta_V\frac{\i-1}{2}\nabla_\text{T}\cdot\bs U
+\delta_T\frac{\omega(\gamma-1)(1+\i)}{2\gamma p_0}p
\qquad\text{at $y=0$,}
\end{equation}
where we have used that $\nabla_\text{T}\cdot\bs U = \nabla_\text{T}\cdot \bs u$.
We thus propose the system
\begin{subequations}\label{1ordBLBC}
\begin{align}
\frac{\i\omega }{c^2}p + \nabla\cdot\rho_0\bs U &= 0 &&\text{for $y>0$}, \label{1ordBLBC_mc}
\\
\i\omega\rho_0 \bs U + \nabla p &= 0 &&\text{for $y>0$},
\label{1ordBLBC_momc}
\\
\bs n\cdot\bs U &= \delta_V\frac{\i-1}{2}\nabla_\text{T}\cdot\bs U
+ \delta_T\frac{\omega(\gamma-1)(1+\i)}{2\gamma p_0}p &&\text{at $y=0$.} \label{1ordBLBC_bc}
\end{align}
\end{subequations}
as a model for acoustic wave propagation over a wall at $y=0$ where thermal and viscous boundary layers form.
Instead of working with the first-order system~\eqref{1ordBLBC}, we suggest, for two reasons, to rewrite it as a second-order equation in $p$.
First, for numerical purposes, the number of unknowns will then be reduced to one scalar variable that can be treated with standard finite elements.
Second, the mathematical analysis of the boundary-value problem appears more straight-forward in the second-order formulation.
To work out the boundary condition for the second order formulation that corresponds to condition~\eqref{1ordBLBC_bc}, we start by evaluating the wall normal component and the tangential divergence of equation~\eqref{wave1ord_momc} at the limit values for the boundary-layer,
\begin{subequations}\label{1ord_momc_split}
\begin{align}
\i\omega\rho_0 \bs n\cdot\bs U ^\infty+ \pder {p^\infty}n1 &= 0,
\\
\i\omega\rho_0 \nabla_\text{T}\cdot\bs U^\infty + \Delta_\text{T} p^\infty &= 0.
\end{align}
\end{subequations}
Eliminating the velocity from system~\eqref{1ordBLBC}, where expressions~\eqref{1ord_momc_split} are used for the boundary condition~\eqref{1ordBLBC_bc}, we obtain the following second-order alternative to~\eqref{1ordBLBC},
\begin{subequations}\label{Hhsystem1}
\begin{align}
-k_0^2p - \Delta p &= 0
&&\text{for $y>0$,}
\\
-\delta_V\frac{\i-1}2 \Delta_\text{T} p + \delta_T k_0^2\frac{(\i-1)(\gamma-1)}2 p + \pder pn1 &= 0 \label{Hhsystem1_bc}
&&\text{at $y=0$.}
\end{align}
\end{subequations}
The solid-wall boundary conditions~\eqref{1ordBLBC_bc} or~\eqref{Hhsystem1_bc} are derived under the assumption of a flat wall, which makes the splitting in tangential and normal directions particularly easy.
However, such a splitting can also be carried out in the case of a smooth non-flat surface.
In that case, the normal field vector $\bs n$ can be extended into the inside of the domain in the vicinity of the wall using the definition
\begin{equation}
\bs n (\bs x) = \frac{\nabla d(\bs x)}{\lvert\nabla d(\bs x)\rvert},
\end{equation}
where $d(\bs x)$ is the \emph{wall distance function}~\cite{KrPa99}.
Taking derivatives in the directions of this extended normal field, we can in the vicinity of the wall split the gradient and divergence operators in their normal and tangential parts, analogously as in the case of a flat wall,
\begin{subequations}\label{tangopt}
\begin{align}
\nabla T &= \nabla_\text{T} T + \bs n\pder Tn1,
\\
\nabla\cdot\bs U &= \nabla_\text{T}\cdot\bs u + \bs n \cdot\pder{\bs U}n1,
\end{align}
\end{subequations}
where $\bs u = \bs U - (\bs U\cdot\bs n)\bs n $ and the tangential operators simply are defined through the expressions above.
However, the splitting of the Laplacian operator, needed in the derivation of the boundary-layer equations, is more complicated in the curved-wall case,
\begin{equation}
\Delta T = \Delta_\text{T} T + \pder Tn2 + H\pder Tn1 ,
\end{equation}
involving an extra term with $H = \nabla_\text{T}\cdot\bs n$, the sum of the principal curvatures of the level surface to $d$ that passes through the point of interest.
However, in many practical situations with a smooth wall, it is reasonable to assume that the minimal radius of the principal curvatures is much larger than the boundary layer thicknesses, which is of the order of 20--400~\textmu{m} in the audio range.
We will therefore here apply boundary conditions~\eqref{1ordBLBC_bc} or~\eqref{Hhsystem1_bc} also for nonplanar smooth boundaries, interpreting the tangential operators as in definitions~\eqref{tangopt}.
Due to the likely scale separation between boundary-layer thickness and the wall's radii of curvature, we conjecture that taking wall curvature into account in the boundary conditions would constitute in many cases at most a second-order correction to the boundary conditions derived above.
\section{Example: a cavity with lossy walls}
Here we exemplify the use of wall boundary condition~\eqref{Hhsystem1_bc} in the context of an acoustic cavity problem.
Let the domain of the cavity, conceptually illustrated in figure~\ref{cavitypicture}, be denoted $\Omega\subset\mathbb R^3$.
The cavity boundary $\partial\Omega$ consists of a solid-wall part $\Gamma_\text{w}$ and a part $\Gamma_\text{io}$ where sound waves can enter and exit.
We assume that both boundary parts are smooth and, as discussed in \S~\ref{BLasBC}, that the radii of the principal curvatures of the surface $\Gamma_{\text w}$ everywhere is much larger than $\delta_V$ and $\delta_T$.
Possible edges and corners of the cavity can therefore only be located at the interfaces between $\Gamma_\text{io}$ and $\Gamma_\text{w}$.
The acoustic pressure amplitude in the cavity is then modeled by the boundary-value problem
\begin{subequations}\label{cavitysystem}
\begin{alignat}{2}
-k_0^2p - \Delta p &= 0
&\qquad&\text{in $\Omega$,} \label{cavitysystem_eq}
\\
\i k_0 p + \pder pn1 &= 2\i k_0 g
&&\text{on $\Gamma_\text{io}$.}
\label{cavitysystem_io}
\\
-\delta_V\frac{\i-1}2 \Delta_\text{T} p + \delta_T k_0^2\frac{(\i-1)(\gamma-1)}2 p + \pder pn1 &= 0 \label{cavitysystem_w}
&&\text{on $\Gamma_\text{w}$,}
\\
\bs n_\text{T}\cdot\nabla_\text{T} p &= 0
&&\text{on $\partial\Gamma_\text{w}$,} \label{cavitysystem_wbc}
\end{alignat}
\end{subequations}
Boundary condition~\eqref{cavitysystem_io} is a simple radiation (or impedance) condition, in which function $g$ supplies an incoming acoustic wave and where outgoing planar waves are absorbed.
Since wall boundary condition~\eqref{cavitysystem_w} in itself constitutes a diffusion problem on the bounded surface $\Gamma_\text{w}$, an extra condition---``a boundary condition to the boundary condition''---is needed to close the system.
Note that if $\Gamma_{\text w}$ would constitute the whole boundary, no such condition would be needed.
Here we choose perhaps the simplest alternative, the homogeneous Neumann condition~\eqref{cavitysystem_wbc}, where $\bs n_\text{T}$ denotes the outward-directed unit normal on the boundary of the surface $\Gamma_{w}$.
(Note that $\bs n_\text{T}$ is directed in the \emph{tangent} direction of $\Gamma_\text{w}$).
Condition~\eqref{cavitysystem_wbc} will constitute a ``natural condition'' in the variational form and the power balance derived below; the interface $\partial\Gamma_{\text w}$ will be transparent in both expressions.
This case is the one treated in the well-posedness analysis of Appendix~\ref{wellposed}.
It would be mathematically possible instead to specify a Robin condition like $\alpha p + \bs n_\text{T}\cdot\nabla_\text{T} p = 0$.
In that case, an integral over $\partial\Gamma_{\text w}$ would appear in the variational form and the power balance, indicating a sink or source of power at $\partial\Gamma_{\text w}$.
However, it is not clear if this condition makes physical sense, and there is no analysis to guide the choice of coefficient $\alpha$.
Assigning a Dirichlet condition at $\partial\Gamma_{\text w}$ is problematic, however, since it likely will lead to a jump discontinuity towards the $\Gamma_\text{io}$ side of the boundary and thus sharp gradients locally around $\partial\Gamma_{\text w}$.
The well-posedness theory in Appendix~\ref{wellposed} is not easily extended to the Dirichlet case, and it is not clear for us whether the case is amenable to analysis at all.
Regardless of these mathematical issues, note that the boundary-layer approximations considered here breaks down at interfaces such as $\Gamma_{\text w}$, as well as at sharp corners within $\Gamma_\text{w}$, so ultimately, how to handle such interfaces is a modeling issue that is an interesting subject for further studies.
\begin{figure}\centering
\includegraphics{cavitypicture}
\caption{Example cavity for problem~\eqref{cavitysystem}, viewed from the outside.
The interior is the domain $\Omega$.
Waves may enter and exit through the surfaces $\Gamma_\text{io}$, whereas the rest of the boundary $\Gamma_{\text w}$ consists of solid material, where thermal and viscous boundary layers form.
}\label{cavitypicture}
\end{figure}
Multiplying equation~\eqref{cavitysystem_eq} with a test function $q$, integrating by parts, and substituting boundary conditions~\eqref{cavitysystem_io}--\eqref{cavitysystem_wbc}, we find that solutions to the system~\eqref{cavitysystem} satisfies the variational expression
\begin{equation}\label{BLvarform}
\begin{aligned}
&-k_0^2\int_\Omega q p
+ \int_\Omega\nabla q\cdot\nabla p
+ \i k_0 \int_{\Gamma_\text{io}} q p
+ \delta_Tk_0^2\frac{(\i-1)(\gamma-1)}2\int_{\Gamma_\text{w}} q p
\\
&\qquad\qquad
+ \delta_V\frac{\i-1}2 \int_{\Gamma_\text{w}}\nabla_\text{T} q\cdot\nabla_\text{T} p
= 2\i k_0 \int_{\Gamma_\text{io}} q g.
\end{aligned}
\end{equation}
\begin{remark}
In this section, as well as in Appendix~\ref{wellposed}, we do not explicitly specify measure symbols (such as $\textit{dV}$ or $\textit{dS}$, for instance) in the integrals, since the type of measure will be clear from the domain of integration.
\end{remark}
In Appendix~\ref{wellposed}, we define weak solutions to system~\eqref{cavitysystem} using the variational form~\eqref{BLvarform} and show that the associated variational problem is well posed.
Variational expression~\eqref{BLvarform} is also the starting point for finite-element discretizations, which can be carried out using the same standard elements as employed for the Helmholtz equation, that is, with finite-element functions that are globally continuous and polynomials on each element of the triangulation.
Variational expression~\eqref{BLvarform} can also be used to derive a power balance law for system~\eqref{cavitysystem}, as follows.
Inserting the test function $q = \bar p$ (complex conjugate) in expression~\eqref{BLvarform}, we find that
\begin{equation}\label{e_VPp}
\int\limits_{\Omega}\lvert\nabla p\rvert^2 - k_0^2 \int\limits_\Omega \lvert p\rvert^2
+ \i k_0\int\limits_{\mathclap{\Gamma_\text{io}}} \lvert p\rvert^2
+ \delta_Tk_0^2 \frac{(\i-1)(\gamma-1)}{2}\int\limits_{\Gamma_\text{w}} \lvert p\rvert^2
+ \delta_V\frac{\i-1}{2}\int\limits_{\Gamma_\text{w}}\lvert\nabla_\text{T} p\rvert^2 = 2\i k_0 \int\limits_{\Gamma_\text{io}}\bar p g.
\end{equation}
The imaginary part of equation~\eqref{e_VPp} divided by $k_0$ is
\begin{equation}\label{e_impart}
\int\limits_{\mathclap{\Gamma_\text{io}}} \lvert p\rvert^2
+ \delta_Tk_0\frac{\gamma-1}{2}\int\limits_{\Gamma_\text{w}} \lvert p\rvert^2
+ \frac{\delta_V}{2k_0}\int\limits_{\Gamma_\text{w}}\lvert\nabla_\text{T} p\rvert^2
= 2\Re \int\limits_{\Gamma_\text{io}}\bar p g.
\end{equation}
Substituting identity
\begin{equation}
\lvert p - g\rvert^2 = \lvert p\rvert^2 + \lvert g \rvert^2 - 2\Re \bar p g,
\end{equation}
into equality~\eqref{e_impart} and dividing by $2\rho_0 c$, to obtain terms in units of power, we obtain
\begin{equation}\label{e_powbal}
\frac1{2\rho_0c} \int\limits_{\Gamma_\text{io}} \lvert g\rvert^2
= \frac1{2\rho_0c} \int\limits_{\Gamma_\text{io}} \lvert p - g\rvert^2
+ (\gamma-1)\frac{\delta_T\omega}{4\rho_0 c^2}\int\limits_{\Gamma_\text{w}}\lvert p\rvert^2
+ \frac{\delta_V}{4\omega\rho_0} \int\limits_{\Gamma_\text{w}}\lvert\nabla_\text{T} p\rvert^2.
\end{equation}
Expression~\eqref{e_powbal} expresses that the incoming power equals the sum of the reflected power and the power losses due to the thermal and viscous boundary layers.
\section{Comparisons with classical results}
\subsection{Boundary-layer theory}\label{BLT}
Searby et al.~\cite{SeNiHaLa08} consider the problem of calculating the visco--thermal boundary-layer damping in acoustic cavities.
The authors review the boundary-layer theory and propose a two-step procedure in which a isentropic Helmholtz solver first calculates the pressure distribution on the solid surfaces.
In a second step, the total power loss from the visco-thermal boundary layers is calculated using boundary-layer theory.
The loss given by Searby et al.~\cite[formulas~(6) and~(10)]{SeNiHaLa08} agree with the two last terms in the power balance law~\eqref{e_powbal}.
Note, however, that the procedure of Searby et al.\ does not predict, for instance, phase shift effects of the boundary layers, nor does it provide, as here, an explicit locally-reacting boundary condition, coupled to the interior problem.
Another approach is to consider a one-dimensional analysis, as presented by Rienstra \& Hirsch\-berg~\cite[\S~4.5]{RiHi15}, who analyze the thermal boundary layer in the case of a plane wave at normal incidence towards an isothermal wall and the viscous boundary layer in the case of a plane wave propagating parallel to the wall.
The so-called \emph{displacement thicknesses} that their analysis yield are then applied to the case of a plane wave propagating in a wave guide.
The final result is an expression for the complex wavenumber $k$ that governs the pressure amplitude, assumed to be of the form $p(z) = \textrm e^{-\i k z}$, where the coordinate $z$ is along the axis of the wave guide, and in which the imaginary part of $k$ represents the visco--thermal damping.
We will now show that if boundary-value problem~\eqref{cavitysystem} in its variational form~\eqref{BLvarform} is applied to such a case of a narrow waveguide, we obtain the same expression as Rienstra \& Hirsch\-berg.
We consider the setup illustrated in figure~\ref{waveguidepicture}, where $\Omega = S\times(0, \ell)$ is a cylindrical wave guide of length $\ell$.
The wave guide's cross section $S$ is fixed, of area $|S|$, and has a smooth boundary $\partial S$ of circumference $|\partial S|$.
In order for the boundary-layer approximations to be valid, we assume that $\delta_V$ and $\delta_T$ are small compared to $\sqrt{|S|}$, and that $\sqrt{|S|}$ is of the same order as $|\partial S|$, precluding overly flattened geometries.
Otherwise, the shape of $S$ can be arbitrary.
Let the waveguide be oriented along the $z$-axis and let $\Gamma_\text{io}$ be the two cross section planes located at $x=0$ and $x=\ell$, respectively.
Furthermore, let $g(x,0) = g_0$ and $g(x,\ell) = 0$ for a given number $g_0$.
According to boundary-layer theory and since the driving signal $g_0$ is constant at the inlet, we may make the ansatz that the pressure field is constant in each cross section, that is, that $p = p(z)$.
\begin{remark}
The assumption that $p = p(z)$ also constitutes a \emph{Galerkin approximation} of variational form~\eqref{BLvarform} such that $p$ and $q$ are constant over each cross section.
\end{remark}
Under this approximation, $\nabla p = \nabla_\text{T} p = \bs e_z p'$, where $\bs e_z$ is a unit vector in the $z$ direction, and variational form~\eqref{BLvarform} reduces to
\begin{equation}\label{BL1Dvarform}
\begin{aligned}
&\int_0^{\ell} q'\left[\lvert S\rvert + \delta _V\frac{\i-1}2\lvert\partial S\rvert\right] p'\,\textit{dz}
-k_0^2\int_0^{\ell} q\left[ \lvert S\rvert - \delta_T\frac{\i - 1}2(\gamma-1) \lvert\partial S\rvert\right] p\, \textit{dz}
\\
&\qquad + \i k_0 \lvert S\rvert\bigl[ q(\ell)p(\ell) + q(0)p(0)\bigr] = 2\i k_0 q(0) g_0.
\end{aligned}
\end{equation}
Variational expression~\eqref{BL1Dvarform} holds for each test functions such that itself and its derivative is square integrable.
In particular, for smooth test functions that vanish at $x = 0$ and $x = \ell$ (that is, $q\in C_0^\infty(0,\ell)$), variational expression~\eqref{BL1Dvarform} reduces to
\begin{equation}
\int_0^{\ell} q'\left[\lvert S\rvert + \delta _V\frac{\i-1}2\lvert\partial S\rvert\right] p'\,\textit{dz}
-k_0^2\int_0^{\ell} q\left[ \lvert S\rvert - \delta_T\frac{\i - 1}2(\gamma-1) \lvert\partial S\rvert\right] p\, \textit{dz} = 0.
\end{equation}
Integration by parts yield that
\begin{equation}\label{wguideODE}
-\left[\lvert S\rvert + \delta _V\frac{\i-1}2\lvert\partial S\rvert\right] p'' - k_0^2\left[\lvert S\rvert - \delta_T\frac{\i - 1}2(\gamma-1) \lvert\partial S\rvert\right] p = 0
\end{equation}
in $(0,\ell)$.
The assumed boundary conditions imply that solutions of equation~\eqref{wguideODE} are of the form $p(z) = A e^{-i k z}$, which after substitution into equation~\eqref{wguideODE} implies that
\begin{equation}\label{k2BL}
k^2 = k_0^2\frac{2\lvert S\rvert - \delta_T(\i - 1)(\gamma-1) \lvert\partial S\rvert}{2\lvert S\rvert + \delta _V(\i-1)\lvert\partial S\rvert},
\end{equation}
which is the same expression as obtained by Rienstra \& Hirsch\-berg~\cite[\S~4.5.3]{RiHi15}.
Note that expression~\eqref{k2BL} reveals that in order for $k$ to differ considerably from $k_0$, the surface-area to air-volume ratio $\lvert \partial S\rvert/\lvert S\rvert$ should be large.
\begin{figure}\centering
\includegraphics{waveguidepicture}
\caption{Example of a cylindrical wave guide of the type considered in \S~\ref{BLT}}
\label{waveguidepicture}
\end{figure}
\subsection{Solutions for special geometries}
Instead of relying on boundary-layer analysis, it is possible to obtain exact or approximate solutions to the linearized Navier--Stokes equations for a few special geometries.
Keefe~\cite{keefe:84}, among others, presents results based on Kirchhoff's classical solution for propagation of the first mode inside an isothermal cylinder with a circular cross section.
The analysis is made in terms of the pressure and volume velocity over cross sections in the cylinder.
The wavelength is assumed to be sufficiently long for only the fundamental mode to propagate.
Note that, as opposed to what was the case in \S~\ref{BLT}, no assumption is made regarding the ratio between the tube radius and the boundary layer thickness.
For a circular tube of radius $a$, expression~(9) given by Keefe~\cite{keefe:84} implies the complex wavenumber
\begin{equation}\label{eq:keefes_k2}
k^2 = k_0^2 \frac{1 + (\gamma-1)F_t}{1-F_v},
\end{equation}
where
\begin{equation}\label{FvFt}
\begin{aligned}
F_v &= \frac{2}{r_v\sqrt{-\i}} \frac{J_1(r_v\sqrt{-\i})}{J_0(r_v\sqrt{-\i})} ,\\
F_t &= \frac{2}{r_t\sqrt{-\i}} \frac{J_1(r_t\sqrt{-\i})}{J_0(r_t\sqrt{-\i})} ,
\end{aligned}
\end{equation}
in which $J_0$ and $J_1$ are the Bessel functions of order 0 and 1 and
\begin{align}
r_v &= a \sqrt{\omega/\nu}=
\sqrt{2}a/\delta_V,
\\
r_t &= r_v\sqrt{\nu \rho_0 c_p/\kappa} =
\sqrt{2} a/\delta_T.
\end{align}
We will now show that expression~\eqref{eq:keefes_k2} approaches the boundary layer expression~\eqref{k2BL} in the large radius or high frequency limit.
Such an analysis requires consideration of the behavior of the Bessel functions in expressions~\eqref{FvFt} in the limit of large arguments~\cite[Eq.~9.2.1]{AbSte72}.
For the Bessel functions of a complex argument, asymptotic
expansions are meaningful only for a constant phase angle.
For such a complex argument $w$, $-\pi<\arg(w)<0$,
\begin{equation}\label{J0expansion}
J_0(w) = \frac{1}{\sqrt{2\pi w}}e^{\i(w-\pi/4)} + \mathcal{O}(1/w)
\end{equation}
and
\begin{equation}\label{J1expansion}
J_1(w) = \frac{1}{\sqrt{2\pi w}}e^{\i(w-3\pi/4)} + \mathcal{O}(1/w).
\end{equation}
We can thus use the large-radius/high-frequency approximation
\begin{equation}
\frac{J_1(w)}{J_0(w)}\approx e^{-\i \pi/2} = -\i
\end{equation}
and find that
\begin{equation}
\begin{aligned}
F_v &\approx (1-i)\delta_V/a,\\
F_t &\approx (1-i)\delta_T/a,
\end{aligned}
\end{equation}
whence
\begin{equation}
\begin{split}
k^2 &= k_0^2 \frac{1 + (\gamma-1)F_t}{1-F_v} \approx
k_0^2\frac{1+(\gamma-1)(1-i)\delta_T/a}{1-(1-i)\delta_V/a} =\\
&= k_0^2\frac{2\pi a^2 + (\gamma-1)(1-i)2\pi a\delta_T}{2\pi a^2 - (1-i)2\pi a\delta_V} =
k_0^2\frac{2|S|+(\gamma-1)(1-i)|\partial S|\delta_T}{2|S|-(1-i)|\partial S|\delta_V},
\end{split}
\end{equation}
which means that expression~\eqref{eq:keefes_k2} yields in the large-radius/high-frequency limit the same expression~\eqref{k2BL} as when using the boundary-layer approximations.
\begin{table}\centering
\caption{Air properties used for evaluation of damping models}\label{AirProp}
\begin{tabular}{lll}
\hline
Density & $\rho_0$ & $1.204$ $\text{kg}\cdot\textrm{m}^{-3}$ \\
Kinematic viscosity & $\nu$ & $1.506\cdot10^{-5}$ $\text{m}^2\cdot\text{s}^{-1}$ \\
Speed of sound &$c_0$ & $343.20$ $\text{m}\cdot\text{s}^{-1}$ \\
Prandtl number &$N_\text{Pr}$ & $0.708$ \\
Specific heat, constant pressure & $c_P$ &$1.0054\times10^3$ J$\cdot\text{kg}^{-1}\cdot\text{K}^{-1}$ \\
Ration of specific heats & $\gamma$ & $1.4$ \\
\hline
\end{tabular}
\bigskip
\includegraphics{k_compare_1000um}\includegraphics{k_compare_100um}
\captionof{figure}{The real (blue) and imaginary (red) part of the relative wavenumber for lossy propagation in circular tubes of radii 1~mm (left) and 0.1~mm (right).
Dashed: ``exact'' wavenumber~\eqref{eq:keefes_k2} according to Keefe~\cite{keefe:84}.
Solid: boundary layer approximation~\eqref{k2BL}.
Air property parameters as in table~\ref{AirProp}.}\label{k_compare}
\end{table}
\subsection{Limits of applicability}
The dispersion relation~\eqref{k2BL} was derived from our approach of taking visco--thermal losses into account through boundary condition~\eqref{cavitysystem_w}.
Corresponding expression~\eqref{eq:keefes_k2} holds only for circular cross sections, but it holds also for tubes that are so narrow that the boundary-layer analysis ceases to be valid.
By comparing expressions~\eqref{k2BL} and~\eqref{eq:keefes_k2} for typical parameter values, it will therefore be possible to assess limits of applicability of our boundary-condition approach.
Figure~\ref{k_compare} shows two examples, for tubes of radii 1~mm and 0.1~mm, respectively, of applying the ``exact'' formula~\eqref{eq:keefes_k2} and the boundary-layer approximation~\eqref{k2BL}.
For the 1~mm tube, the agreement is almost perfect, whereas the plot for the 0.1~mm tube reveals that the asymptotics for low frequencies differ, so that the boundary-layer approximation will be inaccurate for frequencies below, say, 2~kHz.
Thus, the boundary-layer approximations do break down, starting at low frequencies, for narrow enough tubes.
However, capillary tubes need to be very long and narrow in order for these inaccuracies to be noticeable.
\section{Numerical case study: a generic compression driver}\label{sec:driver}
This section attempts to demonstrate the usefulness of the above model in practical numerical computations.
As have been alluded to in \S~\ref{BLT}, acoustic devices for which visco--thermal boundary-layer losses need to be taken into account are typically characterized by a large ratio of the area of solid surfaces to the air volume.
One example of such a device is the \emph{compression driver}.
Due to the demands of high acoustic power, such drivers are commonly used in public address systems to feed mid-to-high-frequency horns.
In a compression driver, a stiff vibrating membrane is placed in a very shallow chamber, from which the sound exits typically through a number of narrow slits in the radial or circumferential directions.
The quotient of the membrane area to the sum of the slits' area constitutes the \emph{compression ratio} of the driver.
A high compression ratio improves the acoustic loading of the membrane, particularly at higher frequencies, and substantially increases the efficiency of the driver.
A so-called \emph{phase plug} collects the acoustic output from the slits of the compression chamber and expands it to a circular waveguide, on which the throat of the horn will be mounted.
The design of the slits and the phase plug is delicate in order to minimize the effects of internal resonances~\cite{Oclee-Brown2012}.
The presence of a shallow chamber and several thin slits
means that visco--thermal boundary-layer losses are potentially important to account for in a numerical simulation of a typical compression driver.
However, as we will see, even a hybrid strategy, where the full Navier--Stokes equations are solved only in the narrow passages of the domain, whereas the pressure Helmholtz equation is used for the rest of the system, tends to lead to large problems and simulations which are expensive in terms of memory and CPU time.
Here, for the generic 3-inch compression driver design shown in figure~\ref{driver}, we compare such a hybrid strategy to a strategy where the boundary losses are modeled by the proposed boundary condition \eqref{cavitysystem_w}.
The membrane diameter is 84~mm, the depth of the compression chamber is 0.5~mm, and the area of each of the 9 radial slits is 51~mm\textsuperscript2, which yields a compression ratio of 12.
The length of the phase plug is 25~mm and the diameter of the final wave guide is 38~mm.
The geometry of this driver is much simplified compared to actual commercially available devices, but the dimensions above are representative for real-life drivers~\cite{Oclee-Brown2012}.
The air properties of table~\ref{AirProp} are used also here.
With the hybrid strategy, the Navier--Stokes equations~\eqref{linNS} are used in the compression chamber and the phase plug, and these equations are coupled to Helmholtz equation \eqref{cavitysystem_eq} in the waveguide.
The sound-hard walls are modeled with isothermal and no-slip boundary conditions in the Navier--Stokes region and with a homogeneous Neumann condition in the Helmholtz region.
Finite-element approximations of these equations are provided in the Acoustics Module (``Thermoviscous Acoustics, Frequency Domain'') of Comsol Multiphysics, which we use for the numerical experiments.
The use of a hybrid strategy is strongly recommended by the software provider to reduce the computationally cost, compared to using the Navier--Stokes equations everywhere~\cite{Co17AM}.
For the alternative strategy derived in this article, Helmholtz equation \eqref{cavitysystem_eq} is used throughout the whole domain, and with boundary conditions \eqref{cavitysystem_w}--\eqref{cavitysystem_wbc} on the walls of the compression chamber and phase plug.
We implement these equations also in Comsol Multiphysics, using the software's so-called Weak Form PDE Interface, in which variational forms like expression~\eqref{varprob} directly can be specified~\cite[Ch.~16]{Co17Ref}.
A homogeneous Neumann bundary condition is used in the waveguide, as in the hybrid strategy.
A lossless model, with the homogeneous Neumann boundary condition on all walls, is used for comparison.
The membrane is regarded as a rigid piston oscillating with a fixed amplitude.
There are nine slits in the phase plug, and thus, due to its symmetry, it suffices to consider a $20^\circ$ segment with appropriate symmetry boundary conditions.
In all three models, the waveguide is terminated at a planar cross section $\Gamma_\text{out}$ supplied with boundary condition \eqref{cavitysystem_io} (with $g\equiv 0$) to model an infinite waveguide.
This condition, equivalent to imposing the acoustic impedance condition $Z_0 = \rho_0 c$, absorbs plane waves, and the acoustic power exiting the waveguide may then simply be computed as
\begin{equation}
P_\text{o} = \frac12\Re\int_{\Gamma_\text{out}} \bs n\cdot\bar{\bs u} p = \frac{1}{2\rho_0 c} \int_{\Gamma_\text{out}} |p|^2.
\label{eq:driver_Po}
\end{equation}
Note that non-planar modes does not propagate for the frequencies considered.
At $10$~kHz, at the upper end of the frequency interval, the first non-planar mode decays with a factor of about $10^{-4}$ over the length of the waveguide, and therefore contributes a smaller error than the discretization.
\begin{figure}
\centering
\includegraphics[scale = 0.5]{driver}
\caption{
Left and middle: The generic compression driver geometry used to compare the hybrid Navier--Stokes/Helmholtz strategy with our boundary-approximation approach.
The circular membrane and the thin compression chamber is to the left, the phase plug in the middle consists of nine radial slits that expand into a circular wave guide.
Right: The rotationally symmetric compression driver geometry used to tune the mesh.
}\label{driver}
\centering
\includegraphics[scale = 1]{driver_mesh}
\caption{Left: the mesh used in the hybrid strategy solution. Middle: close up of a cut through a part of the chamber and phase plug, where the boundary layer elements can be seen. Right: the same close up, but showing the mesh used in the boundary approximation solution. Note that this mesh, apart from the absence of boundary layer elements, is very similar to the mesh used for the hybrid strategy solution.
}\label{driver_mesh}
\end{figure}
\begin{figure}\centering
\includegraphics{Pout}
\hfill
\caption{Output power $P_\text{o}$ for the compression driver in figure~\ref{driver}, computed with the Navier--Stokes/Helmholtz hybrid strategy, our boundary approximations, and when ignoring visco--thermal losses.
The 0-dB reference power is the output power at $f=625$~Hz.
}\label{Pout}
\bigskip
\captionof{table}{
Details from the Comsol log-files for the simulations of the compression driver in figure~\ref{driver}, on meshes tuned to give a relative accuracy of around $0.01$.
\\$^\ast$The solution time is the total wall clock time excluding I/O.
Two computer cores were used for the boundary-approximation solution, while all 24 available cores were used for the hybrid solution.
}\label{table:driver}
\bigskip
\begin{tabular}{l|rrr}
& Degrees of freedom & Memory used & Solution time$^\ast$ \\
& & & per frequency \\ \hline
Hybrid strategy & 1 033 276 & 101 613 MB & 2 111 s \\
Boundary approximation & 63 725 & 1 242 MB & 12 s \\
Quotient Hybrid/Boundary & 16.21 & 81.8 & 180 \\
\end{tabular}
\end{figure}
A good resolution of the rapidly varying velocity and temperature fields in the very thin boundary layers of the Navier--Stokes model requires a very fine mesh near the solid walls.
(The pressure, however, is essentially constant across the boundary layer.)
Anisotropic boundary-layer elements that are elongated in the tangential direction of the wall reduces the number of elements, compared to a uniform refinement, but the presence of the boundary layer will nevertheless lead to a large number of degrees of freedom.
We use quadratic elements for the acoustic pressure in all models, and cubic elements for the velocity components and the temperature in the solution of Navier--Stokes equations, relying on Comsol's choice of Taylor--Hood-like element orders.
To tune the mesh and estimate the accuracy of the solutions, we considered first a rotationally symmetric compression driver design in which the slits are annular rather than radial; see the right part of figure~\ref{driver}.
In this reference design there are three slits, placed according to Smith's guidelines~\cite{Smith1953}, as further described in Ref.~\cite{Oclee-Brown2012}, and the compression ratio and other physical dimensions are identical to the design with the radial slits.
This geometry possesses axial symmetry and allows thus a 2D solution with an extremely fine mesh to be computed and used as a reference.
The chosen mesh settings for the 3D model are as follows.
The maximum element length in interior of the model is $3.43 $~mm, which corresponds to one tenth of the wavelength at $10$~kHz.
At the boundaries of the compression chamber and phase plug, there are six layers of elongated boundary elements, each with thickness $36.6$~\textmu m and maximum length $1.27 $~mm tangential to the boundary.
The boundary layer element thickness is chosen to resolve the viscous and thermal boundary layer effects, and should be compared to the viscous and thermal acoustic boundary-layer thicknesses $\delta_V$ and $\delta_T$.
These are $\delta_V = 21.9 $~\textmu m ($87.6$~\textmu m) and $\delta_T = 26.0$~\textmu m $(104$~\textmu m) at $10$~kHz ($625$~Hz), which is the upper (lower) end of the frequency interval considered.
Recall that the velocity and temperature, which are changing rapidly close to the boundaries, are discretized by polynomials of degree three.
The maximum relative difference in output power $P_\text{o}$ between the fine mesh 2D model and a $20^\circ$-segment 3D model on the chosen mesh is $\approx 0.01$.
The same 3D mesh settings are then used for the design to the left in figure~\ref{driver}, resulting in the mesh in figure~\ref{driver_mesh}, and it seems reasonable to expect an accuracy in the same order $\approx 0.01$ as for the axially-symmetric case.
A similar procedure is used to tune a coarser mesh, without elongated boundary layer elements, to use in our boundary approximation model.
Also in this case the mesh is chosen so that the estimated accuracy for the output power is $\approx 0.01$, to obtain a fair comparison between the hybrid and boundary approximation strategies.
In this model, the maximum element length is $0.40$~mm in the compression chamber, $1.27 $~mm at the boundaries of the phase plug, and $3.43$~mm in the interior of the phase plug and waveguide.
The results can be found in figure~\ref{Pout}.
First of all, we note that the output power is radically different when boundary effects are ignored, especially around the resonance at about 6.5~kHz.
Moreover, the agreement between the hybrid strategy and our proposed boundary-approximation model is very good: the maximum relative difference in output power $P_\text{o}$ is less than $0.01$, and thus in the same order as the estimated relative numerical accuracy.
Both the CPU time and memory consumption are radically lower for the boundary approximation model; the details can be found in table~\ref{table:driver}.
\section{Discussion}
Our proposed approach to account for visco--thermal boundary losses takes the form of the Wentzell condition~\eqref{Hhsystem1_bc} for the acoustic pressure.
The acoustical effects on both the amplitude and the phase of the thin viscous and thermal boundary layers, which are in the order of 20--400~\textmu{m} in the audio range, are taken into account by this boundary condition.
Thus, these layers do not need to be explicitly resolved by the mesh in numerical simulations, and the computational cost becomes essentially the same as for the lossless case.
Therefore, we believe that this approach is a very attractive alternative to full or hybrid solutions to the linearized Navier--Stokes equations for many---maybe most---cases when visco--thermal boundary losses need to be taken into account.
The boundary-layer approximations break down for extremely thin capillaries or slits.
However, as can be seen from figure~\ref{k_compare}, the break down starts at long wavelengths, which means that errors in our model will only be noticeable for devices with extreme geometries, such as very long submillimeter capillaries.
We also saw in the numerical experiments of \S~\ref{sec:driver} that in the frequency range where the boundary-layer approximation becomes questionable, the losses were anyway very small.
The boundary-layer analysis presented here assumes that the radii of the principal curvatures of the wall is much larger than the boundary-layer thickness.
In particular, the effects of corners and edges along solid walls are not accounted for.
Supported by the numerical experiments in \S~\ref{sec:driver}, it seems reasonable to assume that the effects of such geometric features may often only be of second order at low amplitudes.
The situation may be quite different, however, if the geometry includes a very large number of such features, or if significant wall roughness or microscale patterns in the order of the boundary-layer thickness are present.
The proposed boundary condition will likely not be a good model for such situations.
Out of scope for the present investigation is also nonlinear effects such as flow separation at edges, which will become significant at high amplitudes.
\clearpage
|
\section{Introduction}
Double perovskites (DPs) A$_2$BB'O$_6$ display interesting electronic and magnetic properties, strongly depending on the degree of the B-site ordering, which as well is determined both by size and valence mismatch of the involved B-site cations \cite{RevDP,Ohtomo,Anderson}. In addition, epitaxial stabilization in thin films could lead to an improvement of B-site ordering as observed for example in the La$_2$FeCrO$_6$ system \cite{Chakraverty}. The most prominent ordering type of the B-site cations is the rock salt like structure with alternating B and B' planes along the pseudo cubic [111]$_\text{pc}$ direction \cite{RevDP}, shown in Figure \ref{Fig_Crystal} for the title compound of this study. Besides the ferrimagnetic halfmetal Sr$_2$FeMoO$_6$ (Ref. \cite{Sr2FeMoO6}) there are other highly insulating ferromagnetic ordered DPs, like multiferroic La$_2$CoMnO$_6$ (Ref. \cite{La2CoMnO6}) and Ba$_2$CuOsO$_6$ (Ref. \cite{Ba2CuOsO6}), a magnetic insulator recently synthesized und
er high-pressure.
Recently, correlated oxides with strong spin orbit (SO) coupling attracted great attention. In particular, the SO coupling in iridates with Ir$^{4+}$ ($5d^5$ configuration) ions in octahedral coordination results in four occupied $J_\text{eff} = 3/2$ and two half-filled $J_\text{eff} = 1/2$ states~\cite{Kim}. As was first verified for the layered perovskite Sr$_2$IrO$_4$ (Ref. \cite{KimScience}), already a moderate Coulomb repulsion is sufficient to induce a SO Mott insulating state, with magnetic $J_\text{eff} = 1/2$ moments. SO coupling is also important for semimetalic iridates \cite{PRL114}. As realized by Jackeli and Khalliulin \cite{kitaev} novel magnetic exchange interactions, in particular the honeycomb Kitaev exchange~\cite{Kitaev-model}, can arise from $J_\text{eff} = 1/2$ magnetic moments. This initiated strong interest in two- and three-dimensional honeycomb iridates~ \cite{arXiv}.
\begin{figure}[t]
\includegraphics[width=0.95\linewidth]{Fig1.pdf}
\caption{\label{Fig_Crystal}Crystal structure of Sr$_2$CoIrO$_6$ visualized with VESTA \cite{VESTA} based on Ref. \cite{Mikhailova}. The light blue cell indicates the pseudo-cubic (pc) unit with rock-salt type ordering at the B-sites. Note the superstructure with alternating Ir and Co layers along the [111]$_\text{pc}$ direction.}
\end{figure}
Iridate DPs offer another novel playground to investigate the electronic and magnetic properties arising from the competition of SO coupling, electronic correlations and structural distortions. This leads, for example, in Sr$_2$CeIrO$_6$ to an insulating state with weak antiferromagnetic (AFM) orbital order~ \cite{Sr2CeIrO6}. Also both La$_2$MgIrO$_6$ and La$_2$ZnIrO$_6$ host SO Mott insulating magnetic states~ \cite{La2XIrO6}. Sr$_2$YIrO$_6$ represents a class of DPs with Ir$^{5+}$ configuration and evidence of novel magnetism has been reported~\cite{PRL112}. However, subsequent work \cite{PRB95-Sr2YIrO6}, also on the isoelectronic Ba$_2$YIrO$_6$ \cite{PRB96-Ba2YIrO6} related these observations to diluted paramagnetic impurities.
Strontium-iridate derived DPs may also be interesting from the perspective of tuning the properties of the semimetallic three-dimensional perovskite SrIrO$_3$ to a topological state. A tight-binding model for this material in Ref.~\cite{Carter} revealed a symmetry protected nodal line made of $J_\text{eff}=1/2$ bands below the Fermi level. The same work proposed that the line node can be lifted enabling strong topological insulator behavior if sublayer reflection symmetry could be broken and SrIrO$_3$-based B-site ordered DPs such as Sr$_2$CoIrO$_6$ (SCIO) were suggested as possible route for realization.
There were very few reports~\cite{Mikhailova,Narayanan} on the bulk SCIO, prepared as polycrystalline samples by solid-state reaction from mixtures of oxides and carbonates (SrCO3). SCIO crystallizes in the monoclinic strcuture with the space group I2/m and pseudocubic lattice constants $a=0.3909$~nm, $b=0.3925$~nm and $c=0.3921$~nm. The Co/Ir ordered structure is stabilzed due to a large difference in cation radii between Co$^{2+}$ (0.0745 nm) and Ir$^{4+}$ (0.0625 nm) ions. The antiferromagnetic (AFM) ordering with $T_N \sim 70$~K (AFM Curie-Weiss temperature $\Theta=-138$~K) and very small spontaneous magnetization, $M\approx 0.005 \mu_B$/f.u. was detected. The electron transport was found to by an insulating-like obeying a Mott variable-range hopping scenario with a gap $\sim 0.05$~eV.
Here we report for the first time the growth, structure and electronic properties of SCIO thin films, prepared by metalorganic aerosol deposition (MAD) technique, which employs an oxygen-rich growth atmosphere and enables to prepare high-quality perovskite thin films\cite{SchneiderPRL,Moshnyaga,Jungbauer}. The Co$^{3+}$ high spin state and, consequently, the Ir$^{5+}$ configuration was obtained, that is different as compared to the SrIrO$_3$ perovskite. Another motivating question was whether an incomplete B-site order with about 13.2\,\% site mixing in previously reported bulk SCIO~\cite{Narayanan} could be improved by the in-plane epitaxy strain in thin films and how the electronic and magnetic properties will be influenced by epitaxial strain. Coherently strained SCIO thin films were epitaxially grown on various perovskite substrates and demonstrate a remarkable strain control of magnetotransport. By changing from tensile to compressive strain a sign reversal of the magnetoresistance due to
a change of the magnetic easy axis from in- to out-of-plane configuration was observed.
\section{Experimental}
\label{sec:experimental}
Thin films of SCIO were grown by MAD on (111) oriented SrTiO$_3$ (STO), as well as on various pseudo-cubic (pc) (001) oriented substrates.
The lattice mismatch of the used substrates ranges from -1.51\,\% (GdScO$_3$) over -0.99\,\% (DyScO$_3$), 0.05\,\% (STO) and 0.96\,\% (LSAT) to 1.09\,\% (NdGaO$_3$), thus, covering a broad range in both compressive and tensile direction.
To protect the SCIO film an STO capping layer was grown also by MAD directly after the SCIO film.
Phase purity, crystal structure and strain states were determined at room temperature by x-ray diffraction, using a PHILIPS X'PERT diffractometer, operated with Cu-K$_{\alpha1,2}$ radiation. The thickness of the films was determined from XRR measurements, performed by means of a BRUKER D8 ADVANCE diffractometer, and further simulated with the \textsc{ReMagX} \cite{ReMagX} program.
The far-field Raman spectra were measured at room temperature by using a HORIBA Jobin Yvon LabRAM HR Evolution confocal Raman spectrometer in
the back-scattering geometry.
A neodymium-doped yttrium aluminum garnet laser with a wavelength of $\lambda = 532\,\text{nm}$ (second
harmonic generation; Laser Quantum torus 532 with 100\,mW, limited to 1\,mW to avoid heating effects) is appliedfor excitation; the size of the laser spot was $\sim1\,\relax{\mbox{\usefont{U}{psy}{m}{n}{m}}}\text{m}$.
For the polarization-dependent Raman measurements, the incident linear polarization of the laser can be tuned between P- and S-polarization by rotating a $\uplambda$/2 wave plate. The scattered light polarization is determined by an analyzer with two switchable configurations (P-polarization and S-polarization) in front of the detector.
The high-resolution scanning transmission electron microscopy (HR-STEM) studies were carried out on in a cross-section geometry using a FEI Titan 80-300 environmental transmission electron microscope operated at 300\,kV
and equipped with a Gatan Quantum ER image filter. The cross-section lamella sample was prepared using a focus ion-beam (FIB) machine (FEI Nova NanoLab 600 DualBeam instrument).
The polarization-dependent X-ray absorption spectroscopy (XAS) measurements were performed at the Dragon beamline of the National Synchrotron Radiation Research Center (NSRRC) in Taiwan. The spectra were recorded at 300 K using the total electron yield method (TEY) from SCIO films which were capped with 2 nm STO. The photon energy resolution at the Co $L_{2,3}$ edges was set at 0.3 eV and the degree of linear polarization was 99 \%. The samples were mounted on a holder which was tilted with respect to the incoming beam, such that the Poynting vector of the light makes an angle of 70$^{\circ}$ with respect to the [001]$_\text{pc}$ surface normal. By rotating the sample around this Poynting vector, the polarization of the electric field can be varied continuously from E $\parallel$ 20$^{\circ}$ off the [001]$_{pc}$ surface normal, i.e. E $\parallel$ c (20$^{\circ})$, to E $\perp$ the [001]$_\text{pc}$ surface normal, i.e. E $\parallel$ ab. A CoO single crystal were measured simultaneously in a separate chamber to serve as en
ergy reference for Co $L_{2,3}$ edge.
The temperature and field dependences of electrical resistance were measured by means of a Physical Property Measurement System (PPMS) using a four-probe van der Pauw geometry within external measurement setup due to the high resistance, $\sim$ M$\Omega$, of films at low temperatures.
The STO capped samples were micro-structured for the resistance measurement by an \textit{in-situ} optical lithography without any contact to air between the subsequent steps. For this purpose, a special four step method was developed, described in detail in the supplementary material (SM) \cite{SM}.
The temperature and magnetic field dependent magnetization were measured in a 250\,nm thick SCIO film using a Magnetic Property Measurement System (MPMS) equipped with a 7\,T magnet in a stabilized DC mode. For each measuring point the raw background signal of the substrate was carefully subtracted from the measured raw signal of the sample and the resulting signal was analyzed in the standard way of the MPMS.
\section{Results and Discussion}
\subsection{Structural investigation}
XRD measurements on (111)$_\text{pc}$ oriented thin films on (111)$_\text{c}$ STO (see Figure \ref{Fig_RSM} (a)) indicate an out-of-plane epitaxy. The presence of B-site ordering in films was evidenced by observation of (1/2 1/2 1/2) superstructure reflections, marked by red arrows in Figure \ref{Fig_RSM} (a). The extracted pseudo cubic out-of-plane lattice parameter, $d(111)_\text{pc} = 2.271(1)\,\mathring{\text{A}}$, for all studied SCIO/STO films with $d<250\,\text{nm}$ is slightly larger than that measured for bulk SCIO, $d(111)_\text{pc} = 2.256\,\mathring{\text{A}}$~\cite{Mikhailova}. The reason is the in-plane compressive strain $\sim+0.05\,\%$ due to the lattice mismatch between SCIO and STO. Reciprocal space mapping around the (112)-STO peak (see Figure \ref{Fig_RSM} (c)) confirms the fully strained state of the film. Small angle XRR measurements in Figure \ref{Fig_RSM} (b) indicate a large scale homogeneity of all SCIO films, $d(\text{SCIO}) = 2-250\,\text{nm}$, as
well as for the STO capping layer, $d(\text{STO}) = 0 - 20\,\text{nm}$. The XRR signal can be well simulated with \textsc{ReMagX} \cite{ReMagX} yielding extracted interface roughness of less then 2\,nm.
\begin{figure}[t]
\includegraphics[width=0.95\linewidth]{Fig2.pdf}
\caption{\label{Fig_RSM}(Color online) (a) $\theta-2\theta$ XRD scan: Green stars indicating peaks from the (111)-STO substrate, the red cross shows the (222)-SCIO reflection and the arrows mark peaks due to ordered superstructure. (b) XRR measurement of a 84\,nm thick SCIO film with 21\,nm STO capping including simulation with \textsc{ReMagX} \cite{ReMagX}. (c) Reciprocal space map around the (112) peak of STO indicates fully strained thin film (arrow marks SCIO signal). The double peak feature of STO substrate is owed by the Cu-K$_{\alpha 1}$/K$_{\alpha 2}$ doublet.}
\end{figure}
HR-STEM images and corresponding FFTs in Figure \ref{Fig_STEM} confirm a high structural quality and the B-site ordering of SCIO films grown on (111) and (001) STO substrates. One can clearly see the pseudo-cubic symmetry with extra spots (white arrowheads) along the [111]$_\text{pc}$ direction due to the established B-site ordered superstructure. The SCIO/STO interfaces in look coherent in agreement with results from the reciprocal space map, thus, evidencing a fully strained state of the film. Moreover, no dislocations or other defects are observed. Also the B-site ordering is clearly seen in Figure \ref{Fig_STEM} (b) since the atomic number of Ir is much larger than the atomic number of Co. The degree of Co/Ir ordering is estimated using HR-STEM image simulation to be more than 65\% (see SM Figure S3) \cite{SM}.
\begin{figure}[t]
\includegraphics[width=0.95\linewidth]{Fig3new.pdf}
\caption{\label{Fig_STEM} HAADF HR-STEM images of the interface between SCIO grown on (111) (a) and (001) (b) STO substrate. Both images evidence a fully strained thin film with no defects at the interface. Fourier transformations of the SCIO film (insets) proove the (pseudo) cubic structure with ordered superstructure, indicated by additional $(\frac{1}{2}\frac{1}{2}\frac{1}{2}$)$_\text{pc}$ spots .}
\end{figure}
\begin{figure}[t]
\includegraphics[width=\linewidth]{Fig4a.pdf}\newline
\includegraphics[width=0.95\linewidth]{Fig4b.pdf}
\caption{\label{Fig_RSMs}(Color online) \textit{Top panel}: Strain dependence of the out-of-plane lattice parameter, red line indicates linear behaviour. \textit{Middle panel}: Reciprocal space maps around the (pseudo) cubic (013)$_\text{pc}$ peak of each used substrate indicating a fully strained thin film in each case. The double peak feature for substrate (and in some cases also for the thin film) reflections is owed by the Cu-K$_{\alpha 1}$/K$_{\alpha 2}$ doublet. The black arrows always point towards the origin. The dashed red lines indicate the adaption of the in-plane lattice parameter due to a constant in-plane scattering vector component. \textit{Bottom panel}: HAADF STEM images of the SCIO--substrate interface for various substrates.}
\end{figure}
To realize a biaxial tensile strain along the [100]$_\text{pc}$ and the [010]$_\text{pc}$ direction, we used substrates of GdScO$_3$ (GSO, $a_\text{pc} = 3.967\,\mathring{\text{A}}$) and DyScO$_3$ (DSO, $a_\text{pc} = 3.946\,\mathring{\text{A}}$) in (001)$_\text{pc}$ orientation. To exert biaxial compressive strain the (001)$_\text{pc}$-oriented SrTiO$_3$ (STO, $a_\text{c} = 3.905\,\mathring{\text{A}}$), LSAT ($a_\text{c} = 3.870\,\mathring{\text{A}}$) and NdGaO$_3$ (NGO, $a_\text{pc} = 3.865\,\mathring{\text{A}}$) were used.
In the top panel of Figure \ref{Fig_RSMs} the evaluated linear relation between the out-of-plane film lattice parameter $d_\text{out-of-plane}$ and the pseudo-cubic substrate lattice parameter $a^\text{substrate}_\text{pc}$ is shown for STO(20 nm)/SCIO(20 nm) films. Within the linear elasticity theory this behavior indicates a fully strained state of films with the poisson's ratio of $\nu = 0.407(8)$.
Reciprocal space mapping around each (pseudo) cubic (013)$_\text{pc}$ substrate peak (see Figure \ref{Fig_RSMs} (middle panel)) prove the fully strained state of each film grown on the used substrates. In addition, HAADF STEM images also verify the strained film/substrate (001)$_\text{pc}$ interfaces (see Figure \ref{Fig_RSMs} (bottom panel)), preserved up to thickness $\sim50\,\text{nm}$ of SCIO film.
In the case of STO as substrate material the B-site ordering was investigated by a $\theta-2\theta$ XRD scan in tilted geometry with $\vec{Q}$ parallel to <111>$_\text{c}$ of STO and therefore, regarding to the small lattice mismatch between SCIO and STO, also nearly parallel to <111>$_\text{pc}$ of SCIO.
The visible superstructure peaks in the collected XRD pattern (see Figure \ref{Fig_Bsite} (a)) could be well distinguished from the STO background, comparable to the results of the (111)$_\text{pc}$ oriented thin films (see Figure \ref{Fig_RSM} (a)), and indicating well developed B-site ordering.
A larger lattice mismatch and therefore bigger difference between the <111>$_\text{pc}$ directions of substrate materials compared to strained SCIO, in combination with peaks appearing already from the bare substrate material at the crucial $2\theta$ positions, is a key reason, why this access is denied in case of the other used substrate materials.
\begin{figure}[t]
\includegraphics[width=0.925\linewidth]{Fig5.pdf}
\caption{\label{Fig_Bsite}(Color online) (a) $\theta-2\theta$ XRD scan in tilted geometry with $\vec{Q}$ parallel to <111>$_\text{c}$ of STO: the arrows mark peaks due to the ordered superstructure. Polarized far-field Raman spectra of SCIO films on (b) GSO, (c) DSO, (d) LSAT and (e) NGO at room temperature. The spectra are taken in the parallel PP (P'P') and the crossed scattering PS (P'S') configurations (with an additional in-plane rotation of $\Phi = 45\,^\circ$ around the [001]$_\text{pc}$ direction). (f) Strain dependence of Raman shift of the breathing mode, line indicates fully strained state with linear dependence.}
\end{figure}
To clarify the B-site ordering state in such SCIO films we performed the polarization-dependent Raman spectroscopy (Figure \ref{Fig_Bsite} (b)-(e)) and compare the data with other ordered double perovskites of the same point group, e.g. La$_2$CoMnO$_6$ (LCMO). B-site-ordered LCMO obeys a monoclinic $P12_1/n1$ structure, for which theoretical lattice dynamical calculations \cite{Iliev} predict $A_g$ stretching mode of the (Co/Mn)O$_6$ octahedra (breathing mode) at $\sim 697\,\text{cm}^{-1}$. In contrast, the disordered LCMO obeys an orthorhombic $Pbnm$ structure and possesses the $B_{1g}$ breathing mode. Similarly, we expect a (Co/Ir)O$_6$ breathing mode with $A_g$ symmetry at a Raman shift, $\text{RS}\sim 697\,\text{cm}^{-1}$, for the B-site-ordered SCIO. In this case the $A_g$ mode should be present in Raman spectra, measured in parallel PP- and P'P'-configuration, and it is forbidden in the crossed scattering PS- and P'S'-configuration\cite{Iliev,Truong,Meyer} (the prime i
ndicates measurements with an in-plane rotation of the sample by 45$\,^\circ$ around the [001]$_\text{pc}$ direction and provide an additional tool to test the epitaxy). For the disordered orthorhombic structure, the selection rules are opposite\cite{Iliev,Truong,Meyer}. As one can see in Figure \ref{Fig_Bsite} (b)-(e), all films in our strain series exhibit a strong breathing mode at $\sim 697\,\text{cm}^{-1}$ in the parallel PP- and P'P'-scattering configurations and only a weak (or none) intensity in the crossed PS- and P'S'-configuration. Thus, the $A_g$ symmetry and the B-site ordering can be concluded for all strained films grown on GSO, DSO, LSAT and NGO. Furthermore, the Raman shift of the breathing mode (see Figure \ref{Fig_Bsite} (f)) was found to depend linearly on the strain, $\text{RS}\sim-a_\text{pc}^\text{substrate}$. This is due to the change of the unit cell volume, $V_\text{pc uc}$, and of the phonon energy with strain. Note, that a smaller $V_\text{pc uc}$
leads to a closer packing and, respectively, more energy is necessary. Kumar and Kaur\cite{Kumar}, observed a similar behavior of the breathing mode in a strain relaxation series of the double perovskite La$_2$NiMnO$_6$/LaAlO$_3$(001).
For selected films on GSO, STO and NGO substrates the B-site ordering was also studied on the micro scale by HAADF STEM measurements along the [110]$_\text{pc}$ direction (see supplementary material) \cite{SM}. These measurements show that independent on strain there are regions with a high degree of B-site ordering, which alternate with regions with lower or no B-site ordering at all. Concluding, the epitaxial strain perpendicular to the [001]$_\text{pc}$ direction could neither improve nor weaken the B-site ordering in SCIO.
\subsection{Magnetic properties}
Separation of magnetic properties of thin films from the substrate magnetic contribution is generally difficult. A comparison with data on bulk SCIO~\cite{Narayanan} indicates, that a film with minimal thickness of $\sim$ 200\,nm is required to get magnetic moment corresponding to $\sim$ 5\,\% of the moment of pure STO substrate at room temperature. Because fully strained (001)$_\text{pc}$ oriented SCIO films are stable for d$\leq$50\,nm, we focus here solely on thick ($d\sim250\,\text{nm}$) strained SCIO/STO(111)$_c$ films, cf. Figure 1. However, we identified a Sr$_3$Co$_2$O$_6$ (SCO) second phase inclusion on the level 0.8\,volume-\%. For a quantitative determination of the SCO volume fraction, we have used the field dependent measurements of magnetization at lowest temperature (for more details, see supplemental material \cite{SM}). The foreign phase moments are saturated in a field of 7~T that has been applied in the measurement shown in Figure \ref{Fig_MPMS_MvsT}. Furth
ermore, for our analysis of the difference between zero-field cooled (ZFC) and field cooled (FC) magnetization, it cancels out.
\begin{figure}[t]
\includegraphics[width=0.92\linewidth]{Fig6.pdf}
\caption{\label{Fig_MPMS_MvsT}(Color online) Temperature dependence of FC and ZFC magnetization of $\sim 250\,\text{nm}$ thick (111)$_\text{pc}$ oriented SCIO (after subtraction of the substrate background contribution) in comparison to literature data of bulk polycrystalline SCIO from reference~\cite{Narayanan} (scaled to the sample volume). \textit{Inset}: Temperature dependence of the difference between the FC and ZFC data (dotted line intended to guide the eyes).}
\end{figure}
Previous investigation on polycrystalline SCIO revealed long-range magnetic order at 70~K, with ferromagnetic and antiferromagnetic components of the ordered magnetic moments, yielding to a pronounced difference in the FC and ZFC magnetization~\cite{Narayanan}. Figure \ref{Fig_MPMS_MvsT} compares these literature data on polycrystalline SCIO (from Ref. \cite{Narayanan}, indicated by triangles) with the FC and ZFC magnetization measurements of our thick (111)$_\text{c}$ oriented SCIO film below 150~K. Compared to the polycrystal data, much weaker splitting between FC and ZFC is found and the maximum at 70~K in ZFC mode is absent. Taking the onset of the difference $\Delta M$ between FC and ZFC (see inset of Figure \ref{Fig_MPMS_MvsT}) as measure of magnetic order, an ordering temperature of 43~K would be estimated for the film, which is significantly reduced compared to the polycrystal. We speculate that epitaxial strain and/or better B-site order is responsible for the differ
ence in the magnetic susceptibility behavior.
In strained (111)$_\text{pc}$ oriented thin films of the double perowskite Sr$_2$FeMoO$_6$ Hauser \textit{et al.} found a similar decrease of the ordering temperature \cite{Hauser}.
Unfortunately, the thinner films of the (001)$_\text{pc}$ strain series could not be investigated by respective magnetization measurements and the evolution of the susceptibility behavior with strain remains therefore unknown.
\subsection{Spectroscopic results}
As mentioned in the previous section the magnetic signal for the samples of the (001)$_\text{pc}$ strain series are too weak to be detected in a conventional SQUID magnetometer. In order to investigate the electronic structure and its antiferromagnetic property of the SCIO thin films, we resort to polarization-dependent XAS. Optical measurements were done on the SCIO films capped by 2 nm thick STO~\cite{Esser}, which due to its high integrity and insulating properties~\cite{Belenchuk} protects SCIO films from degradation but still allows electrons to escape. We note that the polarization-dependent XAS is one of few techniques that can determine the magnetic axis of an antiferromagnetic ordered state in thin films. \cite{Haverkort2004,Csiszar2005,NatComms7} Figure \ref{Fig_XAS} (a) and (b) show the experimental polarization-dependent Co $L_{2,3}$ XAS spectra of the most tensile strained SCIO thin films on GSO and the most compressive strained SCIO films on NGO, respectively, in this study, taken at 300 K. The spectra of the SCIO thin films on DSO, STO, and LSAT substrates are shown in Figure S9 of the SM \cite{SM}. The spectra are dominated by the Co $2p$ core-hole spin-orbit coupling which splits the spectrum
roughly in two parts, namely the $L_{3}$ ($h\nu \approx$ 776-784 eV) and $L_{2}$ ($h\nu \approx$ 793-797 eV) white lines regions. The line shape strongly depends on the multiplet structure given by the Co 3$d$-3$d$ and 2$p$-3$d$ Coulomb and exchange interactions, as well as by the local crystal fields and the hybridization with the O 2$p$ ligands. Unique to soft XAS is that the dipole selection rules are very sensitive in determining which of the 2$p^{5}$3$d^{n+1}$ final states can be reached and with what intensity, starting from a particular 2$p^{6}$3$d^{n}$ initial state ($n= 6$ for Co$^{3+}$) \cite{PRL92,PRL102,PRB95}. This makes the technique extremely sensitive to the symmetry of the initial state, i.e., the spin, orbital and valence states of the ions.
\begin{figure}
\includegraphics[width=0.95\linewidth]{Fig7a.pdf}\newline
\includegraphics[width=1\linewidth]{Fig7b.pdf}
\caption{\label{Fig_XAS}(Color online) Experimental polarization-dependent Co $L_{2,3}$ XAS spectra of (a) the most tensile strained SCIO films on GSO, and (b) the most compressive strained SCIO films on NGO in this study, taken at 300 K, together with the theoretical spectra calculated for Co$^{3+}$ high spin (HS) state. The inset of (a) and (b) is a zoom-in view at the $L_3$ main peak showing the favor orbital state under different strains. (c) The energy level diagram of the orbital states of Co$^{3+}$ in the SCIO films as a function of strain, obtained from the configuration interaction cluster model including the full atomic multiplet theory, which are consistent with the experimental polarization-dependent XAS results. (d) and (e) are the calculated magnetic anisotropy energy as a function of the spin direction of the SCIO films on GSO and NGO substrates, respectively. Each inset depicts a schematic drawing of the resulting AFM axis.}
\end{figure}
The experimental XAS spectra show that the Co valence state of these SCIO thin films is mainly 3+.
Please be noted that for comparing with the calculated spectra the experimental spectra in Figure \ref{Fig_XAS} (a) and (b) have been subtracted by 10\% and 8\% of Co$^{2+}$, respectively. The experimental isotropic XAS spectra before subtracting Co$^{2+}$ contributions can be found in Figure S8 of the SM \cite{SM}. Further, the spectral features indicate a high spin (HS) state of the Co$^{3+}$ ions \cite{PRL92,PRL102,PRB95} in the SCIO thin films which is independent of substrates underneath, i.e., a strain independent HS state. On the other hand, the orbital state in these SCIO thin films is quite different as indicated by the opposite sign of the polarization-dependent difference of spectra.\cite{Haverkort2004,Csiszar2005} For example, the intensity of the $L_3$ main peak is always larger for E $\parallel$ c (20$^{\circ})$ than for E $\parallel$ ab in the SCIO/GSO thin film, whereas it is always smaller in the SCIO/NGO thin film. Since these spectra were taken at 300 K whi
ch is much higher than the magnetic ordering temperature of 43\,K, the polarization contrast is caused solely by crystal field effects. Using the E vector of light parallel to the ab plane (black lines), we can reach the unoccupied Co 3$d$ orbital states with xy/x$^2$-y$^2$ characters. With the E vector of light parallel c axis (red lines), we detect the unoccupied Co 3$d$ orbital states with yz/zx/3z$^2$-r$^2$ characters. This indicates that the sign of the crystal field splitting is opposite in the two systems. This inference is further consolidated by the observed peak energy difference of the $L_3$ main line as depicted in the insets of Figure~\ref{Fig_XAS} (a) and (b). The $L_3$ peak position is higher for E $\parallel$ c (20$^{\circ})$ than for E $\parallel$ ab in the SCIO/GSO thin film, while it is lower in the SCIO/NGO thin film. All together, we derive that in the tensile strained SCIO/GSO thin films the orbital states with x/y-character are energetically favorable,
whereas in the compressive strained SCIO/NGO thin films the orbital states with z-character are energetically favorable. This can be understood qualitatively as the applied strain from the substrates underneath induces a tetragonal distortion and causes the corresponding orbital state shift in energy.
For further confirming this orbital energy level diagram and knowing the corresponding magnetic anisotropy, we have simulated the XAS spectra using the well proven configuration interaction cluster model that includes the full atomic multiplet theory \cite{deGroot,Tanaka}. The calculations were performed using the XTLS 8.3 program \cite{Tanaka}. For the calculation details, please see the SM \cite{SM}. As displayed in the bottom of Figure \ref{Fig_XAS} (a) and (b), the calculated spectra based on the HS Co$^{3+}$ model with the energy diagram of the orbital state shown in Fig. 9 (c) can well reproduce the experimental spectra. We can safely conclude that a tensile strain stabilizes a ($3d^{5}\uparrow\uparrow\uparrow\uparrow\uparrow$ + 3$d_{xy}\downarrow$) state and a compressive stain stabilizes a ($3d^{5}\uparrow\uparrow\uparrow\uparrow\uparrow$ + 3$d_{yz/zx}\downarrow$) state in SCIO thin films. We infer that the change of the anisotropy in the crystal field parameters in going from room temperature to low temperatures due to thermal contraction of the substrates is negligible, and that we therefore can use these XAS derived parameters also for the analysis of the low temperature magnetic properties (see the Supplemental Material for details). Accordingly, we can calculate the magnetic anisotropy energy as a function of the spin direction for each case as shown in Figure \ref{Fig_XAS} (d) (tensile, SCIO/GSO) and (e) (compressive, SCIO/NGO).The magnitude of the exchange field was set to 4 meV in accordance with the the magnetic ordering temperature of 43 K. \cite{PRB95} This magnetic anisotropy energy is expressed as $E = K_{0} + K_{1}\sin^{2}(\theta) + K_{2}\sin^{4}(\theta) + K_{3}\sin^{4}(\theta)\sin^{2}(\varphi)\cos^{2}(\varphi)$, where $\theta$ is the angle between the exchange field and the [001]$_{pc}$, and $\varphi$ is the azimuthal angle which is set to 45$^{\circ}$. We find for the SCIO/GSO (tensile) $K_{1} = -1.42$ meV, $K_{2} = 0.14$ meV, and $K_{3} = 0.56$ meV, while for the SCIO/NGO (compressive) we obtain $K_{1} = 0.48$ meV, $K_{2} = 0.08$ meV, and $K_{3} = 0.33$ meV. In other words, for the tensile strained SCIO/GSO thin film, the spin moment favors the in-plane direction with the energy difference of about 1.14 meV between the magnetic easy axis ($\theta = 90^{\circ}$, in the
film plane) and hard axis ($\theta = 0^{\circ}$, perpendicular to the film), whereas for the compressive strained SCIO/NGO thin film, the spin moment favors the out-of-plane direction with the energy difference of about 0.64 meV between the magnetic easy axis ($\theta = 0^{\circ}$, perpendicular to the film) and hard axis ($\theta = 90^{\circ}$, in the film plane). Note that the evaluated strain-induced AFM magnetic anisotropy in SCIO films, i.e. in/out of plane for tensile/com-pressive stress, differs from that observed for ferromagnetic double perovskite films of La$_2$CoMnO$_6$, i.e. in/out of plane for compressive/tensile stress~\cite{Galceran,Lopez-Mir}. The reason is unclear up to know and, likely, is related to the FM exchange interaction between Co$^{2+}$ and Mn$^{4+}$ ions accordiung to the second Goodenough-Kanamori-Anderson rule. This is a piece of useful information for understanding the magnetotransport properties of the SCIO thin films under different strains, see the discussion in the section of Magnetotransport.
\subsection{Zero field transport properties}
\begin{figure}[t]
\includegraphics[width=0.95\linewidth]{Fig8.pdf}
\caption{\label{Fig_RT-111}(Color online) Temperature dependence of the electrical resistivity of SCIO for a sample with STO protection (blue) and without protection (red) in comparision to bulk data (black line) obtained by Narayanan \textit{et al}. \cite{Narayanan}. \textit{Inset}: Thickness dependence of specific resistance at 300\,K (dashed line as guide for eyes).}
\end{figure}
We now turn to the electrical resistivity of SCIO thin films. The temperature dependent $\rho(T)$ measurements for a (111)$_\text{pc}$ oriented $d$(SCIO) $=$ 30\,nm thick SCIO film (with and without STO protective top layer) are shown in Figure \ref{Fig_RT-111}. Similar to the bulk SCIO (Ref. \cite{Narayanan}), thin films also show an insulating behavior. The values of room temperature resistivity for the air protected and unprotected films were $ 3.08(1)\,\text{m}\Omega\,\text{cm}$ and $6.66(1)\,\text{m}\Omega\,\text{cm}$, respectively. The reported bulk value (Ref. \cite{Narayanan}) was about 3 and 1.5 times larger, respectively. Upon cooling to 4~K, the electrical resistance of the unprotected film strongly increases, becoming several orders of magnitude larger than the resistance of the protectedfilm; the $\rho(T)$ of the bulk SCIO lies in between. This indicates the importance of the STO capping layer to protect the (surface) properties of SCIO. As shown in the inset of
Figure \ref{Fig_RT-111} the room temperature resistivity is almost independent on the film thickness for $d>5-8\,\text{nm}$. Unfortunately, it was impossible to probe the thickness dependence at low temperatures.
Fitting the $\rho(T)$ data of the capped SCIO thin film between 250 and 300~K results in a charge gap of $32(5)$~meV \cite{SM}. Below 70~K, $\rho(T)$ indicates variable range hopping behavior \cite{SM}.
The electrical resistivity of the various films, strained along the (001)$_\text{pc}$ direction, is discussed in supplemental material \cite{SM}. The overall insulating behavior does not change, though a systematic enhancement of $\rho$(300 K) with strain is observed \cite{SM}.
\subsection{Magnetotransport within the (001)$_\text{\textbf{pc}}$ strain series}
\begin{figure}[b]
\includegraphics[width=0.95\linewidth]{Fig9.pdf}
\caption{\label{Fig_MR-overview} Isothermal transverse magnetoresistance $MR = (\rho(H)-\rho(0))/\rho(0)$ of (001)$_\text{\textbf{pc}}$ strained SCIO thin films grown on LSAT (compressive strain) and GdScO$_3$ (tensile strain) at various temperatures between 2 and 6 K (fields always applied transverse to the film planes).}
\end{figure}
Considering the information on the magnetic anisotropy, obtained in the ''Spectroscopy'' section, we can now address the magnetotransport properties of the SCIO thin films under different strains.
Next, we focus on the low-temperature isothermal magnetoresistance (MR) of (001)$_\text{\textbf{pc}}$ strained SCIO thin films. As discussed previously in section C, the magnetic easy axis in this series changes from the ''in-plane'' to the ''out-of-plane'' as the strain changes from tensile to compressive. This change of the magnetic anisotropy has a direct influence on the magnetoresistance. As shown in Figure \ref{Fig_MR-overview}, MR below 6~K is {\it positive} for thin films grown under compressive strain on LSAT substrate and {\it negative} for those under tensile strain on GSO substrate. This indicates a direct relation between MR sign and magnetic anisotropy. This is further corroborated by the comparison of the MR at 2~K of all strained SCIO thin films, which, as shown in Figure \ref{Fig_MR-strain-overview}, follow this trend. This observation suggests a strong influence of the anisotropic magnetoresistance (AMR) effect in combination with the reorientation of the AF
M ordered Co sublattice in SCIO in an external transverse magnetic field.
\begin{figure}[t]
\includegraphics[width=0.95\linewidth]{Fig10.pdf}
\caption{\label{Fig_MR-strain-overview} Isothermal transverse magnetoresistance of all investigated (001)$_\text{\textbf{pc}}$ strained SCIO thin films at $T=2\,\text{K}$. Note, that the thin film on DSO substrate broke in a field of $\mu_0H=7\,\text{T}$ due to a high magnetic torque effect from DSO.}
\end{figure}
\begin{figure}[b]
\includegraphics[width=0.5\linewidth]{Fig11.pdf}
\caption{\label{Fig_MR-STO}Field dependence of the MR of SCIO on STO at $T=2\,\text{K}$. The red and green lines indicate the contributions of the WAL and AMR effect to the total signal. For the latter contribution, a smeared spin flip scenario, detailed in supplemental materials \cite{SM}, has been utilized.}
\end{figure}
In addition, structural disorder in combination with large spin orbit coupling will generate a quantum correction to MR due to weak anti-localization (WAL). Based on the Dresselhaus effekt~\cite{Dresselhaus} we used for the description of the WAL term of the MR a Dresselhaus like contribution with isotropic spin orbit scattering, see supplemental material~\cite{SM}. For the AMR contribution the orientation of the moments of Co sublattices in their AF ground state with respect to the field direction is important. For tensile strain at zero field the moments are oriented in the film plane and application of a transverse field leads to a continuous rotation of both sublattice moments out of the plane. This leads to a negative MR. For compressive strain, the moments of the two sublattices initially point parallel and antiparallel to the applied field. If there would be very weak coupling between the two sublattices, only the moments of the antiparallel sublattice would continuous
ly rotate towards the applied field direction with increasing field. In the alternative case, upon increasing magnetic field the moments remain in their orientation until a spin-flip occurs (for a sketch, see supplemental material), followed by continuous rotation towards the field. Since there is no indication for a sharp spin flip, we modeled the MR behavior in a smeared spin flip scenario. For more details, we refer to supplemental material~\cite{SM}.
By combining AMR and WAL a valuable quantitative description of the measured MR is possible for all strained SCIO thin films. This is exemplified in
Figure \ref{Fig_MR-STO} for compressively strained SCIO. Within the smeared spin-flip scenario, a positive AMR contribution to the MR dominates between 10 and 24~T~\cite{SM}. As detailed in supplemental material, the obtained parameter for the WAL and AMR contributions within the smeared spin-flip scenario are more realistic compared to the assumption of very weak coupling between the two Co sublattices. To point out our key arguments: We expected a positive contribution from AMR and a temperature independent spin-orbit scattering field within the WAL contribution. Both is only the case in the smeared spin-flip scenario.
\begin{figure}[t]
\includegraphics[width=0.95\linewidth]{Fig12.pdf}
\caption{\label{Fig_MR-strain}(Color online) Strain dependence of the MR at a magnetic field of $\mu_0H=7\,\text{T}$ at $T=2\,\text{K}$ (cyan circles) in comparison to the energy difference $\Delta e_\text{g} = E_{d_{x^2-y^2}}-E_{d_{z^2}}$ (red squares) from Figure \ref{Fig_XAS} (c).}
\end{figure}
As indicated by Figure \ref{Fig_MR-strain} the strain induced change of the Co d-level splitting, and respective change of moment orientation from in-plane (for tensile strain) to out-of plane (for compressive strain) goes hand in hand with a change of the MR from negative to positive. Due to the counteracting contributions of WAL and AMR effect in the samples with compressive strain, the absolute value of the MR is much smaller than for tensile strained samples, cf. Figure \ref{Fig_MR-strain-overview}.
\section{Conclusion}
For the first time B--site ordered SCIO thin films have been grown on (111)$_\text{c}$ oriented STO substrates and within a strain series also on various (pseudo) cubic (001)$_\text{pc}$ oriented substrates. Our electrical transport measurements of (111)$_\text{c}$ oriented samples with and without an air protection layer out of STO indicated in comparison to literature results from reference \cite{Narayanan} that the SCIO thin films degenerate in direct air contact. This leads us to the development of a four step \textit{in-situ} lithographical process to investigate the temperature and magnetic field dependencies of the electrical transport properties for protected samples.
The complete strain transfer from the substrate to the thin film was checked in all cases of the (001)$_\text{pc}$ strain series by RSMs around the (013)$_\text{pc}$ substrate peak in combination with HAADF STEM images in [100]$_\text{pc}$ direction. The, at least partial, B--site ordering of the SCIO thin films was proven by XRD scans in tilted geometry as well as by polarization dependent Raman spectroscopy. HAADF STEM images in [110]$_\text{pc}$ direction of selected samples reinforce these results.
The magnetic properties were investigated with SQUID magnetometry. The AFM ordering temperature $T_\text{N}\sim 43(10)\, \text{K}$ of the thick SCIO thin films is reduced in comparison to the bulk value ($T_\text{N}\sim 70\, \text{K}$ \cite{Narayanan}).
Using polarization dependent x-ray absorption spectroscopy at the Co $L_{2,3}$ edges, we revealed that the Co $3d$ orbital occupation strongly depends on the strain in the SCIO thin films induced by the substrates. The tensile strained SCIO thin films stabilize the occupation of the minority orbital state with x/y character, whereas the compressive strained SCIO thin films favor the occupation of the minority orbital state with z character. Together with the calculations using the well proven configuration interaction cluster model, we were able to determine the magnetic easy axis change due to the induced strain from the substrates, and the sign and the magnitude of the magnetic anisotropy energy of the antiferromagnetically ordered high-spin Co$^{3+}$ ions. We presented a quantitative model including the anisotropic magnetoresistance and weak anti-localization effects to explain the opposite behavior and its magnitude of the low-temperature magnetoresistance in the tensile
strained from the compressive strained SCIO thin films.
\begin{acknowledgments}
The authors like to thank S. Meir for developing the shadow mask argon ion etching method at Augsburg University and B. Meir for the support with it. We thank R. Pentcheva and P. Seiler for stimulating discussions. The work at Augsburg was supported by the Deutsche Forschungsgemeinschaft (DFG) through SPP 1666 and TRR 80. The research in Dresden was partially supported by the DFG through SFB 1143 and the study in G{\"o}ttingen through SFB 1073 (TP A02).
\end{acknowledgments}
|
\section{Introduction}
Graphene is a two-dimensional (2D) material which enables
the realization of different heterostructures and electronic devices with novel
quantum properties \cite{Geim2013,Perali2013,Perali2014}.
The electronic and transport properties of graphene
in a multilayer stacking can be modified and tuned by changing the number
of layers of graphene in the system. Remarkably, the low energy electronic
band structure of multilayered graphene evolves from Dirac cones in
monolayer graphene with massless Weyl excitations,
to parabolic bands with massive excitations in the case of bilayer graphene,
to more complex band structures and peaked density of states when a
few-layer graphene system is realized with different stacking orders \cite{McCann2006,Min2008,Koshino2009,Zhang2010}.
Few-layer graphene can be fabricated from graphite
by mechanical exfoliation \cite{Ferrari2006,Zhang2005}
and by chemical techniques \cite{Berger2004,Shih2011,Mahanandia2014}
controlling the stacking order. Experimental characterizations of
electronic and transport properties of trilayer graphene have been reported
in Ref. \cite{Craciun2009,Bao2011,Mak2010}.
An electric field perpendicular to the graphene layers has been
shown experimentally to open a band gap in the single-particle electronic
energy spectrum in bilayer \cite{Zhang2009} and
trilayer \cite{Lui2011,Zou2013} graphene, being this effect of
crucial importance in realizing semiconducting behavior for electronics.
Theoretical predictions of a similar band gap opening in the energy
spectrum of few-layer graphene have been reported \cite{Avetisyan2010}.
The electronic band structure of graphene and many other 2D systems can be also tailored by strain of their lattice in different directions, inducing several possible quantum effects.
Strain plays a key role in iron-based superconductors \cite{Poccia2010}, in cuprate
superconductors \cite{Agrestini2003}, and in superconducting diborides \cite{Agrestini2001}.
In cuprate superconductors the displacement of the Cu ions along the transversal
direction induces strain modulations and bending in the copper-oxide planes in the form of lattice stripes leading to amplification effects of the superconducting parameters \cite{Bianconi96a,Bianconi96b}.
Curved 2D systems are a source of different kind of compressive and tensile strains, with
a wide range of strain percentages of the lattice depending on the topology and on the curvature
radius of the surface of the curved systems. Therefore we expect that curvature will be
another possible way to engineer the electronic band structure of 2D systems, as we will
discuss in this work.
Recently the low energy electronic properties of bilayer graphene have been studied in the context of a geometrical approach,
using the Lévy-Leblond equation \cite{LevyLeblond1967} for massive particles \cite{Curvatronics2017} in a curved space.
Although the study of graphene through relativistic approaches is a quite new field of research, a significant amount of literature already exists, in the study of deformed monolayer graphene employing the relativistic 2D Dirac equation in a curved metric \cite{Sitenko2007,Cortijo2007,deJuan2007,Cortijo2007b,Vozmediano2010,Arias2015,Amorim2016}.
This connection between graphene and relativity has been explored also in other contexts, for instance for the evolution of the free electron current density in graphene
with defects \cite{sepheri2017,Capozziello2018}.
On the other hand, the scientific challenge to connect Dirac theory of the spinning electron with gravitation
dates back to the seminal works of Hermann Weyl \cite{Weyl1929}. Interestingly,
Weyl fermions with zero mass and definitive chirality have not been identified
in high-energy physics, but they have been clearly observed in
novel topological materials in condensed matter systems \cite{Wan2011}.
The role of two-dimensionality in generating Weyl states in systems
with parabolic electronic bands has been demonstrated in Ref.\cite{Doria2017},
while the influence of zero helicity states at the interfaces of
layered heterostructures, leading to local magnetization and mass anisotropy,
has been discussed in Ref.\cite{Rodrigues2017}.
The above described example demonstrates that solid state systems with
geometrical constraints and specific topological properties can work as
model systems for physical phenomena with very different energy scales,
as high-energy physics and gravitation in a curved D-dimensional space-time.
The mapping proposed in Ref.\cite{Curvatronics2017}
allowed to study the evolution of the band structure,
both conduction and valence bands and their band-gap, as a function of
geometrical curvature. Positive (spherical-like) curvatures have a local effect of opening a band gap
in the spectrum, while negative (hyperbolic-like) curvatures tend to close
the band gap. The band-gap energy is predicted to be tunable and proportional
to the curvature radius of the deformation. This paves the way to
``Curvatronics'' with bilayer graphene, an interesting possibility for
applications to control in a static way the electronic properties
of layered systems with the curvature.
In this work we extend the analysis of Ref.\cite{Curvatronics2017}, analysing massive and massless fermionic excitations in the more complicated case of few-layer flat graphene systems, while for the case of the curved bilayer we refine the treatment, proposing a set of equations that take into account the exact combination of pseudospin states that defines physical states, i.e. eigenstates of the energy. Similar considerations apply to multi-layer systems and will be studied separately.
The work is organised as follows. In sec. \ref{sec:GalileanFermions} we discuss Galilean fermions in few layers graphene: we begin in \ref{sec:two_layers} recalling the results obtained in \cite{Curvatronics2017} for bilayer graphene, and then in \ref{sec:three_layers} and \ref{sec:four_layers} we show that Galilean fermions are present in the case of three and four layers as well. We obtain the exact spinor solutions for all energies, and study them in the proximity of the \textit{Galilei points}, where the electronic dispersions acquire their extreme values: the results is given by 2D spinors that satisfy the Galilei invariant L\'evy-Leblond equation, and we show where these fermions are localised. Then in sec. \ref{sec:3} we discuss various aspects of the L\'evy-Leblond equation, both in the flat and curved case. We recall its Galilean invariance and discuss its two independent solutions, which correspond to states with different pseudo-spin. It has been discussed in \cite{Curvatronics2017} how solutions of the L\'evy-Leblond equation in 2D can be lifted to solutions of the massless Dirac equation in 4D. We show here how states with a definite pseudospin lift to states with definite chirality, with a perfect match of degrees of freedom: the states corresponding to positive and negative isospin are mapped into the two solutions of the massless 4D Dirac equation, one with left and one with right chirality. Then we discuss in detail the case of the curved bilayer. We present a set of covariant equations that describe the generalization of the flat massive fermions, show how the equations predict that energy eigenstates are given by a well defined mixture of pseudospin states and infer that the scalar curvature of the surface alters the local energy density.
\section{Galilean fermions in few layers graphene\label{sec:GalileanFermions}}
\subsection{Bilayer graphene\label{sec:two_layers}}
In this section we review the results obtained in \cite{Curvatronics2017} for bilayer graphene with AB Bernal stacking, presenting them in a manner that is suitable to generalisation to a higher number of layers.
Bilayer graphene with AB stacking presents one set of metallic bands $E_1^{(\pm)}$ , and another of isolating bands $E_2^{(\pm)}$.
We define a quantity with the dimensions of mass
\begin{equation}
m_0 = \frac{\gamma}{v_F^2} \, ,
\end{equation}
where $\gamma$ is the interlayer hopping parameter, $\gamma \sim 0.4$eV, and $v_F \sim 10^6$ms${}^{-1}$ is the Fermi velocity in graphene.
The energy bands can be described in terms of $m_0$ and of the effective mass $m = \pm \frac{m_0}{2}$ for bilayer graphene, where the $\pm$ factor denotes either positive or negative bands:
\begin{equation} \label{eq:energy_free1}
E_i^{(\pm)} = \pm \left[ (-1)^{i} |m| v_F^2 + \sqrt{ m^2 v_F^4 + |\hbar v_F\kappa|^2} \right] \, , i = 1,2 ,.
\end{equation}
It is useful to define a dimensionless factor $\alpha$ by
\begin{equation} \label{eq:alpha}
\alpha |m| := m_0
\end{equation}
such that, in the bilayer case, $\alpha = 2$.
The $\bf{k} \cdot \bf{p}$ model that describes these bands is defined in terms of a spinor $\lambda$ with 4 components. The Hamiltonian is given by
\begin{equation} \label{eq:H}
H_{\kappa}^{(AB)} = \left( \begin{array}{cccc}
0 & \hbar v_F\bb{\kappa} & 0 & 0 \\
\hbar v_F \kappa & 0 & \gamma & 0 \\
0 & \gamma & 0 & \hbar v_F \bb{\kappa} \\
0 & 0 & \hbar v_F \kappa & 0
\end{array} \right) \, ,
\end{equation}
where $\kappa = \tau k_x + i k_y$ is the wave number of the excitation, with $\tau = \pm 1$ denoting the two inequivalent Fermi points\footnote{Notice that, with respect to \cite{Curvatronics2017}, we have used a different although equivalent basis for the Hamiltonian, exchanging the vectors $1\leftrightarrow 2$ and $3\leftrightarrow 4$. This will result, for instance, in slightly different expressions for eqn. \eqref{eq:sol_bilayer_spinors1}.}. These are the points in the Brillouin zone where the conduction and valence bands touch, and for graphene they are at the level of the Fermi energy. The $\bf{k} \cdot \bf{p}$ model can be obtained in three steps: 1) assuming that both the action of the full Hamiltonian, and the overlap between wavefunctions, is restricted to nearest neighbour contributions (the tight-binding model restricted to the nearest-neighbor hoppings gives a satisfactory description of the electronic structure of bilayer graphene at low energies, see e.g. \cite{Jung2014}), 2) expanding the momentum close to the Fermi points, and 3) assuming that the most relevant modes for the expansion of the full wavefunction $\Psi$ of a single electron are given by $p_z$ orbitals, where $z$ is the direction transverse to the bilayer, as these are the orbitals that contribute the most to conduction. In formulas
\begin{equation}
\psi \sim \lambda_1 \left| p_z ; A1 \right. \rangle + \lambda_2 \left| p_z ; B1 \right. \rangle + \lambda_3 \left| p_z ; A2 \right. \rangle + \lambda_4 \left| p_z ; B2 \right. \rangle \, ,
\end{equation}
where $A, B$ correspond to inequivalent types of carbon atoms in the same plane, and the index $1, 2$ corresponds to different layers. Given the form \eqref{eq:H} of the Hamiltonian, we can recognise that $B1$ and $A2$ are the two atoms that are directly aligned along the transversal direction. Solving the eigenvalue equation $H \lambda = E \lambda$ we find that there are no non-trivial eigenspinors with $\lambda_1 = 0$ when $|\kappa|\neq 0$. The component $\lambda_1$ can then be set to $1$ as the eigenvalue equation is homogeneous. The full solution follows from linear algebra
\begin{eqnarray} \label{eq:sol_bilayer_spinors1}
\lambda_2 &=& \frac{E}{\hbar v_F\bar{\kappa}} \, , \quad \lambda_3 = \sigma \frac{E}{\hbar v_F \bb{\kappa}} \, , \quad
\lambda_4 = \sigma \frac{\kappa}{\bb{\kappa}} \, ,
\end{eqnarray}
where $\sigma=\pm 1$ is a factor labelling different bands that appear in the consistency condition satisfied by the energy
\begin{equation}
E^2 - |\hbar v_F \kappa |^2 - \sigma \gamma E = 0 \, .
\end{equation}
In \cite{Curvatronics2017} we expanded the solution around the Fermi points in the dimensionless parameter $\epsilon= \frac{\hbar v_F|\kappa|}{\gamma}$. For the metallic $E_1^{(\pm)}$ bands one obtains
\begin{eqnarray}
\lambda_2 &=& \frac{\hbar \kappa}{2 m v_F} \label{eq:bilayer_metallic_lambda2} +O\left(\epsilon^2\right) \, , \\
\lambda_3 &=& \mp \frac{2}{\alpha} \frac{\hbar \kappa}{2 m v_F} + O \left( \epsilon^2\right) \, , \label{eq:bilayer_metallic_lambda3} \\
\lambda_4 &=& \mp \frac{2}{\alpha} \frac{\kappa}{\bb{\kappa}} \, . \label{eq:bilayer_metallic_lambda4}
\end{eqnarray}
For simplicity we set $\tau = +1$ and focus on one of the Fermi points. To begin, it is useful noticing that for all values of the constant $A$ the following spinors
\begin{eqnarray}
\label{LL_solution1_E1_tau+}
\chi_1(t) &=&
e^{- i \frac{\hbar k^2 t}{2m}} \left( A \lambda_3 , \lambda_2 \right)^T \, , \\
\label{LL_solution2_E1_tau+}
\chi_2(t) &=&
e^{- i \frac{\hbar k^2 t}{2m}} \left( \lambda_1 ,A \lambda_4 \right)^T \, ,
\end{eqnarray}
satisfy the L\'evy-Leblond equations \cite{LevyLeblond1967}
\begin{eqnarray} \label{eq:LevyLeblond1}
&& i \hbar \, {\partial}_t \chi_2 + i \hbar v_F D \chi_1 = 0 \, , \\
&& D \chi_2 - i \frac{2 m v_F}{\hbar} \chi_1 = 0 \, , \label{eq:LevyLeblond2}
\end{eqnarray}
where $D = i \sigma^j k_j$ is the $2$--dimensional Dirac operator in phase space and the $\sigma^j$ are the Pauli matrices in the standard basis. As we will discuss in sec. \ref{sec:3}, these are coupled, first order equations that are invariant under the Galilei group of non-relativistic mechanics and describe non-relativistic spin $\frac{1}{2}$ fermions. In particular, the spinors with $\lambda_3 = 0 = \lambda_4$ correspond to a solution with definite pseudospin, and those with $\lambda_1 = 0 = \lambda_2$ to an independent solution with opposite pseudospin. We notice from the explicit form of the solution (\ref{eq:bilayer_metallic_lambda2}-\ref{eq:bilayer_metallic_lambda4}) that, if in the solution we exchange at the same time the energy $E_1^+$ with $E_1^-$, keeping $\kappa$ unchanged, then the components $\lambda_1$ and $\lambda_3$ of the solutions are unchanged, while $\lambda_2$ and $\lambda_4$ change sign. Thus, by taking a linear combination of such eigenstates of energies $E_1^+$ and $E_1^-$ it is possible to obtain `checkered' states where either $\lambda_2 = 0 = \lambda_4$, i.e. only sites of type $A$ are excited, or viceversa states with $\lambda_1 = 0 = \lambda_3$, i.e. only sites of type $B$ are excited. Conversely, taking a linear combination of the latter it is possible to build eigenstates of the energy. We also notice that, as explicit electronic excitations are defined by their Fourier coefficients for each value of the momentum $\kappa$, these are free parameters in our analysis, which concerns only the energy spectrum. This is why we are able to set $\lambda_1 = 1$ at $|\kappa | \neq 0$. In sec.\ref{sec:3} we will see how these results are generalized in the case of curved sheets. We will obtain equations for Galilean fermions that are a variation of the L\'evy-Leblond equation, which involve a well defined combination of pseudospin. We will see that, at least in the axially symmetric case, it is still true that linear combinations of states with opposite energies give checkered states.
In analogy with the fact that for monolayer graphene the Fermi points are called Dirac points because of the relativistic emergent symmetry, for the bilayer we will use the wording \textit{Galilei points}. At the Galilei point, i.e. for $\kappa = 0$, the solutions have $\lambda_2 = 0$, $\lambda_3 = 0$. As shown in Figure \ref{fig:overlap-b}, this corresponds physically to activating only the $p_z$ orbitals of the $A_1$ and $B_2$ atoms: these are not directly aligned along the $z$ direction, thus giving a small overlap of the electron orbitals.
\begin{figure}%
\centering
\subfloat[$\lambda_2=\lambda_3=0$]{\label{fig:overlap-b}\includegraphics[width=0.45\textwidth]{1.eps}}\qquad
\subfloat[$\lambda_1=\lambda_4=0$]{\label{fig:overlap-a}
\includegraphics[width=0.45\textwidth]{2.eps}}
\caption{Schematic representation of the overlap of the $p_z$ orbitals of bilayer graphene at the Galilei points. The two panels correspond to different choices of the eigenvectors.\ }
\label{fig:overlap}
\end{figure}
Conversely, when we perform a similar analysis for the insulating bands $E_2^{(\pm)}$ at the same $\tau = +1$ Galilei point we find
\begin{eqnarray}
\lambda_2 &=& \frac{2 m v_F}{\hbar \bb{\kappa}} + O \left( \epsilon^2\right) \label{eq:bilayer_insulating_lambda2} \, , \\
\lambda_3 &=& \pm \frac{2}{\alpha} \frac{2 m v_F}{\hbar \bb{\kappa}}+O \left( \epsilon^2\right) \label{eq:bilayer_insulating_lambda3} \, , \\
\lambda_4 &=& \pm \frac{2}{\alpha} \frac{\kappa}{\bb{\kappa}} \label{eq:bilayer_insulating_lambda4} \, .
\end{eqnarray}
and
\begin{eqnarray}
\label{LL_solution1_E2_tau+}
\chi_1(t) &=&
e^{- i \frac{\hbar k^2 t}{2m}} \left( \lambda_1 , A \lambda_4 \right)^T \, , \\
\label{LL_solution2_E2_tau+}
\chi_2(t) &=&
e^{- i \frac{\hbar k^2 t}{2m}} \left( A \lambda_3 , \lambda_2 \right)^T \, .
\end{eqnarray}
This time, at the Galilei point $\lambda_1 = 0$, $\lambda_4 = 0$ and only the orbitals of the $B_1$, $A_1$ atoms are activated, as shown in Figure \ref{fig:overlap-a}. For the $E_2^{(+)}$ bands the wavefunctions of the two orbitals are in phase, maximising the overlap energy, and for the $E_2^{(-)}$ are out of phase, thus minimising it.
\subsection{Three-layer graphene\label{sec:three_layers}}
Here we extend our approach to the three-layer graphene with the ABA stacking
of the layers, which is represented in Figure \ref{fig:flat3}.
\begin{figure}%
\centering
\subfloat[$3$--layer]{\label{fig:flat3}\includegraphics[width=0.45\textwidth]{5.eps}}\qquad
\subfloat[$4$--layer]{\label{fig:flat4}
\includegraphics[width=0.45\textwidth]{6.eps}}
\caption{(a). Schematic representation of the three-layer graphene lattice with the ABA stacking. (b) Four-layer graphene lattice with the ABAB stacking.}
\label{fig:flat}
\end{figure}
The effective low energy Hamiltonian for ABA stacked graphene around the K point can be found in \cite{KatnelsonEtAl_2011_Trilayer}:
\begin{equation} \label{eq:ABA_Hamiltonian}
H_\kappa^{(ABA)} = \left(
\begin{array}{cccccc}
0 & \hbar v_F \bb{\kappa} & 0 & 0 & 0 & 0 \\
\hbar v_F \kappa & 0 & \gamma & 0 & 0 & 0 \\
0 & \gamma & 0 & \hbar v_F \bb{\kappa} & 0 & \gamma \\
0 & 0 & \hbar v_F \kappa & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & \hbar v_F \bb{\kappa} \\
0 & 0 & \gamma & 0 & \hbar v_F \kappa & 0
\end{array}
\right) \, .
\end{equation}
Here we see that the atoms that are directly aligned along the transverse direction are $B_1$ and $A_2$, as before, and $A_2$ and $B_3$. Analysing the band structure one finds two touching bands linear in momentum
\begin{equation}
E_{ABA,0} = \pm\hbar v_F |\kappa| \, ,
\end{equation}
and a set of metallic and insulating bands given by
\begin{equation} \label{eq:EABA}
\hspace{-0.35cm} E_{ABA,i}^{(\pm)} = \pm \left[ \gamma^2 + \hbar^2 v_F^2 |\kappa|^2 + (-1)^i \gamma \sqrt{\gamma^2 + 2 \hbar^2 v_F^2 |\kappa|^2} \right]^{\frac{1}{2}} \hspace{-0.25cm}, \quad i = 1, 2 \, .
\end{equation}
We begin by analising the metallic and insulating bands: we perform an analysis near the Fermi points in order to describe the detailed structure of the spinors. This will allow us to build spinors that satisfy the L\'evy-Leblond equations. We expand the energy bands in the dimensionless parameter $\epsilon = \frac{\hbar v_F |\kappa|}{\gamma}$, finding for the two metallic bands
\begin{equation} \label{eq:ABA_metallic_low_energy}
E_{ABA,1}^{(\pm)} \sim \frac{\hbar^2 |\kappa|^2}{2m} + O\left( \epsilon^4 \right) \, ,
\end{equation}
with effective mass $m = \pm \frac{\gamma}{\sqrt{2} v_F^2}$, or equivalently, recalling \eqref{eq:alpha}, $\alpha =\sqrt{2}$,
a smaller value than for $AB$ bilayer graphene.
For the two insulating bands
\begin{equation} \label{eq:ABA_insulating_low_energy}
E_{ABA,2}^{(\pm)} \sim 2 m v_F^2 + \frac{\hbar^2 |\kappa|^2}{2 m } + O\left( \epsilon^4 \right) \, .
\end{equation}
This is the same gap, for given mass, found in the bilayer case. The energy is again in non-relativistic form and therefore we expect that non-relativistic fermions will appear.
The full formula \eqref{eq:EABA} for the energy levels can be rewritten in terms of $m$ as
\begin{equation} \label{eq:EABA-2}
\hspace{-0.35cm} E_{ABA,i}^{(\pm)} = \pm \left[ 2 m^2 v_F^4 + \hbar^2 v_F^2 |\kappa|^2 + (-1)^i 2 m^2 v_F^2 \sqrt{1 + \frac{\hbar^2 |\kappa|^2}{m^2 v_F^2}} \right]^{\frac{1}{2}} \hspace{-0.25cm}, \quad i = 1, 2.
\end{equation}
It is a different formula from \eqref{eq:energy_free1} but it has an identical low energy limit.
The spinor solution of the equation $H\lambda = E \lambda$ is given by
\begin{eqnarray} \label{eq:sol_3layer_spinors1}
\lambda_1=1\, , \quad \lambda_2 = \frac{E}{\hbar v_F\bar{\kappa}} \, , \quad \lambda_3 = \sigma \alpha \frac{E}{\hbar v_F \bb{\kappa}} \, , \quad
\lambda_4 = \sigma \alpha \frac{\kappa}{\bb{\kappa}} \, , \quad \lambda_5 = 1 \, , \quad \lambda_6 = \frac{E}{\hbar v_F\bar{\kappa}} \, ,
\end{eqnarray}
where $\sigma=\pm 1$ is a factor labelling different bands that appears in the consistency condition satisfied by the energy
\begin{equation}
E^2 - |\hbar v_F \kappa |^2 - \sigma \alpha \gamma E = 0 \, .
\end{equation}
Now we analyse the spinor solutions close to the Fermi point for the metallic bands:
we substitute the low energy limit \eqref{eq:ABA_metallic_low_energy} into the spinor solution, and retain the lowest order in the parameter $\epsilon$. The result is given by \eqref{eq:bilayer_metallic_lambda2}--\eqref{eq:bilayer_metallic_lambda4} plus
\begin{eqnarray}
\lambda_5 &=& \frac{4}{\alpha^2} -1 \label{eq:trilayer_metallic_5} \, , \\
\lambda_6 &=& \left( \frac{4}{\alpha^2} -1 \right) \frac{\hbar \kappa}{2 m v_F} + O (\epsilon^2) \, . \label{eq:trilayer_metallic_6}
\end{eqnarray}
If we build $\chi_1$ and $\chi_2$ as in \eqref{LL_solution1_E1_tau+}, \eqref{LL_solution2_E1_tau+} then again these satisfy the L\'evy-Leblond equations. There is also another set of L\'evy-Leblond equations with spinors
\begin{eqnarray}
\chi_1(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m}} \left( \lambda_3 , B \lambda_6 \right)^T \, , \\
\chi_2(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m}} \left( B\lambda_5 , \lambda_4 \right)^T \, ,
\end{eqnarray}
where $B$ is another arbitrary constant. As for the bilayer, at the Fermi point $\kappa = 0$ the $p_z$ orbitals of the atoms directly aligned $B1$, $A_2$, $B_3$ are turned off. So we see that the non-relativistic fermions for the metallic bands of three-layer graphene consist of states where interlayer hopping happens between all three layers, generalizing the physics occurring in bilayer graphene: interlayer hopping is a fundamental ingredient in order to generate a mass for the excitations.
The same procedure, applied to the insulating bands yields the expansion \eqref{eq:bilayer_insulating_lambda2}--\eqref{eq:bilayer_insulating_lambda4}, plus the low energy limit
\begin{eqnarray}
\lambda_5 &=& \frac{4}{\alpha^2} -1 \label{eq:trilayer_insulating_5} \, , \\
\lambda_6 &=& \left( \frac{4}{\alpha^2} -1 \right) \frac{2 m v_F}{\hbar \bb{\kappa}} + O \left(\epsilon^2\right) \, . \label{eq:trilayer_insulating_6}
\end{eqnarray}
This time at the Fermi point the contributions of the atoms $A_1$, $B_2$, $A_3$ are turned off. The L\'evy-Leblond spinors in this case are given by
\begin{eqnarray}
\chi_1(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m}} \left( B\lambda_5 , \lambda_4 \right)^T \, , \\
\chi_2(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m}} \left( \lambda_3 , B \lambda_6 \right)^T \, .
\end{eqnarray}
We conclude this section analysing the bands that are linear in momentum. The solution close to the bottom of such Dirac bands is given by
\begin{eqnarray}
&& \lambda_1 = 1 \, , \qquad \lambda_2 = \pm \sqrt{\frac{\kappa}{\bb{\kappa}}} \, , \\
&& \lambda_3 = 0 \, , \quad \lambda_4 = 0 \, , \\
&& \lambda_5 = -1 \, , \quad \lambda_6 = \mp \sqrt{\frac{\kappa}{\bb{\kappa}}} \, .
\end{eqnarray}
Then the spinors
\begin{eqnarray}
\psi^{(\pm)}_1(t) &=& e^{ i v_F |\kappa| t} \left( \begin{array}{c} \lambda_1 \\ \lambda_2 \end{array} \right) \, , \\
\psi^{(\pm)}_2(t) &=& e^{ i v_F |\kappa| t} \left( \begin{array}{c} \lambda_5 \\ \lambda_6 \end{array} \right) \, ,
\end{eqnarray}
satisfy the Weyl equations for massless spinors
\begin{equation}
{\partial}_t \psi^{(\pm)} = \pm v_F D \psi^{(\pm)} \, .
\end{equation}
These solutions are localised on the layers $1$ and $3$, and they vanish in the intermediate layer $2$. On the layers, they describe massless modes moving at the speed of sound $v_F$. One can think of these solutions in this way: the electronic wavefunctions undergo totally destructive inteference in the middle layer, and the remaining electrons are confined to the top and bottom layers. Since interlayer hopping is forbidden in these localized states, then the excitations become massless again as for monolayer graphene.
\subsection{Four-layer graphene\label{sec:four_layers}}
The case of four-layer graphene with the ABAB stacking of the layers
is finally considered. Its schematic representation is reported
in figure \ref{fig:flat4}.
The ABAB Hamiltonian $H_k^{(ABAB)}$ around the K point is given by
\begin{equation} \label{eq:ABAB_Hamiltonian}
H_\kappa^{(ABAB)} = \left(
\begin{array}{cccccccc}
0 & \hbar v_F \bb{\kappa} & 0 & 0 & 0 & 0 & 0 & 0 \\
\hbar v_F \kappa & 0 & \gamma & 0 & 0 & 0 & 0 & 0\\
0 & \gamma & 0 & \hbar v_F \bb{\kappa} & 0 & \gamma & 0 & 0\\
0 & 0 & \hbar v_F \kappa & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & \hbar v_F \bb{\kappa} & 0 & 0\\
0 & 0 & \gamma & 0 & \hbar v_F \kappa & 0 & \gamma & 0 \\
0 & 0 & 0 & 0 & 0 & \gamma & 0 & \hbar v_F \bb{\kappa} \\
0 & 0 & 0 & 0 & 0 & 0 & \hbar v_F \kappa & 0 \\
\end{array}
\right) \, .
\end{equation}
This is the same as for the three-layer case, with the addition of atoms $B_3$, $A_4$ directly aligned. We plain to repeat the analysis done in the previous section: first finding the shape of the bands, then looking for special types of excitations. In this case we will find only L\'evy-Leblond spinors, and no massless excitations.
Analysing the eigenvalues of $H_k^{(ABAB)}$ we find 8 bands, expressed in terms of two masses:
\begin{eqnarray}
m_1 &=& \frac{\sqrt{5} -1}{4} m_0 \sim 0.309 \, m_0 \, , \\
m_2 &=& \frac{\sqrt{5} +1}{4} m_0 \sim 0.809 \, m_0 \, .
\end{eqnarray}
In our notation $\alpha_1 = \frac{4}{\sqrt{5}-1} \sim 3.236$, $\alpha_2 = \frac{4}{\sqrt{5}+1} \sim 1.236$.
For each mass $m_i$ there are two metallic and two insulating bands. For $m_1$:
\begin{equation} \label{eq:energy_free1_4layers}
E_{i,m_1}^{(\pm)} = \pm \left[ (-1)^{i} m_1 v_F^2 + \sqrt{ m_1^2 v_F^4 + |\hbar v_F\kappa|^2} \right] \, , i = 1,2 ,
\end{equation}
And for $m_2$
\begin{equation} \label{eq:energy_free2_4layers}
E_{i,m_2}^{(\pm)} = \pm \left[ (-1)^{i} m_2 v_F^2 + \sqrt{ m_2^2 v_F^4 + |\hbar v_F\kappa|^2} \right] \, , i = 1,2 ,
\end{equation}
The components of the energy eigenspinors are given by
\begin{eqnarray} \label{eq:sol_4layer_spinors1}
\lambda_1 &=&1\, ,\quad\lambda_2 = \frac{E}{\hbar v_F\bar{\kappa}} \, , \quad \lambda_3 = \sigma \frac{2}{\alpha_i} \frac{E}{\hbar v_F \bb{\kappa}} \, , \quad
\lambda_4 = \sigma \frac{2}{\alpha_i} \frac{\kappa}{\bb{\kappa}} \, , \quad
\lambda_5 = \left(\frac{4}{\alpha_i^2}-1\right) \, , \\ \lambda_6 &=& \left(\frac{4}{\alpha_i^2}-1\right) \frac{E}{\hbar v_F\bar{\kappa}} \, , \quad
\lambda_7 = \sigma \frac{2}{\alpha_i} \left(\frac{4}{\alpha_i^2}-1\right) \frac{E}{\hbar v_F\bar{\kappa}} \, , \quad \lambda_8 = \sigma \frac{2}{\alpha_i} \left(\frac{4}{\alpha_i^2}-1\right) \frac{\kappa}{\bar{\kappa}} \, ,
\end{eqnarray}
where $\sigma=\pm 1$ is a factor labelling different bands that appears in the consistency condition satisfied by the energy
\begin{equation}
E^2 - |\hbar v_F \kappa |^2 - \sigma \frac{2}{\alpha_i} \gamma E = 0 \, .
\end{equation}
At the bottom of the metallic bands $E_{1,m_i}^{(\pm)}$ band they take the form \eqref{eq:bilayer_metallic_lambda2}-\eqref{eq:bilayer_metallic_lambda4}, \eqref{eq:trilayer_metallic_5}-\eqref{eq:trilayer_metallic_6} plus
\begin{eqnarray}
&& \hspace{-1cm} \lambda_7 = \mp \frac{2}{\alpha_i} \left(\frac{4}{\alpha_i^2} - 2 \right) \frac{\hbar \kappa}{2m_i v_F}+O\left(\epsilon^2\right) \, , \\
&& \hspace{-1cm} \lambda_8 = \mp \frac{2}{\alpha_i} \left(\frac{4}{\alpha_i^2} - 2 \right) \frac{\kappa}{\bb{\kappa}} \, .
\end{eqnarray}
Similarly, at the bottom of the insulating bands $E_{2,m_i}^{(\pm)}$ the spinor takes the form \eqref{eq:bilayer_insulating_lambda2}-\eqref{eq:bilayer_insulating_lambda4}, \eqref{eq:trilayer_insulating_5}-\eqref{eq:trilayer_insulating_6} plus
\begin{eqnarray}
&& \hspace{-1cm} \lambda_7 = \pm \frac{2}{\alpha_i} \left(\frac{4}{\alpha_i^2} - 2 \right) \frac{2m_i v_F}{\hbar \bb{\kappa}}+O\left(\epsilon^2\right) \, , \\
&& \hspace{-1cm} \lambda_8 = \pm \frac{2}{\alpha_i} \left(\frac{4}{\alpha_i^2} - 2 \right) \frac{k}{\bb{\kappa}} \, .
\end{eqnarray}
Therefore we see repeating the same pattern that we found for the bilayer and three-layer cases: at the Galilei point for the metallic bands the component related to the $A_4$ atom is turned off, together with those of all the atoms that are directly aligned, and at the Galilei point for the insulating bands conversely the $B_4$ component is zero.
The new L\'evy-Leblond spinors for the metallic bands are
\begin{eqnarray}
\chi_1(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m_i}} \left( C\lambda_7 , \lambda_6 \right)^T \, , \\
\chi_2(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m_i}} \left( \lambda_5 , C \lambda_8 \right)^T \, ,
\end{eqnarray}
while for the insulating bands
\begin{eqnarray}
\chi_1(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m_i}} \left( \lambda_5 , C \lambda_8 \right)^T \, , \\
\chi_2(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m_i}} \left( C\lambda_7 , \lambda_6 \right)^T \, ,
\end{eqnarray}
where $C$ is another arbitrary constant. We see here that with four layers, no configurations with totally destructive interference in the middle layers appear. All states present interlayer hopping and are massive.
\section{The curved case\label{sec:3}}
\subsection{The L\'evy-Leblond equation}
The L\'evy-Leblond equations \eqref{eq:LevyLeblond1}, \eqref{eq:LevyLeblond2} are coupled, first order equations. L\'evy-Leblond in his seminal paper \cite{LevyLeblond1967} showed that these are obtained from the theory of representations of the Galilei group applied to particles of spin $\frac{1}{2}$. In general, the L\'evy-Leblond equations can be solved in the following way: solving for $\chi_1$ in \eqref{eq:LevyLeblond2} and plugging it into \eqref{eq:LevyLeblond1} we obtain the second order, uncoupled equation
\begin{equation} \label{eq:Schrodinger_full}
i \hbar \, {\partial}_t \chi_2 = - \frac{\hbar^2}{2m} \nabla^2 \chi_2 \, ,
\end{equation}
which in fact is the Pauli equation in the special case of the external magnetic field set to zero, see \cite{Curvatronics2017} for the case of a non-zero field. This equation admits the independent solutions
\begin{equation} \label{eq:chi2_solution_flat}
\chi_{2,\uparrow} = e^{- \frac{i \hbar |\kappa|^2 t}{2m} } \left( \begin{array}{c} 1 \\ 0 \end{array} \right) \, , \qquad
\chi_{2,\downarrow} = e^{- \frac{i \hbar |\kappa|^2 t}{2m} } \left( \begin{array}{c} 0 \\ 1 \end{array} \right)
\end{equation}
which correspond to different values of $2D$ (pseudo)spin. Then using \eqref{eq:LevyLeblond2} $\chi_1$ is obtained by differentiation of $\chi_2$ as
\begin{equation}
\hspace{-0.2cm} \chi_{1,\downarrow} = \frac{\hbar}{2 m v_F} e^{- \frac{i \hbar |\kappa|^2 t}{2m} } \left( \begin{array}{c} 0 \\ \kappa \end{array} \right) \, , \quad
\chi_{1,\uparrow} = \frac{\hbar}{2 m v_F} e^{- \frac{i \hbar |\kappa|^2 t}{2m} } \left( \begin{array}{c} \bb{\kappa} \\ 0 \end{array} \right)
\end{equation}
By linearity one can then construct the full solution:
\begin{eqnarray}
\chi_1 &=& a \chi_{1,\downarrow} + b \chi_{1,\uparrow} \, , \\
\chi_2 &=& a \chi_{2,\uparrow} + b \chi_{2,\downarrow} \, .
\end{eqnarray}
Let us compare this general formula with, for example, the L\'evy-Leblond spinor obtained for the metallic bands $E_1^{(\pm)}$ in the case of bilayer graphene:
\begin{eqnarray}
\chi_1(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m}} \left( \mp \frac{2}{\alpha} \frac{\hbar \kappa}{2mv_F} , \frac{\hbar \kappa}{2mv_F} \right)^T \, , \\
\chi_2(t) &=&
e^{- i \frac{\hbar |\kappa|^2 t}{2m}} \left( 1 , \mp \frac{2}{\alpha} \frac{\kappa}{\bb{\kappa}} \right)^T \, .
\end{eqnarray}
This can be obtained from the general formula setting $a=1$, $b = \mp \frac{2}{\alpha}\frac{\kappa}{\bb{\kappa}}$. So the L\'evy-Leblond spinor that occurs for a definite value of the energy is given by a linear superposition of spinors with definite values of the pseudospin. Conversely, a linear superposition of L\'evy-Leblond spinors from the bands $E_1^{(+)}$ and $E_1^{(-)}$ corresponds to states of definite value of the pseudospin. Similar formulae hold for L\'evy-Leblond spinors related to the insulating bands, as well as for the three- and four-layer case. States with definite pseudospin in the case of bilayer graphene close to the Galilei points are schematically shown in Figure \ref{fig:superpos}.
\begin{figure}%
\centering
\subfloat[$\lambda_1=\lambda_2=0$]{\label{fig:superpos-a}\includegraphics[width=0.45\textwidth]{4.eps}}\qquad
\subfloat[$\lambda_3=\lambda_4=0$]{\label{fig:superpos-b}
\includegraphics[width=0.45\textwidth]{3.eps}}
\caption{States with definite pseudospin in bilayer graphene close to the Galilei points. From eqs.\eqref{LL_solution1_E1_tau+}, \eqref{LL_solution2_E1_tau+}, \eqref{LL_solution1_E2_tau+}, \eqref{LL_solution2_E2_tau+} it can be seen that for both the metallic and insulating bands the two eigenstates of pseudospin have either $\lambda_1 = 0 = \lambda_2$, or $\lambda_3 = 0 = \lambda_4$. These states correspond to excitations concentrated in only one of the two planes. }
\label{fig:superpos}
\end{figure}
We now recall the relationship with the massless Dirac equation in $4$ dimensions. It has been shown in \cite{Curvatronics2017} that given a solution $\left( \chi_1, \chi_2 \right)$ of the L\'evy-Leblond equations it is possible to construct a massless Dirac spinor in an effective 4--dimensional Minkowski spacetime
\begin{equation} \label{eq:Minkowski}
g_{\mu\nu} dx^\mu dx^\nu = dx^2 + dy^2 + 2 du dv \, ,
\end{equation}
where $u$, $v$ are conjugate null coordinates. The massless 4D spinor is given by
\begin{equation} \label{eq:spinor_4d_decomposition}
\h{\Psi}(u,v,x,y) = e^{i \frac{m v_F}{\hbar} v} \left( \begin{array}{c} \chi_1(u,x,y) \\ \chi_2(u,x,y) \end{array} \right) \, .
\end{equation}
In terms of degrees of freedom, the massless Dirac equation in 4D has two degrees of freedom, corresponding to two different allowed chiralities. In turn, the L\'evy-Leblond equations admit two independent solutions, as just seen. This fact is well known in the literature \cite{Duval1985,Duval1995_1,Duval1995_2,Cariglia2012}, and is related to the concept of the Eisenhart-Duval lift of non-relativistic mechanics \cite{Eisenhart1928,DBKP,DGH91}. The relationship can be clearly seen in terms of the Gamma matrices adapted to the 4D lift:
\begin{eqnarray}
\Gamma^+ &=& \left( \begin{array}{cc} 0 & \mathbb{I} \\ 0 & 0 \end{array} \right) , \, \qquad \Gamma^- = 2 \left( \begin{array}{cc} 0 & 0 \\ \mathbb{I} & 0 \end{array} \right) , \, \qquad
\Gamma^i = \left( \begin{array}{cc} \sigma_i & 0 \\ 0 & - \sigma_i \end{array} \right) \, . \nn
\end{eqnarray}
In terms of these the chirality matrix is calculated as
\begin{equation} \label{eq:Gamma_star}
\Gamma^* = i \, \Gamma^0 \, \Gamma^1 \, \Gamma^2 \, \Gamma^3 = \left( \begin{array}{cc} - \sigma_3 & 0 \\ 0 & \sigma_3 \end{array} \right) \, ,
\end{equation}
showing that the spinor $\h{\Psi}$ built from the combination $a=1$, $b=0$ has a definite chirality, which by convention we can call right, and the one built using $a=0$, $b=1$ has opposite chirality, say left. That chirality in 4D can be related to pseudospin in 2D can be understood in the following way. First, we notice that in 4D we can introduce a timelike variable $t$ and a spacelike variable $z$ according to $2dudv = -dt^2 + dz^2$. Then, from eq.\eqref{eq:Gamma_star} chirality is decomposed into a boost in the $z$ direction, times a rotation around the $z$ axis. Spinors of the form \eqref{eq:spinor_4d_decomposition} with either $\chi_1 = 0$ or $\chi_2 = 0$ are eigenspinors of boosts along the $z$ direction: this is related to the fact that the metric \eqref{eq:Minkowski} is written in null coordinates $u, v$ and that the spinors are adapted to the coordinates. So for these spinors 4D chirality and 2D isospin are directly related.
\subsection{The role of curvature}
We now consider curved few layers of graphene. Our chief example will be the bilayer case, however similar considerations apply to the case of three or four layers.
When the radius of curvature is small enough we consider a Hamiltonian that satisfies the following two requirements. First, the Hamiltonian should be self-adjoint, so that the energy eigenvalues are real. Second, it should be a covariant generalization of \eqref{eq:H}, which reduces to \eqref{eq:H} when the metric is flat. In order to do so, we replace the momentum $k$ with $-i \nabla$, the 2D spinorial covariant derivative that includes the spin connection of the curved metric $g_{ij} dq^i dq^j$,
with $q^i = \{x,y\}$.
\begin{figure}%
\centering
\subfloat[spherical]{\label{fig:elliptic}\includegraphics[width=0.45\textwidth]{8.eps}}\qquad
\subfloat[hyperbolic]{\label{fig:hyperbolic}
\includegraphics[width=0.45\textwidth]{7.eps}}
\caption{Configurations with curvature for the three-layer graphene system exploitable for curvatronics, to tune the electronic properties of few-layer graphene by geometrical effects. (a) Spherical (positive) curvature. (b) Hyperbolic (negative) curvature.}
\label{fig:3curved}
\end{figure}
In sec. \ref{sec:two_layers} we have seen how the L\'evy-Leblond spinors appeared reorganizing the spinor components as per \eqref{LL_solution1_E1_tau+}. Motivated by this insight we make the following change of basis
\begin{eqnarray}
|1\rangle &=& |A2\rangle \, \label{eq:change1} \\
|2\rangle &=& |B1\rangle \, \\
|3\rangle &=& |A1\rangle \, \\
|4\rangle &=& |B2\rangle \label{eq:change4} \, ,
\end{eqnarray}
represented by the following linear function $R: \{A1,B1,A2,B2\} \mapsto \{1,2,3,4\}$
\begin{equation} R = \left( \begin{array}{cccc}
0 & 0 & 1 & 0 \\
0 & 1 & 0 & 0 \\
1 & 0 & 0 & 0 \\
0 & 0 &0 & 1
\end{array}\right).
\end{equation}
The Hamiltonian changes accordingly into
\begin{eqnarray} \label{eq:H_transformation_rule}
H^\prime_{\kappa} &=& R H_{\kappa} R^{-1}
\end{eqnarray}
or
\begin{eqnarray} \label{eq:newH}
H^\prime_{\kappa} &=&
\left( \begin{array}{c|c}
\begin{array}{cc} 0 & \gamma \\ \gamma & 0 \end{array}
& - i \hbar v_F D \\ \hline
- i \hbar v_F D &
\begin{array}{cc} 0 & 0 \\ 0 & 0 \end{array}
\end{array} \right) \, ,
\end{eqnarray}
which is (formally) self-adjoint since $D^\dagger = - D$. The electronic wavefunction is
\begin{eqnarray}
|\Psi\rangle &=& \lambda_1 |A1\rangle + \lambda_2 |B1\rangle + \lambda_3 |A2\rangle + \lambda_4 |B2 \rangle \nn \\
&=& \lambda_3 |1\rangle + \lambda_2 |2\rangle + \lambda_1 |3\rangle + \lambda_4 |4\rangle \, ,
\end{eqnarray}
and similarly to what we have done in the flat case we regroup the components into
\begin{eqnarray}
|\chi_1\rangle = \left( \begin{array}{c} \lambda_3 \\ \lambda_2 \end{array} \right) \, ,
\end{eqnarray}
\begin{eqnarray}
|\chi_2\rangle = \left( \begin{array}{c} \lambda_1 \\ \lambda_4 \end{array} \right) \, .
\end{eqnarray}
The eigenvalue equation for $H^\prime_{\kappa}$ becomes
\begin{eqnarray}
E \chi_1 &=& \left( \begin{array}{cc} 0 & \gamma \\ \gamma & 0 \end{array} \right) \chi_1 - i \hbar v_F D \chi_2 \\
E \chi_2 &=& - i \hbar v_F D \chi_1 \, .
\end{eqnarray}
We can decouple the equations by obtaining $\chi_2$ from the second, and inserting it into the first.
An important case is that of metallic bands, for which the $E\chi_1$ term in the top equation is negligible with respect to the other two: in this limit
\begin{eqnarray} \label{eq:correct1}
&& 0 = \left( \begin{array}{cc} 0 & \gamma \\ \gamma & 0 \end{array} \right) \chi_1 - i \hbar v_F D \chi_2 \\
&& E \chi_2 = - i \hbar v_F D\chi_1 \, , \qquad (\mbox{metallic bands}). \label{eq:correct2}
\end{eqnarray}
Recalling that $\gamma = 2 |m| v_F^2$, these equations are similar to the stationary L\'evy-Leblond equations, with the difference that the term $-2mv_F^2$ is replaced by the matrix \begin{small}$\left( \begin{array}{cc} 0 & \gamma \\ \gamma & 0 \end{array} \right)$\end{small}. To understand the relationship with what we found in the flat case, we isolate $\chi_2$ in \eqref{eq:correct2}, plug it into \eqref{eq:correct1} and obtain
\begin{equation} \label{eq:chi1_quadratic}
E\chi_1 = \frac{\hbar^2 v_F^2}{\gamma} \left( \begin{array}{cc} 0 & 1 \\ 1 & 0 \end{array} \right) D^2 \chi_1 \, .
\end{equation}
This is the eigenvalue equation for the energy in the curved case. In flat space and for metallic bands we found $E=\pm \frac{\hbar^2 v_F^2 |\kappa|^2}{\gamma}$ so that \eqref{eq:chi1_quadratic} became
\begin{equation}
\chi_1 = \left( \begin{array}{cc} 0 & \mp1 \\ \mp1 & 0 \end{array} \right) \chi_1 \, ,
\end{equation}
and in fact our solution was
\begin{eqnarray} \label{eq:chi1_flat_solutions}
\chi_1 = \frac{\hbar}{2m v_F \kappa} \left( \begin{array}{c} \mp1 \\ 1 \end{array} \right) \, .
\end{eqnarray}
Notice that in this case $\chi_1$ is an eigenspinor of the matrix $\sigma_1$ with eigenvalues $\mp 1$, so that we are justified in exchanging the matrix \begin{small}$\left( \begin{array}{cc} 0 & \gamma \\ \gamma & 0 \end{array} \right)$\end{small} with $-2mv_F^2$. However, in general the operator on the right hand side of \eqref{eq:chi1_quadratic} does not commute with $\sigma_1$, and as a consequence eqs.(\ref{eq:correct1}-\ref{eq:correct2}) are a variation of the L\'evy-Leblond equations.
Our theory predicts that there will appear a gap in the spectrum if there are no solutions of \eqref{eq:chi1_quadratic} for the eigenvalue $E=0$. In this particular case the analysis reduces to finding zero modes of $D$ , and it is a known result that if the surface has positive curvature then the spectrum of $D$ does not include the eigenvalue zero.
Eq.\eqref{eq:chi1_quadratic} is in general non-trivial to study, and to gain a better understanding we focus our attention on the axisymmetric case. That is, we consider a metric of the type
\begin{equation}
g_{ij} dq^i dq^j = dr^2 + C(r)^2 d\varphi^2 \, .
\end{equation}
For example for $C(r)^2 = \sin(r)^2$ the geometry is that of a sphere, with constant positive curvature, and for $C(r)^2 = r^2 + a^2$, with $a$ a constant, the geometry describes two flat ends joined by a tube, and has negative, non-constant, curvature. In the generic axisymmetric case the Dirac operator is given by
\begin{eqnarray}
D &=& \sigma_1 \left( {\partial}_r + \frac{C^\prime}{2 C} \right) + \sigma_2 \frac{{\partial}_\varphi}{C} \nn \\
&=& \left( \begin{array}{c|c}
0 & \mathcal{O} \\
\hline - \mathcal{O}^\dagger & 0
\end{array} \right)
\end{eqnarray}
where
\begin{eqnarray}
\mathcal{O} &=& {\partial}_r + \frac{C^\prime}{2C} + \frac{-i{\partial}_\varphi}{C} \nn \\
\mathcal{O}^\dagger &=& -{\partial}_r - \frac{C^\prime}{2C} + \frac{-i{\partial}_\varphi}{C} \, ,
\end{eqnarray}
so that
\begin{equation}
D^2 = -
\left( \begin{array}{c|c}
\mathcal{O} \mathcal{O}^\dagger & 0 \\
\hline 0 & \mathcal{O}^\dagger \mathcal{O}
\end{array} \right) \, .
\end{equation}
We will also assume that the operator acts on a spinor of the form $\chi_1 = e^{i l \varphi} \left(\begin{array}{c} a(r) \\ b(r) \end{array}\right)$, $l = \pm \frac{1}{2}, \pm \frac{3}{2}, \dots$. From our calculation then eq.\eqref{eq:chi1_quadratic} reduces to the pair of equations
\begin{eqnarray}
\frac{\gamma E}{\hbar^2 v_F^2} a &=& - B b \, , \label{eq:gammaE_1} \\
\frac{\gamma E}{\hbar^2 v_F^2} b &=& - A a \, , \label{eq:gammaE_2} \,
\end{eqnarray}
where we defined the second order, self-adjoint, positive operators $A = \mathcal{O} \mathcal{O}^\dagger$, $B = \mathcal{O}^\dagger \mathcal{O}$. These equations can be decoupled, giving
\begin{eqnarray}
\frac{\gamma^2 E^2}{\hbar^4 v_F^4} a &=& B A \, a \, , \label{eq:eigenvalue_a} \\
\frac{\gamma^2 E^2}{\hbar^4 v_F^4} b &=& A B \, b \label{eq:eigenvalue_b} \, ,
\end{eqnarray}
or in other words on the space of $\chi_1$ spinors the operator $\gamma^2 E^2$ is represented by
\begin{equation} \label{eq:energy_squared}
\frac{\gamma^2 E^2}{\hbar^4 v_F^4} = \left( \begin{array}{c|c}
BA & 0 \\
\hline 0 & AB
\end{array} \right) \, .
\end{equation}
To summarize, we obtained a fourth order equation for the square of the energy eigenvalues. Our equation is consistent since it is a known fact that for positive operators $A$ and $B$ the spectrum of $AB$ is the same of that of $BA$, and it is positive \cite{Hladnik1988}. Being fourth order, our equations will be challenging to solve analytically in general: the solution will require a careful but feasible numerical approach, which we postpone to future works. There are however some properties that are easy to discuss. First, since they are quadratic in $E$, then the spectrum will be symmetric under $E \rightarrow -E$. Second, as the operator in \eqref{eq:energy_squared} commutes with the matrix $\sigma_3$, then we can see that if $a$, $b$ are eigenfunctions related to the eigenvalue $E$, then $a$, $-b$ are eigenfunctions related to the eigenvalue $-E$. Thus, in analogy to the flat case, we are able to construct checkered solutions from linear combinations of solutions with opposite energy, and viceversa.
However, in general it will be the case that $a \neq b$, since $A \neq B$. Therefore the solutions cannot be related to L\'evy-Leblond fermions. There is however a regime where we reasonably expect that $A \sim B$. One can calculate $\left[\mathcal{O},\mathcal{O}^\dagger \right] = - 2l \frac{C^\prime}{C^2}$. Suppose the function $\left| 2 \frac{C^\prime}{C^2}\right|$ has a maximum $\mu$. We can substitute $B$ in \eqref{eq:gammaE_1} for $A + \left[\mathcal{O}^\dagger,\mathcal{O} \right]$ and, in those cases where
\begin{equation}
\frac{\gamma E}{\hbar^2 v_F^2} >> l \mu
\end{equation}
we can ignore the extra term and reduce the eigenvalue equations to
\begin{eqnarray}
\frac{\gamma E}{\hbar^2 v_F^2} a &=& - A b \, , \label{eq:coupled_1} \\
\frac{\gamma E}{\hbar^2 v_F^2} b &=& - A a \label{eq:coupled_2} \, .
\end{eqnarray}
These admit two types of solutions: solutions with $b = a$ give
\begin{equation}
\frac{\gamma E}{\hbar^2 v_F^2} a = - A a \, ,
\end{equation}
i.e. $a$ is an eigenfunction of $A$ and the energy is negative, or $b = - a$ which gives
\begin{equation}
\frac{\gamma E}{\hbar^2 v_F^2} a = A a \, ,
\end{equation}
i.e. $a$ is an eigenfunction of $A$ and the energy is positive. In this approximation of low angular momentum to energy ratio the problem simplifies considerably, as one needs to solve 2nd order differential equations. Summarising our results, we expect that deforming the few--layer graphene sheets provides an efficient way to increase or reduce an energy band gap between conduction and valence bands, since few layers graphene has several degrees of freedom with respect to which deformations can be introduced: both along the sheets and perpendicular to it. As an illustrative example, possible positive or negative curvatures
of the three-layer graphene are depicted in Figure \ref{fig:3curved}.
\section{Conclusions}
In this work we have extended the geometrical approach and the results of our previous work on curvatronics with bilayer graphene of Ref.\cite{Curvatronics2017} to the case of three and four layers of AB stacked graphene, and we have refined the role of isospin states in the case of the curved bilayer. We found that for flat layers L\'evy-Leblond fermions are still present at the bottom of the parabolic bands, which we call Galilei points of the Brillouin zone, and we gave a geometrical interpretation of the solutions in terms of the $p_z$ orbitals of graphene associated to them, on the basis of their non-trivial spatial overlap. We have discussed in detail the effective electronic masses present for three and four layers, comparing them with the bilayer graphene case, as well as the explicit form of the solutions. We analysed the L\'evy-Leblond equation and showed how energy eigenstates are given by the superpositions of checkered states. Using the relationship between the 2D L\'evy-Leblond equation and the 4D massless Dirac equation, we were able to relate 2D pseudospin with chirality in 4D. Lastly, we discussed the fact that curving few layers graphene provides a tool to tune the energy band gap, which has implications for tunable electronics based on curvature, or curvatronics with few-layer graphene systems. For instance, a graphene--based device with alternating regions of positive or negative curvature, inducing insulating or conductive states, could be exploited to produce non--volatile memories. It would be interesting for future research to study in detail the consequences of having curved few layers, as well as generalize our results to a higher number of layers, and to study the fermionic excitations away from the Galilei points. Another avenue of possible future research is studying the physics of pseudospin in settings with non-trivial curvature or topology, which might lead to novel effects.
Finally, very recently the class of Weyl metamaterials has been introduced in
Ref.\cite{Weststrom}. In these systems, chiral Weyl fermions
are moving in an artificial 3D curved-space geometry, with applications to
tunable novel electronic devices through curvature engineering and with
connections to cosmological theories. Therefore, the new field of
quantum physics in curved spaces is likely to become of great and practical interest
in the near future.
\vspace{0.2cm}
\textbf{Acknowledgments} --
M. Cariglia is funded by CNPq under project 303923/2015-6, and by a \textit{Pesquisador Mineiro} project n. PPM-00630-17. The authors acknowledge the collaboration within the MultiSuper International Network (http://www.multisuper.org) for exchange of ideas and suggestions.
\vspace*{-1ex}
|
\section{Introduction}
Parton distribution functions (PDFs) are key quantities for gaining an understanding of hadron structure and for making predictions for the cross sections in high-energy scattering experiments. In QCD factorization theorems for hard scattering processes~\cite{Collins:1989gx}, the relevant PDFs are defined in terms of the nucleon matrix elements of light-cone correlation operators. For example, in dimensional regularization with $d=4-2\epsilon$, the bare unpolarized quark PDF is
\begin{align} \label{eq:lcpdf}
q(x,\epsilon) \equiv\! \int\! {d\xi^-\over 4\pi} e^{-ixP^+\xi^-}\! \langle P | \bar{\psi} (\xi^-) \gamma^+ U (\xi^-, 0) \psi (0) | P \rangle ,
\end{align}
where $x$ is the momentum fraction, the nucleon momentum $P^\mu=(P^0,0,0,P^z)$, $\xi^{\pm}=(t\pm z)/\sqrt{2}$ are the light-cone coordinates, and the Wilson line is
\begin{align}
U(\xi^-,0) = P\exp\biggl(-ig\int_0^{\xi^-} d\eta^- A^+(\eta^-)\biggr)\ .
\end{align}
Most often the bare PDF is renormalized in the $\overline{\ensuremath{\operatorname{MS}}}$ scheme to obtain $q(x,\mu)$, and this renormalized PDF is used to make predictions for experiment. The relation is
\begin{eqnarray}
q(x,\epsilon) = \int_x^1 {dy\over y}\: Z^{\overline{\ensuremath{\operatorname{MS}}}}\Bigl({x\over y},\epsilon,\mu\Bigr) q(y,\mu)\ ,
\end{eqnarray}
where $\mu$ is the renormalization scale, and we have suppressed the flavor indices in the renormalization constant $Z^{\overline{\ensuremath{\operatorname{MS}}}}$ and PDFs. In light-cone quantization with $A^+=0$, the $\overline{\ensuremath{\operatorname{MS}}}$ definition has an interpretation as a parton number density.
So far our main knowledge of the PDF is obtained from global fits to deep inelastic scattering and jet data, see for example~\cite{Ball:2014uwa,Dulat:2015mca,Martin:2009iq,Alekhin:2017kpj,Buckley:2014ana}. On the other hand, calculating the PDF from first principles with QCD has been an attractive subject, which can for example provide access to spin and momentum distributions that are hard to determine experimentally. Several different approaches to this have been considered using the lattice theory which is a nonperturbative method to solve QCD. Since the lattice theory is defined in a discretized Euclidean space with imaginary time, it is very difficult to calculate Minkowskian quantities with real-time dependence such as the PDF. The first and most well explored option is calculating the moments of the PDF~\cite{Martinelli:1987zd,Martinelli:1988xs,Detmold:2001dv,Detmold:2002nf,Dolgov:2002zm} that are matrix elements of local gauge-invariant operators. However, since the lattice regularization breaks $O(4)$ rotational symmetry, the consequent mixing between operators of different dimensions makes it difficult to compute higher moments, which in practice has limited the amount of information that can be extracted from this approach. A method to improve this situation by restoring the rotational symmetry has been proposed in Ref.~\cite{Davoudi:2012ya}.
Other proposals include extracting the PDF from the hadronic tensor~\cite{Liu:1993cv,Liu:1998um,Liu:1999ak,Liu:2016djw} and the forward Compton amplitude~\cite{Chambers:2017dov}, possibly with flavor changing currents~\cite{Detmold:2005gg}, and the more general ``lattice cross sections"~\cite{Ma:2014jla,Ma:2017pxb}. Systematic lattice analyses of these approaches are under investigation, but challenges remain.
In Ref {\cite{Ji:2013dva}} Ji proposed that the $x$-dependence of the PDF can be extracted from a Euclidean distribution on the lattice, which can be understood in the language of the large momentum effective theory (LaMET)~\cite{Ji:2014gla}. This Euclidean distribution is referred to as the quasi-PDF, whose bare matrix element is defined using a spatial correlation of quarks along the $z$ direction,
\begin{eqnarray} \label{eq:qpdf}
\tilde{q}(x,P^z,\epsilon) \equiv \int_{-\infty}^\infty {dz\over 4\pi} e^{ixP^zz}\langle P | \bar{\psi} (z) \Gamma U (z, 0) \psi (0) | P \rangle \,,\nonumber\\
\end{eqnarray}
where $\Gamma=\gamma^z$, $z^\mu=ze^\mu$, $e^\mu=(0,0,0,1)$, and the Wilson line is
\begin{eqnarray}
U(z,0) = P\exp\left(-ig\int_0^z dz' A^z(z')\right)\ .
\end{eqnarray}
For finite momentum $P^z$, $\tilde{q}(x,P^z,\epsilon)$ has support in $-\infty < x < \infty$.
According to Ref.~\cite{Hatta:2013gta}, there is a universality class of operators that can be considered. For example, for the quasi-PDF, one could also replace $\Gamma=\gamma^z$ by $\Gamma=\gamma^0$ in Eq.~(\ref{eq:qpdf}) as both definitions reduce to the PDF under an infinite Lorentz boost along the $z$ direction. Unlike the PDF in Eq.~(\ref{eq:lcpdf}) that is invariant under the Lorentz boost, the quasi-PDF depends dynamically on it through the nucleon momentum $P^z$. When the nucleon momentum $P^z$ is much larger that the nucleon mass $M$ and $\Lambda_{\rm QCD}$, which is an attainable window on the lattice, the quasi-PDF can be factorized
into a matching coefficient and the PDF~\cite{Ji:2013dva,Ji:2014gla}. The factorization formula is
\begin{align} \label{eq:momfact}
\tilde{q}(x,P^z,\mu_R)
=& \int_{-1}^1 {dy\over |y|}\: C\Bigl({x\over y},{\mu_R\over \mu},{\mu\over p^z}\Bigr) \, q(y,\mu)
\nonumber\\
& + \mathcal{O}\biggl({M^2\over P_z^2},{\Lambda_{\ensuremath{\operatorname{QCD}}}^2\over P_z^2}\biggr)\ ,
\end{align}
where the renormalized quasi-PDF $\tilde{q}(x,P^z,\mu_R)$ is defined in a particular scheme at renormalization scale $\mu_R$, and $\mathcal{O}(M^2/P_z^2, \Lambda_{\rm QCD}^2/P_z^2)$ are power corrections suppressed by the nucleon momentum. In general the result for $C$ will depend on the choice of $\Gamma=\gamma^z$ or $\gamma^0$ and
renormalization schemes. For $\Gamma=\gamma^z$ the matching coefficient $C$ has been computed for the iso-vector quark quasi-PDF at one-loop level, first with a transverse momentum cutoff in~\cite{Xiong:2013bka}, confirmed in~\cite{Ma:2014jla,Alexandrou:2015rja}, and also recently determined in the regularization-invariant momentum subtraction (RI/MOM) scheme~\cite{Stewart:2017tvs}. Matching for the gluon quasi-PDF is calculated in~\cite{Wang:2017qyg,Wang:2017eel}. The matching coefficient $C$ is independent of the choice of states used for $\tilde q$ and $q$.\footnote{The scheme definition itself may separately involve a choice of state, such as in the RI/MOM scheme, but the result is still an operator renormalization that can be used for different choices of hadronic states for $\tilde q$ and $q$. The transverse cutoff and $\overline{\rm MS}$ schemes do not require even this choice.} Since matching calculations are carried out with quark states of momentum $p^z$, it can be tricky to know what the right choice to make is for $C$, and in some of the literature the choice of $p^z=P^z$ has been suggested when utilizing $C$ for the hadronic nucleon state. This is for example the case in the original quasi-PDF papers~\cite{Ji:2013dva,Ji:2014gla,Xiong:2013bka} and in the pioneering lattice calculations of the PDF from the quasi-PDF in~\cite{Lin:2014zya,Alexandrou:2015rja,Chen:2016utp,Alexandrou:2016jqi,Zhang:2017bzy,Alexandrou:2017huk,Chen:2017mzz,Lin:2017ani,Chen:2017gck}, which was summarized in Ref.~\cite{Lin:2017snn}. In the quasi-generalized parton distribution analysis in~\cite{Ji:2015qla} it was observed that one should take $p^z=|y|P^z$. Through our rigorous analysis of \eq{momfact} we show that the correct result for this equation is indeed $p^z=|y|P^z$.
Recently, a different procedure~\cite{Radyushkin:2017cyf} to extract PDFs from the same lattice QCD matrix element as in~\cite{Ji:2013dva}
has been proposed based on the Lorentz invariant variables of the spatial correlator\xspace (or pseudo-PDF), in place of the quasi-PDF. In this approach, one starts from the spatial correlator\xspace $\tilde{Q}_{\gamma^\mu}$ defined for $\mu=0$ or $\mu=z$ by
\begin{align}
{1\over2}\langle P | \tilde{O}_{\gamma^\mu}(z,\epsilon) | P\rangle
= & P^\mu \tilde{Q}_{\gamma^\mu}(\zeta=P^zz,z^2,\epsilon) \,,
\end{align}
which depends on the two Lorentz invariants $z^2$ and $\zeta=-z\cdot P = P^z z$; the latter is also called the Ioffe time.
For an arbitrary Dirac matrix $\Gamma$ the operator $\tilde{O}_\Gamma$ is defined as
\begin{align} \label{eq:OGamma}
\tilde{O}_{\Gamma}(z,\epsilon) = \bar{\psi} (z) \Gamma U (z, 0) \psi (0) \,.
\end{align}
This is the same spatial correlator (calculable on lattice) used to define the quasi-PDF in \eq{qpdf}, where $P^z$ is fixed and one Fourier transforms with respect to $z$. If instead $z^2$ is fixed, and we Fourier transform from the Ioffe time $\zeta$---which is in principle integrating over $P^z$---to the momentum fraction $x$, then one obtains the pseudo-PDF~\cite{Radyushkin:2017cyf},
\begin{align} \label{eq:pseudo}
\mathcal{P} \left( x, z^2,\epsilon\right)
= \int_{-\infty}^\infty \! \frac{d \zeta}{2\pi}\:
e^{ix \zeta}\: \tilde{Q}_{\gamma^0}\bigl( \zeta, z^2,\epsilon \bigr)
\,.
\end{align}
For arbitrary finite $z$, the pseudo-PDF only has support in $-1\le x \le 1$~\cite{Radyushkin:1983wh,Radyushkin:2016hsy}, but has no parton model interpretation.
(Again the pseudo-PDF can equally well be considered for $\Gamma=\gamma^z$.) The spatial correlator\xspace or pseudo-PDF approach has been explored on the lattice~\cite{Orginos:2017kos,Karpie:2017bzm}, where the short distance behavior was explored.
The PDF corresponds to the situation when $z^\mu$ is light-like, in which case the space-time correlator is referred to as the Ioffe-time distribution~\cite{Ioffe:1969kf},
\begin{align}
Q(\zeta,\epsilon) = \tilde Q_{\gamma^+}(\zeta=-P^+\xi^-,z^2=0,\epsilon)
\,.
\end{align}
When Fourier transformed this correlation gives the PDF
\begin{align} \label{eq:Qdefn}
q(y,\epsilon) =\int_{-\infty}^{\infty} {d\zeta\over 2\pi} \ e^{iy\zeta}\: Q(\zeta,\epsilon)\,.
\end{align}
In short the quasi-PDF and pseudo-PDF are different representations of the Euclidean spatial correlator\xspace, as summarized in Table.~\ref{tab:I}.
\begin{table}
\begin{center}
\begin{tabular}{ | c | c | c |}
\hline
Distribution & Fourier transform & Arguments\\[-2pt]
& from spatial correlator\xspace & \\ \hline
Spatial correlation & & $\zeta = zP^z, z^2$\\ \hline
Quasi-PDF & $z\to xP^z$ & $x, P^z$\\ \hline
Pseudo-PDF & $\zeta \to x$ & $x, z^2$\\ \hline
\end{tabular}
\caption{Summary of the relationship between different Euclidean distributions.}
\label{tab:I}
\end{center}
\end{table}
It was pointed out in Ref.~\cite{Ji:2017rah} that to obtain enough information to extract the PDF for the spatial correlator\xspace with small $z^2$, one has to do lattice calculations with large momenta $P^z$, which is the same requirement as for the quasi-PDF. Ref.~\cite{Ji:2017rah} also proposed that the renormalized pseudo-PDF satisfies the following small $z^2$ factorization,
\begin{align} \label{eq:pseudofact}
\mathcal{P} \left( x, z^2\mu_R^2\right)
&= \int {dy\over |y|}\: \mathcal{C}\left({x\over y}, {\mu_R^2\over \mu^2},z^2\mu_R^2\right) q(y,\mu)
\nonumber\\
&\ + {\cal O}(z^2\Lambda_{\rm QCD}^2, z^2 M^2) \,,
\end{align}
which they verified at order $O(\alpha_s)$ for the unpolarized iso-vector case with $\Gamma=\gamma^0$. (Again the coefficient ${\cal C}$ will depend on the choice of $\Gamma=\gamma^0$ or $\gamma^z$.)
In Ref.\cite{Ma:2014jla} a diagrammatic derivation of the factorization formula in \eq{momfact} for the quasi-PDF was given. Here we derive this factorization formula for the quasi-PDF in an alternate manner, and also show that spatial correlator\xspace and pseudo-PDF are different representations of the same
fundamental factorization. Our approach is based on the operator product expansion (OPE) for spacelike separated local operators~\cite{Wilson:1969zs}. For such operators the OPE has been proven for scalar field theory to all orders in perturbation theory~\cite{Brandt:1967rb,Wilson:1972ee,Zimmermann:1972tv}, and is widely assumed to hold for any renormalizable quantum field theory including QCD. By introducing auxiliary fields in place of the Wilson line~\cite{Dorn:1986dt}, the correlator in \eq{OGamma} is known to be equivalent to a product of local renormalizable operators of this type. Through our derivation we find the explicit form of the large $P^z$ and small $z^2$ factorization formulas in Eq.(\ref{eq:momfact}) and Eq.~(\ref{eq:pseudofact}) respectively, as well as the relationship between the matching coefficients $C$ and $\mathcal{C}$. Since the requirement for large $P^z$ and small $z^2$ is the same for both the quasi-PDF and pseudo-PDF approaches, there is in principle no fundamental difference in applying either one
to lattice calculations of the proton matrix element of $\tilde{O}_{\Gamma}(z)$. It is interesting to compare
both approaches utilizing the same lattice data, although they shall not yield different result in principle.
The rest of this paper is organized as follows: In Sec.~\ref{sec:ope}, we use an OPE of $\tilde{O}_\Gamma(z)$ to derive the large $P^z$ factorization of LaMET in Eq.~(\ref{eq:momfact}) and small $z^2$ factorization of the pseudo-PDF in Eq.~(\ref{eq:pseudofact}). We prove that one must take $p^z=|y|P^z$ in \eq{momfact}, so the corresponding argument in $C$ is $\mu/(|y|P^z)$.
(This OPE approach was used recently in Ref.~\cite{Ma:2017pxb} to prove the factorization theorem for the ``lattice cross sections", and the OPE proof carried out here was done independently and first presented in Ref.~\cite{Zhao:2017int}.)
In Sec.~\ref{sec:equiv}, we derive the spatial correlator\xspace, pseudo-PDF, and quasi-PDF distributions and matching coefficients at one-loop in $\overline{\rm MS}$ and analyze the Fourier-transform relation between the quasi-PDF and pseudo-PDF. Unlike earlier results for the quasi-PDF in $\overline{\rm MS}$, we also use dimensional regularization with minimal subtraction to renormalize divergences at $x=\pm\infty$. In Sec.~\ref{sec:ren}, we discuss how renormalization schemes other than $\overline{\rm MS}$ are easily incorporated into the factorization formulas.
In Sec.~\ref{sec:num} we carry out a numerical analysis of the matching coefficients, by computing the convolution in Eq.~(\ref{eq:momfact}) numerically using the PDF determined by global fits~\cite{Martin:2009iq}. We show that the difference between using $p^z=P^z$ and $p^z=|y|P^z$ in \eq{momfact} is an important effect, and that our $\overline{\rm MS}$ matching coefficients are insensitive to cutoffs in the convolution integral. In Sec.~\ref{sec:lattice}, we discuss the implications of our OPE analysis for the lattice calculation of the PDF in both the quasi-PDF and pseudo-PDF approaches. Finally, we conclude in Sec.~\ref{sec:summary}.
\section{Factorization from the OPE}
\label{sec:ope}
In this section we make use of the operator product expansion to derive the matching relation for the quasi-PDF, as well as the equivalent matching relations for the spatial correlator\xspace and pseudo-PDF. For simplicity these three equivalent cases are presented in separate subsections.
\subsection{OPE and Factorization for the Spatial Correlator}
The OPE is a technique to expand nonlocal operators with separation $z^\mu$ in terms of local ones in the Euclidean limit of $z^2\to 0$.
It can be applied to both bare regulated operators as well as renormalized operators, and our focus will be on the latter. For the gauge-invariant Wilson operator $\tilde{O}_\Gamma(z)$, it was proven that it can be multiplicatively renormalized in coordinate space as~\cite{Ji:2017oey,Ishikawa:2017faj}
\begin{eqnarray} \label{eq:ren}
\tilde{O}_\Gamma(z,\mu) = Z_{\psi,z}\, e^{\delta m |z|} \tilde{O}_\Gamma(z,\epsilon)\,,
\end{eqnarray}
where $\delta m$ captures the power divergence from the Wilson line self-energy, $Z_{\psi,z}$ only depends on the end points $z,0$ and renormalizes the logarithmic divergences. This multiplicative renormalization was also discussed earlier in Refs.~\cite{Ishikawa:2016znu,Chen:2016fxx,Constantinou:2017sej}. For simplicity, in this section we take $\Gamma=\gamma^z$ for \eq{ren}.
In the $\overline{\rm MS}$ scheme, the power divergence vanishes, and using the OPE
the renormalized $\tilde{O}_\Gamma(z,\mu)$ can be expanded
in terms of local gauge-invariant operators as $z^2\to 0$ giving
\begin{align}\label{eq:ope-tq}
\tilde{O}_{\gamma^z}(z,\mu) = & \sum_{n=0}^\infty \left[C_n ({\mu}^2 z^2)\frac{(-iz)^n}{n!} e_{\mu_1}\cdots e_{\mu_n}O_1^{\mu_0\mu_1\cdots\mu_n}(\mu)\right. \nonumber\\
&+C'_n ({\mu}^2 z^2)\frac{(-iz)^n}{n!} e_{\mu_1}\cdots e_{\mu_n}O_2^{\mu_0\mu_1\cdots\mu_n}(\mu)\nonumber\\
& + \text{higher-twist operators}\Big]\,,
\end{align}
where $\mu_0=z$, $C_n=1+O(\alpha_s)$ and $C'_n=O(\alpha_s)$ are Wilson coefficients, and $O_1^{\mu_0\mu_1\cdots\mu_n}(\mu)$ and $O_2^{\mu_0\mu_1\cdots\mu_n}(\mu)$ are the only allowed renormalized traceless symmetric twist-2 quark and gluon operators at leading power in the OPE,
\begin{align}\label{eq:twist-2}
O_1^{{\mu}_0 {\mu}_1 \ldots {\mu}_n}(\mu) = & Z_{n+1}^{qq} \bigl(
\bar{\psi} \gamma^{({\mu}_0 } iD^{{\mu}_1} \cdots iD^{ {\mu}_n)} \psi -
\text{trace} \bigr) \,,
\\
O_2^{{\mu}_0 {\mu}_1 \ldots {\mu}_n}(\mu) = & Z_{n+1}^{qg}\bigl(
F^{(\mu_0\rho} iD^{{\mu}_1} \cdots iD^{ {\mu}_{n-1}} F_\rho^{~\mu_n)} - \text{trace} \bigr)
.\nonumber
\end{align}
Here $Z_{n+1}^{ij}=Z_{n+1}^{ij}(\mu,\epsilon)$ are multiplicative $\overline{\rm MS}$ renormalization factors and $(\mu_0 \cdots \mu_n)$ stands for the symmetrization of these Lorentz indices.
The above OPE is valid for the operator itself, where we implicitly constrain ourselves to the subspace of matrix elements for which the twist expansion is appropriate. In the iso-vector case, the mixing with the gluon operators is absent, which we will stick to for the rest of the paper. When $O_1^{\mu_0\mu_1\cdots\mu_n}$ is evaluated in the nucleon state,
\begin{eqnarray} \label{eq:def-moments}
\langle P | O_1^{{\mu}_0 {\mu}_1 \cdots {\mu}_n} | P\rangle = 2a_{n + 1}(\mu)\left(P^{{\mu}_0} P^{{\mu}_1} \ldots P^{{\mu}_n} - \text{trace}\right), \nonumber\\
\end{eqnarray}
where $a_{n+1}(\mu)$ is the $(n+1)$-th moment of the PDF,
\begin{eqnarray} \label{eq:moment}
a_{n + 1} \left(\mu\right)=\int_{-1}^1 dx\,x^n q \left(x,\mu\right)\,,
\end{eqnarray}
and the explicit expression of the trace term have been derived in Ref.~\cite{Nachtmann:1973mr,Georgi:1976ve,Chen:2016utp}. The inverse relation to \eq{moment} is that $q(x,\mu)$ has an expansion with terms proportional to the n'th derivative of the $\delta$-function, as in $\delta^{(n)}(x)\, a_n(\mu)$, without any nontrivial short distance Wilson coefficient.
As pointed out in Ref.~\cite{Ji:2017rah}, to obtain enough information for the spatial correlator at $|\zeta|= |P^z z|\sim 1$ at small $z^2$, we have to choose $P^z$ large compared to the scale $\Lambda_{\rm QCD}$. When $P_z^2\gg\{\Lambda_{\rm QCD}^2, M^2\}$, the trace terms in Eq.(\ref{eq:def-moments}) are suppressed by powers of $M^2/P_z^2$, while the contributions from higher-twist operators in Eq.~(\ref{eq:ope-tq}) are suppressed by powers of $\Lambda_{\rm QCD}^2/P_z^2$ or $z^2\Lambda_{\rm QCD}^2$. Therefore, the twist-2 contribution is the leading approximation of the nucleon matrix element $\langle P|\tilde{O}_{\gamma^z}(z)|P\rangle$ at large momentum. From now on we will drop all the power corrections from our discussion.
The Wilson coefficients $C_n ({\mu}^2 z^2)$ in the OPE of $\tilde{O}_{\gamma^z}(z)$ can be computed in perturbation theory for $\mu\sim 1/|z|\gg \Lambda_{\rm QCD}$. In
the $\overline{\ensuremath{\operatorname{MS}}}$ scheme, the $C_n$ are log-singular near $z^2 =
0$, and so is $\langle P|\tilde{O}_{\gamma^z}(z,\mu)|P\rangle$. For this reason the $x$-moments of the quasi-PDF $\tilde q(x,P^z,\mu)$ are proportional to $C_n|_{z=0}$ which is
divergent, and the quasi-PDF will not simply become the PDF in the infinite $P^z$ limit. Instead, we need a factorization formula which matches the quasi-PDF to the PDF. In contrast, for the pseudo-PDF the moments do exist since we hold $\mu^2 z^2$ fixed when taking the $x$-moment. However, we still need a factorization formula to match the pseudo-PDF to the PDF. We will comment further about this below.
Based on Eqs.~(\ref{eq:ope-tq}-\ref{eq:moment}), we can write down the leading-twist approximation to the spatial correlator as
\begin{align} \label{eq:fac-tQ-orig}
\tilde{Q}_{\gamma^z} \bigl(\zeta, {\mu}^2 z^2\bigr)
=& \sum_{n} C_n ({\mu}^2 z^2) \frac{(-i \zeta)^n}{n!} a_{n + 1}
\left(\mu\right)
\\
= & \sum_{n} C_n (\mu^2z^2)
\frac{(-i \zeta)^n}{n!} \int_{-1}^1 dy\,y^n q \left( y, \mu\right)
\,. \nonumber
\end{align}
It should be noted that the only approximation we have made so far is ignoring
the higher-twist effects that are suppressed by small $z^2$ and the large momentum $P^z$ of the nucleon. In the limit of $P_z^2\gg M^2,\Lambda_{\rm QCD}^2$, we have $P^0\sim P^z$, so even if $\mu_0=0$ in Eq.~(\ref{eq:ope-tq}), the leading approximation of $\tilde{O}_{\gamma^0}(z)$ is still given by the twist-2 contributions in Eq.~(\ref{eq:fac-tQ-orig}), just with modified coefficients $C_n$.
Based on the OPE results in Eq.~(\ref{eq:fac-tQ-orig}), we can derive a factorization formula for the Euclidean spatial correlator\xspace. First of all, let us
define a function $\mathcal{C} (\alpha, {\mu}^2 z^2)$:
\begin{align} \label{eq:fac-C}
\mathcal{C} (\alpha, {\mu}^2 z^2) \equiv \int\! \frac{d \zeta}{2 \pi} \,
e^{i \alpha \zeta} \sum_n C_n ({\mu}^2 z^2) \frac{(-i \zeta)^n}{n!}
\,.
\end{align}
From \eq{fac-tQ-orig} and the renormalized analog of the Fourier-transform relation in \eq{pseudo} $\mathcal{C}(\alpha,\mu^2z^2)$ corresponds to a pseudo-PDF in the special case where $a_{n+1}(\mu)=1$. The analysis of Refs.~\cite{Radyushkin:1983wh,Radyushkin:2016hsy} implies that the support of $\mathcal{C}(\alpha,\mu^2z^2)$ is $-1 \le \alpha \le 1$. Noting that
\begin{align}
\int d\alpha\, e^{-i\alpha(y \zeta)}\, \mathcal{C} (\alpha, {\mu}^2 z^2)
&= \sum_n C_n(\mu^2 z^2) \frac{(-i\zeta y)^n}{n!} \,,
\end{align}
we find from \eq{fac-tQ-orig} that
\begin{align}
\tilde{Q}(\zeta,\mu^2 z^2)
= &\int_{-1}^1\!\! dy\int_{-1}^1\!\! d\alpha\ e^{-i\alpha(y\zeta)}\, \mathcal{C}(\alpha,\mu^2z^2)\, q(y,\mu) \,.
\end{align}
Finally, using the inverse transform of the renormalized analog of \eq{Qdefn},
\begin{align} \label{eq:FTq}
Q(\zeta,\mu) &= \int_{-1}^1\!\! dy\: e^{-i y\zeta}\,
q(y,\mu) \,.
\end{align}
we obtain
\begin{align} \label{eq:io-Q-fact}
\tilde{Q}(\zeta,\mu^2 z^2)
=& \int_{-1}^1 d\alpha\ \mathcal{C}(\alpha,\mu^2z^2)\, Q(\alpha \zeta,\mu)
+ {\cal O}(z^2\Lambda_{\rm QCD}^2)
\,.
\end{align}
The result in \eq{io-Q-fact} is the factorization formula for the lattice calculable spatial correlator\xspace $\tilde Q(\zeta,\mu^2 z^2)$ which expresses it in terms of the light-cone correlation $Q(\zeta,\mu)$ that defines the PDF. It has the same structure as the factorization formula for the spatial correlator\xspace used for the calculation of the pion distribution amplitude in Ref.~\cite{Braun:2007wv,Bali:2017gfr}.
\subsection{Factorization for the quasi-PDF}
The renormalized quasi-PDF is defined as a Fourier transform of the renormalized spatial correlator\xspace,
\begin{align} \label{eq:fac-tq-orig0}
\tilde{q} \Bigl( x, \frac{\mu}{P_z}\Bigr)
\equiv &\int \frac{d \zeta}{2 \pi}\: e^{ix \zeta}\: \tilde{Q} \biggl(\zeta, \frac{{\mu}^2\zeta^2}{P_z^2}\biggr) \,.
\end{align}
Note that we could use either $\tilde{Q}_{\gamma^z}$ or $\tilde{Q}_{\gamma^0}$ here.
Using the result for the spatial correlator\xspace in Eq.~(\ref{eq:fac-tQ-orig}) this gives
\begin{align} \label{eq:fac-tq-orig}
\tilde{q} \Bigl( & x, \frac{\mu}{P_z}\Bigr)
\\
= & \int_{-1}^1 dy \biggl[ \int \frac{d \zeta}{2 \pi} e^{ix \zeta}
\sum_{n=0} C_n \Bigl( \frac{{\mu}^2\zeta^2}{P_z^2}\Bigr) \frac{(-i
\zeta)^n}{n!} y^n \biggr] q \left( y, \mu \right)
\nonumber\\
= & \int_{-1}^1 {dy\over |y|} \biggl[\int \frac{d \zeta}{2 \pi} e^{i
\frac{x}{y} \zeta} \sum_{n=0} C_n \Bigl( \frac{{\mu}^2\zeta^2 }{(yP^z)^2}\Bigr)\, \frac{(-i \zeta)^n}{n!} \biggr] q \left( y,\mu \right) \,.
\nonumber
\end{align}
Already, one can see that the matching kernel is a function of $x / y$
and $\mu / (|y|P^z)$. We define the kernel as
\begin{align} \label{eq:fac-c}
C \left( {x\over y}, \frac{\mu}{|y|P^z} \right)
\equiv \int\! \frac{d \zeta}{2 \pi} \: e^{i {x\over y} \zeta}
\sum_{n=0} C_n \Bigl(
\frac{{\mu}^2\zeta^2}{(yP^z)^2}\Bigr) \, \frac{(-i \zeta)^n}{n!}
\,,
\end{align}
and then Eq.~(\ref{eq:fac-tq-orig}) can be rewritten as
\begin{align} \label{eq:fac-tq}
\tilde{q} \left( x, \frac{\mu}{P^z}\right)
= \int_{-1}^1 \frac{dy}{|y|}\: C \Bigl(
\frac{x}{y}, \frac{\mu}{|y|P^z} \Bigr)\: q \left( y,\mu\right)
\,,
\end{align}
which is the $\overline{\rm MS}$ factorization formula for the quasi-PDF. This result shows that the factorization formula in Eq.~(\ref{eq:momfact}) must have $p^z=|y|P^z$ for the quasi-PDF in the $\overline{\rm MS}$ scheme. We will show that this remains true for any quasi-PDF renormalization scheme in \sec{ren}. This differs
from the choice $p^z=P^z$ which had been conjectured and used in the early papers on the quasi-PDF~\cite{Ji:2013dva,Ji:2014gla,Xiong:2013bka}. Physically the correct result in \eq{fac-tq} can be understood as the fact that the matching coefficient is only sensitive to the perturbative partonic dynamics, and hence it is the magnitude of the partonic momentum $|y| P^z$ which appears, rather than the hadronic momentum $P^z$.
Taking the moment of the quasi-PDF using \eq{fac-tq-orig0} gives
\begin{align} \label{eq:momentqpdf}
&\int_0^1\!\! dx\: x^n \tilde{q} \left( x, \frac{\mu}{P^z}\right)
= \Big(i\frac{d}{d\zeta}\Big)^n \tilde{Q} \biggl(\zeta, \frac{{\mu}^2\zeta^2}{P_z^2}\biggr) \Bigg|_{\zeta\to 0}
\nonumber\\
& \quad = \sum_{n'} \Big(i\frac{d}{d\zeta}\Big)^n \bigg[ C_{n'} \Big(\frac{{\mu}^2 \zeta^2}{P_z^2}\Big) \frac{(-i \zeta)^{n'}}{n'!} \bigg] \Bigg|_{\zeta\to 0} a_{n' + 1}
\left(\mu\right) \,.
\end{align}
Since the $C_{n'}$ coefficients have $\ln(\zeta^2)$ dependence, the derivative for $n'=n$ will always have a logarithmic singularity as $\zeta\to 0$, and there will be even more singular terms for $n'<n$. This explains why the short distance Wilson coefficient causes the moments not to exist for the quasi-PDF.
\subsection{Factorization for the pseudo-PDF}
The renormalized pseudo-PDF is the Fourier transform of the renormalized spatial correlator\xspace
\begin{align} \label{eq:pseudoRen}
\mathcal{P} \left( x, \mu^2 z^2\right)
= \int_{-\infty}^\infty \! \frac{d \zeta}{2\pi}\:
e^{ix \zeta}\: \tilde{Q}\bigl( \zeta, \mu^2 z^2 \bigr)
\,.
\end{align}
Since both the pseudo-PDF and spatial correlator\xspace are multiplicatively renormalized in a $\zeta$-independent manner, this follows immediately from \eq{pseudo}.
If we take \eq{io-Q-fact} and Fourier transform the spatial correlator\xspace $\tilde Q(\zeta,\mu^2z^2)$ into the pseudo-PDF, and light-cone correlation $Q(\alpha\zeta,\mu)$ into the PDF, then we immediately obtain the factorization formula for the pseudo-PDF,
\begin{align}\label{eq:ps-q-fact}
\mathcal{P}(x,z^2\mu^2) =
&\int_{|x|}^1 {dy\over |y|}\
\mathcal{C}\left({x\over y},\mu^2 z^2\right) q(y,\mu)\nonumber\\
&+\int^{-|x|}_{-1} {dy\over |y|}\
\mathcal{C}\left( \frac{x}{y}, \mu^2z^2\right) q(y,\mu)
\nonumber \\
&+ {\cal O}(z^2\Lambda_{\rm QCD}^2)\,,
\end{align}
which is the small $z^2$ factorization formula in Eq.~(\ref{eq:pseudofact}). The upper and lower limits of the integrals in \eq{ps-q-fact} follow immediately from the support $-1\le \alpha \le 1$ of the matching coefficient $\mathcal{C}(\alpha,z^2\mu^2)$, and we recall that we also have $-1\le x\le 1$ for the pseudo-PDF on the LHS.
Since the range of $x$ is bounded for the pseudo-PDF the terms in the series expansion of the exponential in \eq{pseudoRen} exist,
\begin{align}
\tilde Q(\zeta,\mu^2 z^2)
= \sum_{n=0}^\infty \int_{-1}^{1}\!\! dx\: \frac{(-i\zeta)^n x^n}{n!}
{\cal P}(x,\mu^2 z^2) \,.
\end{align}
Comparing with \eq{fac-tQ-orig} this implies that the moments of the pseudo-PDF are given by
\begin{align} \label{eq:Pmoment}
\int_{-1}^{1}\!\! dx\: x^n \: {\cal P}(x,\mu^2 z^2)
&= C_n(\mu^2 z^2)\, a_{n+1}(\mu) \,.
\end{align}
So far we have proven the large $P^z$ factorization of the quasi-PDF and small $z^2$ factorization of the spatial correlation and pseudo-PDFs. After
deriving one factorization, it immediately leads to the others, since they are just different representations of the same spatial correlator. Indeed, we see that the quasi-PDF and pseudo-PDF are related at leading power by their definitions:
\begin{eqnarray} \label{eq:equiv}
\tilde{q}\Bigl(x,{\mu\over P^z}\Bigr)
= \int\! {d\zeta\over 2\pi}\ e^{ix\zeta}\int_{-1}^1\!\! dy \ e^{-iy\zeta}
\: \mathcal{P}\biggl(y,{\mu^2 \zeta^2\over P_z^2}\biggr)\,,\nonumber\\
\end{eqnarray}
where we have used $z=\zeta/P^z$.
Based on Eq.~(\ref{eq:fac-C}) and Eq.~(\ref{eq:fac-c}), the Wilson coefficients in their factorization theorems also maintain the same relationship,
\begin{align} \label{eq:ftc-C}
C\left(\xi,{\mu\over |y|P^z}\right)
&= \int\! {d\zeta\over 2\pi}\ e^{i\xi\zeta}\int_{-1}^1\!\! d\alpha\: e^{-i\alpha\zeta}\: \mathcal{C}\left(\alpha,{\mu^2 \zeta^2\over (yP^z)^2}\right) .
\end{align}
For the relations in \eqs{equiv}{ftc-C} the same choice of $\Gamma=\gamma^0$ or $\gamma^z$ should be used in the quasi- and pseudo-PDFs, or their corresponding coefficients.
In summary, there is a unique factorization formula that matches the quasi-PDF, spatial correlator\xspace and pseudo-PDF to the PDF. Since their factorizations into the PDF all require small distances and have large nucleon momentum, the setup for their lattice calculations must also be the same. Therefore, the LaMET and pseudo distribution approaches are in principle equivalent to each other, and they differ perhaps only by effects related to their implementation on the lattice.
In Ref.~\cite{Radyushkin:2017cyf} it was speculated that one can study a ratio function
\begin{eqnarray} \label{eq:ratio}
\tilde{Q}(\zeta,z^2,a^{-1})/\tilde{Q}(0,z^2,a^{-1})
\end{eqnarray}
on a lattice with spacing $a$, and the $O(z^2)$ corrections may cancel approximately. This idea was tested in Ref.~\cite{Orginos:2017kos} in lattice QCD, and the results show that the ratio evolves slowly in $z^2$ at small values. It is then interesting to consider what type of non-perturbative information can be extracted from this ratio.
This question can be answered using the small $z^2$ factorization for the spatial correlator\xspace.
According to Eq.~(\ref{eq:fac-tQ-orig}),
\begin{eqnarray}
\tilde{Q}(0,\mu^2 z^2) = C_0(\mu^2z^2) + O(z^2\Lambda_{\rm QCD}^2)\,,
\end{eqnarray}
where in the $\overline{\rm MS}$ scheme to one-loop
\begin{eqnarray} \label{eq:C0}
C_0(\mu^2z^2) = 1 + {\alpha_sC_F\over 2\pi} \biggl[ {3\over2} \ln(\mu^2z^2 e^{2\gamma_E}/4)+{5\over2} \biggr] \,,
\end{eqnarray}
which was also derived recently in \cite{Zhang:2018ggy}.
Then the ratio becomes
\begin{align} \label{eq:Qtratio}
\frac{\tilde{Q}\left(\zeta, {\mu}^2 z^2\right)}{\tilde{Q}\left(0, {\mu}^2 z^2\right)}
=&\sum_n \frac{C_n (\mu^2z^2)}{C_0 (\mu^2z^2)} \frac{(-i \zeta)^n}{n!} a_{n+1}(\mu)\nonumber\\
&+ O(z^2\Lambda_{\rm QCD}^2)\,.
\end{align}
Using \eq{Pmoment} and our $\overline{\rm MS}$ one-loop perturbative pseudo-PDF result in \eq{1looppseudopdfb} below we find for $\Gamma=\gamma^0$ that
\begin{align}
C_n(\mu^2 z^2) &= 1 +\frac{\alpha_s C_F}{2\pi} \biggl[ \Bigl(
\frac{3\!+\!2n}{2\!+\!3n\!+\!n^2}
\!+\! 2H_n\Bigr) \ln\frac{\mu^2z^2 e^{2\gamma_E}}{4}
\nonumber\\
&\quad\quad
+ \frac{5\!+\!2n}{2\!+\!3n\!+\!n^2}+2(1-H_n)H_n - 2 H_n^{(2)}
\biggr] ,
\end{align}
where the Harmonic numbers are $H_n=\sum_{i=1}^n 1/i$ and $H_n^{(2)} =\sum_{i=1}^n 1/i^2$. For the case $\Gamma=\gamma^z$ we have $C_n^{\gamma^z}(\mu^2 z^2) = C_n(\mu^2 z^2) + \Delta C_n^{\gamma^z}(\mu^2 z^2)$ with
\begin{align}
\Delta C_n^{\gamma^z}(\mu^2 z^2)
&= \frac{\alpha_s C_F}{2\pi} \: \frac{2}{2\!+\!3n\!+\!n^2} \,,
\end{align}
which also modifies \eq{C0} for $n=0$.
At small $z^2$ where the perturbative expansion with $\mu\simeq 1/|z|$ is valid, the ratio in \eq{Qtratio} has a weak logarithmic dependence on $|z|$, which is consistent with the lattice findings in Refs.~\cite{Orginos:2017kos,Karpie:2017bzm}. The weak dependence on $|z|$ can be quantitatively described by an evolution equation in $\ln z^2$~\cite{Radyushkin:2017cyf,Radyushkin:2017lvu}. According to our OPE analysis, here $\tilde{Q}(0,\mu^2 z^2)$ only serves as an overall normalization factor which is contaminated by higher-twist corrections, and the $\ln z^2$ evolution can be put in accurate terms with the factorization formula in Eq.~(\ref{eq:io-Q-fact}) that enables us extract the PDF from the ratio function. The same point was demonstrated by work done very recently in~\cite{Zhang:2018ggy}, which appeared simultaneously with our paper.
\section{Equivalence at one-loop order}
\label{sec:equiv}
As has been proven in Sec.~\ref{sec:ope}, the quasi-PDF and pseudo-PDF as well as their matching coefficients are related by a simple Fourier transform in Eq.~(\ref{eq:equiv}) and Eq.~(\ref{eq:ftc-C}). This relation is valid to all orders in perturbation theory. In this section we check the relations in Eq.~(\ref{eq:equiv}) and Eq.~(\ref{eq:ftc-C}) at one-loop order. We choose $\Gamma=\gamma^0$ for our main presentation, but also quote final results for the case $\Gamma=\gamma^z$.
In the Feynman gauge, we calculate the quark matrix elements of the unpolarized iso-vector quasi-PDF, pseudo-PDF, and light-cone PDF at one-loop order in dimensional
regularization with $d=4-2\epsilon$. The external quark state is chosen to be on-shell and massless, and we regularize the UV and collinear divergences by $1/\epsilon_{\mbox{\tiny UV}}\ (\epsilon_{\mbox{\tiny UV}}>0)$ and $1/\epsilon_{\mbox{\tiny IR}}\ (\epsilon_{\mbox{\tiny IR}}<0)$ respectively. The one-loop order Feynman diagrams are shown in Fig.~\ref{fig:diagram}.
\begin{widetext}
\begin{figure*}[t!]
\centering
\includegraphics[width=.9\textwidth]{quasi_diagram.pdf}
\caption{One-loop Feynman diagrams for the quasi-PDF, spatial correlator\xspace and pseudo-PDFs. The first one is named ``vertex", the second and third ones are named ``sail", and the last one ``tadpole". The standard quark self energy wavefunction is also included.}
\label{fig:diagram}
\end{figure*}
In an on-shell quark state with momentum $p^\mu=(p^0=p^z,0,0,p^z)$, for $\Gamma=\gamma^0$, each diagram gives
\begin{align} \label{eq:qtloopint1}
\tilde{Q}^{(1)}_\text{vertex}(\zeta,z^2,\epsilon)
&= {\mu^{2\epsilon} \iota^\epsilon \over 2p^0}\bar{u}(p)\int {d^d k\over (2\pi)^d}(-ig T^a\gamma^\mu) {i\over \slashed k} \gamma^0 {i\over\slashed k} (-i gT^a\gamma^\nu){ - i g_{\mu\nu} \over (p-k)^2}u(p)e^{-i k^z z}\ ,
\\
\tilde{Q}^{(1)}_\text{sail}(\zeta,z^2,\epsilon)
&= {\mu^{2\epsilon} \iota^\epsilon\over 2p^0}\bar{u}(p)\int {d^d k\over (2\pi)^d} (ig T^a \gamma^0) {1\over i(p^z-k^z)} \left( e^{-ip^z z}- e^{-ik^z z}\right) \delta^{\mu z} {i\over \slashed k} (-ig T^a \gamma^\nu) {-ig_{\mu\nu}\over (p-k)^2}u(p)
\nonumber\\
&+ {\mu^{2\epsilon}\iota^\epsilon \over 2p^0}\bar{u}(p) \int {d^d k\over (2\pi)^d} (-ig T^a \gamma^\nu){i\over \slashed k} (ig T^a \gamma^0) {1\over i(p^z-k^z)} \left( e^{-ip^z z}- e^{-ik^z z}\right) \delta^{\mu z} {-ig_{\mu\nu}\over (p-k)^2}u(p)\ ,
\nonumber \\
\tilde{Q}^{(1)}_\text{tadpole}(\zeta,z^2,\epsilon)
&= {\mu^{2\epsilon}\iota^\epsilon \over 2p^0}\bar{u}(p)\int {d^d k\over (2\pi)^d} (-g^2)C_F \gamma^0 \delta^{\mu z}\delta^{\nu z}\left({e^{-ip^z z}- e^{-ik^z z}\over (p^z-k^z)^2} - {ze^{-ip^z z}\over i(p^z-k^z)}\right){-ig_{\mu\nu}\over (p-k)^2}u(p)\ ,
\nonumber
\end{align}
where $\iota = e^{\gamma_E}/(4\pi)$ is included to implement $\mu$ in the $\overline{\rm MS}$ scheme, $C_F=4/3$, and $T^a$ is the SU(3) color matrix in the fundamental representation. The second term in the brackets in the last line, which is proportional $z$, does not contribute to the loop integral as it is odd under the exchange of $p^z-k^z\to -(p^z-k^z)$. The quark self-energy correction is $\tilde{Q}_{\rm w.fn.}^{(1)}(\zeta,z^2,\epsilon) =\delta Z_\psi\: \tilde{Q}^{(0)}(\zeta,z^2)$ with the tree level matrix element $\tilde{Q}^{(0)}(\zeta,z^2)=e^{-i\zeta}$ and on-shell renormalization constant $\delta Z_\psi$,
\begin{eqnarray}
\delta Z_\psi = {\alpha_sC_F\over 2\pi}\left(-{1\over2}\right)\left({1\over\epsilon_{\mbox{\tiny UV}}} - {1\over \epsilon_{\mbox{\tiny IR}}}\right)\ .
\end{eqnarray}
After carrying out the loop integrals in Eq.~(\ref{eq:qtloopint1}) according to the method in Ref.~\cite{Ji:2017rah}, we obtain
\begin{align}\label{eq:1loopdiagram}
\tilde{Q}^{(1)}_{\rm vertex}(\zeta,z^2,\epsilon)
=\,& {\alpha_sC_F\over 2\pi}e^{\epsilon\gamma_E} \int_0^1 du\ (1-\epsilon)(1-u) e^{-iu \zeta}\Gamma(-\epsilon) 4^{-\epsilon} \big(\mu |z|\big)^{2\epsilon}
\\
=\,& {\alpha_sC_F\over 2\pi} e^{\epsilon\gamma_E}\biggl({\mu|z|\over2}\biggr)^{2\epsilon}
\frac{(-1)\Gamma(2-\epsilon)}{\epsilon_{\mbox{\tiny IR}}}
{1-i\zeta - e^{-i\zeta}\over \zeta^2}
\,, \nonumber\\
\tilde{Q}^{(1)}_{\rm sail}(\zeta,z^2,\epsilon)
=\, & {\alpha_sC_F\over 2\pi} e^{\epsilon\gamma_E}\left[(i\zeta) \int_0^1 du \int_0^1 dt (2-u) e^{-i(1-ut)\zeta} \Gamma(-\epsilon) 4^{-\epsilon}(t^2z^2\mu^2)^{\epsilon}\right.
\nonumber\\
&\left. -\int_0^1 du \ e^{-iu \zeta}\Gamma(-\epsilon) 4^{-\epsilon}(z^2\mu^2)^{\epsilon} + \left({1\over\epsilon_{\mbox{\tiny UV}}} - {1\over \epsilon_{\mbox{\tiny IR}}}\right) e^{-i\zeta}\right]
\nonumber\\
=\,& {\alpha_sC_F\over 2\pi} e^{\epsilon\gamma_E} \left({\mu|z|\over2}\right)^{2\epsilon}
\left\{\frac{-\Gamma(1-\epsilon)}{\epsilon_{\mbox{\tiny IR}}}
{2(1-\epsilon)\over 1-2\epsilon} \left[{1-e^{-i\zeta}\over -i\zeta} + e^{-i\zeta}(-i\zeta)^{-2\epsilon}\big(\Gamma(2\epsilon)-\Gamma(2\epsilon,-i\zeta)\big)
\right]\right.
\nonumber\\
&\left. + {\Gamma(1-\epsilon)\over \epsilon_{\mbox{\tiny IR}} } \frac{e^{-i\zeta}}{\epsilon} + \left({1\over\epsilon_{\mbox{\tiny UV}}} - {1\over \epsilon_{\mbox{\tiny IR}}}\right) e^{-i\zeta} \right\}
\,, \nonumber\\
\tilde{Q}^{(1)}_{\rm tadpole}(\zeta,z^2,\epsilon)
=\, & {\alpha_sC_F\over 2\pi} e^{\epsilon\gamma_E}\left({\mu|z|\over2}\right)^{2\epsilon} {\Gamma(1-\epsilon)\over \epsilon_{\mbox{\tiny UV}} (1-2\epsilon)}e^{-i\zeta}
\ . \nonumber
\end{align}
For simplicity we have left out the tree level multiplicative spinor factor $\bar u\gamma^0 u$ when quoting one-loop results in \eq{1loopdiagram}, and will continue to do so for the spatial correlator\xspace, quasi-PDF, and pseudo-PDF results quoted below.
Since $\alpha_s^{\rm bare}=\alpha_s(\mu) \mu^{2\epsilon} Z_g^2 = \alpha_s(\mu) \mu^{2\epsilon} + {\cal O}(\alpha_s^2)$ is $\mu$-independent we do not include $\mu$ as an argument in the bare functions.
In the final result for each term we have also specified whether $1/\epsilon$ factors (that remain after expanding about $\epsilon\to 0$) are IR or UV divergences. Combined with the wavefunction corection, the bare spatial correlator\xspace $\tilde{Q}^{(1)}(\zeta,z^2,\epsilon)$ is
\begin{align} \label{eq:1loopioffe}
\tilde{Q}^{(1)}(\zeta,z^2,\epsilon)
&= {\alpha_sC_F\over 2\pi}e^{\epsilon\gamma_E} \left({\mu|z|\over2}\right)^{2\epsilon} \Bigg\{
{3\over2}\left({1\over \epsilon_{\mbox{\tiny UV}}}-{1\over \epsilon_{\mbox{\tiny IR}}}\right)e^{-i\zeta}
+ \frac{(-1)\Gamma(2-\epsilon)}{\epsilon_{\mbox{\tiny IR}}}
\left[ {(1-i\zeta - e^{-i\zeta})\over \zeta^2}
\right.
\\
& \qquad\qquad\qquad\qquad\qquad\quad \left.
+{2\over 1-2\epsilon}\left\{ {1-e^{-i\zeta}\over -i\zeta} + e^{-i\zeta}(-i\zeta)^{-2\epsilon}
\big(\Gamma(2\epsilon)-\Gamma(2\epsilon,-i\zeta)\big) - \frac{e^{-i\zeta}}{2\epsilon} \right\} \right]
\Bigg\}\ . \nonumber
\end{align}
Note that as $\epsilon\to 0$ the terms in the innermost curly brackets have no $1/\epsilon$ term. Also we can verify that in the local limit of the operator that the bare one-loop correction vanishes as expected by conservation of the vector current:
\begin{align} \label{eq:Qvcc}
\lim_{z\to 0} \tilde{Q}^{(1)}(z P^z,z^2,\epsilon)
= \lim_{z\to 0} {\alpha_sC_F\over 2\pi}e^{\epsilon\gamma_E} \left({\mu|z|\over2}\right)^{2\epsilon} \bigg\{
{3\over2}\left({1\over \epsilon_{\mbox{\tiny UV}}}-{1\over \epsilon_{\mbox{\tiny IR}}}\right)
+ \frac{3}{2\epsilon_{\mbox{\tiny IR}}} \bigg\} = 0
\,,
\end{align}
where we note that it is important that the $1/\epsilon_{\mbox{\tiny IR}}$ terms cancel since the assumption $\epsilon>0$ is only valid for the $1/\epsilon_{\mbox{\tiny UV}}$ term.
The corresponding bare pseudo-PDF is
\begin{align}\label{eq:1looppseudo}
{\cal P}^{(1)}(x,z^2,\epsilon)
&= {\alpha_sC_F\over 2\pi} \left[-1+x + \frac{2}{(1-2\epsilon)} -
\left( {2\over (1-2\epsilon)}{1\over (1-x)^{1-2\epsilon}}\right)^{[0,1]}_{+(1)}\right]e^{\epsilon\gamma_E}\left({\mu|z|\over2}\right)^{2\epsilon} \frac{\Gamma(2-\epsilon)}{\epsilon_{\mbox{\tiny IR}}} \theta(x)\theta(1-x)\nonumber\\
&\quad +{\alpha_sC_F\over 2\pi} \:
e^{\epsilon\gamma_E}\left({\mu|z|\over2}\right)^{2\epsilon}{3\over2}\left({1\over \epsilon_{\mbox{\tiny UV}}}-{1\over \epsilon_{\mbox{\tiny IR}}}\right)\, \delta(1-x)
\ .
\end{align}
Since we will encounter plus functions over different domains below, we define a plus function at $x=x_0$ within a given domain $D$ so that
\begin{eqnarray}
\int_D dx\ \big[ g(x)\big]_{+(x_0)}^D\, h(x) = \int_D dx\ g(x) \left[ h(x) - h(x_0)\right]\,.
\end{eqnarray}
(See \app{ep} for more details.)
It is straightforward to confirm that the bare pseudo-PDF satisfies the local vector current conservation, $\lim_{z\to 0} \int\! dx \, {\cal P}^{(1)}(x,z^2\mu^2,\epsilon) =0$, with the same cancellation as in \eq{Qvcc}.
Now, according to the relations between the quasi-PDF and the spatial correlator\xspace or pseudo-PDFs in Eqs.~(\ref{eq:fac-tq-orig0},\ref{eq:equiv}), we can do a Fourier or double Fourier transform of the results in Eqs.~(\ref{eq:1loopioffe},\ref{eq:1looppseudo}) to get the quasi-PDF. Despite its straightforwardness, the Fourier transform is subtle and the details are provided in \app{ft}. Here we simply quote the result for the bare quasi-PDF,
\begin{align} \label{eq:1loopquasi}
\tilde{q}^{(1)}(x,p^z,\epsilon)
=& {\alpha_sC_F\over 2\pi}\left\{{3\over2}\left({1\over \epsilon_{\mbox{\tiny UV}}}-{1\over \epsilon_{\mbox{\tiny IR}}}\right)\delta(1-x) + {\Gamma(\epsilon+{1\over2})e^{\epsilon\gamma_E}\over \sqrt{\pi}}{\mu^{2\epsilon}\over p_z^{2\epsilon}}{1-\epsilon \over \epsilon_{\mbox{\tiny IR}} (1-2\epsilon)} \right.\nonumber\\
&\times\left.\left[|x|^{-1-2\epsilon}\left(1+x+{x\over2}(x-1+2\epsilon)\right)- |1-x|^{-1-2\epsilon}\left(x+{1\over2}(1-x)^2\right) + I_3(x)\right] \right\}\,,
\end{align}
where
\begin{eqnarray}
I_3(x) = \theta(x-1)\left(x^{-1-2\epsilon}\over x-1\right)^{[1,\infty]}_{+(1)} -\theta(x)\theta(1-x) \left(x^{-1-2\epsilon}\over 1-x\right)^{[0,1]}_{+(1)} - \delta(1-x) \pi\csc(2\pi\epsilon) + \theta(-x) {|x|^{-1-2\epsilon}\over x-1}\,.
\end{eqnarray}
After some algebra one can confirm that the bare quasi-PDF satisfies local vector current conservation, with $\int\! dx\, \tilde{q}^{(1)}(x,p^z,\epsilon) = 0$. To verify this result one must carefully separate out $1/\epsilon_{\mbox{\tiny UV}}$ factors arising from requiring $\epsilon>0$ to obtain convergence at $x=\pm \infty$, and $1/\epsilon_{\mbox{\tiny IR}}$ factors that arise from requiring $\epsilon<0$ to obtain convergence at $x=1$.
An alternate method of obtaining the quasi-PDF is to directly calculate it from the Feynman diagrams by first Fourier transforming $z$ into $xp^z$. As a result, the factors $(e^{-ip^z z}- e^{-ik^z z})$ are transformed into $[\delta(p^z-xp^z)-\delta(k^z-xp^z)]$, and all the loop integrals reduce to $(d-1)$-dimensional ones. This is the procedure for the matching calculations of the quasi-PDF used in Refs.~\cite{Xiong:2013bka,Stewart:2017tvs,Wang:2017qyg}, and is distinct from doing the Fourier transformation after fully carrying out the integrals as in Eqs.~(\ref{eq:1loopdiagram}--\ref{eq:1loopquasi}). As a cross-check we have confirmed in \app{quasi} that we obtain the exact same bare quasi-PDF in \eq{1loopquasi} from both procedures.
Now we consider the $\epsilon$ expansion to obtain $\overline{\rm MS}$ renormalized results for the spatial correlator\xspace, pseudo-PDF, and quasi-PDF. Expanding the spatial correlator\xspace in $\epsilon$ we obtain
\begin{align}
\tilde{Q}^{(1)}(\zeta,z^2,\epsilon)=\delta \tilde{Q}^{(1)}(\zeta,z^2,\mu,\epsilon_{\mbox{\tiny UV}})+\tilde{Q}^{(1)}(\zeta,z^2,\mu,\epsilon_{\mbox{\tiny IR}})+{\cal O}(\epsilon)
\end{align}
with the $\overline{\rm MS}$ counterterm and renormalized spatial correlator\xspace given by
\begin{align} \label{eq:1loopiofferen}
\delta \tilde{Q}^{(1)}(\zeta,z^2,\mu,\epsilon_{\mbox{\tiny UV}})
&= {\alpha_sC_F\over 2\pi}
e^{-i\zeta} {3\over2} {1\over \epsilon_{\mbox{\tiny UV}}}
\,, \\
\tilde{Q}^{(1)}(\zeta,z^2,\mu,\epsilon_{\mbox{\tiny IR}})
&= {\alpha_sC_F\over 2\pi} \Bigg\{
\frac{3}{2} \Big(\ln{\mu^2 z^2 e^{2\gamma_E} \over 4} +1\Big) e^{-i\zeta}
+ \Bigl( -\frac{1}{ \epsilon_{\mbox{\tiny IR}} } - \ln\frac{z^2\mu^2e^{2\gamma_E}}{4} - 1 \Bigr) h(\zeta)
+ \frac{2(1 \!-\!i\zeta\!-\! e^{-i\zeta})}{\zeta^2}
\nonumber\\
&\qquad\qquad\quad
+ 4 i \zeta e^{-i\zeta}\, {}_3F_3(1,1,1,2,2,2,i\zeta)
\Bigg\}
\,. \nonumber
\end{align}
Here ${}_3F_3$ is a hypergeometric function and the Fourier transform of $\big[(1+x^2)/(1-x)\big]_{+(1)}^{[0,1]}$ gives the function
\begin{align}
h(\zeta) &=\frac{3}{2}\, e^{-i\zeta}
+ \frac{1+i\zeta-e^{-i\zeta}-2i\zeta e^{-i\zeta}}{\zeta^2}
- 2 e^{-i\zeta} \big[ \Gamma(0,-i\zeta)\!+\!\gamma_E\!+\!\ln(-i\zeta) \big]
\,.
\end{align}
For the position space PDF we have
\begin{align}
Q^{(1)}(\zeta,\mu,\epsilon_{\mbox{\tiny IR}})
&= - {\alpha_sC_F\over 2\pi} \frac{1}{ \epsilon_{\mbox{\tiny IR}} } h(\zeta) \,.
\end{align}
Next we expand the bare pseudo-PDF from Eq.~(\ref{eq:1looppseudo}) in $\epsilon$ to obtain the $\overline{\rm MS}$ counterterm and renormalized pseudo-PDF as ${\cal P}^{(1)} (x,z^2,\epsilon)= \delta {\cal P}^{(1)} (x,z^2\mu^2,\epsilon_{\mbox{\tiny UV}})+{\cal P}^{(1)} (x,z^2\mu^2,\epsilon_{\mbox{\tiny IR}})+{\cal O}(\epsilon)$ with
\begin{align}\label{eq:1looppseudopdfb}
\delta {\cal P}^{(1)} &(x,z^2\mu^2,\epsilon_{\mbox{\tiny UV}})
= {\alpha_sC_F\over 2\pi}\:
{3\over 2\epsilon_{\mbox{\tiny UV}}} \: \delta(1-x)
\,, \nonumber\\
{\cal P}^{(1)} &(x,z^2\mu^2,\epsilon_{\mbox{\tiny IR}})
\nonumber\\
=&{\alpha_sC_F\over 2\pi}\left\{\left({1+x^2\over 1-x}\right)^{[0,1]}_{+(1)}\left[- \left({1\over \epsilon_{\mbox{\tiny IR}}}+\ln{e^{2\gamma_E}\over 4}\right)-\ln(z^2 \mu^2)-1\right] - \left(4\ln(1-x)\over 1-x\right)^{[0,1]}_{+(1)} +2(1-x)\right\}\theta(x)\theta(1-x)\nonumber\\
& +{\alpha_sC_F\over 2\pi}\left[{3\over2}\ln(z^2\mu^2) + {3\over2}\ln{e^{2\gamma_E}\over 4}+{3\over2}\right]\delta(1-x)
\,.
\end{align}
Note that the renormalized $\overline{\rm MS}$ pseudo-PDF depends explicitly on $\mu^2$, and satisfies the relation to the renormalized $\overline{\rm MS}$ spatial correlator\xspace given in Eq.~(\ref{eq:pseudoRen}).
It is also interesting to note that having expanded in $\epsilon$, local vector current conservation is no longer satisfied by the limit of the renormalized $\overline{\rm MS}$ pseudo-PDF, since
\begin{align} \label{eq:Ppdfconsrv}
\lim_{z\to 0} \int\! dx \, {\cal P}^{(1)} (x,z^2\mu^2,\epsilon_{\mbox{\tiny IR}}) \simeq (3\alpha_s C_F/4\pi) \lim_{z\to 0} \ln(z^2\mu^2)
\end{align}
gives a divergent result. The same divergence is present in the one-loop $\overline{\rm MS}$ renormalized spatial correlator\xspace in \eq{1loopiofferen}. Although this is the case in $\overline{\rm MS}$, it does not need to be the case in other renormalization schemes.
For the quasi-PDF there are two methods that we can consider for the renormalized calculation, either expanding the bare result in \eq{1loopquasi} and renormalizing in $(x,p^z)$ space, or following our preferred definition in \eq{fac-tq-orig0} and Fourier transforming the renormalized spatial correlator\xspace in \eq{1loopiofferen}. Although these two approaches will lead to the same final result for $C$ for practical applications, there is a subtle difference that we will explain.
First consider the renormalization of the quasi-PDF done in $(x,p^z)$ space. Expanding Eq.~(\ref{eq:1loopquasi}) in $\epsilon$, and writing $\tilde{q}^{(1)}(x,p^z,\epsilon) =\delta \tilde{q}^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny UV}})+\tilde{q}^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny IR}}) +{\cal O}(\epsilon)$ allows us to identify the $\overline{\rm MS}$ counterterm and renormalized quasi-PDF as
\begin{align}\label{eq:qPDFren}
\delta \tilde{q}^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny UV}})
=&\, {\alpha_sC_F\over 2\pi} {3\over 2\epsilon_{\mbox{\tiny UV}}}
\left[\delta(1-x) - {1\over2}{1\over x^2}\delta^+\Big({1\over x}\Big) - {1\over2}{1\over (1-x)^2}\delta^+\Big({1\over 1-x}\Big) \right]
\,, \nonumber\\
\tilde{q}^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny IR}})
=&\, {\alpha_sC_F\over 2\pi}\left\{
\begin{array}{ll}
\displaystyle \left({1+x^2\over 1-x}\ln {x\over x-1} + 1 + {3\over 2x}\right)^{[1,\infty]}_{+(1)}- \left({3\over 2x}\right)^{[1,\infty]}_{+(\infty)} &\, x>1\nonumber\\[10pt]
\displaystyle \left({1+x^2\over 1-x}\left[- {1\over \epsilon_{\mbox{\tiny IR}}} - \ln{\mu^2\over 4p_z^2} + \ln\big(x(1-x)\big)\right] - {x(1+x)\over 1-x}\right)^{[0,1]}_{+(1)} &\, 0<x<1\nonumber\\[10pt]
\displaystyle \left(-{1+x^2\over 1-x}\ln {-x\over 1-x} - 1 + {3\over 2(1-x)}\right)^{[-\infty,0]}_{+(1)} - \left({3\over 2(1-x)}\right)^{[-\infty,0]}_{+(-\infty)} \quad &\, x<0
\end{array}\right.\nonumber\\[5pt]
& + {\alpha_sC_F\over 2\pi}\left[\delta(1-x) - {1\over2}{1\over x^2}\delta^+\Big({1\over x}\Big) - {1\over2}{1\over (1-x)^2}\delta^+\Big({1\over 1-x}\Big) \right] \left( {3\over2}\ln{\mu^2\over 4p_z^2} + {5\over2}\right)
\,.
\end{align}
The details of working out the $\epsilon$ expansion of \eq{1loopquasi} are provided in \app{ep}, including definitions of the plus functions and $\delta$-functions at $x_0=\pm \infty$ that appear in the result quoted here. The $\overline{\rm MS}$ quasi-PDF obtained in \eq{qPDFren} still satisfies vector current conservation
\begin{align} \label{eq:qpdfconsrv}
\int\! dx\: \tilde{q}^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny IR}}) =0 \,.
\end{align}
This is obviously the case for the plus function terms which individually integrate to zero, and is also true for the combination of $\delta$-functions which appears in \eq{qPDFren}.
The renormalized $\overline{\rm MS}$ quasi-PDF in Eq.~(\ref{eq:qPDFren}) differs slightly from that obtained using our definition in Eq.~(\ref{eq:fac-tq-orig0}). Using Eq.~(\ref{eq:fac-tq-orig0}) and the renormalized spatial correlator\xspace in \eq{1loopiofferen} we instead obtain
\begin{align}\label{eq:qPDFft}
\delta \tilde{q}'^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny UV}})
=&\, {\alpha_sC_F\over 2\pi} {3\over 2\epsilon_{\mbox{\tiny UV}}}\delta(1-x)\,, \\
\tilde{q}'^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny IR}})
=&\, {\alpha_sC_F\over 2\pi}\left\{
\begin{array}{ll}
\displaystyle \left({1+x^2\over 1-x}\ln {x\over x-1} + 1 + {3\over 2x}\right)^{[1,\infty]}_{+(1)}- \left({3\over 2x}\right)^{[1,\infty]}_{+(\infty)} &\, x>1\nonumber\\[10pt]
\displaystyle \left({1+x^2\over 1-x}\left[- {1\over \epsilon_{\mbox{\tiny IR}}} - \ln{\mu^2\over 4p_z^2} + \ln\big(x(1-x)\big)\right] - {x(1+x)\over 1-x}\right)^{[0,1]}_{+(1)} &\, 0<x<1\nonumber\\[10pt]
\displaystyle \left(-{1+x^2\over 1-x}\ln {-x\over 1-x} - 1 + {3\over 2(1-x)}\right)^{[-\infty,0]}_{+(1)} - \left({3\over 2(1-x)}\right)^{[-\infty,0]}_{+(-\infty)} \quad &\, x<0
\end{array}\right.\nonumber\\[5pt]
& + {\alpha_sC_F\over 2\pi}\left[\delta(1-x)\left( {3\over2}\ln{\mu^2\over 4p_z^2} + {5\over2}\right) +{3\over2}\gamma_E\left({1\over(x-1)^2}\delta^+({1\over x-1}) + {1\over(1-x)^2}\delta^+({1\over 1-x})\right)\right]
\nonumber .
\end{align}
To carry out this calculation we defined the Fourier transformation of the singular function $\ln(\zeta^2)$ as
\begin{align} \label{eq:subtlety}
\int {d\zeta\over 2\pi}\ e^{i x \zeta} \ln\zeta^2
= & \biggl[ \left.{d\over d\eta} \int {d\zeta\over 2\pi}\ e^{i x \zeta} (\zeta^2)^\eta \biggr] \right|_{\eta=0}
=\biggl[ \left.{d\over d\eta} {4^\eta\over \Gamma(-\eta)} {\Gamma(\eta+1/2)\over \sqrt{\pi}} {[\theta(x)+\theta(-x)]\over |x|^{1+2\eta}} \biggr]
\right|_{\eta=0}\nonumber\\
=& \gamma_E \left[\left(-\delta(x) + {1\over x^2} \delta^+\Big({1\over x}\Big)\right) + \left(-\delta(x) + {1\over x^2} \delta^+\Big({1\over -x}\Big)\right)\right] \nonumber\\
& -\left[\left({1\over x}\right)_{+(0)}^{[0,1]} + \left({1\over x}\right)_{+(\infty)}^{[1,\infty]}\right]\theta(x) - \left[\left({1\over -x}\right)_{+(0)}^{[-1,0]} + \left({1\over -x}\right)_{+(\infty)}^{[-\infty,-1]}\right]\theta(-x)\,,
\end{align}
where we have used the results in Eqs.~(\ref{eq:ft},\ref{eq:expansion}) to derive the second and last equalities, and took the limit $\eta\to 0^+$ or $\eta\to 0^-$ when needed. This $\tilde{q}'^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny IR}})$ does not satisfy vector-current conservation, and is different from $\tilde{q}^{(1)}(x,\mu/|p^z|,\epsilon_{\mbox{\tiny IR}})$ in \eq{qPDFren} only by the $\delta$-functions at $x_0=\pm\infty$. Within the function domain $-\infty < x < \infty$, they are exactly the same. We will see below that both \eq{qPDFren} and \eq{qPDFft} eventually lead to the same result for the one-loop matching coefficient.
The final ingredient we need for the matching calculations is the PDF, whose one-loop bare matrix element can be written as a sum of an $\overline{\rm MS}$ counterterm and renormalized matrix element, $q^{(1)}(x,\epsilon) = \delta q^{(1)}(x,\epsilon_{\mbox{\tiny UV}})+q^{(1)}(x,\epsilon_{\mbox{\tiny IR}})$, where
\begin{align}
\delta q^{(1)}(x,\epsilon_{\mbox{\tiny UV}}) &= {\alpha_sC_F\over 2\pi} {1\over \epsilon_{\mbox{\tiny UV}}} \left({1+x^2\over 1-x}\right)^{[0,1]}_{+(1)} \theta(x)\theta(1-x)
\,, \\
q^{(1)}(x,\epsilon_{\mbox{\tiny IR}}) &= {\alpha_sC_F\over 2\pi} {(-1)\over \epsilon_{\mbox{\tiny IR}}} \left({1+x^2\over 1-x}\right)^{[0,1]}_{+(1)} \theta(x)\theta(1-x)
\,. \nonumber
\end{align}
With the above results in hand we can now determine the matching coefficients up to one-loop order. Using \eq{ps-q-fact} we find
\begin{eqnarray} \label{eq:pseudoC}
{\cal C}^{(1)}(\alpha,z^2\mu^2)= {\cal P}^{(1)}(\alpha,z^2\mu^2,\epsilon_{\mbox{\tiny IR}}) - q^{(1)}(\alpha,\epsilon_{\mbox{\tiny IR}}) \,.
\end{eqnarray}
Therefore the matching coefficient relating the pseudo-PDF and PDF in the $\overline{\rm MS}$ scheme with $\Gamma=\gamma^0$ is\footnote{A one-loop analysis of the spatial correlator\xspace in the coordinate space also recently appeared in Refs.~\cite{Radyushkin:2017lvu,Radyushkin:2018cvn}. Our factorization result for the spatial correlator\xspace in \eq{io-Q-fact} has a similar form to the hard part of the reduced spatial correlator\xspace found in Eq.(3.35) of Ref.~\cite{Radyushkin:2017lvu} and Eq.(17) of Ref.~\cite{Radyushkin:2018cvn}. It is therefore interesting to compare our ${\cal C}(\alpha,z^2\mu^2)/C_0(\mu^2z^2)$ and this hard part. Our $\overline{\rm MS}$ result \eq{ps-c} differs from Refs.~\cite{Radyushkin:2017lvu} due to the presence of the $2(1-\alpha)$ term. The result in the final version of Ref.~\cite{Radyushkin:2018cvn} agrees with ours. \eq{ps-c} also agrees with the original result derived in Ref.~\cite{Ji:2017rah}, up to the addition of our $e^{2\gamma_E}$ terms. The result for ${\cal C}(\alpha,z^2\mu^2)$ in \eq{ps-c} should be used to extract an $\overline{\rm MS}$ PDF from an $\overline{\rm MS}$ result for the pseudo-PDF.}
\begin{align} \label{eq:ps-c}
{\cal C}(\alpha,z^2\mu^2)
=& \left[ 1+ {\alpha_sC_F\over 2\pi}\left({3\over2}\ln(z^2\mu^2)+{3\over2}\ln{e^{2\gamma_E}\over 4}+{3\over2}\right)\right]\delta(1-\alpha)
\\
&+ {\alpha_sC_F\over 2\pi}\left\{\left({1+\alpha^2\over 1-\alpha}\right)^{[0,1]}_{+(1)}\left[-\ln(z^2 \mu^2)-\ln{e^{2\gamma_E}\over 4}-1\right] - \left(4\ln(1-\alpha)\over 1-\alpha\right)^{[0,1]}_{+(1)} +2(1-\alpha)\right\}\theta(\alpha)\theta(1-\alpha)
\,. \nonumber
\end{align}
This result is independent of the infrared regulator as it must be.
We have also computed the matching coefficient for the $\Gamma=\gamma^z$ case, and it is ${\cal C}_{\gamma^z}(\alpha,z^2\mu^2) = {\cal C}(\alpha,z^2\mu^2) + \Delta {\cal C}_{\gamma^z}(\alpha,z^2\mu^2)$ with
\begin{eqnarray}
\Delta {\cal C}_{\gamma^z}(\alpha,z^2\mu^2) = {\alpha_sC_F\over 2\pi} 2(1-\alpha)\theta(\alpha)\theta(1-\alpha)\,.
\end{eqnarray}
Due to the $\ln(z^2\mu^2)\delta(1-\alpha)$ term in \eq{ps-c}, the matching coefficient for the $\overline{\rm MS}$ pseudo-PDF again displays the fact that there is not a smooth local limit as $z\to 0$. It is possible to define a scheme other than $\overline{\rm MS}$ to ensure that this limit is smooth, reproducing a renormalization for $z\to 0$ that agrees with the fact that the local operator corresponds with a conserved current. One such scheme would be to simply multiply all $\overline{\rm MS}$ renormalization constants by $C_0(\mu^2 z^2)$, which would lead to a spatial correlator\xspace renormalized in a different scheme, and a corresponding different matching coefficient in \eq{ps-c} with a smooth $z\to 0$ limit. This is equivalent to studying the ratio of \eq{Qtratio} from the start as advocated in Ref.~\cite{Radyushkin:2017cyf,Radyushkin:2017lvu}. We will give explicit results for this scheme choice below. This modified scheme should not be confused with the strict definition of the $\overline{\rm MS}$ scheme.
From \eq{fac-tq} the corresponding relation for the matching coefficient for the quasi-PDF defined in Eq.~(\ref{eq:fac-tq-orig0}) is
\begin{eqnarray} \label{eq:quasiC}
C^{(1)}(\xi,\mu/(|y|P^z))= \tilde q'^{(1)}(\xi,\mu/|y|P^z,\epsilon_{\mbox{\tiny IR}})-q^{(1)}(\xi,\epsilon_{\mbox{\tiny IR}})\,.
\end{eqnarray}
Therefore using \eq{qPDFft} the matching coefficient relating the quasi-PDF and PDF is
\begin{align} \label{eq:quasi-c}
C\left(\xi, {\mu\over |y| P^z}\right)
= &\, \delta\left(1-\xi\right)+{\alpha_sC_F\over 2\pi}\left\{
\begin{array}{ll}
\displaystyle \left({1+\xi^2\over 1-\xi}\ln {\xi\over \xi-1} + 1 + {3\over 2\xi}\right)^{[1,\infty]}_{+(1)}- \left({3\over 2\xi}\right)^{[1,\infty]}_{+(\infty)}
&\, \xi>1
\nonumber\\[10pt]
\displaystyle \left({1+\xi^2\over 1-\xi}\left[
- \ln{\mu^2\over y^2P_z^2} + \ln\big(4\xi(1-\xi)\big)\right] - {\xi(1+\xi)\over 1-\xi}\right)^{[0,1]}_{+(1)}
&\, 0<\xi<1
\nonumber\\[10pt]
\displaystyle \left(-{1+\xi^2\over 1-\xi}\ln {-\xi\over 1-\xi} - 1 + {3\over 2(1-\xi)}\right)^{[-\infty,0]}_{+(1)} - \left({3\over 2(1-\xi)}\right)^{[-\infty,0]}_{+(-\infty)} \quad
&\, \xi<0
\end{array}\right.\nonumber\\[5pt]
& + {\alpha_sC_F\over 2\pi}\left[\delta(1-\xi) \left( {3\over2}\ln{\mu^2\over 4y^2P_z^2} + {5\over2}\right) + {3\over2}\gamma_E\left({1\over(\xi-1)^2}\delta^+({1\over \xi-1}) + {1\over(1-\xi)^2}\delta^+({1\over 1-\xi})\right)\right]\,.
\end{align}
Again this result is independent of the IR regulator as it must be. Here the plus function terms $\big[ g_1(\xi) \big]_{+(1)}^{[1,\infty]}$ and $\big[ g_2(\xi) \big]_{+(1)}^{[-\infty,0]}$ have integrands that converge for $\xi\to \pm \infty$, behaving as $g_i(\xi)\sim 1/\xi^2$. Note that if we had instead used the renormalized $\overline{\rm MS}$ quasi-PDF calculated in Eq.~(\ref{eq:qPDFren}), we would obtain a different matching coefficient $C$ with different $\delta$-functions at $\xi=\pm\infty$. However, the $\delta$-functions do not contribute to the convolution integral in Eq.~(\ref{eq:fac-tq}) for any integrable PDFs. For example, to carry out the convolution with $1/\xi^2 \delta^+(1/\xi)$ we can use $\delta^+\big({1\over \xi}\big) = \lim_{\beta\to0^+} \delta\big({1\over \xi} - \beta\big)$, which
when plugged into the factorization formula gives
\begin{align}
&\lim_{\beta\to0^+}\int {dy\over |y|} {y^2\over x^2}
\, \delta\Big({y\over x} -\beta\Big) f_{u-d}(y)
=\lim_{\beta\to0^+}\beta f_{u-d}(\beta x)\,.
\end{align}
For the plus-function at $\infty$ using \eqs{plusinfinity}{plus1} we have
\begin{align}
\int_{-1}^{+1}\! {dy\over |y|} \left[{1\over (x/y)}\right]_{+(\infty)}^{[1,\infty]} f_{u-d}(y)
&= \lim_{\beta\to 0^+} \int_{-1}^{+1}\! {dy\over |y|}
\bigg[ {\theta(x/y-\beta) \over x/y}
+ {y^2\over x^2}\delta\Bigl({y\over x} -\beta\Bigr)\ln\beta \bigg] f_{u-d}(y)
\nonumber\\
&= \int_{-1}^{+1}\! {dy\over x} {y\over |y|} f_{u-d}(y)
+ \lim_{\beta\to 0^+}\beta f_{u-d}(\beta x)\ln\beta
\,. \nonumber
\end{align}
In the last line we dropped the $\theta(x/y-\beta)$ since at small $y$ our PDF behaves as $f_{u-d}(y)\sim y^{-1+a}$ with $0<a<1$. This also implies
\begin{align}
\lim_{\beta\to0}\beta f_{u-d}(\beta x) \propto\: & x^{-1+a}\lim_{\beta\to0} \beta^a =0\,, \nonumber\\
\lim_{\beta\to0}\beta f_{u-d}(\beta x)\ln\beta \propto\: & x^{-1+a}\lim_{\beta\to0} \beta^a \ln\beta =0\,,
\end{align}
which means that the distribution contributions evaluated at $\xi=\pm\infty$ in the matching coefficient $C$ give zero contribution.
Therefore, the matching coefficients calculated from the quasi-PDFs in Eq.~(\ref{eq:qPDFren}) and Eq.~(\ref{eq:qPDFft}) are the same in effect, and we can simply drop all the $\delta$-functions at $\xi=\pm\infty$ when plugging them into the factorization formula:
\begin{align} \label{eq:quasi-matching}
C^{\overline{\rm MS}}\left(\xi, {\mu\over |y| P^z}\right)
= &\, \delta\left(1-\xi\right)+{\alpha_sC_F\over 2\pi}\left\{
\begin{array}{ll}
\displaystyle \left({1+\xi^2\over 1-\xi}\ln {\xi\over \xi-1} + 1 + {3\over 2\xi}\right)^{[1,\infty]}_{+(1)}- {3\over 2\xi}
&\, \xi>1
\nonumber\\[10pt]
\displaystyle \left({1+\xi^2\over 1-\xi}\left[
- \ln{\mu^2\over y^2P_z^2} + \ln\big(4\xi(1-\xi)\big)\right] - {\xi(1+\xi)\over 1-\xi}\right)^{[0,1]}_{+(1)}
&\, 0<\xi<1
\nonumber\\[10pt]
\displaystyle \left(-{1+\xi^2\over 1-\xi}\ln {-\xi\over 1-\xi} - 1 + {3\over 2(1-\xi)}\right)^{[-\infty,0]}_{+(1)} - {3\over 2(1-\xi)}\quad
&\, \xi<0
\end{array}\right.\nonumber\\[5pt]
& + {\alpha_sC_F\over 2\pi}\delta(1-\xi) \left( {3\over2}\ln{\mu^2\over 4y^2P_z^2} + {5\over2}\right)\,.
\end{align}
The use of \eq{quasi-matching} in the factorization formula is valid for any PDF that behaves as $\lim_{y\to 0} f(y,\mu) \sim y^{-1+a}$ with $a>0$.
We have also computed the matching coefficient for the $\Gamma=\gamma^z$ case, and it is given by $C_{\gamma^z}(\xi, \mu/(|y| P^z)) = C(\xi, \mu/(|y| P^z)) + \Delta C_{\gamma^z}(\xi, \mu/(|y| P^z))$ with
\begin{align}
\Delta C_{\gamma^z}(\xi, \mu/(|y| P^z)) = {\alpha_s C_F\over 2\pi} 2(1-\xi)
\,\theta(\xi)\theta(1-\xi) \,.
\end{align}
Note that our result for the quark matching coefficient in $\overline{\rm MS}$ differs from that of Ref.~\cite{Wang:2017qyg} which is a pure plus function, but gives a convolution that does not converge, just as in the case of the quasi-PDF with a transverse momentum cutoff, see Ref.~\cite{Stewart:2017tvs}.
Since the renormalized pseudo-PDF and quasi-PDF satisfy the relation in Eq.~(\ref{eq:equiv}) by definition, $C\left(\xi, \mu/(|y| P^z)\right)$ and ${\cal C}(\alpha,z^2\mu^2)$ that are given by Eqs.~(\ref{eq:ps-c}, \ref{eq:quasi-c}) automatically satisfy the relation in Eq.~(\ref{eq:ftc-C}).
Besides, if one uses a scheme other than $\overline{\rm MS}$ for the quasi-PDF, such as the scheme obtained by absorbing $C_0$ into the $\overline{\rm MS}$ renormalization constant, then this will lead to a result for the matching coefficient that is a pure plus function and hence satisfies current conservation. Starting with \eq{quasi-c} and using \eq{C0} together with \eq{subtlety} we obtain
\begin{align} \label{eq:quasi-matching-alt}
C^{\rm ratio\!}\left(\xi, {\mu\over |y| P^z}\right)
= &\, \delta\left(1-\xi\right)+{\alpha_sC_F\over 2\pi}\left\{
\begin{array}{ll}
\displaystyle \left({1+\xi^2\over 1-\xi}\ln {\xi\over \xi-1} + 1 - {3\over 2(1-\xi)}\right)^{[1,\infty]}_{+(1)}
&\, \xi>1
\nonumber\\[10pt]
\displaystyle \left({1+\xi^2\over 1-\xi}\left[
- \ln{\mu^2\over y^2P_z^2} + \ln\big(4\xi(1-\xi)\big)-1\right] +1+ {3\over 2(1-\xi)}\right)^{[0,1]}_{+(1)}
&\, 0<\xi<1
\nonumber\\[10pt]
\displaystyle \left(-{1+\xi^2\over 1-\xi}\ln {-\xi\over 1-\xi} - 1 + {3\over 2(1-\xi)}\right)^{[-\infty,0]}_{+(1)}\quad
&\, \xi<0
\end{array}\right.\,,\nonumber\\
\end{align}
and for the $\Gamma=\gamma^z$ case,
\begin{eqnarray}
\Delta C^{\rm ratio}_{\gamma^z}(\xi, \mu/(|y| P^z)) = {\alpha_s C_F\over 2\pi} \big[2(1-\xi)\big]^{[0,1]}_{+(1)} \,.
\end{eqnarray}
While retaining current conservation in the renormalized quasi-PDF, \eq{quasi-matching-alt} can be used for example as input to the two-step matching procedure in the lattice calculation of PDF in Refs.~\cite{Alexandrou:2017huk}.
For the matching step, an equivalent procedure is to study the ratio given in \eq{Qtratio} in the $\overline{\rm MS}$ scheme from the start, as advocated in Ref.~\cite{Radyushkin:2017cyf,Radyushkin:2017lvu}, performing its matching onto the PDF, which will yield \eq{quasi-matching-alt}. This concludes our discussion of matching results and the equivalence between the quasi-PDF and pseudo-PDF at one-loop order.
\end{widetext}
\section{Other Renormalization Schemes}
\label{sec:ren}
Although we derive the above matching formula assuming that the quasi-PDF is renormalized in the $\overline{\ensuremath{\operatorname{MS}}}$ scheme, this is not a limitation to our result.
Since the gauge-invariant Wilson line operator $\tilde{O}_\Gamma(z)$ has been proven to be multiplicatively renormalizable in the coordinate space~\cite{Ji:2017oey,Ishikawa:2017faj}, one can convert $\tilde{Q}_{\Gamma}(z)$ from any other scheme to the $\overline{\ensuremath{\operatorname{MS}}}$ scheme before using the above factorization formula. The renormalization of the
quasi-PDF has been studied in many recent papers~\cite{Ji:2015jwa,Ishikawa:2016znu,Chen:2016fxx,Xiong:2017jtn,Constantinou:2017sej,Alexandrou:2017huk,Chen:2017mzz,Green:2017xeu,Stewart:2017tvs,Wang:2017eel}. We will discuss some
of these results and show how they can be incorporated into the factorization formula in Eq.~(\ref{eq:fac-tq}).
The $\overline{\ensuremath{\operatorname{MS}}}$ scheme is convenient for our discussion of the OPE as it guarantees Lorentz and gauge invariances, but it is not practical for lattice renormalization. Since the lattice theory has a natural UV cut-off $1/a$ with $a$ being the lattice spacing, the unrenormalized spatial correlator\xspace $\tilde{Q}$ inherits the power divergence from the Wilson line self-energy according to Eq.~(\ref{eq:ren}). For an arbitrary scheme $X$, the renormalized spatial correlator\xspace
\begin{align}
\tilde{Q}^X(\zeta,z^2\mu^2_R)
= \lim_{a\to0} Z^{-1}_X(z^2\mu_R^2,a^2\mu^2_R)\, \tilde{Q}(\zeta,z^2/a^2)
\end{align}
should be free of all the UV divergences and have a well-defined continuum limit as $a\to0$. This continuum limit, in particular, is independent of the UV regulator, so
\begin{align}
&\lim_{a\to0} Z^{-1}_X(z^2\mu_R^2,a^2\mu^2_R)\tilde{Q}(\zeta,z^2/a^2) \nonumber\\
&= Z^{-1}_X(z^2\mu^2_R,\epsilon)\tilde{Q}(\zeta,z^2,\epsilon)\,.
\end{align}
As a result, we can relate $\tilde{Q}^X(\zeta,z^2\mu^2_R)$ to the $\overline{\rm MS}$ scheme by the conversion
\begin{align}
\tilde{Q}^X(\zeta,z^2\mu^2_R) =& \frac{Z_{\overline{\rm MS}}(\epsilon,\mu)}{Z_X(z^2\mu^2_R,\epsilon)}\, \tilde{Q}^{\overline{\rm MS}}(\zeta,z^2\mu^2)\nonumber\\
=& Z'_X(z^2\mu_R^2,\mu_R^2/\mu^2)\, \tilde{Q}^{\overline{\rm MS}}(\zeta,z^2\mu^2)\,,
\end{align}
where the regulator $\epsilon$ dependence is completely canceled out between $Z_{\overline{\rm MS}}$ and $Z_X$. The ratio $Z_X'$ can be calculated perturbatively in QCD, which was done in~\cite{Constantinou:2017sej} for several lattice schemes and the RI/MOM scheme. Thus the factorization formula we have proven in Sec.~\ref{sec:ope} still applies to $\tilde{Q}^X$ with a slight modification to the coefficient function,
\begin{align}
\tilde{Q}^X(\zeta,z^2\mu^2_R)
= \int_{-1}^1 \!\! d\alpha \:
\mathcal{C}^X(\alpha,\mu_R^2/\mu^2,\mu^2 z^2)\, Q(\alpha\zeta,\mu)\,,
\end{align}
where the matching coefficient for the scheme $X$ is related to that of $\overline{\rm MS}$ by
\begin{align}
\mathcal{C}^X(\alpha,\mu_R^2/\mu^2,\mu^2 z^2) = Z'_X(z^2\mu_R^2,\mu_R^2/\mu^2)\, \mathcal{C}(\alpha,\mu^2 z^2) \,.
\end{align}
For the pseudo-PDF the modified result also involves this same coefficient
\begin{align}
\mathcal{P}^X(x,z^2\mu^2_R)
=&\int_{|x|}^1 {dy\over |y|}\
\mathcal{C}^X\Bigl({x\over y},{\mu_R^2\over \mu^2},\mu^2 z^2\Bigr) q(y,\mu)
\\
&+\int^{-|x|}_{-1} {dy\over |y|}\
\mathcal{C}^X\Bigl({x\over y},{\mu_R^2\over \mu^2},\mu^2 z^2\Bigr) q(y,\mu)
\,. \nonumber
\end{align}
Meanwhile, for the quasi-PDF we have,
\begin{align}\label{eq:rimomfac}
\tilde{q}_X\left(x,{\mu_R^2\over P_z^2}\right)
&\equiv \int \frac{d \zeta}{2 \pi}\: e^{ix \zeta}\: \tilde{Q}^X \biggl(\zeta, \frac{{\mu}^2\zeta^2}{P_z^2}\biggr)
\nonumber\\
&=\int_{-1}^1 {dy\over |y|}
\: C^X\Bigl(\frac{x}{y},\frac{\mu_R}{\mu},\frac{\mu}{|y| P^z}\Bigr)\, q(y,\mu)\,.
\end{align}
Here the modified coefficient for the $X$ scheme is related to coefficient in the $\overline{\rm MS}$ scheme by
\begin{align}
& C^X\Bigl(\frac{x}{y}, \frac{\mu_R}{\mu},\frac{\mu}{|y| P^z}\Bigr)
\\
&\ =\!\! \int\!\! d\eta\ \bar{Z}'_X\Bigl(\eta^2, {\mu_R^2\over \mu^2}\Bigr)\:
C\Bigl( \frac{x}{y} - \frac{\eta}{|y|}\frac{\mu_R}{P^z},{\mu\over |y|P^z}\Bigr)
, \nonumber
\end{align}
where here $\bar{Z}'_X$ is defined by the Fourier transform
\begin{align}
\bar{Z}'_X\left( \eta^2,{\mu_R^2\over \mu^2}\right)
\equiv \int {d\tau\over 2\pi}\ e^{i\eta \tau }\, Z'_X\Bigl(\tau^2,\frac{\mu_R^2}{\mu^2}\Bigr)
\,.
\end{align}
Depending on the scheme $X$ we note that slightly modified definitions of $\bar{Z}'_X$ may be more appropriate.
One undesirable feature of the $\overline{\rm MS}$ scheme for the renormalized spatial correlator\xspace is that it does not have a smooth $z\to 0$ limit, and hence no simple connection with the fact that the local operator for $z=0$ is a conserved current. To avoid this one can simply make use of a different scheme that has a simple relation to $\overline{\rm MS}$, such as by adding $C_0(\mu^2z^2)$ to the $\overline{\rm MS}$ renormalization constant. This removes the offending $\ln(\mu^2 z^2)$ terms and yields a scheme with a smooth connection to the conserved current.
\begin{figure*}
\centering
\includegraphics[width=.49\textwidth]{Compare_qPDF_MSbar_Pzycut_TeX}
\vspace{-0.3cm}
\caption{
The $\overline{\rm MS}$ scheme PDF $xf_{u-d}$ and the $\overline{\rm MS}$ quasi-PDF obtained from $x\, C^{\overline{\rm MS}}(p^z)\otimes f_{u-d}$, comparing results obtained with $p^z=yP^z$ and $p^z=P^z$.
}
\label{fig:quasi}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=.49\textwidth]{Compare_pPDFvsPDF_TeX}
\hspace{0.1cm}
\includegraphics[width=0.49\textwidth]{Compare_pPDFvsPDF_z_TeX} \\
\vspace{-0.3cm}
\caption{(Left) Comparison between the PDF $xf_{u-d}$ and the pseudo-PDF $x({\cal C}^{\overline{\rm MS}}\otimes f_{u-d})$ in the $\overline{\rm MS}$ scheme. The orange and blue bands indicate the results from varying the factorization scale $\mu=4\,{\rm GeV}$ by a factor of two. (Right) Same but now showing only central pseudo-PDF curves for different values of $z$. }
\label{fig:pseudo}
\end{figure*}
Besides the $\overline{\rm MS}$ scheme, the quasi-PDF has also been defined with a transverse momentum cut-off~\cite{Xiong:2013bka,Lin:2014zya,Ma:2014jla,Alexandrou:2015rja} and in the RI/MOM scheme~\cite{Constantinou:2017sej,Alexandrou:2017huk,Chen:2017mzz,Green:2017xeu,Stewart:2017tvs}. The RI/MOM scheme has attracted strong interest recently as it can be implemented nonperturbatively on the lattice, so we consider it as an explicit example of the above relations. In this scheme, the renormalization constant $Z_{\rm OM}$ is determined by imposing a condition on the spatial correlator\xspace in an off-shell quark state,
\begin{align}
&\left.Z^{-1}_{\rm OM}\,
\tilde{Q}(\zeta=zp^z,z^2/a^2,-p^2a^2)\right|_{p^2=-\mu_R^2,p^z=p_R^z}
\nonumber\\
& \quad = \tilde{Q}^{(0)}_q(zp_R^z,z^2/a^2,z^2\mu_R^2) = e^{-izp_R^z}\,,
\end{align}
where $q$ denotes the quark state, $p^\mu$ is the external momentum, and ``$(0)$" in the superscript stands for the tree-level matrix element. As a result,
\begin{align}
Z_{\rm OM} &= Z_{\rm OM} (zp_R^z, z^2/a^2,a^2\mu_R^2)\,,\nonumber\\
Z'_{\rm OM}&=Z'_{\rm OM}(zp_R^z, z^2\mu_R^2, \mu_R^2/\mu^2)\,,
\end{align}
and here we define
\begin{align}
&\bar{Z}'_{\rm OM}\left(\eta,{\mu_R^2\over (p_R^z)^2},{\mu_R^2\over \mu^2}\right)\nonumber\\
&\equiv
p_R^z \int {dz\over 2\pi}\ e^{i\eta p_R^z z}\ Z'_{\rm OM}(zp_R^z, z^2\mu_R^2, \mu_R^2/\mu^2)\,.
\end{align}
Then the matching coefficient in Eq.~(\ref{eq:rimomfac}) becomes
\begin{align}
& C^{\rm OM}\left(\frac{x}{y}, {\mu_R\over p_R^z},{\mu_R\over \mu}, {\mu\over yP^z}\right)
\\
=& \int d\eta\ \bar{Z}'_{\rm OM}\left(\eta,{\mu_R^2\over (p_R^z)^2},{\mu_R^2\over\mu^2}\right) C\left({x\over y}-{\eta\over y}{p_R^z\over P^z}, {\mu\over |y|P^z}\right)
\,. \nonumber
\end{align}
The choices of $\mu_R$ and $p_R^z$ are independent of $\mu$ and $P^z$, and $p_R^z=P^z$ was used in Refs.~\cite{Chen:2017mzz,Stewart:2017tvs}.
It should be noted that on the lattice, due to the breaking of chiral symmetry, the vector-like quark Wilson line operator $\tilde{O}_{\gamma^\mu}(z)$ can mix with the scalar operator $\tilde{O}_{\bf 1}(z)$, as has been discussed in Refs.~\cite{Constantinou:2017sej,Alexandrou:2017huk,Chen:2017mzz,Green:2017xeu,Chen:2017mie}. After considering the mixing effects, the same factorization formula can still be applied to the RI/MOM quasi-PDF from lattice QCD.
\section{Numerical results}
\label{sec:num}
In this section we numerically analyze the quasi-PDF, spatial correlator\xspace and pseudo-PDF by studying how the matching coefficients in Eqs.~(\ref{eq:fac-tq},\ref{eq:ps-q-fact}) change the PDF. The quasi-PDF has already been studied in this manner for the $\overline{\rm MS}$, transverse momentum cut-off, and RI/MOM schemes in Ref.~\cite{Stewart:2017tvs}. Our new $\overline{\rm MS}$ result for the matching is given in \eq{quasi-matching}, and leads to stable convolution integrals. We also compare the differences between using hadron momentum $p^z=P^z$ and the parton momentum $p^z=|y|P^z$ for the matching coefficient in the $\overline{\rm MS}$ scheme. We take $\Gamma =\gamma^0$ for the results here.
As an example we use for our analysis the unpolarized iso-vector parton distribution,
\begin{align}
f_{u-d}(x,\mu) = f_u(x,\mu) - f_d(x,\mu) - f_{\bar{u}}(-x,\mu) + f_{\bar{d}}(-x,\mu),
\end{align}
where we include
$f_{\bar{u}}(-x,\mu) = - f_{\bar{u}}(x,\mu)$ and $f_{\bar{d}}(-x,\mu) = - f_{\bar{d}}(x,\mu)$, the anti-parton distributions. For ease of comparison, we use the next-to-leading-order iso-vector PDF $f_{u-d}$ from MSTW 2008~\cite{Martin:2009iq} with the corresponding running coupling $\alpha_s(\mu)$.
To implement the plus functions in the numerical calculation, we impose a soft cutoff $|y-x|<10^{-m}$ and test the sensitivity of results to $m$. Since the limit of $y\to0$ corresponds to the asymptotic region $|x/y|\to\infty$, we also impose a UV cutoff $|y|>10^{-n}$ to test the convergence of the convolution integral. We find that all the results presented below are insensitive to $m$ and $n$. The fact that our result in \eq{quasi-matching} has terms outside the plus function at $1$ in each of the $\xi\in [1,\infty]$ and $\xi\in [-\infty,0]$ intervals is important for ensuring that our $\overline{\rm MS}$ result for $C$ is insensitive to the $|y|>10^{-n}$ cutoff. This was not the case for the quasi-PDF that was defined with a transverse momentum cutoff~\cite{Xiong:2013bka}. The RI/MOM scheme result~\cite{Stewart:2017tvs} also does not suffer from this issue.
In Fig.~\ref{fig:quasi} we compare the PDF with the quasi-PDF in the $\overline{\rm MS}$ scheme obtained from the convolution in \eq{fac-tq} using our one-loop result in \eq{quasi-matching}. We observe that changing from $p^z=P^z$ to the correct $p^z=|y|P^z$ shifts the result in the physical region by a considerable amount.
\begin{figure*}
\centering
\includegraphics[width=.49\textwidth]{Compare_iPDFvsPDFRe_TeX}
\hspace{0.1cm}
\includegraphics[width=0.49\textwidth]{Compare_iPDFvsPDFRe_z_TeX}
\\
\includegraphics[width=.49\textwidth]{Compare_iPDFvsPDFIm_TeX}
\hspace{0.1cm}
\includegraphics[width=0.49\textwidth]{Compare_iPDFvsPDFIm_z_TeX} \\
\vspace{-0.3cm}
\caption{(left) Comparison between the light-cone time distribution $Q_{u-d}$ and spatial correlator\xspace from $({\cal C}^{\overline{\rm MS}}\otimes Q_{u-d})$ in the $\overline{\rm MS}$ scheme. The orange and blue bands indicate the results from varying the factorization scale $\mu=4\,{\rm GeV}$ by a factor of two. (right) Same but now showing only central spatial correlator\xspace curves for different values of $z$. The top panels show the real part, while bottom panels show the imaginary part. }
\label{fig:ioffe}
\end{figure*}
The same type of comparison can be made for the pseudo-PDF in the $\overline{\rm MS}$ scheme by applying the factorization formula in Eq.~(\ref{eq:ps-q-fact}) and matching coefficient in Eq.~(\ref{eq:ps-c}). In Fig.~\ref{fig:pseudo}a we compare the PDF and pseudo-PDF and their dependence on the factorization scale $\mu$, while in Fig.~\ref{fig:pseudo}b we include the dependence of the pseudo-PDF on the distance $|z|$. Since the matching coefficient in Eq.(\ref{eq:ps-c}) is similar to the parton splitting function except for the nontrivial finite constants, matching the PDF to the pseudo-PDF is analogous to evolving the PDF from $\mu$ to the scale of $1/|z|$. This evolution has been calculated in Refs.~\cite{Radyushkin:2017cyf}. The variation of $|z|$ has a similar effect to the PDF evolution, as is observed in the right panel of Fig.~\ref{fig:pseudo}. When $|z|\mu=1$, the logarithm is zero, and the matching effect from ${\cal C}$ is determined by the nontrivial constants in Eq.(\ref{eq:ps-c}), which shifts the PDF downward in the large-$x$ region and upward in the small-$x$ region.
Finally, we can make a similar comparison for the spatial correlator\xspace in the $\overline{\rm MS}$ scheme obtained with Eq.~(\ref{eq:io-Q-fact}) and Eq.~(\ref{eq:ps-c}). Its real and imaginary parts are even and odd functions of $\zeta$ respectively, and are shown in Fig.~\ref{fig:ioffe}. Again we show the residual dependence on $\mu$ and $|z|$ which are similar to that for the pseudo-PDF. The matching broadens the curves in the coordinate space. The spatial correlator\xspace renormalized in the $\overline{\rm MS}$ scheme does not exhibit vector current (or particle number) conservation, which can be clearly seen from the fact that the real part of the distribution is not equal to 1 at $\zeta=0$ (except for the special case where $|z|\mu$ is tuned to cancel the constant terms in the one-loop ${\cal C}$).
\section{Implications for lattice calculations}
\label{sec:lattice}
Our proof in \sec{ope} makes clear the relationship between the renormalized quasi-PDF, spatial correlator\xspace, and pseduo-PDF distributions. As a practical matter there are a few different ways in which these equations can be used to convert a lattice calculation of the spatial correlator\xspace $\tilde Q$ into a PDF. Three examples are 1) first Fourier transform to the quasi-PDF with \eq{fac-tq-orig0}, and then use \eq{fac-tq}, 2) first Fourier transform to the pseudo-PDF with \eq{pseudoRen}, and then use \eq{ps-q-fact}, and 3) first match to the Fourier transform the position space PDF $Q(\zeta,\mu)$ using \eq{io-Q-fact}, and then transform it to the PDF with the inverse of \eq{FTq}. Since the numerical implementation of these steps may have slightly different systematics it is interesting to compare them, or to use more than one approach in order to reduce uncertainties.
According to the analysis in Sec.~\ref{sec:ope}, for the factorization formula of the Euclidean distributions to work, one must calculate the same spatial correlator\xspace with small distance $z^2$ and large momentum $P^z$ so that the dynamical and kinematic higher-twist effects are suppressed. For practical lattice calculations, this means that there is only a finite number of useful data points in $(z, P^z)$ that we can use to extract the PDF.
To illustrate this, consider a $48^3\times64$ lattice with spacing $a=0.09$ fm. The distance of the spatial correlation $z$ is in units of $a\sim1/2.2\ {\rm GeV}^{-1}$, and the nucleon momentum $P^z$ is in units of $2\pi/(48a)\sim0.29$ GeV. Let us take $\Lambda_{\rm QCD}\sim 0.3$ GeV. In principal the target mass corrections can be subtracted. If we consider various values $z=ma$ and $P^z=n*2\pi/(48a)$ for integer $m$ and $n$, then to satisfy $z\Lambda_{\rm QCD}\ll1$ and $P^z\gg \Lambda_{\rm QCD}$, we must have
\begin{eqnarray}
m\ll 11\,,\,\,\,\quad n\gg 1\,.
\end{eqnarray}
To control the higher-twist correction at $20\%$, i.e. $z^2\Lambda_{\rm QCD}^2\sim0.2,\ \Lambda_{\rm QCD}^2/P_z^2\sim0.2$, we can only choose
\begin{eqnarray}
m = \{0,1,2,3,4\}\,,\,\,\,\, n=\{3,4,5,6,\cdots\}\,,
\end{eqnarray}
where the largest value for $n$ is limited by what is practical in current lattice simulations. Six is the largest number of units attained in Ref.~\cite{Orginos:2017kos}. For quasi-PDF calculations, there are $4\times2+1=9$ useful data points for each fixed momentum $|P^z|$; for pseudo-PDF calculation, there are only $4\times2=8$ useful data points for each fixed $|z|$. In either case, it is anticipated that a direct Fourier transform with respect to $z$ or $\zeta=zP^z$ will lead to oscillation in $x$-space and incorrect prediction for the small-$x$ region due to the truncation in coordinate space. This has been observed in a recent lattice calculation of the quasi-PDF in Ref.~\cite{Chen:2017mzz}. Methods have been developed in recent works to eliminate the oscillation from the truncation effect~\cite{Lin:2017ani,Zhang:2017gau} in the quasi-PDF, while the higher-twist contributions at large $z$ still need to be systematically corrected. It should be noted that the above is a rough estimate of the higher-twist corrections since the prefactor of $z^2\Lambda_{\rm QCD}^2$ could be smaller than 1. Their actual significance can only be quantitatively determined from lattice simulations.
To fully take advantage of all the useful data points, we can evolve them to either the same $z^2$ or $P^z$ according to the perturbative analysis, which has been studied in Refs.~\cite{Orginos:2017kos,Karpie:2017bzm,Radyushkin:2017lvu} for the spatial correlator\xspace. However, since the evolution equation of the spatial correlator\xspace in $\ln z^2$ or $\ln P_z^2$ follows a nonlocal convolution in $\zeta=zP^z$ or $z$, one has to know the full information in coordinate space to do the evolution. With limited number of data points, either large uncertainties or adopting a model-dependent assumption about the shape is inevitable.
To improve the precision of either approach, the only way forward is to have finer lattice spacing $a$ so that we could have more data points which satisfy $|z|\ll \Lambda_{\rm QCD}^{-1}$ and larger nucleon momentum $P^z$.
With increasing $P^z$, the valence distribution of the nucleon is contracted in the $z$ direction, so the spatial correlation of valence quarks is shrinked into smaller distance in $z$. If $P^z$ is large enough, the spatial correlation will fall off quickly within $|z|< \Lambda_{\rm QCD}^{-1}$, then the truncation error from Fourier transform will be significantly reduced. On the other hand, if we interpret the spatial correlation as the spatial correlator\xspace, its shape will not change under a Lorentz boost because it is a scalar function of $\zeta=zP^z$ and $z^2$. Nevertheless, finer lattice spacing $a$ allows for calculation with a wider range of $P^z$, thus covering larger values of $\zeta=zP^z$ to reduce the truncation error. Since the number of useful data points increases quadratically with $1/a$, a more precise lattice calculation with controlled systematic errors will be available in the future.
\section{Conclusions}
\label{sec:summary}
Starting with a Euclidean operator product expansion for products of gauge invariant operators in QCD, we have derived the factorization formulas for the quasi-PDF, spatial correlator\xspace and pseudo-PDF. The three Euclidean distribution functions are related observables, and all follow from the same fundamental factorization. For the spatial correlator\xspace this derivation implies that the ratio in Eq.~(\ref{eq:ratio}) does scale in $z^2$, but needs the small $z^2$ factorization formula in \eq{io-Q-fact} to extract the PDF.
Our derivation for the factorization formula applies when the renormalized spatial correlator\xspace is defined in any scheme.
The OPE used here could also be used to systematically derive factorization formulas for power corrections to \eq{momfact}, which will involve matching onto higher-twist parton distributions. (The numerical relevance of these corrections is considered in Ref.~\cite{Chen:2016utp}.)
Note that LaMET is not equivalent to the expansion from the OPE, as the former is more general and can be applied to the lattice calculations of other quantities, for example the TMD-PDF where a simple OPE does not exist.
Our derivation of the factorization formula for the quasi-PDF also verifies that the parton momentum $p^z$ in the matching coefficient in Eq.~(\ref{eq:momfact}) has to be $p^z=yP^z$, which makes a considerable difference for the $\overline{\rm MS}$ matching result when compared with $p^z=P^z$ (see Fig.~\ref{fig:quasi}). The proper $p^z$ should therefore be used in lattice calculations of the PDF in the LaMET approach.
As a non-trivial test of relations between the various distributions and factorization formulas we have considered results at one-loop in the $\overline{\rm MS}$ scheme. We have derived a corrected results for the coefficient $C$ for the $\overline{\rm MS}$ scheme, given in \eq{quasi-matching}, which leads to convergent results in the convolution integral. We have also computed the one-loop $\overline{\rm MS}$ result for the Wilson coefficient ${\cal C}$ appearing in the spatial correlator\xspace and pseudo-PDF factorizations. A numerical analysis of these one-loop corrections in $\overline{\rm MS}$ has also been provided. The one-loop matching coefficient ${\cal C}$ has a smaller effect for the pseudo-PDF than $C$ does for quasi-PDF, as can seen by comparing \figs{quasi}{pseudo}. Given systematic uncertainties in manipulating the lattice data, it is potentially interesting to consider using the same lattice data on the spatial correlator\xspace to extract the parton distribution function using both the quasi-PDF and pseudo-PDF approaches.
There are several different ways of implementing the factorization formula to calculate the PDF from lattice data for the spatial correlator\xspace $\tilde Q$, which we have discussed in \sec{lattice}. One always has to work with short distance correlation and large nucleon momentum to reduce higher-twist corrections. This limits the number of useful data points from lattice calculations as described in \sec{lattice}. To achieve precision calculations without making model assumptions it will be highly desirable to move towards finer lattice spacing to increase the number of effective data points.
\section*{Acknowledgments}
The authors are thankful for discussions with J.~H.~Zhang and Y.~B.~Yang. This material was partially supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics from DE-FG02-93ER-40762, DE-SC0011090 and DE-SC0012704, by the Laboratory Directed Research and Development (LDRD) funding of BNL under contract DE-EC0012704, by a grant from the National Science Foundation of China (No.~11405104), and within the framework of the TMD Topical Collaboration.
I.S. was also supported in part by the Simons Foundation through the Investigator grant 327942.
T.I. was also supported by JSPS KAKENHI Grant Numbers JP26400261, JP17H02906, and also MEXT as ”Priority Issue on Post-K computer” (Elucidation of the Fundamental Laws and Evolution of the Universe) and JICFuS.
|
\section{Introduction}
It is often assumed that most (if not all) stars are formed in dense, embedded, gravitationally bound clusters.
In this monolithic model of star formation, observed low density associations of OB stars must have been formed as single dense stellar clusters (i.e. \textquotedblleft singularly monolithic\textquotedblright) or multiple stellar clusters (i.e. \textquotedblleft multiply monolithic\textquotedblright) that subsequently underwent a period of expansion to form the large scale structures visible today \citep{LadaLada1991,Brown1997,Kroupa2001}. It is postulated that following the formation of stars, gas is expelled from clusters through feedback, rendering the clusters super-virial and allowing dispersal \citep{Hills1980,Goodwin2006,Baumgardt2007}.
This mechanism has often been used to explain the low proportion of gas-free, gravitationally bound clusters clusters after a few Myr.
However, star formation is observed to take place in a wide range of environments including large scale hierarchical structures and isolated young stellar objects which are not associated with clusters (e.g. \citealt{Gomez1993, Lamb2010,Allen2007,Gutermuth2008, Evans2009}).
In an alternative, hierarchical model of star formation, (e.g. \citealt{Elmegreen2002, Elmegreen2008, Bastian2007,Bonnell2011,Kruijssen2012b})
stars form in a substructured and scale-free distribution of initial conditions (as expected from a supersonically turbulent interstellar medium, ISM), which can naturally explain the wide range of observed stellar densities.
\citet{Bressert2010} find that present-day, nearby young stellar objects (YSOs) are observed over a continuous density distribution, without features corresponding to individual clusters, concluding that this is likely a result of star formation occurring over a continuous density distribution. However, this conclusion has been contested by studies showing that expanding clusters can reproduce a similar result \citep{Gieles2012,Parker2012,Pfalzner2012}.
Recent observational studies of the progenitors of young massive clusters (YMCs) have concluded that the density is insufficient to form the observed stellar densities of YMCs in-situ and that therefore the mass must become more centrally concentrated as the (proto)cluster evolves (\citealt{Walker2015,Walker2016}, see also the review of \citealt{Longmore2014}), in strong contrast with the idea of gas expulsion-driven expansion towards the present-day densities.
Recent theoretical work has also called into question the effectiveness of gas expulsion as a means of cluster disruption, finding that the highest-density regions achieve gas exhaustion rather than gas expulsion, due to the short free-fall times and correspondingly high star formation efficiencies \citep{Kruijssen2012c,Girichidis2012}. As a result, the dynamical effect of gas expulsion is small. This work is supported by the observations presented by \citet{Ginsburg2016}, who find that the W51 protoclusters are evolving towards a state of gas exhaustion rather than gas expulsion.
The above concepts have been combined in the analytical theory for the fraction of star formation occurring in bound clusters (i.e. the cluster formation efficiency) by \citet{Kruijssen2012b}, who translates the local star formation efficiency to a bound fraction {by integrating over the continuous density spectrum of the ISM. The model predicts that the fraction of star formation occurring in bound clusters increases with gas pressure and surface density, leading to the conclusion that clusters are a possible outcome of star formation, rather than a fundamental unit.
The monolithic model of star formation firmly predicts that, if gas expulsion is a rapid process (operating on a time-scale smaller than one crossing time), the stars outside of the Lagrangian radius containing 20\% of the stars in a cluster will acquire strongly radially anisotropic velocities. For the first 20\,Myr, the $N$-body simulations of \citet{Baumgardt2007} predict that expanding clusters are strongly radially anisotropic and super-virial. This anisotropy initially takes values in the range $\beta=0.4$--$1$, but is expected to persist to a value of $\beta\sim0.2$ up to hundreds of initial crossing times (see Figures 8 and 9 in \citealt{Baumgardt2007}).
If clusters expand into transient associations before ultimately dispersing into the Galactic field, then we would expect OB associations to typically exhibit strongly radially anisotropic velocity fields with (on average) positive outward radial motions.
Regardless of the number of expanding clusters, and their geometry, expansion of clusters within an association must result in a net expansion of the association.
Therefore, observing the dynamics of present-day OB associations offers a powerful diagnostic to distinguish between the two paradigms of star and cluster formation, i.e. monolithic or hierarchical star formation.
This is not the first paper aiming to address this question by studying the dynamics of OB associations.
\citet{Preibisch1999} presented a detailed analysis of the distribution and kinematics in the Upper Sco OB association; however, they make no conclusions as to the origin of the association.
In a detailed analysis of the structure and kinematics of the OB association Cygnus OB2, \citet{Wright2014,Wright2016} found that the level of dynamical evolution experienced by this association is low, concluding that Cygnus OB2 most likely formed as a highly sub-structured and globally unbound system, i.e.~it was always an association.
While previous studies have targeted individual associations, our study capitalises on the unprecedented surveying power of Gaia to greatly expand the sample of OB associations, providing a systematic and comprehensive analysis.
In this work we test the hypothesis that all OB associations result from the expansion of one or more gravitationally bound clusters. We focus only on the generalised behaviour of OB associations and do not fit models to any specific OB associations. The main goal of this paper is to assess the validity of the singularly monolithic and multiply monolithic models of star formation as defined in this introduction.
We test the hypothesis that all OB associations are expanded (massive) star clusters using the first {\it Gaia} \citep{GAIA2016} data release (DR1, \citealt{GAIADR12016}).
Using the Tycho-{\it Gaia} Astrometric Solution (TGAS, \citealt{Michalik2015}) catalogue, we have carried out a 5-dimensional (position, distance, proper motion) association member selection of stars in the vicinities of known OB associations. We have performed a 4-dimensional analysis for each association, measuring the positions and velocities in the plane of the sky in order to determine whether the present day OB associations are indeed undergoing gas expulsion-driven expansion.
In Section 2, we introduce our sample selection and association membership criteria as well as our initial data reduction and the generation of model data sets with which to compare the observed data.
In Section 3, we quantify four key kinematic properties of the OB associations and compare these to model distributions.
The implications of these results and remaining caveats are discussed in Section 4. Finally, our conclusions are summarised in Section 5.
\section{Sample selection, data reduction and model distributions}
Accurate and precise distances towards OB association members are crucial for both membership selection and the analysis of dynamical properties of those associations. This study is therefore limited to a strictly enforced 3$\sigma$ confidence limit on the distances (i.e.~$\sigma_D\leq D/3$) derived from the TGAS catalogue parallaxes.
We obtained parallaxes and corresponding uncertainties from the {\it Gaia} archive\footnote{https://gea.esac.esa.int/archive/} for the entire TGAS catalogue (entries for $\sim$2 million stars). From these parallax measurements, we calculate distances and distance uncertainties.
Figure \ref{tgas_distance_fig} shows the uncertainty in distance relative to the distance (derived from TGAS parallaxes) against distance in kpc for the entire TGAS catalogue.
Beyond a distance of 1.6\,kpc, no stars have distances determined to the precision required in this study.
This motivates an absolute distance cut of 1.6\,kpc as no stars have distances measured to a 3$\sigma$ confidence interval beyond this distance. Only OB associations with literature distances of less than or equal to 1.6\,kpc are therefore considered in the remainder of this work.
For OB associations within this distance, we only consider the stars with distance uncertainties smaller than 1/3 of the distance.
This defines the parent sample from which OB associations will be extracted based on their clustering in 5-dimensional phase space, yielding much smaller numbers of stars ($N_*=100$--$400$).
\begin{figure}
\includegraphics[width=0.99\linewidth]{distance_uncertainty_20170918.eps}
\caption{\label{tgas_distance_fig} Relative uncertainties in the derived distances against distance based on parallax measurements for the entire TGAS catalogue. The dashed lines mark the 3$\sigma$ distance uncertainty cut-off imposed in this work, motivating a maximum distance cut of 1.6\,kpc. Only those stars that fall in the lower left quadrant of this Figure are considered in this study. This defines the parent sample from which OB associations will be extracted based on their clustering in 5-dimensional phase space, yielding much smaller numbers of stars ($N_*=100$-$400$).}
\end{figure}
\subsection{TGAS data}
\label{TGAS_data}
We use data from the TGAS catalogue of stellar parallaxes and proper motions to characterise the kinematic properties of nearby OB associations.
We obtained parallaxes and proper motions (along with corresponding uncertainties) for potential members of each of the 43 OB associations within 1.6\,kpc of the sun based on the distances of \citet{Melnik2009}. These data were obtained using a cone search with a radius of 150\,pc at the literature distance for each association.
For each OB association, data for all stars with distances constrained to less than 3$\sigma$ have been excluded from the remainder of our study.
The remaining sample was then cross-referenced with the Simbad database, determining the nearest counterparts within a positional tolerance of 1$^{\prime\prime}$ to define our sample of OB-type stars. A cut has been imposed to exclude sources with distances greater than 150\,pc from the expected association distances from \citet{Melnik2009}.
Using the distances determined from the TGAS parallaxes, we calculate 2-dimensional physical distances ($X$,$Y$) in the plane of the sky from the median position of the OB stars in the sample in pc. Likewise, the same distances are used to convert observed proper motions from mas/yr to physical velocities in km\,s$^{-1}$.
We define over-abundances of OB stars as regions with peak number densities of at least three times the observed background level of O- and B- type stars measured across the 150\,pc radius region with bin sizes of 2\,pc.
We then determine the position and size of the associations in each axis by fitting Gaussian profiles to the over-densities in each of the 5 dimensions available: $X$, $Y$, $D$, $v_x$, and $v_y$.
Association members are selected as satisfying
$|x| < \sigma_X$ and $|y| < \sigma_Y$, where $x \equiv X-\mu_X$ and $y \equiv Y-\mu_Y$.
A strict 1$\sigma$ criterion is imposed where possible in $X$ and $Y$ in order to minimise the likelihood of selecting contaminating field stars as association members. In cases where the 1$\sigma$ criterion did not yield at least 100 stars including at least 10 O- or B- type stars, a 3$\sigma$ criterion was imposed and if insufficient numbers of stars were selected within 3$\sigma$ the association was excluded from the sample.
Similarly, we impose a line-of-sight distance selection criterion of $|Z| < (\Delta D^{2} + \sigma_Z^{2})^{1/2}$ where $Z \equiv D - \mu_{D} $ and $\Delta D$ is the uncertainty as derived from the uncertainty in the TGAS parallax measurement. This then removes the outliers in line-of-sight distance from the sample. The dispersions used to carry out the above selection criteria are listed in Table \ref{disp_tbl}.
\begin{table}
\caption{\label{disp_tbl} Distances (D) and measured dispersions for each cluster in the five dimensions used in the selection criteria of Section 2.1: dispersions in the plane of the sky ($\sigma_{X}$,$\sigma_{Y}$), dispersion in the line-of-sight distance derived from TGAS parallaxes ($\sigma_{Z}$), and velocity dispersions in the plane of the sky ($\sigma_{v_{X}}$, $\sigma_{v_{Y}}$).}
\begin{center}
\begin{tabular}{l c c c c c c}
\hline
OB assoc. & D & $\sigma_{X}$ & $\sigma_{Y}$ & $\sigma_{Z}$ & $\sigma_{v_{X}}$ & $\sigma_{v_{Y}}$ \\
& kpc & pc & pc & pc & km\,s$^{-1}$ & km\,s$^{-1}$ \\
\hline
CamOB1 & 0.82 & 15 & 9 & 12 & 8 & 8 \\
CasOB14 & 0.85 & 12 & 9 & 27 & 13 & 5 \\
CepOB2 & 0.71 & 19 & 8 & 11 & 10 & 6 \\
CepOB3 & 0.68 & 35 & 14 & 14 & 11 & 6 \\
CepOB4 & 0.72 & 7 & 16 & 6 & 12 & 7 \\
Coll140 & 0.35 & 45 & 7 & 28 & 7 & 7 \\
Coll359 & 0.20 & 22 & 7 & 9 & 8 & 7 \\
CygOB7 & 0.68 & 14 & 12 & 46 & 11 & 6 \\
CygOB9 & 0.94 & 15 & 22 & 13 & 10 & 3 \\
MonOB1 & 0.60 & 3 & 14 & 32 & 7 & 6 \\
OriOB1 & 0.40 & 14 & 28 & 22 & 4 & 5 \\
PerOB3 & 0.18 & 23 & 8 & 31 & 12 & 10 \\
ScoOB2 & 0.15 & 13 & 10 & 12 & 3 & 3 \\
SctOB3 & 1.25 & 26 & 67 & 45 & 9 & 7 \\
SgrOB1 & 1.27 & 38 & 87 & 29 & 7 & 4 \\
Tr27 & 0.90 & 10 & 15 & 18 & 8 & 10 \\
VelOB2 & 0.38 & 30 & 14 & 42 & 6 & 6 \\
VulOB4 & 0.80 & 30 & 13 & 16 & 8 & 10 \\
\hline
\end{tabular}
\end{center}
\end{table}
As we are interested in the kinematic properties of the association members, kinematic outliers which are not true association members are potentially a serious problem as they can artificially skew the diagnostics employed in Section 3. For this reason, we impose a velocity cut of $|v_{x}| < 3\sigma_{v_{X}}$ and $|v_y| < 3\sigma_{v_{Y}}$ with $v_x \equiv v_X-\mu_{v_X}$ and $v_y \equiv v_Y-\mu_{v_Y}$, removing the most extreme velocity sources while maintaining the vast majority of the sample. OB associations were rejected if they did not contain at least 100 stars including at least 10 O- or B-type stars within the 3$\sigma$ limit. This yields a final sample of 18 OB associations. The number of selected members and the number of dispersions in $X$,$Y$ and distance used to select the members of each association are listed in Tables \ref{param_tbl} and \ref{param_tbl_relall}. These kinematic criteria typically filter out less than 1\% of the stars in each association, implying that it has a limited influence on the diagnostics presented in Section 3.
The position-velocity analysis employed in this work depends on a reasonable determination of the centres of the OB associations. The centres of mass of the associations require known masses for all association members, which are not available. Therefore we define two association centres and use both in the subsequent analysis. The first centre is determined to be the mean 4-dimensional ($X$,$Y$,$v_x$,$v_y$) position of the OB stars - this is of course biased towards the most massive and least numerous sources in the association.
The properties of the OB associations using the mean position and velocity of the OB-type stars are given in Table \ref{param_tbl}.
The second centre is set as the mean 4-dimensional position of all stars in the association, biased towards the more numerous low mass stars.
Table \ref{param_tbl_relall} shows the properties of the OB associations when using the mean position and velocity of all stars.
Given that one centre is biased towards low mass stars and the other is biased towards high mass stars, it is expected that the centre of mass of each association lies between these two points.
\begin{table*}
\begin{minipage}{170mm}
\caption{\label{param_tbl} Table of parameters for all 18 OB associations relative to the centre of OB stars. The first column gives the association names. This is followed by the number of dispersions ($\sigma$) to which association members were selected in 3-dimensional spatial coordinates in order to provide at least 100 stars, at least 10 of which are OB-stars (see Section \ref{TGAS_data}). The third column contains the number of member stars in each association. Column 4 gives the ratio $N_{v_{r}>0}/N_{v_{r}<0}$ as described in Section 3.1. Columns 5 and 6 give the median radial velocities and radial velocity dispersions. The average velocity uncertainty for each association is given in column 7. The ratio between the radial velocity and the absolute of the tangential velocity (see Section 3.2) and the dispersion in this value are given in the 8th and 9th columns. The final (10th) column displays the radial velocity anisotropy parameter as described in Section 3.3 for each of the associations, where positive (negative) values denote radial (tangential) anisotropy.}
\begin{center}
\begin{tabular}{l c c c c c c c c c c}
\hline
assoc. & $N_{\sigma}(XYZ)$ & $N_{\star}$ & $N_{v_{r}>0}/N_{v_{r}<0}$ & med. $v_{\text{r}}$ (km\,s$^{-1}$) & $\sigma_{v_{r}}$ & $<\Delta v>$ (km\,s$^{-1}$) & med. $v_{\text{r}} / |v_{\text{t}}|$ & $\sigma_{v_{r} / |v_{\text{t}}|}$ & $\beta$ \\
\hline
CamOB1 & 3 & 260 & 1.28 & 1.6 & 10.0 & 4.0& 0.24 & 1.45 & -0.055$\pm$0.006 \\
CasOB14 & 3 & 134 & 1.39 & 3.1 & 10.7 & 4.2 & 0.50 & 1.47 & 0.043$\pm$0.008 \\
CepOB2 & 1 & 127 & 1.15 & 1.2 & 8.8 & 4.5 & 0.34 & 2.23 & 0.480$\pm$0.099 \\
CepOB3 & 1 & 279 & 1.5 & 2.8 & 9.3 & 4.6 & 0.67 & 1.84 & 0.339$\pm$0.054 \\
CepOB4 & 3 & 388 & 0.90 & -0.4 & 10.0 & 3.6 & -0.07 & 1.19 & -0.590$\pm$0.070 \\
Coll140 & 1 & 124 & 1.10 & 0.6 & 9.3 & 2.0 & 0.14 & 1.17 & -0.868$\pm$0.138 \\
Coll359 & 3 & 180 & 0.92 & -0.62 & 10.0 & 1.2 & -0.05 & 1.17 & -0.526$\pm$0.074 \\
CygOB7 & 1 & 104 & 0.93 & -0.2 & 7.0 & 4.7 & -0.04 & 1.52 & 0.069$\pm$0.017 \\
CygOB9 & 3 & 314 & 1.15 & 0.5 & 8.0 & 6.7 & 0.09 & 1.34 & -0.264$\pm$0.036 \\
MonOB1 & 3 & 126 & 1.10 & 0.5 & 9.8 & 4.8 & 0.07 & 1.56 & 0.219$\pm$0.036 \\
OriOB1 & 1 & 190 & 0.88 & -0.3 & 3.0 & 1.6 & -0.14 & 1.51 & 0.035$\pm$0.008 \\
PerOB3 & 3 & 136 & 0.92 & -1.0 & 11.8 & 1.1 & -0.22 & 1.47 & 0.118$\pm$0.024 \\
ScoOB2 & 3 & 325 & 1.18 & 0.5 & 8.6 & 0.8 & 0.19 & 1.50 & 0.088$\pm$0.013 \\
SctOB3 & 3 & 100 & 1.13 & 0.4 & 9.9 & 12.7 & 0.04 & 1.44 & -0.102$\pm$0.021 \\
SgrOB1 & 3 & 101 & 1.15 & 0.8 & 9.8 & 10.7 & 0.24 & 1.45 & -0.197$\pm$0.041 \\
Tr27 & 3 & 117 & 0.92 & -0.9 & 12.3 & 8.2 & -0.10 & 1.80 & 0.166$\pm$0.034 \\
VelOB2 & 1 & 264 & 0.68 & -1.3 & 6.8 & 2.7 & -0.27 & 1.22 & -0.769$\pm$0.135 \\
VulOB4 & 1 & 181 & 0.83 & -1.0 & 10.2 & 4.1 & -0.21 & 1.39 & -0.168$\pm$0.026 \\
\hline
\end{tabular}
\end{center}
\end{minipage}
\end{table*}
\begin{table*}
\begin{minipage}{170mm}
\caption{\label{param_tbl_relall} Table of parameters for all OB associations relative to the centre of all stars. See also the caption of Table \ref{param_tbl}.}
\begin{center}
\begin{tabular}{l c c c c c c c c c}
\hline
assoc. & $N_{\sigma}(XYZ)$ & $N_{\star}$ & $N_{v_{r}>0}/N_{v_{r}<0}$ & med. $v_{\text{r}}$ (km\,s$^{-1}$) & $\sigma_{v_{r}}$ & $<\Delta v>$ (km\,s$^{-1}$) & med. $v_{r} / |v_{\text{t}}|$ & $\sigma_{v_{r} / |v_{\text{t}}|}$ & $\beta$ \\
\hline
CamOB1 & 3 & 260 & 1.26 & 2.5 & 10.5 & 4.0 & 0.31 & 1.47 & -0.089$\pm$0.009 \\
CasOB14 & 3 & 134 & 1.48 & 2.6 & 9.8 & 4.2 & 0.50 & 1.39 & -0.017$\pm$0.003 \\
CepOB2 & 3 & 127 & 1.19 & 0.6 & 7.8 & 4.5 & 0.42 & 1.94 & 0.417$\pm$0.081 \\
CepOB3 & 1 & 279 & 1.47 & 3.1 & 10.3 & 4.6 & 0.71 & 1.93 & 0.369$\pm$0.055 \\
CepOB4 & 1 & 388 & 1.10 & 0.7 & 9.5 & 3.6 & 0.08 & 1.16 & -0.623$\pm$0.076 \\
Coll140 & 1 & 124 & 1.38 & 1.5 & 8.5 & 2.0 & 0.31 & 1.32 & -0.796$\pm$0.123 \\
Coll359 & 3 & 180 & 0.97 & -0.1 & 9.5 & 1.2 & -0.01 & 1.26 & -0.585$\pm$0.082 \\
CygOB7 & 1 & 104 & 0.79 & -0.9 & 7.8 & 4.7 & -0.32 & 1.53 & 0.126$\pm$0.031 \\
CygOB9 & 3 & 314 & 1.12 & 0.5 & 7.2 & 6.7 & 0.06 & 1.39 & -0.315$\pm$0.043 \\
MonOB1 & 3 & 126 & 1.10 & 0.4 & 10.3 & 4.8 & 0.09 & 1.59 & 0.145$\pm$0.021 \\
OriOB1 & 1 & 190 & 1.11 & 0.2 & 3.4 & 1.6 & 0.11 & 1.60 & 0.059$\pm$0.014 \\
PerOB3 & 3 & 136 & 0.86 & -3.9 & 19.1 & 1.1 & -0.38 & 1.58 & 0.077$\pm$0.011 \\
ScoOB2 & 3 & 325 & 1.10 & 0.4 & 9.5 & 0.8 & 0.11 & 1.52 & 0.088$\pm$0.013 \\
SctOB3 & 3 & 100 & 1.22 & 0.6 & 11.8 & 12.7 & 0.07 & 1.57 & -0.014$\pm$0.002 \\
SgrOB1 & 3 & 101 & 1.06 & 1.2 & 9.1 & 10.7 & 0.26 & 1.60 & -0.090$\pm$0.019 \\
Tr27 & 3 & 117 & 1.17 & 1.6 & 12.1 & 8.2 & 0.19 & 1.49 & 0.130$\pm$0.025 \\
VelOB2 & 1 & 264 & 0.80 & -1.0 & 6.3 & 2.7 & -0.14 & 1.17 & -0.739$\pm$0.137 \\
VulOB4 & 1 & 181 & 0.97 & 0.0 & 10.0 & 4.1 & -0.03 & 1.48 & -0.143$\pm$0.023 \\
\hline
\end{tabular}
\end{center}
\end{minipage}
\end{table*}
Once the velocities are in the association centre frame, the radial velocities $v_{\text{r}}$ are determined by taking the dot product of the velocity vector with the radial unit vector. The tangential velocity is then obtained by subtraction as $v_{\text{t}}=(v^2-v_{\text{r}}^2)^{1/2}$.
We define a positive radial velocity as moving away from the centre of the OB association and a negative value as moving towards the centre of the association. Similarly, we (arbitrarily) define the tangential velocity as being positive if the motion is anti-clockwise with respect to the previously determined association centre and negative if the motion is clockwise with respect to the association centre.
\subsection{Model distributions}
\begin{figure*}
\begin{minipage}{170mm}
\begin{center}
\includegraphics[width=0.95\linewidth]{model_case_examples_20171108.eps}
\end{center}
\caption{\label{case_examples} Top: positions of all selected members of Cas OB14 with velocity vectors. Middle row, left to right: Case I model association based on Cas OB14, case II model association based on Cas OB14, and case III model association based on Cas OB14. Bottom row, left to right: Case IV random model association, Case V model association, Case VI model association. Black stars represent original, unaltered position and velocity vectors. Blue stars are those of which the positions and velocities are randomly generated. Red stars are either original or generated stars which have been forced to exhibit randomised directions of motion (case I), or expansion either relative to a local centre (cases II and III) or the association centre (cases V and VI). All positions are in pc. }
\end{minipage}
\end{figure*}
In order to assess whether the observed kinematics are consistent with a net radial expansion, we compare our results for the observed OB associations with those of randomly generated velocity distributions.
Six sets of models will be calculated: two distributions representing random velocity fields, two distributions representing localised expansion within associations, and two distributions representing globally expanding velocity fields. These models are therefore representative of no systematic expansion or contraction, expansion from multiple centres within an association (multiply monolithic), and globalised expansion from a single centre (monolithic).
For reference, we will refer to these six model data sets as cases I, II, III, IV, V, and VI. They are explained below and visualised in Fig. \ref{case_examples}.
The first model set (case I) takes the real positions of the measured OB association data and the magnitude of the motions. The direction of motion is then randomised according to a uniform distribution in azimuthal angle. This process was repeated 20 times to increase the number of renditions yielding a total of 360 associations with randomised directions of motion.
As in the case I models, the case II and III models retain the positions and the velocity magnitudes of the observed OB associations. In case II, the 20 nearest neighbours (in $X$ and $Y$) are selected for each star. We define a local centre as the mean position and velocity of those 20 neighbours, and the radial and tangential components of the stellar velocity are calculated relative to the local centre. One third of the stars within each association are forced to exhibit outwards radial motion with respect to the local centre by taking $v_{\text{r}}$ to be the absolute value of the measured $v_{\text{r}}$. Therefore, this model case represents associations that exhibit localised expansion. In case III we assign the radial velocity to be the absolute of either the radial or tangential velocity components ($v_{\text{r}}=\max{(|v_{\text{r}}|,|v_{\text{t}}|)}$) for the same third of the stars, with the new tangential component being the smaller of the two components. Case III is representative of radially anisotropic localised expansion from multiple centres as is expected from a multiply monolithic formation model.
For the case IV distribution, we use the median and the standard deviations of the position and velocity dispersions of the observed sample of associations to generate a set of random position and velocity dispersions.
These random values were drawn from four Gaussian distributions centred on the median of the measured dispersions in $X$, $Y$, $v_{X}$, and $v_{Y}$ (Table \ref{disp_tbl}) and using the standard deviation of the measured dispersions.
These were subsequently used to populate 300 model OB associations with random numbers of stars drawn from a uniform distribution occupying the range 100--400.
A subset of OB stars is defined within each model association by randomly drawing fractions of OB stars from a Gaussian distribution based on the observed median (0.11) and dispersion (0.10) of the OB star fractions in the observed sample.
Each of the stars is assigned a random position and proper motion, drawn from Gaussian distributions centred on zero using the randomly generated dispersion values.
Cases V and VI are representative of globally expanding configurations of the case IV models.
In case V, one third of the stars within each association are forced to exhibit outwards radial motion with respect to the association centre by taking $v_{\text{r}}$ to be the absolute value of the originally generated $v_{\text{r}}$. On average, this causes $2/3$ of the stars to have positive radial velocities. In case VI, one third of the stars within each association are assigned radial velocities equal to the largest absolute value of the original $v_{\text{r}}$ and $v_{\text{t}}$, i.e.~$v_{\text{r}}=\max{(|v_{\text{r}}|,|v_{\text{t}}|)}$. The tangential velocity is set to the other, smaller component while keeping its original sign. This model ensures that radial motion dominates over tangential motion. This reproduces the strong radial anisotropy ($\beta\sim0.4$) that is expected from a monolithic formation scenario.
\begin{figure*}
\begin{minipage}{170mm}
\begin{center}
\includegraphics[width=0.99\linewidth]{pos_vel_CasOB14_20170719.eps}
\end{center}
\caption{\label{CasOB_exmpl} X and Y positions (in pc) of all selected members in the OB association Cas OB14 with vectors representing the velocities of all sources relative to the association centre determined using all stars (see text). Left: total velocities. Centre: radial velocity components. Right: tangential velocity components. The blue points in the left panel are OB stars with other stars marked in red. In the centre panel, the points are colour-coded according to $v_{r} < 0$ (blue) and $v_{\text{t}} > 0$ (red). In the right panel, blue points are moving clockwise and red points are moving anti-clockwise with respect to the association centre. In each panel, the relative position and velocity vector of the association based on only OB stars is marked in grey.}
\end{minipage}
\end{figure*}
The model cases can be summarised as follows:
\begin{enumerate}
\item Observed stellar positions and absolute velocities with random directions of motion (case I).
\item Observed stellar positions and absolute velocities where $1/3$ of the stars are guaranteed to have positive radial velocities with respect to the mean position and velocity of the 20 nearest neighbours (case II).
\item Observed stellar positions and absolute velocities where $1/3$ of the stars are guaranteed to have positive radial velocities and dominant radial motions with respect to the mean position and velocity of the 20 nearest neighbours (case III).
\item Random stellar positions and velocities drawn from a four-dimensional Gaussian distribution (case IV).
\item Random stellar positions and velocities drawn from a four-dimensional Gaussian distribution where $1/3$ of the stars are guaranteed to have positive radial velocities (case V).
\item Random stellar positions and velocities drawn from a four-dimensional Gaussian distribution where $1/3$ of the stars are guaranteed to have positive radial velocities and dominant radial motions (case VI).
\end{enumerate}
Together, these cases span a physically appropriate range of configurations for comparison to the observed kinematic diagnostics discussed below.
\section{Quantifying the kinematics of OB associations}
\label{results_section}
Through analysis of the radial velocities of the member stars of OB associations, it is possible to determine whether the associations are showing signs of strong radial expansion and, by extension, distinguish between the monolithic or hierarchical models of star formation.
The left panel of Figure \ref{CasOB_exmpl} shows the positions and velocities for the OB association Cas OB14. These velocities are then separated into their radial and tangential components in the middle and right panels of Figure \ref{CasOB_exmpl}. These diagrams are shown for all 18 OB associations in Figure \ref{fig:A1}.
In the left panels of Figures \ref{CasOB_exmpl} and \ref{fig:A1}, the OB stars are marked in blue with all other stars marked in red. In the centre panels, the points are colour-coded according to the radial velocity direction: blue for $v_{\text{r}} < 0$ and red for $v_{\text{r}} > 0$. In the right hand panels the points are colour-coded by tangential motion direction: blue for clockwise ($v_{\text{t}} < 0$) and red for anti-clockwise \textquoteleft rotation' ($v_{\text{t}} > 0$).
In all panels, positions and velocities are shown relative to the association centre defined as the mean position and velocity of all stars. The relative position and velocity vector of the centre defined using only the OB stars is marked in grey in each panel.
The general impression of these maps is one of largely stochastic motion, albeit with some substructure in many cases. In the remainder of this section, we employ a number of quantitative tests by considering the cumulative distributions of several key kinematic diagnostics across the sample of OB associations. This allows us to assess the degree to which their kinematics are best described by a net expansion or by random motion.
\subsection{Number ratios}
\begin{figure*}
\begin{minipage}{165mm}
\begin{center}
\includegraphics[width=0.99\linewidth]{Nrat_20171124.eps}
\caption{\label{Nrat_fig} Cumulative distribution of the ratio of the number of stars with positive radial velocities over those with negative radial velocities relative to the centre of OB stars (left) and the center of all stars (right). The solid black lines are the measured data for the 18 OB associations analysed in this work.
The solid red lines indicate the distributions for the 18 OB associations calculated with real X and Y positions and velocity magnitudes but with randomised directions of motion (case I).
Additional lines indicate the case II distribution (blue dashed), case III (orange dotted), case IV (green dashed), case V (magenta dash-dot) and case VI (maroon dotted) distributions. The case V and case VI distributions are indistinguishable in this figure.
The observed OB associations are best reproduced by the random motion models (case I and IV).}
\end{center}
\end{minipage}
\end{figure*}
If there is a widespread, systematic expansion of present-day OB associations, then it is likely to be evident in the ratio of stars moving away from the association centre to stars moving towards the association centre.
Figure \ref{Nrat_fig} shows the cumulative distribution of the number ratio of sources with positive (outwards) radial velocities ($N_{v_{\text{r}}>0}$) to the number of sources with negative (inward) radial velocities ($N_{v_{\text{r}}<0}$) for all OB associations. Also shown in this Figure are the normalised cumulative distributions for each of the six model cases described in the previous section.
It is immediately clear from this figure that, based on number ratios alone, the observed distribution of OB associations are far better represented by purely random distributions of stellar velocities rather than those distributions representing either locally expanding or globally expanding associations.
There is a departure from the case IV curve tending slightly towards expansion; however, this effect is relatively small and the maximum observed ratio is $\sim$1.5, falling short of what may be expected from a widespread, systematic expansion.
Qualitatively, the case I distributions are the closest approximation to the observed distributions.
When the centres of the OB star populations are used, the case I approach of randomising proper motion directions recreates the departure from the case IV distribution of random positions and velocities. The locally expanding case II and III model distributions fall between the distributions based on random velocity fields and the two globally expanding distributions.
When the centres of the entire stellar populations are used (right panel), the observed systems show a small expansion relative to the random motion models (case I and IV), but this small excess is reproduced by forcing just $\sim$5\% of the stars to exhibit outward radial motion as in cases II, III, V, and VI (rather than the 1/3 used there). This is in strong contrast with the expectation from monolithic models, in which most of the stellar population exhibits expansion \citep{Baumgardt2007}.
Based on these number ratios alone, the observed velocities are certainly consistent with a population of stars with randomly distributed velocities and inconsistent with both global and localised expansion scenarios.
Whilst giving the qualitative impression of a random distribution, the information gained from this simple analysis is limited, providing no indication of the magnitudes of the radial velocities and by itself it is not a definitive measure of the dynamical behaviour of the sample.
\subsection{Median velocity distributions}
To quantify the typical expansion (or contraction) velocities of the OB associations in our sample, we take the median value of the radial velocities of the stellar populations for each OB association.
The cumulative distributions of the median radial velocities are shown in Figure \ref{PMXp_fig} relative to both the centre of the OB stars and the centre of all stars, along with the six model cases.
Uncertainties in the observed distributions are calculated by randomly generating 100 cumulative distributions of the median radial velocities by drawing values from a Gaussian distribution for each OB association, using the median radial velocities and the radial velocity dispersions shown in Tables~\ref{param_tbl} and \ref{param_tbl_relall}. These uncertainties are represented by the grey shaded area in Figure \ref{PMXp_fig}.
It is again immediately clear that, although there is a tendency towards positive (outwards) radial velocities, the observed distribution falls closer to the case I (randomised velocity directions) and case IV (randomised positions and velocities) distributions than to any of the expanding models, regardless of which association centre is used.
Within the uncertainties of the observed distributions (represented by the grey shaded area in Figure \ref{PMXp_fig}), the lower end of the observed distributions are consistent with the case IV random distribution. However, the case I distribution provides a closer match than the case IV distribution over the positive radial velocity region of parameter space.
The extreme positive ends of the distributions partially overlap with the case II, III, and V model distributions; however, the positive parts of the distributions are far more consistent with the case I distribution overall.
Therefore, the tendency towards positive values could be a result of positional substructure in the apparently high $v_{\text{r}}$ associations rather than a genuine radial velocity effect, as the case I distribution retains the original geometry of the observed associations, and therefore any existing positional substructure.
\begin{figure*}
\begin{minipage}{165mm}
\begin{center}
\includegraphics[width=0.99\linewidth]{vr_20171124.eps}
\caption{\label{PMXp_fig} Cumulative distributions showing the median radial velocities for all associations relative to the centre of OB stars (left) and the centre of all stars (right).
The solid black lines are the measured data for the 18 OB associations analysed in this work. Uncertainties in the observed distributions are represented by the grey shaded area.
The solid red lines indicate the distributions for the 18 OB associations calculated with real X and Y positions and velocity magnitudes but with randomised directions of motion (case I).
Additional lines indicate the case II distribution (blue dashed), case III (orange dotted), case IV (green dashed), case V (magenta dash-dot) and case VI (maroon dotted) distributions. The observed OB associations are best reproduced by a combination of the random motion models (cases I and IV).}
\end{center}
\end{minipage}
\end{figure*}
Overall, the median radial velocities in Figure \ref{PMXp_fig} are much smaller than the radial velocity dispersions (see Tables \ref{param_tbl} and \ref{param_tbl_relall}). This indicates that any expansion of the associations
is unlikely to have resulted from an impulsive event (such as gas expulsion) that caused a possible progenitor cluster to become unbound. Instead, these radial velocities are more consistent with the gradual, secular evolution of the nascent velocity field, as inherited from the turbulent ISM.
In Tables \ref{param_tbl} and \ref{param_tbl_relall}, the seventh column shows the mean velocity uncertainty of the members of each association. In two of the 18 OB associations (Sct OB3 and Sgr OB1), the radial velocity dispersions are lower than the average uncertainties. These two associations are at relatively high distances where the velocity uncertainties are dominated by the uncertainties in parallax in the TGAS catalogue.The TGAS parallax uncertainties are likely to be overestimated (see the discussion in Section 4.2) which can account for the measured velocity dispersions being smaller than the average uncertainties.
While the median velocity distributions convincingly rule out systematic rapid ($>3$\,km\,s$^{-1}$) expansion, they fail to take into account any information regarding the tangential component of the stellar velocities.
In Figure \ref{PMXp_PMYp_fig}, the median is taken of the ratio of the radial velocity to the absolute value of the tangential velocity ($v_{\text{r}} / |v_{\text{t}}|$) across all stars in the OB associations. This probes the expansion rates in units of the typical velocity in the tangential direction.
When the tangential velocities are taken into account, the case VI distribution is clearly distinguished from the other expanding distributions (cases II, III, and V), indicating that this is a powerful diagnostic of global, strongly anisotropic expansion.
The overall trends of Figures \ref{PMXp_fig} and \ref{PMXp_PMYp_fig} are very similar. However, the normalised expansion velocities of Figure \ref{PMXp_PMYp_fig} provide a better match to cases I and IV than the absolute expansion velocities in Figure \ref{PMXp_fig}. The upper end of the observed distribution relative to the centre of all stars falls between the randomly distributed case IV and the locally expanding case II distributions but is well characterised by the case I distribution. At its most extreme, the observed distribution is consistent with both locally expanding model distributions. However, this only concerns up to three OB associations out of the 18 analysed in this work.
The radially anisotropic, globally expanding case VI distribution remains entirely inconsistent with observations.
\begin{figure*}
\begin{minipage}{165mm}
\begin{center}
\includegraphics[width=0.99\linewidth]{vr_vt_20171124.eps}
\caption{\label{PMXp_PMYp_fig} Cumulative distributions showing the median $v_{\text{r}}$ / $| v_{\text{t}} |$ values for all associations relative to the centre of OB stars (left) and the center of all stars (right). Lines and symbols have the same meaning as for Figure \ref{PMXp_fig}. Again, the observed OB associations are best reproduced by a combination of the random motion models (cases I and IV).}
\end{center}
\end{minipage}
\end{figure*}
\subsection{Anisotropy}
The radial anisotropy of a system represents the degree to which the radial velocity component dominates over the tangential velocity component. A radially anisotropic velocity field, retained for many initial crossing times (or tens of Myr), is a firm prediction of the monolithic star formation model.
The 2-dimensional anisotropy parameter ($\beta$) determined for the OB associations in this work is directly analogous to the 3-dimensional parameter and takes the form:
\begin{equation}
\label{anieqn}
\beta = 1-\frac{<v_{\text{t}}^{2}>}{<v_{\text{r}}^{2}>} \text{.}
\end{equation}
While the three-dimensional form carries an additional factor of 2 in the denominator due to the double dimensionality of the tangential component, this is omitted when using a 2-dimensional space. A value of zero indicates that the radial and tangential velocity components are equal, while negative values and positive values indicate dominant tangential and radial components, respectively. In a system experiencing a period of systematic expansion (or contraction), $\beta$ is expected to have a positive value. This is a critical quantity for answering whether OB associations are expanded clusters, because $N$-body simulations show that the anisotropy remains systematically high ($\beta > 0.2$) for long periods (tens to hundreds of initial crossing times, or tens of Myr) after gas expulsion \citep{Baumgardt2007}.
The cumulative distributions for the anisotropy parameter are shown in Figure \ref{beta_fig}, using both the centre of OB stars and the centre of all stars.
The model case where we take the absolute $v_{\text{r}}$ value for 1/3 of a set of randomly generated sources (case V) does not affect the anisotropy parameter, because it enters squared in equation (\ref{anieqn}). Only the case III and VI distributions represent strongly radially anisotropic expansion.
As for the previous quantities, the case I distribution fits well to observed data in the positive region of parameter space, whereas the case II, III, IV and V distributions provide a far better fit to the data at negative anisotropies.
Qualitatively, the observed cumulative distributions of $\beta$ are well matched by the case I, II, III, IV, and V distributions, while they are entirely inconsistent with case VI (where the radial velocity component is dominant).
The case III anisotropic localised expansion models fall very close to the case II non-anisotropic locally expanding models; however, there is always a slight anisotropy excess in the case III models with respect to the case II models.
The median of the anisotropy parameters of the OB associations in this sample is approximately zero. The radial component is therefore certainly not dominant, indicating no systematic tendency towards expansion or contraction. Likewise, there is no observed bias towards tangential motion, which implies that the observed stars are not moving in well ordered orbits around a common centre of mass. Therefore, it is clear that, even if none of the model cases presented here fit the observed data well, the observed population of OB associations cannot be produced by singularly monolithic expansion. However, expansion from multiple centres cannot be ruled out by the anisotropy parameter alone. Multiply monolithic cluster formation is ruled out by the combination of quantities discussed previously.
\begin{figure*}
\begin{minipage}{165mm}
\begin{center}
\includegraphics[width=0.99\linewidth]{beta_20171124.eps}
\caption{\label{beta_fig} Cumulative distributions of the anisotropy parameter $\beta$, for all associations relative to the centre of OB stars (left) and the center of all stars (right). Lines and symbols have the same meaning as for Figure \ref{PMXp_fig}. Again, the observed OB associations are best reproduced by a combination of the random motion models (cases I and IV). The anisotropy does not allow a distinction between case IV and case V, but case VI expansion is firmly ruled out.}
\end{center}
\end{minipage}
\end{figure*}
\subsection{Kolmogorov-Smirnov tests}
\begin{table*}
\begin{minipage}{170mm}
\caption{\label{KS_table} Kolmogorov-Smirnov statistics and associated {\it p}-values calculated using the centre of the OB star members for each association. The quantity for which the test is carried out is given in the first column. The following columns show the KS statistic and the corresponding {\it p}-values for each of the model cases.
Case I uses the measured positions and absolute velocities of the stars in the 18 measured OB associations, but with randomised proper motion directions.
Case IV uses randomly generated model associations, with each following a Gaussian distribution in X,Y,$v_x$ and $v_y$. In case V one third of the stars in the case IV associations are forced to be expanding by using the absolute value of $v_{\text{r}}$. In case VI, the velocity vectors of one third of the stars were changed such that the radial velocity is set by the largest absolute value of $v_{\text{r}}$ and $v_{\text{t}}$, ensuring that radial motion is dominant. The final column shows the KS test results for when the case I distribution is used above the threshold for expansion (1 for $N_{v_{\text{r}}>0}/N_{v_{\text{r}}<0}$ and 0 for all other parameters) and the case IV distribution is used at values lower than the threshold.}
\begin{center}
\begin{tabular}{l c c c c c c c c c c c c c c}
\hline
& \multicolumn{2}{c}{case I} & \multicolumn{2}{c}{case II} & \multicolumn{2}{c}{case III} & \multicolumn{2}{c}{case IV} & \multicolumn{2}{c}{case V} & \multicolumn{2}{c}{case VI} & \multicolumn{2}{c}{case I $+$ case IV} \\
Quantity & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$\\
\hline
$N_{v_{\text{r}}>0}/N_{v_{\text{r}}<0}$ & 0.239 & 0.245 & 0.519 & 0.000 & 0.583 & 0.000 & 0.316 & 0.053 & 0.963 & 0.000 & 0.963 & 0.000 & 0.244 & 0.227 \\
$v_{\text{r}}$ & 0.306 & 0.065 & 0.525 & 0.000 & 0.572 & 0.000 & 0.312 & 0.057 & 0.758 & 0.000 & 0.816 & 0.000 & 0.146 & 0.831 \\
$v_{\text{r}}/ |v_{\text{t}}|$ & 0.272 & 0.131 & 0.439 & 0.002 & 0.475 & 0.001 & 0.212 & 0.385 & 0.807 & 0.000 & 0.921 & 0.000 & 0.112 & 0.977 \\
$\beta$ & 0.244 & 0.222 & 0.156 & 0.765 & 0.231 & 0.283 & 0.231 & 0.303 & 0.231 & 0.303 & 0.753 & 0.000 & 0.169 & 0.681 \\
mean & 0.265 & 0.166 & 0.410 & 0.192 & 0.465 & 0.071 & 0.268 & 0.200 & 0.690 & 0.076 & 0.863 & 0.000 & 0.168 & 0.679 \\
\hline
\end{tabular}
\end{center}
\end{minipage}
\end{table*}
\begin{table*}
\begin{minipage}{170mm}
\caption{\label{KS_table_relall} KS statistics and associated {\it p}-values calculated using the centre of all stars for each association. See also the caption of Table \ref{KS_table}.}
\begin{center}
\begin{tabular}{l c c c c c c c c c c c c c c}
\hline
& \multicolumn{2}{c}{case I} & \multicolumn{2}{c}{case II} & \multicolumn{2}{c}{case III} & \multicolumn{2}{c}{case IV} & \multicolumn{2}{c}{case V} & \multicolumn{2}{c}{case VI} & \multicolumn{2}{c}{case I $+$ case IV} \\
Quantity & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$ & KS & $p$ \\
\hline
$N_{v_{\text{r}}>0}/N_{v_{\text{r}}<0}$ & 0.414 & 0.004 & 0.667 & 0.000 & 0.739 & 0.000 & 0.420 & 0.003 & 0.970 & 0.000 & 0.970 & 0.000 & 0.242 & 0.235 \\
$v_{\text{r}}$ & 0.303 & 0.069 & 0.625 & 0.000 & 0.689 & 0.000 & 0.423 & 0.003 & 0.650 & 0.000 & 0.760 & 0.000 & 0.308 & 0.063 \\
$v_{\text{r}}/ |v_{\text{t}}|$ & 0.297 & 0.078 & 0.517 & 0.000 & 0.578 & 0.000 & 0.389 & 0.008 & 0.770 & 0.000& 0.921 & 0.000 & 0.295 & 0.084 \\
$\beta$ & 0.231 & 0.283 & 0.281 & 0.111 & 0.353 & 0.021 & 0.280 & 0.126 & 0.280 & 0.126 & 0.813 & 0.000 & 0.197 & 0.487 \\
mean & 0.311 & 0.109 & 0.523 & 0.028 & 0.590 & 0.005 & 0.378 & 0.035 & 0.668 & 0.032 & 0.866 & 0.000 & 0.261 & 0.217 \\
\hline
\end{tabular}
\end{center}
\end{minipage}
\end{table*}
After qualitatively comparing the kinematic properties of the observed OB associations to our model distributions, we now quantify this comparison further.
In order to assess the likelihood that model and observations are drawn from the same parent samples, we carry out two-sample Kolmogorov--Smirnov (KS) tests between the observed OB association data and each of the model case distributions.
The KS tests return a KS statistic which is a normalised maximum distance between the two cumulative distributions and a {\it p}-value which represents the likelihood that the two distributions were drawn from the same parent sample.
The resulting KS statistics and corresponding {\it p}-values are given in Tables \ref{KS_table} and \ref{KS_table_relall} using the centres of OB stars and the centres of all stars, respectively.
For the number ratio of sources with positive versus negative values of the radial velocity ($N_{v_{\text{r}}>0} / N_{v_{\text{r}}<0}$) relative to the OB star centre, the case I distribution has the lowest KS statistic (0.239) alongside the highest {\it p}-value (0.245). While neither the KS statistic, nor the {\it p}-value indicate a particularly good fit, this indicates that it is the most likely model from which the observed data were drawn out of those tested. The case I distribution and the observed distribution relative to the centre of all stars do not agree as well, with a KS statistic and {\it p}-value of 0.414 and 0.004, respectively. Regardless of which centre is used, the case I distribution is the closest match to the observations and both of the globally expanding configurations (cases V and VI) are highly inconsistent with the observed distribution, exhibiting KS statistics of 0.96 and 0.97 as well as extremely low {\it p}-values. The locally expanding model distributions (cases II and III) are also inconsistent with observations, with KS statistics in the range $0.5-0.8$ and extremely low {\it p}-values. The case IV distribution lies between the case I and the expanding distributions in terms of both KS statistics and {\it p}-values.
For the median radial velocity distributions, the case I distribution is again the best fit to the observed data, with relatively low KS values of $\sim$0.3 and {\it p}-values of $\sim$0.07 regardless of which centre is used. While case IV does not provide as good a fit as case I, it represents a significant improvement on all of the expanding model cases in both KS statistic and {\it p}-value.
While none of the model distributions provide particularly good fits to the observed data, with {\it p}-values of $\sim 0$, all of the expanding model cases are ruled out based on the observed median velocity distributions, regardless of whether the expansion is global or localised.
When the median radial velocities are normalised by the tangential velocities, the KS statistics again suggest that case I and IV are the best fits to the data, with cases V and VI being strongly inconsistent with the observations.
For the anisotropy parameter, the case I, II, IV and V models provide similar fits to the observed distribution, consistently providing KS statistics below 0.3 and {\it p}-values greater than 0.1. Case III exhibits similar KS statistics and {\it p}-values relative to the centre of OB stars but the {\it p}-value is significantly lower (0.021) when the centre is defined as the mean position and velocity of all stars.
Meanwhile, the strongly anisotropic globally expanding case VI results in high KS statistics (0.753 and 0.813) and extremely low {\it p}-values, indicating that case VI is inconsistent with the observed distributions. While none of the model distributions fit the observed data well, only the globally expanding and anisotropic case VI model distribution is conclusively ruled out by the anisotropy parameter.
The mean {\it p}-values of the six model cases in Tables~\ref{KS_table} and~\ref{KS_table_relall} summarise the above discussion, in that case I and IV best reproduce the observed kinematics.\footnote{The mean value of case II, when using the centre of OB stars, is also quite high but this is owing only to the anisotropy parameter.} The globally expanding models of case V and VI are firmly ruled out by the observations. While the locally expanding models are not ruled out by anisotropy, they are firmly ruled out by the remaining three parameters.
Regardless of the parameter being tested, or whether the centre of the OB stars or the centre of all stars is used as the reference frame, the cases that represent simple random distributions (case I and case IV) yield lower KS statistics than the globally expanding scenarios (case V and case VI). This also holds in the case of localised expansion from multiple points (case II and III), except for the anisotropy parameter. This is in good agreement with the qualitative interpretations of Figures \ref{Nrat_fig}--\ref{beta_fig}.
In general, the sample of OB associations presented in this work are therefore best described as close to random velocity fields with no evidence of systematic expansion or contraction.
It is clear from Figures \ref{Nrat_fig}--\ref{beta_fig} that a combination of cases I (at high, expansion-like values) and IV (at low, expansion-averse values) will qualitatively provide the best fit to the data in every case.
In the final columns of Tables \ref{KS_table} and \ref{KS_table_relall}, we present KS test results for a combination of the case I and case IV models where the values from case I are used above the threshold value between expansion and contraction (1 for the $N_{v_{\text{r}}>0}/N_{v_{\text{r}}<0}$ ratio and 0 for all other parameters), and case IV below that threshold.
When these two distributions are combined in such a manner, it provides the lowest KS statistics and highest {\it p}-values for all parameters except for $N_{v_{\text{r}}>0}/N_{v_{\text{r}}<0}$ when the centre is defined as the mean position and velocity of OB-type stars. The case I models retain the original geometries of the observed associations. It is therefore plausible that the departure of the observed distribution from the purely random case IV distribution towards expanding velocity fields is due to a geometric effect such as positional substructure.
\section{Discussion}
In this paper, we have quantified the kinematics of the members of 18 OB associations with high-quality astrometry from the {\it Gaia}-TGAS catalogue of stellar parallaxes and proper motions. The position--position figures with velocity vectors shown in Figure \ref{fig:A1} show that there are no immediately obvious trends in velocity fields or in the radial or tangential components of these velocities. The degree to which these associations can be described as being expanding has been quantified using four key kinematic diagnostics, finding no evidence of systematic rapid expansion of nearby OB associations.
In conclusion, these results show that the observed kinematics of nearby OB associations are inconsistent with the gas-expulsion driven expansion of the singularly and multiply monolithic cluster formation scenarios. While hierarchical star formation is neither ruled out nor confirmed by this work, the elimination of monolithic cluster formation scenarios must favour hierarchical star formation.
\subsection{Implications of this work}
We have used the radial and tangential velocity components in the plane of the sky ($v_{\text{r}}$ and $v_{\text{t}}$, respectively) of the association members to quantify four kinematic properties for each of the associations: the stellar number ratio $N_{v_{\text{r}}>0}/N_{v_{\text{r}}<0}$, the median radial velocity ($v_{\text{r}}$), the median radial velocity normalised by the tangential velocity ($v_{\text{r}}/|v_{\text{t}}|$), and the radial anisotropy parameter ($\beta$).
The cumulative distribution of each of these values over all OB associations is tested against model distributions generated using randomised velocity distributions and expanding distributions (both from a single centre and from multiple centres).
In each of these tests, the measured sample lies much closer to the random velocity field distributions than the distributions resulting from the globally expanding models. The localised expansion models are ruled out in all parameters except for the anisotropy parameter. This is quantified through the use of KS tests performed between each model distribution and its observed counterpart, shown in Tables \ref{KS_table} and \ref{KS_table_relall}. While none of the models provide a particularly good fit to the observed data, it is clear from these tests that the observed distributions in radial velocity and anisotropy are inconsistent with a scenario in which OB associations exhibit any significant net expansion due to having been much more compact in the past. Even though some OB associations do show signs of expansion, these can easily be reconciled with a random velocity distribution when the original association geometries are retained. This is most likely the result of substructure in the $XY$ plane, as this is the only difference between the case I and case IV distributions. There is no widespread, systematic expansion of present-day OB associations.
\citet{Baumgardt2007} find that in the event of early cluster disruption resulting from rapid gas expulsion, stars acquire strongly radially anisotropic velocity fields. This is not seen in the sample of OB associations presented here.
Indeed, no evidence has been found to support the view that OB associations are undergoing dispersal by radial expansion following gas expulsion.
The consistency of the observed velocity fields with random velocity distributions (within the uncertainties) indicates that there is no systematic expansion or contraction of OB associations.
This is inconsistent with a picture of star formation in which high density, gravitationally bound clusters are a fundamental unit of star formation and subsequently expand in response to gas expulsion.
This sample of associations is therefore far more consistent with a hierarchical model of star formation than singularly or multiply monolithic star formation models.
From this, and in light of previous studies of OB associations (e.g. \citealt{Wright2014,Wright2016}), we draw the conclusion that it is unlikely that OB associations are the expanded remnants of dense clusters and that therefore not all stars form in high-density clustered environments. Of the two widely adopted models of association formation (monolithic and hierarchical), our findings favour a hierarchical star formation scenario, in which stars are formed over a scale-free, hierarchically structured continuum of environments. This continuum allows for the in-situ formation of the full range of young stellar populations, from dense clusters at the high-density end, through OB associations at intermediate densities, to relatively isolated star formation at low densities.
\subsection{Caveats and other considerations}
Following the release of the TGAS catalogue, the appropriateness of the large parallax uncertainties presented in the catalogue has been called into question.
The parallax uncertainties in the TGAS catalogue have been inflated with respect to the internally derived formal uncertainties \citep{Lindegren2016} by:
\begin{equation}
\begin{aligned}
\sigma_{\text{TGAS}}(\pi) = \sqrt{[A\sigma_{int}(\pi)]^{2}+\sigma_{0}^{2}} \text{\,\,} \text{\, , \,} (A,\sigma_{0}) = (1.4,0.20 \text{\,mas}) \text{.}
\end{aligned}
\end{equation}
Since the release of the catalogue, a number of studies have shown that these uncertainties (which are conservative by design) are significantly overestimated. By comparing the TGAS parallax measurements with those determined from the RR Lyrae period-luminosity relation, \citet{Gould2016} found that the uncertainties in TGAS are overestimated and suggest that more appropriate values would be $(A,\sigma_{0}) = (1.1,0.10 \text{\,mas})$. By contrast, when comparing the TGAS parallaxes to those derived from the RAVE, \citet{McMillan2017} find that the random uncertainties in the TGAS catalogue are overestimated by $\sim$0.2~mas.
We choose to use the original, conservative uncertainties, because the disagreement between the Gould and McMillan papers shows that the extent to which the uncertainties are overestimated is not uniquely characterised. The overestimation of the parallax uncertainties can account for the small radial velocity dispersions compared to the average velocity uncertainties of the two most distant OB associations in this study (see Section 3.2).
In addition, \citet{McMillan2017} show that for a small region of the sky, the parallaxes in the TGAS catalogue are systematically higher than those determined by the Radial Velocity Experiment (RAVE, \citealt{Steinmetz2006}) by up to 0.4\,mas. At the distance of 1\,kpc this would mean that the distances in the TGAS catalogue could be underestimated by up to $\sim$300\,pc. The affected region lies at approximately $\lambda\sim0$\degree\,and $\beta\sim-60$\degree\,in ecliptic coordinates. We have verified that none of the OB associations in our sample lie in a conservatively-chosen region between $-30$\degree and $30$\degree\,in ecliptic longitude and lower than $-20$\,\degree\,in ecliptic latitude. Therefore our results are unaffected by this discrepancy in parallax.
Substructure and mass segregation are powerful diagnostics of the formation and past evolution of associations, which in turn can be key in distinguishing between different models of star formation \citep{Parker2014,Wright2014}. While we speculate that positional substructure retained in the case I models may explain the better fit to observational data than the purely random case IV models, kinematic and positional substructure has not been directly addressed in this study. This is because the limited number of OB association members that have been selected makes any detailed analysis of such small structures highly uncertain. It is anticipated that future {\it Gaia} data releases will allow high precision, kinematic studies of individual substructures and clusters within OB associations.
\citet{Melnik2017} have also made use of the TGAS catalogue to study the kinematics of OB associations, finding that their sample of observed OB associations are globally unbound. Their result does not fundamentally conflict with the results of this paper. An obvious potential point of contention is the high expansion velocities found for Per OB1 and Car OB1 of $6.3~{\rm km}~{\rm s}^{-1}$. These associations were not deemed to contain a sufficient number of stars for our study after the membership selection criteria imposed in Section 2, so we cannot quantitatively comment on these velocities directly. However, our findings do suggest that random velocity fields can easily reproduce apparent median expansion velocities of at least 3\,km\,s$^{-1}$ when original association geometries are retained, including any possible substructure.
It should also be noted that \citet{Melnik2017} find velocity dispersions of the order of 4\,km\,s$^{-1}$, around half the average radial velocity dispersion of our larger sample of association members (9\,km\,s$^{-1}$). This could be related to the smaller samples of association members used in \citet{Melnik2017}, making all averages and dispersion measurements more susceptible to outliers.
In view of the above considerations the finding of expanding OB associations from \citet{Melnik2017} should be treated with some caution, because the physical interpretation of the adopted metrics is not unambiguous.
The sample of associations analysed in this study is not complete, but it does represent the subset that can be studied up to the maximum considered distance of 1.6\,kpc at the desirable precision using the available data from DR1. Moreover, due to the relatively high uncertainties in distances derived from TGAS parallaxes (compared to the upcoming {\it Gaia} DR2), it is highly likely that the membership selection is both incomplete and has a number of contaminants.
The upcoming {\it Gaia} data releases will allow a study of parallaxes and proper motions for OB association members derived entirely from {\it Gaia} data, allowing for a significant improvement in precision. This will allow for an unbiased approach to selecting OB associations without the prior selection criterion of an OB association appearing in the literature. Moreover, the improved precision of distances derived from parallaxes in future {\it Gaia} data releases will increase the maximum distance to which the techniques employed in this work can be applied, and increase the number of successfully selected association members.
With the combination of {\it Gaia} DR2 and line-of-sight velocities from large scale radial velocity surveys, we anticipate a future high precision six-dimensional study of a significantly larger sample than of the OB associations presented here.
\section{Conclusions}
In this paper, we quantify the kinematics of 18 nearby OB associations using the {\it Gaia} TGAS catalogue for which we could identify at least 100 stellar members. The goal of this analysis is to test whether these associations are the relics of single or multiple expanded compact clusters. We measure four key parameters: the stellar number ratio $N_{v_{\text{r}}>0}/N_{v_{\text{r}}<0}$, the median radial velocity ($v_{\text{r}}$), the median radial velocity normalised by the tangential velocity ($v_{\text{r}}/|v_{\text{t}}|$), and the radial anisotropy parameter ($\beta$). We compare the observed cumulative distributions of these parameters with those of model OB associations that are representative of random velocity fields, expansion from a single point, and expansion from multiple points. A systematic global expansion of OB associations from a single point, as would be expected from a singularly monolithic cluster formation scenario, is firmly ruled out by each of the four parameters measured in this study. This is the case regardless of whether the expansion is anisotropic or non-anisotropic. Likewise, a multiply monolithic cluster formation model, in which expansion occurs from many points, is ruled out by the number ratio, median radial velocity, and normalised median radial velocity parameters. Within the uncertainties, the observed associations are largely consistent with the model distributions that represent randomised motions of stars with no systematic expansion or contraction. Where small net expansions ($<3$\,km\,s$^{-1}$) are present, these can easily be reconciled with random velocity fields when retaining the geometric configurations of the observed associations, which preserve any positional substructure.
Based on the above results, we infer that OB associations most likely formed as large-scale systems, with hierarchical structures similar to those of molecular clouds. Therefore, we must conclude that only a subset of stars form in gravitationally-bound clusters.
Given that, in the solar neighbourhood, most stars older than a few Myr reside in associations \citep{Blaauw1964,LadaLada2003}, this implies that a minority of all stars form in bound clusters under solar neighbourhood conditions.
The perspectives for the continuation of our analysis using the diagnostic power of future {\it Gaia} releases are highly encouraging. By using only 18 OB associations from the first Gaia data release, the present study already conclusively rules out both the singularly and multiply monolithic cluster formation scenarios where all OB associations are the expanded relics of either single or multiple dense clusters following their expansion due to gas expulsion. Instead, it favours a model in which OB associations originate in-situ and inherit the structure and dynamics of the hierarchical ISM. Given that many more stars are found in OB associations than in dense clusters (e.g. \citealt{LadaLada2003}), this immediately implies that most stars do not form in clusters in the solar neighbourhood. While this work has expanded the handful of OB associations studied this way in the literature to about 20, future Gaia releases will enable systematic surveys of hundreds of OB associations. The next couple of years promise to mark great progress in our understanding of spatially clustered star formation.
\section*{Acknowledgements}
JLW and JMDK acknowledge support from Sonderforschungsbereich SFB 881 \textquotedblleft The Milky Way System\textquotedblright\, (subproject P1) of the German Research Foundation (DFG).
JMDK gratefully acknowledges funding from the DFG in the form of an Emmy Noether Research Group (grant number KR4801/1-1, PI Kruijssen), from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation programme via the ERC Starting Grant MUSTANG (grant agreement number 714907, PI Kruijssen). JLW and JMDK are grateful to the participants of the SFB881 Workshop in Kloster Sch\"{o}ntal in April 2017, where the idea for this paper was conceived.
We thank Hans-Walter Rix for stimulating discussions during the early stages of this work.
We also thank Adam Ginsburg and Alvaro Hacar for their comments in the later stages of the project, which significantly improved this paper.
This work has made use of data from the European Space Agency (ESA)
mission {\it Gaia} (\url{https://www.cosmos.esa.int/gaia}), processed by
the {\it Gaia} Data Processing and Analysis Consortium (DPAC,
\url{https://www.cosmos.esa.int/web/gaia/dpac/consortium}). Funding
for the DPAC has been provided by national institutions, in particular
the institutions participating in the {\it Gaia} Multilateral Agreement.
This research has made use of the SIMBAD database,
operated at CDS, Strasbourg, France.
\bibliographystyle{mnras}
|
\section{Introduction}
Mutual information is a fundamental quantity for measuring the relationship between random variables. In data science it
has found applications in a wide range of domains and tasks, including biomedical sciences \citep{maes1997multimodality}, blind source separation~\citep[BSS, e.g., independent component analysis,][]{hyvarinen2004independent}, information bottleneck~\citep[IB,][]{tishby2000information}, feature selection~\citep{kwak2002input, peng2005feature}, and causality~\citep{butte2000mutual}.
Put simply, mutual information quantifies the dependence of two random variables $X$ and $Z$. It has the form,
\begin{align}
I(X; Z) = \int_{\mathcal{X} \times \mathcal{Z}} \log{\frac{d\PJ{X}{Z}}{d\PI{X}{Z}}} d\PJ{X}{Z},
\label{eq:mutual_information}
\end{align}
where $\PJ{X}{Z}$ is the joint probability distribution, and $\PX =
\int_{\mathcal{Z}} d\PJ{X}{Z}$ and $\PZ = \int_{\mathcal{X}} d\PJ{X}{Z}$ are
the marginals. In contrast to correlation, mutual information captures non-linear statistical dependencies between variables, and thus can act as a measure of true dependence~\citep{kinney2014equitability}.
\begin{comment}
In contrast to correlation, mutual information captures non-linear statistical dependencies between variables, and thus can act as a measure of true dependence.
Put simply, mutual information is the shared information of two random variables, $X$ and $Z$, defined on the same probability space, $(\mathcal{X} \times \mathcal{Z}, \mathcal{F})$, where $\mathcal{X} \times \mathcal{Z}$ is the domain over both variables (such as $\mathbb{R}^m \times \mathbb{R}^n$), and $\mathcal{F}$ is the set of all possible outcomes over both variables. The mutual information has the form\footnote{We assume the convention that $\log$ is the natural log, so that our units of information are in \emph{nats}.}:
\begin{align}
I(X; Z) = \int_{\mathcal{X} \times \mathcal{Z}} \log{\frac{d\PJ{X}{Z}}{d\PI{X}{Z}}} d\PJ{X}{Z}
\label{eq:mutual_information}
\end{align}
where $\PJ{X}{Z} : \mathcal{F} \rightarrow [0, 1]$ is a probabilistic measure (commonly
known as a joint probability distribution in this context), and $\PX =
\int_{\mathcal{Z}} d\PJ{X}{Z}$ and $\PZ = \int_{ \mathcal{X}} d\PJ{X}{Z}$ are
the marginals.
\end{comment}
Despite being a pivotal quantity across data science, mutual information has historically been difficult to compute~\citep{paninski2003estimation}.
Exact computation is only tractable for discrete variables (as the sum can be computed exactly), or for a limited family of problems where the probability distributions are known.
For more general problems, this is not possible.
Common approaches are non-parametric~\citep[e.g., binning, likelihood-ratio estimators based on support
vector machines, non-parametric kernel-density estimators; see,][]{fraser1986independent, darbellay1999estimation, suzuki2008approximating, kwak2002input, moon1995estimation, kraskov2004estimating}, or rely on approximate gaussianity of data distribution~\citep[e.g., Edgeworth expansion,][]{van2005edgeworth}.
Unfortunately, these estimators typically do not scale well with sample size or dimension~\citep{Gao2014}, and thus cannot be said to be general-purpose.
Other recent works include \citet{Kandamay2017MI, Singh2016MI, Moon2017MI}.
In order to achieve a general-purpose estimator, we rely on the well-known characterization of the mutual information as the Kullback-Leibler (KL-) divergence~\citep{kullback1997information} between the joint distribution and the product of the marginals (i.e., $I(X; Z) = \KL{\PJ{X}{Z}}{\PI{X}{Z}}$).
Recent work uses a dual formulation to cast the estimation of $f$-divergences~\citep[including the KL-divergence, see][]{nguyen2010estimating} as part of an adversarial game between competing deep neural networks~\citep{nowozin2016f}. This approach is at the cornerstone of generative adversarial networks~\citep[GANs, ][]{goodfellow2014generative}, which train a generative model without any explicit assumptions about the underlying distribution of the data.
In this paper we demonstrate that exploiting dual optimization to estimate divergences goes beyond the minimax objective as formalized in GANs.
We leverage this strategy to offer a general-purpose parametric neural estimator of mutual information based on dual representations of the KL-divergence~\citep{ruderman2012tighter}, which we show is valuable in settings that do not necessarily involve an adversarial game.
Our estimator is scalable, flexible, and completely trainable via back-propagation.
The contributions of this paper are as follows:
\begin{itemize}
\item We introduce the Mutual Information Neural Estimator (MINE), which is scalable, flexible, and completely trainable via back-prop, as well as provide a thorough theoretical analysis.
\item We show that the utility of this estimator transcends the minimax objective as formalized in GANs, such that it can be used in mutual information estimation, maximization, and minimization.
\item We apply MINE to palliate mode-dropping in GANs and to improve reconstructions and inference in Adversarially Learned Inference~\citep[ALI,][]{dumoulin2016adversarially} on large scale datasets.
\item We use MINE to apply the Information
Bottleneck method~\cite{tishby2000information} in a continuous setting, and show that this approach outperforms variational bottleneck methods~\citep{Alemi2016deep}.
\end{itemize}
\section{Background}
\subsection{Mutual Information}
Mutual information is a Shannon entropy-based measure of dependence between random variables.
The mutual information between $X$ and $Z$ can be understood as the decrease of the uncertainty in $X$ given $Z$:
\begin{equation}
I(X;Z) := H(X) - H(X \mid Z),
\end{equation}
where $H$ is the Shannon entropy, and $H(X\, |\, Z)$ is the conditional entropy of $Z$ given $X$.
As stated in \Eq~\ref{eq:mutual_information} and the discussion above, the mutual information is equivalent to the Kullback-Leibler (KL-) divergence between the joint, $\PJ{X}{Z}$, and the product of the marginals $\PI{X}{Z}$:
\begin{equation} \label{MIdiv}
I(X,Z) = \KL{\PJ{X}{Z}}{\PI{X}{Z}},
\end{equation}
where $D_{KL}$ is defined as\footnote{Although the discussion is more general, we can think of $\PP$ and $\QQ$ as being distributions on some compact domain $\Omega \subset \RR^d$, with density $p$ and $q$ respect the Lebesgue measure $\lambda$, so that $D_{KL} = \int p \log \frac{p}{q} d\lambda$.},
\begin{align} \label{KL_div}
\KL{\PP}{\QQ} :=
\EE_{\PP}\left[ \log\frac{d\PP}{d\QQ}\right].
\end{align}
whenever $\PP$ is absolutely continuous with respect to $\QQ$\footnote{and infinity otherwise.
}.
The intuitive meaning of \Eq~\ref{MIdiv} is clear: the larger the divergence between the joint and the product of the marginals, the stronger the dependence between $X$ and $Z$. This divergence, hence the mutual information, vanishes for fully independent variables.
\begin{comment}
There is also a strong connection between the mutual information and the structure between random variables. \footnote{We briefly touch upon this subject in Appendix~\ref{ssec:on_structure}.}
\end{comment}
\subsection{Dual representations of the KL-divergence.}
A key technical ingredient of MINE are {\it dual representations} of the KL-divergence.
We will primarily work with the Donsker-Varadhan representation~\citep{DonskerVaradhan}, which results in a tighter estimator; but will also consider the dual $f$-divergence representation~\citep{keziou2003fdivergence,nguyen2010estimating, nowozin2016f}.
\begin{comment}
The KL-divergence between two probability distributions $\PP$ and $\QQ$ on a measure space $\Omega$, with $\PP$ absolutely continuous with respect to $\QQ$ (i.e., $\PP \ll \QQ$), is defined as
\begin{align} \label{KL_div}
\KL{\PP}{\QQ} := \int_{\Omega} \log\left(\frac{d\PP}{d\QQ}\right) d\PP = \EE_{\PP}\left[ \log\frac{d\PP}{d\QQ}\right],
\end{align}
where the argument of the log is the density ratio\footnote{Although the discussion is more general, we can think of $\PP$ and $\QQ$ as being distributions on some compact domain $\Omega \subset \RR^d$, with density $p$ and $q$ respect the Lebesgue measure $\lambda$, so that $D_{KL} = \int p \log \frac{p}{q} d\lambda$.} and $\EE_{\PP}$ denotes the expectation with respect to $\PP$. It follows from Jensen's inequality that the KL-divergence is always non-negative
and vanishes if and only if $\PP = \QQ$.
\end{comment}
\paragraph{The Donsker-Varadhan representation.}
The following theorem gives a representation of the KL-divergence ~\citep{DonskerVaradhan}:
\begin{theo}[Donsker-Varadhan representation] \label{DVtheorem}
The KL divergence admits the following dual representation:
\begin{align}
\label{eq:donsker}
\KL{\PP}{\QQ} = \sup_{T : \Omega \to \RR} \EE_{\PP}[T] - \log(\EE_{\QQ}[e^{T}]),
\end{align}
where the supremum is taken over all functions $T$ such that the two expectations are finite. \end{theo}
\proof{See the Supplementary Material.}
\begin{comment}
\ab{}{move proof to Supplementary material}
\begin{proof} A simple proof
goes as follows. For a given function $T$, consider the Gibbs distribution $\mathbb{G}$ defined by $d \mathbb{G} = \frac{1}{Z} e^T d\QQ$, where $Z = \EE_{\QQ}[e^{T}]$. By construction,
\begin{equation} \label{Gibbsequ}
\EE_{\PP}[T] - \log Z = \EE_{\PP} \left[\log \frac{d\mathbb{G}}{d\QQ}\right]
\end{equation}
Let $\Delta$ be the gap between the two sides of Equ \ref{eq:donsker}:
\begin{equation} \Delta:= \KL{\PP}{\QQ} - \left(\EE_{\PP}[T] - \log(\EE_{\QQ}[e^{T}])\right)
\end{equation}
Using Eqn \ref{Gibbsequ}, we can write $\Delta$ as a KL-divergence:
\begin{align} \label{gap}
\Delta &= \EE_{\PP}\left[ \log\frac{d\PP}{d\QQ} - \log \frac{d\mathbb{G}}{d\QQ}\right] = \EE_{\PP} \log\frac{d\PP}{d\mathbb{G}} \nonumber \\
&= \KL{\PP}{\mathbb{G}} \geq 0
\end{align}
The positivity of the KL-divergence gives $\Delta \geq 0$. We have thus shown that for any $T$,
\begin{equation}
\KL{\PP}{\QQ} \geq \EE_{\PP}[T] - \log(\EE_{\QQ}[e^{T}])
\end{equation}
and the inequality is preserved upon taking the supremum over the right-hand side.
Finally, the identity (\ref{gap}) also shows that this bound is {\it tight} whenever $\mathbb{G} = \PP$, namely for an optimal function $T^\ast$ that takes the form
\begin{equation}
T^\ast = \log \frac{d\PP}{d\QQ} + C
\end{equation}
for some constant $C \in \R$.
\end{proof}
\end{comment}
A straightforward consequence of Theorem \ref{DVtheorem} is as follows. Let $\mathcal{F}$ be {\it any} class of functions $T : \Omega \to \RR$ satisfying the integrability constraints of the theorem.
We then have the lower-bound\footnote{The bound in \Eq~\ref{eq:donskerbound} is known as the {\it compression lemma} in the PAC-Bayes literature ~\citep{Banerjee}.}:
\begin{equation}
\label{eq:donskerbound}
\KL{\PP}{\QQ} \geq \sup_{T \in \mathcal{F}} \EE_{\PP}[T] - \log(\EE_{\QQ}[e^{T}]).
\end{equation}
Note also that the bound is {\it tight} for optimal functions $T^\ast$ that relate the distributions to the \emph{Gibbs density} as,
\begin{equation} d\PP = \frac{1}{Z} e^{T^\ast} d\QQ, \,\, \mbox{where} \,\, Z = \EE_{\QQ}[e^{T^\ast}].
\end{equation}
\paragraph{The $f$-divergence representation.}
It is worthwhile to compare the Donsker-Varadhan representation to the $f$-divergence representation proposed in \citet{nguyen2010estimating, nowozin2016f}, which leads
to the following bound:
\begin{equation} \label{eq:nguyenKLbound}
\KL{\PP}{\QQ} \geq \sup_{T \in \mathcal{F}} \EE_{\PP}[T] - \EE_{\QQ}[e^{T-1}].
\end{equation}
Although the bounds in \Eqs~\ref{eq:donskerbound} and~\ref{eq:nguyenKLbound} are tight for sufficiently large families $\mathcal{F}$, the Donsker-Varadhan bound is {\it stronger} in the sense that, for any fixed $T$, the right hand side of \Eq~\ref{eq:donskerbound} is larger\footnote{To see this, just apply the identity $x \geq e \log x$ with $x = \EE_\QQ[e^T]$.}
than the right hand side of \Eq~\ref{eq:nguyenKLbound}.
We refer to the work by ~\citet{ruderman2012tighter} for a derivation of both representations in \Eqs~\ref{eq:donskerbound} and~\ref{eq:nguyenKLbound} from the unifying perspective of Fenchel duality.
In Section \ref{MItheo} we discuss versions of MINE based on these two representations, and numerical comparisons are performed in Section~\ref{Sec:MIestimate}.
\section{The Mutual Information Neural Estimator}
\label{MItheo}
In this section we formulate the framework of the Mutual Information Neural Estimator (MINE).
We define MINE and present a theoretical analysis of its consistency and convergence properties.
\subsection{Method}
Using both \Eq~\ref{MIdiv} for the mutual information and the dual representation of the KL-divergence, the idea is to choose $\mathcal{F}$ to be the family of functions $T_\theta :\mathcal{X} \times \mathcal{Z} \to \R$ parametrized by a deep neural network with parameters $\theta \in \Theta$. We call this network the {\it statistics network}.
We exploit the bound:
\begin{equation} \label{MIbound}
I(X; Z) \geq I_{\Theta}(X,Z),
\end{equation}
where $I_{\Theta}(X,Z)$ is the {\it neural information measure} defined as
\begin{equation} \label{Fdiv}
I_{\Theta}(X,Z) = \sup_{\theta\in\Theta} \EE_{\PJ{X}{Z}}[T_\theta] - \log(\EE_{\PI{X}{Z}}[e^{T_\theta}]).
\end{equation}
The expectations in \Eq~\ref{Fdiv} are estimated using empirical samples\footnote{Note that samples $\bar{x} \sim \mathbb{P}_X$ and $\bar{z} \sim \mathbb{P}_Z$ from the marginals are obtained by simply dropping $x, z$ from samples $(\bar{x}, z)$ and $(x, \bar{z}) \sim \PJ{X}{Z}$.} from $\PJ{X}{Z}$ and $\PI{X}{Z}$ or by shuffling the samples from the joint distribution along the batch axis. The objective can be maximized by gradient ascent.
It should be noted that \Eq~\ref{Fdiv} actually {\it defines} a new class information measures,
The expressive power of neural network insures that they can approximate the mutual information with arbitrary accuracy.
In what follows, given a distribution $\PP$, we denote by $\hat \PP^{(n)}$ as the empirical distribution associated to $n$ {\it i.i.d.} samples.
\begin{definition}[Mutual Information Neural Estimator (MINE)]
\label{DefMINE}
Let $\mathcal{F}=\{T_{\theta}\}_{\theta\in \Theta}$ be the set of functions parametrized by a neural network. MINE is defined as,
\begin{equation} \label{donskeremp}
\widehat{I(X;Z)}_n =
\sup_{\theta \in \Theta} \EE_{\PP^{(n)}_{XZ}}[T_\theta] - \log(\EE_{\PP^{(n)}_{X} \otimes \hat \PP^{(n)}_{Z}}[e^{T_\theta}]).
\end{equation}
\end{definition}
\begin{algorithm}[ht]
\begin{algorithmic}
\STATE $\theta \gets \text{initialize network parameters}$
\REPEAT
\STATE Draw $b$ minibatch samples from the joint distribution:
\STATE $(\bm{x}^{(1)}, \bm{z}^{(1)}), \ldots, (\bm{x}^{(b)}, \bm{z}^{(b)}) \sim \PJ{X}{Z}$
\STATE Draw $n$ samples from the $Z$ marginal distribution:
\STATE $\bar{\bm{z}}^{(1)}, \ldots, \bar{\bm{z}}^{(b)} \sim \PP_{Z}$
\STATE Evaluate the lower-bound:
\STATE \hspace{-3mm} {\footnotesize $\mathcal{V}(\theta) \gets \frac{1}{b} \sum_{i=1}^{b}\FGen_{\theta}(\bm{x}^{(i)},
\bm{z}^{(i)}) - \log(\frac{1}{b} \sum_{i=1}^{b} e^{\FGen_{\theta}(\bm{x}^{(i)},
\bar{\bm{z}}^{(i)})})$}
\STATE Evaluate bias corrected gradients (e.g., moving average):
\STATE $\widehat{G}(\theta) \gets \widetilde{\nabla}_{\theta}\mathcal{V}(\theta)$
\STATE Update the statistics network parameters:
\STATE $\theta \gets \theta + \widehat{G}(\theta)$
\UNTIL{convergence}
\end{algorithmic}
\caption{MINE \label{alg:mi_donsker_estimation}}
\end{algorithm}
Details on the implementation of MINE are provided in Algorithm~\ref{alg:mi_donsker_estimation}.
An analogous definition and algorithm also hold for the $f$-divergence formulation in \Eq~\ref{eq:nguyenKLbound}, which we refer to as MINE-$f$.
Since \Eq~\ref{eq:nguyenKLbound} lower-bounds \Eq~\ref{eq:donskerbound}, it generally leads to a {\it looser} estimator of the mutual information, and numerical comparisons of MINE with MINE-$f$ can be found in Section~\ref{Sec:MIestimate}.
However, in a mini-batch setting, the SGD gradients of MINE are biased. We address this in the next section.
\subsection{Correcting the bias from the stochastic gradients}
A naive application of stochastic gradient estimation leads to the gradient estimate:
\begin{equation}
\widehat{G}_B = \EE_{B}[\nabla_\theta T_\theta] -
\frac{\EE_{B}[\nabla_\theta T_\theta \, e^{T_\theta}]}{\EE_B \, [e^{T_{\theta}}]}.
\end{equation}
where, in the second term, the expectations are over the samples of a minibatch $B$, leads to a biased estimate of the full batch gradient\footnote{From the optimization point of view, the $f$-divergence formulation has the advantage of making the use of SGD with unbiased gradients straightforward.}.
Fortunately, the bias
can be reduced by replacing the estimate in the denominator by an exponential moving average. For small learning rates, this improved MINE gradient estimator can be made to have arbitrarily small bias.\\
We found in our experiments that this improves all-around performance of MINE.
\subsection{Theoretical properties}
\label{theo_analysis}
In this section we analyze the consistency and convergence properties of MINE. All the proofs can be found in the Supplementary Material.
\begin{comment}
In a nutshell, we show that:
$(i)$ For a large enough statistics networks, the quantity
\begin{equation}
I_\mathcal{F}(X,Z) := D_{\mathcal{F}}(\PJ{X}{Z}\mid\mid \PI{X}{Z})
\end{equation}
can approximate $I(X,Z)$ with arbitrary accuracy.
$(ii)$ For a given statistics network, the empirical estimations
\begin{equation}
\widehat{I(X;Z)}_n = D_{\mathcal{F}}( \hat \PP^{(n)}_{XZ} \mid\mid \hat \PP^{(n)}_{X} \otimes \hat \PP^{(n)}_{Z})
\end{equation}
converge to $I_\mathcal{F}(X,Z)$ almost surely for the choice of samples. \ab{}{add sample complexity}
\end{comment}
\subsubsection{Consistency}
\label{consistency}
MINE relies on a choice of $(i)$ a statistics network and $(ii)$ $n$ samples from the data distribution $\PP_{XZ}$.
\begin{definition}[Strong consistency] \label{def-consistency}
The estimator $\widehat{I(X;Z)}_n$ is strongly consistent if for all $\epsilon >0$, there exists a positive integer $N$
and a choice of statistics network such that:
\[
\vspace{-1mm}
\forall n\geq N, \quad | I(X, Z)-\widehat{I(X;Z)}_n | \leq \epsilon, \, a.e.
\]
where the probability is over a set of samples.
\end{definition}
In a nutshell, the question of consistency is divided into two problems: an {\it approximation} problem related to the size of the family, $\mathcal{F}$, and an {\it estimation} problem related to the use of empirical measures.
The first problem is addressed by universal approximation theorems for neural networks~\citep{Hornik1989approxtheorem}.
For the second problem, classical consistency theorems for extremum estimators apply~\citep{deGeer2006Mestimators} under mild conditions on the parameter space.
This leads to the two lemmas below.
The first lemma states that the neural information measures $I_{\Theta}(X,Z)$, defined in \Eq~\ref{Fdiv}, can approximate the mutual information with arbitrary accuracy:
\begin{lemma}[approximation] \label{lemma:approximation}
Let $\epsilon >0$. There exists a neural network parametrizing functions $T_\theta$ with parameters $\theta$ in some compact domain $\Theta \subset \R^k$, such that
\[
|I(X, Z) - I_\Theta(X,Z)| \leq \epsilon, \, a.e.
\]
\end{lemma}
The second lemma states the almost sure convergence of MINE to a neural information measure as the number of samples goes to infinity:
\begin{lemma}[estimation] \label{lemma:estimation}
Let $\epsilon >0$. Given a family of neural network functions $T_{\theta}$ with parameters $\theta$ in some bounded domain $\Theta \subset \R^k$, there exists an $N \in \mathbb{N}$, such that
\begin{equation} \label{eps-estim}
\forall n\geq N, \quad \mid \widehat{I(X;Z)}_n - I_\Theta(X,Z) \mid \leq \epsilon, \, a.e.
\end{equation}
\end{lemma}
Combining the two lemmas with the triangular inequality, we have,
\begin{theo} \label{theo:consistency}
MINE is strongly consistent.
\end{theo}
\begin{comment}
\begin{proof}
Let $ \epsilon > 0$. We apply the two Lemmas to find a a family of neural network function $\mathcal{F}$ and $N \in \NN$ such that (\ref{eps-estim}) and (\ref{eps-approx}) hold with $\eta = \epsilon/2$.
By the triangular inequality, for all $n\geq N$ and with probability one, we have:
\begin{align}
| I(X, Z)-\widehat{I(X;Z)}_n | \quad \leq \quad &|I(X, Z) - \sup_{T_\theta \in \mathcal{F}} \hat{I}(T_\theta)| + \nonumber \\ &|\widehat{I(X;Z)}_n - I_\mathcal{F}(X,Z) | \nonumber \\ \leq \quad &\epsilon \nonumber
\end{align}
which proves consistency.
\end{proof}
\end{comment}
\subsubsection{Sample complexity}
\label{samp_complexity}
In this section we discuss the {\it sample complexity} of our estimator.
Since the focus here is on the empirical estimation problem, we assume that the mutual information is well enough approximated by the neural information measure $I_\Theta(X, Z)$.
The theorem below is a refinement of Lemma~\ref{lemma:estimation}: it gives how many samples we need for an empirical estimation of the neural information measure at a given accuracy and with high confidence.
We make the following assumptions: the functions $T_{\theta}$ are $M$-bounded (i.e., $|T_\theta|\leq M$) and $L$-Lipschitz with respect to the parameters $\theta$. The domain $\Theta \subset \R^d$ is bounded, so that $\|\theta\| \leq K$ for some constant $K$. The theorem below shows a sample complexity of $\widetilde{O}\left(\frac{d \log d}{\epsilon^2}\right)$, where $d$ is the dimension of the parameter space.
\begin{theo} \label{theo:rate}
Given any values $\epsilon,\delta$ of the desired accuracy and confidence parameters, we have,
\begin{equation}
\mathrm{Pr}\left(| \widehat{I(X;Z)}_n - I_\Theta(X, Z)| \leq \epsilon \right) \geq 1- \delta,
\end{equation}
whenever the number $n$ of samples satisfies
\begin{equation}
n \geq \frac{ 2M^2 (d\log (16K L \sqrt{d} /\epsilon) + 2d M + \log(2/\delta))}{\epsilon^2}.
\end{equation}
\end{theo}
\begin{comment}
\begin{itemize}
\item We present in a proposition the optimal statistics network for both the Donsker and f-divergence representations
\item We show why the Donsker Representation is tighter that the f-divergence one by understanding f-divergences as convex functional acting on the space of all likelihood ratios with respect to $\QQ$.
\item We raise the following question: Which restrictions of the space of likelihood ratio should we consider? (Donsker is f-divergence restricted to all likelihood ratio integrable to 1. What if we require the energy of the likelihood ratio to be finite? What if we add restrictions based on the spectrum of the likelihood ratio.)
\end{itemize}
We close this section by pointing out that the previous construction can be extended to more general information measures based on so-called $f$-divergences~\citep{Ali1966f-div}:
\begin{align} \label{eq:f-div}
\FD{\PP}{\QQ} := \int_{\Omega} f\left(\frac{d\PP}{d\QQ}\right) d\QQ
\end{align}
indexed by a convex
function
$f\colon[0, \infty) \rightarrow \RR$ such that $f(1) = 0$. The KL-divergence is a special case of $f$-divergence with $f(u) = u\log(u)$.
Just as the mutual information can be understood as the KL-divergence between the joint and product of marginals distributions, we can define a family of $f$-information measures as $f$-divergences:
\begin{equation}
I_{f}(X; Z) := \FD{\PJ{X}{Z}}{\PI{X}{Z}}
\end{equation}
An analogue for $f$-divergences of the Donsker-Varadhan representation of Theorem \ref{DVtheorem} can be found in ~\citet{ruderman2012tighter}. The key idea is to express $f$-divergences in terms of convex operators, and to leverage Fenchel-Legendre duality to obtain variational representation in terms of the convex conjugate~\citep{Rockafellarconvexduality}.
This allows a straightforward extension of MINE to a mutual $f$-information estimator, following the construction of the previous section.
The study of such information measures and their estimators is left for future work.
\end{comment}
\section{Empirical comparisons}
\label{Sec:MIestimate}
Before diving into applications, we perform some simple empirical evaluation and comparisons of MINE. The objective is to show that MINE is effectively able to estimate mutual information and account for non-linear dependence.
\subsection{Comparing MINE to non-parametric estimation}
We compare MINE and MINE-$f$ to the $k$-NN-based non-parametric estimator found in \citet{kraskov2004estimating}.
In our experiment, we consider multivariate Gaussian random
variables, $X_a$ and $X_b$, with componentwise correlation, $corr(X^i_a, X^j_b) = \delta_{ij} \, \rho$, where $\rho \in (-1,1)$ and $\delta_ij$ is Kronecker's delta.
As the mutual information is invariant to continuous bijective transformations
of the considered variables, it is enough to consider standardized
Gaussians marginals. We also compare MINE (using the Donsker-Varadhan representation in \Eq~\ref{eq:donskerbound}) and MINE-$f$ (based on the $f$-divergence representation in \Eq~\ref{eq:nguyenKLbound}).
Our results are presented in \Figs~\ref{fig:omie_mutual_information_test}. We observe that both MINE and Kraskov's estimation are virtually indistinguishable from the ground truth when estimating the mutual information between bivariate Gaussians. MINE shows marked improvement over Krakov's when estimating the mutual information between twenty dimensional random variables. We also remark that MINE provides a tighter estimate of the mutual information than MINE-$f$.
\begin{figure}[th]
\centering
\begin{tabular}[c]{cc}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{2d-smile}
\end{subfigure}
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{20d-smile}
\end{subfigure}
\end{tabular}
\vspace{-1em}
\caption{Mutual information between two multivariate Gaussians with component-wise correlation $\rho \in (-1,1)$.
}
\label{fig:omie_mutual_information_test}
\end{figure}
\subsection{Capturing non-linear dependencies}
An important property of mutual information between random variables with relationship $Y = f(X) + \sigma \odot \epsilon$, where $f$ is a deterministic non-linear transformation and $\epsilon$ is random noise, is that it is invariant to the deterministic nonlinear transformation, but should only depend on the amount of noise, $\sigma \odot \epsilon$. This important property, that guarantees the quantification dependence without bias for the relationship, is called \emph{equitability}~\citep{kinney2014equitability}.
Our results (\Fig~\ref{fig:nonlinear-mine}) show that MINE captures this important property.
\begin{figure}[th]
\centering
\includegraphics[width=0.45\textwidth]{nonlinear-mine}
\vspace{-1.5em}
\caption{MINE is invariant to choice of deterministic nonlinear transformation. The heatmap depicts mutual information estimated by MINE between 2-dimensional random variables $X \sim \mathcal{U}(-1, 1)$ and $Y = f(X) + \sigma \odot \epsilon$, where $f(x) \in \{x, x^3, sin(x)\}$ and $\epsilon \sim \mathcal{N}(0, I)$.}
\vspace{-1.5em}
\label{fig:nonlinear-mine}
\end{figure}
\section{Applications}
\label{Sec:applications}
In this section, we use MINE to present applications of mutual information and compare to competing methods designed to achieve the same goals.
Specifically, by using MINE to maximize the mutual information, we are able to improve mode representation and reconstruction of generative models.
Finally, by minimizing mutual information, we are able to effectively implement the information bottleneck in a continuous setting.
\subsection{Maximizing mutual information to improve GANs}
\label{sec:Gan1}
Mode collapse~\citep{che2016mode,dumoulin2016adversarially,donahue2016adversarial,Salimans2016gan,Metz2017Unrolled,saatchi2017bayesian,nguyen2017dual,lin2017pacgan,ghosh2017multi} is a common pathology of generative
adversarial networks~\citep[GANs,][]{goodfellow2014generative}, where the
generator fails to produces samples with sufficient diversity (i.e., poorly represent some modes).
GANs as formulated in \citet{goodfellow2014generative} consist of two components: a discriminator, $D:\mathcal{X} \rightarrow [0, 1]$ and a generator, $G: \mathcal{Z} \rightarrow \mathcal{X}$, where $\mathcal{X}$ is a domain such as a compact subspace of $\RR^n$.
Given $Z \in \mathcal{Z}$ follows some simple prior distribution (e.g., a spherical Gaussian with density, $\PP_Z$), the goal of the generator is to match its output distribution to a target distribution, $\PP_X$ (specified by the data samples). The discriminator and generator are optimized through the \emph{value function},
\begin{align}
\min_G &\max_D V(D,G) :=
\nonumber \\
&\EE_{\PP_X}[D(X)] + \EE_{\PP_Z}[\log{(1 - D(G(Z))}].
\label{eq:gan_d}
\end{align}
\begin{comment}
As observed in ~\citet{nowozin2016f}, maximizing the value function amounts to maximizing the variational lower-bound of $2 * D_{JS}(\PP || \QQ) - 2 \log{2}$, where $D_{JS}$ is the Jensen-Shannon divergence. The generator is then optimized to minimize $V$ alternatively as the discriminator maximizes it. In practice, however, we will use a \emph{proxy} to be maximized by the generator, $\EE_{p_{\tiny \mbox{gen}}}[\log(D(x)]$, which can palliate vanishing gradients.
\end{comment}
A natural approach to diminish mode collapse would be regularizing the generator's loss with the neg-entropy of the samples. As the sample entropy is intractable, we propose to use the mutual information as a proxy.
Following~\citet{chen2016infogan}, we write the prior as the concatenation of noise and code variables, $Z = [\bm{\epsilon}, \bm{c}]$. We propose to palliate mode collapse by maximizing the mutual information between the samples and the code. $I(G([\bm{\epsilon}, \bm{c}]); \bm{c}) = H(G([\bm{\epsilon}, \bm{c}])) - H(G([\bm{\epsilon}, \bm{c}])\mid\bm{c})$.
The generator objective then becomes,
\begin{align}
\argmax_G \EE[\log(D(G([\bm{\epsilon}, \bm{c}])))] + \beta I(G([\bm{\epsilon}, \bm{c}]); \bm{c}).
\label{eq:MINEGAN}
\end{align}
As the samples $G([\bm{\epsilon}, \bm{c}])$ are differentiable w.r.t. the parameters of $G$, and the statistics network being a differentiable function, we can maximize the mutual information using back-propagation and gradient ascent by only specifying this additional loss term. Since the mutual information is theoretically unbounded, we use adaptive gradient clipping (see the Supplementary Material) to ensure that the generator receives learning signals similar in magnitude from the discriminator and the statistics network.
\paragraph{Related works on mode-dropping}
Methods to address mode dropping in GANs can readily be found in the literature.
\citet{Salimans2016gan} use mini-batch discrimination.
In the same spirit, \citet{lin2017pacgan} successfully mitigates mode dropping in GANs by modifying the discriminator to make decisions on multiple real or generated samples.
\citet{ghosh2017multi} uses multiple generators that are encouraged to generate different parts of the target distribution.
\citet{nguyen2017dual} uses two discriminators to minimize the KL and reverse KL divergences between the target and generated distributions.
\citet{che2016mode} learns a reconstruction distribution, then teach the generator to sample from it, the intuition being that the reconstruction distribution is a de-noised or \emph{smoothed} version of the data distribution, and thus easier to learn.
\citet{srivastava2017veegan} minimizes the reconstruction error in the latent space of bi-directional GANs~\citep{dumoulin2016adversarially,donahue2016adversarial}.
\citet{Metz2017Unrolled} includes many steps of the discriminator's optimization as part of the generator's objective. While \citet{chen2016infogan} maximizes the mutual information between the code and the samples, it does so by minimizing a variational upper bound on the conditional entropy~\citep{barber2003algorithm} therefore ignoring the entropy of the samples. \citet{chen2016infogan} makes no claim about mode-dropping.
\paragraph{Experiments: Spiral, 25-Gaussians datasets}
We apply MINE to improve mode coverage when training a generative adversarial network~\citep[GAN,][]{goodfellow2014generative}.
We demonstrate using \Eq~\ref{eq:MINEGAN} on the spiral and the 25-Gaussians datasets, comparing two models, one with $\beta = 0$ (which corresponds to the orthodox GAN as in \citet{goodfellow2014generative}) and one with $\beta = 1.0$, which corresponds to mutual information maximization.
\begin{figure}[th]
\centering
\begin{tabular}[c]{cc}
\begin{subfigure}[l]{0.22\textwidth}
\includegraphics[width=\textwidth, height=30mm, clip, trim=1cm .8cm 1cm 1cm]{5000gan.png}
\caption{GAN}\label{subfig:omiegan_spiral_baseline_5000}
\end{subfigure}
\begin{subfigure}[r]{0.22\textwidth}
\includegraphics[width=\textwidth, height=30mm, clip, trim=0 .2cm 0 .2cm]{omie_5000.png}
\caption{GAN+MINE}\label{subfig:omieomiegan_spiral_5000}
\end{subfigure}
\end{tabular}
\caption{The generator of the GAN model without mutual information maximization after $5000$ iterations suffers from mode collapse (has poor coverage of the target dataset) compared to GAN+MINE on the spiral experiment.}
\label{fig:omie_gan_spiral}
\end{figure}
Our results on the spiral (\Fig~\ref{fig:omie_gan_spiral}) and the $25$-Gaussians (\Fig~\ref{fig:omie_gan_25_gaussians}) experiments both show improved mode coverage over the baseline with no mutual information objective.
This confirms our hypothesis that maximizing mutual information helps against mode-dropping in this simple setting.
\begin{figure}[t!]
\centering
\begin{tabular}[c]{ccc}
\begin{subfigure}[b]{0.15\textwidth}
\includegraphics[width=\textwidth, clip, trim=2.5cm 2cm 2cm 2cm]{0_25_gaussinas_0_1_0_1_spftplus.png}
\caption{Original data }\label{subfig:omiegan_25gaussian_baseline_0}
\end{subfigure}
\begin{subfigure}[b]{0.15\textwidth}
\includegraphics[width=\textwidth, clip, trim=2.5cm 2cm 2cm 2cm]{25gaussian_baseline.png}
\caption{GAN}\label{subfig:omiegan_25gaussian_baseline_20000}
\end{subfigure}
\begin{subfigure}[b]{0.15\textwidth}
\includegraphics[width=\textwidth, clip, trim=2.5cm 2cm 2cm 2cm]{80000_25_gaussinas_0_1_0_spftplus.png}
\caption{GAN+MINE}\label{subfig:80000_25gaussian_MINE_2}
\end{subfigure}
\end{tabular}
\caption{Kernel density estimate (KDE) plots for GAN+MINE samples and GAN samples on 25 Gaussians dataset.
}
\label{fig:omie_gan_25_gaussians}
\end{figure}
\paragraph{Experiment: Stacked MNIST}
Following \citet{che2016mode, Metz2017Unrolled, srivastava2017veegan,lin2017pacgan}, we quantitatively assess MINE's ability to diminish mode dropping on the stacked MNIST dataset which is constructed by stacking three randomly sampled MNIST digits. As a consequence, stacked MNIST offers 1000 modes. Using the same architecture and training protocol as in \citet{srivastava2017veegan,lin2017pacgan}, we train a GAN on the constructed dataset and use a pre-trained classifier on 26,000 samples to count the number of modes in the samples, as well as to compute the KL divergence between the
sample and expected data distributions. Our results in Table~\ref{table:stacked_mnist} demonstrate the effectiveness of MINE in preventing mode collapse on Stacked MNIST.
\begin{table}[th]
\small
\begin{center}
\begin{tabular}{ lcc}
& \multicolumn{2}{c}{Stacked MNIST} \\ \cline{2-3}
& \makecell{Modes \\(Max 1000)}& KL \\\midrule
DCGAN & $99.0$ & $3.40$ \\%& 0.00844$\pm$0.0020 \\
ALI & $16.0$ & $5.40$ \\%& 0.00670$\pm$0.0040 \\
Unrolled GAN & $48.7$ & $4.32$ \\%& 0.01300$\pm$0.0009\\
VEEGAN & $150.0$ & $2.95$ \\
PacGAN & $1000.0\pm0.0$& $0.06\pm1.0\mathrm{e}^{-2}$\\\midrule
GAN+MINE (Ours) & $1000.0\pm0.0$& $0.05\pm6.9\mathrm{e}^{-3}$ \\\bottomrule
\end{tabular}
\end{center}
\caption{Number of captured modes and Kullblack-Leibler divergence between the training and samples distributions for DCGAN~\citep{radford2015unsupervised}, ALI~\citep{dumoulin2016adversarially}, Unrolled GAN~\citep{Metz2017Unrolled}, VeeGAN~\citep{srivastava2017veegan}, PacGAN~\citep{lin2017pacgan}.}
\label{table:stacked_mnist}
\end{table}
\begin{figure*}[t!]
\centering
\begin{tabular}[c]{ccc}
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth, clip, trim={0 5cm 0 0}]{stackedmnist_trainingset.png}
\caption{Training set}\label{subfig:stackedmnist_trainingset}
\end{subfigure}&
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth, clip, trim={0 5cm 0 0}]{stackedmnist_dcgan.png}
\caption{DCGAN}\label{subfig:stackedmnist_dcgan}
\end{subfigure}&
\begin{subfigure}[b]{0.3\textwidth}
\includegraphics[width=\textwidth, clip, trim={0 5cm 0 0}]{stackedmnist_minegan.png}
\caption{DCGAN+MINE}\label{subfig:stackedmnist_minegan}
\end{subfigure}
\end{tabular}
\caption{Samples from the Stacked MNIST dataset along with generated samples from DCGAN and DCGAN with MINE.
While DCGAN only shows a very limited number of modes, the inclusion of MINE generates a much better representative set of samples.}
\label{fig:stacked_mnist}
\end{figure*}
\subsection{Maximizing mutual information to improve inference in bi-directional adversarial models}
\label{sec:Gan2}
Adversarial bi-directional models were introduced in Adversarially Learned Inference~\citep[ALI,][]{dumoulin2016adversarially} and BiGAN~\citep{donahue2016adversarial} and are an extension of GANs which incorporate a reverse model, $F: \mathcal{X} \rightarrow \mathcal{Z}$ jointly trained with the generator.
These models formulate the problem in terms of the value function in \Eq~\ref{eq:gan_d} between two joint distributions, $p(\bm{x}, \bm{z}) = p(\bm{z} \mid \bm{x}) p(\bm{x})$ and $q(\bm{x}, \bm{z}) = q(\bm{x} \mid \bm{z}) p(\bm{z})$ induced by the forward (encoder) and reverse (decoder) models, respectively\footnote{We switch to density notations for convenience throughout this section.}.
One goal of bi-directional models is to do inference as well as to learn a good generative model.
Reconstructions are one desirable property of a model that does both inference and generation, but in practice ALI can lack fidelity
~\citep[i.e., reconstructs less faithfully than desired, see][]{li2017towards,ulyanov2017adversarial,belghazi2018hierarchical}.
To demonstrate the connection to mutual information, it can be shown (see the Supplementary Material for details) that the reconstruction error, $\mathcal{R}$, is bounded by,
\begin{align} \label{equ:recons_bound}
\mathcal{R} \leq \KL{q(\bm{x}, \bm{z})}{p(\bm{x}, \bm{z})} - I_q(\bm{x}, \bm{z}) + H_q(\bm{z})
\end{align}
If the joint distributions are matched, $H_q(\bm{z})$ tends to $H_p(\bm{z})$, which is fixed as long as the prior, $p(\bm{z})$, is itself fixed. Subsequently, maximizing the mutual information minimizes the expected reconstruction error.
Assuming that the generator is the same as with GANs in the previous section, the objectives for training a bi-directional adversarial model then become:
\begin{align}
\argmax_{D} &\EE_{q(\bm{x}, \bm{z})} [\log{D(\bm{x}, \bm{z})}] + \EE_{p(\bm{x}, \bm{z})}[\log{(1 - D(\bm{x}, \bm{z}))}]
\nonumber \\
\argmax_{F, G} &\EE_{q(\bm{x}, \bm{z})} [\log{(1 - D(\bm{x}, \bm{z}))}] + \EE_{p(\bm{x}, \bm{z})}[\log{D(\bm{x}, \bm{z})}]
\nonumber \\
&+ \beta I_q(\bm{x}, \bm{z}).
\end{align}
\paragraph{Related works}
\citet{ulyanov2017adversarial} improves reconstructions quality by forgoing the discriminator and expressing the adversarial game between the encoder and decoder.
\citet{kumar2017improved} augments the bi-directional objective by considering the reconstruction and the corresponding encodings as an additional fake pair.
\citet{belghazi2018hierarchical} shows that a Markovian hierarchical generator in a bi-directional adversarial model provide a hierarchy of reconstructions with increasing levels of fidelity (increasing reconstruction quality).
\citet{li2017towards} shows that the expected reconstruction error can be diminished by minimizing the conditional entropy of the observables given the latent representations.
The conditional entropy being intractable for general posterior, \citet{li2017towards} proposes to augment the generator's loss with an adversarial cycle consistency loss~\citep{CycleGAN2017} between the observables and their reconstructions.
\vspace{-1cm}
\\
\paragraph{Experiment: ALI+MINE}
In this section we compare MINE to existing bi-directional adversarial models.
As the decoder's density is generally intractable, we use three different metrics to measure the fidelity of the reconstructions with respect to the samples; $(i)$ the euclidean reconstruction error, $(ii)$ \emph{reconstruction accuracy}, which is the proportion of labels preserved by the reconstruction as identified by a pre-trained classifier; $(iii)$ the Multi-scale structural similarity metric~\citep[MS-SSIM,][]{MS-SSIM01} between the observables and their reconstructions.
We train MINE on datasets of increasing order of complexity: a toy dataset composed of 25-Gaussians, MNIST~\citep{lecun1998mnist}, and the CelebA dataset~\citep{liu2015deep}.
\Fig~\ref{fig:bi_directional} shows the reconstruction ability of MINE compared to ALI. Although ALICE does perfect reconstruction (which is in its explicit formulation), we observe significant mode-dropping in the sample space. MINE does a balanced job of reconstructing along with capturing all the modes of the underlying data distribution.
Next, we measure the fidelity of the reconstructions over ALI, ALICE, and MINE.
\Tbl~2 compares MINE to the existing baselines in terms of
euclidean reconstruction errors, reconstruction accuracy, and MS-SSIM. On MNIST,
MINE outperforms ALI in terms of reconstruction errors by a good margin
and is competitive to ALICE with respect to reconstruction accuracy and
MS-SSIM. Our results show that MINE's effect on reconstructions
is even more dramatic when compared to ALI and ALICE on the CelebA dataset.
\begin{figure*}[t]
\title{Reconstructions and Samples}
\begin{minipage}{0.67\textwidth}
\begin{minipage}{0.24\textwidth}
\centering
(a) ALI
\includegraphics[width=0.99\columnwidth, clip=True, trim=40 20 40 40]{ali_reconstructions}
\end{minipage}
\begin{minipage}{0.24\textwidth}
\centering
(b) ALICE ($l_2$)
\includegraphics[width=0.99\columnwidth, clip=True, trim=40 20 40 40]{ALICEL2_recon}
\end{minipage}
\begin{minipage}{0.24\textwidth}
\centering
(c) ALICE (A)
\includegraphics[width=.99\linewidth, clip=True, trim=40 20 40 40]{ALICEAdv_recon}
\end{minipage}
\begin{minipage}{0.24\textwidth}
\centering
(d) ALI+MINE
\includegraphics[width=.99\linewidth, clip=True, trim=40 20 40 40]{100000_donsker6_0_reconstructions}
\end{minipage}
\vspace{.3cm}
\begin{minipage}{0.24\textwidth}
\centering
\includegraphics[width=0.99\columnwidth, clip=True, trim=40 20 40 40]{ali_samples}
\end{minipage}
\begin{minipage}{0.24\textwidth}
\centering
\includegraphics[width=0.99\columnwidth, clip=True, trim=40 20 40 40]{ALICEL2_samples}
\end{minipage}
\begin{minipage}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth, clip=True, trim=40 20 40 40]{ALICEAdv_samples}
\end{minipage}
\begin{minipage}{0.24\textwidth}
\centering
\includegraphics[width=.99\linewidth, clip=True, trim=40 20 40 40]{OMIE_samples}
\end{minipage}
\end{minipage}
\begin{minipage}{0.31\textwidth}
\caption{Reconstructions and model samples from adversarially learned inference (ALI) and variations intended to increase improve reconstructions.
Shown left to right are the baseline (ALI), ALICE with the $l_2$ loss to minimize the reconstruction error, ALICE with an adversarial loss, and ALI+MINE.
Top to bottom are the reconstructions and samples from the priors.
ALICE with the adversarial loss has the best reconstruction, though at the expense of poor sample quality, where as ALI+MINE captures all the modes of the data in sample space.}
\label{fig:bi_directional}
\end{minipage}
\end{figure*}
\begin{table}[ht]
\small
\centering
\begin{tabular}{l|c|c|c}
\toprule
Model & \makecell{Recons. \\Error} & \makecell{Recons. \\Acc.(\%)} & MS-SSIM \\% \Tstrut\Bstrut\\
\midrule
\multicolumn{4}{c}{\textbf{MNIST}} \\% \Tstrut\Bstrut\\
\midrule
ALI & 14.24 & 45.95 & 0.97\\
ALICE($l_2$) & 3.20 &99.03 &0.97 \\
ALICE(Adv.) & 5.20 & 98.17 & 0.98\\
MINE & 9.73& 96.10 & 0.99\\
\midrule
\multicolumn{4}{c}{\textbf{CelebA}} \\% \Tstrut\Bstrut\\
\midrule
ALI & 53.75 & 57.49 & 0.81\\
ALICE($l_2$) & 8.01 & 32.22 &0.93\\
ALICE(Adv.) & 92.56 & 48.95 & 0.51\\
MINE & 36.11 & 76.08 & 0.99\\
\bottomrule
\end{tabular}
\label{table:mnist}
\caption{Comparison of MINE with other bi-directional adversarial models in terms of euclidean reconstruction error, reconstruction accuracy, and MS-SSIM on the MNIST and CelebA datasets. MINE does a good job compared to ALI in terms of reconstructions.
Though the explicit reconstruction based baselines (ALICE) can sometimes do better than MINE in terms of reconstructions related tasks, they consistently lag behind in MS-SSIM scores and reconstruction accuracy on CelebA. }
\end{table}
\subsection{Information Bottleneck}
The Information Bottleneck~\citep[IB,][]{tishby2000information} is an information theoretic method for extracting relevant information, or
yielding a representation, that
an input $X \in \mathcal{X}$ contains about an output $Y \in \mathcal{Y}$. An optimal representation of $X$ would capture the relevant factors and
compress $X$ by diminishing the irrelevant parts which do not contribute to
the prediction of $Y$. IB was recently covered in the context of deep learning~\citep{tishby2015deep}, and as such can be seen as a process to construct an
approximation of the minimally sufficient statistics of the data.
IB seeks an encoder, $q(Z \mid X)$, that induces the
Markovian structure $X \rightarrow Z \rightarrow Y$. This is done by minimizing the IB Lagrangian,
\begin{align}
\mathcal{L}[q(Z \mid X)] = H(Y|Z) + \beta I(X,Z),
\end{align}
which appears as a standard cross-entropy loss augmented with a regularizer promoting minimality of the representation~\citep{AchilleSoatto2017}. Here we propose to estimate the regularizer with MINE.
\paragraph{Related works} In the discrete setting, \cite{tishby2000information} uses
the Blahut-Arimoto Algorithm \cite{arimoto1972algorithm}, which can be
understood as cyclical coordinate ascent in function spaces.
While IB is successful and popular in a discrete
setting, its application to the continuous setting was stifled by the
intractability of the continuous mutual information. Nonetheless,
IB was applied in the case of jointly Gaussian random
variables in \cite{chechik2005information}.
In order to overcome the intractability of $I(X; Z)$ in the continuous setting, \citet{Alemi2016deep, kolchinsky2017nonlinear, chalk2016relevant} exploit the variational
bound of \citet{barber2003algorithm} to approximate the conditional entropy in $I(X; Z)$.
These approaches differ only on their treatment of the marginal distribution of the bottleneck variable: \citet{Alemi2016deep} assumes a
standard multivariate normal marginal distribution,
\citet{chalk2016relevant} uses a Student-t distribution, and \citet{kolchinsky2017nonlinear} uses non-parametric estimators.
Due to their reliance on a variational approximation, these methods require a tractable density for the approximate posterior, while MINE does not.
\vspace{-1cm}
\\
\paragraph{Experiment: Permutation-invariant MNIST classification}
Here, we demonstrate an implementation of the IB objective on
permutation invariant MNIST using MINE. We compare to the Deep Variational Bottleneck~\citep[DVB,][]{Alemi2016deep} and use the same empirical setup.
As the DVB relies on a variational bound on the conditional entropy, it therefore requires a tractable density. \citet{Alemi2016deep} opts for a conditional Gaussian encoder $\bm{z} =
\mu(\bm{x}) + \sigma \odot \epsilon$, where $\epsilon \sim \mathcal{N}(0, I)$.
As MINE does not require a tractable density, we consider three type of encoders: $(i)$ a Gaussian encoder as in \citet{Alemi2016deep}; $(ii)$ an \emph{additive noise encoder}, $\bm{z} = enc(\bm{x} + \sigma \odot \epsilon)$; and $(iii)$ a \emph{propagated noise encoder}, $\bm{z} = enc([\bm{x}, \epsilon])$.
Our results can be seen in \Tbl~\ref{table:mnist_mi_bottleneck}, and this shows MINE as being superior in these settings.
\begin{table}[h!]
\label{table:info_bottleneck}
\centering
\begin{tabular}{c|c}
\toprule
Model & Misclass. rate(\%) \\% \Tstrut\Bstrut\\
\midrule
Baseline & 1.38\% \\
Dropout & 1.34\% \\
Confidence penalty & 1.36\% \\
Label Smoothing & 1.40\% \\
DVB & 1.13\% \\
DVB + Additive noise & 1.06\% \\
\midrule
MINE(Gaussian) (ours) & 1.11\% \\
MINE(Propagated) (ours) & 1.10\% \\
MINE(Additive) (ours) & 1.01\% \\
\bottomrule
\end{tabular}
\caption{Permutation Invariant MNIST misclassification rate using ~\citet{Alemi2016deep} experimental setup for regularization by confidence penalty~\citep{pereyra2017regularizing}, label smoothing~\citep{pereyra2017regularizing}, Deep Variational Bottleneck(DVB)~\citep{Alemi2016deep} and MINE.
The misclassification rate is averaged over ten runs. In order to control for the regularizing impact of the additive Gaussian noise in the additive conditional, we also report the results for DVB with additional additive Gaussian noise at the input. All non-MINE results are taken from~\citet{Alemi2016deep}.}
\label{table:mnist_mi_bottleneck}
\end{table}
\section{Conclusion}
We proposed a mutual information estimator, which we called the mutual information neural estimator (MINE), that is scalable in dimension and sample-size. We demonstrated the efficiency of this estimator by applying it in a number of settings.
First, a term of mutual information can be introduced alleviate mode-dropping issue in generative adversarial networks~\citep[GANs,][]{goodfellow2014generative}. Mutual information can also be used to improve inference and reconstructions in adversarially-learned inference~\citep[ALI,][]{dumoulin2016adversarially}. Finally, we showed that our estimator allows for tractable application of Information bottleneck methods~\citep{tishby2000information} in a continuous setting.
\section{Acknowledgements}
We would like to thank Martin Arjovsky, Caglar Gulcehre, Marcin Moczulski, Negar Rostamzadeh, Thomas Boquet, Ioannis Mitliagkas, Pedro Oliveira Pinheiro for helpful comments,
as well as Samsung and IVADO for their support.
\small
|
\section{Introduction}\label{sec11}
Our starting point is the classical linear boundary value problem modelling the
deformation of a homogeneous isotropic elastic body,
\begin{subequations} \label{os1}
\begin{align} \label{os1a}
-\nabla\cdot\mbox{\boldmath$\sigma$}& =\bm{f} \quad \mbox{in } \Omega &&\hspace*{-24pt}(\mbox{equilibrium of forces}),\\
\bm{u} &= {\bm{g} \quad \mbox{on } \partial \Omega} &&\hspace*{-24pt}(\mbox{essential boundary condition}),
\label{os1c}
\end{align}
\end{subequations}
where $\Omega \subset \mathbb{R}^{2}$ is a bounded Lipschitz polygon.
Here, \rbl{the deformation} is written in terms of the
stress tensor $\mbox{\boldmath$\sigma$} : \hbox{{\msbm \char "52}}^2\rightarrow \hbox{{\msbm \char "52}}^{2\times 2}$ and the body force
$\bm{f}: \hbox{{\msbm \char "52}}^2\rightarrow \hbox{{\msbm \char "52}}^{2}$, where
\begin{align*}
\mbox{\boldmath$\sigma$}=2 \mu \bm{\varepsilon}(\bm{u})+\lambda ({\nabla\cdot \bm{u}}){\bm{I}},
\end{align*}
${\bm{I}}$ is the $2\times 2$ identity matrix, $\bm{\varepsilon} : \hbox{{\msbm \char "52}}^2\rightarrow \hbox{{\msbm \char "52}}^{2\times 2}$ is the strain
tensor, $\bm{u}: \hbox{{\msbm \char "52}}^2\rightarrow \hbox{{\msbm \char "52}}^{2}$ is the displacement, and $\bm{\varepsilon}(\bm{u})=\frac{1}{2}(\nabla \bm{u}+(\nabla \bm{u})^{\top})$. The Lam\'{e} coefficients $\mu$ and $\lambda$ satisfy $0<\mu_1<\mu<\mu_2 <\infty$
and $0<\lambda<\infty$ and can be written in
terms of the Young's modulus $E$ and the Poisson ratio $\nu$ as
\begin{align*}
\mu=\frac{E}{2(1+\nu)}, \quad \lambda=\frac{E\nu}{(1+\nu)(1-2\nu)}.
\end{align*}
The coefficient $\lambda$ becomes unbounded in the incompressible limit $\nu \to 1/2$, leading to the
well-known phenomenon of locking for standard finite element methods. A popular remedy is to introduce
an additional unknown, \rbl{rewrite} \eqref{os1a}--\eqref{os1c} as a system and then apply an appropriate
\emph{mixed} finite element method.
We consider mixed approximation methods that are robust with respect to the Lam\'{e} coefficients which arise
from the {\it Herrmann} or {\it Hydrostatic} formulations \cite{DR,RLH} of (\ref{os1a})--(\ref{os1c}). Introducing $p=-\kappa \nabla \cdot \mathbf{u}$ \rbl{we rewrite} the problem as
\begin{subequations} \label{os2a}
\begin{align}
-\nabla\cdot\mbox{\boldmath$\sigma$}& =\bm{f} \quad\mbox{in } \Omega, \\
\nabla\cdot\bm{u}+\frac{p}{\kappa} &=0\quad\mbox{in } \Omega,\\
\bm{u}&= {\bm{g}\quad \mbox{on } \partial\Omega},
\end{align}
\end{subequations}
where either $\kappa=\lambda$ (in the Herrmann formulation) or $\kappa=\mu+\lambda$ (in the Hydrostatic formulation in two dimensions). The stress tensor can then be written as
\begin{align} \label{stress-def}
\mbox{\boldmath$\sigma$}(\bm{u}, p) = \left\{\begin{array}{ll}
2 \mu \bm{\varepsilon}(\bm{u})- p{\bm{I}} & \quad \mbox{ (Herrmann)},\\
2 \mu (\bm{\varepsilon}(\bm{u})-\frac{1}{2} (\nabla \cdot \bm{u}) \bm{I})- p{\bm{I}} & \quad \, \mbox{(Hydrostatic)}.
\end{array}\right.
\end{align}
There is an extensive literature on finite element approximation of elasticity problems; see Boffi et al.\cite{DFM} and Hughes\cite{TJH} for a comprehensive overview and \rbl{Houston et al.\cite{PDT} and Kouhia \& Stenberg\cite{KS}
for specific details}.
In Ref.~\refcite{KPS}, the authors provide a posteriori error analysis for conforming mixed finite element approximations of the Herrmann formulation using \rbl{stable} rectangular elements. \rbl{A variety of} local problem error estimators for the energy error are considered and proved to be robust \rbl{when} $\nu \to 1/2$. Those results can be extended to the Hydrostatic formulation
\rbl{whenever} the chosen finite element spaces satisfy \rbl{minimal conditions, as discussed by Boffi \& Stenberg\cite{DR}}.
\rbl{In this work, we extend the analysis in Ref.~\refcite{KPS} to cover the {\it lowest-order} $\bm{P}_{1}$--$P_{0}$
approximation defined on triangular elements. {An important} issue that will be addressed is the requirement
for pressure stabilisation. While pressure stabilisation of the lowest order mixed methods for
the Stokes equations has been extensively studied (for example by
Dohrmann \& Bochev\cite{CP}, Burman \& Fern{\' a}ndez\cite{BF} and Barrenechea \& Valentin\cite{BV})
the application of stabilised methods to elasticity equations appears to be a new development.}
In Section~\ref{weak_stuff} we review the weak formulation of \eqref{os2a}. In Section~\ref{Hdivmethsec},
we discuss $\bm{P}_{1}$--$P_{0}$ approximation, \rbl{review our} local stabilisation strategy and
\rbl{establish an} a priori error bound.
\rbl{The stabilisation strategy that is adopted was developed in Refs.}~\refcite{DD,ks92} \rbl{in the context of
the Stokes equations}.
\rbl{The distinctive feature of this contribution is the identification of a suitable energy norm---which
removes the requirement to specify (or ``tune'') a stabilisation parameter.}
In \rbl{Section~\ref{apost}} we discuss a \rbl{conventional} residual-based a posteriori error estimator and
\rbl{we introduce} a local Poisson problem estimator. Both estimators are shown to
be {\it robust} in the sense that the material parameters do not appear in the error bounds.
This \rbl{robustness is significantly} more challenging to achieve than for the Stokes problem, which only
\rbl{involves a single (viscosity) parameter.} \rbl{Some numerical results that reinforce the theory are
discussed in Section~\ref{Numres}.}
\rbl{In the rest of the paper we will use the symbols $\lesssim$ and $\gtrsim$ to denote bounds that are
valid up to positive constants that are independent of the Lam\'{e} coefficients and the mesh parameters.}
\section{Weak Formulation}\label{weak_stuff}
Our notation is standard: $H^s(\omega)$ denotes the usual Sobolev space with norm $||\cdot||_{s,\omega}$ for $s\ge0$. When $\omega=\Omega$, we use $||\cdot||_{s}$ instead of
$||\cdot||_{s,\Omega}$ and we denote vector-valued Sobolev spaces by boldface letters
$\bm{H}^{s}(\omega)=\bm{H}^{s}(\omega;\hbox{{\msbm \char "52}}^2)$. We also define
\begin{align*}
\bm{H}^1_E(\Omega):=\bigl\{\bm{v}\in \bm{H}^1(\Omega)
\;\big|\; \bm{v}|_{\partial\Omega}=\bm{g}\bigr\}, \quad
\bm{H}^{1\over 2}(\partial \Omega):=\bigl\{\bm{v} \, |\, \bm{v}=\bm{u}|_{\partial\Omega}, \bm{u}\in \bm{H}^1(\Omega)\bigr\} ,
\end{align*}
and the test spaces
${\bm{V}=\bm{H}_{0}^{1}(\Omega) := \bigl\{\bm{v}\in \bm{H}^1(\Omega) \;\big|\; \bm{v}|_{\partial\Omega}=\bm{0} \bigr\}}$ and
$Q=\rbl{L^{2}(\Omega)}$.
The standard weak formulation of (\ref{os2a}) is:
find $(\bm{u},p)\in \bm{H}^1_E\times Q$ such that
\begin{subequations} \label{scm11a}
\begin{align
a(\bm{u},\bm{v})+b(\bm{v},p)&=f(\bm{v})\quad \forall \bm{v}\in \bm{V}, \\
b(\bm{u},q)-c(p,q)&=0\quad\quad\,\forall q\in Q, \label{scm11b}
\end{align}
\end{subequations}
where
$$ b(\bm{v},p)=-\int_{\Omega} p \nabla\cdot \bm{v},\quad c(p,q)=\frac{1}{\kappa}\int_{\Omega} pq,\quad f(\bm{v})=\int_{\Omega}\bm{f}\,\bm{v},$$
and either
\begin{align*}
a(\bm{u},\bm{v})=a_{H}(\bm{u}, \bm{v}) = 2\mu\int_{\Omega}\bm{\varepsilon}(\bm{u}):\bm{\varepsilon}(\bm{v}),
\end{align*}
(in the Herrmann formulation) or
\begin{align*}
a(\bm{u},\bm{v})= a_{S}(\bm{u}, \bm{v}) = 2\mu\left(\int_{\Omega}\bm{\varepsilon}(\bm{u}):\bm{\varepsilon}(\bm{v})-\frac{1}{2}\int_{\Omega}(\nabla \cdot \bm{u})(\nabla \cdot \bm{v}) \right)
\end{align*}
(in the Hydrostatic formulation). Note that, where it is necessary to make a distinction, we will use the notation $a_{H}(\cdot, \cdot)$ and $a_{S}(\cdot, \cdot)$, but where a stated result holds for both, we will simply use $a(\cdot, \cdot)$. We assume $\bm{f}\in ( L^{2}(\Omega))^2$ and that $\bm{g}\in \bm{H}^{1\over 2}(\partial\Omega)$ is a polynomial of degree at most one in each component so that no error is incurred in approximating the essential boundary condition. As usual, we define
\begin{align}
\mathcal{B}(\bm{u},p; \bm{v},q)=a(\bm{u},\bm{v})+b(\bm{v},p)+b(\bm{u},q)-c(p,q),
\end{align}
so as to express (\ref{scm11a}) in the more compact form:
find $(\bm{u},p)\in \bm{H}^1_E\times Q$ such that
\begin{align}\label{scm12}
\mathcal{B}(\bm{u},p; \bm{v},q)=f(\bm{v}), \quad \forall (\bm{v},q)\in\bm{V}\times Q.
\end{align}
Finally, we define the following \textit{energy} norm for the error analysis
\begin{align}\label{energy_def}
|||(\bm{u},p)|||^2&=2\mu\, {|| \nabla \bm{u}||^2_{0}} +(2\mu)^{-1}||p||^2_{0}+ \kappa^{-1} ||p||^2_{0}.
\end{align}
\rbl{One} can establish the well-posedness of the weak formulation for $\nu \in (0, 1/2)$ by considering \eqref{scm11a}
or \eqref{scm12}. We will work with the latter. Note that when $\nu = 1/2$, $c(\cdot, \cdot)$ disappears
from \eqref{scm11a} and the problem can be analysed as a saddle point problem in the standard way
(similar to Stokes problems). However, since we impose $\bm{u}=\bm{g}$ on the whole boundary,
the pressure solution is only unique up to a constant in that case.
\rbl{We start by reviewing some useful results.}
{For both} formulations, it is is easy to show that
\begin{align}
a(\bm{u},\bm{v})\le {2\mu} \, { ||\nabla \bm{u} ||_0} \,
\, { ||\nabla \bm{v} ||_0} \quad \forall \bm{u},\bm{v}\in \bm{V}. \label{abd}
\end{align}
It is also known that that there exists an (inf-sup) constant $C_{\Omega}>0$ such that
\begin{align}
\sup_{0\neq\bm{v}\in \bm{V}}
\frac{b(\bm{v},q)}{ ||\nabla\bm{v}||_0 }\ge C_{\Omega} ||q||_{0}, \quad { \forall q\in {Q}, q\neq {\rm constant},}
\label{binfsup}
\end{align}
see, for example, p.\,128 of Ref.~\refcite{HDA}.
\rbl{Next,} in the Herrmann formulation, we \rbl{know that}
\begin{align}
a_{H}(\bm{v},\bm{v}) & \ge C_K 2\mu \, { ||\nabla \bm{v} ||_0^2} \quad \forall \bm{v}\in \bm{V},
\label{aell}
\end{align}
by Korn's inequality, so that $a_{H}(\cdot, \cdot)$ is coercive on $\bm{V}$.
{Similarly,} in the Hydrostatic case\cite{DR} \rbl{we have that}
\begin{align}
a_{S}(\bm{v},\bm{v}) & \ge(1/2) 2\mu \, { ||\nabla \bm{v} ||_0^2} \quad \forall \bm{v}\in \bm{V}.
\label{aell-B}
\end{align}
\rbl{We note in passing that the coercivity estimate \eqref{aell-B} does not hold} if
$\partial \Omega = \partial \Omega_{D} \cup \partial \Omega_{N}$,
where $\partial \Omega_{N} \neq \emptyset$ is a portion of the boundary where $\bm{\sigma} \bm{n} =\bm{0}$.
\rbl{(This case requires a separate treatment, exploiting the fact
that $a_{S}(\cdot,\cdot)$ is coercive on an appropriate nullspace $\bm{V}_{0} \subset \bm{V}$.)}
The following stability result ensures well-posedness of (\ref{scm12}).
\begin{lemma}\label{Sinsuplem12}
Let $\rblx{Q_0}:=\left\{ q \in \rblx{L^2(\Omega)}, \int_{\Omega} q = 0 \right\}$.
For any $(\bm{u},p)\in \bm{V}\times \rblx{Q_0}$,
there exists a pair of functions
$(\bm{v},q)\in \bm{V} \times \rblx{Q_0}$, with $|||(\bm{v},q)|||\lesssim |||(\bm{u},p)|||$, satisfying
$$ \mathcal{B}(\bm{u},p; \bm{v},q)\gtrsim |||(\bm{u},p)|||^{{2}}. $$
\end{lemma}
\begin{proof}
For the Herrmann case, the result follows from \eqref{abd}, \eqref{binfsup} and \eqref{aell}; see Lemma 3.3 in Ref.~\refcite{KPS}. In the Hydrostatic case, the same proof can be applied, using \eqref{aell-B} instead of \eqref{aell} (which is the same result with $C_{K}=1/2$). Since the energy norm \eqref{energy_def} is defined with respect to $\kappa$, the constant in the bound $\gtrsim$ is the same (up to the value of $C_{K}$).
\end{proof}
\begin{remark}
\rblx{To check the uniqueness of the pressure solution for $\nu \in (0, 1/2)$ we}
test \eqref{scm11b} with a constant function $\rblx{q=1}$ and use the divergence theorem. This gives
\begin{align}
\frac{1}{\kappa}\int_{\Omega} p \> &= -\int_{\Omega} \nabla\cdot \bm{u}
= \rblx{-} \int_{\partial \Omega} \bm{u} \cdot \bm{n} \, ds ,
\quad\hbox{thus}\quad \int_{\Omega} p \> = - \kappa \int_{\partial \Omega} \bm{g} \cdot \bm{n} \, ds . \label{uniquep}
\end{align}
The characterisation \eqref{uniquep} guarantees the uniqueness of the pressure \rblx{satisfying \eqref{scm12}}
using either of the two formulations.
\end{remark}
\section{Stabilised $P_{1}$--$P_{0}$ approximation } \label{Hdivmethsec}
Let $\{\mathcal{T}_{h}\}$ denote a family of shape-regular triangular meshes of $\rbl{\overline{\Omega}}$ into triangles $K$ of diameter $h_K$. For each mesh $\mathcal{T}_h$, we let $\mathcal{E}_h$ denote the set of all edges and $h_E$
\rbl{denote} the length of an edge $E\in\mathcal{E}_h$. Next, we introduce finite-dimensional subsets $\bm{X}^h_E \subset \bm{H}^1_{E}$, $\bm{X}^h_0 \subset \bm{V}$ and {$Q^h \subset Q$}. The discrete weak formulation of \eqref{scm11a} is as follows: find $(\bm{u}_h,p_h)\in \bm{X}^h_E\times Q^h$ such that
\begin{subequations} \label{FEA11}
\begin{align}
a(\bm{u}_h,\bm{v}_h)+b(\bm{v}_h,p_h)&=f(\bm{v}_h)\quad \forall \bm{v}_h\in \bm{X}^h_0, \\
b(\bm{u}_h,q_h)-c(p_h,q_h)&=0\quad\quad\;\;\forall q_h\in Q^h.
\end{align}
\end{subequations}
Specifically, we choose $\bm{X}^{h}_{0}$ to be the space of vector-valued functions that are piecewise linear in each component and globally continuous ($\bm{P}_{1}$), and we choose $Q^{h}$ to be the subset of $Q$ that contains piecewise constant functions ($P_{0}$). The solution space $\bm{X}^h_E$ is obtained from $\bm{X}^h_0$ by construction in the usual way, by augmenting the basis with additional $\bm{P}_{1}$ functions associated with Dirichlet boundary nodes (where $\bm{g} \neq \bm{0}$). For more details about $\bm{P}_1$--${P}_{0}$ approximation, see Refs.~\refcite{HDA,DD,DFM,ks92,ns98}.
\rbl{We note that, while the simplicity of the low-order scheme is very attractive} from a computational point of view,
stabilisation \rbl{of the underlying approximation is essential when} working with values of $\nu$ close to $1/2$.
Given a mesh $\mathcal{T}_{h}$, to define our stabilisation strategy, we first select a macroelement partitioning $\mathcal{M}_h$ which satisfies:
\begin{enumerate}
\item Each macroelement $M \in \mathcal{M}_{h}$ is a connected set of adjoining elements from $\mathcal{T}_h$.
\item $M_i\cap M_j = \emptyset$ for all $M_i, M_j \in \mathcal{M}_{h}, \, i\neq j$.
\item For any two neighboring macroelements $M_1$ and $M_2$ with $\int_{M_1\cap M_2}ds\neq 0$, there exists $\bm{v}\in
\rblx{\bm{X}^h_0}$ such that supp $\bm{v}\subset \overline{M}_1\cup \overline{M}_2 $
and $\int_{M_1\cap M_2} \bm{v}\cdot \bm{n}\, ds\neq 0$.
\item $\cup_{M\in \mathcal{M}_{h}} \overline{M}= \rbl{\overline{\Omega}}$.
For each $M \in \mathcal{M}_{h}$, the set of \rbl{interior interelement edges}
will be denoted by $\Gamma_M$. That is,
$$\Gamma_M=\{ E\in \mathcal{E}_{h} \setminus \partial \Omega, E\subset M\}.$$
\end{enumerate}
With the above definition, a locally stabilised version of the discrete weak problem (\ref{FEA11}) is as follows: find $(\bm{u}_h,p_h)\in \bm{X}^h_E\times Q^h$ such that
\begin{subequations} \label{LSFEA11}
\begin{align}
a(\bm{u}_h,\bm{v}_h)+b(\bm{v}_h,p_h)&=f(\bm{v}_h)\quad \forall \bm{v}_h\in\bm{X}^h_{0}, \\
b(\bm{u}_h,q_h)-c(p_h,q_h)-\mathcal{C}_{loc}(p_h,q_h)&=0,\quad\quad\;\;\forall q_h\in Q^h, \label{LSFEA12}
\end{align}
\end{subequations}
where
\begin{align*}
\mathcal{C}_{loc}(p_h,q_h)=\frac{1}{2\mu}\sum_{M\in\mathcal{M}_h}\sum_{E\in\Gamma_M}h_E
\int_{E}\llbracket p_h \rrbracket \llbracket q_h \rrbracket ds, \qquad p_{h}, q_{h} \in Q^{h},
\end{align*}
and $\llbracket\cdot\rrbracket$ denotes the jump across $E \in \Gamma_{M}$.
\begin{remark}
The choice of the stabilisation parameter $(1/2\mu)$ in the definition of $\mathcal{C}_{loc}(\cdot, \cdot)$
is motivated by the \rblx{a priori error analysis presented next}.
\end{remark}
The discrete \rblx{pressure} $p_h\in Q^h$ that solves \eqref{LSFEA11} is not uniquely defined
in the limiting case $\nu=1/2$. The associated linear algebra system is {\it singular} in this case.\footnote{In
the generation of the computational results with $\nu=0.49999$ (discussed later in Section~\ref{Numres})
the near-singular linear algebra systems were solved using $\backslash$ within MATLAB.}
\rblx{Define the constrained pressure approximation space $Q_0^h=Q^h\cap Q_0$.}
We will assume that for any partitioning ${\mathcal{M}}_{h}$, each macroelement $M \in {\mathcal{M}}_{h}$
belongs to one of a finite number of possible equivalence classes $\mathcal{E}_{\hat{M}_{1}}, \ldots, \mathcal{E}_{\hat{M}_{N}}$.
The next result immediately follows from Lemma 3.1 in Ref.~\refcite{ks92}.
\begin{lemma}\label{alpha1-result}
Let $\Pi_h$ be the $L^2$ projection operator {from $\rblx{Q_0^h}$} onto the subspace
\begin{align}
\rbl{\overline Q}^{\,h}=\{ q\in \rblx{Q_0}, \, q|_M\, \mbox{ is \rbl{constant} } \forall M\in \mathcal{M}_h \}.
\end{align}
Then, \rbl{there exists} $\alpha_{1}>0$ independent of $h$ and the Lam\'e coefficients \rbl{satisfying}
\begin{align*}
\mathcal{C}_{loc}(q,q)\ge\alpha_1\frac{1}{2\mu}||(I-\Pi_h)q||_{0}^2 \quad \forall q\in \rblx{Q_0^h}.
\end{align*}
\end{lemma}
The stabilised discrete formulation (\ref{LSFEA11}) can {also}
be written as: find $(\bm{u}_h,p_h)\in \bm{X}^h_E\times Q^h$ such that
\begin{align}\label{LSFEA13}
\mathcal{B}_S(\bm{u}_h,p_h; \bm{v}_h,q_h)=f(\bm{v}_h), \quad \forall (\bm{v}_h,q_h)\in\bm{X}^h_{0}\times Q^h,
\end{align}
\rbl{which involves the stabilised bilinear} form
\begin{align*}
\mathcal{B}_{S}(\bm{u}_{h},p_{h}; \bm{v}_{h},q_{h})=a(\bm{u}_{h},\bm{v}_{h})+b(\bm{v}_{h},p_{h})+b(\bm{u}_{h},q_{h})-c(p_{h},q_{h})-\mathcal{C}_{loc}(p_h,q_h).
\end{align*}
We are now ready to prove a stability result for \eqref{LSFEA13}.
\begin{lemma}\label{dinsuplem11}
For any $(\bm{u},p)\in \bm{X}^h_0\times \rblx{Q_0^h}$, there exists a pair of functions
$(\bm{v},q)\in \bm{X}^h_0\times \rblx{Q_0^h}$
with $|||(\bm{v},q)|||\lesssim |||(\bm{u},p)|||$ satisfying
\begin{align*}
\mathcal{B}_S(\bm{u},p; \bm{v},q)\gtrsim |||(\bm{u},p)|||^2.
\end{align*}
\end{lemma}
\begin{proof}
\rbl{A consequence of \eqref{binfsup}} \rbl{is that} there exists a constant $\alpha_2$,
independent of $h$ and the Lam\'e coefficients, and a function $\bm{w}\in \bm{X}^h_0$ satisfying
\begin{align}\label{deq11}
(\Pi_h p, \nabla\cdot \bm{w})=(2\mu)^{-1}||\Pi_h p||_{0}^2, \quad (2\mu)^{1/2}||\nabla\bm{w}||_{0}\le\alpha_2(2\mu)^{-1/2}||\Pi_h p||_{0}.
\end{align}
Since $(\bm{u},p)\in \bm{X}^h_{0}\times \rblx{Q_0^h}$ and $\bm{X}_{0}^{h} \subset \bm{V}$, $\rblx{Q_0^h \subset Q^h}$,
\rbl{using} the definition of $\mathcal{B}_{S}(\cdot, \cdot)$ gives,
\begin{align}
\mathcal{B}_S(\bm{u},p;\bm{u},-p) & \ge C_{K} 2\mu||\nabla \bm{u}||^2_0+ \kappa^{-1}||p||^2_0+\mathcal{C}_{loc}(p,p), \label{dinfsup1}
\end{align}
by \eqref{aell} (Herrmann case) or \eqref{aell-B} (Hydrostatic case). Next, using \eqref{deq11} and \eqref{abd}, for any $\epsilon >0$ we have,
\begin{align}\label{dinfsup2}
\mathcal{B}_S(\bm{u},p; -\bm{w},0) & = -a(\bm{u},\bm{w})-b(\bm{w}, (I-\Pi_h)p)-b(\bm{w}, \Pi_h p)\nonumber\\
&\ge -(2\mu)^{\frac{1}{2}}||\nabla\bm{u}||_0\;(2\mu)^{\frac{1}{2}}||\nabla\bm{w}||_0\nonumber\\
&\quad-(2\mu)^{-\frac{1}{2}}||(I-\Pi_h)p||_{0}(2\mu)^{\frac{1}{2}}||\nabla\bm{w}||_{0}\ + (2\mu)^{-1}||\Pi_h p||^2_{0} \nonumber\\
&\ge -(2\mu)^{1/2}||\nabla\bm{u}||_0\;\alpha_2(2\mu)^{-1/2}||\Pi_h p||_0\nonumber\\
&\quad-(2\mu)^{-1/2}||(I-\Pi_h)p||_{0}(2\mu)^{-1/2}\alpha_2||\Pi_h p||_{0} + (2\mu)^{-1}||\Pi_h p||^2_{0}, \nonumber\\
&\ge -\epsilon(2\mu)||\nabla\bm{u}||^2_0- 2\epsilon^{-1} \alpha_2^{2}(2\mu)^{-1}||\Pi_h p||_0^{2} \nonumber\\
&\quad-\epsilon(2\mu)^{-1}||(I-\Pi_h) p||_0^2 + (2\mu)^{-1}||\Pi_h p||^2_{0}.
\end{align}
Now we introduce a parameter $\delta$. Using Lemma \ref{alpha1-result}, (\ref{dinfsup1}) and (\ref{dinfsup2}) we have,
\begin{align*}
\mathcal{B}_S(\bm{u},p;\bm{u} -\delta \bm{w},-p)&=\mathcal{B}(\bm{u},p;\bm{u},-p)+\delta\mathcal{B}(\bm{u},p; -\bm{w},0)\\
&\ge C_K 2\mu||\nabla\bm{u}||^2_0 + \kappa^{-1}||p||^2_0 + \mathcal{C}_{loc}(p,p)+\delta (2\mu)^{-1}||\Pi_h p||^2_{0}\nonumber\\
&-2\mu\delta\epsilon||\nabla\bm{u}||^2_0
- 2\delta \epsilon^{-1} \alpha_2^2(2\mu)^{-1}||\Pi_h p||_0^2 - \delta\epsilon(2\mu)^{-1}||(I-\Pi_h) p||_0^2,\\
&\ge (C_K-\delta\epsilon)2\mu||\nabla\bm{u}||^2_0+\kappa^{-1}||p||^2_0+\delta \left(1- 2\alpha_2^2 \epsilon^{-1}\right)(2\mu)^{-1}||\Pi_h p||^2_{0}\nonumber\\
&\quad-\delta\epsilon(2\mu)^{-1}||(I-\Pi_h) p||_0^2+\alpha_1(2\mu)^{-1}||(I-\Pi_h) p||_0^2,\\
&\ge (C_K-\delta\epsilon)2\mu||\nabla\bm{u}||^2_0+\kappa^{-1}||p||^2_0+\delta \left(1-\ 2\alpha_2^2 \epsilon^{-1}\right)(2\mu)^{-1}||\Pi_h p||^2_{0}\nonumber\\
&\quad+(\alpha_1-\delta\epsilon)(2\mu)^{-1}||(I-\Pi_h) p||_0^2.
\end{align*}
Making the specific choices $\epsilon =4\alpha_2^2$ and $\delta=\frac{1}{4\alpha_2^2}\min\{C_K/2,\alpha_1/2 \}$, it follows:
\begin{align}\label{infsup11}
\mathcal{B}_S(\bm{u},p;\bm{u} -\delta \bm{w},-p)&\ge C \left(2\mu||\nabla\bm{u}||^2_0+\frac{1}{\kappa}||p||^2_0 +(2\mu)^{-1}\left(||\Pi_h p||^2_{0}+||(I-\Pi_h) p||_0^2\right)\right),\nonumber\\
&\ge C \left(2\mu||\nabla\bm{u}||^2_0+\frac{1}{\kappa}||p||^2_0+(2\mu)^{-1}||p||^2_{0}\right),
\end{align}
where $C= \min\left\{1,\frac{C_K}{2},\frac{\alpha_1}{2},\frac{\delta}{2}\right\}$. Hence, the result holds with $\bm{v}=\bm{u}-\delta \bm{w}$ and $q=-p$. Finally, using the definition of $||| \cdot ||| $ and \eqref{deq11} gives
\begin{align}\label{infsup21}
|||(\bm{v},q)|||^2 & = 2 \mu \| \nabla\left(\bm{u}-\delta \bm{w} \right)\|_{0}^{2} + (2\mu)^{-1} \| p \|_{0}^{2} + \kappa^{-1}\|p\|_{0}^{2} \nonumber \\
& \le 2(2\mu)||\nabla\bm{u}||_{0}^2+ 2(2\mu)\delta^2||\nabla\bm{w}||_{0}^{2} +\left(\kappa^{-1}+(2\mu)^{-1}\right)||p||^{2}_{0},\nonumber\\
&\le \left(2+\frac{\delta^2\alpha_2^2}{2}\right)|||(\bm{u},p)|||^2.
\end{align}
The constants in (\ref{infsup11}) and (\ref{infsup21}) are independent of the Lam\'e coefficients. \end{proof}
We \rbl{can} now establish an a priori bound for the energy norm of the error associated with the
stabilised $\bm{P}_{1}$--$P_{0}$ approximation.
\begin{theorem}\label{apriori-thm} Let $\rblx{(\bm{u},p)\in \bm{H}^1_E\times Q}$
be the solution to (\ref{scm11a})
and let $(\bm{u}_h,p_h)\in\bm{X}^h_E\times Q^h$ satisfy \eqref{LSFEA11}. \rblx{Suppose that
$\int_{\partial\Omega} \bm{g} \cdot \bm{n} \, ds =0$ so that $\int_\Omega p = 0 = \int_\Omega p_h$
from \eqref{uniquep}.}
If $\bm{u}\in \bm{H}^2(\Omega)$ and $p\in H^1(\Omega)$, then
\begin{align}
|||(\bm{u}-\bm{u}_h,p-p_h)|||\lesssim h \left((2\mu)^{1/2} |\bm{u} |_{2}+\left((2\mu)^{-1/2}+\kappa^{-1/2}\right) | p |_1\right).
\end{align}
\end{theorem}
\begin{proof}
\rblx{Let $\tilde{\bm{u}} \in \bm{X}^h_E$ represent the piecewise linear interpolant of $\bm{u}\in \bm{H}^1_E$} and let
$\tilde{p} \in Q^h_0$ be the piecewise constant projection of $p\inQ$ \rblx{with mean value zero}.
\rblx{Using} the triangle inequality gives
\begin{align}\label{tri-bound}
|||(\bm{u}-\bm{u}_h,p-p_h)||| \, \lesssim \, |||(\bm{u}-\tilde{\bm{u}},p-\tilde{p})||| + |||(\tilde{\bm{u}}-\bm{u}_{h},\tilde{p}- p_{h})|||,
\end{align}
and the interpolation error satisfies
\begin{align}\label{interp-bound}
||| (\bm{u} - \tilde{\bm{u}}, p - \tilde{p}) ||| \, \lesssim \, h \left( (2\mu)^{1/2} |\bm{u} |_{2}+\left((2\mu)^{-1/2}+\kappa^{-1/2}\right) | p |_1\right).
\end{align}
Now, for all $(\bm{v},q)\in \bm{X}^{h}_{0}\times Q^h$, using (\ref{scm12}) and (\ref{LSFEA13}), \rbl{gives}
\begin{align*}
\mathcal{B}_S(\bm{u}_h-\tilde{\bm{u}},p_h-\tilde{p} ; \bm{v},q)& = \mathcal{B}_S(\bm{u}_h,p_h ; \bm{v},q)-\mathcal{B}(\tilde{\bm{u}},\tilde{p} ; \bm{v},q)+\mathcal{C}_{loc}(\tilde{p},q) \\
&= \mathcal{B}(\bm{u},p;\bm{v},q)-\mathcal{B}(\tilde{\bm{u}},\tilde{p} ; \bm{v},q)+\mathcal{C}_{loc}(\tilde{p},q) \nonumber\\
&= \mathcal{B}(\bm{u}-\tilde{\bm{u}},p-\tilde{p} ; \bm{v},q)+\mathcal{C}_{loc}(\tilde{p},q).
\end{align*}
Since $(\bm{u}_{h}-\tilde{\bm{u}}, p_{h}-\tilde{p}) \in \rblx{\bm{X}_{0}^{h} \times \rblx{Q_0^h}}$,
applying Lemma \ref{dinsuplem11} in the usual way gives
\begin{align*}
|||(\bm{u}_h-\tilde{\bm{u}},p_h-\tilde{p})||| \lesssim ||| \left(\bm{u}-\tilde{\bm{u}},p-\tilde{p} \right) |||
+ \sup_{\rblx{q\in Q_0^h},(2\mu)^{-1/2}||q||_{0}=1}\mathcal{C}_{loc}(\tilde{p},q).
\end{align*}
If $p\in H^1(\Omega)$, then $\mathcal{C}_{loc}({p},q)=0$ and the Cauchy--Schwarz inequality gives
\begin{align*}
\mathcal{C}_{loc}(\tilde{p},q)=\mathcal{C}_{loc}(\tilde{p}-p,q)\lesssim\Bigg(\sum_{K\in \mathcal{T}_h} \frac{h_K}{{2\mu}}||p-\tilde{p}||_{0,\partial K}^2\Bigg)^{1/2}\Bigg(\sum_{K\in \mathcal{T}_h} \frac{h_K}{{2\mu}}||q||_{0,\partial K}^2\Bigg)^{1/2}.\nonumber
\end{align*}
Following the proof of Theorem 3.1 in Ref.~\refcite{ks92}, it follows that
\begin{align*}
\left(\sum_{K\in \mathcal{T}_h} \frac{h_K}{2\mu}||q||_{0,\partial K}^2\right)^{1/2} \lesssim\frac{1}{\sqrt{2\mu}}||q||_{0}, \quad
\Bigg(\sum_{K\in \mathcal{T}_h} \frac{h_K}{{2\mu}}||p-\tilde{p}||_{0,\partial K}^2\Bigg)^{1/2} \lesssim \frac{h}{\sqrt{2\mu}}|p|_1,
\end{align*}
and hence
\begin{align}\label{other-bound}
|||(\bm{u}_h-\tilde{\bm{u}},p_h-\tilde{p})||| \lesssim ||| \left(\bm{u}-\tilde{\bm{u}},p-\tilde{p} \right) |||+ \frac{h}{\sqrt{2\mu}}|p|_1.
\end{align}
Combining \eqref{tri-bound} with \eqref{interp-bound} and \eqref{other-bound} gives the final result.
\end{proof}
\section{A posteriori error analysis}\label{apost}
\rbl{Two alternative a posteriori {energy} error estimation strategies will be discussed here. Both
estimation strategies are {robust} in the sense that material parameters do not appear in the error bounds.
The proofs are presented here for completeness---they are a minor extension of the results
established in Ref.~\refcite{KPS}.}
\subsection{Residual error estimation}\label{rposest}
\rbl{We discuss a residual-based error estimator first. The definition involves three} {distinct} parameters:
\begin{align}\label{gridparams}
\rho_K=h_K(2\mu)^{-\frac{1}{2}},\quad \rho_E=h_E(2\mu)^{-1},\quad\rho_d={1/( \kappa^{-1} + (2\mu)^{-1})}.
\end{align}
Let $(\bm{u}_{h}, p_{h}) \in\bm{X}^h_E\times Q^h$ satisfy \eqref{LSFEA11} and let $\bm{f}_h$ be the $L^2$-projection of $\bm{f}$ onto the space of piecewise constant \rbl{functions}. For each element $K$ in the finite element mesh $\mathcal{T}_{h}$, we define the local data oscillation error $\bm{\Theta}_{K}$ satisfying
\begin{align}
\label{dataapp1}
\bm{\Theta}_{K}^2=\rho_{K}^2||\bm{f}-\bm{f}_{h}||^{2}_{0,K},
\end{align}
and a local error indicator $\eta_{K}$ satisfying
$\eta^2_{K}=\eta^2_{R_K}+\eta^2_{E_K}+\eta_{J_K}^2$, where
\begin{align}\label{components}
\eta^2_{R_K}=\rho_{K}^2||\bm{R}_K||^2_{0,K}, \quad
\eta^2_{J_K}=\rho_d||R_K||^2_{0,K}\quad
\mbox{and}\quad
\eta^2_{E_K}=\sum_{E\in \rbl{\partial K}}\rho_E||\bm{R}_E||^2_{0,E}.
\end{align}
The two {\it element} residuals associated with \eqref{os2a} are given by
\begin{align} \label{elt_resid}
\bm{R}_K=\bm{f}_{h}\big|_K,
\quad R_K=\rblx{\left\{\nabla\cdot \bm{u}_h+ \frac{p_{h}}{\kappa} \right\} \Big|_K},
\end{align}
and the {\it edge} residual $\bm{R}_{E}$ is associated with the normal stress jump. That is,
\begin{align} \label{stress-jump-def}
\bm{R}_E=\left\{\begin{array}{ll}
\frac{1}{2}\llbracket -\bm{\sigma}(\bm{u}_{h}, p_{h}) \bm{n}\rrbracket_E & \quad E\in \mathcal{E}_h\setminus\partial\Omega ,\\
0 & \quad {E \in \partial \Omega},
\end{array}\right.
\end{align}
where $\bm{\sigma}(\bm{u}_{h}, p_{h})$ is defined via \eqref{stress-def}. Note that since $\bm{u}_{h} \in \bm{P}_{1}$ and $p_{h} \in P_{0}$, $\nabla \cdot \bm{u}_{h}$ and \rbl{$R_K$} are constant on each element, \rbl{as is}
the normal stress jump on each edge (in both formulations). Hence, $\eta_{K}$ is straightforward to compute.
\rbl{Finally,} we sum the element contributions to give the \rbl{residual error} estimator and data oscillation error respectively,
\begin{align}\label{errest1}
\eta=\left(\sum_{K\in\mathcal{T}_{h}}\eta_{K}^2 \right)^{1/2} \quad \hbox{and} \quad
\bm{\Theta}=\left(\sum_{K\in\mathcal{T}_{h}}\bm{\Theta}_{K}^2\right)^{1/2}.
\end{align}
\begin{remark}
{The nonuniqueness of the pressure solution in the incompressible limit
is not seen by the error estimator ($p_h$ drops out of $R_k$ when $\lambda\to\infty$
and $\bm{R}_E$ measures inter-element {\it jumps} in the pressure).}
\end{remark}
Theorems \ref{realiab} and \ref{efficie} show that $\eta$ is a reliable and efficient estimator for the energy error associated with locally stabilised $\bm{P}_{1}$--$P_{0}$ approximations of \eqref{scm11a}. The following standard result is needed for Theorem \ref{realiab}.
\begin{lemma}[Cl\'{e}ment interpolation]\label{approxlem11}
Given $\bm{v}\in \bm{V},$ let $\bm{v}_h\in \bm{X}^h_0$ be the quasi-interpolant of $\bm{v}$ defined by averaging\cite{CLA}. For any $K\in\mathcal{T}_h$,
\begin{align*}
\rho^{-1}_K||\bm{v}-\bm{v}_h||_{0,K}&\lesssim (2\mu)^{1/2} |\bm{v}|_{1,\omega_K},
\end{align*}
{where $|\cdot |_{1,\omega_K}$ is the $H^1(\omega_K)$ seminorm. Moreover,} for all $E\in\partial K$ {we have}
\begin{align*}
\rho^{-1/2}_E||\bm{v}-\bm{v}_h||_{0,E}&\lesssim (2\mu)^{1/2} |\bm{v}|_{1,\omega_K},
\end{align*}
where $\omega_K$ is the set of triangles sharing at least one vertex with $K$.
\end{lemma}
\begin{theorem}\label{realiab}
Suppose that $\rblx{(\bm{u},p) \in \bm{H}^1_E\times Q}$ is the weak solution satisfying
(\ref{scm11a}) and $(\bm{u}_h,p_h)\in\bm{X}^h_E\times Q^h$ satisfies \eqref{LSFEA11}. {Suppose further that
$\int_{\partial\Omega} \bm{g} \cdot \bm{n} \, ds =0$ so that $\int_\Omega p = 0 = \int_\Omega p_h$
from \eqref{uniquep}.}
Defining $\eta$ and $\bm{\Theta}$ as in \eqref{errest1}, we have
\begin{align}
|||(\bm{u}-\bm{u}_h, p-p_h)|||\lesssim \eta +\Theta.
\end{align}
\end{theorem}
\begin{proof} Since $(\bm{u}-\bm{u}_{h}, p-p_{h}) \in \rblx{\bm{V} \times Q_0}$, from Lemma \ref{Sinsuplem12}, we have
\begin{align*}
|||(\bm{u}-\bm{u}_h,p-p_h )|||^{2} \lesssim \mathcal{B}(\bm{u}-\bm{u}_h,p-p_h;\bm{v},q)
\end{align*}
for some $(\bm{v}, q) \in \rblx{\bm{V} \times Q_0}$ with
$|||(\bm{v},q)|||\lesssim |||(\bm{u} -\bm{u}_h,{p-p_h})|||$. For this $\bm{v}$, choose $\bm{v}_{h} \in \bm{X}_{0}^{h}$ to be defined as in Lemma \ref{approxlem11}. Then, we have $\mathcal{B}(\bm{u}-\bm{u}_{h}, p-p_{h}, \bm{v}_{h}, 0)=0$ by (\ref{scm11a}a) and (\ref{LSFEA11}a). Hence, since $\nabla \cdot \bm{u}+\kappa^{-1}p=0$ and using (\ref{scm11a}a) again,
\begin{align}\label{rea11}
\mathcal{B}(\bm{u}-\bm{u}_h,p-p_h;\bm{v},q)&=\mathcal{B}(\bm{u}-\bm{u}_h,p-p_h;\bm{v}-\bm{v}_h,q),\nonumber\\
&=(\bm{f},\bm{v}-\bm{v}_h)- a(\bm{u}_h, \bm{v}-\bm{v}_h)
+ (p_h, \nabla\cdot (\bm{v}-\bm{v}_h)) \nonumber\\
& \quad - (q, \nabla\cdot \bm{u}) + (q, \nabla\cdot \bm{u}_h)
- \kappa^{-1}(q, p) + \kappa^{-1}(q, p_h),\nonumber\\
& = (\bm{f},\bm{v}-\bm{v}_h)-a(\bm{u}_h, \bm{v}-\bm{v}_h)
+ (p_h, \nabla\cdot (\bm{v}-\bm{v}_h)) \nonumber\\
& \quad + (q, \nabla\cdot \bm{u}_h + \kappa^{-1} p_{h}),\nonumber\\
&=(\bm{f} - \bm{f}_h,\bm{v}-\bm{v}_h) \, + \! \sum_{K\in \mathcal{T}_h}\!
\Big\{ \big(\bm{f}_h,(\bm{v} - \bm{v}_h) \big)_{{0,K}} \nonumber\\
&\quad + \! \sum_{E\in \partial K} \!\big \langle \bm{R}_E, \bm{v}-\bm{v}_h \big\rangle_E
+ \big(q, R_{K} \big)_{{0,K}} \Big\}
\end{align}
where $ {\big\langle} \bm{R}_E, \bm{v}-\bm{v}_h {\big\rangle}_E=\int_E \bm{R}_E\cdot (\bm{v}-\bm{v}_h) $.
Applying Cauchy--Schwarz to (\ref{rea11}) and then using Lemma \ref{approxlem11} gives
\begin{align}\label{rea12}
||| (\bm{u}-\bm{u}_h,p-p_h )|||^{2} &\lesssim \mathcal{B}(\bm{u}-\bm{u}_h,p-p_h;\bm{v},q
\nonumber \\
& \lesssim \; { |||(\bm{v},q )||| \;
\Bigg(\sum_{K\in \mathcal{T}_h}\Big(\eta_{K}^2+\Theta_K^2\Big)\Bigg)^{1\over 2} }.
\end{align}
\end{proof}
\begin{theorem}\label{efficie}
Suppose that $\rblx{(\bm{u},p) \in \bm{H}^1_E\times Q}$ is the weak solution satisfying \eqref{scm11a} and
$(\bm{u}_h,p_h)\in\bm{X}^h_E\times Q^h$ satisfies \eqref{LSFEA11}. Defining $\eta$ and $\bm{\Theta}$ as in \eqref{errest1}, we have
\begin{align}\label{elowerbd}
\eta\lesssim\, |||(\bm{u}-\bm{u}_h, p-p_h)||| +\Theta.
\end{align}
\end{theorem}
To establish the bound (\ref{elowerbd}), we need to establish efficiency bounds for each of the component residual terms $\eta_{R_K}^{2}$, $\eta_{J_K}^{2}$ and $\eta_{E_K}^{2}$ defined in \eqref{components}.
\begin{lemma}\label{efficie12}
Let $K$ be an element of $\mathcal{T}_h$. The local equilibrium residual satisfies
\begin{align*}
\eta^2_{R_K}&\lesssim \Big(2\mu \; |\bm{u}-\bm{u}_h|_{1,K}^2+ (2\mu)^{-1}||p-p_h||_{0,K}^2+\Theta_K^2\Big).
\end{align*}
\end{lemma}
\begin{proof} The proof follows the same lines as that of Lemma $3.5$ in Ref.~\refcite{KPS}, here using $\bm{R}_{K}= \bm{f}_{h} = \bm{f}_{h} + \nabla\cdot \bm{\sigma}(\bm{u}_{h}, p_{h})$ (since $\bm{u}_{h} \in \bm{P}_{1}$ and $p_{h} \in P_{0}$) and noting that $ \left(\bm{f} + \nabla \cdot \bm{\sigma}(\bm{u}, p)\right)|_{K}=0$ for a classical solution $(\bm{u},p)$ (in both the Herrmann and Hydrostatic formulations). In the Hydrostatic formulation, equation (3.22) in Ref.~\refcite{KPS} has the additional term $\mu (\nabla \cdot(\bm{u} - \bm{u}_{h}), \nabla \cdot \bm{w})_{K}$. Applying the Cauchy--Schwarz inequality to this term as well as the others, leads to the stated result.
\end{proof}
\begin{lemma}\label{efficieR2}
Let $K \in \mathcal{T}_h$. The local mass conservation residual satisfies
\begin{align*}
\eta^2_{J_K}&\lesssim \Big(2\mu \; |\bm{u}-\bm{u}_h|_{1,K}^2+ (2\mu)^{-1}||p-p_h||_{0,K}^2
+ \kappa^{-1} ||p-p_h||_{0,K}^2\Big).
\end{align*}
\end{lemma}
\begin{proof}
Noting that $(\nabla\cdot \bm{u}+ \kappa^{-1} p)|_K=0$ for a classical solution $(\bm{u},p)$, we have
\begin{align*}
\rho_d||\nabla\cdot \bm{u}_{h}+ \kappa^{-1} p_h||_{0,K}^2&=\rho_d||\nabla\cdot (\bm{u}-\bm{u}_{h})+ \kappa^{-1} (p-p_h)||_{0,K}^2\\
&\lesssim \rho_d||\nabla\cdot (\bm{u} -\bm{u}_h ) \, ||_{0,K}^2+ \frac{\rho_d}{\kappa^2} ||( p-p_h )||_{0,K}^2\\
&\lesssim 2\mu \; |\bm{u} -\bm{u}_h |_{1,K}^2+ \kappa^{-1} ||(p-p_h)||_{0,K}^2,
\end{align*}
where the last line follows from the definition of $\rho_d$ in \eqref{gridparams}.
\end{proof}
\begin{lemma}\label{efficieED1}
Let $K \in \mathcal{T}_h$. The stress jump residual satisfies
\begin{align*}
{\eta_{E_K}^2} \lesssim {\sum_{E\in \partial K} \left(2\mu\, |\bm{u}-\bm{u}_h|_{1,\omega_E}^2
+(2\mu)^{-1}||p-p_h||_{0,\omega_E}^2+\Theta_{\omega_E}^2 \right)},
\end{align*}
where $\Theta_{\omega_E}^2 = \sum_{K\in \omega_E}\Theta_K^2$ is the localised data oscillation term and $\omega_{E}$ is the patch of elements that share the edge $E$.
\end{lemma}
\begin{proof} The proof follows the same lines as that of Lemma $3.7$ in Ref.~\refcite{KPS}, but with $\bm{R}_{E}$ defined as in \eqref{stress-jump-def}, replacing $\llbracket (p_{h}\bm{I} - 2 \mu\bm{\varepsilon}(\bm{u}_{h}))\bm{n}\rrbracket_E$ with $\llbracket -\bm{\sigma}(\bm{u}_{h}, p_{h}) \bm{n}\rrbracket_E $
and choosing $\Lambda=\rho_{E} \llbracket -\bm{\sigma}(\bm{u}_{h}, p_{h}) \bm{n}\rrbracket_{E} \, \chi_{E}$. We again exploit the fact that the classical solution $(\bm{u}, p)$ satisfies $-\nabla \cdot \bm{\sigma}(\bm{u}, p) = \bm{f}$ and $\nabla \cdot \bm{\sigma}(\bm{u}_{h}, p_{h})=0.$ To obtain the upper bound for $\rho_{E} \| \bm{R}_{E}\|_{0,E}^{2}$ in the proof of Lemma $3.7$ there is an additional term to bound for each of the terms $T_{1}$ and $T_{3}$. However the same upper bounds hold.
\end{proof}
\rbl{The desired local lower bound}
(\ref{elowerbd}) follows by consolidating the estimates from Lemma~\ref{efficie12}, Lemma~\ref{efficieR2} and Lemma~\ref{efficieED1}.
\subsection{\rbl{A Poisson problem local error estimator}}\label{SLEPE}
Having established that the residual error estimator $\eta$ in (\ref{errest1}) is reliable and efficient, the framework established by Verf\"urth\cite{RV} makes it straightforward to construct equivalent {\it local problem} estimators that are equally reliable but potentially more efficient. For the Herrmann formulation with $\bm{Q}_{2}$ (biquadratic) displacement approximation, four local problem error estimators were discussed in Ref. \refcite{KPS}. Of these, the so-called Poisson problem estimator \rbl{proved to be} the most attractive
\rbl{from a computational perspective.}
\rbl{This strategy will be extended to cover stabilised $\bm{P}_{1}$--$P_{0}$ approximation herein.}
\rbl{We compute} a local estimator $\bm{e}_{P,K} \in ({\mathcal P}(K))^{2}$ for the displacement error
that is \rbl{{\it super-quadratic}} in each component and a local estimator $\epsilon_{S,K} \in P_{1}(K)$ for the pressure error
that is linear.
\rbl{More specifically,} for the displacement error, we define
\begin{align}
\mathcal{P}(K) = \textrm{span}\{ \psi_{E}, \, E \in \partial K \cap \left( {\mathcal{E}}_{h} \setminus \partial \Omega\right) \}
\rbl{\>\oplus\> B_T},
\end{align}
where $\psi_{E}$ is a quadratic bubble function associated with an interior edge $E$
\rbl{and $B_T$ is the space spanned by the cubic bubble function that is zero on the three boundary edges.}
\rbl{We assume} that every triangle $K \in {\mathcal T}_{h}$ has at least two edges in the interior of $\Omega$.
\rbl{See} Kay \& Silvester\cite{DD} (and references therein) where the same error estimation strategy
\rbl{is applied to} Stokes problems.
The Poisson problem estimator is now defined by
$$\eta_{P} = \sqrt{\sum_{K \in {\cal T}_{h}} \eta_{P,K}^{2}},$$
where the local contributions are given by
\begin{align}\label{poiloc1}
\eta_{P,K}^2= 2\mu \, ||\nabla \bm{e}_{P,K}||^2_{0,K} +\rho_d^{-1}||\epsilon_{P,K}||^2_{0,K},
\end{align}
and $(\bm{e}_{P,K},\epsilon_{S,K})\in (\mathcal{P}(K))^{2} \times {P}_1(K)$ is the solution to the following problem
\begin{subequations} \label{poiloc}
\begin{align}\label{poiloc2}
2\mu \, (\nabla\bm{e}_{P,K},\nabla\bm{v})_K &=(\bm{R}_K,\bm{v})_K
-\sum_{E\in\partial K} \langle \bm{R}_E, \bm{v} \rangle_E, \quad \forall \bm{v}\in (\mathcal{P}(K))^{2}, \\
\rho_d^{-1}(\epsilon_{P,K}, q)_K &=(R_K,q)_K, \quad \forall q\in P_1(K).\label{poiloc3}
\end{align}
\end{subequations}
Recall that $\rho_{d}$ is defined in \eqref{gridparams}, $\bm{R}_{K}$ and $R_{K}$ are defined in \eqref{elt_resid} and $\bm{R}_{E}$ is defined in \eqref{stress-jump-def}. With the exception of $\bm{R}_{K}$, these quantities are slightly different depending on which mixed formulation is used. In both cases, (\ref{poiloc2}) decouples into a
pair of local \rbl{$4 \times 4$} Poisson problems and since $R_{K} \in P_1(K)$, the
solution of (\ref{poiloc3}) is immediate: $\epsilon_{P,K} = \rho_d R_K = \rho_d (\nabla\cdot \bm{u}_h + \kappa^{-1} p_h)$. Hence, (\ref{poiloc1}) simplifies to
$$\eta_{P,K}^2 =2\mu \, ||\nabla \bm{e}_{P,K}||^2_{0,K} + \rho_d || \nabla\cdot \bm{u}_h + \kappa^{-1} p_h ||^2_{0,K}.$$
We note that this strategy of decoupling the components of local problem error estimators in a mixed setting it not new; it was pioneered by Ainsworth \& Oden\cite{MJ}.
\rbl{Using the arguments that are} sketched in Ref.~\refcite{KPS}, \rbl{the equivalence result}
$$\eta_{P,K} \lesssim \eta_{K} \lesssim \eta_{P,K}, \qquad K \in {\mathcal T}_{h},$$
\rbl{is easily established.}
\section{Computational results}\label{Numres}
\rbl{In this section we}
compare the performance of the estimators $\eta$ and $\eta_P$ for the Herrmann and Hydrostatic formulations of three test problems.
\rbl{All results} were computed using locally stabilised $\bm{P}_{1}$--$P_{0}$ approximation with software adapted from the MATLAB
toolbox TIFISS\cite{TIFISS}. To define the stabilisation term, we group the elements in the meshes into disjoint macroelements
consisting of four neighbouring triangles, {with a central element connected to three neighbours\cite{ks92}}.
In some experiments we use uniform meshes and in others we use the local contributions
$\eta_{K}$ and $\eta_{P,K}$ to drive \emph{adaptive} mesh refinement.
More precisely, starting with an initial mesh $\mathcal{T}_{0},$ we apply the iterative refinement loop
\begin{align*}
\mbox{Solve}\rightarrow\mbox{Estimate}\rightarrow\mbox{Mark}\rightarrow\mbox{Refine}
\end{align*}
to generate a sequence of (nested) regular meshes $\{\mathcal{T_\ell}\}$ with mesh size $h_{\ell}$. For each $\mathcal{T_\ell}$ and the associated finite element approximation, we compute $\eta_{\ell}^{2}=\sum_{K\in {\mathcal{T}}_{\ell}} \eta_{K}^{2}$ (if using the residual estimator), or else replace $\eta_{K}$ with $\eta_{P,K}$ (if using the Poisson estimator). Then, in the usual way\cite{Doerfler}, using a bulk parameter $\theta\in(0,1)$ (here $\theta=1/2$), we determine a minimal subset $\mathcal{M}_\ell$ of marked triangles such that $\sum_{K\in {\mathcal{M}}_{\ell}} \eta_{K}^{2} \ge \theta\eta_\ell^2$ (and similarly with $\eta_{P,K}$). Mesh refinement is then done using the red-green-blue strategy\cite{RV}. We denote the number of degrees of freedom associated with the mesh $\mathcal{T}_{\ell}$ by $N_{\ell}.$ Hence, for uniform meshes we have $\mathcal{O} (N^{-r}_{\ell})\approx \mathcal{O}(h^{2r}_{\ell})$ where $r>0$.
{From Theorem \ref{apriori-thm} we know that, if the solution} $(\bm{u}, p)$ is sufficiently smooth, \rbl{then the
energy error $e=|||(\bm{u}-\bm{u}_{h}, p-p_{h}) |||$ will decay to zero with} rate $r=0.5$.
\subsection{\rbl{An analytic solution}}
The first test problem is taken from Ref.~\refcite{CJ}. We choose $\Omega=(0,1)\times (0,1)$ and a zero essential boundary
condition; \rbl{that is,} $\bm{g}=\bm{0}$ on $\partial \Omega$. In addition,
\begin{align*}
\bm{f}= \left(\begin{array}{c} -2\mu\pi^3\cos(\pi y)\sin(\pi y)(2\cos(2\pi x) -1) \\ 2\mu\pi^3\cos(\pi x)\sin(\pi x)(2\cos(2\pi y)-1)
\end{array} \right).
\end{align*}
The exact solution is $p=0$ and $\textbf{u}=(u_1,u_2)^{\top}$
where
\begin{align*}
u_1&=\pi\cos(\pi y)\sin^2(\pi x)\sin(\pi y),\quad
u_2= -\pi\cos(\pi x)\sin^2(\pi y)\sin(\pi x).
\end{align*}
Figures~\ref{Ex1mu100pes} and \ref{Ex1mu100res} show the convergence behaviour of the exact error $e$ as well as the estimated errors obtained with $\eta_{P}$ and $\eta$, respectively, using adaptively generated meshes.
(The initial mesh ${\mathcal{T}}_{0}$ was \rbl{generated} with $N_{0}=1,090$ degrees of freedom.)
Here, $\mu$ is fixed and we consider two values of the Poisson ratio $\nu$. The estimated errors converge to zero at the optimal rate ($r=0.5$). \rbl{While both estimators are obviously efficient and reliable for either formulation, the results in Figure \ref{Ex1mu100pes}
show that the Poisson estimator is the more accurate of the two---the effectivity indices} for the Poisson estimator
are close to unity even when $\nu \to 1/2$.
Identical results (not reported) were obtained when the experiments were repeated with $\mu=1$ and $\mu=0.01$. We conclude that both \rbl{estimation strategies} are robust with respect to variations in the parameters $\mu$ and $\nu$.
\begin{figure}[th!]
\centering
\includegraphics[width=.41\textwidth]{./P1P0_mu100nu4_Pesterr}
\label{ex1mu100nu4}
\centering
\includegraphics[width=.41\textwidth]{./P1P0_mu100nu49999_Pesterr}
\label{ex1mu100nu49999}
\centering
\includegraphics[width=.41\textwidth]{./P1P0_mu100nu4_Pesterrhydro}
\label{ex1mu100nu4hydro}
\centering
\includegraphics[width=.41\textwidth]{./P1P0_mu100nu49999_Pesterrhydro}
\label{ex1mu100nu49999hydro}
\caption
{Exact ($e$) and estimated (using the local Poisson estimator $\eta_{P}$) energy errors computed using adaptive meshes, \rbl{for} Herrmann (top) and Hydrostatic (bottom) formulations of test problem 1,
with $\mu=100$ and $\nu=0.4$ (left); $\mu=100$ and $\nu=0.49999$ (right).}
\label{Ex1mu100pes}
\includegraphics[width=.38\textwidth]{./P1P0_mu100nu4_Resterr}
\label{ex1mu100nu4res}
\centering
\includegraphics[width=.41\textwidth]{./P1P0_mu100nu49999_Resterr}
\label{ex1mu100nu49999res}
\centering
\includegraphics[width=.41\textwidth]{./P1P0_mu100nu4_Resterrhydro}
\label{ex1mu100nu4reseff}
\centering
\includegraphics[width=.41\textwidth]{./P1P0_mu100nu49999_Resterrhydro}
\label{ex1mu100nu49999reseff}
\caption{
Exact ($e$) and estimated (using the residual estimator $\eta$) energy errors computed using adaptive meshes, \rbl{for}
Herrmann (top) and Hydrostatic (bottom) formulations of test problem 1,
with $\mu=100$ and $\nu=0.4$ (left); $\mu=100$ and $\nu=0.49999$ (right).}
\label{Ex1mu100res}
\end{figure}
\subsection{\rbl{A nonsmooth solution}}
The second \rbl{test problem is taken from} Ref.~\refcite{wihler2004locking}.
Again, $\Omega=(0,1)\times (0,1)$ but now $\bm{f}=\bm{0}$ and we impose the
condition $\bm{u}=(g,0)^{\top}$ on $\partial \Omega$, where
\begin{align*}
g=\Big\{\begin{array}{lc}
(1-4(x-\frac{1}{2})^2)^{\frac{1}{2}+\alpha}, &\mbox{on }\; [0,1] \times \{1\},\\
0,& \hbox{elsewhere on } \partial \Omega.
\end{array}
\end{align*}
If $\alpha\in(0,\frac{1}{2})$, then the displacement exhibits $H^{\frac{3}{2}+\alpha}$--regularity. Specifically, there are
singularities at the {top two corners of the domain}. We set the specific value $\alpha=0.1$ so that $\bm{u} \in \rbl{\bm H^{1.6}}(\Omega)$.
This lack of smoothness is reflected in the convergence behaviour of the estimated energy error.
Results obtained with the Poisson estimator $\eta_P$ on \emph{uniformly} refined meshes are shown in Figure~\ref{Ex2nu4uni}.
Our results suggest that for both the Herrmann and Hydrostatic formulations, the error converges to zero at \rbl{the anticipated}
suboptimal rate ($\rbl{r= 0.3}$). However, when we use \emph{adaptively} refined meshes, for both the Herrmann and Hydrostatic
formulations, we recover the optimal convergence rate \rbl{of $r=1/2$}, \rbl{as shown in Figure \ref{Ex2c2}.}
\rbl{Starting from an initial mesh with $N_{0}=1,090$ degrees of freedom,
the singular solution behaviour is detected and strong refinement occurs near the top corners.
Figure \ref{Ex2c2adamesh} shows the meshes that are generated at the first refinement step where $N_{\ell} \ge 10^{4}$.}
\begin{figure}
\centering
\includegraphics[width=.47\textwidth]{./P1P0_mu1nu4_Pesterrdisuni}
\label{exc2nu4PEEuni}
\includegraphics[width=.46\textwidth]{./P1P0_mu1nu4_Pesterrdishydrouni}
\label{exc2nu49999PEEuni}
\caption
{Estimated energy errors (using the estimator $\eta_{P}$) computed with \emph{uniform} meshes, \rbl{for}
Herrmann (left) and Hydrostatic (right) formulations of test problem 2 with $\mu=1, \nu=0.4$.}
\label{Ex2nu4uni}
\includegraphics[width=.44\textwidth]{./P1P0_mu1nu4_Pesterrdissing}
\includegraphics[width=.43\textwidth]{./P1P0_mu1nu49999_Pesterrdissing}
\centering
\includegraphics[width=.45\textwidth]{./P1P0_mu1nu4_Pesterrdissinghydro}
\includegraphics[width=.45\textwidth]{./P1P0_mu1nu49999_Pesterrdissinghydro}
\caption{Estimated energy errors (using $\eta_{P}$) computed using \emph{adaptive} meshes, \rbl{for}
Herrmann (top) and Hydrostatic (bottom) formulations of test problem 2 with $\mu=1$ and $\nu=0.4$ (left); $\mu=1$
and $\nu=0.49999$ (right).}
\label{Ex2c2}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=.46\textwidth]{./P1P0_mu1nu49999_Pestmeshdissing12762}
\label{exc2nu49999ada}
\centering
\includegraphics[width=.45\textwidth]{./P1P0_mu1nu49999_Pestmeshdissing11846hydro}
\label{exc2nu49999hydada}
\caption
{Adaptive meshes generated using $\eta_{P}$ \rbl{for}
Herrmann (left) and Hydrostatic (right) formulations of test problem 2 with $\nu=0.49999$.
$N_{\ell}=12,762$ (left), and $N_{\ell}=11,846$ (right).}
\label{Ex2c2adamesh}
\end{figure}
\subsection{\rbl{A singular solution}}
\rbl{To conclude, we discuss a test problem that is considered} in Refs.~\refcite{CJ} and \refcite{wihler2004locking}.
\rbl{The problem is posed in} an L-shaped domain $\Omega =(-1,1)\times (-1,1)\setminus (-1,0]\times(-1,0]$. In polar coordinates, the exact displacement is
\begin{align*}
\bm{u} = \frac{r^\alpha}{2\mu} \left( \begin{array}{c} -(\alpha+1)\cos((\alpha+1)\phi)+(C_2-\alpha+1)C_1\cos((\alpha-1)\phi) \\
(\alpha+1)\sin((\alpha+1)\phi)+(C_2+\alpha-1)C_1\sin((\alpha-1)\phi) \end{array} \right),
\end{align*}
where $\alpha=0.544483736782$ is a positive solution of $\alpha\sin(\omega)+\sin(2\omega\alpha)=0$ with
\begin{align*}
\omega=\frac{3\pi}{4}, \quad C_1=-\frac{\cos((\alpha+1)\omega)}{\cos((\alpha-1)\omega)},\quad C_2=\frac{2(\lambda+2\mu)}{\lambda+\mu}.
\end{align*}
The body force is $\bm{f}=\bm{0}$ and the \rbl{nonzero} essential boundary data $\bm{g}$ \rbl{is represented by the piecewise linear interpolant of the given solution}. \rblx{Note that $\int_{\partial \Omega} \bm{g} \cdot \bm{n} \, ds \neq 0$.}
To compute the Lam\'{e} constants $\lambda$ and $\mu$, we choose $E=10^5$ and \rbl{set $\nu=0.4$ or $0.49999$}.
Note that the exact displacement $\bm{u}$ is analytic inside the domain $\Omega$ but $\nabla \bm{u}$ is singular at the origin, so $\bm{u}\notin \bm{H}^2(\Omega)$. This lack of smoothness is reflected in the convergence behaviour of the estimated energy error. Results computed with the Poisson estimator $\eta_P$ on \emph{uniform} meshes are shown in Figure~\ref{LshapeE105uni}.
As \rbl{in the second test problem, we observe the estimated errors converge at a suboptimal rate (here
$r\approx 0.27$)}.
{Moreover, when we use \emph{adaptively} refined meshes, we recover the optimal rate of convergence of $r=0.5$.
This is shown in Figure~\ref{LshapeE105ada}.}
The singular solution behaviour is detected and strong refinement is \rbl{generated around the re-entrant corner}.
{While the effectivity indices in Figure~\ref{LshapeE105ada}
are not quite as impressive as those in Figure~\ref{Ex1mu100pes} they remain close to unity
(approximately 1.35 when $\nu=0.4$ and 1.6 when $\nu=0.49999$).}
{We infer from these results that $\eta_P$ provides \rblx{an} efficient and reliable error estimate for both
Herrmann and Hydrostatic formulations.}
\begin{figure}
\centering
\includegraphics[width=.4\textwidth]{./P1P0_E1000000nu4_Pesterr}
\includegraphics[width=.4\textwidth]{./P1P0_E1000000nu49999_Pesterr}
\includegraphics[width=.4\textwidth]{./P1P0_E1000000nu4_Pesterrhydro}
\includegraphics[width=.4\textwidth]{./P1P0_E1000000nu49999_Pesterrhydro}
\label{LshapeE105nu49999hydro}
\caption
{Estimated energy errors (using $\eta_{P}$) computed with \emph{uniform} meshes \rbl{for}
Herrmann (top) and Hydrostatic (bottom) formulations of test problem 3 with $\nu=0.4$ (left); $\nu=0.49999$ (right).}
\label{LshapeE105uni}
\includegraphics[width=.42\textwidth]{./P1P0_E1000000nu4_Pesterrada}
\includegraphics[width=.42\textwidth]{./P1P0_E1000000nu49999_Pesterrada}
\includegraphics[width=.42\textwidth]{./P1P0_E1000000nu4_Pesterrhydroada}
\includegraphics[width=.42\textwidth]{./P1P0_E1000000nu49999_Pesterrhydroada}
\caption
{Estimated energy errors (using $\eta_{P}$) computed with \emph{adaptive} meshes \rbl{for}
Herrmann (top) and Hydrostatic (bottom) formulations of test problem 3 with $\nu=0.4$ (left); $\nu=0.49999$ (right).}
\label{LshapeE105ada}
\end{figure}
\section{\rbl{Concluding remarks}}\label{conclusions}
\rbl{There are two important contributions in this paper.
First, we have developed a low-order mixed finite element method
for computing locking-free approximations of linear elasticity problems. The method is computationally cheap
and challenges the conventional wisdom that it is necessary to start from an inf-sup stable pair of
finite element spaces. The stabilisation term is weighted by the problem specific factor {of $1/2\mu$}
but is otherwise parameter-free. {Our} a priori error analysis shows that the method provides a
robust approximation of the energy error.
That is, the constants in the error bounds do not depend on the Lam\'e coefficients.
Second, we have described a practical error estimation strategy---based on
solving uncoupled Poisson problems for each displacement component---that
give effectivity indices that are close to unity in all cases that have been tested.
Ensuring robustness in the error estimation process is fundamentally important
when solving problems with large variability in the measurement of material
parameters. Extending this work to enable the adaptive solution of elasticity problems
with {\it uncertain} material parameters is the subject of ongoing research.}
\bibliographystyle{plain}
|
\section{Introduction}\label{sec:intro}
\noindent The size of the digital footprint left behind by users interacting
with online services is increasing at a rapid pace, due to the popularity of
location based services, social networks and related online services. An
important portion of this footprint contains spatio-temporal references and is
a fertile resource for applications for social good and business
intelligence~\cite{ref:jensen}. We refer to the services that create spatio-
temporal records of their usage as \emph{Location Enhanced Services} (LES).
For instance, Foursquare/Swarm\footnote{www.foursquare.com / www.swarmapp.com}
--- a popular social networking service, records the locations of users when
they check-in at a point-of-interest (POI) registered in the system.
We consider two varieties of LES based on a user's level of involvement in the
production of spatio-temporal usage records. Users of \emph{explicit LES}
actively participate in sharing their spatio-temporal information. Location-
based social network services, like Foursquare/Swarm, are well-known examples
of such services, where the user explicitly checks-in to a particular POI at a
particular time. On the other hand, \emph{implicit LES} produce spatio-
temporal records of usage as a byproduct of a different activity, whose focus
is not sharing the location. For instance, when a user makes a payment with
her credit card, a record is produced containing time of the payment and
location of the store. Same applies for the cell phone calls, since
originating cell tower location is known to the service provider.
We postulate that the spatial-temporal usage records belonging to the same
real-world entities can be matched across records from two different LESs.
Linking spatio-temporal records enables data scientists and service providers
to obtain information that they cannot derive by analyzing only one set of
usage records. For example, a LES provider can combine user segmentation
results derived from its own usage records with social segmentation results
derived from the publicly available Foursquare/Swarm records, assuming that
the linking of users across these two datasets can be performed effectively.
Data from multiple LES providers can be combined using common spatio-temporal
information to identify general patterns and improve urban life. While
possible, there are a number of challenges associated with performing such
linkage of entities across two spatio-temporal usage record datasets.
First, unlike in traditional record linkage~\cite{ref:survey1, ref:survey2,
ref:survey3}, where it is easier to formulate linkage based on a traditional
similarity measure defined over records (such as Minkowski distance or Jaccard
similarity), in spatio-temporal linkage similarity needs to be defined based
on time, location, and the relationship between the two. For a pair of
entities from two different datasets to be considered similar, their usage
history must contain records that are close both in space and time. Equally
importantly there must not be \emph{negative matches}, such as records that
are close in time, but far in distance. We call such negative matches,
\emph{alibi}s. To address these challenges, in this paper, we introduce a
novel \emph{linkage model} based on $k$-$l$ \emph{diversity} --- a concept we
developed to capture both spatial and temporal diversity aspects of the
linkage. Informally, a pair of entities, one from each dataset, is called
$k$-$l$ diverse if they have at least $k$ co-occurring records (both
temporally and spatially) in at least $l$ different locations. However, as we
will detail later not all co-occurring records contribute fully and equally to
the overall aggregation. Furthermore, number of alibi events of such pairs
should not exceed a predefined threshold.
Second, na\"{i}ve record linkage algorithms that compare every pair of records
take $\mathcal{O}(n^2)$ time~\cite{ref:tutorial}, where $n$ is the number of
records. However, such a computation would not scale to large dataset sizes
that are typically involved in LES. Considering that location-based social
networks get millions of updates every day, processing of hundreds of days of
data for the purpose of linkage would take impractically long amount of time.
In order to link entities in a reasonable time, the \emph{ST-Link} algorithm
we have developed uses two filtering steps before pairwise comparisons of
candidate entities are performed to compute the final linkage. Taking
advantage of the spatio-temporal structure of the data, \emph{ST-Link} first
distributes entities over coarse-grained geographical regions that we call
\emph{dominating grid cells}. Such grid cells contain most of the activities
of their corresponding entities. For two entities to link, they must have a
common dominating grid. Once this step is over, the linkage is independently
performed over each dominating grid cell. During the temporal filtering step,
\emph{ST-Link} uses a sliding window based scan to build candidate entity
pairs, while also pruning this list as alibis are encountered for the current
candidate pairs. It then performs a reverse scan to further prune the
candidate pair set by finding and applying alibis that were not known during
the forward scan. Finally, our complete linkage model is evaluated over
candidate pairs of entities that remain following the spatial and temporal
filtering steps. Pairs of entities that satisfy $k$-$l$ diversity are linked
to each other.
This paper makes the following contributions:
\smallskip\\
\indent$\bullet$\textbf{Model.} We introduce a novel spatio-temporal linkage model based on
the concept of $k$-$l$ diversity for matching.
\smallskip\\
\indent$\bullet$\textbf{Algorithm}. To realize the linkage model in practice, we develop the
\emph{ST-Link} algorithm. \emph{ST-Link} applies spatial and temporal
filtering techniques to effectively prune the candidate entity pairs in order
to scale to large datasets. It also performs mostly sequential I/O to further
improve performance.
\smallskip\\
\indent$\bullet$ \textbf{Evaluation.} We
provide an experimental study using several datasets to showcase the
effectiveness of the $k$-$l$ diversity based linkage model and the efficiency
of the \emph{ST-Link} algorithm.
\smallskip\\
The rest of this paper is organized as follows.
Section~\ref{sec:formalization} gives the formalization of the $k$-$l$
diversity based linkage model. Section~\ref{sec:ST-Link} explains the \emph
{ST-Link} algorithm for an effective realization of our linkage model.
Section~\ref{sec:experiments} presents the experimental evaluation.
Section~\ref{sec:related} gives the related work and
Section~\ref{sec:conclusion} concludes the paper.
\section{Spatio-temporal Linkage}
\label{sec:formalization}
\noindent
In this section, we introduce our $k$-$l$ diversity based spatio-temporal
linkage model. We first present the preliminaries, including the notation used,
and then present the detailed formalization of the linkage model.
\subsection{Notation and Preliminaries}\label{sec:prelim}
\noindent
\textbf{Datasets.} We denote the two spatio-temporal usage record
datasets from the two LES across which the linkage is to be performed as
$\mathcal{I}$ and $\mathcal{E}$.
\medskip
\\
\noindent \textbf{Entities and events.} \emph{Entities, or users,} are real-world systems or people who
use LES. Throughout this paper, the terms user and entity will be used interchangeably.
They are represented in the datasets with their ids, which are
different for the two LES. \emph{Events} correspond to usage records generated
by a LES as a result of users interacting with the service. For an event
$e\in\mathcal{E}$ (or $i\in\mathcal{I}$), $e.u$ (or $i.u$) represents the user
associated with the event. We use $U_\mathcal{E}$ and $U_\mathcal{I}$ to
denote the set of user ids in the datasets $\mathcal{E}$ and $\mathcal{I}$,
respectively. We have $U_\mathcal{E}=\{e.u: e\in\mathcal{E}\}$ and
$U_\mathcal{I}=\{i.u: i\in\mathcal{I}\}$.
\medskip
\\
\noindent \textbf{Location and time.} Each event in the dataset contains
location and time information. The location information is in the form of a
region, denoted as $e.r$ for event $e$. We do not use a point for location, as
for most LES the location information is in the form of a region (e.g., POI in
check-ins, cell towers in calls). The time information is a point in time,
denoted as $e.t$ for event $e$. Although an event might contain a time period
as well (e.g., call start time and duration), frequently those records contain
only the start location (e.g., originating cell tower), and thus it would be
incorrect to assume the same location for the entire duration. However, if the
time information of an event is a period, and the associated locations are
known, this event could be represented as multiple events,
each with its own location information and time point (details given in Section~
\ref{sec:timePeriod}).
\medskip
\\
\noindent \textbf{Linkage.} Our goal is to come up with a linkage
function $\mathcal{L}$, where $\mathcal{L}(\mathcal{E},
\mathcal{I})\subseteq U_\mathcal{E} \times U_\mathcal{I}$. Each pair in
the result, that is $(u_1, u_2)\in \mathcal{L}(\mathcal{E},\mathcal{I})$,
represents a potential linkage. We only consider user pairs $(u_1, u_2)$ for
which there is no ambiguity in the linkage, that is $\nexists u\neq u_1 \mbox{
s.t. } (u, u_2)\in
\mathcal{L}(\mathcal{E}, \mathcal{I}) \,\wedge\, \nexists u\neq u_2 \mbox{ s.t.
} (u_1, u)\in \mathcal{L}(\mathcal{E},\mathcal{I})$.
\subsection{$k$-$l$ Diversity based Linkage}
\noindent
The core idea behind our linkage model is to locate pairs of users
whose events satisfy $k$-$l$ diversity. Stated informally, a pair of users is
called $k$-$l$ diverse if they have at least $k$ co-occurring events (both
temporally and spatially) in at least $l$ different locations. Furthermore,
number of alibi events of such pairs should not exceed a predefined threshold.
In what follows we provide a
number of definitions that help us formalize the proposed $k$-$l$ diversity.%
\medskip
\\
\noindent \textbf{Co-occurrence.} Two events from different datasets are
called co-occurring if they are close in space and time.
Eq.~\ref{eq:co-place} defines the $P$ relationship to capture the closeness
in space. For two records $i\in\mathcal{I}$ and $e\in\mathcal{E}$, $P$ is
defined as:
\begin{equation}
P(i,e) \equiv (i.r \cap e.r) \neq \emptyset,
\label{eq:co-place}
\end{equation}
\noindent where $i.r$ and $e.r$ are the regions of the two events. While we
defined the closeness in terms of intersection of regions, other approaches are
possible, such as the fraction of the intersection being above a threshold:
$|i.r \cap e.r|/\text{min}(|i.r|, |e.r|) \geq \epsilon$. Our methods are equally
applicable to such measures.
Eq.~\ref{ref:co-time} defines the $T$ relationship to capture the closeness of
events in time:
\begin{equation}
T(i,e) \equiv |i.t-e.t| \le \alpha.
\label{ref:co-time}
\end{equation}
\noindent Here, we use the $\alpha$ parameter to restrict the matching events
to be within a window of $\alpha$ time units of each other. Using
Eq.~\ref{eq:co-place} and Eq.~\ref{ref:co-time}, we define the
\emph{co-occurrence} function $C$ as:
\begin{equation}
C(i,e) \equiv T(i,e) \wedge P(i,e)
\label{eq:co-occur}
\end{equation}
\noindent
\textbf{Alibi.} While a definition of similarity is necessary to link events
from two different datasets, a definition of dissimilarity is also required
to rule out pairs of users as potential matches in our linkage. Such
\emph{negative matches} enable us to rule out incorrect matches and also
reduce the space of possible matches throughout the linkage process. We refer
to these negative matches as \emph{alibi}s.
By definition alibi means \textit{``A claim or piece of evidence that one was
elsewhere when an act is alleged to have taken place''}. In this paper we use
alibi to define events from two different datasets that happened around the
same time but at different locations, such that it is not possible for a user to
move from one of these locations to the other within the duration defined by the
difference of the timestamps of the events. To formalize this, we define a
\emph{runaway} function $R$, which indicates whether locations of two events
are close enough to be from the same user based on their timestamps. We define
$R$ as follows:
\begin{equation}
R(i,e) \equiv d(i.r, e.r) \leq \lambda \cdot |i.t - e.t|
\label{ref:runaway}
\end{equation}
\noindent Here, $\lambda$ is the maximum speed constant and $d$ is a function
that gives the shortest distance between two regions. If the distance between
the regions of two events is less than or equal to the distance one can travel
at the maximum speed, then we cannot rule out linkage of users associated with
these two events. Otherwise, and more importantly, these two events form an
alibi, which proves that they cannot belong to the same user. Based on this,
we define an alibi function, denoted by $A$, as follows:
\begin{equation}
A(i,e) \equiv T(i,e) \wedge \neg P(i,e) \wedge \neg R(i,e)
\label{eq:alibi}
\end{equation}
\noindent\textbf{User linkage.} The definitions we have outlined so far are on
pairs of events, and with these definitions at hand, we can now move on to
definitions on pairs of users. Let $x\in U_{\mathcal{I}}$ and $y\in
U_{\mathcal{E}}$ be two users. We use $\mathcal{I}_x$ to denote the events of
user $x$ and $\mathcal{E}_y$ to denote the events of user $y$. In order to be
able to decide whether two users are the same person or not, we need to define
a matching between their events.
Initially, let us define the set of all co-occurring events of users $x$
and $y$, represented by the function $F$. We have:
\begin{equation} \label{eq:set}
F(x,y) = \left\{ {(i,e) \in\mathcal{I}_x \times \mathcal{E}_y: C(i,e) } \right\}
\end{equation}
\noindent $F$ is our \emph{focus set} and contains pairs of co-occurring
events of the two users. However, in this set, some of the events may be
involved in more than one co-occurring pairs. We restrict the matching between
the events of two users by disallowing multiple co-occurring event pairs
containing the same events. Accordingly, we define $\mathcal{S}$ as the set
containing all possible subsets of $F$ satisfying this restriction. We call
each such subset an \emph{event linkage set}. Formally, we have:
\begin{equation}
\begin{split}
\mathcal{S}(x,y) &= \{S \subseteq F(x,y):\\
& \nexists \{ (i_1, e_1), (i_2, e_2)\} \subseteq S \text{ s.t. } i_1=i_2 \vee e_1=e_2\}
\end{split}
\end{equation}
We say that the user pair $(x, y$) satisfy $k$-$l$ diversity if there is at
least one event linkage set $S\in \mathcal{S}(x, y)$ that contains $k$
co-occurring event pairs and at least $l$ of them are at different locations.
However, each co-occurring event pair does not count as $1$, since there could
be many other co-occurring event pairs outside of $S$ or even $F$
that involve the same events. As such, we weight these co-occurring event
pairs (detailed below).
Figure~\ref{fig:match} shows a sample event linkage set with weights
for the co-occurring event pairs.
\medskip
\begin{figure}[!t]
\centering
\includegraphics[width=.85\linewidth]{matching.png}
\caption{Sample event linkage set (solid lines) for users $x$ and $y$.
The co-occurring event pairs are shown using dashed lines. Events from a given
user are shown within circles. Users $a$, $b$, $c$, and $y$ are from one LES,
and the users $d$, $e$, $f$, and $x$ are from the other LES.}
\label{fig:match}
\end{figure}
\noindent\textbf{$k$ co-occurring event pairs.}
Let $S$ be an event linkage set in $\mathcal{S}(x,y)$ and let $\mathcal{C}$ be
a function that determines whether the co-occurring event pairs in $S$
satisfy the co-occurrence condition of $k$-$l$ diversity. We have:
\begin{equation} \label{eq:cooccur}
\mathcal{C}(S) \equiv \sum_{(i,e) \in S} w(i,e) \geq k
\end{equation}
The weight of a co-occurring event pair is defined as:
\begin{equation}
\label{eq:weight}
\begin{split}
w(i,e) = &|\{i_1.u: C(i_1,e) \wedge i_1\in \mathcal{I}\}|^{-1} \cdot \\
&|\{e_1.u: C(i,e_1) \wedge e_1\in \mathcal{E}\}|^{-1}
\end{split}
\end{equation}
Here, given a co-occurring event pair between two users, we check how many
possible users' events could be matched to the these events. For instance,
in the figure, consider the solid line at the top with the weight $1/6$.
The event on its left could be matched to events of $2$ different users,
and the event on its right could be matched to events of $3$ different users.
To compute the weight of a co-occurring pair, we multiply the inverse of these
user counts, assuming the possibility of matching from both sides are independent.
As such, in the figure, we get $1/2\cdot 1/3=1/6$.
\medskip
\\
\noindent\textbf{$l$ diverse event pairs.}
\noindent For $S\in \mathcal{S}$$(x,y)$ to be $l$-diverse, there needs to be
at least $l$ unique locations for the co-occurring event pairs in it. However,
for a location to be counted towards these $l$ locations, the weights of the
co-occurring event pairs for that location must be at least 1. Let
$\mathcal{D}$ denote the function that determines whether the co-occurring
event pairs in $S$ satisfy the diversity condition of $k$-$l$ diversity. We
have:
\begin{equation}
\label{eq:diversity}
\mathcal{D}(S) \equiv | \{ p\in \mathcal{P} : \sum_{\stackrel{(i,e) \in S \text{ s.t. }}
{p\,\cap\, i.r \,\cap\, e.r \neq \emptyset }} w(i,e) \geq 1 \} | \geq l
\end{equation}
Here, one subtle issue is defining a unique location. In
Eq.~\ref{eq:diversity} we use $\mathcal{P}$ as the set of all unique locations. This
could simply be a grid-based division of the space. In our experiments, we use
the regions of the Voronoi diagram formed by cell towers as our set of unique
locations.
\begin{figure*}[!t]
\centering
\includegraphics[width=\linewidth]{er_architecture.png}
\caption{Data processing pipeline of ST-Link.}\label{fig:data-pipeline}
\end{figure*}
\iffalse
Before we can formally state the $k$-$l$ diversity based linkage, we have to
define the alibi relation for user pairs. Let $\mathcal{A}$ denote a function
that determines whether there exists at least one alibi for a given pair of
users.Note that even having a single alibi is enough to decide that user $x$ and
$y$ are not the same person. We have:
\begin{equation}\label{eq:noAlibi}
\mathcal{A}(x,y) \equiv \exists (i,e) \in \mathcal{I}_x \times
\mathcal{E}_y \text{ s.t. } A(i,e)
\end{equation}
\fi
Before we can formally state the $k$-$l$ diversity based linkage, we have to
define the alibi relation for user pairs. Let $\mathcal{A}$ denote a function
that determines whether there are more than $a$ alibi events for a given pair of
users. Intuitively, having a single alibi is enough to decide that user $x$ and
$y$ are not the same person, but when there is inaccurate information,
disregarding candidate pairs with a single alibi event might lead to false negatives.
We have:
\begin{equation}\label{eq:noAlibi}
\mathcal{A}(x,y) \equiv |{i,e} \in \mathcal{I}_x \times
\mathcal{E}_y, s.t. A(i,e) | \leq a
\end{equation}
With these definitions at hand, we can define the spatio-temporal linkage
function $\mathcal{M}$ that determines whether users $x$ and $y$ satisfy
$k$-$l$ diversity as follows:
\begin{equation}
\label{eq:k-l-match}
\mathcal{M}(x,y) \equiv \neg \mathcal{A}(x,y) \,\wedge\, S\in \mathcal{S}(x,
y) \;s.t.\; (\mathcal{C}(S) \wedge \mathcal{D}(S))
\end{equation}
Finally, the linkage function $\mathcal{L}$ from the original problem
formulation from Section~\ref{sec:prelim} can be defined to contain only
matching pairs of users based on $\mathcal{M}$, such that there is no
ambiguity. Formally:
\begin{equation}
\label{eq:final}
\begin{split}
\mathcal{L}(\mathcal{E},\mathcal{I}) &= \{(x, y)\in \mathcal{E}\times\mathcal{I}:
\mathcal{M}(x,y) \,\wedge\,\\
&\nexists z\neq x \text{ s.t. } \mathcal{M}(z,y) \,\wedge\,
\nexists z\neq y \text{ s.t. } \mathcal{M}(x,z)\}
\end{split}
\end{equation}
\subsection{Example Scenario}
Consider three colleagues Alice, Bob, and Carl who are working in the same
office. Assume that they all use two \emph{LES}s: \emph{les1} and \emph{les2}.
Both services generate spatio-temporal records only when they are used. The
service provider would like to link the profiles of users common in both
services. However, Bob uses the services only when he is at the office. On the
other hand, Alice and Carl use the services frequently while at work, at home,
and during vacations. Let us also assume that Alice and Carl live on the same
block, but they take vacations at different locations.
When records of Alice from \emph{les1} are processed against records of Carl
from \emph{les2}, we will encounter \emph{co-occurrences} with some amount of
diversity, as they will have matching events from work and home locations. However,
we will encounter alibi events during vacation time. In this case,
\emph{alibi} checks will help us rule out the match.
When records of Alice from \emph{les1} are processed against records of Bob
from \emph{les2}, the number of \emph{co-occurrence}s will be high, as they
are working in the same office. Yet, \emph{diversity} will be low, as Bob does
not use the services outside of the office. This also means we will not
encounter any \emph{alibi} events with Alice. In this case, diversity will
help us rule out the match.
In contrast to these cases, when Alice's own usage records from \emph{les1}
and \emph{les2} are processed, the resulting \emph{co-occurrences} will
contain high diversity since Alice uses the services at work, home, and during
vacations, and will contain no \emph{alibi}s.
In this example scenario, high number of \emph{co-occurrences} helped us
distinguish between mere coincidences and potential candidate pairs. The
\emph{alibi} definition helped us to eliminate a false link between Alice and
Carl. Finally, \emph{diversity} helped us to eliminate a false link between
Alice and Bob, even in the absence of alibi events.
\section{ST-Link}
\label{sec:ST-Link}
\noindent In this section, we describe how the \emph{ST-Link} algorithm implements
$k$-$l$ diversity based spatio-temporal linkage in practice. At a high-level,
\emph{ST-Link} algorithm performs filtering to reduce the space of possible entity
matches, before it performs a more costly pairwise comparison of entities
according to the formalization given in Section~\ref{sec:formalization}. The
filtering phase is divided into two steps: \emph{temporal filtering} and
\emph{spatial filtering}. The final phase of pairwise comparisons is called
\emph{linkage}.
\subsection{Overview}\label{sub:overview}
\noindent Na\"{i}ve algorithms for linkage repeatedly compare pairwise
records, and thus take $\mathcal{O}(n^2)$~\cite{ref:tutorial} time, where $n$
is the number of records. Such algorithms do not scale to large datasets. To
address this issue, many linkage algorithms introduce some form of pruning,
typically based on blocking~\cite{ref:blocking1, ref:blocking2, ref:payger} or indexing~\cite{ref:index1, ref:index2}. Identifying
the candidate user pairs on which the full linkage algorithm is to be run can
significantly reduce the complexity of the end-to-end algorithm. Accordingly,
\emph{ST-Link} algorithm incorporates pruning strategies, which are integrated into
the spatial filtering and temporal filtering steps.
Figure~\ref{fig:data-pipeline} shows the pipelined processing of the
\emph{ST-Link} algorithm. Given two sources of data for location-enhanced services (DS1
and DS2 in the figure), spatial filtering step maps users to coarse-grained
geographical grid cells that we call \emph{dominating grid cells}. Such cells
contain most activities of the corresponding entities. Once this step is over, the remaining
steps are independently performed for each grid.
The temporal filtering step slides a window over the time ordered events to
build a set of candidate entity pairs. During this processing, it also prunes as
many entity pairs as possible based on alibi events. As we will detail later in
this section, a reverse window based scan is also performed to make sure that
all relevant alibis are taken into account.
Following the spatial and temporal filtering steps, the complete linkage is
performed over the set of candidate entity pairs. With a significantly reduced
entity pair set, the number of compared events decreases significantly as well.
Given two datasets $\mathcal{I}$ and $\mathcal{E}$, the linkage step
calculates $\mathcal{L}(\mathcal{E},\mathcal{I})$ as given in
Eq.~\ref{eq:final} without considering all possible entity pairs.
\subsection{Spatial Filtering}
\noindent By their nature, spatio-temporal data are distributed
geographically. Spatial filtering step takes advantage of this, by
partitioning the geographical region of the datasets into coarse-grained grid
cells using a modified version of quad trees~\cite{ref:quadTree}. Each entity
is assigned to one (an in rare cases to a few) of the grid cells, which
becomes that entity's dominating grid. The dominating grid of an entity is the
cell that contains the most events from the entity. Entities that do not share
their dominating grid cells are not considered for linkage. The intuition
behind this filtering step is that, if entity $x$ from dataset $\mathcal{E}$
and entity $y$ from dataset $\mathcal{I}$ have differing dominating grids,
then they cannot be the same entity.
\subsubsection{Coarse Partitioning} \noindent For quad-tree generation in the
\emph{ST-Link} algorithm, we continue splitting the space until the grid cells
size hits a given minimum. For our experiments, we make sure that the area of
the grid cells is at least 100 km squares. For users, the grid cells should be
big enough to cover a typical user's mobility range around his home and work
location. If the minimum grid cell size is too small, then the spatial
filtering can incorrectly eliminate potential matches, as the dominating grids
from different datasets may end up being different. A concrete example is a
user that checks in to coffee shops and restaurants around his work location,
but uses a location-based match-making application only when he is at home.
\begin{figure}[h!t]
\centering
\begin{subfigure}{0.475\linewidth}
\centering
\includegraphics[height=0.6\linewidth]{sea.png}
\caption{Grid cells.}
\label{fig:nopoints}
\end{subfigure}
\hspace{0.025\linewidth}
\begin{subfigure}{0.475\linewidth}
\centering
\includegraphics[height=0.6\linewidth]{istanbul.png}
\caption{Most popular venues.}
\label{fig:istanbul}
\end{subfigure}
\caption{Grids cells and top 1K venues}
\label{fig:spatial}
\end{figure}
We also do not split grid cells that do not contain any events. As a result,
not all grid cells are the same size. Figure~\ref{fig:spatial} shows the grid
cells for two selected areas in Turkey and the top 1K venues in those areas in
terms of check-in counts, based on Foursquare check-in data.
\subsubsection{Determining Dominating Grids}
\noindent The determination of the dominating grid for an entity is simply
done by counting the entity's events for different cells and picking the cell with the
highest count. A subtle issue here is about entities whose events end up being
close to the border areas of the grid cells. As a specific example, consider a
user who lives in one cell and works in another. In this case, it is quite
possible that a majority of the user's check-ins happen in one cell and the
majority of the calls in another cell. This will result in missing some of
the potential matches. To avoid this situation, we make two adjustments:%
\smallskip\\
\indent$\bullet$ If an event is close to the border, then it is counted towards the sums
for the neighboring cell(s)\footnote{An event can count towards at most 3
neighbors, in case it is at the corner of the grid.} as well. We use a strip
around the border of the cell to determine the notion of `close to the
border'. The width of the strip is taken as the 1/8th of the minimum cell's
edge width. This means that around $43\%$ of a grid overlaps with one or
more neighboring grids. This adjustment resembles the loose quad
trees~\cite{ref:looseTree}.
\smallskip\\
\indent$\bullet$ An entity can potentially have multiple dominating grid cells. We have
found this to be rare for users in practice.
\smallskip\\
Figure~\ref{fig:istanbul} shows the resulting grids over selected areas in Turkey,
and the most popular venues from our dataset. Red pins are
showing the venues and the blue ones are showing the ones that count towards
neighboring grids.
\subsubsection{Forming Partitioned Datasets}
\noindent Once the dominating grid cells of users are determined, we
create grid cell specific datasets. For a given grid cell $c$, we take only
the events of the entities who has $c$ as a dominating grid cell. These events
may or may not be in the grid cell $c$. Determination of the dominating grid cells
of entities requires a single scan over the time sorted events from entities. The
forming of the partitioned datasets requires a second scan.
\subsection{Temporal Filtering}
\noindent Temporal filtering aims at creating a small set of candidate user pairs
on which the full linkage algorithm can be executed. To create this set,
temporal filtering looks for user pairs that have co-occurring events, as
expressed by Eq.~\ref{eq:co-occur}. Importantly, temporal filtering also
detects alibi events, based on Eq.~\ref{eq:alibi}, and prevents user
pairs that have such alibi events from taking part in the candidate pair set.
Temporal filtering is based on two main ideas. First, a temporal window is
slided over the events from two different datasets to detect user pairs with
co-occurring events. Since co-occurring events must appear within a given time
duration, the window approach captures all co-occurring events. Second, as the
window slides, alibi events are tracked to prune the candidate user pair
set. However, since the number of alibis is potentially very large, alibis
are only tracked for the user pairs that are currently in the candidate set.
This means that some relevant alibis can be missed if the user pair was added
into the candidate set after an alibi event occurred. To process such alibis
properly, a reverse window scan is performed, during which no new candidate
pairs are added, but only alibis are processed. Algorithm~\ref{algo:slide} gives
the pseudo-code of temporal filtering.
\begin{algorithm}[!t]
\begin{small}
\KwData{$SR_\mathcal{I}$, $SR_\mathcal{E}$: Time sorted datasets of events}
\KwResult{$CS$: A set of candidate user pairs}
$CS \leftarrow \varnothing$ \tcp*{Candidates, $CS[u]$ is the list of pair users of $u$}
$AS \leftarrow \varnothing$ \tcp*{Alibis, $AS[u]$ is the list of alibi users of $u$}
$UI_x \leftarrow \varnothing, x \in \{ \mathcal{I}, \mathcal{E} \}$ \tcp*{User index over window}
\tcp{$UI_x[u]$: events from $x$ in window belonging to user $u$}
$LS_x \leftarrow \varnothing, x \in \{ \mathcal{I}, \mathcal{E} \}$ \tcp*{Spatial index over window}
\tcp{$LS_x.query(e.r)$: events from $x$ in window intersecting event $e$}
$W \leftarrow window(SR_\mathcal{I}, SR_\mathcal{E}, \alpha)$ \tcp*{Window over the datasets}
\tcp{Forward scan phase}
\While(\tcp*[f]{While more events after window}){$W.hasNext()$}{
\tcp{Get events inserted into and removed from the window}
$(N^+_\mathcal{I}, N^+_\mathcal{E}, N^-_\mathcal{I}, N^-_\mathcal{E}) \leftarrow W.next()$\\
\For(\tcp*[f]{In both directions}){$x \in \{\mathcal{I}, \mathcal{E}\}$}{
\For(\tcp*[f]{For each removed event}){$i \in N^-_x$}{
$LS_x.remove(i.r, i)$ \tcp*{Remove from spatial index}
$UI_x[i.u] \leftarrow UI_x[i.u] \setminus i$ \tcp*{Remove from user index}
}
}
\For(\tcp*[f]{In both directions}){$(x, \bar{x}) \in \{(\mathcal{I}, \mathcal{E}), (\mathcal{E}, \mathcal{I})\}$}{
\For(\tcp*[f]{For each inserted event}){$i \in N^+_x$}{
\tcp{Query spatially close elements}
\For{$e \in LS_{\bar{x}}.query(i.r)$}{
\If(\tcp*[f]{If users are not alibi}){$e.u \not\in AS[i.u]$}{
\If(\tcp*[f]{If events co-occur}){$C(i,e)$}{
\tcp{Add to the candidate set}
$CS[i.u] \leftarrow CS[i.u] \cup \{e.u\}$\\
$CS[e.u] \leftarrow CS[e.u] \cup \{i.u\}$
}
}
}
\For(\tcp*[f]{For each candidate user}){$u \in CS[i.u]$}{
\tcp{For each event of the user in the window}
\For{$e \in UI_{\bar{x}}[u]$}{
\If(\tcp*[f]{If $i$ and $e$ is an alibi}){$A(i,e)$}{
\tcp{Add to the alibi set}
$AS[i.u] \leftarrow AS[i.u] \cup \{u\}$\\
$AS[u] \leftarrow AS[u] \cup \{i.u\}$\\
\tcp{Remove from the candidate set}
$CS[i.u] \leftarrow CS[i.u] \setminus \{u\}$\\
$CS[u] \leftarrow CS[u] \setminus \{i.u\}$
}
}
}
$LS_x.insert(i.r, i)$ \tcp*{Add to spatial index}
$UI_x[i.u] \leftarrow UI_x[i.u] \cup \{i\}$ \tcp*{Add to user index}
}
}
}
\tcp{Reverse scan phase}
$W \leftarrow reverse\_window(SR_\mathcal{I}, SR_\mathcal{E}, \alpha)$ \tcp*{Reverse sliding window}
\While(\tcp*[f]{While more events after window}){$W.hasNext()$}{
$(N^+_\mathcal{I}, N^+_\mathcal{E}, N^-_\mathcal{I}, N^-_\mathcal{E}) \leftarrow W.next()$\\
\For(\tcp*[f]{In both directions}){$x \in \{\mathcal{I}, \mathcal{E}\}$}{
\For(\tcp*[f]{For each removed event}){$i \in N^-_x$}{
$UI_x[i.u] \leftarrow UI_x[i.u] \setminus i$ \tcp*{Remove from user index}
}
}
\For(\tcp*[f]{In both directions}){$(x, \bar{x}) \in \{(\mathcal{I}, \mathcal{E}), (\mathcal{E}, \mathcal{I})\}$}{
\For(\tcp*[f]{For each inserted event}){$i \in N^+_x$}{
\For(\tcp*[f]{For each candidate user}){$u \in CS[i.u]$}{
\If(\tcp*[f]{If $i$ and $e$ is an alibi}){$A(i,e)$}{
\tcp{Add to the alibi set}
$AS[i.u] \leftarrow AS[i.u] \cup \{u\}$\\
$AS[u] \leftarrow AS[u] \cup \{i.u\}$\\
\tcp{Remove from the candidate set}
$CS[i.u] \leftarrow CS[i.u] \setminus \{u\}$\\
$CS[u] \leftarrow CS[u] \setminus \{i.u\}$
}
}
}
$UI_x[i.u] \leftarrow UI_x[i.u] \cup \{i\}$ \tcp*{Add to user index}
}
}
\textbf{return} $CS$ \tcp*{Return the candidate set}
\caption{Candidate Set Calculation}
\label{algo:slide}
\end{small}
\end{algorithm}
\begin{figure*}[t]
\centering
\begin{subfigure}{0.54\linewidth}
\includegraphics[width=\linewidth]{sliding_window.png}
\caption{Sliding window algorithm.}\label{fig:sliding_window}
\end{subfigure}
\hspace{0.01\linewidth}
\begin{subfigure}{0.42\linewidth}
\includegraphics[width=\linewidth]{algo.png}
\caption{Calculation of the candidate set.}\label{fig:algo}
\end{subfigure}
\caption{Temporal Filtering}\label{fig:temporal}
\end{figure*}
\subsubsection{Data Structures}
\noindent
A window of size $\alpha$ (see Eq.~\ref{ref:co-time}) is slided jointly over both
time sorted datasets.
Figure~\ref{fig:sliding_window} depicts this visually. Each time the window
slides, some events from both datasets may enter and exit the window. We utilize
two types of data structures to index the events that are currently in the
window.
The first type of index we keep is called the \emph{user index}, denoted by
$UI_x$, where $x\in\{\mathcal{I}, \mathcal{E}\}$. In other words, we keep
separate user indexes for the two datasets. $UI_x$ is a hash map indexed by
the user. $UI_x[u]$ keeps all the events (from dataset $x$) of user $u$ in
the window. As we will see, this index is useful for quickly checking alibis.
The second type of index we keep is called the \emph{spatial index}, denoted
by $LS_x$, where $x\in\{\mathcal{I}, \mathcal{E}\}$. Again, we keep separate
indexes for the two datasets. $LS_x$ could be any spatial data structure like R-trees.
$LS_x.query(r)$ gives all events whose region intersect with
region $r$. As we will see, this index is useful for quickly locating
co-occurring events.
In addition to these indexes, we maintain a global candidate set $CS$ and a
global alibi set $AS$. For a user $u$ (from either dataset, assuming user ids
are unique), $CS[u]$ keeps the current set of candidate pair users for $u$;
and $AS[u]$ keeps the current known alibis users for $u$. It is important to
note that $AS$ is not designed to be exhaustive. For a user $u$, $AS[u]$ only
keeps alibi users that have co-occurring events with $u$ in the dataset.
\subsubsection{Processing Window Events}
\noindent
The algorithm operates by reacting to events being inserted and removed from
the window as the window slides over the dataset. As a result, an
outermost while loop that advances the window until the entire dataset is
processed. At each iteration, we get a list of events inserted
($N^+_\mathcal{I}$ and $N^+_\mathcal{E}$) and removed ($N^-_\mathcal{I}$ and
$N^-_\mathcal{E}$) from the window. We first process the removed events, which
consists of removing them from the spatial and user indexes. We then process
the inserted events. We first process $N^+_\mathcal{I}$ against
$UI_{\mathcal{E}}$ and $LS_{\mathcal{E}}$, then insert the events in
$N^+_\mathcal{I}$ into $UI_{\mathcal{I}}$ and $LS_{\mathcal{I}}$, then process
$N^+_{\mathcal{E}}$ against $UI_{\mathcal{I}}$ and $LS_{\mathcal{I}}$, and
finally insert the events in $N^+_{\mathcal{E}}$ into $UI_{\mathcal{E}}$ and
$LS_{\mathcal{E}}$. This ensures that all the events are compared, and no
repeated comparisons are made. Figure~\ref{fig:algo} depicts the order of
events visually.
To compare a new event $i$ from dataset $x$ against the events from dataset
$\bar{x}$ that are already indexed in the window (where $\{x, \bar{x}\} =
\{\mathcal{E}, \mathcal{I}\}$), we use the indexes $UI_{\bar{x}}$ and $LS_{\bar{x}}$.
First, we find events that co-occur with $i$ by considering events $e$ in $LS_{\bar{x}}.query(i.r)$.
These are events whose regions intersect with that of
$i$. If the user of such an event $e$ is not already a known alibi of the user
of $i$ (not in $AS[i.u]$) and if the co-occurrence condition $C(i, e)$ is
satisfied, then the user $e.u$ and user $i.u$ are added as candidate pairs of
each other. Second, and after all the co-occurrences are processed, we
consider all candidate users of the event $i$'s user, that is $CS[i.u]$, for
alibi processing. For each user $u$ in this set, we check if any of its events
result in an alibi. To do this, we iterate over user $u$'s events with the
help of the index $UI_{\bar{x}}$. In particular, for each event $e$ in
$UI_{\bar{x}}[u]$, we check if $i$ and $e$ are alibis, using the condition
$A(i, e)$. If they are alibis, then we remove $u$ and $i$'s user ($i.u$) from
each other's candidate sets, and add them to their alibi sets.
This completes the description of the forward scan of the window. An important
point to note is that, during the forward window scan, we only check alibis
for user pairs that are in the set of candidate pairs. It is possible that
there exists an alibi event pair for users $x$ and $y$, that appears before
the first co-occurring event pair for these users. In such a case, during the
processing of the alibi events we won't have this pair of users in our
candidate set and thus their alibi will be missed. To fix this problem, we
perform a reverse scan. During the reverse scan, we only process alibis, as no
new candidate pairs can appear. Furthermore, we need to process alibi events
for a user pair only if the events happened before the time this pair was
added into the candidate set. For brevity, we do not show this detail in
Algorithm~\ref{algo:slide}. At the end of the reverse scan, the set $CS$
contains our final candidate user pairs, which are sent to the linkage step.
Temporal filtering is highly effective in reducing the number of pairs for
which complete linkage procedure is executed. The experimental results show the
effectiveness of this filtering.
When there is inaccurate information in the datasets,
disregarding candidate pairs due to only a single alibi event might lead to
false negatives. However, the algorithm is easily modifiable to use a
threshold for alibi values. In this modified version, we update the structure
of the alibi set $AS$ to keep the number of alibi events of a pair as well.
Now $AS[u]$ keeps the current known alibi users of user $u$ with alibi event
counts for each. Just like in the original algorithm, when two events $i$ and
$e$ are compared we first check if the number of alibi events of users $i.u$
and $e.u$ exceeds the threshold. To avoid double counting, we reset
the counters before the reverse scan. Since all alibi events of current
candidate pairs will be counted in reverse scan, candidates whose count of
alibi events exceed threshold will not be included in the resulting candidate
set $CS$.
So far we have operated on time sorted event data and our algorithms used only
sequential I/O. However, during the linkage step, when we finally decide
whether a candidate user pair can be linked, we will need the time sorted
events of the users at hand. For that purpose, during the forward scan, we
also create a disk-based index sorted by the user id and event time. This
index enables us to quickly iterate over the events of a given user in
timestamp order, which is an operation used by the linkage step. For this
purpose, we use LevelDB~\cite{ref:leveldb} as an index, which is a
log-structured merge-tree supporting fast insertions.
While writing the event to the disk-based index, we also include information
about the number of unique users the event has matched throughout its stay in
the forward scan window. This information is used as part of the weight
calculation (recall Eq.~\ref{eq:weight}) in the linkage step.
\subsubsection{Handling Time Period in Events}\label{sec:timePeriod}
The temporal filtering step scans time-ordered events by sliding a window of
size $\alpha$ over them. This operation assumes that the time information is a
point in time. Yet, there could be scenarios where the time information is a
period (e.g., a start time and a duration). However, frequently, these records
contain only start location of the event. For example, although Call Detail
Records (CDR) have the start time of the call and the duration, they usually
contain only the originating cell tower information. Considering mobility of
the users, assuming a fixed location during this period would lead to location
ambiguity.
If we have events with time periods and accurate location information is
present during this period, we can adapt our approach to handle this. In
particular, we need to avoid false negative candidate pairs when the event
contains a time period. Since events are processed via windowing, making sure
that the event with the time period information stays in the window as long as
its time period is valid would guarantee that all co-occurrences will be
processed. This requires creating multiple events out of the original event,
with time information converted into a point in time and the correct location
information attached to it. The number of such events is bounded by the time
duration divided by the window size, $\alpha$.
\subsection{Linkage}
\label{sec:linkage}
\noindent The last step of the \emph{ST-Link} algorithm is the linkage of the
entities that are determined as possible pairs as a result of spatial and
temporal filtering. This linkage is a realization of the $k$-$l$ diversity
based linkage model introduced in Section~\ref{sec:formalization}. Given two
entities from two datasets, the linkage step uses the events of them to
determine whether they can be linked according to Eq.~\ref{eq:k-l-match}.
Thanks to efficient filtering steps applied on the data beforehand, the number
of entity pairs for which this linkage computation is to be performed is
significantly reduced.
For each entity pair, their events are retrieved from the disk-based index
created as part of the forward scan during the temporal filtering. These
events are compared for detecting co-occurring events. Co-occurring events are
used to compute the $k$ value, via simple accumulation of the co-occurrence
weights. They are also used to accumulate weights for the places where
co-occurring events occur. This helps us compute the $l$ value, that is
diversity. After all events of a pair of entities are compared, we check if they
satisfy the $k$-$l$ diversity requirement. Note that, it is not possible to
see an alibi pair event at this step, as they are eliminated by the temporal
filtering step.
There are a number of challenges in applying the $k$-$l$-diversity based
linkage. The first is to minimize the number of queries made to the disk-based
index to decrease the I/O cost. Events from the same entity are stored in
a timestamped order within the index, which makes this access more efficient.
Also, if one of the datasets is more sparse than the other, then the linkage
can be performed by iterating over the entities of the dense datasets first,
making sure their events are loaded only once. This is akin to the classical
join ordering heuristic in databases.
Another challenge is the definition of the place ids to keep track of
diversity. A place id might be a venue id for a Foursquare dataset, store id
for credit card payment records, cell tower id for Call Detail Records, or a
geographic location represented as latitude and longitude. An important
difference is the area of coverage for these places. Consider two datasets of
Foursquare check-ins and Call Detail Records, and places based on venues. If a
user visits several nearby coffee shops and makes check-ins and calls, these
will be considered as diverse even though they are not geographically diverse.
The use of cell tower coverage areas is a more practical choice for
determining places.
The last challenge is about matching events. Recall from
Figure~\ref{fig:match} that events of two entities can be matched in multiple
different ways, resulting in different weights for the co-occurrences. Ideally,
we want to maximize the overall total weight of the matching, however this
would be quite costly to compute, as the problem is a variation of the
\emph{bipartite graph assignment problem}. As a result, we use a greedy
heuristic. We process events in a timestamped order and match them to the
co-occurring event from the other entity that provides the highest weight. Once
a match is made, event pairs are removed from the dataset so that they are not
re-used.
Different $k$-$l$ value pairs may perform significantly different in terms of
precision and recall, depending on the frequencies of the events in the
datasets. An ad-hoc approach is to decide the $k$ and $l$ values based on
observation of results from multiple experimental runs. A more robust
technique we used is to detect the best \emph{trade-off} point (a.k.a
\emph{elbow point}) on a curve. Given the \emph{co-occurrence} and
\emph{diversity} distributions, we independently detect the elbow point of
each, and set the $k$ and $l$ values accordingly. Although there is no
unambiguous solution for detecting an elbow point, the maximum absolute second
derivative is an approximation. Let $A$ be an array of co-occurrence (or
diversity) values with size $n$. Second derivative, $SD$, of point at index
$i$ can be approximated with a central difference as follows:
\begin{equation}
SD[i] = A[i+1] + A[i-1] - 2* A[i]
\end{equation}
The value at index $A[i]$, such that $i$ has the maximum
absolute $SD[i]$ value, is selected as the \emph{elbow point} and $k$ (or $l$)
value is set accordingly.
\section{Experimental Evaluation}\label{sec:experiments}
\begin{figure*}[t]
\centering
\begin{subfigure}{.30\linewidth}
\includegraphics[width=\linewidth]{WindowSizevsRuntime.png}
\caption{Running time vs. dataset size.\hfill\eject$\,$}
\label{fig:runtime}
\end{subfigure}
\hspace{0.01\linewidth}
\begin{subfigure}{0.30\linewidth}
\includegraphics[width=\linewidth]{WindowSizevsOperations.png}
\caption{Number of comparisons vs. dataset size.}
\label{fig:comparisons}
\end{subfigure}
\hspace{0.01\linewidth}
\begin{subfigure}{.30\linewidth}
\includegraphics[width=\linewidth]{WindowSizevsPairsCount.png}
\caption{Number of candidate user pairs vs. dataset size.}
\label{fig:pairs}
\end{subfigure}
\caption{Performance Results}
\label{fig:performance}
\end{figure*}
\noindent In this section, we present an evaluation of the
proposed $k$-$l$ diversity based linkage method and the
\emph{ST-Link} algorithm. We implemented the \emph{ST-Link} algorithm
using Java 1.7. All experiments were executed on a Linux server with $2$
Intel Xeon E5520 $2.27$GHz CPUs and $64$GB of RAM.
We present two sets of experiments. In the first set of
experiments, we measure the performance and the scalability of the
\emph{ST-Link} algorithm. By increasing the size of the input data, we
test the change in the running time, number of event comparisons, and the
number of candidate user pairs, for different window sizes. In the second set of experiments, we analyze the quality of the $k$-$l$ diversity based linkage. To measure quality, we use two metrics. The first is the
precision, which measures the fraction of correctly linked pairs in the list
of user pairs produced by \emph{ST-Link}. The second is the number of true
positives, which is the number of user pairs correctly linked by the
\emph{ST-Link} algorithm.
\subsection{Datasets Used}\noindent
For the performance, scalability, and accuracy evaluations we used three datasets. The first is a Foursquare dataset
of check-ins. The second is anonymized call detail records in a telecommunication
provider. For privacy concerns, we did not perform any linkage across these two datasets.
As a result, we were not able to compute accuracy results when using these two datasets.
However, they are used for the evaluation of running time performance.
To evaluate accuracy, linkage is performed
between a third dataset belonging to a hypothetical LES and the call dataset. This dataset was synthetically
derived to protect privacy, from the call dataset by ($i$) picking a predefined fraction $f$ of the
callers at random as active users of the second LES, ($ii$) generating usage
records for the selected users by assuming that they generate such a record
with probability $p$, within a $15$ minute time window of a call, inside a
location within the same cell tower of the call. We change the parameters $p$
and $f$ to experiment with different scenarios. Lower values for $p$ result in
a sparser usage record dataset for the second LES. We call the parameter $p$,
the \emph{check-in probability}. As not all users have the same check-in
probability in practice, we pick the value of the check-in probability for a given user
from a Gaussian distribution with mean $p$. We call the parameter $f$, the
\emph{usage ratio}.
\begin{table}[h!t]
\centering
\begin{tabular}{|r||c|c|}\hline
\emph{Datasets $\Rightarrow$}
& Foursquare & Call \\\hline\hline
\# of activities & 1,903,674 & 1,890,107,057 \\\hline
\# of venues/cell towers & 300,685 & 109,780 \\\hline
\# of users & 284,856 & 3,357,069 \\\hline
\end{tabular}
\caption{Dataset statistics}\label{tbl:data}
\end{table}
The Foursquare dataset consists of check-ins that were shared publicly on
Twitter, collected via the Twitter streaming
API\footnote{www.dev.twitter.com} and the Foursquare
API\footnote{www.developer.foursquare.com}. This dataset spans $40$ days and
only contains check-ins from Country X. Each row contains the acting user's
Foursquare id, venue id, geographical location (lat/lon) of the
venue, and the time of the check-in. The call dataset spans
the same $40$ days in Country X. Each row contains an anonymized id, time of the call, and geographical location
(lat/lon) of the handling cell tower. The anonymized id is the same across all usage
of the same user. Table~\ref{tbl:data} shows the statistics about both the Foursquare and the
call datasets. For the runtime performance and filtering effectiveness experiments, we used
the two real datasets. However, since it is not possible to
verify the accuracy of the results using these two datasets, we used the synthetic dataset which is derived from the
anonymous call data for the evaluation of \emph{ST-Link}'s accuracy.
\subsection{Running Time Performance}
\begin{figure*}[t]
\centering
\begin{minipage}{.30\linewidth}
\includegraphics[width=\linewidth]{filteringEfficiency.png}
\caption{Reduction in the number of possible pairs.}
\label{fig:efficiency}
\end{minipage}
\hspace{0.01\linewidth}
\begin{minipage}{0.30\linewidth}
\includegraphics[width=\linewidth]{cp_precision.png}
\caption{Precision as a function of check-in probability.}
\label{fig:meanPrecision}
\end{minipage}
\hspace{0.01\linewidth}
\begin{minipage}{0.30\linewidth}
\includegraphics[width=\linewidth]{cp_recall.png}
\caption{Number of true positives as a function of check-in probability.}
\label{fig:meanTP}
\end{minipage}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{minipage}{0.30\linewidth}
\includegraphics[width=\linewidth]{au_precision.png}
\caption{Precision as a function of usage ratio.}
\label{fig:activePrecision}
\end{minipage}
\hspace{0.01\linewidth}
\begin{minipage}{0.30\linewidth}
\includegraphics[width=\linewidth]{au_recall.png}
\caption{Number of true positives as a function of usage ratio.}
\label{fig:activeTP}
\end{minipage}
\hspace{0.01\linewidth}
\begin{minipage}{0.30\linewidth}
\centering
\begin{tabular}{|r||c|c|}\hline
& Precision & Recall \\ \hline
2-2 & 0.19 & 0.31 \\ \hline
3-2 & 0.36 & 0.32 \\ \hline
\rowcolor{Gray} 3-3 & 0.89 & 0.61 \\ \hline
4-3 & 0.93 & 0.47 \\ \hline
4-4 & 0.99 & 0.58 \\ \hline
5-4 & 0.99 & 0.47 \\ \hline
5-5 & 0.99 & 0.50 \\ \hline
\end{tabular}
\caption{Precision and recall using unweighted linkage}
\label{tbl:weightTest}
\end{minipage}
\end{figure*}
\begin{figure*}[t]
\centering
\begin{minipage}{0.30\linewidth}
\includegraphics[width=\linewidth]{ws_precision.png}
\caption{Precision as a function of window size.}
\label{fig:wsPrecision}
\end{minipage}
\hspace{0.01\linewidth}
\begin{minipage}{0.30\linewidth}
\includegraphics[width=\linewidth]{ws_recall.png}
\caption{Number of true positives as a function of window size.}
\label{fig:wsTP}
\end{minipage}
\hspace{0.01\linewidth}
\begin{minipage}{0.30\linewidth}
\includegraphics[width=\linewidth]{distribution.png}
\caption{$k$-$l$ values distribution\hfill\eject$\;$}
\label{fig:distribution}
\end{minipage}
\end{figure*}
\noindent We observe the running time, the number of candidate pairs, and the number of
event comparisons as a function of the dataset size. The dataset size is
increased by increasing the number of days of data included in the linkage
analysis. Furthermore, for these experiments we also change the window size.
Recall that the window size is used during the temporal filtering step to
locate co-occurring events.
Figure~\ref{fig:performance} presents our running time related results. In all
the figures, the $x$-axis represents the dataset size in days and the $y$-axis
represents a performance metric. Different series represent varying window
sizes.
One of the main challenges is the scalability of the linkage solution.
Processing many days of data should complete in reasonable amount of time for
the resulting analysis to be valuable. Figure~\ref{fig:runtime} plots the
running time as a function of the dataset size. We make two observations from
the figure. First, the running time of \emph{ST-Link} is
linear in the dataset size. For 5 days of data, the running time is around $1$
hour and for $40$ days of data it is around $7$ hours, all for $30$ minute
windows. Second, the running time increases with increasing window size, yet
the running time is linear in the dataset size for all window sizes.
Figure~\ref{fig:comparisons} plots the number of event-to-event comparisons as
a function of the dataset size. In our experimental evaluation, every time we
compare two location based events for either co-location or alibi check, the
number of event comparisons is increased by one. We observe that up to 15 days
of data, the number of comparisons grows at an increasing rate. Yet, after 15
days the rate starts to go down and eventually the growth of the number of
operations happens at a relatively low fixed rate. This can be explained by
the alibi checks performed by \emph{ST-Link}. Recall that when a user pair is
marked as an alibi, their records are not compared with each other anymore.
Also, if two users are marked as a candidate pair, their future records are
only compared to see if they are an alibi or not. Considering this, we can say
that within 15 days most of the candidate pairs and alibi pairs are
identified. As an important difference from the running time experiment, the
gaps between the series corresponding to the three window sizes are
considerably larger. This is because larger windows require more event to
event comparisons. Since event comparisons are not necessarily the only cost
of the algorithm (there is I/O, window processing, window index maintenance,
etc.), the running time experiment has narrower gaps between the running times
for different window sizes. The impact of these extra costs can be seen in
Figure~\ref{fig:runtime} as well; although the number of comparisons stabilize
after 15 days, the linear increase in the runtime continues.
Figure~\ref{fig:pairs} plots the number of candidate user pairs as a function
of the dataset size. Just like for the number of comparisons experiment, up to
15 days, the number of candidate pairs grows with an increasing rate and after
15 days the rate starts to decrease and eventually stabilizes at a low value.
For the case of candidate pairs, the eventual rate of increase is very low,
suggesting that observing additional data brings diminishing returns in terms
of being able to find new candidate pairs. However, this does not imply that
we are unable to perform additional linkages, because the number of linked
pairs within the candidate set can still grow (we will observe such growth
in the quality experiments).
Figure~\ref{fig:efficiency} shows the number of candidate user pairs after
each filtering step. It illustrates the effectiveness of the spatial and
temporal filtering steps of \emph{ST-Link}. If no filtering was applied on the
data, every user pair from the two datasets would have constituted a candidate
user pair. By applying only spatial filtering, the number of candidate user
pairs decreases by $43$ times compared to the no filtering case. It is
possible to say that spatial filtering is an effective step. Intuitively, if
data was spread over a wider geographical area, this step would be have been
even more effective (our datasets are limited to the geographic area of
Country X). After applying temporal filtering, the candidate user set
decreases by an additional $1,836$ times after spatial filtering.
Cumulatively, the number of candidate user pairs without any filtering is
$78,948$ times of the number of pruned candidate user pairs, which gets close
to $5$ orders of magnitude reduction in the number of pairs.
\subsection{Quality of Linkage}
\noindent We observe the precision and the number of true positives as a
function of the usage ratio, check-in probability, and window size. We also
observe the precision and recall values for a variation of the \emph{ST-Link}
algorithm that does not use weights, thus trades off precision for better
recall.
\subsubsection{Impact of Check-in Probability}
\noindent Figures~\ref{fig:meanPrecision}~and~\ref{fig:meanTP} plot the
precision and number of true positives, respectively, for the results produced
by the \emph{ST-Link} algorithm as a function of the mean check-in probability.
Different series in the figure represent different $k$-$l$ settings. We set
$k\geq l$, as the co-occurrence counts has to be greater than the diversity
counts. For these experiments the usage ratio is set to $50\%$, which means
only half of the call users are performing check-ins.
Figure~\ref{fig:meanPrecision} shows that precision is very close to $1$ for
all $k$-$l$ settings but for $1$-$1$. We see that using $1$-$1$ diversity
results in very poor precision for low values of the check-in probability. As
the check-in probability increases, then the precision of $1$-$1$ diversity
increases as well, but never reaches $1$. The increase is understandable, as
more events on the check-in side will help rule out incorrect candidate pairs
via alibis. Surprisingly, the precision for higher $k$-$l$ values are all
close to $1$. This is due the impact of alibi detection, and strong weight
constraint. As we will see shortly, not using weights trades off precision
for better recall. Even if two users have events that are only $2$-$2$
diverse, they can be correctly linked if they have no alibis.
Figure~\ref{fig:meanTP} shows that the number of true positives in the linkage
increases with the increasing check-in probability. This is expected, as more
events help in increasing the co-occurrence and diversity counts. We also
observe that higher $k$-$l$ values result in reduced number of linkages. Given
that $2$-$2$ diversity has very good precision, and has the second highest
true positive count (after $1$-$1$ diversity, which has unacceptable
precision), it can be considered a good setting for getting the best out of
the linkage. We see that for a check-in probability as low as $0.01$, it can
match many hundreds of users, and for probability $0.1$, it can match up to
$10$ thousand users. As we will see shortly, these numbers can be further
increased by trading off some accuracy.
\subsubsection{Impact of Usage Ratio}
\noindent Figures~\ref{fig:activePrecision}~and~\ref{fig:activeTP} plot the precision
and the number of true positives, respectively, for the results produced by the
\emph{ST-Link} algorithm as a function of the usage ratio. Different series in
the figure represent different $k$-$l$ settings, as before. For these
experiments check-in probability is taken as $0.01$.
Figure~\ref{fig:activeTP} shows that the precision of all $k$-$l$ settings is
close to $1$ throughout the entire range of the usage ratio, except for
$1$-$1$. The $2$-$2$ setting has precision values that are slightly lower than
$1$, but not lower than $0.95$. Figure~\ref{fig:meanTP} shows the true
positive counts for the same settings. As we can see clearly from the figure,
increased usage ratio results in increased number of successful linkages.
Again, this could be attributed to increasing weights for co-occurrence and
diversity, as well as increased effectiveness of alibi detection.
Interestingly, even when only $1$ percent of the call users are synthetically set to making
check-ins, and when the check-in probability around a call is set as low as $1$ in
$100$, one can still match some users (around $10$). This could also be looked
at from a privacy standpoint. In other words, being able to perform
spatio-temporal linkage across two datasets successfully even for only $10$
users may be considered a privacy breach. We plan to investigate the privacy
protection mechanisms against this kind of linkage in our future work.
\subsubsection{Unweighted Linkage}
\noindent The results so far have demonstrated high precision, but the number of
users one could match is relatively low compared to the number of total users.
In order to show the trade-off between precision and accuracy, we have also
performed experiments where the linkage model is slightly modified to use
weights that are equal to $1$. That is, we count each event co-occurrence
between two users as $1$, without considering other possible co-occurrences
these events may have with events of other users. In other words, the weight
function from Eq.~\ref{eq:weight} is taken as $1$. As it was discussed before,
there are two different approaches for deciding the values of the $k$-$l$
parameters. The first one is deciding after multiple experimental runs, and
the second one is by detecting the \emph{elbow point} of distributions of
co-occurrence and diversity values.
In this experiment we applied both to show the effectiveness of the
\emph{elbow point} detection technique as well. According to maximum absolute
second derivative test results, the values of $k$-$l$ parameters based on
elbow detection are $3$-$3$.
Table~\ref{tbl:weightTest} shows the precision and recall results for the
unweighted linkage. The recall values here represent the fraction of users
from the check-in dataset that were successfully linked. It is important to note
that we only considered users that have enough number of events. A user is said
to have enough number of events only when she has at least $l$ diverse events,
for each $k$-$l$ setting.
The table shows an interesting result: With unweighted linkage we see a clear
tradeoff, where with increasing $k$-$l$ values the precision improves, but the
recall drops. With the $3$-$3$ setting, we get a precision of $0.89$ and can
link $61\%$ of the users that have enough number of events. Considering all
users from the check-in dataset this value is $23~\%$. Recall that $3$-$3$
setting was identified using the \emph{elbow point}s. Increasing the diversity
setting to $5$-$5$, one gets almost perfect accuracy ($0.99$), but the recall
drops to $50\%$ of the users. When absolute accuracy is not required, such as
for machine learning to extract overall patterns, the unweighted linkage model could be more
effective in practice.
\subsubsection{Impact of Alibi}
\begin{figure}
\begin{minipage}{0.5\linewidth}
\begin{tabular}{|r||c|c|c|c|}\hline
& Runtime (m) & Precision & Recall & Cand. Count \\ \hline
1 & 30 & 0.78 & 0.68 & 1,998,491 \\ \hline
2 & 34 & 0.75 & 0.74 & 2,651,746 \\ \hline
4 & 39 & 0.71 & 0.82 & 3,511,090 \\ \hline
8 & 45 & 0.68 & 0.88 & 4,446,937 \\ \hline
16 & 58 & 0.65 & 0.91 & 5,311,043 \\ \hline
$\infty$ & 122 & 0.62 & 0.99 & 6,765,345 \\ \hline
\end{tabular}
\end{minipage}
\caption{Alibi threshold experiment results.}
\label{tbl:alibiThreshold}
\end{figure}
Alibis are used to improve both the running time performance and the accuracy.
For these experiments the check-in probability is taken as $0.5$ and the
usage ratio is taken as $50\%$. Only one grid is considered, which contains
15,268 users and 1,956,734 events in total. As it was discussed before, a threshold value
on the number of alibi events can be used before disregarding a candidate pair. In this
experiment we evaluate the impact of alibi in terms of performance and accuracy as a function of the alibi threshold.
\medskip\\
\noindent \textbf{Performance.}
Table~\ref{tbl:alibiThreshold} shows the running time, precision, recall, and
the number of candidate pairs for the alibi threshold experiment. When the
threshold is set to $\infty$, effectively disabling alibi detection, we
observed that the algorithm took 122 minutes to complete. At the end of the
temporal filtering step, there were 6,765,345 possible pairs. On the other
hand, when alibi is used and the threshold is set to $1$, the running time
decreased down to 30 minutes and the number of possible pairs were 1,998,491.
Almost $70~\%$ of the possible pairs were pruned with the help of alibi
detection and further processing is avoided. When larger threshold values are
used, we observe slight increase in the running time. For the threshold values
of $2$ and $16$, the processing time is $34$ and $58$ minutes, respectively.
\medskip\\
\noindent\textbf{Accuracy.} Precision of the $k$-$l$ diversity based linkage can
be increased by setting sufficiently large $k$ and $l$ values. Larger $k$ and
$l$ values decrease the probability of different users satisfying the linkage
requirements. However when at least one of the datasets is sparse, setting
larger $k$ and $l$ values will result in low recall, as many true positive
pairs will be missed. In such datasets, alibi definition prevents many false
positive pairs that satisfy the co-occurrence and diversity requirements. Our
experiments showed that when alibi is not used (threshold value $\infty$), $99
\%$ recall can be reached, yet with $62\%$ precision. In contrast, setting
alibi threshold to $1$, increases the precision to $78\%$, with recall
decreased to $68\%$. The reason behind this decrease has to do with the lack
of precise location information in our datasets. For example, when two
temporally close events of a user are from two neighboring cell towers, their
locations end up being the centers of the cell towers, as the location
information is not sufficiently fine grained. This results in
incorrectly identifying a pair of events as alibis, as the distance between
the event locations is relatively high when considering their close
timestamps. This is when the alibi threshold becomes crucial. We observe that
the recall increases to $0.74\%$ when alibi threshold set to $2$. Increasing
threshold further increases the recall values with a cost of sacrificed
precision. For alibi threshold $4$ recall is $0.82\%$ and precision is $71\%$.
\subsubsection{Window Size}
\noindent Figures~\ref{fig:wsPrecision}~and~\ref{fig:wsTP} plot the precision and
the number of true positives, respectively, as a function of the window size.
Different series in the figure represent different $k$-$l$ settings, as
before. For these experiments the check-in probability is taken as $0.01$ and the
usage ratio is taken as $50\%$. Window sizes start from $15$ minutes and
increases up to $75$ minutes in increments of $15$ minutes.
Figure~\ref{fig:wsPrecision} shows that the precision stays at $1$ is not
effected by the window size except for lines corresponding to lower $k$-$l$
values. In particular $2$-$2$ and $3$-$2$ are impacted negatively from larger
window sizes. $1$-$1$ is not shown in this experiment, as it already has a
very low precision. Note that the window size does not impact only the size
the temporal window we slide over the events, but also the definition of
co-occurrence (recall the $\alpha$ parameter from Eq.~\ref{ref:co-time}).
Increasing the window size makes it possible to match potentially
unrelated events from different real-world users and the results reflect that.
However, due to the alibi processing, the negative impact of increasing window
size on the precision is milder that it would otherwise be.
Figure~\ref{fig:wsTP} shows that the number of true positives drops with
increasing window sizes. Again this can be attributed to the increasing number
of unrelated event matches due to the larger window. Recall that if the same
user is matched to more than one user from the other dataset, we remove such
users from the linkage results. The increased window size results in ambiguity
in the results. Assuming users $x$ and $y$ are linked for a given window size,
increasing the window size does not change the fact that $x$ and $y$ are
matched, but it may result in additional matches, such as between user $x$ and
some other user $z$, and thus eliminating the correct linkage between $x$ and
$y$ from the results.
\subsubsection{$k$-$l$ value Distribution}
\noindent Figure~\ref{fig:distribution} shows the $k$-$l$ value distribution of
user pairs after spatial and temporal linkage. The usage ratio is taken as
$50\%$ and the check-in probability is $0.01$ for this experiment. For a given
pair in the results, we find the highest $k$-$l$ diversity values it supports
and maintain these counts. In the figure, the areas of the circles are
proportional to number of pairs with the given $k$-$l$ diversity. Since the
number of pairs for $k$-$l$ lower than $2$-$2$ is too high (and precision very
low as we have seen earlier), we do not present them in the results. As
expected, the number of linked pairs is decreasing as the $k$-$l$ values are
increasing. It is interesting to note that for extreme values such as $11$-$7$
diversity, it is still possible to find user pairs. We also observe that
increasing diversity has a higher filtering power than increasing occurrence.
\subsection{Integration with SERF}
In addition to evaluating our approach under different settings, we also
attempted to integrate our linkage model with the Stanford Entity Resolution
Framework (SERF). SERF implements the R-Swoosh~\cite{ref:swoosh} algorithm.
For this integration, users are arranged as entities and their events are
considered as attributes. Given two entities, if they have enough number of
co-occurring attributes satisfying the $k$-$l$ diversity model, they are
marked as a match.
Starting with pairwise comparison of entities, R-Swoosh algorithm gradually
decreases the number of entities by merging the matching records, and deleting
the dominated ones. While this is an effective method to decrease the number
of comparisons on match heavy datasets, for datasets that contain few matching
entities, the run-time is still $\mathcal{O}(N^2)$. Merging of two
records is valid only when there is merge associativity between records. Given
three records, $r_1$, $r_2$, and $r_3$, comparisons of $\langle r_1, \langle
r_2, r_3\rangle\rangle$ and $\langle\langle r_1,r_2\rangle, r_3\rangle$ may
result in different linkage decisions~\cite{ref:koosh}. To alleviate this
problem, the SERF framework also implements the Koosh
algorithm~\cite{ref:koosh}. Different than the R-Swoosh algorithm, when the
Koosh algorithm finds a matching pair of entities, it does not merge them
immediately, unless confidence is above a threshold. However, defining the
confidence to use our spatio-temporal linkage model
in SERF is not straightforward and requires further research, which we
leave as future work.
Applying Koosh algorithm without using merges is almost brute force and using
a small subset our dataset (15,268 users, 1,956,734 events, in total),
SERF takes more than $50$ hours of processing time in the same setting. In comparison,
our algorithm finds the matching users in the same dataset in $30$ minutes.
\subsubsection*{Summary}
In this experimental study, we evaluated various aspects of the $k$-$l$
diversity based linkage model and the \emph{ST-Link} algorithm. We studied the
scalability of the algorithm and showed that it scales linearly with the
dataset size. We also studied the effectiveness of the linkage and showed that
high precisions can be achieved. Using the unweighted version of our
model, some of that precision can be traded off in order to achieve better
recall values as well.
\section{Related Work}\label{sec:related}
\noindent \textbf{Record Linkage.}
One of the earliest appearances of the term \emph{record linkage} is by
Newcombe et al.~\cite{ref:blocking1, ref:ER}.
Several surveys exist on the
topic~\cite{ref:survey1, ref:survey2, ref:survey3}. Most of the work in this
area focus on a single type of databases and define the linked records with
respect to a similarity metric. The input to such a record linkage algorithm
is a set of records and the output from it is a clustering of records. In
contrast, our problem involves linking users from two datasets, where each
user can have multiple spatio-temporal records. A theoretical approach, and
its validation, on linking users across domains is studied
recently~\cite{ref:riederer}. The first phase this work is computing a score
for every candidate pair. In a second phase they construct a bipartite graph
of users and reduce the problem into bipartite assignment problem. Their
experiments validate the accuracy of this two phase computation. While many
works on record linkage focus on accuracy~\cite{ref:dong, ref:citations, ref:riederer}
and a few on scalability~\cite{ref:joint}, our work must consider both. In our case,
successful linkage does not rely solely on the similarity of records and as
such \emph{ST-Link} algorithm searches multiple diverse matches, aka $k$-$l$
diversity, and also makes sure that there are no negative matches, aka alibis.
To the best of our knowledge, this is a novel approach for record linkage,
specifically targeted at spatio-temporal datasets
\\
\noindent\textbf{Temporal Record Linkage and Entity Evolution.} Temporal record linkage
differs from traditional record linkage in that it takes entity evolution into
account (e.g., a person can change her phone number). The \emph{time decay}
model captures the probability of an entity changing its attribute value
within a given time interval~\cite{ref:timedecay}. The \emph{mutation model}
learns the probability of an attribute value re-appearing over
time~\cite{ref:mutation}. The \emph{transition model} learns the probability
of complex value transitions over time~\cite{ref:transition}. Furthermore,
declarative rules can be used to link records
temporally~\cite{ref:declarative}. Transition model can also capture complex
declarative rules. Temporal record linkage algorithms are able to capture the
entity evolution and determine if an entity has changed the value of one or
more of its attributes. Our problem has some resemblance to entity evolution,
since the location attributes of the users change over time. However, this
change can be better described as entity mobility, rather than entity
evolution. Application of aforementioned models to spatio-temporal datasets
might be effective in predicting a user's next stop or calculating the
probability of whether a user will return back to a given location. Yet, they
would fell short of linking spatio-temporal records of users
\\
\noindent \textbf{Spatial Record Linkage and Spatial Joins.} Many join and
self-join algorithms are proposed in the literature for spatial
data~\cite{ref:spatialJoin}. Sehgal et al.~\cite{ref:spatialER} proposes a
method to link the spatial records by integrating spatial and non-spatial
(e.g. location name) features. However, spatial record linkage and spatial
join algorithms are not extensible to spatio-temporal data as they are based
on intersection of minimum bounding boxes, one-sided nearest join, or string
similarity. Spatio-temporal joins are more complex with constrains on both
spatial and temporal domains~\cite{ref:cstj}. Yet our problem involves more
than spatio-temporal records, it involves matching spatio-temporal record
series from two datasets
\\
\noindent \textbf{Trajectory Join.}
Bakalov et al.~\cite{ref:bakalov1} define the trajectory joins as the
identification of all pairs of similar trajectories given two datasets. They
represent an object trajectory as a sequence of symbols. Based on the symbol
similarity, they prune the pairwise trajectory comparisons. Effective
evaluation of symbol similarity is supported by a tree-like index scheme.
In~\cite{ref:cstj}, the authors extend the problem to continuous queries over
streaming spatio-temporal trajectories. An important difference between
trajectory join algorithms and our work is that trajectory similarity is not
necessarily an indication of a linkage and vice verse. If one of the datasets
is denser than the other, trajectories would be dissimilar, yet we still can
have matching user pairs based on $k$-$l$ linkage. However, some indexing
structures of trajectory join algorithms are closely related to our approach.
There are multiple indexing schemes for spatio-temporal data.
In~\cite{ref:gedik,ref:gridPart, ref:gridPart2, ref:sina} various grid based
structures are used for indexing. Our spatial filtering approach is similar in
its use of a grid-based index, but instead of associating objects with grid
cells, we associate users with grid cells based on the frequency of their
events residing in these cells. There are also tree-like spatio-temporal
indexing structures, surveyed in~\cite{ref:tree-like}. A common theme of these
works is the reduction of the update cost, which is not a concern in our
work
\\
\noindent \textbf{User Identification.} Our work has
commonalities with the work done in the area of user identification. For
instance, de~Montjoye et al.~\cite{ref:4point} has shown that, given a spatio-
temporal dataset of call detail records, one can uniquely identify the 95~\%
of the population by using $4$ randomly selected spatio-temporal points. Similar to our
discussion, the authors mention that spatio-temporal points do not contribute
to information gain equally. In our work, we cover this by introducing a
\emph{weight} function. Unlike our work, \cite{ref:4point} does not consider
the linkage problem,
instead, they study how users can be uniquely identified within a single
dataset using a small subset of their records. Another related work
is~\cite{ref:ccard}, in which authors show that using the credit card
metadata, they can identify unique users and group the transactions with
respect to users. In addition to spatio-temporal reference data, they use the
transaction price and gender as auxiliary information. Another related work is
from Rossi et al.~\cite{ref:st-identity}, in which user identification
techniques for GPS mobility data is presented. They
use a classification based algorithm rather than pairwise comparison of
records. Importantly, our
algorithm does not use any auxiliary information but only spatio-temporal
data, and it aims to match entities across datasets.
\section{Conclusion}\label{sec:conclusion}
\noindent In this paper, we studied the problem of matching
real-world entities using spatial-temporal usage records from two different
LESs. By introducing the $k$-$l$--diversity model, a
novel concept that captures both spatial and temporal diversity aspects of the
linkage, we study the challenge of defining similarity between usage records
of entities from different datasets. As part of this model, we introduced
the concept of an \emph{alibi}, which effectively filters out negative
matches and significantly improves the linkage quality.
To realize the $k$-$l$--diversity model, we developed the scalable
\emph{ST-Link} algorithm that makes use of effective filtering steps.
Taking advantage of the spatial nature of the data, users are associated with
\emph{dominating grids} --- grids that contain most activities of their entities.
This enables processing each grid independently, improving scalability. Taking
advantage of the temporal nature of the data, we slide a window over both
datasets jointly and maintain set of candidate users that have co-occurring
events but no alibis. The set of candidate entities are pruned as the window is
slided.
Our experimental evaluation, conducted with several data sets showed that the running
time of the \emph{ST-Link} algorithm scales linearly with the dataset size.
Moreover, precision of the linkage results is practically $1$ for most $k$-$l$
settings. We also observed that using an unweighted version of our linkage
model, the precision can be sacrificed to achieve higher recall values.
Spatio-temporal linkage can enable gathering large dynamic data sets for many
social good applications, such as smart cities and environmental monitoring.
We are also investigating privacy preserving methodologies, those are needed
to prevent information leakage while analyzing and sharing location based
information~\cite{ref:gedikPrivacy, ref:emreTDSC2017}.
\bibliographystyle{abbrv}
|
\subsection*{Scientific Background}
If we examine the information that drives machine
learning today, we find that it is almost entirely
statistical. In other words, learning machines improve
their performance by optimizing parameters over a stream of
sensory inputs received from the environment.
It is a slow process, analogous in many respects
to the natural selection process that drives Darwinian evolution.
It explains how species like eagles and snakes have developed
superb vision systems over millions of years.
It cannot explain however the super-evolutionary
process that enabled humans to build eyeglasses and telescopes
over barely one thousand years. What humans possessed that
other species lacked was a mental representation, a
blue-print of their environment which they could manipulate
at will to {\em imagine} alternative hypothetical environments
for planning and learning. Anthropologists like N.\ Harari,
and S.\ Mithen are in general agreement that the decisive
ingredient that gave our Homo sapiens ancestors the
ability to achieve global dominion, about 40,000
years ago, was their ability to choreograph a mental representation
of their environment, interrogate that representation, distort
it by mental acts of imagination and finally
answer ``What if?'' kind of questions.
Examples are interventional questions: ``What if I act?''
and retrospective or explanatory questions:
``What if I had acted differently?''
No learning machine in operation today can answer such questions
about interventions not encountered before, say, ``What if we ban cigarettes.''
Moreover, most learning machines today do not provide a representation
from which the answers to such questions can be derived.
I postulate that the major impediment to achieving accelerated
learning speeds as well as human level performance should be
overcome by removing these barriers and equipping
learning machines with causal reasoning tools. This postulate
would have been speculative
twenty years ago, prior to the mathematization of counterfactuals.
Not so today.
Advances in graphical and
structural models have made counterfactuals
computationally manageable and thus
rendered model-driven reasoning a more promising direction
on which to base strong AI.
In the next section, I will describe the impediments facing
machine learning systems using a three-level hierarchy that governs inferences
in causal reasoning.
The final section summarizes how these impediments were
circumvented using modern tools of causal inference.
\newpage
\subsection*{The Three Layer Causal Hierarchy}
\begin{figure*}[h]
\centering
\begin{tabular}{| p{3.25cm} | p{2.45cm} | p{4.5cm} | p{4.5cm} |}
\hline
Level & Typical & Typical Questions & Examples \\
(Symbol) & Activity & & \\
\hline
{1. Association \hfill} \linebreak $P(y|x)$ &
Seeing &
{What is? \hfill}
\linebreak {How would seeing $X$ \hfill}
\linebreak {change my belief in$Y$? \hfill} &
{What does a symptom tell me about a disease? \hfill}
\linebreak What does a survey tell us about the election results? \\ \hline
{2. Intervention \hfill} \linebreak $P(y|do(x),z)$ &
{Doing \linebreak Intervening} &
{What if?\hfill}
\linebreak What if I do $X$? &
{What if I take aspirin, will my headache be cured? \hfill}
\linebreak What if we ban cigarettes? \\ \hline
{3. Counterfactuals \hfill} \linebreak $P(y_{x}|x',y')$ &
{Imagining, \linebreak Retrospection} &
{Why? \hfill}
\linebreak {Was it $X$ that caused $Y$? \hfill}
\linebreak {What if I had acted \hfill}
\linebreak differently? &
{Was it the aspirin that stopped my headache? \hfill}
\linebreak {Would Kennedy be alive had Oswald not shot him? \hfill}
\linebreak What if I had not been smoking the past 2 years? \\ \hline
\end{tabular}
\caption{The Causal Hierarchy. Questions at level $i$
can only be answered if information from level
$i$ or higher is available. \label{fig1}}
\end{figure*}
An extremely useful insight unveiled by the
logic of causal reasoning is the existence of a sharp
classification of causal information, in terms
of the kind of questions that each class is
capable of answering. The classification forms
a 3-level hierarchy in the sense that questions
at level $i$ $(i=1,2,3)$ can only be answered if
information from level $j$ ($j \geq i$)
is available.
Figure \ref{fig1} shows the 3-level hierarchy, together with the
characteristic
questions that can be answered at each level.
The levels are titled 1.\ Association, 2.\ Intervention, and 3.\
Counterfactual. The names of these layers were chosen to emphasize
their usage.
We call the first level Association, because it invokes
purely statistical relationships, defined by the naked data.\footnote{Other names used
for inferences at this layer are: ``model-free,''
``model-blind,'' ``black-box,'' or ``data-centric.''
\citet{darwiche:17}
used ``function-fitting,''
for it amounts to fitting data by a complex function
defined by the neural network architecture.}
For instance, observing a customer
who buys toothpaste makes it more likely that he/she buys floss; such association can be inferred directly
from the observed data using conditional expectation.
Questions at this layer, because they require no causal information, are
placed at the bottom level on the hierarchy.
The second level, Intervention, ranks higher than Association
because it involves not just seeing what is, but changing what we
see. A typical question at this level would be:
What happens if we double the price? Such questions cannot
be answered from sales data alone, because they involve
a change in customers behavior, in reaction to the new pricing.
These choices may differ substantially from those taken
in previous price-raising situations. (Unless
we replicate precisely the market conditions that existed
when the price reached double its current value.)
Finally, the top level is called Counterfactuals, a term
that goes back to the philosophers David Hume and John
Stewart Mill, and which has been given computer-friendly
semantics in the past two decades. A typical question in the
counterfactual category is ``What if I had acted differently,''
thus necessitating retrospective reasoning.
Counterfactuals are placed at the top of the hierarchy because
they subsume interventional and associational questions.
If we have a model that can answer counterfactual queries,
we can also answer questions about interventions
and observations. For example, the interventional question,
What will happen if we double the price?
can be answered by asking the counterfactual question:
What would happen had the price been twice its current value?
Likewise, associational questions can be answered once we can
answer interventional questions; we simply ignore the
action part and let observations take over.
The translation does not work in the opposite direction.
Interventional questions cannot be answered from purely
observational information (i.e., from statistical data alone).
No counterfactual question involving retrospection can be answered
from purely interventional information, such as that acquired
from controlled experiments; we cannot re-run an experiment on
subjects who were treated with a drug and see how they behave
had they not given the drug.
The hierarchy is therefore directional, with the top level
being the most powerful one.
Counterfactuals are the building blocks of scientific thinking
as well as legal and moral reasoning.
In civil court, for example, the defendant is considered to
be the culprit of an injury if, {\em but for} the
defendant's action, it is more likely than not that the
injury would not have occurred.
The computational meaning of {\em but for} calls for
comparing the real world to an alternative world in which
the defendant action did not take place.
Each layer in the hierarchy has
a syntactic signature that characterizes the
the sentences admitted into that layer. For example, the association
layer is characterized by conditional probability
sentences, e.g., $P(y|x)= p$ stating that:
the probability of event $Y=y$ given that we observed
event $X=x$ is equal to $p$. In large systems, such evidential sentences
can be computed efficiently using Bayesian Networks,
or any of the neural networks that support
deep-learning systems.
At the interventional layer we find sentences
of the type $P(y|do(x),$\newline$z)$, which denotes
``The probability of event $Y=y$ given that we intervene
and set the value of $X$ to $x$ and subsequently observe event $Z=z$.
Such expressions can be estimated experimentally from randomized
trials or analytically using Causal Bayesian
Networks \citep[Chapter 3]{pearl:2k}.
A child learns the effects of interventions through
playful manipulation of the environment (usually in a deterministic
playground), and AI planners obtain interventional
knowledge by exercising their designated sets of actions.
Interventional expressions cannot be inferred from
passive observations alone, regardless of how big the data.
Finally, at the counterfactual level, we have expressions
of the type
$P(y_{x} |x',y')$ which stand for
``The probability that event $Y=y$ would be observed
had $X$ been $x$, given that
we actually observed $X$ to be $x'$ and and $Y$ to be $y'$.
For example, the probability that Joe's salary would
be $y$ had he finished college, given that his actual
salary is $y'$ and that he had only two years of college.''
Such sentences can be computed only when
we possess functional or Structural Equation models,
or properties of such models \citep[Chapter 7]{pearl:2k}.
This hierarchy, and the formal restrictions it entails,
explains why statistics-based
machine learning systems are prevented from
reasoning about actions, experiments and
explanations. It also informs us what extra-statistical information
is needed, and in what format, in order to support those
modes of reasoning.
Researchers are often surprised that
the hierarchy denegrades the impressive achievements
of deep learning to the level of Association, side
by side with textbook curve-fitting exercises. A popular
stance against this comparison argues that, whereas the objective
of curve-fitting is to maximize ``fit,''
in deep learning we try to minimize ``over fit.''
Unfortunately, the theoretical barriers that separate
the three layers in the hierarchy tell us that the
nature of our objective function does not matter.
As long as our system optimizes some property
of the observed data, however noble or sophisticated,
while making no reference to the world
outside the data, we are back to level-1 of the hierarchy
with all the limitations that this level entails.
\subsection*{The Seven Pillars of the Causal Revolution
(or What you can do with a causal model that you could not
do without?)}
Consider the following five questions:
\begin{itemize}
\item How effective is a given treatment in preventing a disease?
\item Was it the new tax break that caused our sales to go up?
\item What is the annual health-care costs attributed to obesity?
\item Can hiring records prove an employer guilty of sex discrimination?
\item I am about to quit my gob, but should I?
\end{itemize}
The common feature of these questions is that they
are concerned with cause-and-effect relationships. We can
recognize them through words such as ``preventing,'' ``cause,''
``attributed to,'' ``discrimination,'' and ``should I.''
Such words are common in everyday language, and our society
constantly demands answers to such questions.
Yet, until very recently science gave us no means even to
articulate them, let alone answer them.
Unlike the rules of geometry, mechanics, optics or probabilities,
the rules of cause and effect have been denied the benefits of mathematical analysis.
To appreciate the extent of this denial,
readers would be stunned to know that
only a few decades ago scientists were unable to
write down a mathematical equation for the obvious fact that
``mud does not cause rain.'' Even today, only the top
echelon of the scientific community can write such an
equation and formally distinguish ``mud causes rain'' from
``rain causes mud.'' And you would probably be even more
surprised to discover that your favorite college professor is
not among them.
Things have changed dramatically in the past three decades,
A mathematical
language has been developed for managing causes and effects,
accompanied by a set of tools that turn causal analysis
into a mathematical game, not unlike solving algebraic
equations, or finding proofs in high-school geometry.
These tools permit us to express causal questions formally
codify our existing knowledge in both diagrammatic
and algebraic forms, and then leverage our data to
estimate the answers. Moreover, the theory warns us when
the state of existing knowledge or the available data
are insufficient to answer our questions; and then
suggests additional sources of knowledge or data to make the
questions answerable.
Harvard professor Garry King gave this
transformation a historical perspective:
``More has been learned about causal inference in the last
few decades than the sum total of everything that had been
learned about it in all prior recorded history'' \citep{morgan:win15}.
I call this transformation ``The Causal Revolution,'' \citep{pearl:mac18}
and the mathematical framework that led to it
I call ``Structural Causal Models (SCM).''
The SCM deploys three parts
\begin{enumerate}
\item Graphical models,
\item Structural equations, and
\item Counterfactual and interventional logic
\end{enumerate}
Graphical models serve as a language
for representing what we know about the world,
counterfactuals help us to articulate what we
want to know, while structural equations serve to
tie the two together in a solid semantics.
\begin{figure}[h]
\begin{center}
\epsfig{figure=r475-boxes-test.eps,width=3in}
\end{center}
\caption{How the SCM ``inference engine'' combines data with causal model (or assumptions) to produce
answers to queries of interest.\label{fig2}}
\end{figure}
Figure \ref{fig2} illustrates the operation of SCM in the form of
an inference engine. The engine accepts three inputs:
Assumptions, Queries, and Data, and produces three outputs:
Estimand, Estimate and Fit indices. The Estimand $(E_S)$ is a
mathematical formula that, based on the Assumptions, provides
a recipe for answering the Query from any hypothetical data,
whenever they are available. After receiving the Data, the
engine uses the Estimand to produce an actual Estimate
$(\hat{E}_S)$
for the answer, along with statistical estimates of the
confidence in that answer (To reflect the limited size of
the data set, as well as possible measurement errors or
missing data.) Finally, the engine produces a list of
``fit indices'' which measure how compatible the data are
with the Assumptions conveyed by the model.
To exemplify these operations, let us assume that our Query
stands for the causal effect of $X$ on $Y$, written $Q=P(Y|do(X))$,
where $X$ and $Y$ are two variables of interest.
Let the modeling assumptions be encoded in the graph below,
\begin{figure}[h]
\begin{center}
\epsfig{figure=r475-cacm-fig2.eps,width=1.25in}
\end{center}
\end{figure}
where $Z$ is a third variable affecting both $X$ and $Y$.
Finally, let the data
be sampled at random from a joint distribution $P(X,Y,Z)$.
The Estimand $(E_S)$ calculated by the engine will be the
formula $E_S = \sum_z P(Y|X,Z)P(Z)$. It defines
a property of P(X,Y,Z) that, if estimated,
would provide a correct answer to our Query. The answer itself,
the Estimate $\hat{E}_S$, can be produced by any number of techniques
that produce a consistent estimate of $E_S$
from finite samples of $P(X,Y,Z)$. For example, the sample
average (of $Y$) over all cases satisfying the specified
$X$ and $Z$ conditions, would be a consistent estimate.
But more efficient estimation techniques can be devised to
overcome data sparsity \citep{rosenbaum:rub83}.
This is where deep learning excels and where most work in machine learning
has been focused, albeit with no guidance of a model-based estimand. Finally, the Fit
Index in our example will be NULL. In other words,
after examining the structure of the graph, the
engine should conclude that the assumptions encoded do not
have any testable implications. Therefore, the veracity of
resultant estimate must lean entirely on the
assumptions encoded in the graph -- no refutation
nor corroboration can be obtained from the data.\footnote{The assumptions encoded in the graph
are conveyed by its missing arrows. For example,
$Y$ does not influence $X$ or $Z$, $X$ does not influence $Z$
and, most importantly, $Z$ is the only variable affecting
both $X$ and $Y$. That these assumptions lack testable implications can
be concluded from the fact that the graph is complete, i.e.,
no edges are missing.}
The same procedure applies to more sophisticated
queries, for example, the counterfactual query $Q=P(y_x|x',y')$
discussed before. We may also permit some of the data to
arrive from controlled experiments, which would take the form
$P(V|do(W))$, in case W is the controlled variable.
The role of the Estimand would remain that
of converting the Query into the syntactic format of the
available data and, then, guiding the choice of the estimation
technique to ensure unbiased estimates. Needless to state,
the conversion task is not always feasible, in which case the Query will
be declared ``non-identifiable'' and the engine should
exit with FAILURE. Fortunately, efficient and complete
algorithms have been developed to decide identifiability
and to produce estimands for a variety of counterfactual
queries and a variety of data types \citep{bareinboim:pea16-r450}.
Next we provide a bird's eye view of seven
accomplishments of the SCM framework and discuss the unique
contribution that each pillar brings to the art of automated reasoning.
\section*{Pillar 1: Encoding Causal Assumptions -- Transparency and Testability \label{sec1}}
The task of encoding assumptions in a compact and
usable form, is not a trivial matter once we take
seriously the requirement of transparency and testability.\footnote{Economists, for example, having chosen algebraic
over graphical representations, are deprived of elementary testability-detecting features \citep{pearl:15-r391}.}
Transparency enables analysts to discern whether the assumptions encoded
are plausible (on scientific grounds), or whether additional
assumptions are warranted.
Testability permits us (be it an analyst or a machine) to
determine whether the assumptions encoded are compatible
with the available data and, if not, identify those
that need repair.
Advances in graphical models have
made compact encoding feasible. Their transparency
stems naturally from the fact that all assumptions are encoded
graphically, mirroring the way researchers perceive of
cause-effect relationship in the domain;
judgments of counterfactual or
statistical dependencies are not required, since these can
be read off the structure of the graph.
Testability is facilitated through
a graphical criterion called $d$-separation, which
provides the fundamental connection between
causes and probabilities.
It tells us, for any given pattern
of paths in the model, what pattern of dependencies we
should expect to find in the data \citep{pearl:88a}.
\section*{Pillar 2: $Do$-calculus and the control of confounding \label{sec2}}
Confounding, or the presence of unobserved causes of two or
more variables, has long been consider the the major obstacle to
drawing causal inference from data, This obstacle had been
demystified and ``deconfounded'' through a graphical
criterion called ``back-door.'' In particular, the task of
selecting an appropriate set of covariates to control for
confounding has been reduced to a simple ``roadblocks''
puzzle manageable by a simple algorithm \citep{pearl:93ss}.
For models where the ``back-door'' criterion does not hold,
a symbolic engine is available, called $do$-{\em calculus},
which predicts the effect of policy interventions
whenever feasible, and exits with failure whenever
predictions cannot be ascertained with the specified
assumptions \citep{pearl:95,tian:pea02,shpitser:pea08-r336}.
\section*{Pillar 3: The Algorithmization of Counterfactuals \label{sec3}}
Counterfactual analysis deals with behavior of specific
individuals, identified by a distinct set of characteristics,
For example, given that Joe's salary is $Y = y$, and that he
went $X=x$ years to college, what would Joe's salary be had he
had one more year of education.
One of the crown achievements of the Causal Revolution has been
to formalize counterfactual reasoning
within the graphical representation, the
very representation researchers use to encode scientific knowledge.
Every structural equation model determines
the truth value of every counterfactual
sentence. Therefore, we can determine analytically
if the probability of the sentence is estimable from experimental
or observational studies, or combination thereof [\citealp{balke:pea94a}; \citealp[Chapter 7]{pearl:2k}].
Of special interest in causal discourse are counterfactual
questions concerning ``causes of effects,'' as opposed to
``effects of causes.'' For example, how likely it is that
Joe's swimming exercise was a necessary (or
sufficient) cause of Joe's death \citep{pearl:15-r431,halpern:pea05-r266a}.
\section*{Pillar 4: Mediation Analysis and the Assessment of Direct and
Indirect Effects \label{sec4}}
Mediation analysis concerns the mechanisms
that transmit changes from a cause to its effects.
The identification of such intermediate
mechanism is essential for generating explanations and
counterfactual analysis must be invoked to facilitate
this identification. The graphical
representation of counterfactuals enables us to define
direct and indirect effects and to decide when
these effects are estimable from data, or experiments \citep{robins:gre92, pearl:01,vanderweele:15}.
Typical queries answerable by this analysis are: What
fraction of the effect of $X$ on $Y$ is mediated by variable $Z$.
\section*{Pillar 5: External Validity and Sample Selection Bias \label{sec5}}
The validity of every experimental study is challenged
by disparities between the experimental and implementational
setups. A machine trained in one environment cannot
be expected to perform well when environmental conditions
change, unless the changes are localized and identified.
This problem, and its various manifestations are well
recognized by machine-learning researchers, and enterprises
such as ``domain adaptation,'' ``transfer learning,''
``life-long learning,'' and ``explainable AI,'' are just some
of the subtasks identified by researchers and funding agencies in an
attempt to alleviate the general problem of robustness.
Unfortunately, the problem of robustness requires a
causal model of the environment, and cannot be handled at the
level of Association, in which most remedies were tried.
Associations are not sufficient for identifying
the mechanisms affected by changes that occurred.
The $do$-calculus discussed above now offers
a complete methodology for overcoming bias due
to environmental changes. It can be used both
for re-adjusting learned policies to circumvent environmental
changes and for controlling bias due to non-representative
samples \citep{bareinboim:pea16-r450}.
\section*{Pillar 6: Missing Data \label{sec6}}
Problems of missing data plague every branch of experimental
science. Respondents do not answer every item on a
questionnaire, sensors fade as environmental conditions change,
and patients often drop from a clinical study for
unknown reasons. The rich literature on this problem is
wedded to a model-blind paradigm of statistical analysis
and, accordingly, it is severely limited to situations where
missingness occurs at random, that is, independent of values
taken by other variables in the model. Using causal models of the
missingness process we can now formalize the conditions under which
causal and probabilistic relationships can be recovered from
incomplete data and, whenever the conditions are satisfied,
produce a consistent estimate of the desired relationship \citep{mohan:pea17-r473}.
\section*{Pillar 7: Causal Discovery \label{sec7}}
The $d$-separation criterion described above
enables us to detect and enumerate the testable
implications of a given causal model. This opens
the possibility of inferring, with mild assumptions,
the set of models that are compatible with
the data, and to represent this set compactly.
Systematic searches have been developed which,
in certain circumstances,
can prune the set of compatible models significantly
to the point where causal queries can be
estimated directly from that set \citep{spirtes:etal00,pearl:2k,peters:etal17}.
\section*{Conclusions}
The philosopher Stephen Toulmin (\citeyear{toulmin:61})
identifies model-based vs.\ model-blind dichotomy as the key
to understanding the ancient rivalry between Babylonian
and Greek science. According to Toulmin,
the Babylonians astronomers were masters of black-box
prediction, far surpassing their Greek rivals in accuracy
and consistency \cite[pp.\ 27--30]{toulmin:61}.
Yet Science favored the
creative-speculative strategy of the Greek astronomers which was
wild with metaphysical imagery: circular tubes full of fire,
small holes through which celestial fire was visible as stars,
and hemispherical earth riding on turtle backs.
Yet it was this wild modeling strategy, not Babylonian rigidity,
that jolted Eratosthenes (276-194 BC) to perform one of the most
creative experiments in the ancient world and measure the
radius of the earth. This would never have occurred to
a Babylonian curve-fitter.
Coming back to strong AI, we have seen that
model-blind approaches have intrinsic
limitations on the cognitive tasks that they can perform.
We have described some of these tasks and demonstrated
how they can be accomplished in the SCM framework, and why a model-based approach
is essential for performing these tasks. Our general conclusion is that human-level AI cannot emerge
solely from model-blind learning machines;
it requires the symbiotic collaboration of data and models.
Data science is only as much of a science
as it facilitates the interpretation of data --
a two-body problem, connecting data to reality.
Data alone are hardly a science, regardless how big they get and
how skillfully they are manipulated.
\section*{Acknowledgement}
This research was supported in parts by grants from Defense Advanced Research Projects Agency [\#W911NF-16-057], National Science Foundation [\#IIS-1302448, \#IIS-1527490, and \#IIS-1704932], and Office of Naval Research [\#N00014-17-S-B001].
|
\section{Introduction}
Extremely metal-poor (EMP) stars are the shining witnesses of a primordial
Universe
corresponding to a redshift higher than 10.
They belong to the second generation of stars, objects formed from the gas
enriched
only by the first generation of the stars, Pop\,III stars.
Direct observation at high redshift of the individual
massive Pop\,III stars is not feasible because they are too
faint for the current and next generation of telescopes.
There is one possible exception.
In the case of Pop\,III stars formed with masses of a few hundred M$_\odot$
\citep{hirano14},
they could have ended their life with a pair-instability supernovae (PISN)
event.
According to \citet{bromm14}, this kind of event could be visible for the
James Webb Space Telescope (JWST) planned to be launched in 2019.
However, even in this event it is unlikely that the low-resolution spectra
that could be obtained from JWST would be able to provide a detailed chemical
inventory of PISN ejecta.
At the moment, EMP stars are the only way to deduce the characteristics
of the Universe at the time they formed.
The TOPoS project \citep{topos1} has been designed to increase the number
and
chemical inventory of EMP turn-off stars. The TOPoS targets were selected
from the stars observed spectroscopically in the SDSS/SEGUE/BOSS survey
with the following colour cuts: $0.18 \le (g-z)_0 \le 0.70$ and $(u-g)_0
> 0.70$.
This allows the identification of stars likely to be close to the main sequence
turn-off of halo stars.
\section{Observations and data reduction}
Target stars were observed using the ESO Ultraviolet and Visual Echelle Spectrograph
\citep[UVES,][]{dekker00} mounted at the Kueyen 8.2m telescope, the second
unit telescope
of the Very Large Telescope (VLT, Cerro Paranal, Chile). Observations of
five stars were collected in service
mode in the course of Large Programme 189.D-0165 (PI E. Caffau)
between April 2013 and March 2016. We adopted setting Dic\,1 390+580 and
therefore observed simultaneously with the blue and red UVES arms, centred
at 390\,nm and 580\,nm, and covering the spectral ranges 326-454\,nm and
476-684\,nm, respectively. The data were binned $2\times 2$ (spectral $\times$
spatial direction),
and a 1.4$^{\prime\prime}$ slit width was adopted. This resulted in a
resolving power of R$\simeq$30000 or larger when the image quality was better
than
1.4$^{\prime\prime}$. A different number of 3005\,s frames were taken for
six of the
target stars, namely 1, 5, 11, 5, 3, and 6 for stars SDSS\,J0140+2344,
SDSS\,J1034+0701, SDSS\,J1035+0641, SDSS\,J1247--0341,
SDSS\,J1442--0015, and SDSS\,J1507+0051, respectively. The
signal-to-noise ratios per pixel of individual spectra varied in the range
3-30 for the blue arm and 8-61 for
the red arm.
\begin{figure*}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[draft = false, clip=true]
{plot_cak.eps}}
\end{center}
\caption[]{Spectral ranges of the \ion{Ca}{ii}-H and -K lines of the sample
stars.
}
\label{plot_HK}
\end{figure*}
\begin{figure*}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[draft = false, clip=true]
{cfe-topos4.ps}}
\end{center}
\caption[]{ Carbon abundances of the programme stars (filled red circles)
as a function of [Fe/H]. Other stars
published by our group are shown as open red circles.
Other stars from the literature
\citep{sivarani,Plez05,PCM05,frebel05,frebel06,thompson08,aoki08,behara,masseron10,masseron12,yong2013,Cohen13,Li2015}
are shown as blue squares.
The violet star is the upper limit on SDSS\,J1029+1729 \citep{leostar,leostaraa}.
The horizontal dotted lines delimit the low-carbon band.
}
\label{plot_c}
\end{figure*}
\begin{table}
\caption{ SDSS DR12 $^1$ photometry and exposure times for the programme
stars.}
\label{data}
\begin{center}
\renewcommand{\tabcolsep}{1pt}
\begin{tabular}{lccccccccr}
\hline \noalign{\smallskip}
Star & $u$ & $g$ & $r$ & $i$ & $z$
& T$_{exp}$ & N$_{\rm exp}$\\
& & & & &
& s & \\
\hline \noalign{\smallskip}
SDSS\,J014036.22+234458.0 & 16.78 & 15.82 & 15.35 & 15.12 & 15.03 & 3005
& 1 \\
SDSS\,J103402.71+070116.6 & 18.34 & 17.52 & 17.23 & 17.13 & 17.12 & 3005
& 5 \\
SDSS\,J103556.11+064143.9 & 19.53 & 18.65 & 18.37 & 18.31 & 18.26 & 3005
& 29 \\
SDSS\,J124719.46--034152.4 & 19.32 & 18.50 & 18.24 & 18.15 & 18.14 & 3005
& 5 \\
SDSS\,J134922.91+140736.9 & 17.44 & 16.65 & 16.33 & 16.24 & 16.19 & 3600
& 9 \\
SDSS\,J144256.37--001542.8 & 18.76 & 17.96 & 17.65 & 17.51 & 17.49 & 3005
& 3 \\
SDSS\,J150702.02+005152.6 & 19.75 & 18.77 & 18.57 & 18.48 & 18.41 & 3005
& 6 \\
\noalign{\smallskip}
\hline
\end{tabular}
$^1$In this work we used the SDSS DR12 photometry; the magnitudes of the
latest release DR14, differ from these values by less that 0.015
mag.
\end{center}
\end{table}
Star SDSS\,J1035+0641 was also observed with setting Dic\,2 437+760
(programme ID: 096.D-0468 PI Sbordone), with the blue and red arms covering
the spectral ranges
373-499\,nm and 565-946\,nm, respectively. A total of 18\,$\times$3005\,sec
frames were acquired. The same slit width (1\farcs{4}) and on-chip
binning ($2\times 2$) were chosen. Signal-to-noise ratios per pixel of individual
spectra varied in
the range 5-10 and 4-8 for the blue and red arms, respectively.
SDSS\,J1349+1407 was observed in service mode in the course
of programmes 090.D-0306 and 093.D-0136 (PI Sbordone) between March 2013
and
June 2014. The setting used in this case was Dic2 437+760 with
a $1\times 1$ binning and a 1\farcs{0} slit, providing a resolving power
of R$\sim 48000$. A total of nine integrations
of 3600s each were acquired. The coadded spectrum attains a S/N $\sim 40$
per pixel at 400\,nm
and a S/N $\sim 60$ per pixel at 590\,nm.
The data were reduced by the ESO staff, and retrieved through the Phase\,3
spectral
data products query
form\footnote{\url{http://archive.eso.org/wdb/wdb/adp/phase3_spectral/form}}.
\section{Analysis}
The chemical analysis has been done with the code MyGIsFOS\ \citep{mygisfos}.
This code derives the chemical abundances by fitting a selected set of features.
The best fit profile is obtained by interpolating in a series of pre-calculated
synthetic spectra.
These synthetic spectra have been computed with {\tt turbospectrum} \citep{alvarez_plez,2012ascl.soft05004P}
and are based on a grid of 1D plane-parallel hydrostatic model atmospheres
previously computed with MARCS \citep{G2008}.
The procedure to derive the atmospheric parameters and chemical composition
is described by \citet{topos1} and has
also been used in the other papers of this series \citep{topos2,topos3}.
There were two exceptions for which MyGIsFOS\ was not well adapted: carbon
in all stars, and magnesium
in the star SDSS\,J1349+1407. Carbon was derived by $\chi$-squared fitting
of the
G-band with a specifically computed grid of synthetic spectra with fixed
metallicity, $T\rm_{eff}$,\ and \ensuremath{\log g}\
and with varying C abundance. This technique is the same as that
used in our previous papers \citep{spite13,topos2}.
To take into account the known
granulation effects on the G-band \citep{Collet2007,behara,Gallagher2016,Gallagher2017},
we used the corrections provided by \citet{Gallagher2016}, assuming a metallicity
of --3.0.
The stellar parameters and abundances of C and Li are provided in Table\,\ref{param}.
Because
its [Mg/Fe] ratio is so extreme (Mg enhanced by more than 1\,dex over Fe),
one concern for SDSS\,J1349+1407 was that its temperature
structure could be different
from
that of a normal solar-scaled and $\alpha$-enhanced model.
Magnesium
is one of the
main electron donors for solar-type stars, and its abundance may affect the
ionisation
structure of the atmosphere.
We verified that the high [Mg/H] does not affect the abundances derived with
models enhanced in Mg by
the canonical 0.4\,dex
by computing both the ATLAS\,9 and ATLAS\,12 models \citep{kurucz93,kurucz05,SBC04,sbordone05},
the latter with [Mg/Fe]=+1.4 to
derive
the abundances.
From this test we conclude that
the differences are negligible; the largest difference reaches 0.05\,dex
for Mn.
At the very low metallicities relevant for the stars
analysed in this paper, Mg, Ca, and any metals cease to be important electron
donors;
as already discussed by \citet{nc04},
the main electron donor is, in fact, hydrogen.
\subsection{Individual stars and comparison with previous analyses}
All of our targets have already been analysed by our group
on the basis of X-Shooter spectra, and some have also been analysed by other
groups.
\begin{itemize}
\item SDSS\,J0140+2344.
The stellar parameters and [Fe/H] we derive are in excellent agreement with
those determined
by \citet{topos1} from the analysis of an X-Shooter spectrum.
Our estimated upper limit of A(C) from the X-Shooter spectrum was too optimistic:
the carbon abundance we derive in this work is 0.3\,dex larger than the X-Shooter
upper limit.
The Mg and Si abundances
measured from the X-Shooter and UVES spectra are in excellent agreement.
The Ca abundance measurements are more problematic.
We detect two \ion{Ca}{i} lines, the resonance line at 422.67\,nm and the
subordinate line
at 442.54\,nm. The two lines are very discrepant, as is generally the case
in EMP stars, even when NLTE effects are taken into
account \citep[see][and references therein]{spite12}.
The 422.67\,nm resonance line provides $[{\rm Ca/H}]=-3.97$, in good agreement
with the X-Shooter
spectrum where the only \ion{Ca}{i} line detected was in fact the 422.67\,nm
line,
which implied $[{\rm Ca/H}]=-3.78$.
The subordinate line yields a Ca abundance that is over 1\,dex higher
($[{\rm Ca/H}])=-2.79$). In Table\,\ref{abbo2} we only provide the value
from the subordinate line,
which, as discussed by \citet{spite12}, is more reliable.
The \ion{Ca}{ii} is derived from the K-line alone and it is a factor of three
lower than
the value we had estimated from the X-Shooter spectrum; however, in that
case we relied only on the \ion{Ca}{ii} infrared triplet lines. The abundance
from the K-line is also strongly discrepant from
the \ion{Ca}{i} subordinate line.
\citet{yong2013} analysed a Keck/HIRES spectrum of this star. They derived
$T\rm_{eff}$ = ${5703}$\,K from a combination of spectrophotometry and
Balmer line analysis. This is in agreement, within the errors, with our
adopted $T\rm_{eff}$. The surface gravity was derived from isochrones and two
solutions were found, depending on whether the star is a main sequence
or a subgiant star, this corresponds to \ensuremath{\log g} = ${3.36}$
or \ensuremath{\log g} = 4.68. In the two cases, they derive [Fe/H]=$-4.11$ or
[Fe/H]=$-4.02$ (with our adopted solar Fe abundance \citealt{abbosun}),
in excellent agreement with our result.
For Ca, they only measure the \ion{Ca}{i} 422.67\,nm resonance line
and derive [Ca/H]=$-3.82$ or [Ca/H]=$-3.88$, again in excellent
agreement with our results for the same line.
\citet{aoki13} analysed a Subaru spectrum of this star
and used $T\rm_{eff}$ = ${6103}$\,K, which was derived from the SDSS low-resolution
spectrum by the SEGUE Stellar Parameter Pipeline \citep[SSPP,][]{SSPP},
even though both $V-K$ and $g-r$ imply a temperature close to
that used by us. For the surface gravity they assumed
\ensuremath{\log g} = 4.0 and, like us, they derived [Fe/H]=$-3.67$ (on our adopted
solar abundance scale). Considering their adopted higher effective
temperature, this metallicity is consistent both with ours and with that
of \citet{yong2013}. For Ca they derive [Ca/H]=$-3.82$ from a single line.
Although it is not explicitly stated, it is highly probable that this
refers to the \ion{Ca}{i} 422.67\,nm resonance line.
Finally, \citet{aguado16} analysed the SDSS spectrum and a ISIS WHT spectrum
and derived
the stellar parameters ${6090\pm 200}$\,K/${4.7\pm 0.3}$/$-3.6\pm 0.2$.
Unsurprisingly, the $T\rm_{eff}$\ is close to the SSPP value and the derived metallicity
is consistent.
We conclude that the different analyses of this star are all consistent,
but we point
out the discrepancy of the Ca abundance derived from different lines. This
will have
to be revisited in the light of 3D NLTE computations in the future.
\smallskip
\item SDSS\,J1034+0701.
The X-Shoooter spectrum is analysed in \citet{pfr17}.
The [Fe/H] from the UVES spectrum is about 0.4\, dex lower than the value
derived from the X-Shooter spectrum.
The Mg abundance is in agreement within errors.
We did not derive the Ca abundance from the UVES spectrum, but the Ca abundance
from the X-Shooter spectrum, based on the \ion{Ca}{ii} IR triplet, and the
X-Shooter based Fe abundance implies [Ca/Fe]=+0.33.
\smallskip
\item SDSS\,J1035+0641.
\citet{topos2} investigated an X-Shooter spectrum, but they could only derive
A(C) and an upper limit for the Fe abundance.
The A(C) derived from the UVES spectrum
is a factor of two higher than the value we derived from the X-Shooter spectrum,
which illustrates
the limit of measuring the G-band for a warm star like this at moderate resolution.
The non-detection of the 382.0425\,nm line (equivalent width $< 0.4$ pm
at $1 \sigma$)
provides an upper limit of the Fe abundance [Fe/H]$< -5.2$, tighter than
the value
we derived from the X-Shooter spectrum.
A fit to the \ion{Ca}{ii} K line
provides an abundance that is consistent with that derived from the X-Shooter
spectrum.
We found a mistake in Table 7 of \citet{topos2} and in their Figure 10.
The upper limit on lithium, corresponding to an upper limit of 1.8\,pm,
is A(Li)$< 2.05$ and not 1.1, as incorrectly listed in Table 7. Our measured
equivalent
width of 1.3\,pm is consistent with that upper limit and corresponds to A(Li)=1.90.
\smallskip
\item SDSS\,J1247--0341.
The metallicity determined by \citet{topos1} from the X-Shooter spectrum
is in good agreement with that derived from the UVES spectrum.
Also in this case the C abundance derived from
the UVES spectrum is larger by 0.6\,dex than the value we estimated
from the X-Shooter spectrum. The Mg abundances from X-Shooter and UVES spectra
are consistent
within 1$\sigma$. \citet{topos1} provided the Ca abundance from
two lines of the \ion{Ca}{ii} IR triplet, which we considered more reliable
than the \ion{Ca}{i} 422.67\,nm resonance line. That line implied [Ca/H]=$-3.96$,
which is consistent with the abundance provided here for the same line.
Also in this case we point out how different lines provide different abundances
for Ca; this is likely due to the inadequacy of the 1D LTE analysis of this
element
in these extremely metal-poor stars.
The Gaia data release 1 \citep{GaiaDR1} shows that this star
has a close-by companion at a distance of 0\farcs{01} and 0.5 mag
fainter in $G$. Certainly the light of this companion contaminates our
spectrum, but since we do not have any colour information at this time we
did not
attempt to correct our spectrum for the veiling.
We do not detect any sign of the spectrum of the companion star in our spectrum.
\smallskip
\item SDSS\,J1349+1407. \citet{sbordone12} announced it as a Mg-rich star.
The quality of the spectrum available at that time was sufficient only for
giving abundances of a few elements.
The star was independently recovered as a Mg-rich star by
\citet{li14}, who analysed the SDSS\,DR9 spectra.
Their temperature for the star is higher by 230\,K and consequently their
metallicity is higher ([Fe/H]=--2.83).
In the UVES spectrum of SDSS\,J1349+1407
we identified six \ion{Mg}{i} lines in the blue spectrum (382.9, 383.2, 383.8,
405.7, 416.7, 470.2\,nm).
We fitted the line profiles and
we derived ${\rm A(Mg)}=5.30\pm 0.16$ and by removing the line which is
in the wing of a Balmer line,
we derive ${\rm A(Mg)}=5.36\pm 0.11$.
The star is also enhanced in Na, with [Na/Fe]=+0.86.
\smallskip
\item SDSS\,J1442--0015.
We compared our spectrum with the X-Shooter spectrum of \citet{gto12}.
The metallicities from the analysis of the two spectra
are in good agreement within errors of less than $1 \sigma $. The Mg abundances
are in reasonable agreement with $1.5 \sigma$ errors.
The situation is slightly worse for the Ca abundance; the
abundances derived from the \ion{Ca}{i}
422.67\,nm resonance line from the two spectra are consistent to within
$1.8 \sigma$, i.e. 0.7\,dex. As usual, this is discrepant with the \ion{Ca}{ii}
IR
triplet lines, measured in the X-Shooter spectrum.
This example is a recommendation not to overinterpret the abundances
that rely on a single, weak line.
It is important to note that both here and in the study of \citet{gto12}
we adopted the effective temperature derived by fitting the wings of H$\alpha$,
which
is considerably lower than the temperature implied by the $g-z$ colour (6161\,
K).
\smallskip
\item SDSS\,J1507+0051. The X-Shooter spectrum was analysed by \citet{topos1}.
Both the metallicity and the Mg abundance of the two analyses
are in agreement within less than $1 \sigma$. Instead, we have a strong discrepancy
for the abundances of Ca derived from \ion{Ca}{ii} lines.
In the UVES spectrum we detect the 370.6024\,nm line, while in the X-Shooter
spectrum, we relied on the IR triplet lines. This discrepancy needs to be
further investigated.
\end{itemize}
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[draft = false, clip=true]
{pl_li_feh.eps}}
\end{center}
\caption[]{Lithium abundance in unevolved extremely metal-poor stars.
The different symbols refer to different carbon abundances.
The filled hexagons refer to carbon normal stars.
CEMP stars of the low- and high-carbon bands are shown as star symbols and
crossed squares, respectively.
Measurements and upper limits of the programme stars are shown in red.
Measurements and upper limits from our group's previous papers \citep{topos2,topos3}
are shown in blue.
Black symbols are stars for which metallicity, lithium abundance,
and carbon abundance are taken from the literature
\citep{norris97,Lucatello03,
sivarani04,Ivans05,sivarani,frebel07,frebel08,thompson08,aoki08,
sbordone10,behara,leostaraa,
Carollo12,masseron12,aoki13,
Ito13,Carollo13,spite13,Roederer14,Aoki15,sdss_uves,Li2015,terese,topos3,
Placco16,Matsuno17}.
The two components of the binary system CS 22876-32 \citep{jonay08} are shown
as black crosses.
The green dashed line is the level of the {\em} Spite plateau
as determined by \citet{sbordone10}.
\label{plot_Li}
}
\end{figure}
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[draft = false, clip=true]
{pl_teff_li.eps}}
\end{center}
\caption[]{ Lithium abundance in the unevolved extremely metal-poor stars
as a function of effective temperature. The meaning of the symbols
is the same as in Fig.\,\ref{plot_Li}. The high-carbon band CEMP stars have
not been included in this plot.
\label{plot_Li_teff}
}
\end{figure}
\begin{figure}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[draft = false, clip=true]
{plot_lim4.eps}}
\end{center}
\caption[]{ Lithium doublet in SDSS\,J1034+0701 (squares) and
SDSS\,J1247--0341 (triangles) compared to a 3D NLTE theoretical profile
($T\rm_{eff}$ = 6270\,K \ensuremath{\log g} = 4.0 [M/H]=--3.0, A(Li)=2.1). The observed spectra
have been smoothed with a Gaussian of
4 $\mathrm{km\, s^{-1}}$\ FWHM to increase the S/N.
\label{plot_lim4}
}
\end{figure}
\section{Results and discussion}
The main result of this investigation is the confirmation, based on
higher resolution spectra, of the very low metallicities that we derived
for these stars from the analysis of the X-Shooter spectra. Two stars have
[Fe/H] below $-4.3$,
three stars around $-4.0$, and two stars around $-3.5$. These numbers confirm
the high efficiency of the TOPoS strategy for target selection.
The five stars with [Fe/H]$\le -4.0$ discussed in this paper,
SDSS\,J1742+2531 ( [Fe/H]=$-4.80$, \citealt{topos2} ) and
SDSS\,J0929+0238 ([Fe/H]=$-4.97$, \citealt{topos3})
are the most iron-poor stars we found in the course of the TOPoS project
and they are all strongly C-enhanced. To date, among the stars with
[Fe/H]$\le -4.5$ the only `non C-enhanced star' found is SDSS\,J1029+1729
\citep{leostar,leostaraa}.
\subsection{Carbon abundances}
It is interesting to note that all the C-enhanced stars that we have found
belong
indeed to the low-carbon band discussed by \citet{topos2}, as illustrated
in Fig.\,\ref{plot_c}.
These stars do not seem to be enhanced in $s$-process elements and we suggest
that they are indeed
CEMP-no stars. This view is supported also by the
recent study of \citet{Hansen16} who analysed a sample of 27 metal-poor stars
and found that 20 of them
are CEMP,
3 of which are CEMP-no
stars that belong to the low-carbon
band.
We suggest here that a useful classification of metal-poor stars
can be made using only their C abundance without any reference
to their abundance of n-capture elements.
This is related to the fact that, for unevolved stars, it is very difficult
to secure data quality high enough
to derive measurements or significant upper limits for the heavy elements.
Our proposed classification scheme is as follows:
\begin{itemize}
\item `carbon normal': for [Fe/H]$\ge -4$ [C/Fe]$< 1.0$,
for [Fe/H]$< -4$ A(C)$<5.5$;
\item low-carbon band CEMP stars: stars that do not fulfil the carbon normal
criterion and have A(C)$\le 7.6$;
\item high-carbon band CEMP stars: stars that do not fulfil the carbon normal
criterion and have A(C)$ > 7.6$.
\end{itemize}
This classification is qualitatively similar to that proposed by \citet{Yoon},
except that their Group II is partly included in our
low-carbon band and mostly in our carbon normal stars, their
Group I is by and large coincident to our high-carbon band, except for the
stars with the lowest C abundances in their Group I, which we assign to the
low-carbon
band.
In \citet{topos2} we also used the working hypothesis that all the stars
on the high-carbon band
are binaries and the high carbon abundance is the consequence of mass-transfer
from a companion
during its AGB phase. This hypothesis, however, does not place any lower
limit to the [Fe/H]
of the high-carbon band stars. The fact that, observationally, we do not
find
any high-carbon band stars below [Fe/H]=--3.6 is intriguing. In Fig.\,\ref{plot_c}
we have 26 stars with a metallicity below --3.6, of which 25 are C-enhanced,
and none is a high-carbon
band star. It therefore seems unlikely to admit that this is just the result
of small number statistics.
The binary systems that can give rise to a high-carbon band star, in our
hypothesis, are severely constrained regarding the masses of the components.
In particular, the primary star should be larger than about $0.9 M_{\odot}$
to experience efficient third-dredge-up episodes during the AGB phase
\citep{stancliffe08, karakas10}, and consequently enrich its surface with
the carbon produced in the stellar interior.
The secondary should be less massive than about $0.8-0.85 M_{\odot}$, because
higher-mass stars would become white dwarfs
on a timescale shorter than $10-12$ billion years, which is the approximate
age of the Galactic halo. Finally, mass transfer
in binary systems with AGB donors is efficient in a limited range of orbital
periods between approximately $10^2$ and $10^5$ days \cite[e.g.][]{Izzard,
Abate}.
\citet{CL2008} computed the yields of AGB stars of extremely low and zero
metallicity,
and they found that the [C/Fe] ratios of their yields increase
monotonically with decreasing [Fe/H] (as illustrated in their Figure 5).
The carbon abundance in the ejecta
of the AGB star varies within a factor of four \relax in the metallicity
range $-6.5\le$[Fe/H]$\le -3.0$
\citep{CL2008}. Therefore, theoretically, there is no indication that
the nucleosynthetic yields of AGB stars vary in metallicity in such
a way as to disfavour the formation of high-carbon band stars.
A recent review on the nucleosynthesis of AGB stars is given
by \citet{Karakas}.
\begin{figure*}
\begin{center}
\resizebox{\hsize}{!}{\includegraphics[draft = false, clip=true]
{plalpha_paper.eps}}
\end{center}
\caption[]{Ratio of $\alpha$ elements to iron
for the programme stars (red stars) compared to those measured
by our group in other extremely metal-poor dwarf stars \citep{bonifacio09,gto11,sdss_uves,gto12,topos1}.
Black squares are measurements from X-Shooter spectra, while black hexagons
are
measurements from UVES spectra. In the panel giving the [Ca/Fe] ratio, the
blue star symbols
are \ion{Ca}{ii} measures. We restricted the x-axis so as to exclude
SDSS\,J1035+0641; which only has an upper limit on iron and thus a lower
limit on [Ca/Fe] and is not informative.
}
\label{plot_alpha}
\end{figure*}
A possible explanation of the lack of high-carbon band stars at low [Fe/H]
is that at these
metallicities, such systems either do not form at all or are exceedingly
rare.
The existence of a rather sharp cut-off suggests the existence of a physical
mechanism that prevents the formation of these systems.
Because the enriched material synthesised by the AGB star needs to be expelled
in the stellar wind
and accreted by a low-mass companion, the lack of high-carbon band stars
could indicate that at low [Fe/H]
binary mass-transfer processes become rarer or less efficient. Although the
binary-star fraction is not
well constrained at low metallicity, there is no evidence that it decreases
\cite[e.g.][]{gao14}.
The data on the binarity among stars with
[Fe/H]$<-3.6$ is scanty, yet the discovery of the binarity
of SDSS\,J0929+0238 \citep{topos3} suggests that it cannot
be much lower than among stars of higher metallicity.
It has been suggested that at low metallicity the ejection of the AGB envelope
becomes increasingly
inefficient \cite[][]{wood11}. If only a small mass is ejected by the primary,
the amount of
material accreted by the secondary may be insufficient to enhance the carbon
abundance up to
a typical high-carbon band value.
Another effect to take into account is that when a star accretes
material and becomes more massive, it evolves more quickly than a single
star of equal initial mass.
In the simulations of \cite{Abate} the synthetic population
of CEMP-$s$ stars has typically accreted a mass of approximately $0.1 M_{\odot}$.
Stars with
initial mass $0.8 M_{\odot}$ at low metallicity have a lifetime of approximately
the age of the Universe;
however, if they experienced binary mass-transfer they would become high-carbon
band CEMP-$s$
stars of mass $\approx 0.9 M_{\odot}$ and consequently their lifetimes would
be shorter than 12 Gyr.
Hence, they would not be visible any longer. Similarly, stars of initial
mass of $0.7M_{\odot}$
would reach masses of around $0.8 M_{\odot}$ and still be visible after 12
Gyr, but for a shorter
time than single stars of the same initial mass.
Also, the results of \cite{Moe} suggest that large initial mass ratios $q=M_{2,\mathrm{i}}/M_{1,\mathrm{i}}$
could be slightly more likely among low-mass binaries, hence increasing the
proportion of systems with initial
secondary mass $M_{2,\mathrm{i}}>0.8 M_{\odot}$.
The probability of observing a star at a given distance during its evolution
is
proportional to the time its apparent luminosity is above our detection
threshold. If a star is visible in its main sequence phase, it will also
be visible in the RGB phase, but not once it reaches the white dwarf stage,
unless it is very near to us.
A faster evolving star reaches the white dwarf stage more quickly
and is thus less likely to be observed.
The decrease in lifetime of the high-carbon band population, due to their
increase in mass, naturally occurs at every metallicity,
but its consequences become more evident at [Fe/H]$\leq-3.5$ because fewer
stars are known in this range.
The determination
of the binary fraction and the period distribution for different metallicity
bins
down to the lowest metallicities may provide a fundamental insight into the
properties of star formation at different metallicities.
\subsection{Lithium abundances}
In Fig. \ref{plot_Li} we show the Li abundances as a function of [Fe/H]
for our programme stars and for stars taken from the literature.
We classify the stars in three groups: low-carbon band CEMP stars (star symbols);
high-carbon band CEMP stars (crossed squares); and carbon normal stars (filled
hexagons), which includes all stars with [C/Fe]$< 1.0$ or for which
the C abundance has not been measured.
From this figure it appears that the Li abundances in the low-carbon band
CEMP stars have a behaviour
that is indistinguishable from that of the carbon normal stars.
On the other hand the high-carbon band CEMP stars are preferentially Li-depleted,
although there are a few measurements on the Spite plateau.
This is consistent with our hypothesis that the high-carbon
CEMP stars are the result of mass transfer from an AGB companion.
In most cases Li is depleted as a result of processing in the H-burning AGB
shell,
although in some cases it may also be produced through
the Cameron--Fowler mechanism \citep{CF71}.
Mass transfer from an AGB companion has been invoked
to explain the occurrence of Li-rich main sequence stars
in globular clusters \citep{koch11,monaco12}, although pollution
from nearby RGB stars having undergone the Li-flash is
an alternative explanation \citep{Pasquini14}.
The point is that transfer from an AGB always results
in an enhancement of carbon, but Li may be either depleted
or enhanced.
The stars SDSS\,J1034+0701 and SDSS\,J1247--0341 both have
[Fe/H]\,$\sim -4.0 $. The measurements of the equivalent width
of the \ion{Li}{i} resonance doublet are very uncertain. In Fig. \ref{plot_lim4}
we show the observed spectra compared to a 3D NLTE synthetic spectrum.
We stress that this is not a fit; we just plot the synthetic spectrum to
provide guidance to the position and expected strength of the
\ion{Li}{i} resonance doublet. Although
the measurement is uncertain, we believe that the detection of Li is robust.
Taken at face value, our measurements place both stars
on the Spite plateau. All the other stars have either an upper limit
or a measurement that places them well below the Spite
plateau. While this behaviour, also referred to as `meltdown' of the Spite
plateau
(\citealt{sbordone10}, but see also \citealt{bonifacio07,aoki09}), is well
known,
these two stars on the Spite plateau appear exceptional.
In fact, recently \citet{Matsuno17} pointed out that
`no star in the literature
has comparable Li abundance to the Spite plateau below [Fe/H] of--3.5, except
for the primary of the
double-lined binary system CS\,22876--032.'
The two components of CS\,22876--032 are shown as $\times$ symbols in
Figs.\,\ref{plot_Li} and \ref{plot_Li_teff}.
While SDSS\,J1034+0701 and SDSS\,J1247--0341 have
an effective temperature well above 6000\,K, as does
CS\,22876-032\,A \citep{jonay08}, the star SDSS\,J0140+2344 has an effective
temperature below 6000\,K;
it has roughly the same metallicity ([Fe/H]=--4.0) and,
like CS\,22876-032\,B, it has a Li abundance well below the Spite plateau.
Therefore, our observations call into question the claim by
\citet{Matsuno17}, that Li abundances in the interval
$\rm -4.5 \le [Fe/H] \le -3.5$ are almost constant at a value
lower than the Spite plateau.
For stars with $T\rm_{eff}$ $> 6000$\,K,
there are some stars on the Spite plateau,
at least down to [Fe/H]=--4.0. On the
other hand, among the stars with $T\rm_{eff}$ $\le $6000\,K and
[Fe/H]$\le -3.5$, lithium is always below
the Spite plateau. This can be seen in Fig.\,\ref{plot_Li_teff},
where the only star at effective temperature below 6000\,K
on the Spite plateau is SDSS\,J0907+0246 ([Fe/H]=--3.44).
The current observations suggest that
the $T\rm_{eff}$\ for which lithium destruction in the stellar
atmospheres becomes important, increases as metallicity decreases,
although it is unlikely that this is the only effect that drives Li destruction
since there are several
stars hotter than 6000\,K that are found well below the Spite plateau.
\citet{BM97} argued that for $T\rm_{eff}$ $\ge $5700\,K Li destruction
was negligible; however, their sample had only three stars
with metallicity below --3.0. From Fig.\,\ref{plot_Li_teff} it is clear
that with increasing $T\rm_{eff}$\, there are fewer stars below the Spite plateau,
and, at any rate, the average Li abundance is higher.
Our detection of Li in SDSS\,J1035+0641 is to date the measurement of Li
in the most
iron-poor unevolved star. It poses a strong constraint on any theory that
aims
to explain the cosmological Li problem. Any theory involving stellar destruction
of Li must accommodate the fact that in SDSS\,J1035+0641 Li is measurable.
In other stars of comparable iron abundance, it is either completely
destroyed or, at least, more severely depleted.
Alternatively, if one tries to explain the Li abundance of this star
by a prompt enrichment in Li, one has to explain why such a prompt enrichment
has not taken place in the other stars with [Fe/H]$\le -5.0$.
\subsection{Abundances of other elements}
One of the main goals of the UVES observations was to increase the chemical
inventory
of the stars at very low metallicity.
The data presented in Table\,\ref{abbo2} fulfil this goal,
although the low metallicities of the stars imply that we have many
upper limits. We have two measurements
of Co that confirm that Co is enhanced over Fe in EMP stars, as already shown
by \citet{cayrel04} and \citet{bonifacio09}. In the two cases where Sr is
measured,
it is enhanced over Fe; however, this is certainly a selection effect since
for a given
effective temperature, the higher the abundance, the stronger the line
and thus the
easier it is to detect it.
In Fig.\,\ref{plot_alpha} we provide the ratios of the measured $\alpha$
elements, Mg, Si, and Ca, to Fe.
In the upper panel, SDSS\,J1349+1407 clearly stands out as being extremely
enhanced in Mg, although its Ca abundance seems quite normal. There is another
star in our sample, SDSS\,J1050+2421 \citep{topos1}, that is highly enhanced
in Mg,
although not as much as SDSS\,J1349+1407.
We note that there is a discrepancy of about 1\,dex in the Ca abundance
of SDSS\,J1050+2421 derived from the \ion{Ca}{i} resonance line
or from the \ion{Ca}{ii} lines. This discrepancy
may be partly due to an incorrect gravity, since a
discrepancy of about 0.5\, dex is found between the iron abundances derived
from the \ion{Fe}{i}
and \ion{Fe}{ii} lines \citep{topos1}.
Based on
its proper motion, it has been pointed out that this star
is most probably a main sequence star with a \ensuremath{\log g}\ $\sim$ 4.7, rather
than 4.0,
as was assumed in \citet{topos1} (M. Bessell, private communication).
The [\ion{Ca}{i}/\ion{Fe}{i}]
and the [\ion{Ca}{ii}/\ion{Fe}{ii}] ratios differ by `only' about 0.6\,dex.
Thus, a change in gravity may alleviate the problem, but not solve it.
In Fig.\,\ref{plot_alpha} we plot the value
derived from the \ion{Ca}{i} resonance line. Had we chosen the value derived
from \ion{Ca}{ii}, SDSS\,J1050+2421 would be highly enhanced in Ca as well.
The abundances of SDSS\,J1349+1407 resemble very closely those
of CS\,22949-037 \citep{depagne,norris02}; in fact, they have almost the
same [Mg/Ca] ($\sim 1.2$).
Among the EMP stars, a [Mg/Ca] almost as high is
found for HE\,1327-2326 (\citealt{frebel08} [Mg/Ca]=0.99) and an even higher
value is found
in SMSS\,J0313-6708 (\citealt{Keller14}) [Mg/Ca]=2.95).
Both CS\,22949-037 and HE\,1327-2326 are enhanced in Na: [Na/Fe]=+1.57 \citep{Andrievski}
and [Na/Fe]=+0.94 \citep{frebel08}.
Star HE\,1012-1540, analysed at high resolution by the 0Z project
\citep{Cohen13}, also seems
to belong to this group ([Mg/Ca]=+0.83, [Na/Fe]=+1.02).
Quite intriguingly, a star with similar properties has been found
in the Hercules ultra faint dwarf galaxy \citep{Koch08}.
For Her-2, with [Mg/Ca]=+0.93 and [Na/Fe]=+0.78, \citet{Koch08}
argued that this may be understood in terms of enrichment from
very massive stars with masses of the order of 30 M\sun.
The other feature that stands out in Fig.\ref{plot_alpha}
is that we find additional stars with low $\alpha$-to-iron ratios,
for example SDSS\,J1442--0015 and SDSS\,J1247--0341.
This confirms our earlier claim \citep{gto12}, based on X-Shooter
spectra, regarding the existence of a population of extremely metal-poor
$\alpha$-poor stars.
The presence of metal-poor stars with low $\alpha$-to-iron ratios,
in some cases quite extreme \citep{Ivans03}, is also supported
by other investigations \citep{Cohen13,Susmitha16}. Recently
\citet{hayes17} also recognised the presence of an $\alpha$-poor population
of stars belonging to the Galactic Halo,
albeit at a higher metallicity regime than the one considered in this paper.
The presence of these $\alpha$-poor stars has so-far
received little theoretical attention.
While \citet{Ivans03} has invoked a prompt enrichment by
type Ia supernovae, \citet{Cohen13} has suggested
that the diversity in nucleosynthesis products of
extremely metal-poor type II supernovae is greater than
that predicted by current models.
These stars could have formed in low-mass dwarf galaxies and
subsequently accreted to the Milky Way Halo, carrying memory of
a different chemical evolution, as has been claimed
by \citet{hayes17}.
\section{Conclusions}
We have provided detailed abundances for a sample of seven EMP turn-off stars,
based on high-resolution spectra.
All the stars for which carbon could be measured are CEMP stars
and belong to the low-carbon band.
For the most iron-poor star,
SDSS\,J1035+0641, we are unable to detect any iron lines, but we
were able to place a stringent upper limit at [Fe/H]$\le -5.2$.
It is remarkable that we measure Li in this star, albeit
well below the Spite plateau. This is the first measurement of Li in
an unevolved star with [Fe/H]$\le -5$, and it places a strong constraint on any theory
to explain the cosmological Li problem.
Three of the programme stars have [Fe/H]$\sim -4.0$, and lithium
has been measured in all three. Quite interestingly, the Li abundance of
the two warmer stars
places them squarely on the Spite plateau, while the cooler one lingers below.
This situation is reminiscent of what is observed in the binary system
CS\,22876-32 \citep{jonay08}. We have argued that this suggests that
the effective temperature at which Li depletion begins increases
with decreasing [Fe/H].
We confirmed that SDSS\,J1349+1407 is extremely enhanced in Mg but not
in Ca, and in fact its abundance pattern resembles that of CS\,22949-037
and
its Mg/Ca is also similar to that of HE\,1327-2326.
This may suggest that there is a class of Mg-rich stars among EMP stars.
Our high-resolution observations have confirmed the existence of stars
with low $\alpha$-to-iron ratios, as pointed out by \citet{gto12}, based
on medium-resolution X-Shooter spectra.
The TOPoS project has been very successful in increasing the number of stars
with detailed abundances at and below a [Fe/H] of --4.0, yet the numbers
are still
smaller than desirable for making robust statistical inferences.
Hopefully the Pristine survey \citep{Pristine1,Pristine2,Pristine3}
and its massive follow-up with WEAVE \citep{WEAVE} will soon boost the numbers,
and the LAMOST survey as well \citep{Li2015a,Li2015}.
\begin{table*}
\caption{Stellar parameters and abundances of C and Li}
\label{param}
\begin{center}
\begin{tabular}{lclrrcccll}
\hline \noalign{\smallskip}
Star & $T\rm_{eff}$\ & \ensuremath{\log g}\ & $\xi$ & [Fe/H] & A(C) & A(C) & A(Ca) & A(Li) & EW(Li)\\
& K & [C's] & Km/s & & 1D & 3D & \ion{Ca}{ii}-K & & pm \\
\hline \noalign{\smallskip}
SDSS\,J0140+2344 & 5848 & 4.0 & 1.5 & \phantom{$<$}$-4.00\pm 0.14$ & 6.33 & 6.11 & & \phantom{$<$}1.86 &\phantom{$<$}2.2 \\
SDSS\,J1034+0701 & 6224 & 4.0 & 1.5 & \phantom{$<$}$-4.01\pm 0.14$ & 6.28 & 5.96 & & \phantom{$<$}2.24:&\phantom{$<$}2.8: \\
SDSS\,J1035+0641 & 6262 & 4.0 & 1.5 & $<-5.20$ & 7.16 & 6.95 & 1.5 & \phantom{$<$}1.90 &\phantom{$<$}1.3\\
SDSS\,J1247--0341 & 6332 & 4.0 & 1.5 & \phantom{$<$}$-4.01\pm 0.32$ & 6.64 & 6.35 & & \phantom{$<$}2.20:&\phantom{$<$}2.2:\\
SDSS\,J1349+1407 & 6112 & 4.0 & 1.5 & \phantom{$<$}$-3.60\pm 0.12$ & 7.00 & 6.82 & & $< 0.9$ & $< 0.3$ \\
SDSS\,J1442--0015 & 5850 & 4.0 & 1.5 & \phantom{$<$}$-4.37\pm 0.16$ & 6.02 & 5.71 & & $<1.58$ &$<1.2$ \\
SDSS\,J1507+0051 & 6555 & 4.0 & 1.5 & \phantom{$<$}$-3.41\pm 0.20$ & & & $<2.19$ &$<1.6$ \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{table*}
\caption{Abundances relative to hydrogen [X/H] with solar abundances from \citet{lodders09}
for all elements except Fe and C whose solar abundances are from \citet{abbosun}.}
\label{abbo2}
\begin{center}
\begin{tabular}{llllllllll}
\hline
\noalign{\smallskip}
El & Sun & J0140+2344 & J1034+0701 & J1035+0641 & J1247--0341 & J1349+1407 & J1442--0015 & J1507+0051 \\
\hline \noalign{\smallskip}
\ion{C}{i} & 8.50 & $-2.17$ & $-2.22$ & \phantom{$<$}$-1.33$ & $-1.86$ & $-1.50$ & $-2.48$ & \\
\ion{Na}{i} & 6.30 & $-4.22$ & & & & $-2.74\pm 0.12$ & & \\
\ion{Mg}{i} & 7.54 & $-3.37\pm 0.02$& $-3.28\pm 0.10$& & $-3.77\pm 0.23$& $-2.18\pm 0.11$ & $-4.19\pm 0.05$& $-3.30\pm 0.13$ \\
\ion{Al}{i} & 6.47 & $-4.29$ & $-4.27$ & & $<-4.23$ & $-3.97$ & $<-4.68$ & $<-4.03$ \\
\ion{Si}{i} & 7.52 & $-3.91$ & $-4.12$ & & $-4.27$ & & $-4.53$ & $-3.67$ \\
\ion{Ca}{i} & 6.33 & $-2.79$ & $-4.17$ & & & $-3.42\pm 0.16$ & $-4.48$ & $-3.35$ \\
\ion{Ca}{ii}& 6.33 & $-3.66$ & & \phantom{$<$}$-4.83$ & $-4.07$ & & & $-3.05$ \\
\ion{Sc}{ii}& 3.10 & & $-3.38$ & & & $-3.54$ & & \\
\ion{Ti}{i} & 4.90 & & & & & & $-2.33$ & $-3.04\pm 0.12$ \\
\ion{Ti}{ii}& 4.90 & $-3.75\pm 0.14$& $-3.84\pm 0.04$& & & $-3.22\pm 015$ & $-2.44$ & \\
\ion{Cr}{i} & 5.64 & $-4.17\pm 0.12$& & & & $-3.73$ & & \\
\ion{Mn}{i} & 5.64 & & & & & $-3.87\pm 0.03$ & & \\
\ion{Fe}{i} & 7.52 & $-4.00\pm 0.14$& $-4.01\pm 0.14$& $<-5.20$ & $-4.01\pm 0.32$& $-3.60\pm 0.12$ & $-4.37\pm 0.16$& $-3.41\pm 0.20$ \\
\ion{Fe}{ii}& 7.52 & & & & & $-3.58\pm 0.16$ & & $-3.56$ \\
\ion{Co}{i} & 4.92 & $-3.13\pm 0.34$& & & & $-2.87$ & & \\
\ion{Ni}{i} & 6.23 & $-3.86\pm 0.12$& $-3.99$ & & & $-3.43$ & & \\
\ion{Sr}{ii}& 2.92 & $-3.43\pm 0.06$& & & & & & $-3.16\pm 0.02$ \\
\noalign{\smallskip}
\hline
\end{tabular}
\end{center}
\end{table*}
\begin{acknowledgements}
We are grateful to our referee Mike Bessell for his very thorough and useful
report of our paper.
The project is supported by fondation MERAC.
N.C., H.G.L., and A.J.G. were supported by Sonderforschungsbereich SFB 881
``The Milky Way System'' (subproject A4) of the German Research
Foundation (DFG).
R.S.K. acknowledges support from the European Research Council via the Advanced
Grant
``STARLIGHT: Formation of the First Stars'' (project number 339177).
C.A. acknowledges funding from the Alexander von Humboldt Foundation
\end{acknowledgements}
\bibliographystyle{aa}
|
\section*{Acknowledgment}
This work has been supported by the Federal Ministry of Education and Research of the Federal Republic of Germany (Foerderkennzeichen 16KIS0267, HiFlecs). The authors alone are responsible for the content of the paper.
\section{Introduction}
The ability to efficiently support different traffic types with different requirements is a crucial one for today's wireless systems and is expected to gain even more importance in the future, along with but not limited to the success of the Internet of Things (IoT) and the proceeding of industrial communication technologies. One specific example of a system, which would benefit of supporting different traffic types simultaneously and which we would like to address, would be a wireless communication system in an industrial environment. In today's production sites, wireless communication systems are serving either human purposes (in the following referred to as ``best-effort traffic'') or non-human purposes (in the following referred to as ``machine-type traffic'') exclusively, because a flexible support of both traffic types is not possible, which leads to a waste of spectrum.
In accordance with what intuition would tell us, analysis of machine-to-machine (M2M) traffic reveals that it has got specific features which hold true for most cases \cite{Nikaein2013}. In table \ref{tab:traffic}, the different message characteristics for best-effort and for machine-type traffic are shown.
\begin{table
\caption{Characteristics of best-effort and machine-type traffic}
\centering
\begin{tabular}{l|l|l}
\textbf{Parameter} & \textbf{best-effort traffic} & \textbf{machine-type traffic}\\
\hline
packet length & long & short\\
number of packets & low & high\\
duty cycle & high & low\\
traffic direction & downlink- & uplink-\\
& dominant & dominant\\
latency constraints & medium & medium up to\\
& & real-time\\
traffic generation & human triggered & periodic or\\
& & event-driven\\
Bit Error Rate & medium & high up to\\
requirements & & ultra-reliable
\end{tabular}
\label{tab:traffic}
\end{table}
Our idea, that we would like to present in the following sections, is to use a CDMA-based MAC which possesses two subsets of spreading sequences, as depicted in figure \ref{fig:codes}, which can be assigned to a corresponding traffic type. The first subset of sequences is consisting of sequences which are mutually orthogonal to every other sequence in this subset and in ever other subset of sequences. The second subset of sequences is consisting of sequences which are mutually orthogonal to every sequence in the first subset but interfere with sequences in the second subset of sequences. Intuition suggests to assign sequences of the first group of sequences to traffic with high bit error rate requirements, i.e. machine-type traffic, while assigning sequences of the second group of sequences to traffic with lower reliability requirements, i.e. best-effort traffic.
\begin{figure
\centering
\includegraphics[width=0.8\columnwidth]{figures/Codes.pdf}
\caption{Partition of spreading sequences}
\label{fig:codes}
\end{figure}
In a second step, we would like to combine this idea with a variable spreading factor to gain additional flexibility. Assuming a constant frame length, we then have a total of three degrees of freedom: the FEC coding rate, the spreading factor and the amount of overloading in the CDMA system.
The rest of this paper is organized as follows: in the following section, an overview of related work is given. In section III, the generation of our proposed spreading sequences is explained. Section IV explains the support of variable spreading factors within the proposed CDMA system. In section V, numerical evaluations using computer simulations are performed. A conclusion is drawn in section VI.
\section{Related Work}
The CDMA MAC we are proposing in this paper is related to different topics already known in the literature which we would like to shortly mention in the following.
\subsubsection{Hadamard Matrices}
The spreading sequences used in this paper are based on a well-known family of orthogonal spreading sequences called Hadamard sequences \cite{Sylvester1867}\cite{Hadamard1893}. A family $F=\{f_i(t)|i=1,2...N\}$ of $N$ binary sequences is called orthogonal, if any two sequences $x(v)$ and $y(v)$ of length $N=2^n$ are mutually orthogonal, i.e. their synchronous cross-correlation
\begin{equation}
C_{x,y} = \sum_{v=0}^{N-1}{x(v)\cdot y(v)},
\end{equation}
is zero.
A matrix $H$ consisting of $N$ different sequences of length $N$, forming an $N \times N$ matrix is called a Hadamard matrix, if the following equation applies:
\begin{equation}
H \cdot H^T=H^T \cdot H=N \cdot I
\end{equation}
with $I$ being the identity matrix and $H^T$ being the transposed matrix $H$.
\subsubsection{Orthogonal Variable Spreading Factor}
To support multi-rate operation within a CDMA system, the orthogonal variable spreading factor (OVSF) code has been introduced in the literature \cite{Adachi1997}. The idea is to use Hadamard codes of different lengths and to organize them in a tree structure according to figure \ref{fig:ovsf}.
The generation of the sequences differs slightly from Sylvester's method \cite{Sylvester1867} and is defined recursively leading to the following sequence of matrices:
\begin{equation}
H_1 = \begin{pmatrix}[r]
1
\end{pmatrix},
H_{2^{n}} = \begin{pmatrix}[l]
H_{2^{n-1}}\otimes(1,1)\\
H_{2^{n-1}}\otimes(1,-1)
\end{pmatrix},
\end{equation}
with $\otimes$ being the Kronecker product. As we can easily check using equation 2, the generated matrices are Hadamard matrices.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/ovsf.pdf}
\caption{Orthogonal Variable Spreading Factor (OVSF) code tree}
\label{fig:ovsf}
\end{figure}
As a consequence, codes of the same layer are orthogonal. Additionally, any two codes of different layers are also orthogonal as long as one of them is not a mother code of the other.
\subsubsection{Overloaded CDMA}
As we could see, an orthogonal set of Hadamard codes of length $N$ consists of $N$ different sequences which are mutually orthogonal. If we are in need of more than $N$ different spreading sequences and are not intending to increase the code length, i.e. the spreading factor (SF), we have to add sequences which are not mutually orthogonal to every other sequence of the Hadamard set and thus bring additional interference with them. This technique is called overloaded CDMA and is well known in the literature \cite{Hosseini2011}.
Following the notation of \cite{Yang2000} and \cite{Amadei2002}, the maximum cross-correlation between an additional sequence $f$ of length $N$ and every sequence $h_i$ in the Hadamard code set $H$ is defined as
\begin{equation}
C_f=\displaystyle\max_{i} C_{f, (h_i+i_0)}
\end{equation}
where $i_0$ is the all-zero or the all-one sequence of length $N$. Furthermore, the authors of \cite{Yang2000} and \cite{Amadei2002} also stated the minimum achievable correlation value for all sequences f, $C_{min}(N)=\displaystyle\min_{f}C_f$ which is given by
\begin{equation}
C_{min}(N)=2^{n/2}=\sqrt{N}
\end{equation}
for any even integer n.
Additional sequences fulfilling equation (5) and thus minimizing additional interference are referred to as ``Quasi-orthogonal sequences'' in the literature. In \cite{Yang2000}, quasi-orthogonal sequences for the use in CDMA systems are described. The authors of \cite{Amadei2002} combine this approach with a variable spreading factor and use these sequences in a multi-carrier CDMA system supporting multiple classes of users with respect to different data rates only. To the best of the authors knowledge, a CDMA system using only partly overloaded spreading sequences and supporting multiple QoS levels is not yet published in the literature.
\subsubsection{Trade-Off between CDMA and FEC}
Improving signal robustness by adding additional information can either be done by applying channel coding or by increasing the spreading factor. This trade-off is investigated for a multi-carrier CDMA system in\cite{Kaiser1996}.
\section{Requirements and System Parameters}
As already mentioned, wireless industrial communication systems are expected to provide at least comparable performance to wire-line systems. In the following, we aim to shortly depict the requirements and system parameters of a wireless industrial communication system for the use in a closed-loop control application.
\subsection{Reliability and Latency}
A target packet error rate of $10^{-9}$ is assumed as ultra-reliable, for example in \cite{Osman.2015}. This is, of course, only possible if adequate channel coding, such as forward error correction (FEC), is applied. For end-to-end latency a maximum delay of $1$ ms is required to serve fast control processes (up to $1$ kHz update frequency) in real time. This allows a delay budget of $500$ $\mu$s for one UL and DL transmission respectively yielding to a maximum transmission time of approximately $100$ $\mu$s \cite{Johansson.}. Regarding both constraints, it seems impossible to make use of any retransmission schemes for real-time transmissions. We expect, that closed-loop wireless industrial communication systems will not only consist of transmissions with real-time requirements, but will also have transmissions of best-effort quality. This best-effort traffic on the other hand could use a retransmission scheme within its channel in the code domain. In table \ref{tab:requirements}, the requirements of a closed-loop wireless industrial communication system for real-time as well as best-effort transmissions is depicted.
\begin{table}[h]
\caption{Requirements of Closed-Loop-Control Wireless Industrial Communication Systems}
\begin{tabular}{l|l|l}
\textbf{Parameter} & \textbf{Real-time} & \textbf{Best-Effort}\\
\hline
Packet Error Rate (PER) & $10^{-9}$ & application dependent\\
Latency [ms] & 0.5-10 & 10 - 50\\
Maximum number of users & 128 & 128\\
Payload size [bit] & 48 - 1024 & 120 - 2048\\
Coverage [m] & several 10 & several 10\\
Automatic Repeat Request (ARQ) & No & Yes\\
\end{tabular}
\label{tab:requirements}
\end{table}
\subsection{Topology of a Closed-Loop Wireless Communication System}
\label{fig:LLWCS}
For the general system setup one central node is considered which is referred to as the access point (AP) in the following and carries out the network management of the cell. The users, which are transceiver nodes connected to sensors and/or actuators, share each a direct link with the AP respectively yielding a starlike topology according to Fig. \ref{fig:system_arch}.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/arch.pdf}
\caption{Topology of a Low Latency Wireless Communication System}
\label{fig:system_arch}
\end{figure}
Due to the tight delay requirements in wireless industrial communication systems, other topologies such as meshed networks are not feasible. Users receive or transmit application data from sensors or actuators via their dedicated links to the AP which may be connected to backend devices such as conventional industrial fieldbus controllers or application servers via industrial ethernets.
\section{Partly Overloaded Spreading Sequences}
For our approach, we generate the spreading sequences as follows. First of all we start with the generation of spreading sequences with $SF=2^n$ according to equation (3), which leads to a Hadamard matrix. On the left-hand side of figure \ref{fig:QOSS} the corresponding Hadamard matrix of $SF=8$ is shown as an example. We are considering every row of this matrix as being one spreading sequence. As it is known in the literature \cite{Koukouvinos2008}, we are allowed to interchange columns (as well as rows) of a Hadamard matrix which will lead to another Hadamard matrix of $SF=8$. Since, as we can observe, neighboring columns of the above half of our Hadamard matrix are equal, this subset of spreading sequences does not change, when we interchange these columns. In the below subset however, we create additional spreading sequences by interchanging these columns. Using equation (2), it can be shown, that this new matrix, consisting of the original Hadamard matrix with interchanged columns is also a Hadamard matrix. The new spreading sequences are of course also orthogonal to the unchanged above part of the spreading matrix, but bring additional interference with the original below part of the matrix.
A code set consisting of sequences of length 8 is depicted in figure \ref{fig:QOSS}.
\begin{figure
\centering
\includegraphics[width=\columnwidth]{figures/QOSS.jpg}
\caption{Overloaded Hadamard set}
\label{fig:QOSS}
\end{figure}
\section{Partly Overloaded Spreading Sequences with Variable Spreading Factor}
As we already stated, in a second step, we now would like to combine the proposed partly overloaded spreading sequences with the idea of a variable spreading factor, which leads to a maximum of flexibility. To do so, we arrange the sequences described in the preceding section on a code tree, as it is known with standard OVSF codes (cf. figure \ref{fig:POVSF}). This leads us to a code tree, which has the following properties:
\begin{itemize}
\item All spreading sequences of the upper half of the code tree are mutually orthogonal to every other sequence in the same layer and to every other code in the tree, as long as one sequence is not a mother code of the second sequence.
\item In the lower half of the code tree, overloading is occurring as soon as more than one code per branch and layer is assigned. If not, the code tree is equivalent to a standard OVSF code tree.
\item Depending on our constraints concerning latency, throughput and bit error rate, when can now choose, whether to invest resources in a larger spreading factor $SF$, in a lower code rate or in a more robust modulation order.
\end{itemize}
\begin{figure
\centering
\includegraphics[width=\columnwidth]{figures/povsf.jpg}
\caption{Partly Overloaded Variable Spreading Factor code tree}
\label{fig:POVSF}
\end{figure}
\section{Simulation}
As already stated, we have several parameters which affect the system performance w.r.t. BER, PER and throughput:
\begin{itemize}
\item the spreading factor (SF)
\item the signal to noise ratio (SNR)
\item the number of users ($N_U$), which effects the amount of overloading in the system
\item the sending probability of every user ($P_S$), which is also related to the overloading
\item the FEC code rate
\item the modulation order.
\end{itemize}
In order to enable a numerical evaluation the presented spreading sequences, a simulation has been performed which will be described in the following.
\subsection{Setup}
We simulated a setup with the parameters shown in table \ref{tab:parameters}. Please note, that these parameters can vary in some simulations, so that the values stated in table \ref{tab:parameters} hold true if not otherwise stated.
\begin{table
\caption{Simulation Setup}
\centering
\begin{tabular}{l|l|l}
\textbf{Parameter} & \textbf{Machine-type} & \textbf{Best-Effort}\\
\hline
Number of users & $N_{MT}=4$ & $N_{BE}=6$\\
Number of iterations&\multicolumn{2}{c}{}\\ per simulation&\multicolumn{2}{c}{10.000}\\
Sending probability & \multicolumn{2}{c}{$P_S = 0.5$}\\
Packet length & \multicolumn{2}{c}{128 Bit}\\
Spreading Factor & \multicolumn{2}{c}{$SF=8$}\\
Modulation & \multicolumn{2}{c}{QPSK}\\
FEC code rate & \multicolumn{2}{c}{1}\\
Signal to noise ratio (SNR) & \multicolumn{2}{c}{10 dB}\\
Synchronization & \multicolumn{2}{c}{perfect}\\
Channel Estimation & \multicolumn{2}{c}{perfect}\\
\end{tabular}
\label{tab:parameters}
\end{table}
The signal processing chain consists of the following components: a data source, the digital modulation, the spreading, an AWGN channel, the despreading, the demodulation and a data sink. These components will now be described in detail.
\subsubsection{Data Source}
In our simulation, we use a simple packet source, which is generating packets with a length of 128 Bit and with a probability of $P_S$.
\subsubsection{Digital Modulation}
We use QPSK modulation where not stated otherwise.
\subsubsection{Spreading}
The modulated user data vectors $A_i$ are then spread using the spreading code vectors $C_i$ by applying the Kronecker product, yielding to the spread signal $S_i$:
\begin{equation}
S_i = A_i \otimes C_i.
\end{equation}
\subsubsection{Channel Model}
Since in this paper, we focus on the performance of the proposed spreading sequences, we assume, that channel distortions have been equalized by some preceding processing steps (e.g. when CDMA is used within a MC-CDMA system) and that only AWGN is applied to the CDMA part of the system. The spreaded user signals $S_i$ are however superposed yielding to the signal $S$:
\begin{equation}
S = \sum_{i=1}^{N_U} S_i.
\end{equation}
\subsubsection{Despreading}
The despreading of the spreaded signal $S$ is done by multiplying the repeated code vectors $\hat{C_i} = (C_i, C_i, ..., C_i)$ for $N_B$ times component-wise with $S$, yielding to the despread user signal $\tilde{D_i}$:
\begin{equation}
\tilde{D_i} = \hat{C_i} \odot S,
\end{equation}
with $\odot$ being the Hadamard product.
Finally the despread user signal for the $i$th user $D_i$ is achieved by applying ``integrate and dump'' to $\tilde{D_i}$:
\begin{equation}
D_i = (d_{i,1}, ..., d_{i,N_B})
\end{equation}
with
\begin{equation}
d_{i,j} = \sum_{x=(j-1) \cdot SF + 1}^{j \cdot SF} \tilde{d_{i,x}} / SF.
\end{equation}
\subsubsection{Data Sink}
In the data sink, the Bit Error Rate ($BER$) is calculated. BER over all $N_I$ iterations and all users $N_U$ is calculated as
\begin{equation}
BER = \frac{\sum_{a=1}^{N_I} \sum_{i=1}^{N_U} \sum_{j=1}^{N_B} \|d_{i,j}^a - u_{i,j}^a\|}{N_I \cdot N_U \cdot N_B}
\end{equation}
with $d_{i,j}^a$ being the $j$th received bit of the $a$th packet of user $i$.
\subsection{Results}
Next, some simulation results are presented. To this purpose, the system performance is investigated as a function of the SNR, number of users, and sending probability. In a second step, the trade-off between FEC code rate, spreading factor and modulation order is investigated.
\subsubsection{Variation of the SNR}
In a first simulation, we would like to investigate the system performance for different Signal-to-noise ratios (SNRs). Therefore, the SNR is varied from $0dB$ to $20dB$ in steps of $1dB$. The result of this simulation is depicted in figure \ref{fig:BERvsSNR}.
\begin{figure
\input{figures/BERvsSNR_inkl_Hadamard}
\caption{BER vs. SNR}
\label{fig:BERvsSNR}
\end{figure}
As we can observe, the bit error rate (BER) of best-effort traffic in this scenario ($N_U=10$ and $P_S=0.5$, i.e. overloading occurs) is hardly improving with better SNR, whereas the BER of machine-type traffic improves dramatically. This shows, that orthogonality of machine-type users and their spreading sequences can be maintained while overloading best-effort users. The dashed curve shows the BER performance of a pure Hadamard set with $SF=8$ for the sake of comparison. Since a Hadamard set of spreading sequences with $SF=8$ is able to support up to $N_H=8$ different users, the sending probability $\widehat{P_S}$ has been set to
\begin{equation}
\widehat{P_S} = \frac{N_U}{N_H} \times P_S=\frac{10}{8} \times 0.5 = 0.625
\end{equation} to ensure a realistic comparison. As we can observe, the BER performance of this pure Hadamard set with $SF=8$ is identical to the performance of our proposed sequences for the case of machine-type traffic.
\subsubsection{Variation of the sending probability}
In a second step, we are intending to investigate the impact of the sending probability on the system performance. For this purpose, $P_S$ is varied from $0\%$ to $100\%$ in steps of $10\%$ and the result is depicted in figure \ref{fig:BERvsPS}.
\begin{figure
\input{figures/BERvsPS_inkl_Hadamard}
\caption{BER vs. Sending Probability}
\label{fig:BERvsPS}
\end{figure}
As one would expect, the Bit Error Rate (BER) for both machine-type as well as best-effort traffic worsens with a growing sending probability. Due to its orthogonality, machine-type traffic has a better BER with about two to three orders of magnitude difference.
\subsubsection{Variation of the number of users}
Furthermore, we investigated the influence of the number of users in the system, both for best-effort type users as well as for machine-type traffic users. The results are depicted in figure \ref{fig:BERvsNU_RT} for the BER of the machine-type users and in figure \ref{fig:BERvsNU_NRT} for best-effort traffic users.
\begin{figure
\input{figures/BERvsNU_RT}
\caption{BER of machine-type traffic users vs. number of users}
\label{fig:BERvsNU_RT}
\end{figure}
\begin{figure
\input{figures/BERvsNU_NRT}
\caption{BER of best-effort traffic users vs. number of users}
\label{fig:BERvsNU_NRT}
\end{figure}
As we can see, the BER in both cases worsens, as soon as overloading occurs, i.e. $N_{BE} > 4$. Due to the orthogonality in the case of machine-type traffic, this group of users experiences a better BER with at least two orders of magnitude compared to the users with best-effort traffic.
\subsubsection{Investigation of the trade-off between FEC code rate and spreading factor}
In a last simulation, we investigated the the trade-off between FEC code rate and spreading factor. In order to do so, we calculated the BER vs. the sending probability of ten users. The results are depicted in figure \ref{fig:BERvsFECvsSF}. We observe, that for best-effort traffic, investing in a higher spreading factor returns with lower BER with four orders of magnitudes. This is because in this case of $SF=16$ the system is not overloaded and orthogonality is maintained for the best-effort type users as well. Applying a convolutional code with code rate $1/2$ brings no benefit in terms of BER for best-effort users, as we can see when comparing to the orange dashed line, which is equivalent to the illustration in figure \ref{fig:BERvsPS} and was included for the sake of comparison.
For machine-type traffic, an increased spreading factor of $SF=16$ results in a BER gain of two orders of magnitude compared to the overloaded case, which is represented by the blue dashed line. Since in the case of $SF=16$ the user spreading sequences are equivalent to standard Hadamard sequences, the performance w.r.t. to BER for machine-type as well as best-effort type users is identical.
The application of the FEC with code rate $1/2$ to machine-type traffic resulted in no bit errors and is therefore not shown in figure \ref{fig:BERvsFECvsSF}. Since machine-type users usually have strict requirements concerning the BER, the application of FEC would be the better choice compared to an increased spreading factor in this case.
\begin{figure
\input{figures/FEC_0.5_SF_8_QPSK}
\caption{BER vs Sending Probability for different values of SF and FEC code rate}
\label{fig:BERvsFECvsSF}
\end{figure}
\section{Conclusion}
When thinking of wireless industrial communication systems, the challenge of efficiently combining machine-type traffic as well as best effort traffic with one MAC method is a crucial one. In this paper, we have presented a CDMA-based approach which looks promising in this context. As we have shown with the proposed CDMA MAC it is possible to overload parts of the system, enabling a higher number of users with lower requirements w.r.t. BER, while maintaining orthogonality for users with strict BER requirements. This means that with the proposed sequences, we are able to support Hadamard-equivalent performance regarding BER for one part of the users while allowing to increase the overall number of users, in particular users with less strict BER requirements and rather low sending probability, at the same time. The proposed MAC is therefore particularly appropriate, if the user group with lower BER requirements has a low sending probability as for example in wireless sensor networks.
\section{Future Work}
As we have shown in this paper, the proposed CDMA-based MAC is a promising candidate when the efficient combination of machine-type traffic as well as best effort traffic in the same system is required. In the future it would be preferable to investigate this using higher spreading factors. It is expected, that the advantages of the proposed MAC increase with higher spreading factors, mainly for two reasons:
\begin{enumerate}
\item For higher spreading factors, the additional interference per spreading sequence is relatively small.
\item With higher spreading factors, more additional spreading sequences, generated according to this paper, are available.
\end{enumerate}
It would be interesting to investigate scenarios with lots of users with only best-effort requirements and low sending probability.
|
\section{Introduction}
Even though our knowledge about the Galaxy has remarkably grown in the last century, we are still struggling with fundamental questions regarding its formation, structure and evolution. Understanding the chemodynamical evolution of the Galaxy from its birth to the present day is challenging because we can only directly observe a single frame in time, from an otherwise ever--changing complex scenario.
One way to infer the properties of the Galaxy at different epochs, and then recover its chemodynamical evolution, is by using individual stellar ages. The chemical abundances of a star reflects the interstellar medium abundances at the time in which the star was formed, and its kinematical properties may contain information about the galactic structure at that particular epoch.
Individual stellar ages are also important for studies of single stars, as the whole physical structure of a star may be determined exclusively from its mass, chemical composition and age \citep{Vogt1926, Russell+1927}. Mass and chemical abundances can be directly measured but ages can only be inferred from observable properties that are known to change with time \citep{Soderblom2010}.
The task of estimating stellar ages has been addressed by several authors and lots of different methods are found in the literature. For instance, there are: (i) empirical methods, which uses a deterministic relation between a given parameter and the age of a star (first proposed by \citealp{Skumanich1972}); this is the case of gyrochronology \citep[e.g.][]{Barnes2003, Barnes2007, Mamajek+Hillenbrand2008, CollierCameron+2009}, decay of cromosferic activity \citep[e.g.][]{Soderblom+1991, Rocha-Pinto+Maciel1998, Rocha-Pinto+2000, Lyra+PortodeMello2005, Pace+Pasquini2004, Pace+2009, Pace2013, Zhao+2011}, lithium depletion \citep{Sestito&Randich2005, Jackson+Jeffries2014, Carlos+2016} and ``magnetochronology'' (proposed by \citealp{Vidotto+2014}). (ii) model dependent methods, which are based on the comparison between measurable physical quantities and the ones expected from stellar structure models that use age as one of its parameters. Isochrone fitting \citep{Edvardsson+1993, Pont+Eyer2004, Nordstrom+2004, Jorgensen+Lindegren2005, Silaj+Landstreet2014, Maxted+2015} and asteroseismology \citep{Cunha+2007, Vauclair2009, Metcalfe+2010, Silva-Aguirre+2017} are classified in this category. (iii) semi-fundamental methods, those that are based on well known fundamental physics and employ only few assumptions; these are the cases for the method of cluster expansion \citep[e.g.][]{Makarov2007}, and nucleocosmocronology, known to predict unreliable ages \citep{Ludwig+2010}. (iv) Statistical methods, which uses statistical relations, like the age-metallicity relation (AMR) and the age--velocity dispersion relation (AVR), between a given property and the age. These relations have not been much explored in the literature as a direct tool to estimate stellar ages. Some few examples of its usage are found in \citet{Lachaume+1999} and \citet{Maciel+2011} (for the AVR) and \citet{Spina+2016} (for the AMR, specially [Y/Mg]$\times$age and [Y/Al]$\times$age).
\citet{Maciel+2011} developed a kinematical method that is based on the difference between the actual stellar rotational velocity, $\Theta$, and the one expected from the rotational curve. This difference is attributed to occasional disturbances on the stellar orbit and, therefore, stars that show greater differences must be older because they have participated in more collisions. In their work, \citet{Maciel+2011} adopt a deterministic relation between a given kinematical parameter and the stellar age. But the fact that stars of all ages may possess almost any velocity at all (e.g. the low velocities old stars, and the rare high velocities young stars) goes against their method. What is for sure related to the age is the probability of a star having a given velocity, because the velocity distribution changes based on the average age of the stellar sample. \st{} \citep{Nordstrom+2004, Casagrande+2011}.
It is known that the velocity dispersion of stars increases with the average age of the stellar groups \citep{Wielen1974, Wielen1977, Nordstrom+2004, Koval+2009, Casagrande+2011, Gontcharov2012, Martig+2014}. This relation has also been observed in external galaxies \citep{Beasley+2015, Dorman+2015}. The cause for this effect may be due to (i) stars being born dynamically hotter in the past or (ii) to stars being dynamically heated through time. The former can be explained by the disk being more gas-rich and more turbulent in the past \citep[e.g.][]{Bournaud+2009}, which is supported by observations of high-redshift galaxies \citep{Forster+2009}. Several causes have been proposed to explain the latter mechanism and are supported by cosmological simulations \citep[e.g.][]{House+2011}: heating caused by giant molecular clouds \citep[e.g.][]{Lacey1984, Hanninen+Flynn2002}, interaction with non-axissymetric galactic structures as transient and recurring spiral arms \citep[e.g.][]{Carlberg+Sellwood1985, Martinez-Medina+2015} or the bar \citep[e.g.][]{Saha+2010, Grand+2016}, and also interactions with satelite galaxies \citep[e.g.][]{Velazquez+White1999}.
Differently from \citet{Maciel+2011}, who obtained ages from a deterministic relation, we make use of the whole velocity distribution function and its well studied relation with age, to derive a new statistical age dating method, achieved through a Bayesian approach.
It is important to notice, that not all authors agree that the velocity dispersion parametrization is adequate to represent the velocity distribution in the $UV$ plane \citep[e.g.][]{Seabroke+Gilmore2007}. The reason for this is the existence of several substructures with typical sizes of $\approx 10 \, \, \mathrm{km} \, \mathrm{s}^{-1}$ in this plane. The mapping and study of the origins of these substructures are a very hot topic in the present days \citep{Dehnen1998, Famaey+2007, Bovy+Hogg2010, Bobylev+Bajkova2016}. For instance, it is not clear if these substructures are only local, or if they cover the whole disk \citep{Kushniruk+2017}. If their origin is dynamical, they would cover a large spread of ages, or would be concentrate within a single age if they originate from the disruption of stellar clusters \citep{Famaey+2008}. Considering that their origins and relation with age are still not well understood, and that modelling all these substructures in the velocity distributions would result in a very large number of free parameters, we still choose to work with the age-velocity dispersion. However, our model allows for an age-dependent correlation between the $U$ and $V$ velocities (described by the vertex deviation, $\ell_v$), which takes into account the existence of these groups as a first order approximation.
We work with F and G dwarfs because they have a long life expectancy. Therefore, their properties reflect properties of the disk at many different epochs. These stars also do not live too long to cause their internal changes to be insufficient to result in significant variations on observational properties that are used to derive their ages. Since orbital diffusion is not expected to depend on the stellar masses, the methods we derived for F and G stars are also likely to be valid for K and M dwarfs, which generally cannot have ages estimated by isochrones. As these stars have a lifetime that is greater than the age of the disk, they may contain key information about its chemodynamical state in different epochs and they are also numerous enough for us to work with averaged properties instead of individual stars that may be somehow peculiar. In Section \ref{sec:trappist1} we exemplify the flexibility of the method by calculating the age of Trappist-1, an M8V star for which other methods are not reliable.
We show how this Bayesian approach can be used to obtain a probability density function from which independent statistical ages may be estimated. We also show that from these estimated ages, important constraints on the evolution of the Galaxy may be obtained, as the age distribution and the age-metallicity relation.
The paper is structured as follows: Section \ref{sec:sample} presents the sample used to perform our analysis. In Section \ref{sec:methods} we develop three kinematical methods that can be used independently. The discussion of the method and some of our results are shown in Section \ref{sec:discussion}. Finally, Section \ref{sec:conclusions} presents our conclusions.
\section{Sample}
\label{sec:sample}
In order to calibrate the relations between the stellar kinematical parameters and the ages, we need a sample of stars that have known ages and velocities. The Geneva-Copenhagen Survey \citep{Nordstrom+2004}, an all-sky survey of F and G dwarf stars in the Solar Neighbourhood, is ideal for this task. It contains isochronal ages for most of the stars as well as unbiased kinematical information.
The first version of the catalog \citep{Nordstrom+2004} presented around $63\,000$ new radial velocities measurements. Together with published \textit{uvby$\beta$} photometry, Hipparcos parallaxes and Tycho-2 proper motions, this data allowed the authors to calculate, for most of the stars, their spatial velocities and also derive effective temperatures and metallicities from photometric calibrations. With these astrophysical parameters and theoretical isochrones from the Padova group \citep{Girardi+2000, Salasnich+2000} they applied the Bayesian method described by \citet{Jorgensen+Lindegren2005} to derive isochronal ages.
Since its first publication, there were three revisions of this survey. \citet{Holmberg+2007} improved the photometric calibrations for early F stars, \citet{Holmberg+2009} re-derived the astrophysical parameters implementing the revision of the Hipparcos parallaxes by \citet{vanLeeuwen2007}, and \citet{Casagrande+2011} revised the data using the effective temperature scale presented in \citet{Casagrande+2010} and also obtained new metallicity scales. From this improved data, \citet{Casagrande+2011} derived isochronal ages for the sample stars using both the Padova \citep{Bertelli+2008, Bertelli+2009} and BaSTI \citep{Pietrinferni+2004,Pietrinferni+2006} grids.
We built our sample from the data of \citet{Casagrande+2011} after applying the following selection criteria: (i) first, we removed all the stars for which the kinematical data was not complete; (ii) to avoid unknown binaries for which the derived astrophysical parameters would be unreliable, we discarded stars that had radial velocity determined from a single observation; (iii) as we are interested in fitting relations between kinematical parameters and age, we considered only the stars that could have its age derived by \citet{Casagrande+2011}; (iv) in order to work only with stars that have well defined ages, we have removed all those whose difference between the 84\% and 16\% percentiles and the median of the age probability density function exceeds 3 Gyr; (v) also to avoid unreliable ages, we have selected only the stars whose difference between the isochronal ages from the BaSTI and Padova grids are smaller than 1 Gyr. After applying all these selection criteria, our final sample consists of $9\,102$ F and G dwarf stars having complete kinematical data and well defined isochronal ages.
For simplicity, we consider further only the ages derived by the Padova grid. This will not affect our results because, by our sample definition, Padova and BaSTI ages cannot be significantly different.
\section{Methods}
\label{sec:methods}
The relation between age and velocity dispersion means that the probability of a star having a given velocity depends on the stellar age. Through the Bayes theorem, this probability relation can be reversed and one can obtain a probability density function for the age of a star given their velocities.
In its most general form, the Bayes Theorem for the probability of a parameter $\theta$, including multi-dimensional data $\vect{d}$ and general background knowledge $I$ can be written as
\begin{equation}
\label{eq:bayes_general_form}
p(\theta|\vect{d},I) \propto p(\vect{d}|\theta,I) \, p(\theta|I) \,.
\end{equation}
In the case the observational data consists of two observables (i.e. $d_1$ and $d_2$), the pdf for the parameter $\theta$ is given by \citep[eq. 48]{dAgostini2003}
\begin{equation}
\label{eq:bayes_general_two_obs}
p(\theta|d_1, d_2 ,I) \propto p(d_2|d_1,\theta,I) \, p(d_1|\theta,I) \, p(\theta|I) \,.
\end{equation}
When two observables are independent, it is also valid that
\begin{equation}
\label{eq:bayes_general_two_obs}
p(d_2|d_1,\theta,I) = p(d_2|\theta,I) \,.
\end{equation}
In this section, we show how these relations can be used to derive a probability density function for the age of a star from its measured spatial heliocentric velocities $U$, $V$, $W$ (named here as Method $UVW$). We also show how other kinematical properties, such as the eccentricity, can be used in the formalism, which is the case of Methods $eVW$ and $eUW$.
\subsection[Method $UVW$]{Method $\boldsymbol{UVW}$}
\label{sec:method_UVW}
In this method, the observational data consists of the spatial heliocentric velocity components $U$, $V$ and $W$ \footnote{In this work we define the $U$ axis as directed toward the galactic center.}. An important factor that must be taken into account is the existence of correlation between the $U$ and $V$ components, that might also depend on age \citep[see, for instance,][]{Rocha-Pinto+2004}. The correlations involving the $W$ component are usually smaller than the errors and therefore can be ignored \citep{Binney+Merrifield1998}.
The equations in this method are simplified by working with the components of the velocity ellipsoid, $v_1$, $v_2$ and $v_3$, instead of $U$, $V$ and $W$. This is the case because, by definition, they have no correlation between them. The $W$ and $v_3$ components are equivalent except for the displacement that is necessary to ensure zero mean for $v_3$, which is known to be caused by the Solar velocity with respect to the Local Standard of Rest ($W_{\sun}$). Calculating $v_1$ and $v_2$ then involves only the $U$ and $V$ components and can be described in terms of the vertex deviation ($\ell_v$). To calculate $v_1$ and $v_2$ one also needs to know the solar velocities $U_{\sun}$ and $V'_{\sun}$, with special care in the case of $V'_{\sun}$ because it may also depend on age. The transformation between $U$, $V$ and $W$ is then given by
\begin{subequations}
\label{eq:v1v2v3}
\begin{align}
v_1 &= (U+U_{\sun}) \, \cos{\ell_v} + (V+V'_{\sun}) \, \sin{\ell_v} \,, \\
v_2 &= -(U+U_{\sun}) \, \sin{\ell_v} + (V+V'_{\sun}) \, \cos{\ell_v} \,, \\
v_3 &= W+W_{\sun} \,,
\end{align}
\end{subequations}
where
\begin{equation}
\ell_v = \frac{1}{2}\arctan{\left( \frac{2\,\sigma_{UV}}{\sigma^2_U-\sigma^2_V} \right)}\,.
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=1]{Images/sigma_fit3_bw.pdf}
\caption{Velocity dispersion as a function of age for the $U$ (squares, solid line), $V$ (triangles, dashed line) and $W$ (circles, dot-dashed line) components, and also for the principal components, $v_1$ (open squares, long-dashed line), $v_2$ (open triangles, double dashed line), for each of the 30 bins divided by age. Uncertainties were estimated by bootstrap re-sampling. For each component, the line represents the best fit as a function of age following a relation $\sigma_i = b_i \, t^{a_i}$. The first and last group were excluded from the fit for reasons explained in the text. The values fitted for the parameters $b_i$ and $a_i$ are shown in Table \ref{tab:fit_sigma}.}
\label{fig:sigma_fit}
\end{figure}
The transformation between the probability density functions is given in terms of the Jacobian of the coordinate transformation as $p(t|U,V,W) = J(v_1,v_2,v_3) \, p(t|v_1,v_2,v_3)$, where, through the remaining of the paper, $t$ denotes age. In this case this transformation is very simple for it involves only translations and rotations so the Jacobian is equal to unity. It then follows from Equation \ref{eq:bayes_general_form} (omitting the term denoting general background knowledge, $I$) that
\begin{equation}
p(t|U,V,W) \propto p(v_1|t) \, p(v_2|t) \, p(v_3|t) \, p(t) \,.
\end{equation}
The probability $p(t)$ corresponds to the prior age probability, before including the observed data. The only known information we are considering is that no star in the galaxy should be older than 14 Gyr. For all lower ages we consider a uniform probability distribution, so that $p(t)$ is given by
\begin{equation}
p(t) = \begin{cases}
1, & \mathrm{if} \; 0 < t < 14 \; \textrm{Gyr.} \\
0, & \mathrm{otherwise.}
\end{cases}
\end{equation}
To describe the probabilities $p(v_i|t)$ we approximate the distributions as Gaussians having dispersions that depends on age, $\sigma_i(t)$. This simplification is necessary because we need to work with distributions that can be described by few parameters, for which we can fit relations as a function of age from our data of $9\,102$ stars. A more realistic description of the stellar distribution, that also includes moving groups, is beyond the scope of this work.
The parameters that are a function of age are the dispersions $\sigma_1(t)$, $\sigma_2(t)$, $\sigma_U(t)$, $\sigma_V(t)$ and $\sigma_W(t)$, and also the
vertex deviation $\ell_v(t)$ and the $V$ component of the Solar motion $V'_{\sun}$. In order to fit the age dependencies of these parameters we use a procedure similar to the one used by \citet{Nordstrom+2004} for the original GCS sample. First, we have divided the sample in 30 bins according to the stellar ages, then, the aforementioned parameters were calculated for each of the bins and the uncertainties were estimated by bootstrap re-sampling. Figure \ref{fig:sigma_fit} shows the velocity dispersion for each component calculated for each one of the groups. As usually done in the literature, we have chosen to fit the relation between dispersion and age as single power laws in the form $\sigma_i = b_i \, t^{a_i}$, with $t$ in Gyr. In each case, we excluded the first and last bins, shown in grey, before fitting the relation. The reason for this is that the last bin may be contaminated with stars from the thick disk, biasing the dispersion to higher values, and the stars that belong to the first bin are too young and have not completed a significant amount of orbits so they may reflect the kinematical properties of not yet dissolved local structures. We present the values obtained for $a_i$ and $b_i$ for all components in Table \ref{tab:fit_sigma}. The results corroborate what has already been established from the GCS: there is an increase in the velocity dispersion for all components in all age ranges considered for the disk.
\begin{table}
\centering
\caption{Parameters $b_i$ and $a_i$ that parametrize the relation between velocity dispersion and age, $\sigma_i = b_i \, t^{a_i}$, for each of the components $U$, $V$ and $W$ and for the principal components $v_1$ and $v_2$. Values obtained by \citet[GCSI]{Nordstrom+2004} and \citet[GCSIII]{Holmberg+2009} are shown for comparison.}
\label{tab:fit_sigma}
\begin{tabular}{ccccc}
\hline
& $b$ & $a$ & $a_{GCSI}$ & $a_{GCSIII}$ \\
\hline
$U$ & $21.2 \pm 1.0$ & $0.35 \pm 0.02$ & 0.31 & 0.39\\
$V$ & $13.0 \pm 1.0$ & $0.36 \pm 0.02$ & 0.34 & 0.40\\
$W$ & $9.1 \pm 1.0$ & $0.48 \pm 0.04$ & 0.47 & 0.50\\
$v_1$ & $22.0 \pm 1.0$ & $0.33 \pm 0.02$ & & \\
$v_2$ & $11.9 \pm 1.0$ & $0.42 \pm 0.02$ & & \\
\hline
\end{tabular}
\end{table}
The relations between the remaining parameters, $\ell_v$ and $V'_{\sun}$, and age were found in a similar manner and are represented in Figure \ref{fig:meanv_lv_fit}. The expressions that best describe the behaviour of these parameters as a function of age are
\begin{align}
\ell_v(t) &= 0.41 \, \exp(-0.37 \, t) \\
V'_{\sun}(t) &= 0.17 \, t^2 + 0.63 \, t + 12.5 \, \mathrm{.}
\end{align}
The increase found for $V'_{\sun}$ with age is expected and is caused by the asymmetric drift.
There is still much discussion regarding the cause of the vertex deviation and its dependence on age. Here we find that the vertex deviation is maximum for the youngest stars, and tends to zero for the older stars. A possible cause for the vertex deviation are the presence of moving groups \citep{Dehnen1998}.
For instance, an investigation regarding the existence and origin of the main moving groups in the Solar neighbourhood was done by \citet{Famaey+2008}. The positions of the four moving groups found in his study (Hercules, Pleiades, Hyades and Sirius) are aligned with the necessary rotation we found for the principal component of the velocity elipsoid. The stars in these groups are mostly associated with young clusters of a few hundreds of Myr, which corroborates the decline of the vertex deviation with age.
\begin{figure}
\centering
\includegraphics[scale=0.9]{Images/muV_lv_bw_edited.pdf}
\caption{Vertex deviation, $\ell_v$ (bottom), and $V'_{\sun}$ (top) calculated for each of the 30 groups dived by ages. Uncertainties were estimated by bootstrap re-sampling. The relations that describe this parameters as function of age (solid lines) are fitted as $V'_{\sun}(t) = 0.17 \, t^2 + 0.63 \, t + 12.5$ and $\ell_v(t) = 0.41 \, \exp(-0.37 \, t)$. For reasons explained in the text, the first and last groups were removed before fitting the relations.}
\label{fig:meanv_lv_fit}
\end{figure}
\begin{figure*}
\centering
\includegraphics[scale=0.9]{Images/dens_map2_bw.pdf}
\caption{(left) $eV$ plane density map built from a simulated sample of $2\,000\,000$ stars having $t = 4$ Gyr. The eccentricity was considered a function of $U$ and $V$, which were sampled from a distribution described by the parameters presented in Section \ref{sec:method_UVW}. (right) The probability density functions obtained from the density map for the eccentricity given the age (4 Gyr) and three different velocities $V = 20, -20, -60 \, \mathrm{km} \, \mathrm{s}^{-1}$, respectively (also represented on the left panel by dashed lines).}
\label{fig:dens_map}
\end{figure*}
The Solar velocity components are found from the averages of the velocities of stars in the Solar Neighbourhood. Our sample results in $U_{\sun} = 9.8 \pm 0.3 \, \mathrm{km} \, \mathrm{s}^{-1}$ and $W_{\sun} = 7.2 \pm 0.2 \, \mathrm{km} \, \mathrm{s}^{-1}$. As the average for the azimuthal component, $V$, has an age dependence, deriving this component of the Solar velocity is not as simple. If we define the $V_{\sun}$ component to be the velocity obtained from an idealized sample of stars with zero age, equivalent to $V'_{\sun}(t=0)$, we obtain $V_{\sun} = 12.5 \pm 0.9 \, \mathrm{km} \, \mathrm{s}^{-1}$. Table \ref{tab:sun_uvw} shows a comparison between this values and those obtained by other authors. As can be seen, the values obtained by different authors differ considerably, but the values calculated in this work are within the usual ranges. In Section \ref{sec:uncertainties-Solar-motion} we investigate how different input Solar Velocities affects the ages determinations.
\begin{table}
\centering
\caption{Comparisson between the values found for the Solar velocity components in this work, and the values obtained by \citet[F\&A14]{Francis+Anderson2014}, \citet[B\&B14]{Bobylev+Bajkova2014}, \citet[Co{\c s}+11]{Coskunoglu+2011}, \citet[Sch+10]{Schonrich+2010} and \citet[Kov+2009]{Koval+2009}.}
\label{tab:sun_uvw}
\begin{tabular}{lccc}
\hline
Author & $u_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$ & $v_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$ & $w_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$\\
\hline
This work & $9.8 \pm 0.3$ & $12.5 \pm 0.9$ & $7.2 \pm 0.2$\\
F\&A14 & $14.1 \pm 1.1$ & $14.6 \pm 0.4$ & $6.9 \pm 0.1$\\
B\&B14 & $6.0 \pm 0.5$ & $10.6 \pm 0.8$ & $6.5 \pm 0.3$\\
Co{\c s}+11 & $8.83 \pm 0.24$ & $14.19 \pm 0.34$ & $6.57 \pm 0.21$\\[+0.03in]
Sch+10 & $11.1^{+0.69}_{-0.75}$ & $12.24^{+0.47}_{-0.47}$ & $7.25^{+0.37}_{-0.36}$\\[0.03in]
Kov+09 & $5.1 \pm 0.4$ & $7.9 \pm 0.5$ & $7.7 \pm 0.2$\\
\hline
\end{tabular}
\end{table}
We then build the age pdf calculating the probability $p(t|U,V,W)$ for different ages. As we assume a Gaussian distribution, this probability is given in terms of the adjusted parameters as\footnote{The symbols $\sigma_1$ and $\sigma_2$ represent the velocity dispersions for the components $v_1$ and $v_2$. It should also be noted that Equation \ref{eq:v1v2v3} implies that $\sigma_3$ = $\sigma_W$.}
{\medmuskip=0mu
\thinmuskip=0mu
\thickmuskip=0mu
\begin{equation}
\label{eq:prob_t_method1}
p(t|U,V,W) \, \propto \mathlarger{\mathlarger{\prod_{\mathsmaller{\mathsmaller{i = 1, 2, 3}}}}} \left[ \frac{1}{(2\,\pi)^{\nicefrac{1}{2}} \, \sigma_i(t)} \exp{ \left(- \frac{v_i^2}{2 \sigma_i(t)^2}\right)} \right] \,,
\end{equation}}
where $v_1$, $v_2$ and $v_3$ are obtained from Equation \ref{eq:v1v2v3}.
\subsection[Method $eVW$]{Method $\boldsymbol{eVW}$}
\label{sec:method_eVW}
One can also use other kinematical parameters instead of the stellar velocity components, as long as the distribution for these parameters are known and varies with age. In this work, we show how this can be done including the orbital eccentricity. We call it Method $eVW$. Now the probability is obtained from the eccentricity and the velocities $V$ and $W$. In this case, these variables are no longer independent, since the eccentricity correlates with the $UV$ velocities, so the Bayes equation gives us
\begin{equation}
p(t|e,V,W) = p(e|V,t) \, p(V|t) \, p(W|t) \, p(t) \,.
\end{equation}
As we did for the Method $UVW$, we use a uniform $p(t)$ for ages between 0 and 14 Gyr and approximate the velocity distribution as Gaussians. The relation between the velocity distribution parameters and age were already obtained in Section \ref{sec:method_UVW} and all that is left to know is the probability $p(e|V,t)$.
In this work, we obtain the probability $p(e|V,t)$ from density maps in the $eV$ plane for different ages. To create the density maps, we generate random pairs of values ($U$, $V$) following the distribution described by the parameters obtained in Section \ref{sec:method_UVW} for a given age $t$. Then, we estimate the eccentricity for the simulated $U$ and $V$ values and build a density map using a 2d kernel density estimator method. The approximate expression used to estimate the eccentricities was
{\medmuskip=0mu
\thinmuskip=0mu
\thickmuskip=0mu
\begin{equation}
e = 2.98 \cdot 10^{-3} \, \left( 155 + 20.0\,U + 19.6\,V + U^2 + 1.95\,V^2 \right)^{\nicefrac{1}{2}} - 7.23 \cdot 10^{-4} \,,
\label{eq:ecc}
\end{equation}}
which was obtained with the automated model building software Eureqa \citep{Schmidt+Lipson2009}. This software iteratively tries a combination of polynomial, linear, exponential, logarithmic, trignometric and power functions, penalizing complexity, to find the best relation to describe a given parameter as a function of multiple others in a data set. In this case, we obtained the expression for the eccentricity as a function of the velocities $U$, $V$ and $W$ using the data set of \citet{Casagrande+2011}, therefore, this eccentricity corresponds to the one obtained when using the same galactic potential these authors have used in their paper.
This simple expression for the eccentricity eliminates the need for orbital integration, simplifying the process of building the density maps. Compared to the eccentricity calculated by \citet{Casagrande+2011} for the GCS stars, we found that, for 98\% of them, the eccentricity calculated by this expression differs from the one given in the catalog by less then 0.01, which we consider to be good enough for our purposes.
To exemplify how drastically the probability $p(e|V,t)$ may change for different values of $V$ and same age $t$ we present in Figure \ref{fig:dens_map} the density map in the $eV$ plane for the age of 4 Gyr. Also in Figure \ref{fig:dens_map}, the probability functions $p(e|V,t)$, for assigned values of $V$, are represented.
Considering the known values of $e$, $V$ and $W$ for a star, the probabilities $p(e|V,t)$, $p(V,t)$ and $p(W,t)$ are then calculated for different ages $t$ and the results are used to build the probability density function for the stellar age. While $p(e|V,t)$ is obtained through the density maps, $p(V|t)$ and $p(W|t)$
are obtained approximating the distributions as independent Gaussians and using the relations between age and the Gaussian parameters found in Section \ref{sec:method_UVW}:
\begin{equation}
\label{eq:prob_t_method2}
p(V|t) \cdot p(W|t) = \frac{1}{2\pi} \, \mathlarger{\mathlarger{\prod_{\mathsmaller{\mathsmaller{k = V, W}}}}}\left[\frac{1}{\sigma_k(t)} \, \exp{\left( - \frac{\left[k + k_{\sun}(t)\right]^2}{2\,\sigma_k^2(t)} \right)}\right]
\end{equation}
\subsection[Method eUW]{Method $\boldsymbol{eUW}$}
Method $eUW$ is essentially the same as Method $eVW$, but for this method, we are considering the $U$ velocity component instead of the $V$ component. In this case, the density maps needed are the ones in the $eU$ plane. The density maps were build from the same simulated data used for the ones described in Method $eVW$. The Bayesian equation that gives the age probability from the kinematical parameters is
\begin{equation}
p(t|e,U,W) = p(e|U,t) \, p(U|t) \, p(W|t) \, p(t) \,.
\end{equation}
As before, we obtain the age pdf calculating the probability $p(t|e,U,W)$ for different ages. In each case, $p(e|U,t)$ is obtained through the density maps and $p(U|t)$ and $p(W|t)$ are obtained approximating gaussians and using the relations found for the parameters in Section \ref{sec:method_UVW}:
\begin{equation}
\label{eq:prob_t_method3}
p(U|t) \cdot p(W|t) = \frac{1}{2\pi} \, \mathlarger{\mathlarger{\prod_{\mathsmaller{\mathsmaller{k = U, W}}}}}\left[\frac{1}{\sigma_k(t)} \, \exp{\left( - \frac{\left[k + k_{\sun}\right]^2}{2\,\sigma_k^2(t)} \right)}\right]
\end{equation}
\section{Discussion}
\label{sec:discussion}
We have applied the methods described in Section \ref{sec:methods} to derive the pdf for the age of all $9\,102$ stars from the sample defined in Section \ref{sec:sample}. Nine out of these stars could not have their ages determined by Methods $eVW$ and $eUW$ because their velocities were higher than the range we considered while building the density maps (i.e., $\vert U \vert \, \mathrm{or} \, \vert V \vert > 200 \, \mathrm{km} \, \mathrm{s}^{-1}$). As most of these stars also have high $W$ velocity, they should belong to other components of the Galaxy rather than the thin disk and may not affect our conclusions.
While it is best to use all information available in the posterior age pdf, it is often necessary to use a single age estimator in order to study the relation of the stellar parameters and age. As our first goal is to compare the results obtained from the kinematical method to those obtained from the isochronal one, we use the most likely age $t_{\mathrm{ML}}$ and expected age $t_{\mathrm{E}}$ obtained from the pdf to characterize single ages:
\begin{subequations}
\begin{align}
t_{\mathrm{ML}} &= \argmax_t f(t) \,, \\
t_{\mathrm{E}} &= \int_{-\infty}^{\infty} \! \! \! \! t \, f(t) \, \mathrm{d}t \,,
\end{align}
\end{subequations}
where $f(t)$ represents the posterior probability density function $p(t|U,V,W)$, $p(t|e,V,W)$ or $p(t|e,U,W)$, respectively for Methods $UVW$, $eVW$ and $eUW$.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/pdf_examples2_bw_edited.pdf}
\caption{Probability density function obtained by Method $UVW$ (solid line), Method $eVW$ (dashed line) and Method $eUW$ (dotted line), for four representative stars. Panel a) HD 1101, which has velocities $(U,V,W) = (-12, -23, -6) \, \mathrm{km} \, \mathrm{s}^{-1}$ and eccentricity $e = 0.07$, b) HD 1343, for which the velocities are $(U,V,W) = (-22, 16, 3) \, \mathrm{km} \, \mathrm{s}^{-1}$ and the eccentricity is $e = 0.09$, c) HD 852, that has $(U,V,W) = (-58, 39, -9) \, \mathrm{km} \, \mathrm{s}^{-1}$ and $e = 0.22$ and d) HD 12387, with velocities $(U,V,W) = (-3, -88, 60) \, \mathrm{km} \, \mathrm{s}^{-1}$ and eccentricity $e = 0.32$. The vertical marks at the top axis represents the most likely age for each method and those at bottom axis represents the expected ages.}
\label{fig:pdf_example}
\end{figure}
Figure \ref{fig:pdf_example} shows examples of pdfs obtained for four different stars, using the three methods previously described. Also plotted are the most likely age (upper axis lines) and the expected age obtained for each method (lower axis lines). These stars were selected as examples of the different situations one can find when calculating the pdf. In panel a), we can see that there is no defined lower limit for the stellar age, therefore, the most probable age for this star is defined as the lowest age considered in this work (0.1 Gyr). Panel b) represents the case for which both the expected age, and the most likely age can be properly defined. In the case of Panel c), even though a most likely age can be obtained, the expected age will be biased towards lower ages because of the truncation our prior $p(t)$ imposes on the pdf. Panel d) represents a case similar to Panel a), but in this case the limitation is caused by the fact that the most likely age would be greater than the age of the universe.
Since removing all stars whose age pdfs happens to be the cases a), c) or d) would seriously diminish our sample, we have chosen instead to work with a combination of the expected age ($t_{\mathrm{E}}$) and the most likely age ($t_{\mathrm{ML}}$). This way, the sample is less affected by the bias discussed above as in most cases at least one of the ages is well defined. We then define the kinematical age $t_{\mathrm{kin}}$ as a weighted average between $t_{\mathrm{ML}}$ and $t_{\mathrm{E}}$. We found that the expression giving the most similar results as isochronal ages is:
\begin{equation}
\label{eq:t_kin_definition}
t_{\mathrm{kin}} = \frac{3\,t_{\mathrm{ML}} + t_{\mathrm{E}}}{4} \mathrm{.}
\end{equation}
To tabulate the results, we have calculated from each stellar age pdf the expected age ($t^{(i)}_{\mathrm{E}}$), the most likely age ($t^{(i)}_{\mathrm{ML}}$) and the ages corresponding to the $2.5\%$, $16\%$, $50\%$, $84\%$ and $97.5\%$ percentiles, respectively $t^{(i)}_{2.5}$, $t^{(i)}_{16}$, $t^{(i)}_{50}$, $t^{(i)}_{84}$ and $t^{(i)}_{97.5}$, where $i = 1, 2, 3$ designates the method used to derive the age pdf, respectively methods $UVW$, $eVW$ and $eUW$. The results for the first 75 sample stars are shown in Table \ref{tab:results50}. The full data sample is only available in electronic format.
\subsection{Uncertainties for individual ages}
\label{sec:uncertainties}
Due to the statistical nature of this method, uncertainties arise from two different sources: (i) the observational uncertainties that affects the $UVW$ values and, consequently, the resulting pdf and (ii) uncertainties caused by our lack of knowledge of the exact value of the Solar peculiar velocity. We analyze how the latter affects the results by changing the Solar motion and, as the former depends on the quality of the data, we investigate how the results change when considering improved observational precision.
\subsubsection{Uncertainties due to statistical nature of the Method}
\begin{figure}
\centering
\includegraphics[scale=1]{Images/delta_distribution_bw_edited.pdf}
\caption{Distribution of individual uncertainties for the ages obtained by Method $UVW$ (solid line), Method $eVW$ (dashed line) and Method $eUW$ (dotted line) using Equation \ref{eq:delta_t_definition}. All methods are very similar and the $\delta_t$ distributions peak at about 3.25 Gyr, with no uncertainty higher than 4 Gyr.}
\label{fig:dist_delta}
\end{figure}
To estimate the uncertainties, related to the pdf spread, for the ages obtained for single stars through Equation \ref{eq:t_kin_definition}, we have used an expression involving the ages of the percentiles described above. The expression for the uncertainty was chosen in a way that, for a Gaussian distribution, $\delta_t$ would correspond to 1 sigma, and is given by
\begin{equation}
\label{eq:delta_t_definition}
\delta_t = \frac{1}{4} \left[ (t_{84} - t_{16}) + \frac{t_{97.5} - t_{2.5}}{2} \right] \, \mathrm{.}
\end{equation}
This definition is a measure of the spread that takes into account both the dispersion closer to the median age (from $t_{84}$ and $t_{16}$) and the dispersion at the tails of the distribution (from $t_{2.5}$ and $t_{97.5}$). It is designed to allow a direct comparison between the spread of the calculated kinematical method and original isochronal pdf obtained by \citet{Casagrande+2011}, whose data contains the data for these percentiles. It's important to keep in mind that, although a valid measure of the spread, it's arbitrary and not always corresponds to the Gaussian $1\sigma$, as the pdf can be very different from a normal distribution.
The distribution of individual uncertainties for each of the three methods is represented in Figure \ref{fig:dist_delta}. It can be seen that, concerning individual uncertainties, the methods presented here are very similar to each other. Figure \ref{fig:dist_delta} also shows that, for individual stars, uncertainties are very high. The mean uncertainty for Methods $UVW$ and $eVW$ is 3.1 Gyr, and for Method $eUW$, 3.0 Gyr. For comparison, the mean uncertainty for isochronal ages, defined in a similar way, would be about 0.7 Gyr for this sub-sample of the GCS. This means the kinematical ages are useful for individual stars only when the isochronal method is certain to result in extremely high uncertainties, as is the case for M dwarfs, or as an independent age indicator that can complement other methods.
\subsubsection{Impact of observational uncertainties}
In order to understand how the observational uncertainties that leads to the calculation of the $UVW$ velocities affect the obtained pdf, we have performed a set of Monte Carlo simulations and analyzed the effect of this in the calculated $t_\mathrm{ML}$ and $t_\mathrm{E}$.
For each star, the parallax, proper motions and radial velocities were re-sampled 1000 times, considering Gaussian individual errors, and the $UVW$ velocities were recalculated in each case.
The top panels of Figure \ref{fig:pdf_monte_carlo} shows the pdfs obtained for each of the 1000 Monte Carlo simulations (grey lines) and the original pdf (black line) for three representative star: HD 3598 (left) which has uncertainties considered small for the GCS survey ($\sigma_\pi = 0.53$ mas, $\sigma_\mu = 1.0$ mas/year and $\sigma_{r_v} = 0.3 \, \, \mathrm{km} \, \mathrm{s}^{-1}$); HD 578 (middle) which has uncertainties representative of the most stars in the survey ($\sigma_\pi = 0.66$ mas, $\sigma_\mu = 2.0$ mas/year and $\sigma_{r_v} = 0.4 \, \, \mathrm{km} \, \mathrm{s}^{-1}$); and HD 180748 (right) which has large uncertainties compared to other stars in the survey ($\sigma_\pi = 1.8$ mas, $\sigma_\mu = 4.0$ mas/year and $\sigma_{r_v} = 3 \, \, \mathrm{km} \, \mathrm{s}^{-1}$). Also represented in each plot, is the distribution of the obtained $t_\mathrm{ML}$ (dark histograms) and $t_\mathrm{E}$ (light grey histograms). The dispersions of the histograms are: $\sigma_{t_\mathrm{ML}} = 0.26$ Gyr and $\sigma_{t_\mathrm{E}} = 0.15$ for HD 3598; $\sigma_{t_\mathrm{ML}} = 0.39$ Gyr and $\sigma_{t_\mathrm{E}} = 0.27$ for HD 578; and $\sigma_{t_\mathrm{ML}} = 1.25$ Gyr and $\sigma_{t_\mathrm{E}} = 0.61$ Gyr for HD 180748.
It's clear that, the higher the observational uncertainties are, the greater is the uncertainties in the calculated age point estimator. For the whole sample, considering the GCS uncertainties, we obtain on average $\sigma_{t_\mathrm{ML}} = 0.47$ and $\sigma_{t_\mathrm{E}} = 0.28$. Therefore, considering the observational uncertainties of the GCS, the expected age is less affected by uncertainties than the most likely age.
\begin{figure}
\centering
\includegraphics[scale=0.52]{Images/pdf_variations_edited.png}
\caption{Effects of the observational uncertainties on the pdf considering the GCS uncertainties (top) and Gaia's predicted uncertainties (bottom) for the stars HD 3598 (left), HD 507 (middle) and HD 180748 (right). The black solid line represents the original pdf, while the grey lines are the results of 1000 Monte Carlo simulations. The dark grey histogram represents the distribution of most-likely ages and the light grey histogram, the distribution of expected ages.}
\label{fig:pdf_monte_carlo}
\end{figure}
Also, to verify how much the upcoming Gaia data will improve the results, we perform another set of Monte Carlo simulations considering the uncertainties that Gaia aims to achieve by the end of the mission \citep{Gaia2005}: $\sigma_\pi = 25 \cdot 10^{-3}$ mas, $\sigma_\mu = 13.15 \cdot 10^{-3}$ mas/year and $\sigma_{r_v} = 1 \, \, \mathrm{km} \, \mathrm{s}^{-1}$. The lower panels of Figure \ref{fig:pdf_monte_carlo} shows the obtained pdfs for the stars HD 3598 (left), HD 578 (middle) HD 180748 (right). As can be seen, the quality of the expected Gaia data practically eliminates the intrinsic uncertainty of the method due to observational uncertainties. On average, for the whole sample, we obtain $\sigma_{t_\mathrm{ML}} = 0.09$ Gyr and $\sigma_{t_\mathrm{E}} = 0.07$ Gyr.
\subsubsection{Effect of changing the Solar peculiar velocity}
\label{sec:uncertainties-Solar-motion}
We have also investigated the impact of changing the Solar peculiar velocity in the obtained pdfs and consequently in the estimated values of $t_\mathrm{ML}$ and $t_\mathrm{E}$. Figure \ref{fig:pdf_solar_velocity} shows the pdf obtained for the solar velocity obtained in this work (black line) and the pdfs obtained using the other values of solar velocity shown in Table \ref{tab:sun_uvw} (grey lines). We see that, considering the GCS uncertainties, the results are less affected by changing the Solar peculiar velocity than by the observational uncertainties (which will no longer be the case for the data quality expected by Gaia).
Compared to the Solar velocity of this work, the standard deviation of the age differences of the whole sample are: $\sigma_{t_\mathrm{ML}} = 0.26$ Gyr and $\sigma_{t_\mathrm{E}} = 0.18$ Gyr for the Solar peculiar velocity of F\&A14; $\sigma_{t_\mathrm{ML}} = 0.25$ Gyr and $\sigma_{t_\mathrm{E}} = 0.18$ Gyr for B\&B14; $\sigma_{t_\mathrm{ML}} = 0.21$ Gyr and $\sigma_{t_\mathrm{E}} = 0.14 $ for Co{\c s}+11, $\sigma_{t_\mathrm{ML}} = 0.08$ Gyr and $\sigma_{t_\mathrm{E}} = 0.06 $ for Sch+10 and $\sigma_{t_\mathrm{ML}} = 0.52$ Gyr and $\sigma_{t_\mathrm{E}} = 0.31 $ for Kov+09.
\begin{figure}
\centering
\includegraphics[scale=0.52]{Images/change_solar_motion.pdf}
\caption{Effects of changing the Solar peculiar velocity on the calculated pdfs of the stars HD 3598 (left), HD 507 (middle) and HD 180748 (rigth). The black line represents the pdf obtained using the Solar motion derived in this work, while the grey lines represents the pdfs obtained using the Solar velocities displayed in Table \ref{tab:sun_uvw}.}
\label{fig:pdf_solar_velocity}
\end{figure}
\subsection{Comparison between different methods}
Our method uses the isochronal ages derived from the Padova grid by \citet{Casagrande+2011} to fit the relations between the kinematical parameters and age, which are then used in Equations \ref{eq:v1v2v3}, \ref{eq:prob_t_method1}, \ref{eq:prob_t_method2} and \ref{eq:prob_t_method3}. To obtain a more independent comparison, we compare our derived kinematical ages against those derived by \citet{Casagrande+2011} using the BaSTI grid \footnote{While using the BaSTI ages for comparison is certainly better than using the Padova ages (which were used to fit the kinematical equations) they can't be seeing as fully independent age estimators. One of the criteria used to select the stars with well determined ages was that Padova and BaSTI ages had to agree within 1 Gyr.}. As these isochronal ages are also derived from a probability density function, we defined individual isochronal ages in the same way as before:
\begin{equation}
\label{eq:t_iso_definition}
t_{\mathrm{iso}} = \frac{3\,t^{(\mathrm{iso})}_{\mathrm{ML}} + t^{(\mathrm{iso})}_{\mathrm{E}}}{4} \, \mathrm{,}
\end{equation}
where $t^{(\mathrm{iso})}_{\mathrm{ML}}$ is the most likely age and $t^{(\mathrm{iso})}_{\mathrm{E}}$ is the expected age derived from the age pdf.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/Delta_iso_kin_distributions_bw_edited.pdf}
\caption{Distribution of difference between isochronal age and kinematical age, obtained by Method $UVW$ (left), Method $eVW$ (middle) and Method $eUW$ (right). The distributions are similar for all three methods and peak at zero, showing that the ages display a tendency of good agreement. The spreads can be explained by the individual uncertainties of both isochronal and kinematical methods.}
\label{fig:Delta_iso_kin}
\end{figure}
We plotted in Figure \ref{fig:Delta_iso_kin} the distributions of the differences between the isochronal and the kinematical ages derived by the three Methods. The behaviour of all distributions is similar, the peak is close to zero, meaning that the most common case is an agreement between isochronal and kinematical ages; the median of the distributions is also close to zero, being $-$0.33 for the case of Method $UVW$ (left), $-$0.34 for Method $eVW$ (middle) and $-$0.24 for Method $eUW$ (right). The spread in the distribution is explained by the high uncertainties of the kinematical methods, coupled with the also considerable uncertainties of the isochronal method.
Also noticeable is the long tail towards negative values that appears in all distributions, meaning that, in some cases, kinematical ages might be overestimated. This can be explained by the large spread in the pdf, which pushes the expected age towards the center of the age interval. As there are more young than old stars in the sample, there are more stars having their expected age overestimated than underestimated. The reason we chose to give less weight to the expected age in our kinematical age definition, Eq. \ref{eq:t_kin_definition}, was to reduce this effect.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/Delta_kin_kin_distributions_bw_edited.pdf}
\caption{Distributions of differences between ages obtained by the different kinematical methods. In all cases the peak of the distribution is close to zero, and the differences rarely exceed 1 Gyr, showing good agreement between the methods.}
\label{fig:Delta_kin_kin}
\end{figure}
We have also compared the ages obtained by the different methods present here, as can be seen in Figure \ref{fig:Delta_kin_kin}. There is very good agreement between all Methods as the differences rarely exceed 1 Gyr. The peak of the differences distributions is also very close to zero, the mean being 0.02 in the case of Methods $UVW$ and $eVW$, 0.14 for Methods $UVW$ and $eUW$ and 0.12 for Methods $eVW$ and $eUW$.
\subsection{Age distribution}
As the individual uncertainties $\delta_t$ are too high we only recommend this analysis for individual stars as an independent age indicator or when other precise methods are not available. Even though, the kinematical age method may be useful when the properties of large samples are considered. One of its applications is to derive the age distribution of stars in a given sample. Figure \ref{fig:age_dist} shows the distributions of stellar ages obtained by Method $UVW$ (top), Method $eVW$ (middle) and Method $eUW$ (bottom). Also, for comparison, we plotted the isochronal age distribution in all panels. In all cases, especially for Methods $eVW$ and $eUW$, the derived age distribution agrees very well to the isochronal case.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/age_distribution_bw_edited.pdf}
\caption{Distribution of ages obtained by Method $UVW$ (top), Method $eVW$ (middle) and Method $eUW$ (bottom). For comparison, the distribution of isochronal ages is plotted in all panels (dashed line). As can be seen, the ages distributions obtained by the kinematical methods are very similar to the ones obtained by the isochronal case. The agreement is even better for the cases that include eccentricity.}
\label{fig:age_dist}
\end{figure}
The age distribution in a sample is an important parameter, closely related to the star formation history in the Galaxy. The best stars to observe for this task are the ones that evolves slowly enough so that those born in the early phases of our Galaxy formation are still around. While slow evolution is what makes these stars live long enough to trace the Galactic history, it also makes it harder to obtain the stellar ages through their internal properties since they slowly change with time. This slow internal evolution does not affect the evolution of the parameters used by the kinematical method (spatial velocities and eccentricity), which varies with time in the same manner for star of different masses. Therefore, the kinematical method has a role in helping the understanding of important aspects of the evolution of our Galaxy.
In this work, we simply compare the kinematical age distribution with the isochronal age distribution. This distribution cannot be directly interpreted as a star formation history because the selection of stars for the Geneva-Copenhagen Survey \citep{Nordstrom+2004} was based in photometric cuts, which favours certain regions of the $T_\mathrm{eff}-M_{\mathrm{bol}}-[\mathrm{Me}/\mathrm{H}]$ space, directly biasing the age distribution. To translate the age distribution to star formation history, a correction for this bias must be applied, which is beyond the scope of this work.
\subsection{Age-metallicity relation}
The chemical evolution of the Galaxy remains largely unknown. Studies directly relating chemical abundances of different elements to specific epochs are still restricted to very few stars \citep{Spina+2016}. Therefore, to understand this evolution with the available data, we have to rely on models and observational constraints.
The first models considered a simple closed box scenario \citep{Talbot+Arnett1971} in which stars expel their material enriching the interstellar medium, causing newborn stars to be more metal rich than previous generations. The metallicity distribution predicted for present-day stars from this model does not agree well with observations \citep[e.g. the G-dwarf problem][]{Schmidt1963, Pagel+Patchett1975, Wyse+Gilmore1995, Rocha-Pinto+Maciel1996, Haywood2001, Nordstrom+2004}, indicating that more complex processes are involved. When inflows and outflows of gas \citep{Larson1972, Hartwick1976, Schindler+Diaferio2008}, and radial stellar migration \citep{Wielen+1996, Sellwood+Binney2002, Minchev+2013} are considered, the agreement between model predictions and observations is remarkably enhanced.
Considering the complexity of the models, more observational constraints are required. One property that can elucidate aspects of the Galactic chemical evolution is how the overall metallicity of stars changes with time, the so called age-metallicity relation (AMR). The most critical part of obtaining the AMR is determining stellar ages for low mass stars, especially the ones that can live enough to tell us about the early history of the Galaxy. The feature of the AMR that models must be able to explain is the low \citep[or none at all, see][]{Edvardsson+1993, Nordstrom+2004} increase of metallicity with time and the increase of the metallicity dispersion with age \citep{Casagrande+2011}.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/AMR_bw_edited.pdf}
\caption{Relation between the mean metallicity and the ages, lower than 7 Gyr, obtained through isochronal method (left) and kinematical methods (right). The behaviour of the relation is similar in both cases, suggesting that the kinematical method is useful for finding relations between chemical features and age.}
\label{fig:AMR}
\end{figure}
To check whether the kinematical method can be useful to derive the age-metallicity relation, we divided our stellar sample by bins of age 0.5 Gyr wide and calculated the mean $[\mathrm{Fe}/\mathrm{H}]$ for each group. The same was done for the isochronal ages for comparison. As groups older than 7 Gyr would have very few stars, we have restricted our analysis for ages lower than this. The results are plotted in Figure \ref{fig:AMR}. This raw analysis of the data shows a decay of metallicity with age both for the isochronal and the kinematical cases. The relation is slightly flatter for the kinematical ages. This is probably caused by the higher uncertainties for individual ages, which causes a mixture between stars that should, in reality, belong to other age bins.
As there is also a bias caused by selection criteria, we do not interpret the result as a true age-metallicity relation in the Solar Neighbourhood. In this work, we are just interested in showing that the results obtained from the kinematical ages are similar to those obtained using isochronal ones, therefore concluding that the kinematical method might be used for this task for samples of stars for which ages cannot be determined by isochrones.
\subsection{The age of Trappist-1}
\label{sec:trappist1}
Trappist-1 is the system that hosts the largest number of known earth-size planets in its habitable zone \citep{Gillon+2016}. Recent results show that the star harbors 5 planets of sizes similar to that of the Earth (planets b, c, e, f and g) and 2 planets slightly smaller ($\approx0.75$ $R_{\mathrm{\oplus}}$, planets d and h; \citealp{Gillon+2017}). The planets e, f and g are in the habitable zone and, therefore, could harbour water oceans.
After its discovery in 2016 and confirmation of 3 earth-size planets in the habitable zone in 2017, this star received a lot of attention and has been the central investigation of several studies consisting of further analysis of light curves \citep[e.g.][]{Wang+2017}, dynamical investigations \citep[e.g.][]{Tamayo+2017, Quarles+2017}, and habitability plausibility \citep[e.g.][]{OMalley+Kaltenegger2017, Wolf2017}. Despite of this, one very important information is still missing: the stellar age.
The reason for the lack of age information is the fact that Trappist-1 has a mass of only 8\% the mass of the Sun. It causes the stellar evolution to be too slow, making the atmospheric parameters practically unchanged since the stellar birth. Because of this, most of the traditional age dating methods can not be applied. Since the kinematical method is expected to be independent from the internal evolution of the star and depends only on its orbital evolution, we can apply the method to obtain the age p.d.f. for the star and estimate its expected value and uncertainty.
From the stellar coordinates, proper motions and parallax \citep{Costa+2006} and radial velocity \citep{Burgasser+2015}, we derived its heliocentric peculiar velocities using the method described by \citet{Johnson+Soderblom1987}. The obtained values were $U = -43.24 \, \mathrm{km} \, \mathrm{s}^{-1}$, $V = -66.25 \, \mathrm{km} \, \mathrm{s}^{-1}$ and $W = 13.87 \, \mathrm{km} \, \mathrm{s}^{-1}$. The eccentricity was calculated using Equation \ref{eq:ecc} and found to be 0.27. We used these velocities and eccentricity to apply Methods UVW, eVW and eUW to derive the age pdfs represented in Figure \ref{fig:Trappist_age}.
\begin{figure}
\centering
\includegraphics[scale=0.75]{Images/Trappist_age.pdf}
\caption{Age pdfs obtained for the Trappist-1 star using Method UVW (solid line), Method eVW (dashed line) and Method eUW (dotted line). The vertical lines in the top axis represent the most-likely age and the lines in the bottom axis, the expected age.}
\label{fig:Trappist_age}
\end{figure}
Figure \ref{fig:Trappist_age} shows that the most likely age (vertical marks at the top axis) and the expected age (vertical marks at the lower axis) for the three different methods. The most-likely ages for methods $UVW$, $eVW$ and $eUW$ are respectivelly 13.4, 13.13 and 13.06 Gyr and the expected ages are respectivelly 9.78, 9.86 and 9.73, showing good agreement between the three kinematical methods. We then apply Equation \ref{eq:t_kin_definition} to calculate the defined kinematical ages, and obtain 12.50 for Method $UVW$, 12.31 for Method $eVW$ and 12.23 for Method $eUW$.
In order to estimate upper and lower limits, we also calculate the 16\% and 84\% percentile ages, and obtained $t_\mathrm{kin}^{(UVW)} = 12.50^{+0.29}_{-6.23}$, $t_\mathrm{kin}^{(eVW)} = 12.31^{+0.53}_{-6.05}$, $t_\mathrm{kin}^{(eUW)} = 12.23^{+0.56}_{-6.22}$. In this case the kinematical Method only provides a good upper-limit for the age and has a large tail towards lower ages, as can also be seen in Figure \ref{fig:Trappist_age}. The results allows us to conclude that the age of the system most-likelly lies between $\approx$6--12.5 Gyrs. Considering the observational uncertainties, the results are in good agreement with the age derived by \citet{Burgasser+Mamajek2017}: $7.6 \pm 2.2$ Gyr. In this case, the authors have considered several age constrains: the stellar cmd, average density, lithium absorption, surface gravity, rotation, magnetic activity, as well as kinematics.
We believe the derived age further increases the astrobiological interest for the system: it not only contains 7 planets with sizes similar to the earth (3 of which are in the habitable zone) but is also old enough for life as we know it to develop and evolve. Nevertheless an important remark must be made: as the system is extremely compact (the highest period being of 18.77 days; \citealp{Luger+2017}) the planets are all expected to be tidally locked \citep{Gillon+2017}, causing the temperatures to vary significantly between the face towards the star and the opposite one, which would constitute another barrier for the formation of life.
\section{Conclusions}
\label{sec:conclusions}
We showed how a probability density function (pdf) for the age of a star may be obtained from its spatial velocity components $U$, $V$ and $W$ and also from other kinematical parameters like the eccentricity. We characterize individual ages from the pdf using the most likely and the expected ages. This individual age estimation has uncertainties of about 3 Gyr, which, although higher than classical methods, may be the best estimate available for most very low mass stars.
The methods are based on the growth of velocity dispersion with age. We adopted a relation between age and velocity dispersion as $\sigma_i(t) = b_i \, t^{a_i}$. The parameters $b_i$ and $a_i$ were found using a sub-sample of the Geneva-Copenhagen Survey, which contains isochronal ages derived by \citep{Casagrande+2011}. We show that other parameters that define the velocity distribution are also a function of stellar age: the $V$ component of the Solar motion ($V'_{\sun}$) and the vertex deviation ($\ell_v$). From this sample, we obtained the solar motion to be $(U,V,W)_{\sun} = (9.8\pm0.3, 12.5\pm0.9, 7.2\pm0.2)_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$. Using the example of eccentricity, we show how other kinematical parameters may be included in the analysis, provided the density of stars in the kinematical parameter-velocity space is known for all ages.
We applied the Kinematical Method for the stars of the Geneva-Copenhagen survey and compared the results with those obtained by isochronal estimation. We show that differences between the obtained ages peak at zero and, despite a slightly overestimation of kinematical ages, there is no significant bias. This means that, even though uncertainties are large, the kinematical method is useful to derive statistical parameters for groups of stars. A comparison between the kinematical methods based purely on spatial velocity and the ones that include eccentricity shows that there is a very good agreement in the ages derived and the differences rarely exceed 1 Gyr.
We investigated how the results are affected by observational uncertainties and by changing the Solar peculiar velocity. We conclude that, for the GCS uncertainties, the results are more affected by the observational uncertainties than by changes in the Solar peculiar velocity, but it will no longer be the case for the data quality that Gaia aims to achieve.
The distribution of stellar ages obtained through the kinematical methods agrees very well with the one obtained through isochrones, principally for the methods that include the eccentricity. We have not attempted to derive the star formation history from this data, because it would require the consideration of a detailed analysis of the age bias imposed by the sample selected criteria, which is left for future work. However, the agreement between kinematical and isochronal distribution leads us to conclude that the kinematical method may be as useful as the isochronal for this task, with the advantage of being applicable to the numerous very low mass stars.
The relation between metallicity and age, which is an important constraint for chemical evolution models, has also been investigated using the kinematical method. The behaviour of the relation shown for stars younger than 7 Gyr, is similar to the one obtained through isochrones, suggesting the kinematical method can also be used to derive this relation. An interpretation of this relation as a true age-metallicity relation also needs a detailed analysis of bias imprinted by selection criteria.
We showed that the method can also be applied for M stars, as it does not depend on the stellar mass. As an example, we derived the age for the Trappist-1 system, which is the known system that hosts the largest number of earth-size planets (seven, three of which are in the habitable zone). Our analysis of the obtained pdfs indicated an age between $\approx$6--12.5 Gyr.
We intent to explore this method to obtain kinematical ages pdfs for stars in large surveys, such as the Radial Velocity Experiment (RAVE) \citep{Steinmetz+2006} and Gaia \citep{Gaia+2016}, in the cases the other methods cannot be applied or would be unreliable. These pdfs can also be used as an independent age indicator to corroborate the results obtained by other means.
\textit{Acknowledgements}
For providing support with a PhD grant, Almeida-Fernandes F. wants to thank CAPES --- The Brazilian Federal Agency for Support and Evaluation of Graduate Education within the Ministry of Education of Brazil, and for the support, when the author was an undergrad, Almeida-Fernandes F. wants to thank CNPq --- The Brazilian National Council for Scientific and Technological Development.
\bibliographystyle{mnras}
\section{Introduction}
Even though our knowledge about the Galaxy has remarkably grown in the last century, we are still struggling with fundamental questions regarding its formation, structure and evolution. Understanding the chemodynamical evolution of the Galaxy from its birth to the present day is challenging because we can only directly observe a single frame in time, from an otherwise ever--changing complex scenario.
One way to infer the properties of the Galaxy at different epochs, and then recover its chemodynamical evolution, is by using individual stellar ages. The chemical abundances of a star reflects the interstellar medium abundances at the time in which the star was formed, and its kinematical properties may contain information about the galactic structure at that particular epoch.
Individual stellar ages are also important for studies of single stars, as the whole physical structure of a star may be determined exclusively from its mass, chemical composition and age \citep{Vogt1926, Russell+1927}. Mass and chemical abundances can be directly measured but ages can only be inferred from observable properties that are known to change with time \citep{Soderblom2010}.
The task of estimating stellar ages has been addressed by several authors and lots of different methods are found in the literature. For instance, there are: (i) empirical methods, which uses a deterministic relation between a given parameter and the age of a star (first proposed by \citealp{Skumanich1972}); this is the case of gyrochronology \citep[e.g.][]{Barnes2003, Barnes2007, Mamajek+Hillenbrand2008, CollierCameron+2009}, decay of cromosferic activity \citep[e.g.][]{Soderblom+1991, Rocha-Pinto+Maciel1998, Rocha-Pinto+2000, Lyra+PortodeMello2005, Pace+Pasquini2004, Pace+2009, Pace2013, Zhao+2011}, lithium depletion \citep{Sestito&Randich2005, Jackson+Jeffries2014, Carlos+2016} and ``magnetochronology'' (proposed by \citealp{Vidotto+2014}). (ii) model dependent methods, which are based on the comparison between measurable physical quantities and the ones expected from stellar structure models that use age as one of its parameters. Isochrone fitting \citep{Edvardsson+1993, Pont+Eyer2004, Nordstrom+2004, Jorgensen+Lindegren2005, Silaj+Landstreet2014, Maxted+2015} and asteroseismology \citep{Cunha+2007, Vauclair2009, Metcalfe+2010, Silva-Aguirre+2017} are classified in this category. (iii) semi-fundamental methods, those that are based on well known fundamental physics and employ only few assumptions; these are the cases for the method of cluster expansion \citep[e.g.][]{Makarov2007}, and nucleocosmocronology, known to predict unreliable ages \citep{Ludwig+2010}. (iv) Statistical methods, which uses statistical relations, like the age-metallicity relation (AMR) and the age--velocity dispersion relation (AVR), between a given property and the age. These relations have not been much explored in the literature as a direct tool to estimate stellar ages. Some few examples of its usage are found in \citet{Lachaume+1999} and \citet{Maciel+2011} (for the AVR) and \citet{Spina+2016} (for the AMR, specially [Y/Mg]$\times$age and [Y/Al]$\times$age).
\citet{Maciel+2011} developed a kinematical method that is based on the difference between the actual stellar rotational velocity, $\Theta$, and the one expected from the rotational curve. This difference is attributed to occasional disturbances on the stellar orbit and, therefore, stars that show greater differences must be older because they have participated in more collisions. In their work, \citet{Maciel+2011} adopt a deterministic relation between a given kinematical parameter and the stellar age. But the fact that stars of all ages may possess almost any velocity at all (e.g. the low velocities old stars, and the rare high velocities young stars) goes against their method. What is for sure related to the age is the probability of a star having a given velocity, because the velocity distribution changes based on the average age of the stellar sample. \st{} \citep{Nordstrom+2004, Casagrande+2011}.
It is known that the velocity dispersion of stars increases with the average age of the stellar groups \citep{Wielen1974, Wielen1977, Nordstrom+2004, Koval+2009, Casagrande+2011, Gontcharov2012, Martig+2014}. This relation has also been observed in external galaxies \citep{Beasley+2015, Dorman+2015}. The cause for this effect may be due to (i) stars being born dynamically hotter in the past or (ii) to stars being dynamically heated through time. The former can be explained by the disk being more gas-rich and more turbulent in the past \citep[e.g.][]{Bournaud+2009}, which is supported by observations of high-redshift galaxies \citep{Forster+2009}. Several causes have been proposed to explain the latter mechanism and are supported by cosmological simulations \citep[e.g.][]{House+2011}: heating caused by giant molecular clouds \citep[e.g.][]{Lacey1984, Hanninen+Flynn2002}, interaction with non-axissymetric galactic structures as transient and recurring spiral arms \citep[e.g.][]{Carlberg+Sellwood1985, Martinez-Medina+2015} or the bar \citep[e.g.][]{Saha+2010, Grand+2016}, and also interactions with satelite galaxies \citep[e.g.][]{Velazquez+White1999}.
Differently from \citet{Maciel+2011}, who obtained ages from a deterministic relation, we make use of the whole velocity distribution function and its well studied relation with age, to derive a new statistical age dating method, achieved through a Bayesian approach.
It is important to notice, that not all authors agree that the velocity dispersion parametrization is adequate to represent the velocity distribution in the $UV$ plane \citep[e.g.][]{Seabroke+Gilmore2007}. The reason for this is the existence of several substructures with typical sizes of $\approx 10 \, \, \mathrm{km} \, \mathrm{s}^{-1}$ in this plane. The mapping and study of the origins of these substructures are a very hot topic in the present days \citep{Dehnen1998, Famaey+2007, Bovy+Hogg2010, Bobylev+Bajkova2016}. For instance, it is not clear if these substructures are only local, or if they cover the whole disk \citep{Kushniruk+2017}. If their origin is dynamical, they would cover a large spread of ages, or would be concentrate within a single age if they originate from the disruption of stellar clusters \citep{Famaey+2008}. Considering that their origins and relation with age are still not well understood, and that modelling all these substructures in the velocity distributions would result in a very large number of free parameters, we still choose to work with the age-velocity dispersion. However, our model allows for an age-dependent correlation between the $U$ and $V$ velocities (described by the vertex deviation, $\ell_v$), which takes into account the existence of these groups as a first order approximation.
We work with F and G dwarfs because they have a long life expectancy. Therefore, their properties reflect properties of the disk at many different epochs. These stars also do not live too long to cause their internal changes to be insufficient to result in significant variations on observational properties that are used to derive their ages. Since orbital diffusion is not expected to depend on the stellar masses, the methods we derived for F and G stars are also likely to be valid for K and M dwarfs, which generally cannot have ages estimated by isochrones. As these stars have a lifetime that is greater than the age of the disk, they may contain key information about its chemodynamical state in different epochs and they are also numerous enough for us to work with averaged properties instead of individual stars that may be somehow peculiar. In Section \ref{sec:trappist1} we exemplify the flexibility of the method by calculating the age of Trappist-1, an M8V star for which other methods are not reliable.
We show how this Bayesian approach can be used to obtain a probability density function from which independent statistical ages may be estimated. We also show that from these estimated ages, important constraints on the evolution of the Galaxy may be obtained, as the age distribution and the age-metallicity relation.
The paper is structured as follows: Section \ref{sec:sample} presents the sample used to perform our analysis. In Section \ref{sec:methods} we develop three kinematical methods that can be used independently. The discussion of the method and some of our results are shown in Section \ref{sec:discussion}. Finally, Section \ref{sec:conclusions} presents our conclusions.
\section{Sample}
\label{sec:sample}
In order to calibrate the relations between the stellar kinematical parameters and the ages, we need a sample of stars that have known ages and velocities. The Geneva-Copenhagen Survey \citep{Nordstrom+2004}, an all-sky survey of F and G dwarf stars in the Solar Neighbourhood, is ideal for this task. It contains isochronal ages for most of the stars as well as unbiased kinematical information.
The first version of the catalog \citep{Nordstrom+2004} presented around $63\,000$ new radial velocities measurements. Together with published \textit{uvby$\beta$} photometry, Hipparcos parallaxes and Tycho-2 proper motions, this data allowed the authors to calculate, for most of the stars, their spatial velocities and also derive effective temperatures and metallicities from photometric calibrations. With these astrophysical parameters and theoretical isochrones from the Padova group \citep{Girardi+2000, Salasnich+2000} they applied the Bayesian method described by \citet{Jorgensen+Lindegren2005} to derive isochronal ages.
Since its first publication, there were three revisions of this survey. \citet{Holmberg+2007} improved the photometric calibrations for early F stars, \citet{Holmberg+2009} re-derived the astrophysical parameters implementing the revision of the Hipparcos parallaxes by \citet{vanLeeuwen2007}, and \citet{Casagrande+2011} revised the data using the effective temperature scale presented in \citet{Casagrande+2010} and also obtained new metallicity scales. From this improved data, \citet{Casagrande+2011} derived isochronal ages for the sample stars using both the Padova \citep{Bertelli+2008, Bertelli+2009} and BaSTI \citep{Pietrinferni+2004,Pietrinferni+2006} grids.
We built our sample from the data of \citet{Casagrande+2011} after applying the following selection criteria: (i) first, we removed all the stars for which the kinematical data was not complete; (ii) to avoid unknown binaries for which the derived astrophysical parameters would be unreliable, we discarded stars that had radial velocity determined from a single observation; (iii) as we are interested in fitting relations between kinematical parameters and age, we considered only the stars that could have its age derived by \citet{Casagrande+2011}; (iv) in order to work only with stars that have well defined ages, we have removed all those whose difference between the 84\% and 16\% percentiles and the median of the age probability density function exceeds 3 Gyr; (v) also to avoid unreliable ages, we have selected only the stars whose difference between the isochronal ages from the BaSTI and Padova grids are smaller than 1 Gyr. After applying all these selection criteria, our final sample consists of $9\,102$ F and G dwarf stars having complete kinematical data and well defined isochronal ages.
For simplicity, we consider further only the ages derived by the Padova grid. This will not affect our results because, by our sample definition, Padova and BaSTI ages cannot be significantly different.
\section{Methods}
\label{sec:methods}
The relation between age and velocity dispersion means that the probability of a star having a given velocity depends on the stellar age. Through the Bayes theorem, this probability relation can be reversed and one can obtain a probability density function for the age of a star given their velocities.
In its most general form, the Bayes Theorem for the probability of a parameter $\theta$, including multi-dimensional data $\vect{d}$ and general background knowledge $I$ can be written as
\begin{equation}
\label{eq:bayes_general_form}
p(\theta|\vect{d},I) \propto p(\vect{d}|\theta,I) \, p(\theta|I) \,.
\end{equation}
In the case the observational data consists of two observables (i.e. $d_1$ and $d_2$), the pdf for the parameter $\theta$ is given by \citep[eq. 48]{dAgostini2003}
\begin{equation}
\label{eq:bayes_general_two_obs}
p(\theta|d_1, d_2 ,I) \propto p(d_2|d_1,\theta,I) \, p(d_1|\theta,I) \, p(\theta|I) \,.
\end{equation}
When two observables are independent, it is also valid that
\begin{equation}
\label{eq:bayes_general_two_obs}
p(d_2|d_1,\theta,I) = p(d_2|\theta,I) \,.
\end{equation}
In this section, we show how these relations can be used to derive a probability density function for the age of a star from its measured spatial heliocentric velocities $U$, $V$, $W$ (named here as Method $UVW$). We also show how other kinematical properties, such as the eccentricity, can be used in the formalism, which is the case of Methods $eVW$ and $eUW$.
\subsection[Method $UVW$]{Method $\boldsymbol{UVW}$}
\label{sec:method_UVW}
In this method, the observational data consists of the spatial heliocentric velocity components $U$, $V$ and $W$ \footnote{In this work we define the $U$ axis as directed toward the galactic center.}. An important factor that must be taken into account is the existence of correlation between the $U$ and $V$ components, that might also depend on age \citep[see, for instance,][]{Rocha-Pinto+2004}. The correlations involving the $W$ component are usually smaller than the errors and therefore can be ignored \citep{Binney+Merrifield1998}.
The equations in this method are simplified by working with the components of the velocity ellipsoid, $v_1$, $v_2$ and $v_3$, instead of $U$, $V$ and $W$. This is the case because, by definition, they have no correlation between them. The $W$ and $v_3$ components are equivalent except for the displacement that is necessary to ensure zero mean for $v_3$, which is known to be caused by the Solar velocity with respect to the Local Standard of Rest ($W_{\sun}$). Calculating $v_1$ and $v_2$ then involves only the $U$ and $V$ components and can be described in terms of the vertex deviation ($\ell_v$). To calculate $v_1$ and $v_2$ one also needs to know the solar velocities $U_{\sun}$ and $V'_{\sun}$, with special care in the case of $V'_{\sun}$ because it may also depend on age. The transformation between $U$, $V$ and $W$ is then given by
\begin{subequations}
\label{eq:v1v2v3}
\begin{align}
v_1 &= (U+U_{\sun}) \, \cos{\ell_v} + (V+V'_{\sun}) \, \sin{\ell_v} \,, \\
v_2 &= -(U+U_{\sun}) \, \sin{\ell_v} + (V+V'_{\sun}) \, \cos{\ell_v} \,, \\
v_3 &= W+W_{\sun} \,,
\end{align}
\end{subequations}
where
\begin{equation}
\ell_v = \frac{1}{2}\arctan{\left( \frac{2\,\sigma_{UV}}{\sigma^2_U-\sigma^2_V} \right)}\,.
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=1]{Images/sigma_fit3_bw.pdf}
\caption{Velocity dispersion as a function of age for the $U$ (squares, solid line), $V$ (triangles, dashed line) and $W$ (circles, dot-dashed line) components, and also for the principal components, $v_1$ (open squares, long-dashed line), $v_2$ (open triangles, double dashed line), for each of the 30 bins divided by age. Uncertainties were estimated by bootstrap re-sampling. For each component, the line represents the best fit as a function of age following a relation $\sigma_i = b_i \, t^{a_i}$. The first and last group were excluded from the fit for reasons explained in the text. The values fitted for the parameters $b_i$ and $a_i$ are shown in Table \ref{tab:fit_sigma}.}
\label{fig:sigma_fit}
\end{figure}
The transformation between the probability density functions is given in terms of the Jacobian of the coordinate transformation as $p(t|U,V,W) = J(v_1,v_2,v_3) \, p(t|v_1,v_2,v_3)$, where, through the remaining of the paper, $t$ denotes age. In this case this transformation is very simple for it involves only translations and rotations so the Jacobian is equal to unity. It then follows from Equation \ref{eq:bayes_general_form} (omitting the term denoting general background knowledge, $I$) that
\begin{equation}
p(t|U,V,W) \propto p(v_1|t) \, p(v_2|t) \, p(v_3|t) \, p(t) \,.
\end{equation}
The probability $p(t)$ corresponds to the prior age probability, before including the observed data. The only known information we are considering is that no star in the galaxy should be older than 14 Gyr. For all lower ages we consider a uniform probability distribution, so that $p(t)$ is given by
\begin{equation}
p(t) = \begin{cases}
1, & \mathrm{if} \; 0 < t < 14 \; \textrm{Gyr.} \\
0, & \mathrm{otherwise.}
\end{cases}
\end{equation}
To describe the probabilities $p(v_i|t)$ we approximate the distributions as Gaussians having dispersions that depends on age, $\sigma_i(t)$. This simplification is necessary because we need to work with distributions that can be described by few parameters, for which we can fit relations as a function of age from our data of $9\,102$ stars. A more realistic description of the stellar distribution, that also includes moving groups, is beyond the scope of this work.
The parameters that are a function of age are the dispersions $\sigma_1(t)$, $\sigma_2(t)$, $\sigma_U(t)$, $\sigma_V(t)$ and $\sigma_W(t)$, and also the
vertex deviation $\ell_v(t)$ and the $V$ component of the Solar motion $V'_{\sun}$. In order to fit the age dependencies of these parameters we use a procedure similar to the one used by \citet{Nordstrom+2004} for the original GCS sample. First, we have divided the sample in 30 bins according to the stellar ages, then, the aforementioned parameters were calculated for each of the bins and the uncertainties were estimated by bootstrap re-sampling. Figure \ref{fig:sigma_fit} shows the velocity dispersion for each component calculated for each one of the groups. As usually done in the literature, we have chosen to fit the relation between dispersion and age as single power laws in the form $\sigma_i = b_i \, t^{a_i}$, with $t$ in Gyr. In each case, we excluded the first and last bins, shown in grey, before fitting the relation. The reason for this is that the last bin may be contaminated with stars from the thick disk, biasing the dispersion to higher values, and the stars that belong to the first bin are too young and have not completed a significant amount of orbits so they may reflect the kinematical properties of not yet dissolved local structures. We present the values obtained for $a_i$ and $b_i$ for all components in Table \ref{tab:fit_sigma}. The results corroborate what has already been established from the GCS: there is an increase in the velocity dispersion for all components in all age ranges considered for the disk.
\begin{table}
\centering
\caption{Parameters $b_i$ and $a_i$ that parametrize the relation between velocity dispersion and age, $\sigma_i = b_i \, t^{a_i}$, for each of the components $U$, $V$ and $W$ and for the principal components $v_1$ and $v_2$. Values obtained by \citet[GCSI]{Nordstrom+2004} and \citet[GCSIII]{Holmberg+2009} are shown for comparison.}
\label{tab:fit_sigma}
\begin{tabular}{ccccc}
\hline
& $b$ & $a$ & $a_{GCSI}$ & $a_{GCSIII}$ \\
\hline
$U$ & $21.2 \pm 1.0$ & $0.35 \pm 0.02$ & 0.31 & 0.39\\
$V$ & $13.0 \pm 1.0$ & $0.36 \pm 0.02$ & 0.34 & 0.40\\
$W$ & $9.1 \pm 1.0$ & $0.48 \pm 0.04$ & 0.47 & 0.50\\
$v_1$ & $22.0 \pm 1.0$ & $0.33 \pm 0.02$ & & \\
$v_2$ & $11.9 \pm 1.0$ & $0.42 \pm 0.02$ & & \\
\hline
\end{tabular}
\end{table}
The relations between the remaining parameters, $\ell_v$ and $V'_{\sun}$, and age were found in a similar manner and are represented in Figure \ref{fig:meanv_lv_fit}. The expressions that best describe the behaviour of these parameters as a function of age are
\begin{align}
\ell_v(t) &= 0.41 \, \exp(-0.37 \, t) \\
V'_{\sun}(t) &= 0.17 \, t^2 + 0.63 \, t + 12.5 \, \mathrm{.}
\end{align}
The increase found for $V'_{\sun}$ with age is expected and is caused by the asymmetric drift.
There is still much discussion regarding the cause of the vertex deviation and its dependence on age. Here we find that the vertex deviation is maximum for the youngest stars, and tends to zero for the older stars. A possible cause for the vertex deviation are the presence of moving groups \citep{Dehnen1998}.
For instance, an investigation regarding the existence and origin of the main moving groups in the Solar neighbourhood was done by \citet{Famaey+2008}. The positions of the four moving groups found in his study (Hercules, Pleiades, Hyades and Sirius) are aligned with the necessary rotation we found for the principal component of the velocity elipsoid. The stars in these groups are mostly associated with young clusters of a few hundreds of Myr, which corroborates the decline of the vertex deviation with age.
\begin{figure}
\centering
\includegraphics[scale=0.9]{Images/muV_lv_bw_edited.pdf}
\caption{Vertex deviation, $\ell_v$ (bottom), and $V'_{\sun}$ (top) calculated for each of the 30 groups dived by ages. Uncertainties were estimated by bootstrap re-sampling. The relations that describe this parameters as function of age (solid lines) are fitted as $V'_{\sun}(t) = 0.17 \, t^2 + 0.63 \, t + 12.5$ and $\ell_v(t) = 0.41 \, \exp(-0.37 \, t)$. For reasons explained in the text, the first and last groups were removed before fitting the relations.}
\label{fig:meanv_lv_fit}
\end{figure}
\begin{figure*}
\centering
\includegraphics[scale=0.9]{Images/dens_map2_bw.pdf}
\caption{(left) $eV$ plane density map built from a simulated sample of $2\,000\,000$ stars having $t = 4$ Gyr. The eccentricity was considered a function of $U$ and $V$, which were sampled from a distribution described by the parameters presented in Section \ref{sec:method_UVW}. (right) The probability density functions obtained from the density map for the eccentricity given the age (4 Gyr) and three different velocities $V = 20, -20, -60 \, \mathrm{km} \, \mathrm{s}^{-1}$, respectively (also represented on the left panel by dashed lines).}
\label{fig:dens_map}
\end{figure*}
The Solar velocity components are found from the averages of the velocities of stars in the Solar Neighbourhood. Our sample results in $U_{\sun} = 9.8 \pm 0.3 \, \mathrm{km} \, \mathrm{s}^{-1}$ and $W_{\sun} = 7.2 \pm 0.2 \, \mathrm{km} \, \mathrm{s}^{-1}$. As the average for the azimuthal component, $V$, has an age dependence, deriving this component of the Solar velocity is not as simple. If we define the $V_{\sun}$ component to be the velocity obtained from an idealized sample of stars with zero age, equivalent to $V'_{\sun}(t=0)$, we obtain $V_{\sun} = 12.5 \pm 0.9 \, \mathrm{km} \, \mathrm{s}^{-1}$. Table \ref{tab:sun_uvw} shows a comparison between this values and those obtained by other authors. As can be seen, the values obtained by different authors differ considerably, but the values calculated in this work are within the usual ranges. In Section \ref{sec:uncertainties-Solar-motion} we investigate how different input Solar Velocities affects the ages determinations.
\begin{table}
\centering
\caption{Comparisson between the values found for the Solar velocity components in this work, and the values obtained by \citet[F\&A14]{Francis+Anderson2014}, \citet[B\&B14]{Bobylev+Bajkova2014}, \citet[Co{\c s}+11]{Coskunoglu+2011}, \citet[Sch+10]{Schonrich+2010} and \citet[Kov+2009]{Koval+2009}.}
\label{tab:sun_uvw}
\begin{tabular}{lccc}
\hline
Author & $u_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$ & $v_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$ & $w_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$\\
\hline
This work & $9.8 \pm 0.3$ & $12.5 \pm 0.9$ & $7.2 \pm 0.2$\\
F\&A14 & $14.1 \pm 1.1$ & $14.6 \pm 0.4$ & $6.9 \pm 0.1$\\
B\&B14 & $6.0 \pm 0.5$ & $10.6 \pm 0.8$ & $6.5 \pm 0.3$\\
Co{\c s}+11 & $8.83 \pm 0.24$ & $14.19 \pm 0.34$ & $6.57 \pm 0.21$\\[+0.03in]
Sch+10 & $11.1^{+0.69}_{-0.75}$ & $12.24^{+0.47}_{-0.47}$ & $7.25^{+0.37}_{-0.36}$\\[0.03in]
Kov+09 & $5.1 \pm 0.4$ & $7.9 \pm 0.5$ & $7.7 \pm 0.2$\\
\hline
\end{tabular}
\end{table}
We then build the age pdf calculating the probability $p(t|U,V,W)$ for different ages. As we assume a Gaussian distribution, this probability is given in terms of the adjusted parameters as\footnote{The symbols $\sigma_1$ and $\sigma_2$ represent the velocity dispersions for the components $v_1$ and $v_2$. It should also be noted that Equation \ref{eq:v1v2v3} implies that $\sigma_3$ = $\sigma_W$.}
{\medmuskip=0mu
\thinmuskip=0mu
\thickmuskip=0mu
\begin{equation}
\label{eq:prob_t_method1}
p(t|U,V,W) \, \propto \mathlarger{\mathlarger{\prod_{\mathsmaller{\mathsmaller{i = 1, 2, 3}}}}} \left[ \frac{1}{(2\,\pi)^{\nicefrac{1}{2}} \, \sigma_i(t)} \exp{ \left(- \frac{v_i^2}{2 \sigma_i(t)^2}\right)} \right] \,,
\end{equation}}
where $v_1$, $v_2$ and $v_3$ are obtained from Equation \ref{eq:v1v2v3}.
\subsection[Method $eVW$]{Method $\boldsymbol{eVW}$}
\label{sec:method_eVW}
One can also use other kinematical parameters instead of the stellar velocity components, as long as the distribution for these parameters are known and varies with age. In this work, we show how this can be done including the orbital eccentricity. We call it Method $eVW$. Now the probability is obtained from the eccentricity and the velocities $V$ and $W$. In this case, these variables are no longer independent, since the eccentricity correlates with the $UV$ velocities, so the Bayes equation gives us
\begin{equation}
p(t|e,V,W) = p(e|V,t) \, p(V|t) \, p(W|t) \, p(t) \,.
\end{equation}
As we did for the Method $UVW$, we use a uniform $p(t)$ for ages between 0 and 14 Gyr and approximate the velocity distribution as Gaussians. The relation between the velocity distribution parameters and age were already obtained in Section \ref{sec:method_UVW} and all that is left to know is the probability $p(e|V,t)$.
In this work, we obtain the probability $p(e|V,t)$ from density maps in the $eV$ plane for different ages. To create the density maps, we generate random pairs of values ($U$, $V$) following the distribution described by the parameters obtained in Section \ref{sec:method_UVW} for a given age $t$. Then, we estimate the eccentricity for the simulated $U$ and $V$ values and build a density map using a 2d kernel density estimator method. The approximate expression used to estimate the eccentricities was
{\medmuskip=0mu
\thinmuskip=0mu
\thickmuskip=0mu
\begin{equation}
e = 2.98 \cdot 10^{-3} \, \left( 155 + 20.0\,U + 19.6\,V + U^2 + 1.95\,V^2 \right)^{\nicefrac{1}{2}} - 7.23 \cdot 10^{-4} \,,
\label{eq:ecc}
\end{equation}}
which was obtained with the automated model building software Eureqa \citep{Schmidt+Lipson2009}. This software iteratively tries a combination of polynomial, linear, exponential, logarithmic, trignometric and power functions, penalizing complexity, to find the best relation to describe a given parameter as a function of multiple others in a data set. In this case, we obtained the expression for the eccentricity as a function of the velocities $U$, $V$ and $W$ using the data set of \citet{Casagrande+2011}, therefore, this eccentricity corresponds to the one obtained when using the same galactic potential these authors have used in their paper.
This simple expression for the eccentricity eliminates the need for orbital integration, simplifying the process of building the density maps. Compared to the eccentricity calculated by \citet{Casagrande+2011} for the GCS stars, we found that, for 98\% of them, the eccentricity calculated by this expression differs from the one given in the catalog by less then 0.01, which we consider to be good enough for our purposes.
To exemplify how drastically the probability $p(e|V,t)$ may change for different values of $V$ and same age $t$ we present in Figure \ref{fig:dens_map} the density map in the $eV$ plane for the age of 4 Gyr. Also in Figure \ref{fig:dens_map}, the probability functions $p(e|V,t)$, for assigned values of $V$, are represented.
Considering the known values of $e$, $V$ and $W$ for a star, the probabilities $p(e|V,t)$, $p(V,t)$ and $p(W,t)$ are then calculated for different ages $t$ and the results are used to build the probability density function for the stellar age. While $p(e|V,t)$ is obtained through the density maps, $p(V|t)$ and $p(W|t)$
are obtained approximating the distributions as independent Gaussians and using the relations between age and the Gaussian parameters found in Section \ref{sec:method_UVW}:
\begin{equation}
\label{eq:prob_t_method2}
p(V|t) \cdot p(W|t) = \frac{1}{2\pi} \, \mathlarger{\mathlarger{\prod_{\mathsmaller{\mathsmaller{k = V, W}}}}}\left[\frac{1}{\sigma_k(t)} \, \exp{\left( - \frac{\left[k + k_{\sun}(t)\right]^2}{2\,\sigma_k^2(t)} \right)}\right]
\end{equation}
\subsection[Method eUW]{Method $\boldsymbol{eUW}$}
Method $eUW$ is essentially the same as Method $eVW$, but for this method, we are considering the $U$ velocity component instead of the $V$ component. In this case, the density maps needed are the ones in the $eU$ plane. The density maps were build from the same simulated data used for the ones described in Method $eVW$. The Bayesian equation that gives the age probability from the kinematical parameters is
\begin{equation}
p(t|e,U,W) = p(e|U,t) \, p(U|t) \, p(W|t) \, p(t) \,.
\end{equation}
As before, we obtain the age pdf calculating the probability $p(t|e,U,W)$ for different ages. In each case, $p(e|U,t)$ is obtained through the density maps and $p(U|t)$ and $p(W|t)$ are obtained approximating gaussians and using the relations found for the parameters in Section \ref{sec:method_UVW}:
\begin{equation}
\label{eq:prob_t_method3}
p(U|t) \cdot p(W|t) = \frac{1}{2\pi} \, \mathlarger{\mathlarger{\prod_{\mathsmaller{\mathsmaller{k = U, W}}}}}\left[\frac{1}{\sigma_k(t)} \, \exp{\left( - \frac{\left[k + k_{\sun}\right]^2}{2\,\sigma_k^2(t)} \right)}\right]
\end{equation}
\section{Discussion}
\label{sec:discussion}
We have applied the methods described in Section \ref{sec:methods} to derive the pdf for the age of all $9\,102$ stars from the sample defined in Section \ref{sec:sample}. Nine out of these stars could not have their ages determined by Methods $eVW$ and $eUW$ because their velocities were higher than the range we considered while building the density maps (i.e., $\vert U \vert \, \mathrm{or} \, \vert V \vert > 200 \, \mathrm{km} \, \mathrm{s}^{-1}$). As most of these stars also have high $W$ velocity, they should belong to other components of the Galaxy rather than the thin disk and may not affect our conclusions.
While it is best to use all information available in the posterior age pdf, it is often necessary to use a single age estimator in order to study the relation of the stellar parameters and age. As our first goal is to compare the results obtained from the kinematical method to those obtained from the isochronal one, we use the most likely age $t_{\mathrm{ML}}$ and expected age $t_{\mathrm{E}}$ obtained from the pdf to characterize single ages:
\begin{subequations}
\begin{align}
t_{\mathrm{ML}} &= \argmax_t f(t) \,, \\
t_{\mathrm{E}} &= \int_{-\infty}^{\infty} \! \! \! \! t \, f(t) \, \mathrm{d}t \,,
\end{align}
\end{subequations}
where $f(t)$ represents the posterior probability density function $p(t|U,V,W)$, $p(t|e,V,W)$ or $p(t|e,U,W)$, respectively for Methods $UVW$, $eVW$ and $eUW$.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/pdf_examples2_bw_edited.pdf}
\caption{Probability density function obtained by Method $UVW$ (solid line), Method $eVW$ (dashed line) and Method $eUW$ (dotted line), for four representative stars. Panel a) HD 1101, which has velocities $(U,V,W) = (-12, -23, -6) \, \mathrm{km} \, \mathrm{s}^{-1}$ and eccentricity $e = 0.07$, b) HD 1343, for which the velocities are $(U,V,W) = (-22, 16, 3) \, \mathrm{km} \, \mathrm{s}^{-1}$ and the eccentricity is $e = 0.09$, c) HD 852, that has $(U,V,W) = (-58, 39, -9) \, \mathrm{km} \, \mathrm{s}^{-1}$ and $e = 0.22$ and d) HD 12387, with velocities $(U,V,W) = (-3, -88, 60) \, \mathrm{km} \, \mathrm{s}^{-1}$ and eccentricity $e = 0.32$. The vertical marks at the top axis represents the most likely age for each method and those at bottom axis represents the expected ages.}
\label{fig:pdf_example}
\end{figure}
Figure \ref{fig:pdf_example} shows examples of pdfs obtained for four different stars, using the three methods previously described. Also plotted are the most likely age (upper axis lines) and the expected age obtained for each method (lower axis lines). These stars were selected as examples of the different situations one can find when calculating the pdf. In panel a), we can see that there is no defined lower limit for the stellar age, therefore, the most probable age for this star is defined as the lowest age considered in this work (0.1 Gyr). Panel b) represents the case for which both the expected age, and the most likely age can be properly defined. In the case of Panel c), even though a most likely age can be obtained, the expected age will be biased towards lower ages because of the truncation our prior $p(t)$ imposes on the pdf. Panel d) represents a case similar to Panel a), but in this case the limitation is caused by the fact that the most likely age would be greater than the age of the universe.
Since removing all stars whose age pdfs happens to be the cases a), c) or d) would seriously diminish our sample, we have chosen instead to work with a combination of the expected age ($t_{\mathrm{E}}$) and the most likely age ($t_{\mathrm{ML}}$). This way, the sample is less affected by the bias discussed above as in most cases at least one of the ages is well defined. We then define the kinematical age $t_{\mathrm{kin}}$ as a weighted average between $t_{\mathrm{ML}}$ and $t_{\mathrm{E}}$. We found that the expression giving the most similar results as isochronal ages is:
\begin{equation}
\label{eq:t_kin_definition}
t_{\mathrm{kin}} = \frac{3\,t_{\mathrm{ML}} + t_{\mathrm{E}}}{4} \mathrm{.}
\end{equation}
To tabulate the results, we have calculated from each stellar age pdf the expected age ($t^{(i)}_{\mathrm{E}}$), the most likely age ($t^{(i)}_{\mathrm{ML}}$) and the ages corresponding to the $2.5\%$, $16\%$, $50\%$, $84\%$ and $97.5\%$ percentiles, respectively $t^{(i)}_{2.5}$, $t^{(i)}_{16}$, $t^{(i)}_{50}$, $t^{(i)}_{84}$ and $t^{(i)}_{97.5}$, where $i = 1, 2, 3$ designates the method used to derive the age pdf, respectively methods $UVW$, $eVW$ and $eUW$. The results for the first 75 sample stars are shown in Table \ref{tab:results50}. The full data sample is only available in electronic format.
\subsection{Uncertainties for individual ages}
\label{sec:uncertainties}
Due to the statistical nature of this method, uncertainties arise from two different sources: (i) the observational uncertainties that affects the $UVW$ values and, consequently, the resulting pdf and (ii) uncertainties caused by our lack of knowledge of the exact value of the Solar peculiar velocity. We analyze how the latter affects the results by changing the Solar motion and, as the former depends on the quality of the data, we investigate how the results change when considering improved observational precision.
\subsubsection{Uncertainties due to statistical nature of the Method}
\begin{figure}
\centering
\includegraphics[scale=1]{Images/delta_distribution_bw_edited.pdf}
\caption{Distribution of individual uncertainties for the ages obtained by Method $UVW$ (solid line), Method $eVW$ (dashed line) and Method $eUW$ (dotted line) using Equation \ref{eq:delta_t_definition}. All methods are very similar and the $\delta_t$ distributions peak at about 3.25 Gyr, with no uncertainty higher than 4 Gyr.}
\label{fig:dist_delta}
\end{figure}
To estimate the uncertainties, related to the pdf spread, for the ages obtained for single stars through Equation \ref{eq:t_kin_definition}, we have used an expression involving the ages of the percentiles described above. The expression for the uncertainty was chosen in a way that, for a Gaussian distribution, $\delta_t$ would correspond to 1 sigma, and is given by
\begin{equation}
\label{eq:delta_t_definition}
\delta_t = \frac{1}{4} \left[ (t_{84} - t_{16}) + \frac{t_{97.5} - t_{2.5}}{2} \right] \, \mathrm{.}
\end{equation}
This definition is a measure of the spread that takes into account both the dispersion closer to the median age (from $t_{84}$ and $t_{16}$) and the dispersion at the tails of the distribution (from $t_{2.5}$ and $t_{97.5}$). It is designed to allow a direct comparison between the spread of the calculated kinematical method and original isochronal pdf obtained by \citet{Casagrande+2011}, whose data contains the data for these percentiles. It's important to keep in mind that, although a valid measure of the spread, it's arbitrary and not always corresponds to the Gaussian $1\sigma$, as the pdf can be very different from a normal distribution.
The distribution of individual uncertainties for each of the three methods is represented in Figure \ref{fig:dist_delta}. It can be seen that, concerning individual uncertainties, the methods presented here are very similar to each other. Figure \ref{fig:dist_delta} also shows that, for individual stars, uncertainties are very high. The mean uncertainty for Methods $UVW$ and $eVW$ is 3.1 Gyr, and for Method $eUW$, 3.0 Gyr. For comparison, the mean uncertainty for isochronal ages, defined in a similar way, would be about 0.7 Gyr for this sub-sample of the GCS. This means the kinematical ages are useful for individual stars only when the isochronal method is certain to result in extremely high uncertainties, as is the case for M dwarfs, or as an independent age indicator that can complement other methods.
\subsubsection{Impact of observational uncertainties}
In order to understand how the observational uncertainties that leads to the calculation of the $UVW$ velocities affect the obtained pdf, we have performed a set of Monte Carlo simulations and analyzed the effect of this in the calculated $t_\mathrm{ML}$ and $t_\mathrm{E}$.
For each star, the parallax, proper motions and radial velocities were re-sampled 1000 times, considering Gaussian individual errors, and the $UVW$ velocities were recalculated in each case.
The top panels of Figure \ref{fig:pdf_monte_carlo} shows the pdfs obtained for each of the 1000 Monte Carlo simulations (grey lines) and the original pdf (black line) for three representative star: HD 3598 (left) which has uncertainties considered small for the GCS survey ($\sigma_\pi = 0.53$ mas, $\sigma_\mu = 1.0$ mas/year and $\sigma_{r_v} = 0.3 \, \, \mathrm{km} \, \mathrm{s}^{-1}$); HD 578 (middle) which has uncertainties representative of the most stars in the survey ($\sigma_\pi = 0.66$ mas, $\sigma_\mu = 2.0$ mas/year and $\sigma_{r_v} = 0.4 \, \, \mathrm{km} \, \mathrm{s}^{-1}$); and HD 180748 (right) which has large uncertainties compared to other stars in the survey ($\sigma_\pi = 1.8$ mas, $\sigma_\mu = 4.0$ mas/year and $\sigma_{r_v} = 3 \, \, \mathrm{km} \, \mathrm{s}^{-1}$). Also represented in each plot, is the distribution of the obtained $t_\mathrm{ML}$ (dark histograms) and $t_\mathrm{E}$ (light grey histograms). The dispersions of the histograms are: $\sigma_{t_\mathrm{ML}} = 0.26$ Gyr and $\sigma_{t_\mathrm{E}} = 0.15$ for HD 3598; $\sigma_{t_\mathrm{ML}} = 0.39$ Gyr and $\sigma_{t_\mathrm{E}} = 0.27$ for HD 578; and $\sigma_{t_\mathrm{ML}} = 1.25$ Gyr and $\sigma_{t_\mathrm{E}} = 0.61$ Gyr for HD 180748.
It's clear that, the higher the observational uncertainties are, the greater is the uncertainties in the calculated age point estimator. For the whole sample, considering the GCS uncertainties, we obtain on average $\sigma_{t_\mathrm{ML}} = 0.47$ and $\sigma_{t_\mathrm{E}} = 0.28$. Therefore, considering the observational uncertainties of the GCS, the expected age is less affected by uncertainties than the most likely age.
\begin{figure}
\centering
\includegraphics[scale=0.52]{Images/pdf_variations_edited.png}
\caption{Effects of the observational uncertainties on the pdf considering the GCS uncertainties (top) and Gaia's predicted uncertainties (bottom) for the stars HD 3598 (left), HD 507 (middle) and HD 180748 (right). The black solid line represents the original pdf, while the grey lines are the results of 1000 Monte Carlo simulations. The dark grey histogram represents the distribution of most-likely ages and the light grey histogram, the distribution of expected ages.}
\label{fig:pdf_monte_carlo}
\end{figure}
Also, to verify how much the upcoming Gaia data will improve the results, we perform another set of Monte Carlo simulations considering the uncertainties that Gaia aims to achieve by the end of the mission \citep{Gaia2005}: $\sigma_\pi = 25 \cdot 10^{-3}$ mas, $\sigma_\mu = 13.15 \cdot 10^{-3}$ mas/year and $\sigma_{r_v} = 1 \, \, \mathrm{km} \, \mathrm{s}^{-1}$. The lower panels of Figure \ref{fig:pdf_monte_carlo} shows the obtained pdfs for the stars HD 3598 (left), HD 578 (middle) HD 180748 (right). As can be seen, the quality of the expected Gaia data practically eliminates the intrinsic uncertainty of the method due to observational uncertainties. On average, for the whole sample, we obtain $\sigma_{t_\mathrm{ML}} = 0.09$ Gyr and $\sigma_{t_\mathrm{E}} = 0.07$ Gyr.
\subsubsection{Effect of changing the Solar peculiar velocity}
\label{sec:uncertainties-Solar-motion}
We have also investigated the impact of changing the Solar peculiar velocity in the obtained pdfs and consequently in the estimated values of $t_\mathrm{ML}$ and $t_\mathrm{E}$. Figure \ref{fig:pdf_solar_velocity} shows the pdf obtained for the solar velocity obtained in this work (black line) and the pdfs obtained using the other values of solar velocity shown in Table \ref{tab:sun_uvw} (grey lines). We see that, considering the GCS uncertainties, the results are less affected by changing the Solar peculiar velocity than by the observational uncertainties (which will no longer be the case for the data quality expected by Gaia).
Compared to the Solar velocity of this work, the standard deviation of the age differences of the whole sample are: $\sigma_{t_\mathrm{ML}} = 0.26$ Gyr and $\sigma_{t_\mathrm{E}} = 0.18$ Gyr for the Solar peculiar velocity of F\&A14; $\sigma_{t_\mathrm{ML}} = 0.25$ Gyr and $\sigma_{t_\mathrm{E}} = 0.18$ Gyr for B\&B14; $\sigma_{t_\mathrm{ML}} = 0.21$ Gyr and $\sigma_{t_\mathrm{E}} = 0.14 $ for Co{\c s}+11, $\sigma_{t_\mathrm{ML}} = 0.08$ Gyr and $\sigma_{t_\mathrm{E}} = 0.06 $ for Sch+10 and $\sigma_{t_\mathrm{ML}} = 0.52$ Gyr and $\sigma_{t_\mathrm{E}} = 0.31 $ for Kov+09.
\begin{figure}
\centering
\includegraphics[scale=0.52]{Images/change_solar_motion.pdf}
\caption{Effects of changing the Solar peculiar velocity on the calculated pdfs of the stars HD 3598 (left), HD 507 (middle) and HD 180748 (rigth). The black line represents the pdf obtained using the Solar motion derived in this work, while the grey lines represents the pdfs obtained using the Solar velocities displayed in Table \ref{tab:sun_uvw}.}
\label{fig:pdf_solar_velocity}
\end{figure}
\subsection{Comparison between different methods}
Our method uses the isochronal ages derived from the Padova grid by \citet{Casagrande+2011} to fit the relations between the kinematical parameters and age, which are then used in Equations \ref{eq:v1v2v3}, \ref{eq:prob_t_method1}, \ref{eq:prob_t_method2} and \ref{eq:prob_t_method3}. To obtain a more independent comparison, we compare our derived kinematical ages against those derived by \citet{Casagrande+2011} using the BaSTI grid \footnote{While using the BaSTI ages for comparison is certainly better than using the Padova ages (which were used to fit the kinematical equations) they can't be seeing as fully independent age estimators. One of the criteria used to select the stars with well determined ages was that Padova and BaSTI ages had to agree within 1 Gyr.}. As these isochronal ages are also derived from a probability density function, we defined individual isochronal ages in the same way as before:
\begin{equation}
\label{eq:t_iso_definition}
t_{\mathrm{iso}} = \frac{3\,t^{(\mathrm{iso})}_{\mathrm{ML}} + t^{(\mathrm{iso})}_{\mathrm{E}}}{4} \, \mathrm{,}
\end{equation}
where $t^{(\mathrm{iso})}_{\mathrm{ML}}$ is the most likely age and $t^{(\mathrm{iso})}_{\mathrm{E}}$ is the expected age derived from the age pdf.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/Delta_iso_kin_distributions_bw_edited.pdf}
\caption{Distribution of difference between isochronal age and kinematical age, obtained by Method $UVW$ (left), Method $eVW$ (middle) and Method $eUW$ (right). The distributions are similar for all three methods and peak at zero, showing that the ages display a tendency of good agreement. The spreads can be explained by the individual uncertainties of both isochronal and kinematical methods.}
\label{fig:Delta_iso_kin}
\end{figure}
We plotted in Figure \ref{fig:Delta_iso_kin} the distributions of the differences between the isochronal and the kinematical ages derived by the three Methods. The behaviour of all distributions is similar, the peak is close to zero, meaning that the most common case is an agreement between isochronal and kinematical ages; the median of the distributions is also close to zero, being $-$0.33 for the case of Method $UVW$ (left), $-$0.34 for Method $eVW$ (middle) and $-$0.24 for Method $eUW$ (right). The spread in the distribution is explained by the high uncertainties of the kinematical methods, coupled with the also considerable uncertainties of the isochronal method.
Also noticeable is the long tail towards negative values that appears in all distributions, meaning that, in some cases, kinematical ages might be overestimated. This can be explained by the large spread in the pdf, which pushes the expected age towards the center of the age interval. As there are more young than old stars in the sample, there are more stars having their expected age overestimated than underestimated. The reason we chose to give less weight to the expected age in our kinematical age definition, Eq. \ref{eq:t_kin_definition}, was to reduce this effect.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/Delta_kin_kin_distributions_bw_edited.pdf}
\caption{Distributions of differences between ages obtained by the different kinematical methods. In all cases the peak of the distribution is close to zero, and the differences rarely exceed 1 Gyr, showing good agreement between the methods.}
\label{fig:Delta_kin_kin}
\end{figure}
We have also compared the ages obtained by the different methods present here, as can be seen in Figure \ref{fig:Delta_kin_kin}. There is very good agreement between all Methods as the differences rarely exceed 1 Gyr. The peak of the differences distributions is also very close to zero, the mean being 0.02 in the case of Methods $UVW$ and $eVW$, 0.14 for Methods $UVW$ and $eUW$ and 0.12 for Methods $eVW$ and $eUW$.
\subsection{Age distribution}
As the individual uncertainties $\delta_t$ are too high we only recommend this analysis for individual stars as an independent age indicator or when other precise methods are not available. Even though, the kinematical age method may be useful when the properties of large samples are considered. One of its applications is to derive the age distribution of stars in a given sample. Figure \ref{fig:age_dist} shows the distributions of stellar ages obtained by Method $UVW$ (top), Method $eVW$ (middle) and Method $eUW$ (bottom). Also, for comparison, we plotted the isochronal age distribution in all panels. In all cases, especially for Methods $eVW$ and $eUW$, the derived age distribution agrees very well to the isochronal case.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/age_distribution_bw_edited.pdf}
\caption{Distribution of ages obtained by Method $UVW$ (top), Method $eVW$ (middle) and Method $eUW$ (bottom). For comparison, the distribution of isochronal ages is plotted in all panels (dashed line). As can be seen, the ages distributions obtained by the kinematical methods are very similar to the ones obtained by the isochronal case. The agreement is even better for the cases that include eccentricity.}
\label{fig:age_dist}
\end{figure}
The age distribution in a sample is an important parameter, closely related to the star formation history in the Galaxy. The best stars to observe for this task are the ones that evolves slowly enough so that those born in the early phases of our Galaxy formation are still around. While slow evolution is what makes these stars live long enough to trace the Galactic history, it also makes it harder to obtain the stellar ages through their internal properties since they slowly change with time. This slow internal evolution does not affect the evolution of the parameters used by the kinematical method (spatial velocities and eccentricity), which varies with time in the same manner for star of different masses. Therefore, the kinematical method has a role in helping the understanding of important aspects of the evolution of our Galaxy.
In this work, we simply compare the kinematical age distribution with the isochronal age distribution. This distribution cannot be directly interpreted as a star formation history because the selection of stars for the Geneva-Copenhagen Survey \citep{Nordstrom+2004} was based in photometric cuts, which favours certain regions of the $T_\mathrm{eff}-M_{\mathrm{bol}}-[\mathrm{Me}/\mathrm{H}]$ space, directly biasing the age distribution. To translate the age distribution to star formation history, a correction for this bias must be applied, which is beyond the scope of this work.
\subsection{Age-metallicity relation}
The chemical evolution of the Galaxy remains largely unknown. Studies directly relating chemical abundances of different elements to specific epochs are still restricted to very few stars \citep{Spina+2016}. Therefore, to understand this evolution with the available data, we have to rely on models and observational constraints.
The first models considered a simple closed box scenario \citep{Talbot+Arnett1971} in which stars expel their material enriching the interstellar medium, causing newborn stars to be more metal rich than previous generations. The metallicity distribution predicted for present-day stars from this model does not agree well with observations \citep[e.g. the G-dwarf problem][]{Schmidt1963, Pagel+Patchett1975, Wyse+Gilmore1995, Rocha-Pinto+Maciel1996, Haywood2001, Nordstrom+2004}, indicating that more complex processes are involved. When inflows and outflows of gas \citep{Larson1972, Hartwick1976, Schindler+Diaferio2008}, and radial stellar migration \citep{Wielen+1996, Sellwood+Binney2002, Minchev+2013} are considered, the agreement between model predictions and observations is remarkably enhanced.
Considering the complexity of the models, more observational constraints are required. One property that can elucidate aspects of the Galactic chemical evolution is how the overall metallicity of stars changes with time, the so called age-metallicity relation (AMR). The most critical part of obtaining the AMR is determining stellar ages for low mass stars, especially the ones that can live enough to tell us about the early history of the Galaxy. The feature of the AMR that models must be able to explain is the low \citep[or none at all, see][]{Edvardsson+1993, Nordstrom+2004} increase of metallicity with time and the increase of the metallicity dispersion with age \citep{Casagrande+2011}.
\begin{figure}
\centering
\includegraphics[scale=1]{Images/AMR_bw_edited.pdf}
\caption{Relation between the mean metallicity and the ages, lower than 7 Gyr, obtained through isochronal method (left) and kinematical methods (right). The behaviour of the relation is similar in both cases, suggesting that the kinematical method is useful for finding relations between chemical features and age.}
\label{fig:AMR}
\end{figure}
To check whether the kinematical method can be useful to derive the age-metallicity relation, we divided our stellar sample by bins of age 0.5 Gyr wide and calculated the mean $[\mathrm{Fe}/\mathrm{H}]$ for each group. The same was done for the isochronal ages for comparison. As groups older than 7 Gyr would have very few stars, we have restricted our analysis for ages lower than this. The results are plotted in Figure \ref{fig:AMR}. This raw analysis of the data shows a decay of metallicity with age both for the isochronal and the kinematical cases. The relation is slightly flatter for the kinematical ages. This is probably caused by the higher uncertainties for individual ages, which causes a mixture between stars that should, in reality, belong to other age bins.
As there is also a bias caused by selection criteria, we do not interpret the result as a true age-metallicity relation in the Solar Neighbourhood. In this work, we are just interested in showing that the results obtained from the kinematical ages are similar to those obtained using isochronal ones, therefore concluding that the kinematical method might be used for this task for samples of stars for which ages cannot be determined by isochrones.
\subsection{The age of Trappist-1}
\label{sec:trappist1}
Trappist-1 is the system that hosts the largest number of known earth-size planets in its habitable zone \citep{Gillon+2016}. Recent results show that the star harbors 5 planets of sizes similar to that of the Earth (planets b, c, e, f and g) and 2 planets slightly smaller ($\approx0.75$ $R_{\mathrm{\oplus}}$, planets d and h; \citealp{Gillon+2017}). The planets e, f and g are in the habitable zone and, therefore, could harbour water oceans.
After its discovery in 2016 and confirmation of 3 earth-size planets in the habitable zone in 2017, this star received a lot of attention and has been the central investigation of several studies consisting of further analysis of light curves \citep[e.g.][]{Wang+2017}, dynamical investigations \citep[e.g.][]{Tamayo+2017, Quarles+2017}, and habitability plausibility \citep[e.g.][]{OMalley+Kaltenegger2017, Wolf2017}. Despite of this, one very important information is still missing: the stellar age.
The reason for the lack of age information is the fact that Trappist-1 has a mass of only 8\% the mass of the Sun. It causes the stellar evolution to be too slow, making the atmospheric parameters practically unchanged since the stellar birth. Because of this, most of the traditional age dating methods can not be applied. Since the kinematical method is expected to be independent from the internal evolution of the star and depends only on its orbital evolution, we can apply the method to obtain the age p.d.f. for the star and estimate its expected value and uncertainty.
From the stellar coordinates, proper motions and parallax \citep{Costa+2006} and radial velocity \citep{Burgasser+2015}, we derived its heliocentric peculiar velocities using the method described by \citet{Johnson+Soderblom1987}. The obtained values were $U = -43.24 \, \mathrm{km} \, \mathrm{s}^{-1}$, $V = -66.25 \, \mathrm{km} \, \mathrm{s}^{-1}$ and $W = 13.87 \, \mathrm{km} \, \mathrm{s}^{-1}$. The eccentricity was calculated using Equation \ref{eq:ecc} and found to be 0.27. We used these velocities and eccentricity to apply Methods UVW, eVW and eUW to derive the age pdfs represented in Figure \ref{fig:Trappist_age}.
\begin{figure}
\centering
\includegraphics[scale=0.75]{Images/Trappist_age.pdf}
\caption{Age pdfs obtained for the Trappist-1 star using Method UVW (solid line), Method eVW (dashed line) and Method eUW (dotted line). The vertical lines in the top axis represent the most-likely age and the lines in the bottom axis, the expected age.}
\label{fig:Trappist_age}
\end{figure}
Figure \ref{fig:Trappist_age} shows that the most likely age (vertical marks at the top axis) and the expected age (vertical marks at the lower axis) for the three different methods. The most-likely ages for methods $UVW$, $eVW$ and $eUW$ are respectivelly 13.4, 13.13 and 13.06 Gyr and the expected ages are respectivelly 9.78, 9.86 and 9.73, showing good agreement between the three kinematical methods. We then apply Equation \ref{eq:t_kin_definition} to calculate the defined kinematical ages, and obtain 12.50 for Method $UVW$, 12.31 for Method $eVW$ and 12.23 for Method $eUW$.
In order to estimate upper and lower limits, we also calculate the 16\% and 84\% percentile ages, and obtained $t_\mathrm{kin}^{(UVW)} = 12.50^{+0.29}_{-6.23}$, $t_\mathrm{kin}^{(eVW)} = 12.31^{+0.53}_{-6.05}$, $t_\mathrm{kin}^{(eUW)} = 12.23^{+0.56}_{-6.22}$. In this case the kinematical Method only provides a good upper-limit for the age and has a large tail towards lower ages, as can also be seen in Figure \ref{fig:Trappist_age}. The results allows us to conclude that the age of the system most-likelly lies between $\approx$6--12.5 Gyrs. Considering the observational uncertainties, the results are in good agreement with the age derived by \citet{Burgasser+Mamajek2017}: $7.6 \pm 2.2$ Gyr. In this case, the authors have considered several age constrains: the stellar cmd, average density, lithium absorption, surface gravity, rotation, magnetic activity, as well as kinematics.
We believe the derived age further increases the astrobiological interest for the system: it not only contains 7 planets with sizes similar to the earth (3 of which are in the habitable zone) but is also old enough for life as we know it to develop and evolve. Nevertheless an important remark must be made: as the system is extremely compact (the highest period being of 18.77 days; \citealp{Luger+2017}) the planets are all expected to be tidally locked \citep{Gillon+2017}, causing the temperatures to vary significantly between the face towards the star and the opposite one, which would constitute another barrier for the formation of life.
\section{Conclusions}
\label{sec:conclusions}
We showed how a probability density function (pdf) for the age of a star may be obtained from its spatial velocity components $U$, $V$ and $W$ and also from other kinematical parameters like the eccentricity. We characterize individual ages from the pdf using the most likely and the expected ages. This individual age estimation has uncertainties of about 3 Gyr, which, although higher than classical methods, may be the best estimate available for most very low mass stars.
The methods are based on the growth of velocity dispersion with age. We adopted a relation between age and velocity dispersion as $\sigma_i(t) = b_i \, t^{a_i}$. The parameters $b_i$ and $a_i$ were found using a sub-sample of the Geneva-Copenhagen Survey, which contains isochronal ages derived by \citep{Casagrande+2011}. We show that other parameters that define the velocity distribution are also a function of stellar age: the $V$ component of the Solar motion ($V'_{\sun}$) and the vertex deviation ($\ell_v$). From this sample, we obtained the solar motion to be $(U,V,W)_{\sun} = (9.8\pm0.3, 12.5\pm0.9, 7.2\pm0.2)_{\sun} \, \mathrm{km} \, \mathrm{s}^{-1}$. Using the example of eccentricity, we show how other kinematical parameters may be included in the analysis, provided the density of stars in the kinematical parameter-velocity space is known for all ages.
We applied the Kinematical Method for the stars of the Geneva-Copenhagen survey and compared the results with those obtained by isochronal estimation. We show that differences between the obtained ages peak at zero and, despite a slightly overestimation of kinematical ages, there is no significant bias. This means that, even though uncertainties are large, the kinematical method is useful to derive statistical parameters for groups of stars. A comparison between the kinematical methods based purely on spatial velocity and the ones that include eccentricity shows that there is a very good agreement in the ages derived and the differences rarely exceed 1 Gyr.
We investigated how the results are affected by observational uncertainties and by changing the Solar peculiar velocity. We conclude that, for the GCS uncertainties, the results are more affected by the observational uncertainties than by changes in the Solar peculiar velocity, but it will no longer be the case for the data quality that Gaia aims to achieve.
The distribution of stellar ages obtained through the kinematical methods agrees very well with the one obtained through isochrones, principally for the methods that include the eccentricity. We have not attempted to derive the star formation history from this data, because it would require the consideration of a detailed analysis of the age bias imposed by the sample selected criteria, which is left for future work. However, the agreement between kinematical and isochronal distribution leads us to conclude that the kinematical method may be as useful as the isochronal for this task, with the advantage of being applicable to the numerous very low mass stars.
The relation between metallicity and age, which is an important constraint for chemical evolution models, has also been investigated using the kinematical method. The behaviour of the relation shown for stars younger than 7 Gyr, is similar to the one obtained through isochrones, suggesting the kinematical method can also be used to derive this relation. An interpretation of this relation as a true age-metallicity relation also needs a detailed analysis of bias imprinted by selection criteria.
We showed that the method can also be applied for M stars, as it does not depend on the stellar mass. As an example, we derived the age for the Trappist-1 system, which is the known system that hosts the largest number of earth-size planets (seven, three of which are in the habitable zone). Our analysis of the obtained pdfs indicated an age between $\approx$6--12.5 Gyr.
We intent to explore this method to obtain kinematical ages pdfs for stars in large surveys, such as the Radial Velocity Experiment (RAVE) \citep{Steinmetz+2006} and Gaia \citep{Gaia+2016}, in the cases the other methods cannot be applied or would be unreliable. These pdfs can also be used as an independent age indicator to corroborate the results obtained by other means.
\textit{Acknowledgements}
For providing support with a PhD grant, Almeida-Fernandes F. wants to thank CAPES --- The Brazilian Federal Agency for Support and Evaluation of Graduate Education within the Ministry of Education of Brazil, and for the support, when the author was an undergrad, Almeida-Fernandes F. wants to thank CNPq --- The Brazilian National Council for Scientific and Technological Development.
\bibliographystyle{mnras}
|
\section{Noise Boosting the Expectation-Maximization Algorithm}
We show how carefully chosen and injected noise can speed convergence of the popular expectation-maximization (EM) algorithm.
A general theorem allows \emph{arbitrary} modes of combining signal and noise to improve the speed of parameter estimation.
The result still speeds EM convergence on average at each iteration so long as the injected noise satisfies a positivity condition.
The EM algorithm generalizes maximum-likelihood estimation to the case of missing or corrupted data\cite{dempster-laird-rubin1977,mclachlan-krishnan2007}.
Maximum likelihood maximizes the conditional signal probability density function (pdf) $f(y|\theta)$ for a random signal variable $Y$ given a vector of parameters $\theta$.
It equally maximizes the log-likelihood $\ln f(y | \theta)$ since the logarithm is monotone increasing.
So the maximum--likelihood estimate $\theta_*$ is
\begin{equation}
\theta_*= \argmax{\theta}\ \ln f(y | \theta).
\end{equation}
The parameter vector $\theta$ can contain means or covariances or mixture weights or any other terms that parametrize the pdf $f(y| \theta)$.
The data itself consists of observations or realizations $y$ of the signal random variable $Y$.
The data can be speech samples or image vectors or any type of numerical measurement.
The EM framework allows for missing or hidden data or so-called latent variables.
The random variable $Z$ denotes all such latent variables.
These latent variables can describe unseen states in a hidden Markov model or hidden neurons in a multilayer neural network.
Then $Z$ appears in the log-likelihood $\ln f(y | \theta)$ through the pdf identity $f(y|\theta) = \frac{f(y,z | \theta)}{f(z | y, \theta)}$.
This gives the key EM log-likelihood equality $\ln f(y | \theta) = \ln f(y, z | \theta)- \ln f(z | y, \theta)$.
The EM algorithm estimates the missing information in $Z$ by iteratively maximizing the probability of $Z$ given both the observed data $y$ and the current parameter estimate $\theta_k$ \cite{Hogg2013}.
This involves averaging the log-likelihood $\ln f(y,z | \theta_k)$ over the conditional pdf $f(z|y, \theta_k)$ to form the surrogate likelihood function $Q(\theta | \theta_k)$:
\begin{align}
Q(\theta|\theta_k) &= \E_{Z} \left[ \ln f(y,Z| \theta) \big{|} Y=y, \theta_k \right] \\
&= \int_{\mathcal{Z}} \ln [f(y,z| \theta)] f(z|y,\theta_k) ~\mathrm{d}z.
\end{align}
Then EM's ``ascent property" \cite{dempster-laird-rubin1977} uses Jensen's inequality \cite{cover-thomas91} and the above EM log-likelihood equality to ensure that any $\theta$ that increases the surrogate likelihood function $Q(\theta|\theta_k)$ can only increase the log--likelihood difference $\ln f(y | \theta)- \ln f(y | \theta_k)$: $\ln\frac{f(y|\theta)}{f(y|\theta_k)} \ge Q(\theta | \theta_k) - Q(\theta_k | \theta_k)$.
The result is that EM is a hill-climbing algorithm that can only increase the log-likelihood at each step.
The EM algorithm iteratively climbs a hill of probability or log-likelihood until it reaches the closest peak of maximum likelihood.
The peak or mode corresponds to the locally maximal parameter $\theta_*$.
So the EM algorithm converges to the local likelihood maximum $\theta_*$: $\theta_k \rightarrow \theta_*$.
The EM algorithm halts in practice when its successive estimates $\theta_k$ differ by less than a given tolerance level $\|\theta_k-\theta_{k-1}\| < 10^{-tol}$ or when $|\ln f(y | \theta_k)- \ln f(y | \theta_{k-1})| < \varepsilon$ for some small positive $\varepsilon$.
The EM algorithm generalizes many popular algorithms.
These include the $k$-means clustering algorithm\cite{osoba-kosko2013} used in pattern recognition and big-data analysis, the backpropagation algorithm used to train deep feedforward and convolutional neural networks\cite{audhkhasi2013noise, audhkhasi2014noise, lecun2015}, and the Baum-Welch algorithm used to train hidden Markov models\cite{audhkhasi-osoba-kosko-HMM2013,welch2003}.
But the EM algorithm can converge slowly if the amount of missing data is high or if the number of estimated parameters is large\cite{mclachlan-krishnan2007,tanner1996}.
It can also get stuck at local probability maxima.
Users can run the EM algorithm from several starting points to mitigate the problem of convergence to local maxima.
The Noisy EM (NEM) algorithm~\cite{osoba-kosko2013,osoba-mitaim-kosko2013,osoba-mitaim-kosko2011,osoba-dissertation2013} is a noise-enhanced version of the EM algorithm that carefully selects noise and then \emph{injects} it into the data.
NEM converges faster on average than EM does because on average it takes larger steps up the same hill of probability or of log-likelihood.
NEM never takes shorter steps on average.
The largest noise gains tend to occur in the first few steps.
So NEM enhances the ascent property at each iteration.
This is a type of nonlinear noise benefit or \emph{stochastic resonance}\cite{wiesenfeld1995, bulsara1996, gammaitoni1998, mitaim-kosko1998SR, chapeau-blondeau-rousseau2004, lee2006, kosko2006, mcdonnell2008, patel-kosko-SPL2010, patel-kosko-TSP2011, chen2014, mitaim2014} that does not depend on a threshold \cite{franzke2011}.
NEM injects noise $N$ to the data $Y$ if the noise satisfies the NEM average positivity (nonnegativity) condition:
\begin{equation}
\E_{Y,Z,N|\theta_*} \left[ \ln\left( \frac{f(\phi(Y,N),Z|\theta_k)}{f(Y,Z|\theta_k)} \right) \right] \geq 0\;. \label{eq:NEM-Goal}
\end{equation}
The NEM positivity condition~(\ref{eq:NEM-Goal}) holds when the noise-perturbed likelihood $f(\phi(y,N),z|\theta_k)$ is larger on average than the noiseless likelihood $f(y,z|\theta_k)$ at the $k^{th}$ step of the algorithm~\cite{osoba-mitaim-kosko2013,osoba-dissertation2013}.
This noise-benefit condition for additive noise injection has a simple quadratic form when the data or signal model is a mixture of Gaussian pdfs.
A simple argument gives the intuition behind the NEM positivity condition for additive noise.
This argument holds in much greater generality and underlies much of the theory of noise-boosting both the EM algorithm and Markov chain Monte Carlo algorithms \cite{Franzke2015}.
Consider a noise sample or realization $n$ that makes a signal $y$ more probable: $f(\phi(y,n)|\theta) \geq f(y|\theta)$ for some parameter $\theta$.
The value $y$ is a realization of the signal random variable $Y$.
The value $n$ is a realization of the noise random variable $N$.
Then this pdf inequality holds if and only if $\ln\frac{f(\phi(y,n)|\theta)}{f(y|\theta)} \ge 0$.
Averaging over the signal and noise random variables gives the basic expectation form of the NEM positivity condition.
Averaging implies that the pdf inequality need hold only almost everywhere.
It need not hold on sets of zero probability.
This allows the user to ignore particular values when using continuous probability models.
Particular choices of the signal conditional probability $f(y|\theta)$ can greatly simplify the NEM sufficient condition.
This signal probability is the so-called ``data model" in the EM context of maximum likelihood estimation.
Estimation on Gaussian mixtures data models leads to simple quadratic NEM conditions. An exponential data model leads to an even simpler linear NEM condition.
Theorem 1 presents the generalized form of the NEM Theorem for arbitrary measurable noise injection $\phi(Y,N)$. Corollaries 1 and 2 state the NEM sufficient condition for the special cases of additive and multiplicative injection: $\phi(Y,N) = Y+N$ and $\phi(Y,N) = YN$.
\begin{figure}[h]
\centerline{ \includegraphics[width=3.5in]{./GMM-mNEM-upd.pdf} }
\caption{
{Multiplicative noise benefit when estimating the parameters of a sampled Gaussian mixture model.
The mixture density $f$ equally weighted two Gaussian probability density functions with the same variance of 4: $f(x) = \frac{1}{2}N_1(-2, 4) + \frac{1}{2}N_2(2, 4)$.
The EM algorithm estimated the mixing weights, the means, and the variances of the two Gaussian densities.
Low intensity starting noise decreased the EM convergence time while higher intensity starting noise increased it.
The multiplicative noise had unit mean with different but decaying standard deviations.
The optimal initial noise standard deviation was $\sigma^* = 0.44$.
It gave a $27.6\%~$ speed-up over the noiseless EM algorithm.
Optimal m-NEM needed only 7 iterations on average to converge to the correct mixture parameters while noiseless EM needed 10 iterations on average.
The m-NEM procedure injected multiplicative noise that decayed at an inverse-square rate with the iterations.}
}
\label{fg:GaussNEM}
\end{figure}
Figure~\ref{fg:GaussNEM} shows an EM speed-up of $27.6\%~$ for multiplicative-NEM noise injection in the generic case of a bimodal mixture of two Gaussian pdfs.
Sampling from the mixture corresponds to sampling from two subpopulations that have the same variance but different means.
The task is threefold: Estimate the unknown means of the two mixed Gaussian densities.
Estimate the unknown variances of the mixed densities.
And estimate the unknown mixture weights.
The mixture weights are nonnegative and sum to unity.
The noise-injected EM algorithm estimated all these parameters of the equally weighted two-pdf Gaussian mixture model.
Suppose random variable $X_j$ is Gaussian or normal with mean $\mu_j$ and variance $\sigma_j^2$: $X_j \sim N(\mu_j, \sigma_j^2)$ with pdf $f_j(x|\mu_j, \sigma_j^2)$.
Then the two-mixture density in Figure 1 had the form $f(x) = \alpha f_1(x|\mu_1, \sigma_1^2) + (1 - \alpha) f_2(x|\mu_2, \sigma_2^2) = \frac{1}{2} f_1(x| -2, 4) + \frac{1}{2} f_2(x|2, 4)$.
The data itself came from randomly samples of a Gaussian mixture.
The noise-boosted EM algorithm took on average only 7 iterations to estimate the Gaussian mixture parameters $\alpha, \mu_1, \mu_2, \sigma_1^2, \and \sigma_2^2$ while the noiseless EM algorithm took on average 10 steps.
The optimal initial noise standard deviation was $\sigma_N^* = 0.44$.
The simulations ``cooled" or ``annealed" the noise by multiplying the starting noise standard deviation $\sigma_N$ with the inverse-square term $k^{-2}$ at each iteration $k$.
This gradually shut off the noise injection as we discuss below when we present the details of the n-NEM algorithm.
Ordinary or \emph{blind} noise (not subject to the appropriate NEM condition) only slowed EM convergence.
Blind noise was just noise drawn at random or uniformly from the set of all possible noise.
It was not subject to the NEM condition or to any other condition.
The optimal speed-up using additive noise on the same data model was $30.5\%~$ at an optimal noise power of $\sigma^* = 1.9$. This speed-up was slightly better than the m-NEM speed-up for the same mixture model of two Gaussian pdfs. Figure~\ref{fg:GaussNEM-comp} shows the performance of the additive NEM algorithm on the same model.
\begin{figure}[h]\centerline{ \includegraphics[width=3.5in]{./GMM-NEM-addmult-comp.png} }\caption{{Noise benefit in the same GMM-NEM algorithm using additive noise injection. Low intensity noise decreases convergence time while higher intensity starting noise increases it. The noise decays at an inverse square rate. The optimal initial noise standard deviation is at $\sigma^* = 1.9$ which gives a $30.5\%~$ speed improvement over the regular EM algorithm. This additive noise model results in slightly faster average convergence speed at the optimal noise level than the multiplicative noise model. But a $95\%$-bootstrap confidence interval for the average difference in optimal convergence time is $[-0.45,0.067]$. So the difference in optimal average convergence time is not statistically significant.}}\label{fg:GaussNEM-comp}\end{figure}
A statistical test for the difference in the averaged optimal convergence times found that this difference was not statistically significant at the standard $0.05$ significance level.
Nor was it significant at the $0.10$ or $0.01$ levels.
The hypothesis test for the difference of means gave the very large bootstrap $p$-value (achieved significance level \cite{Hogg2013}) of $0.492$ based on 10,000 bootstraps.
That large $p$-value argues strongly against rejecting the null hypothesis that there was no statistically significant difference in the optimal average convergence times of the additive and multiplicative NEM speed-ups.
A $95\%$-bootstrap confidence interval for the average difference in optimal convergence time was $(-0.44, 0.06)$.
The confidence interval contained zero.
So we cannot reject the null hypothesis that the difference in optimal average convergence times for the two noise-injection modes was statistically insignificant at the $0.05$ level.
Nor can we reject the null hypothesis at the $0.10$ and $0.01$ significance levels because their respective $90\%$ and $99\%$ bootstrap confidence intervals were $(-0.40, 0.02)$ and $(-0.52, 0.13)$.
So there was no statistically significant difference in the noise speed-ups of the additive and mulitplicative cases.
An open and important research question is whether there are general conditions under which one of these noise injection modes outperforms the other.
\section{General Noise Injection for a NEM Benefit}
We next generalize the original proof for additive NEM~\cite{osoba-mitaim-kosko2013,osoba-dissertation2013} to NEM that uses an arbitrary mode of noise injection.
The metrical idea behind the proof remains the same: a noise benefit occurs on average at an iteration if the noisy pdf is closer to the optimal pdf than the noiseless pdf is.
Relative entropy measures the pseudo-distance of a pdf to the optimal pdf in a topological space of pdfs:
\begin{align}
D\left(f(y,z|\theta_*)\Vert f_N(y,z|\theta_k)\right) \leq D\left(f(y,z|\theta_*)\Vert f(y,z|\theta_k)\right)
\end{align}
where
\begin{equation}
f_N(y,z|\theta_k) = f(\phi(y, N),z|\theta_k)
\end{equation}
is the noise-injected pdf.
The literature sometimes refers to the relative entropy as the Kullback-Leibler divergence\cite{cover-thomas91}.
The relative entropy is asymmetric and has the form of an average logarithm
\begin{equation}
D\left(h(u,v)||g(u,v)\right) = \int_{\mathcal{U}} \int_{\mathcal{V}} \ln \left[ \frac{h(u,v)}{g(u,v)} \right] h(u,v) ~\mathrm{d}u~\mathrm{d}v
\end{equation}
for positive pdfs $h$ and $g$ over the same support \cite{cover-thomas91}.
Convergent sums can replace the integrals in the discrete case.
We follow convention in calling the relative entropy a pseudo-metric.
It is technically only a pre-metric because the relative entropy between two pdfs is always nonnegative.
The relative entropy is zero if and only if the two pdfs are equal almost everywhere.
This yields the proof strategy of reducing the relative entropy with respect to the optimal pdf at each iteration $k$.
The key point is that the noise-injection mode $\phi(y, N)$ need be neither addition $\phi(y, N)=y+N$ nor multiplication $\phi(y, N)=yN$.
It can be any measurable function $\phi$ of the data $y$ and the noise $N$.
This generality does not affect the main proofs for a noise benefit.
The above relative entropy inequality is logically equivalent to the EM noise-benefit condition at iteration $k$ if we cast the noise benefit in terms of expectations~\cite{osoba-mitaim-kosko2013}:
\begin{align}
\E \Big[ Q(\theta_*|\theta_*) - Q(\theta_{k}|\theta_*) \Big] &\geq \E \Big[ Q(\theta_*|\theta_*) - Q_N(\theta_{k}|\theta_*) \Big] \label{eq:EQ-NoiseBenefit}
\end{align}
where $Q_N$ is the noise-perturbed surrogate likelihood function
\begin{equation}
Q_N\left( \theta |\theta_k \right) = \E_{Z|Y,\theta_k} \left[ \ln f_N(y,Z| \theta) \right].
\end{equation}
Any noise $N$ that satisfies this EM noise-benefit condition will on average give better parameter estimates at each iteration than will noiseless estimates or those that use blind noise.
The relative-entropy version of the noise-benefit condition allows the same derivation of the generalized NEM condition as in the original case of additive noise.
The result is Theorem 1.
\noindent{\bf{Theorem 1: The Arbitrary-Injection NEM Theorem}}\label{thm:genNEM}\\
Let $\phi(Y,N)$ be an arbitrary mode of combining the signal $Y$ with the noise $N$.
Suppose the average positivity condition holds:
\begin{equation}
\E_{Y,Z,N|\theta_*} \left[ \ln\left( \frac{f(\phi(Y, N),Z|\theta_k)}{f(Y,Z|\theta_k)} \right) \right] \geq 0\;.
\label{eq:genNEM-Goal}
\end{equation}
Then the EM noise benefit
\begin{equation}
\ Q(\theta_{k}|\theta_*) \leq \ Q_N(\theta_{k}|\theta_*)
\end{equation}
holds on average at each iteration $k$:
\begin{multline}
\E_{Y| \theta_k} \Big[ Q \left( \theta_* |\theta_* \right) - Q \left( \theta_k |\theta_* \right) \Big] \geq \\ \E_{N,Y| \theta_k} \Big[ Q \left( \theta_* |\theta_* \right) - Q_N \left( \theta_k |\theta_* \right) \Big] \;.
\end{multline}
\noindent{\bf{Corollary 1: Additive NEM}}\label{thm:aNEM}\\
Suppose the average positivity condition holds for additive noise injection:
\begin{equation}
\E_{Y,Z,N|\theta_*} \left[ \ln\left( \frac{f(Y+N,Z|\theta_k)}{f(Y,Z|\theta_k)} \right) \right] \geq 0\;.
\label{eq:genNEM-Goal}
\end{equation}
Then the EM noise benefit
\begin{equation}
\ Q(\theta_{k}|\theta_*) \leq \ Q_N(\theta_{k}|\theta_*)
\end{equation}
holds on average at each iteration $k$:
\begin{multline}
\E_{Y| \theta_k} \Big[ Q \left( \theta_* |\theta_* \right) - Q \left( \theta_k |\theta_* \right) \Big] \geq \\ \E_{N,Y| \theta_k} \Big[ Q \left( \theta_* |\theta_* \right) - Q_N \left( \theta_k |\theta_* \right) \Big] \;.
\end{multline}
\noindent{\bf{Corollary 2: Multiplicative NEM (m-NEM)}}\label{thm:mNEM}\\
Suppose the average positivity condition holds for multiplicative noise injection:
\begin{equation}
\E_{Y,Z,N|\theta_*} \left[ \ln\left( \frac{f(YN,Z|\theta_k)}{f(Y,Z|\theta_k)} \right) \right] \geq 0\;.
\label{eq:genNEM-Goal}
\end{equation}
Then the EM noise benefit
\begin{equation}
\ Q(\theta_{k}|\theta_*) \leq \ Q_N(\theta_{k}|\theta_*)
\end{equation}
holds on average at each iteration $k$:
\begin{multline}
\E_{Y| \theta_k} \Big[ Q \left( \theta_* |\theta_* \right) - Q \left( \theta_k |\theta_* \right) \Big] \geq \\ \E_{N,Y| \theta_k} \Big[ Q \left( \theta_* |\theta_* \right) - Q_N \left( \theta_k |\theta_* \right) \Big] \;.
\end{multline}
The NEM Theorem and its corollaries give a general method for noise-boosting the EM algorithm.
Theorem 1 implies that on average these NEM variants outperform the noiseless EM algorithm
Algorithm~\ref{algo:m-NEM} gives the {generalized--NEM} algorithm schema.
\noindent The operation \textsc{gNEMNoiseSample($\mathbf{y}, ~k^{-\tau} \sigma_N$) } generates noise samples that satisfy the NEM condition for the current data model.
The noise sampling pdf depends on the vector of random samples $\mathbf{y}$ in the data-generating model.
\begin{algorithm}[H]
\caption{$\hat{\theta}_{gNEM}$ = gen-NEM-Estimate($\mathbf{y}$)}
\label{algo:m-NEM}
\begin{algorithmic}[1]
\REQUIRE $\mathbf{y} = \left( y_1,\ldots, y_M \right)$ : vector of observed incomplete data
\ENSURE $\hat{\theta}_{gNEM}$ : gNEM estimate of parameter $\theta$
\WHILE{($\|\theta_k-\theta_{k-1}\| \geq 10^{-tol}$)}
\STATE $\mathbf{N_S}${\bf -Step:} $\mathbf{n} \leftarrow $ NEMNoiseSample($\mathbf{y}, ~k^{-\tau} \sigma_N$)
\STATE $\mathbf{N_G}${\bf -Step:} $\mathbf{y}_\dagger \leftarrow \phi(\mathbf{y}, \mathbf{n})$
\STATE {\bf E-Step:} $Q \left( \theta |\theta_k \right) \leftarrow\E_{Z|y,\theta_k} \left[ \ln f(\mathbf{y}_\dagger, \mathbf{Z}|\theta) \right] $
\STATE {\bf M-Step:} $\theta_{k+1} \leftarrow \argmax{\theta} \left\{ Q\left( \theta |\theta_k \right) \right\}$
\STATE $k \leftarrow k+1$
\ENDWHILE
\STATE $\hat{\theta}_{gNEM} \leftarrow \theta_k$
\end{algorithmic}
\end{algorithm}
\noindent The E-Step takes a conditional expectation of a function of the noisy data samples $\mathbf{y}_\dagger$ given the noiseless data samples $\mathbf{y}$.
A deterministic decay factor $k^{-\tau}$ scaled the noise on the $k^{th}$ iteration.
It did this by replacing the fixed standard deviation $\sigma_N$ of the noise with the weighted standard deviation $k^{-\tau}\sigma_N$.
So the NEM noise had slightly smaller standard deviation with each successive iteration.
$\tau$ was the noise decay rate~\cite{osoba-mitaim-kosko2013}.
The decay factor drove the noise $N_k$ to zero as the iteration step $k$ increased.
This eventually shut off the noise injection.
We found that the value $\tau = 2$ worked best in the simulations and thus we used an inverse-square scaling $k^{-2}$.
The inverse-square decay factor reduced the NEM estimator's jitter around its final value.
This was important because the EM algorithm converges to fixed-points.
Excessive estimator jitter prolongs convergence time even when the jitter occurs near the final solution.
Our simulations used the inverse-square and thus polynomial decay factor instead of the logarithmic cooling schedules found in annealing applications \cite{kirkpatrick-gelatt-vecchi1983, cerny1985, geman-hwang1986, hajek1988, kosko-nnfs}.
\section{Noise-Boosting Parameter Estimation for Gaussian Mixture Models}
Corollaries 1 and 2 from \cite{osoba-mitaim-kosko2013} lead to NEM conditions for GMMs because the noise condition applies to each mixed normal pdf in the mixture.
We state and prove the NEM GMM results for additive and multiplicative noise injection.
The resulting quadratic NEM conditions depend only on the Gaussian means and not on their variances.
A finite mixture model~\cite{redner-walker1984,mclachlan-peel2004,hogg-tanis2006, Hogg2013} is a convex combination of a finite number of similar pdfs.
So we can view a mixture as a convex combination of a finite set of similar sub-populations.
The sub-population pdfs are similar in the sense that they all come from the same parametric family.
Mixture models apply to a wide range of statistical problems in pattern recognition and machine intelligence.
A Gaussian mixture consists of convex-weighted normal pdfs.
The EM algorithm estimates the mixture weights as well as the means and variances of each normal pdf.
The GMM is by far the most common mixture model in practice \cite{gershenfeld1999}.
The EM algorithm offers a standard way to estimate the parameters of a mixture model.
Let $Y$ be the observed mixed random variable.
Let $K$ be the number of sub-populations.
Let $Z \in \left\{1,\ldots,K\right\}$ be the hidden sub-population index random variable.
The convex population mixing proportions $\alpha_1,\ldots, \alpha_K$ define a discrete pdf for $Z$: $P(Z=j) = \alpha_j$.
The pdf $f(y|Z=j,\theta_j)$ is the pdf of the $j^{th}$ sub-population where $\theta_1,\ldots, \theta_K $ are the pdf parameters for each sub-population.
The sub-population parameter $\theta_j$ can represent the mean or variance of a normal pdf or both.
It can represent any number of quantities that parametrize the pdf.
Let $\Theta$ denote the vector of all model parameters: $\Theta = \left\{\alpha_1, \ldots , \alpha_K, \theta_1, \ldots , \theta_K\right\}$.
The mixing weights $\alpha_1, \ldots , \alpha_K$ are convex coefficients.
So they are nonnegative and add to unity.
And thus they define a discrete probability distribution.
The joint pdf $f(y, z|\Theta)$ is
\begin{equation}
f(y,z|\Theta) =\sum_{j=1}^K \alpha_j ~f(y|j,\theta_j) ~\delta[z-j] \;
\end{equation}
where $\delta[z-j] = 1$ if $z = j$ and $\delta[z-j] = 0$ otherwise.
The $K$ pdfs $f(y|j,\theta_j)$ are the mixed pdfs in the finite mixture.
Their structure determines the sufficient condition for a NEM noise benefit.
EM algorithms for finite mixture models estimate $\Theta$ using the sub-population index $Z$ as the latent variable.
The GMM-EM algorithm uses the following $Q$-function
\begin{align}
Q(\Theta|\Theta_k) =& \E_{Z|y,\Theta_k}[\ln f(y,Z|\Theta)] \\
=& \sum_{j} \ln[\alpha_j f(y|j,\theta_j)] p_Z(j|y,\Theta_k).
\label{eq:Q_Mixture}
\end{align}
We can now state and prove a sufficient condition involving the mixed pdfs $f(y|j,\theta_j)$ for an m-NEM noise benefit in a Gaussian mixture model.
The condition has a simple quadratic form.
\vspace{8pt}
\noindent{\bf{Corollary: NEM Condition for Gaussian Mixture Models}}\label{cor:GMM-NEM} \\
Suppose $Y|_{Z=j} \sim {\cal N}(\mu_j,\sigma^2_j)$ and so $f(y|j,\theta)$ is a normal pdf. Then the pointwise pdf noise benefit for additive noise
\begin{align}
f(y + n|\theta) \geq f(y|\theta)
\end{align}
holds if and only if
\begin{align}
n^2 \leq 2n\left(\mu_j-y\right) \;.
\label{eq:GaussMLECondn}
\end{align}
\noindent{\bf{Corollary: m-NEM Condition for Gaussian Mixture Models}}\label{cor:mGMM-NEM} \\
Suppose that $Y|_{Z=j} \sim {\cal N}(\mu_j,\sigma^2_j)$.
So $f(y|j,\theta)$ is a normal or Gaussian pdf.
Then the pointwise pdf noise benefit for multiplicative noise
\begin{align}
f(y n|\theta) \geq f(y|\theta)
\end{align}
holds if and only if
\begin{equation}
y(n-1)\left[ y(n+1) - 2\mu_j\right] \leq 0 \;.
\label{eq:mGMM-NEMCond}
\end{equation}
\section{Conclusion}
This discussion summarizes some of the basic theorems for speeding up EM algorithms using noise injection. The theorems apply for generalized noise injection modes. We present the specializations to additive and multiplicative noise injection. Our subsequent work has demonstrated the many supervised and unsupervised machine learning algorithms are special cases of EM algorithms. This means they benefit in speed and accuracy from the principled injection of noise (i.e. noise that satisfies the NEM condition). We have demonstrated such noise benefits in unsupervised learning (like clustering\cite{osoba-kosko2013} and hidden markov model training\cite{audhkhasi-osoba-kosko-HMM2013,welch2003}) and in backpropagation training for neural networks\cite{audhkhasi2013noise, audhkhasi2014noise}.
Open research questions include the determination of optimal injective noise, conditions under which either multiplicative or additive noise outperforms the other, and the effect of data sparsity on m-NEM speed-ups and other general modes of NEM-based noise injection.
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:16293}
IRAS 16293--2422\ is a well-studied Class 0 protostellar source in Ophiuchus ($d = 120$ pc) \citep{KnudeHog1998}.
This source is known to be a binary, consisting of Source A and Source B, whose apparent separation is
5\arcsec\ \citep[$\sim$600 au; e.g.][]{Wootten_binary, Mundy_binary, Bottinelli_Mass}.
Molecular gas distribution and dynamics around the binary components have extensively been studied
at a high angular resolution with millimeter- and submillimeter-wave interferometers
\citep[e.g.][]{Mundy_binary, Bottinelli_Mass, Kuan_COMs, Pineda_ALMA, Zapata_infall, Favre_SMA, Jorgensen_SMA, Jorgensen_Cycle1}.
This source is also famous as a prototypical hot corino source,
which is characterized by rich complex organic molecules (COMs) in the vicinity of the protostars
\citep[e.g.][]{Schoier_hotcore, Cazaux_hotcore, Bottinelli_Mass, Kuan_COMs, Pineda_ALMA, Jorgensen_sugar, Jorgensen_Cycle1, Coutens_PILS}.
Recently, ALMA Cycle 1 archival data toward this source was analyzed
to investigate the kinematic structure of the infalling-rotating envelope\ associated with Source A by \citet{Oya_16293A}.
The authors found that the kinematic structure of the infalling-rotating envelope\ is successfully explained by a simple ballistic model \citep{Oya_15398}.
Based on this study, the radius of the \cb\ of the infalling-rotating envelope\ was evaluated to be $\sim$50 au.
Since the \cb\ in the infalling-rotating envelope\ is also identified in other low-mass protostellar sources,
L1527, TMC-1A, L483, BHB07-11, and HH212
\citep{Sakai_1527nature, Sakai_1527apjl, Sakai_TMC1A, Sakai_1527_highres, Oya_1527, Oya_483, Alves_BHB07-11, Lee_HH212},
it seems to be a common occurrence in low-mass star formation.
In addition, a salient result of the above study \citep[i.e.][]{Oya_16293A} is that a drastic chemical change was found around the \cb;
the OCS (carbonyl sulphide) and H$_2$CS\ (thioformaldehyde) lines trace the infalling-rotating envelope\ outside the \cb,
while the COM lines, such as CH$_3$OH\ (methanol) and HCOOCH$_3$\ (methyl formate), mainly highlight the \cb.
The H$_2$CS\ lines also trace the high velocity component inside the \cb, which is likely a circumstellar disk component.
Such enhancement of the COM emission around the \cb\ would be an essential part of the hot corino chemistry in this source.
Recently, a similar distribution of COMs around the \cb\ is also reported for the low-mass protostellar source HH212 by \citet{Lee_HH212}.
Thus, the \cb\ seems to stand for not only the physical transition zone from the infalling-rotating envelope\ to the disk,
but also the transition zone of the chemical composition.
Although the species tracing each part is different,
chemical changes around \cb s are reported for other sources,
L1527, TMC-1A, and L483 \citep{Sakai_1527apjl, Sakai_TMC1A, Oya_483}.
Conversely, chemical diagnostics can be a powerful method to investigate the physical structure of the gas around a protostar at a 100 au scale.
In this study, we apply the chemical diagnostics to IRAS 16293--2422\ Source B, which is the other component of the binary system.
Source B is known to be rich in COMs as well as Source A
\citep[e.g.][]{Bottinelli_Mass, Kuan_COMs, Jorgensen_SMA, Jorgensen_sugar, Jorgensen_Cycle1, Pineda_ALMA}.
Its disk/envelope system is reported to have a nearly face-on geometry in contrast to the edge-on geometry of Source A
\citep[e.g.][]{Pineda_ALMA, Zapata_infall, Oya_16293A}.
For this reason, the molecular line emission shows a narrower linewidth toward Source B than toward Source A.
Furthermore, an inverse P-Cygni profile\ is reported toward Source B
\citep[e.g.][]{Pineda_ALMA, Jorgensen_sugar},
which implies the existence of the infalling gas in front of the protostar along the line-of-sight.
In this study, we analyze molecular distributions and the kinematic structure of this source at a sub-arcsecond resolution,
and compare the results with those of Source A.
\section{Observations} \label{sec:observation}
In this study, we used the ALMA Cycle 1 archival data of IRAS 16293--2422\ in Band 6 (\#2012.1.00712.S),
which covers the frequency ranges from 230 to 250 GHz and from 220 to 240 GHz.
The details of the observations are reported by \citet{Jorgensen_Cycle1} and \citet{Oya_16293A}.
Here, we briefly summarize important points.
In the observations, (42--44) antennas were used,
and the primary beam (half-power beam width) is (24--25)\arcsec.
The largest recoverable scale is $\sim$13\arcsec\ at 240 GHz.
The backend correlator was tuned to a resolution of 122 kHz,
which corresponds to the velocity resolution of 0.15 km s$^{-1}$ at 240 GHz,
and a bandwidth of 468.750 MHz.
The data calibration was performed in the antenna-based manner and uncertainties are less than 10 \% \citep[ALMA Cycle 1 Technical Handbook; ][]{TechHB_Cyc1}.
In addition,
we also analyzed the ALMA Cycle 3 data, which were carried out on 5 March 2016 (\#2015.1.01060.S).
41 antennas were used in these observations with the baseline length ranging from 17 to 636 m.
The largest recoverable scale is $\sim$14\arcsec\ at 260 GHz.
The phase center of these observations is ($\alpha_{2000}$, $\delta_{2000}$) = ($16^{\rm h} 32^{\rm m} 22\fs87$, $-24\degr 28\arcmin 36\farcs3$),
and the primary beam is 25\arcsec.
The total on-source time was 16.38 minutes with typical system temperature of (60--140) K.
The backend correlator was tuned to a resolution of 122 kHz, which corresponds to the velocity resolution of 0.14 km s\inv\ at 260 GHz,
and a bandwidth of 58.6 MHz.
J1625-2527 was used for the phase calibration.
The bandpass calibration was done on the quasar J1427-4206,
and the absolute flux density scale was derived from Titan.
The data calibration was performed in the antenna-based manner,
and uncertainties are less than 10 \% \citep[ALMA Cycle 3 Technical Handbook; ][]{TechHB_Cyc3}.
Observed spectral lines are summarized in Table \ref{tb:lines},
including their rest frequencies, upper state energies, intrinsic line strengths, and synthesized beams.
The 1.3 mm (236 GHz) continuum image was obtained by averaging the line-free channels of the two Cycle 1 data (230 and 240 GHz),
whose total band width is 7.5 GHz.
The line images were obtained by subtracting the continuum data directly from the visibilities.
The Brigg's weighting with the robustness parameter of 0.5 was employed to obtain the images of the continuum and the spectral lines.
Self-calibration using the continuum emission was applied to the ALMA Cycle 1 data (the continuum, OCS, CH$_3$OH, HCOOCH$_3$, and H$_2$CS).
On the other hand, it was not applied to the ALMA Cycle 3 data (SiO),
because the continuum sensitivity is not enough for self-calibration processing
due to a limited number of line-free channels.
\section{Distribution} \label{sec:distribution}
\subsection{Overall Distributions} \label{sec:dist_all}
Figure \ref{fig:cont} shows the 1.3 mm (236 GHz) continuum map.
The synthesized beam is $(0\farcs524 \times 0\farcs463)$ (PA $73\fdegr87$).
There are two intensity peaks corresponding to the two components of the binary, Source A and Source B.
A weak emission bridging between the two peaks can also be seen.
These features are consistent with the previous report by \citet{Jorgensen_Cycle1}.
While the distribution around Source A is slightly elongated along the NE-SW direction,
that around Source B has an almost round shape.
The disk/envelope systems of Source A and B are reported to have nearly edge-on and face-on configurations,
respectively \citep[e.g.][]{Rodriguez_16293B, Chandler_infall},
to which the above distributions are consistent.
The full-width half-maximum (FWHM) sizes, the peak flux densities,
and the peak position for Source A and Source B are
determined by a two-dimensional Gaussian fit to the image.
The FWHM sizes
deconvolved by the beam are evaluated to be
$(1\farcs138 \pm 0\farcs007) \times (0\farcs583 \pm 0\farcs004)$ (PA $36\fdegr1 \pm 0\fdegr35$) and
$(0\farcs429 \pm 0\farcs003) \times (0\farcs357 \pm 0\farcs003)$ (PA $139\fdegr0 \pm 1\fdegr6$)
for Source A and Source B, respectively.
The peak flux densities are $(314.0 \pm 1.4)$ and $(899.9 \pm 1.5)$ mJy/beam for Source A and Source B, respectively.
The derived continuum peak positions are:
($\alpha_{2000}$, $\delta_{2000}$) = ($16^{\rm h} 32^{\rm m} 22\fs8725 \pm 0\fs0001$, $-24\degr 28\arcmin 36\farcs536 \pm 0\farcs002$) and
($\alpha_{2000}$, $\delta_{2000}$) = ($16^{\rm h} 32^{\rm m} 22\fs61531 \pm 0\fs00003$, $-24\degr 28\arcmin 32\farcs5467 \pm 0\farcs0004$),
for Source A and Source B, respectively.
Although they may be affected by the self-calibration, the effect seems to be negligible for a beam size of $\sim0\farcs5$.
Indeed, the continuum peak position of Source B was derived to be
($\alpha_{2000}$, $\delta_{2000}$) = ($16^{\rm h} 32^{\rm m} 22\fs6142 \pm 0\fs0001$, $-24\degr 28\arcmin 32\farcs524\pm 0\farcs002$)
without self-calibration processing,
and the difference from the position derived from the self-calibrated data is very small ($\sim0\farcs03$).
Figure \ref{fig:channelmap_OCS} shows the velocity channel maps of the OCS ($J=19-18$) line.
Absorption toward the continuum peak position can be seen in the channels with $v_{\rm LSR}$\ ranging from 3.4 to 4.3 km s\inv.
Since the systemic velocity is around 3 km s\inv\ \citep{Bottinelli_Mass}, this absorption feature is red-shifted.
It is most naturally interpreted as the inverse P-Cygni profile, as previously reported by \citet{Jorgensen_sugar}, \citet{Pineda_ALMA}, and \citet{Zapata_infall}.
At the systemic velocity, the distribution is extended at a 3\arcsec\ ($\sim$400 au) scale in diameter around the continuum peak,
although it would suffer from the resolving-out effect with the largest recoverable scale of
$\sim$14\arcsec\ in this observation.
Figures \ref{fig:channelmap_MN}, \ref{fig:channelmap_MF}, and \ref{fig:channelmap_TFA}
show the velocity channel maps of the CH$_3$OH\ ($5_{1,5}-4_{1,4}$; A$^+$), HCOOCH$_3$\ ($20_{3,17}-19_{3,16}$; A), and H$_2$CS\ (\tfaa) lines, respectively.
For all the three lines, the red-shifted components of the emission ($v_{\rm LSR}$\ = 3--5 km s\inv)
show absorption features toward the continuum peak as in the OCS case.
The distributions of CH$_3$OH\ and HCOOCH$_3$\ are more compact than that of OCS.
Although the distribution of H$_2$CS\ is also smaller than that of OCS around the protostar,
it is larger than those of CH$_3$OH\ and HCOOCH$_3$.
The H$_2$CS\ emission is slightly extended toward the western side of the continuum peak.
The integrated intensity maps of
the four molecular species are shown in Figure \ref{fig:mom0}.
The intensity distributions, except for that of HCOOCH$_3$, show a ring-like structure around the protostar;
the molecular line intensities are weaker within a radius of 0\farcs25 ($\sim$30 au) toward the continuum peak position
than the surrounding positions.
This is due to the contribution of the inverse P-Cygni profile\ toward the continuum peak position, as shown in the channel maps
(Figures \ref{fig:channelmap_OCS}, \ref{fig:channelmap_MN}, and \ref{fig:channelmap_TFA}).
The different sizes of the molecular distributions seen in the channel maps can also be confirmed in the integrated intensity maps.
Indeed,
the OCS distribution is clearly extended over 1\arcsec\ ($\sim$120 au) around the protostar.
The CH$_3$OH\ and HCOOCH$_3$\ emission is concentrated around the protostar with a radius of 0\farcs6 ($\sim$70 au).
It should be noted that
a weak emission of CH$_3$OH\ is seen
at the angular distance of 2\arcsec\ and 3\arcsec\ in the northern and southern sides of the protostar, respectively.
Although their origin is puzzling, these components may be related to the outflow
originating from Source B (see Section \ref{sec:kinematics}),
or may be enhanced by an interaction between the envelope gas and the outflow from Source A,
as suggested from the CO ($J=6-5$) observations
\citep{Kristensen_16293outflow}.
Probably, extended components of the outflows would be resolved out in the present study.
The H$_2$CS\ (\tfaa) emission is slightly more extended than the CH$_3$OH\ and HCOOCH$_3$\ emission,
but is not so extended as the OCS emission.
The distributions of the higher excitation lines of H$_2$CS\ (\tfab; \tfae) tend to be as compact as those of the CH$_3$OH\ and HCOOCH$_3$\ lines.
\subsection{Chemical Differentiation} \label{sec:dist_diff}
The above results for OCS, CH$_3$OH, HCOOCH$_3$, and H$_2$CS\ show small-scale chemical differentiation in the vicinity of the protostar in Source B.
This differentiation is clearly confirmed by the spatial profiles of the integrated line intensities of these molecules,
as shown in Figure \ref{fig:intprofile}.
Only the blue-shifted components,
which are in the velocity range from 0.9 to 2.9 km s\inv,
are integrated to prevent contamination of the inverse P-Cygni profile.
{\bf
The position axis of the spatial profiles has the origin at the continuum peak,
and is prepared along the line where the disk/envelope system\ is extended.
Its position angle is 110\degr, as described in Section \ref{sec:kin_rot} (Figures \ref{fig:cont} and \ref{fig:highV_OCS}).
}
In IRAS 16293--2422\ Source B,
it is well-known that the molecular line emission
associated with the main molecular isotopologues
is often optically thick \citep[e.g.,][]{Jorgensen_Cycle1}.
In such a case, the line intensities do not reflect their actual abundances,
especially toward the continuum peak.
Thus, we need to consider the optical depth effect carefully.
However, we confirm by a non-LTE excitation calculation
that the optical depth effect does not affect the difference of the emitting regions shown in Figure \ref{fig:intprofile}.
Here, we used the RADEX code \citep{vanderTak_radex}
assuming a kinetic temperature of 100 K and a $n$(\hydro)\ of $10^8$ cm$^{-3}$.
In this calculation, the optical depths of the OCS and CH$_3$OH\ lines are roughly estimated to be about 0.2 and 0.01, respectively,
at the angular offset of $-1$\arcsec\ (corresponding to $-120$ au) from the continuum peak in Figure \ref{fig:intprofile},
where their integrated intensities are 0.45 and 0.03 Jy beam$^{-1}$\ km s\inv, respectively.
The excitation temperatures are 100 and 97 K for the OCS and CH$_3$OH\ lines, respectively,
indicating that the lines are well thermalized.
{\bf
Since the excitation temperatures employed by \citet{Coutens_PILS} are between 100 and 300 K at 0\farcs5 from the protostar,
the kinetic temperature would be higher than 70 K at 1\farcs0.
Even in this condition,
the optical depths of OCS and CH$_3$OH\ are estimated to be lower than 0.3 according to the RADEX code.
It should be noted that
the above results for the optical depths do not change significantly for a $n$(\hydro)\ from $10^6$ to $10^9$ cm$^{-3}$.
}
Thus, the gradual change in the OCS intensity and the CH$_3$OH\ intensity
can be attributed to the difference in their respective distributions.
In addition,
the excitation effect cannot make the OCS distribution more extended
in comparison with that of CH$_3$OH\ and HCOOCH$_3$,
because the upper-state energy for the OCS line is comparable to that of the HCOOCH$_3$\ line
and is even higher than those of the CH$_3$OH\ and H$_2$CS\ lines (Table \ref{tb:lines}).
Thus, the different distributions seem to mainly represent the change in the chemical composition of the gas.
A similar chemical differentiation is also seen in Source A;
the distributions of CH$_3$OH\ and HCOOCH$_3$\ are concentrated around the \cb\ with the radius of 50 au,
while the OCS and H$_2$CS\ distributions are more extended \citep{Oya_16293A}.
In that context,
\citet{Miura_2017} recently reported on the basis of numerical simulations
that the enhancement of the COM lines around the \cb\ can be caused by
the dust heating in accretion shocks in front of the \cb.
\section{Kinematic Structure} \label{sec:kinematics}
\subsection{Rotation Feature} \label{sec:kin_rot}
The velocity widths (FWHM) of these line emissions toward Source B are as narrow as 3 km s\inv.
They are narrower than those toward Source A, which are typically wider than 5 km s\inv.
The narrow velocity widths likely originate from the nearly face-on geometry of the disk/envelope system of Source B,
as mentioned in Section \ref{sec:16293}.
Nevertheless, we can recognize small velocity gradient in the channel maps of OCS (Figure \ref{fig:channelmap_OCS}).
This result indicates that the disk/envelope system\ is slightly inclined.
At the blue-shifted velocity ($v_{\rm LSR}$\ $\sim$1.5 km s\inv),
the OCS distribution shows a slight offset from the continuum peak position toward the northwestern (NW) direction,
while at the red-shifted velocity ($v_{\rm LSR}$\ $\sim$4 km s\inv),
it tends to have a slight offset toward the southeastern (SE) direction.
This trend can be confirmed
in the integrated intensity maps of the high velocity-shifted components (Figure \ref{fig:highV_OCS});
although the red-shifted component may be contaminated by the absorption toward the continuum peak,
the blue-shifted component of the OCS line
seems to be aligned on a straight line along the SE-NW direction with a position angle (PA) of about 110\degr.
This velocity gradient suggests a rotation motion of the disk/envelope system, slightly inclined from the face-on geometry.
In the below analysis, we use the position angle of 110\degr\ as the direction
along which the mid-plane of the disk/envelope system\ is extended (hereafter `the disk/envelope direction').
\subsection{Observed Features} \label{sec:kin_obs}
Figures \ref{fig:PV_OCS} to \ref{fig:PV_others} show the position-velocity (PV) diagrams of OCS, H$_2$CS, CH$_3$OH, and HCOOCH$_3$\ that
allow us to investigate the velocity structure around the protostar.
In the PV diagrams, the inverse P-Cygni profile\ is confirmed toward the protostar.
Along the PA of 110\degr, a slight velocity gradient can be seen in the OCS line (Figure \ref{fig:PV_OCS}),
although it is heavily contaminated with the absorption feature toward the continuum peak position.
The velocity in the northwestern side tends to be lower than that in the southeastern side.
In this diagram, the position of the intensity peak has a slight offset from the continuum peak position.
The PV diagrams along the lines with various PAs also show the absorption feature toward the continuum peak position.
Moreover, the PV diagrams along the lines with the PAs of 140\degr, 170\degr, 200\degr, and 230\degr\
show two intensity peaks with an offset from the continuum peak position.
For instance, the peak intensity ($\sim$0.6 Jy beam$^{-1}$) is about 1.5 times higher than the intensity toward the continuum peak position
($\sim$0.4 Jy beam$^{-1}$) in the PV diagram along the line with the PA of 200\degr.
The PV diagram of the H$_2$CS\ (\tfaa) line along the PA of 110\degr\ shows
a slight velocity gradient similar to the OCS case (Figure \ref{fig:PV_TFA}).
The absorption in the red-shifted component toward the continuum peak can be confirmed.
There is another absorption feature with the velocity higher than 5 km s\inv,
which is likely to be a contamination by an unidentified line.
As in the case of OCS,
the H$_2$CS\ line also shows two intensity peaks in the PV diagrams except for
that along the PA of 200\degr\ (i.e. the direction perpendicular to the disk/envelope direction).
The intensity dip toward the continuum peak position is clearer in H$_2$CS\ than in OCS.
In fact, the peak integrated intensity ($\sim$0.57 Jy beam$^{-1}$\ km s\inv) is
twice higher than the intensity toward the continuum peak (0.45 Jy beam$^{-1}$\ km s\inv)
in Figure \ref{fig:intprofile}.
The peak intensities have asymmetry in the PV diagrams;
the peak intensity at the southern side of the protostar is higher than at the northern side,
except for the PV diagram along the PA of 110\degr.
This may be caused by the asymmetric distribution of the gas.
On the other hand, the PV diagrams of the CH$_3$OH\ and HCOOCH$_3$\ lines do not show such a clear velocity gradient,
as shown in Figures \ref{fig:PV_MN} and \ref{fig:PV_others}(a, b), respectively.
The absence of the apparent velocity gradient for these lines seems to originate from the contamination of the inverse P-Cygni profile.
The distributions of the CH$_3$OH\ and HCOOCH$_3$\ emission along the PA of 110\degr\ is as compact as 1\farcs5.
Their maximum velocity shift from the systemic velocity ($\sim 3$ \kmps) is 2 km s\inv,
which is comparable to that for OCS.
The distributions for the CH$_3$OH\ and HCOOCH$_3$\ lines are so compact that
the red-shifted part of their emission near the continuum peak is heavily obscured by the unresolved absorption feature of the inverse P-Cygni profile.
We also prepare the moment 1 maps (not shown in this article)
of the CH$_3$OH\ and HCOOCH$_3$\ lines to closely inspect the velocity gradient in these lines.
However, no significant features of the velocity gradient can be seen for these lines.
It should be noted that the compact distributions of the CH$_3$OH\ and HCOOCH$_3$\ emission are not due to excitation effect,
because the upper-state energies of these lines are comparable to or even lower than that of OCS (See Table \ref{tb:lines}).
In Figure \ref{fig:PV_MN}, all the six PV diagrams of CH$_3$OH\ show some asymmetry with respect to the continuum peak position.
They are essentially similar to one another in the sizes of the absorption and the emission.
This asymmetry in the intensities seems to be anti-correlated to that seen in the H$_2$CS\ line
for the PV diagrams along the PA of (140--200)\degr.
Although the reason of this anticorrelation is puzzling at present,
such a chemical differentiation could be a key to understand chemical processes occurring there.
\subsection{Comparison of the Observed Molecular Distributions with the Source A Case} \label{sec:sourceA}
Since rotation motion around Source B is suggested by the OCS and H$_2$CS\ emission,
we investigate its kinematic structure in more detail.
As for Source A, we successfully explained the kinematic structure of the infalling-rotating envelope\ at a 100 au scale by a ballistic model \citep{Oya_15398, Oya_16293A}.
Hence, we take the advantage of this knowledge in the analysis of the Source B data.
In Source A, OCS traces the infalling-rotating envelope, while the COMs highlight the \cb, as mentioned in Section \ref{sec:16293}.
This situation is consistent with the chemical differentiation in Source B,
where the OCS distribution is rather extended while the CH$_3$OH\ and HCOOCH$_3$\ distributions are compact around the protostar.
If the OCS line traces the infalling-rotating envelope,
the velocity gradient shown in Figure \ref{fig:PV_OCS} can be interpreted
by a combination of the infall and rotation motions in the envelope.
The size of the CH$_3$OH\ and HCOOCH$_3$\ emission would be close to the size of the \cb, as in the case of Source A.
These features will be examined by our kinematic model in Section \ref{sec:analysis}.
On the other hand,
the distribution of H$_2$CS\ in Source B inside the \cb\ seems
different from that in Source A.
In Source A,
high velocity-shift components tracing the disk component inside the \cb\ are observed.
In contrast, such components are not apparently seen in Source B.
The PV diagram shows a double-peaked structure having an intensity dip at the continuum peak position
(Figure \ref{fig:PV_TFA}).
This dip structure does not originate from the high dust opacity,
because the emission,
particularly at the blue-shifted one,
is indeed observed toward the continuum peak position
in other molecular lines:
the PV diagrams of the CH$_3$OH, HCOOCH$_3$, and H$_2$CS\ (\tfab; \tfae) lines show a compact emission in Figures \ref{fig:PV_MN} and \ref{fig:PV_others}.
The dip is not due to the absorption by the foreground gas either,
because the dip is seen at the velocity shift of
$-1$ km s\inv\ from the systemic velocity ($\sim 3$ \kmps).
Hence, the intensity dip means that H$_2$CS\ would be deficient in the closest vicinity of the protostar.
The two intensity peaks in the PV diagram just correspond to the position within which CH$_3$OH\ and HCOOCH$_3$\ appear (Figures \ref{fig:PV_others}a, b).
Although the double-peaked structure is not clearly seen in OCS (Figure \ref{fig:PV_OCS}),
the intensity peak of OCS seems to be shifted from the continuum peak position
and almost coincides one of the intensity peaks of H$_2$CS.
Considering that CH$_3$OH\ and HCOOCH$_3$\ are concentrated around the \cb\ in Source A,
it is most likely that the two intensity peaks in the PV diagram of H$_2$CS\ represent
the positions of the \cb\ in Source B.
Figure \ref{fig:geometry}(a) shows a schematic illustration of the above configuration of the disk/envelope system.
The structure of the outflow is consistent with the configuration,
as described later (Section \ref{sec:outflow}).
\section{Modelling} \label{sec:analysis}
\subsection{Infalling-Rotating Envelope Model} \label{sec:model_ire}
We investigate the velocity gradient observed in the OCS ($J=19-18$) and H$_2$CS\ (\tfaa) lines,
using the infalling-rotating envelope\ model by \citet{Oya_15398}.
Since the velocity gradient is seen as the double-peaked structure
in the PV diagram of H$_2$CS\ (\tfaa) along the disk/envelope direction (PA 110\degr), as mentioned above,
we evaluate the physical parameters by comparing the model with this PV diagram.
Figures \ref{fig:PV_H2CS-model_PAenv} and \ref{fig:PV_H2CS-model_PAoutflow} show examples of the infalling-rotating envelope\ models
superposed on the PV diagram of the H$_2$CS\ (\tfaa) line along the disk/envelope direction
and the direction perpendicular to it, respectively.
We here adopt the systemic velocity of 2.9 \kmps\ in the model.
In the infalling-rotating envelope\ model, the main physical parameters are the protostellar mass ($M$) and the radius of the \cb\ ($r_{\rm CB}$).
To see how the PV diagrams of the infalling-rotating envelope\ model depend on these parameters,
we conduct the simulations for various sets of the parameters.
We assume an envelope with a constant thickness of 50 au\ and the outer radius of 300 au\
with the inclination angle of 5\degr\ (0\degr\ for a face-on configuration).
We also assume the intrinsic line width of 0.5 \kmps.
The model image is convolved with the synthesized beam.
Since we are interested in the kinematic structure around the protostar,
the emissivity is simply assumed to be proportional to the H$_2$\ density ($n$(\hydro)\ $\propto r^{-1.5}$) in the infalling-rotating envelope,
and no radiative transfer effect is considered in this model.
Namely, we do not make fine tuning of the intensity distribution,
but we rather just focus on the fundamental characteristics of the kinematic structure.
Since the red-shifted components in these observations suffer from the absorption feature of the inverse P-Cygni profile,
we mainly consider the kinematic structure of the blue-shifted components.
As shown in Figure \ref{fig:PV_H2CS-model_PAenv},
the velocity gradient along the disk/envelope direction seems to be reasonably explained by the rotating motion in the infalling-rotating envelope\ model.
Especially, the models with $r_{\rm CB}$\ of 40 au well reproduce the observations except for the absorption feature.
Although the model depends on the protostellar mass ($M$),
it is not well constrained due to the contamination of the absorption feature.
Hence, it is estimated to be around 0.4 \Msun\ within a factor of 2
assuming the inclination angle of the disk/envelope system of 5\degr.
On the other hand, the model result strongly depends on $r_{\rm CB}$.
In the models with $r_{\rm CB}$\ of 60 au, the distance between the two intensity peaks is larger than that observed,
because the positions of the two peaks in the model correspond to those of the \cb.
In the models with $r_{\rm CB}$\ of 20 au, the emission seems to be concentrated toward the protostar.
This is because the emission at the \cb\ is not resolved with the synthesized beam in the model,
and thus the observed double-peaked structure cannot be reproduced.
Hence, $r_{\rm CB}$\ is evaluated to be $(30-50)$ au.
In the PV diagram along the line perpendicular to the disk/envelope direction (PA 200\degr; Figure \ref{fig:PV_H2CS-model_PAoutflow}),
the models with $r_{\rm CB}$\ of 40 au show a better agreement with the observations than those with $r_{\rm CB}$\ of 20 or 60 au,
although the observations show an asymmetric distribution of the emission.
The inclination angle of 5\degr\ is fixed in the above analysis,
where the northeastern side of the disk/envelope system\ faces to us.
The velocity gradient cannot be explained with the inclination angle of 0\degr\ (completely face-on configuration)
or a negative value, where the southwestern side of the disk/envelope system\ faces to us.
Likewise,
the velocity structure cannot be reproduced either with the inclination angle larger than 15\degr.
The comparison between the infalling-rotating envelope\ model and the OCS ($J=19-18$) and CH$_3$OH\ ($5_{1,5}-4_{1,4}$; A$^+$) lines are also shown in Figure \ref{fig:PV_IREbest}.
Here the model with $M$ of 0.4 \Msun\ and $r_{\rm CB}$\ of \parRcb\ au\ is employed as a representative set of the parameters.
Strictly speaking, the OCS ($J=19-18$) line does not show a clear double-peaked feature unlike the H$_2$CS\ (\tfaa) line.
Nevertheless, the model seems to explain the basic velocity feature of the PV diagram of the OCS ($J=19-18$) line,
except for the absorption feature.
On the other hand,
the kinematic structure traced by the CH$_3$OH\ ($5_{1,5}-4_{1,4}$; A$^+$) line cannot be explained by the infalling-rotating envelope\ model
obtained from the above analysis.
The emitting region of the CH$_3$OH\ ($5_{1,5}-4_{1,4}$; A$^+$) line seems to be concentrated around the \cb\ and/or inside it.
This is consistent with the Source A case,
where the CH$_3$OH\ (11$_{0, 11}-10_{1, 10}$; A$^{+}$) line mainly highlights the \cb\ and/or inside it \citep{Oya_16293A}.
\subsection{Origin of the inverse P-Cygni Profile} \label{sec:iPC}
Although the basic feature of the kinematic structure traced by the OCS and H$_2$CS\ lines is reasonably explained
by the infalling-rotating envelope\ model (Figure \ref{fig:PV_IREbest}) as discussed above,
there remains an important problem:
{\it how can we interpret the inverse P-Cygni profile?}
The inverse P-Cygni profile\ means an infall motion along the line of sight toward the protostar.
Since the disk/envelope system is nearly face-on,
the outflow motion would also exist along the line of sight,
as shown in Section \ref{sec:outflow}.
This situation, namely the coexistence of the infall motion and the outflow motion along the line-of-sight in the vicinity of the protostar,
is hardly possible, as far as we consider the thin disk structure such as Figure \ref{fig:geometry}(a).
At least,
a thick disk/envelope structure would be necessary so as
that substantial amount of the infalling gas exists near the protostar.
In fact, such an infall component perpendicular to the mid-plane
near the launching point of the outflow
is seen in numerical simulations of outflows \citep{Machida2013}.
Even in this case, there remains the following difficulty.
Since the velocity shift of the absorption feature from the systemic velocity ($\sim 3$ \kmps) is at most
2 km s\inv\ (Figure \ref{fig:PV_MN}),
the infalling gas at this velocity-shift has to be located at 180 au,
assuming the free-fall with the protostellar mass of 0.4 \Msun.
This distance is much larger than the apparent size of the COM emission ($\sim$50 au) around the protostar,
which accompanies the inverse P-Cygni profile.
Recently, a hint to solve this problem is found in the other protostellar source L1527.
On the basis of the high-resolution molecular line observations with ALMA,
\citet{Sakai_1527_highres} reported that the thickness of the disk/envelope system is broadened at the \cb\ and inward of it
due to the stagnation of the accreting gas.
A part of the stagnant gas at the \cb\ once moves toward the out-of-plane direction, and falls toward the protostar,
if its angular momentum is extracted by some mechanisms (i.e. disk winds and/or low-velocity outflows \citep[c.f.,][]{Alves_BHB07-11}).
If this situation is also the case for Source B,
such an infalling gas may cause the inverse P-Cygni profile.
This situation is schematically illustrated in Figure \ref{fig:geometry}(b).
Hence, we incorporate the free-fall motion in the vicinity of the protostar in the model.
We approximate the gas distribution around the protostar as a spherical clump with a radius of the \cb\ for simplicity.
This model simulates the hypothetical situation that the infalling gas is once stagnated at the distance of \parRcb\ au\ ($r_{\rm CB}$) from the protostar to make a spherical clump,
and then it falls to the protostar by the gravity without the initial radial speed.
Here, the `spherical clump' mimics the gas stagnation,
although the gas distribution is not spherical in reality.
Figure \ref{fig:PV_freefall} shows the results.
This very simplified picture can explain the kinematic structure of CH$_3$OH\ at a \parRcb\ au\ scale around the protostar,
including the velocity of the absorption in the inverse P-Cygni profile.
The kinematic structures of OCS and H$_2$CS, on the other hand,
seem to be explained by the combination of the infalling-rotating envelope\ model for the extended components
and this free-fall picture for the inverse P-Cygni profile.
In this free-fall picture, the absorption gas for the inverse P-Cygni profile\ having the infall velocity of 2 km s\inv\ is located at 33 au from the protostar,
which is smaller than the size of their apparent distribution.
However, it should be stressed that this model is just a simplified one representing
the above hypothetical physical picture.
We need to resolve the structure of the vicinity of the protostar to verify its validity.
\section{Outflow} \label{sec:outflow}
We have also analyzed the SiO ($J=6-5$) data.
Figure \ref{fig:highV_SiO} shows the integrated intensity maps of the high velocity-shift components traced by SiO.
This molecule is often employed as a shock tracer \citep[e.g.,][]{Mikami_L1157-SiO, Bachiller_shockL1157}.
Both the blue-shifted and red-shifted components show a shell-like feature surrounding the continuum peak,
which are spatially overlapped with each other even at a 300 au scale.
\citet{Kristensen_16293outflow} have suggested by using the CO ($J=6-5$) ALMA observation
that there is an interaction of the outflow from Source A with Source B, resulting in shocks there.
In addition,
\citet{Girart_16293outflow} have also detected the SiO ($J=8-7$) line around Source B,
and have suggested
that the SiO morphology is consistent with an
interaction of the outflow from Source A with Source B.
Nevertheless, the authors did not investigate the kinematics.
Although this possibility cannot be ruled out,
these SiO components can also be attributed to the outflow from Source B with a nearly pole-on geometry;
namely, its kinematic structure is consistent with the situation that
we are looking at the outflow cavities as a shell-like feature.
We favor this possibility,
because the CO outflow from Source A is blue-shifted while the SiO emission has the red-shift component.
Moreover, the interaction cannot be seen clearly in the other lines.
For instance, no enhancement of CH$_3$OH\ is seen in the southeastern side (Figure \ref{fig:PV_IREbest}c).
Furthermore, the SiO emission can be seen in the back side of Source B
with respect to the direction to Source A.
When we look at the two components of SiO closely (Figure \ref{fig:highV_SiO}),
the blue-shifted components are slightly extended toward the northeastern direction from the continuum peak.
On the other hand, the red-shifted emission is stronger than the blue-shifted one in the southwestern side of the continuum peak.
Hence, the outflow axis might be inclined slightly (positive inclination angle),
as illustrated in Figure \ref{fig:geometry}.
Thus, the blue-shifted outflow lobe seems to be overlapped on the protostar along the line of sight.
This interpretation is natural,
because the disk/envelope system\ has a face-on geometry.
The positive inclination angle is consistent with that suggested by the analysis of the disk/envelope system.
In the OCS, CH$_3$OH, HCOOCH$_3$, and H$_2$CS\ lines, the inverse P-Cygni profile\ is seen toward the protostar.
As discussed in Section \ref{sec:iPC},
the picture shown in Figure \ref{fig:geometry}(b) allows the coexistence of the blue-shifted outflow lobe and the infalling gas toward the protostar.
In this case, the outflow responsible for the above SiO distribution would be not launched directly from the protostar,
but could be from the inner part of the disk/envelope system possibly around the \cb.
In fact, there is a hint for such a situation in the kinematic structure of the SiO emission.
The SiO emission shows the bipolar outflow lobes along the PA of 200\degr, as shown in Figure \ref{fig:PV_SiO-OCS_PAoutflow}.
In this figure, we note that there are the two absorption features toward the continuum peak position,
which are likely due to the contamination by other molecular lines with the inverse P-Cygni profile.
The blue-shifted and red-shifted lobes are seen in both the northeastern and southwestern sides of the continuum peak,
which is consistent with the configuration shown in Figure \ref{fig:geometry}(b).
If the outflow is accelerated as it propagates away from the protostar \citep[e.g.][]{Lee_outflow},
the systemic velocity ($\sim 3$ \kmps) component
may approximately regarded as
the outflow component at its launching point.
At the systemic velocity, the SiO emission is seen at the position with an offset of $\sim$0\farcs5 from the continuum peak,
but not toward the protostar.
The radial size of the SiO dip is larger than those of the OCS and H$_2$CS\ emission in
Figure \ref{fig:PV_SiO-OCS_PAoutflow}.
Hence, this feature is not likely to be due to the optical depth effect toward the continuum peak.
This implies that the launching point of the outflow has a radial offset from the protostar (Figure \ref{fig:geometry}b).
Moreover, the SiO emission at the systemic velocity seems to appear near to the intensity peak of
the OCS and H$_2$CS\ emission.
This feature may suggest that the launching point of the outflow could be around the \cb\ traced by OCS and H$_2$CS.
A similar situation is recently reported for BHB07-11 \citep{Alves_BHB07-11}.
Large-scale outflows blowing out from this binary system have extensively been studied
\citep[e.g.][]{Mizuno_largeoutflow, Hirano_largeoutflow}.
Their directions are different from that found in this study.
This contradiction may be due to the complexity of the binary system.
The dynamical timescale of the outflow lobes at a larger scale $(10^4 - 10^5)$ au
is estimated to be $(10^4 - 10^5)$ years \citep[]{Mizuno_largeoutflow, Hirano_largeoutflow}.
On the other hand, it is reported that Source A and Source B are rotating around each other with the period of
$\sim 2 \times 10^4$ years \citep[]{Bottinelli_Mass}.
Thus the directions of the two outflow can be modulated in a complex way.
Hence, the small-scale outflow structure in the vicinity of a protostar has to be studied
in order to explore the relation between the outflow and the disk/envelope system.
It should be noted that a small-scale outflow structure of IRAS 16293--2422\ is investigated by using the CO lines with ALMA
\citep{Kristensen_16293outflow, Loinard_16293outflow, Girart_16293outflow}.
However, the outflow morphology seen in the CO lines
in this source is very complicated.
The northwest-southeast outflow is seen in Source A,
while the outflow from Source B is unclear in these observations.
In contrast, our kinematic analysis suggests
a possibility that the SiO emission traces the outflow launched from Source B.
\section{Gas Kinetic Temperature} \label{sec:Tkin}
For Source A, \citet{Oya_16293A} evaluated the gas kinetic temperature from the intensities of the two lines (\tfaa; \tfab) of H$_2$CS.
The lines of H$_2$CS\ with different $K_{\rm a}$\ can be used as a good tracer of the gas kinetic temperature.
As a result, it is found that the gas kinetic temperature around Source A once rises from the infalling-rotating envelope\ to the \cb,
and then drops in the disk component inside the \cb.
Here, we also evaluate the gas kinetic temperature around Source B from the H$_2$CS\ line intensities.
Figures \ref{fig:PV_others}(c--f) show the PV diagrams of the high excitation lines of H$_2$CS\ (\tfab; \tfae)
on which those of the H$_2$CS\ line (\tfaa) is superposed.
The distributions of the higher excitation lines of H$_2$CS\ seem to be
more concentrated to the continuum peak position
than the H$_2$CS\ line (\tfaa).
Thus the intensity ratio of a higher excitation line relative to the lower excitation line
becomes higher toward the protostar.
This implies that the gas kinetic temperature is raised as approaching to the continuum peak position.
The gas kinetic temperature is evaluated
by using the non-LTE code RADEX \citep{vanderTak_radex}, as shown in Table \ref{tb:Tkin}.
Temperatures are calculated for the positions with the radii of 0, 60, and 120\ au
from the continuum peak along the disk/envelope direction (PA 110\degr).
For this calculation, we prepare the integrated intensity maps of the three H$_2$CS\ lines (\tfaa; \tfab; \tfae) with the velocity width of 1 km s\inv.
The velocity-shift range from $-1.5$ to $-0.5$ km s\inv\ is taken at the continuum peak position
in order to extract the high velocity-shift component excluding the absorption effect.
The H$_2$CS\ (\tfae) line emission is detected with this velocity range (Figure \ref{fig:PV_others}).
On the other hand, the velocity-shift range from $-0.5$ to $+0.5$ km s\inv\ is taken for the positions at the distance of
60\ and 120\ au
from the continuum peak position.
The above velocity-ranges and the positions are shown in the PV diagram of H$_2$CS\ (\tfaa) (Figure \ref{fig:PV_H2CS_Tkin}).
It should be noted that the H$_2$CS\ (\tfaa) line may possibly be optically thick toward the continuum peak.
Hence, the intensity of the H$_2$CS\ (\tfaa) line may be attenuated,
which would overestimate the gas kinetic temperature.
Thus the values derived from the \tfab\ and \tfae\ lines (the second row in Table \ref{tb:Tkin})
may be more reliable than those derived from the \tfaa\ and \tfab\ lines (the first row).
The gas kinetic temperature is as high as $\sim$100 K within the radius of 60\ au.
Moreover, the temperature tends to be higher as getting close to the protostar probably due to the thermal heating by the protostar.
In Source B, we do not find any particular enhancement of the gas kinetic temperature at the \cb\ in contrast to the Source A case,
probably because of the limited resolution of the observations.
Higher angular resolution observations are required for further investigation
on the small-scale temperature structure of this source.
\section{Summary} \label{sec:summary}
We analyzed the OCS, CH$_3$OH, HCOOCH$_3$, H$_2$CS, and SiO data observed toward IRAS 16293--2422\ Source B at a sub-arcsecond resolution with ALMA.
Major findings are as follows:
\noindent(1)
The chemical differentiation observed for the above molecules is similar to that found in Source A;
OCS and H$_2$CS\ have more extended distributions than CH$_3$OH\ and HCOOCH$_3$.
\noindent(2)
Although the disk/envelope system of Source B has a nearly face-on configuration,
a modest
rotation feature is implied in the extended component of the H$_2$CS\ and OCS lines.
It is reasonably interpreted in terms of the infalling-rotating envelope\ model assuming the ballistic motion.
On the other hand, the CH$_3$OH\ and HCOOCH$_3$\ lines do not show a rotation feature.
It might be due to the serious contamination of the unresolved absorption feature of the inverse P-Cygni profile.
\noindent(3)
The bipolar outflow lobes near the protostar are likely detected in the SiO emission,
although a possibility of the impact of the Source A outflow on Source B cannot be ruled out
as the origin of the SiO emission.
The blue and red lobes are largely overlapped,
and hence, this feature is consistent with
a nearly pole-on geometry
of the disk/envelope system\ of this source.
\noindent(4)
The molecular lines, except for HCOOCH$_3$, show absorption features toward the protostar.
It is interpreted as the inverse P-Cygni profile\ by the infalling gas along the line of sight, as previously reported.
The infall motion is reasonably explained by the free-fall motion of the gas stagnated around the \cb,
although it is a tentative and quite simplified interpretation.
\noindent(5)
The coexistence of the outflow and the infall motion toward the protostar
may suggest that the current launching point of the outflow responsible for the SiO emission likely has an offset from the protostar.
\noindent(6)
The gas kinetic temperature is evaluated by using the H$_2$CS\ lines.
The gas kinetic temperature is found to be as high as $\sim$100 K,
which is consistent to the hot corino character of IRAS 16293--2422\ Source B.
The temperature is systematically higher toward the continuum peak position than toward the outer positions.
No enhancement of the temperature is seen around the \cb\ in contrast to the Source A case.
\acknowledgments
We thank the anonymous reviewer for valuable comments.
This paper makes use of the ALMA data set ADS/JAO.ALMA\#2012.1.00712.S and \#2015.1.01060.S.
ALMA is a partnership of the ESO (representing its member states),
the NSF (USA) and NINS (Japan), together with the NRC (Canada) and the NSC and ASIAA (Taiwan),
in cooperation with the Republic of Chile.
The Joint ALMA Observatory is operated by the ESO, the AUI/NRAO and the NAOJ.
The authors are grateful to the ALMA staff for their excellent support.
Y.O. acknowledges the JSPS fellowship.
This study is supported by Grant-in-Aid from the Ministry of Education, Culture, Sports, Science, and Technologies of Japan (21224002, 25400223, 25108005, and 15J01610).
N.S. and S.Y. acknowledge financial support by JSPS and MAEE under the Japan--France integrated action program (SAKURA: 25765VC).
Y.O., N.S, Y.W., and S.Y. also acknowledge financial support by JSPS and MAEE under the Japan--France integrated action program.
C.C. and B.L. acknowledge the financial support by CNRS under the France--Japan action program.
C.F. acknowledges the support from the Italian Ministry of Education, Universities and Research, project SIR (RBSI14ZRHR).
|
\section{Introduction}
Planets orbiting the nearest stars to the Sun are the most highly prized of all exoplanets, since they represent the most accessible targets for follow-up characterization studies. The measurement of precision radial velocities has allowed us to begin to build up a collection of planets orbiting the nearest stars, while also characterizing their orbital parameters. In particular, discoveries like 51 Peg $b$ \citep{MayorQueloz1995}, 47 UMa $b$ \citep{ButlerMarcy1996}, 70 Vir $b$ \citep{MarcyButler1996}, HD143361 $b$ and HD154672 $b$ \citep{Jenkins2009}, HD86226 $b$, HD164604 $b$, HD175167 $b$ \citep{Arriagada2010}, HD128356 $b$, HD154672 $b$ and HD224538 $b$ \citep{Jenkins2017}, GJ 876 $b$, $c$, $d$, $e$ \citep{Rivera2010}, and $\upsilon$ And $b$, $c$, $d$ \citep{Wright2009,Curiel2011}, among others, have allowed us to explore the wide diversity of gas giant planetary systems.
In the last few years, the advances in radial velocity precision that have been driven by technology improvements and better analysis methods have allowed the discovery of the first batch of low-mass planets orbiting nearby stars, e.g., GJ 876 $d$ \citep{Rivera2010}, HD40307 $b$, $c$, $d$, $e$, $f$ and $g$ \citep{Mayor2009, Tuomi2013a}, GJ 581 $d$ \citep{Vogt2010}, GJ 667C $b$, $c$ and $d$ \citep{AngladaEscude2012, AngladaEscude2013}, the candidates orbiting $\tau$ Ceti, planets $b$, $c$, $d$, $e$ and $f$ \citep{Tuomi2013b, Feng2017} and more recently Proxima Centauri $b$ \citep{AngladaEscude2016} represent a new population of super-Earth planets not witnessed in the Solar System, and are defined as being small planets with masses $\sim$2-10$M_{\oplus}$ that can either be primarily rocky objects or more fluffy, atmosphere dominated worlds \citep{Valencia2007, Kaltenegger2011}.
In comparison to the gas giants, super-Earths seem to have some dramatically different characteristics, likely related to their formation and early evolution. They generally appear to be orbiting on mostly circular orbits \citep{TuomiAngladaEscude2013}, come in tightly packed planetary systems \citep{Lissauer2011, Latham2011}, and do not seem to follow the same metallicity bias as the gas giants \citep{Buchhave2012, Courcol2016}. In fact, there may be a lack of low-mass planets orbiting nearby and super metal-rich Sun-like stars \citep{Jenkins2009, Jenkins2013}. Models that invoke core accretion as the dominant planet formation scenario predict some of these trends, with mass functions rising heavily toward the lowest masses \citep{Mordasini2008}, also shown by analysis of the radial velocity sample of detected planets \citep{LopezJenkins2012}.
Planetary formation models also predict a damping of the metallicity bias in planet fraction for low-mass objects, since the stellar metallicity is an observational proxy of the dust content in the inner disk when the planets were undergoing formation. However the picture may be less clear, since \citet{Mulders2016} have shown that there might be an increase in the occurrence of small rocky planets around host stars with super-solar metallicities and orbital periods $<$ 10 days.
Although the radial velocity method has been very successful at planet detection, it is an indirect method and therefore care must be taken when trying to confirm any signal with an amplitude at the few m s$^{-1}$ level (like many super-Earth signals), since this is the domain where stellar activity effects that are correlated with the rotation of the star can impact the data \citep{Boisse2011, Boisse2012}. In numerous cases, both large and small planet candidates have been challenged as being due to the effects of stellar activity (e.g., HD166435 \citealt{Queloz2001}; HIP13044, \citealt{JonesJenkins2014}; HD41248, \citealt{Santos2014}; GJ 581 $d$, \citealt{Robertson2014}; Kapteyn $b$ \citealt{Robertson2015}; $\alpha$ Cen B $b$, \citealt{Rajpaul2016}), with most of these challenges leading to counter-claims (e.g. HD41248, \citealt{JenkinsTuomi2014}; GJ 581 $d$, \citealt{AngladaEscudeTuomi2015}; Kapteyn $b$ and $c$, \citealt{Anglada2016}). Therefore due care must be taken to ensure any signal has been well inspected for the effects of stellar activity and/or stellar rotation.
Once a planet has been confirmed orbiting a nearby star, there exists the ability to perform detailed secondary follow-up studies, like measuring accurate stellar atomic and molecular abundances \citep{Schuler2015, Melendez2017} that could be sign-posts of planetary systems, or searching for transits and secondary eclipse measurements \citep{Baskin2013,Chen2014,vonParis2016}. The combination of mass from the velocities and radius from any detected transit allows the bulk density of the planet to be measured (e.g., BD+20594 $b$, \citealt{Espinoza2016}; GJ 1214 $b$, \citealt{Valencia2013}; GJ 436 $b$ \citealt{vonBraun2012, Lanotte2014}; 55 Cnc $e$, \citealt{deMooij2014, Winn2011}) and from there, model comparisons can be made to infer the bulk composition. Therefore, gaining a better understanding of the population of low-mass planets requires the detection of more of these worlds orbiting bright stars in the solar neighborhood.
Here we present data from a 16 year precision radial velocity monitoring campaign, using multiple high resolution optical spectrographs, of the nearby ($\sim$5 pc) K0.5 star HD26965.
\section{HD 26965 - Stellar properties}\label{sec:prop}
HD26965 (HIP19849, GJ 166A) is classified as a K0.5V star \citep{Gray2006} with a visual magnitude of $V$=4.43 and an optical color of $B-V$=$0.82$. An activity index of log$\,R_{\rm HK}^{'}$=$-4.99$ is reported by \citet{Jenkins2011}. This value is also consistent with measurements found in other sources in the literature (e.g., -5.09, \citealt{Gray2006}; -4.97, \citealt{Murgas2013}) and a comparison with the Sun's mean activity value of log$\,R_{\rm HK \odot}^{'}$=$-4.91$ \citep{MamajekHillenbrand2008} tells us that HD26965 is a chromospherically quiet star.
The remaining stellar parameters were estimated using the {\bf S}pectroscopic {\bf P}arameters and atmosphe{\bf E}ric {\bf C}hem{\bf I}stri{\bf E}s of {\bf S}tars code ({SPECIES}; Soto \& Jenkins, submitted). {SPECIES} derives the effective temperature, surface gravity, metallicity, microturbulence, macroturbulence and rotational velocity, mass, age and chemical abundances for 11 elements in a self-consistent and automatic manner. The parameters were derived using high-resolution, high signal-to-noise spectra as input for the code, where in this particular case, we have used spectra from HARPS to derive the stellar parameters with SPECIES. The first four parameters were found by measuring the equivalent widths (EWs) for a set of iron lines using the {ARES} code \citep{Sousa2007}. These values, along with a stellar atmosphere model \citep{Kurucz1993}, were then input to {MOOG} \citep{Sneden1973}, which solves the radiative transfer equation by imposing excitation and ionization equilibrium. Following on from this, we then derived the chemical abundances, measuring the EWs of a set of lines for each element. Macroturbulence and rotation velocity were computed by measuring the broadening of spectral lines using a Fourier analysis. Finally, mass and age were found by fitting isochrones \citep{Dotter2008} using the luminosity and temperature of the star. Figure \ref{fig:specieschains} shows the final distributions for the stellar mass, age and log $g$ using a Markov chain Monte Carlo (MCMC) method within SPECIES. More details about this code, in particular on the treatment of correlations between parameters and uncertainties, can be found in Soto \& Jenkins (submitted).
\begin{figure}\label{fig:specieschains}
\centering
\includegraphics[scale=0.47]{F1.pdf}
\caption{Corner plot showing the one and two dimensional projections of the posterior probability distributions for the mass, age and log $g$ parameters estimated via MCMC samples with SPECIES. The plot has been generated using the Python pa\-ckage \texttt{corner.py} \citep{corner}.}
\end{figure}
We found HD26965 to have a metallicity [Fe/H] of -0.29 $\pm$ 0.13 dex, consistent with previously reported values (e.g., -0.28 dex \citealt{Gray2006, ValentiFischer2005, Turnbull2015}), and significantly poorer in metals than the Sun. SPECIES finds a rotational velocity of $v\,$sin$\,i$=1.23 $\pm$ 0.28 km s$^{-1}$, which is in agreement with the values of 1.1 $\pm$ 1.0, 1.4 $\pm$ 0.8, 1.6 $\pm$ 0.8 km s$^{-1}$ reported by \citet{Glebocki2005} calculated via cross-correlation, calibrated line full width at half maximum (FWHM), and convolution with rotational broadening profiles, respectively.
The $v\,$sin$\,i$ value we report is consistent with the old age of the star found by {SPECIES} if we consider that stars on the main sequence spin-down with time due to the loss of angular momentum from winds and the increase in stellar radius with time that is required to maintain hydrostatic equilibrium as the core changes due to nuclear burning. The Sun has a rotational velocity of only 1.6 $\pm$ 0.3 km s$^{-1}$ \citep{Pavlenko2012} and we classify it as a slow rotator. In summary, the values found for the parameters make HD26965 a good candidate for radial velocity planet search since it can be considered a quiescent and slowly rotating star. The properties and derived parameters for HD26965 are summarized in Table \ref{tab:properties}.
\begin{table}\label{tab:properties}
\center
\caption{Stellar Parameters of HD26965.}
\label{tab:params}
\begin{tabular}{lcc}
\hline\hline
\multicolumn{1}{l}{Parameter}& \multicolumn{1}{c}{ Value } & \multicolumn{1}{c}{Source}\\ \hline
R.A. (J2000)& 04:15:16.32 & SIMBAD\\
Dec. (J2000)& -07:39:10.34 & SIMBAD\\
$m_{V}$ & 4.43 & SIMBAD\\
$B$-$V$ & 0.82& SIMBAD \\
Distance (pc) & 4.98 $\pm$ 0.01 & \citealt{vanLeeuwen2007}\\ \hline
Spectral type & K0.5V & \citealt{Gray2006} \\
Mass ($M_{\odot}$) & 0.76 $\pm$ 0.03 & This work (SPECIES)\\
Age (Gyr) & 9.23 $\pm$ 4.84 & This work (SPECIES)\\
Luminosity ($L_{\odot}$)& 0.44 & \citealt{Anderson2012}\\
T$_{\rm eff}$ (K) & 5151 $\pm$ 55 & This work (SPECIES)\\
\lbrack Fe/H\rbrack & -0.29 $\pm$ 0.12 & This work (SPECIES)\\
log $g$& $4.54 \pm$ 0.04 & This work (SPECIES)\\
$v$ sin $i$ (km s$^{-1}$) & 1.23 $\pm$ 0.28 & This work (SPECIES)\\
log $R^{'}_{HK}$ & -4.99 & \citealt{Jenkins2011}\\ \hline
\end{tabular}
\vspace{0.5cm}
\end{table}
\section{Spectroscopic Observations}\label{sec:obs}
High-precision Doppler measurements of HD26965 were carried out using 4 different spectrographs: The High Resolution Echelle Spectrograph (HIRES) installed on the 10 m Keck Telescope in Hawaii, the Carnegie Planet Finder Spectrograph (PFS) mounted on the 6.5 m Magellan II (Clay) telescope at Las Campanas Observatory, CHIRON mounted on the 1.5 m telescope from the Small to Moderate Aperture Research Telescopes (SMARTS) consortium in Cerro Tololo Interamerican Observatory and the High Accuracy Radial velocity Planet Searcher (HARPS) installed on the 3.6 m ESO telescope at La Silla Observatory.
\begin{table*}\label{tab:obs_params}
\center
\caption{Summary of instrumental and observational parameters for the different instruments.}
\label{tab:params}
\begin{tabular}{lcccccc}
\hline\hline
\multicolumn{1}{l}{Instrument}& \multicolumn{1}{c}{ Resolution } & \multicolumn{1}{c}{$<$SNR$>$/Resol. element} &\multicolumn{1}{c}{$<$Exposure time$>$}&\multicolumn{1}{c}{$<$RV error$ >$}&\multicolumn{1}{c}{N$_{\rm obs}$}&\multicolumn{1}{c}{ Time baseline}\\
\multicolumn{1}{l}{}& \multicolumn{1}{c}{} & \multicolumn{1}{c}{} &\multicolumn{1}{c}{(s)}&\multicolumn{1}{c}{(m s$^{-1}$)}&&\multicolumn{1}{c}{(yrs)}
\\ \hline
HIRES & 45,000 & 270 $\times$ 4 exp & 11 & 1.2& 230 &12 \\
PFS & 80,000 & 235 $\times$ 4 exp & 40 &1.0 & 65 &5 \\
CHIRON & 95,000 & 120 $\times$ 3 exp & 300 & 1.6& 259 &2 \\
HARPS & 115,000 & 150 $\times$ 4 exp & 100& 0.4 & 437 &10 \\
\hline
\end{tabular}
\vspace{0.5cm}
\\
\end{table*}
\subsection{HIRES Observations}
The full HIRES \citep{Vogt1994} dataset comprises 229 individual Doppler measurements with an observational baseline of almost twelve years, between November 22nd 2001 and August 25th 2013. These individual radial velocities have been binned nightly to produce 90 measurements. One outlier point with a velocity value more than 3-$\sigma$ away from the mean of the series has been rejected as it was acquired under poor weather conditions.
HIRES uses the iodine cell method to deliver high precision radial velocities. The method employs a cell containing molecular gaseous iodine (I$_{2}$) that is mounted before the slit of the spectrograph so that the incoming starlight is imprinted with thousands of I$_{2}$ absorption lines, between $\sim$4800\AA{} and $\sim$6200\AA{} that are used for both very precise wavelength reference points and also in the determination of the instrumental point spread function (PSF).
The HIRES spectrograph covers a wavelength range of 3700-8000\AA{}. For most of the observations the B5 Decker (0.86''$\times$3.5'') was used, delivering a spectral resolving power of $R\sim$45,000. The C2 Decker (0.86''$\times$14'', $R\sim$45,000) was also used for a smaller number of observations. I$_{2}$-free template observations were carried out with the B3 Decker (0.574''$\times$14'') at $R\sim$60,000.
For the template observations we acquire multiple shots (typically 3) of the target star without I$_{2}$ with the narrow slit and we bracket these observations with the spectra of a bright, fast rotating B star observed through the I$_{2}$ cell. These I$_{2}$-free shots are then combined to create a high signal-to-noise, high resolution spectrum of the star that is later used for the computation of the radial velocities following the spectral synthesis procedure explained in \citet{Butler1996}, where the I$_{2}$ region is divided into $\sim$700 chunks of about 2\AA{} each to produce an independent measure of the wavelength, PSF, and Doppler shift.
This procedure is also carried out for PFS and CHIRON observations.
Exposure times varied with nightly weather conditions, but we obtained a formal mean\footnote{Weighted means using the radial velocity uncertainties as weights; $w_{i}= 1/\sigma_{i}$} uncertainty of $\sigma_{\rm BIN}$= 1.21 m s$^{-1}$ and $\sigma$=1.18 m s$^{-1}$ for the binned nightly and unbinned radial velocities, respectively, with this spectrograph.
From individual HIRES spectra we have calculated the S-indices from the Ca \sc ii \rm H and K line cores (at 3968.47\AA{} and 3933.66\AA{}, respectively) following the prescription of \citet{Duncan1991} also described in \citet{Arriagada2011}. S-indices can be used for chromospheric activity analysis of the stars \citep{Arriagada2011, Boisse2011} since they are known to be correlated with spot activity on the surface of the star that can mimic planetary signals, or at best, introduce noise into the data.
\subsection{PFS Observations}
Observations were carried out using PFS \citep{Crane2006, Crane2008, Crane2010} between October 18th 2011 and March 5th 2016. We obtained a total of 65 individual radial velocity measurements, translated into 19 binned velocities. PFS is also equipped with an I$_{2}$ cell for precise radial velocity measurements and it delivers a resolution of $R\sim$80,000 in the I$_{2}$ region when observing with the 0.5''$\times$2.5'' slit. I$_{2}$-free template observations were acquired with the 0.3''$\times$2.5'' slit at a resolution of $R\sim$127,000.
We routinely expose for a typical signal-to-noise ratio of $\sim$300 per spectral resolution element required to achieve a level of $\sim$1-2 m s$^{-1}$ radial velocity precision. For bright targets, such as HD26965, we take consecutive multiple exposures -usually 4 or 5- within a timespan of 5 minutes, to both average over the strongest stellar p-mode oscillations ($\sim$5 min for solar-type stars; \citealt{Leighton1962, EvansMichard1962,Ulrich1970}) and avoid saturation. For monitoring the stellar activity, S-indices were derived using individual spectra using the same approach described for HIRES.
We report a mean uncertainty of $\sigma$=0.97 m s$^{-1}$ from this instrument. Mean uncertainty for the nightly binned data is $\sigma_{\rm BIN}$= 0.98 m s$^{-1}$.
\subsection{CHIRON Observations}
All observations with the fiber-fed high-resolution echelle spectrograph CHIRON \citep{Tokovinin2013} were performed in service mode at $R\sim$95,000 using the `{\it Slit}' mode and 3$\times$1 pixel binning. CHIRON is installed in a thermally controlled space that allows the instrument to be stabilized to temperatures drifts of $\pm$ 2 K.The spectrograph covers a fixed wavelength range between 4150\AA{} and 8800\AA{} which, unfortunately, does not allow any measurement of calcium lines to monitor the chromospheric activity. CHI\-RON also employs an I$_{2}$ absorption cell for wavelength calibration. The CHIRON team provides reduced data corresponding to wavelength calibrated spectra \citep{Brewer2014}. We also acquired higher resolution I$_{2}$-free templates taken in `{\it Narrow}' mode at $R\sim$136,000 with the same pixel sampling as in `{\it Slit}' mode. Then we used our pipeline to compute the final Doppler shifts with a modified routine similar to the ones used in the PFS and HIRES reduction.
In 2014 we started a high-cadence campaign using this instrument to monitor nearby bright FGK stars with V$\leq$6. When observing with CHIRON, we have found that the linearity regime for the CCD ends once the counts per pixel reach $\sim$30,000, so we have routinely exposed every target up to a maximum level of 25,000 counts to avoid reaching this non-linearity regime.
Since this target is a bright star, we use the same observational strategy that was used on both PFS and HIRES, meaning we take multiple short single exposures of the star that are combined into a single high-precision measurement.
Previous work by \citet{Jones2016} have shown precision of $\sim$5 m s$^{-1}$ using the high efficiency slicer mode to look for planets orbiting around giant stars, at a lower resolution of $R\sim$79,000 and for targets fainter than $V$=6.
Recent results by \citet{Zhao2018A} using the same observing mode we describe in this work have also shown consistent short-term (nightly) radial velocity precision on the $\sim$1 m s$^{-1}$ level for the very bright stars. They obtain a mean error of 1.1 m s$^{-1}$ and 1.2 m s$^{-1}$ for $\alpha$ Centauri A ($V$=-0.01) and B ($V$=1.13), respectively.
Results from our analysis give a mean radial velocity error for this bright star of $\sigma$=1.60 m s$^{-1}$ for the unbinned dataset consisting of 258 velocities taken between October 11th 2014 and January 15th 2016. The mean radial velocity error for the nightly binned velocities is $\sigma_{\rm BIN}$=1.62 m s$^{-1}$.
\subsection{HARPS Observations}
We used public data obtained with the HARPS spectrograph \citep{Mayor2003} available from the ESO HARPS archive\footnote{\url{http://archive.eso.org/wdb/wdb/eso/repro/form}}. All the data have been processed with the HARPS-Data Reduction Software (hereafter DRS) Version 3.5 pipeline which performs all the required reduction steps from bias, flat fielding and wavelength calibration of the high resolution spectra. HARPS is a pressure and temperature stabilized spectrograph that covers a wavelength range between 3800\AA{} and 6900\AA{} with a spectral resolving power of $R\sim$115,000. We note that HARPS does not make use of an I$_{2}$ cell for precise Doppler spectroscopy. Instead, exposures of a Thorium-Argon lamp are taken at the same time as each observation to get a precise wavelength reference for the science spectra (one spectrum on each of the two fibers that feed the instrument).
Radial velocities are a product of a post-reduction analysis consisting of cross-correlating each echelle order with a binary mask that is chosen depending on the spectral type of each star. This produces cross-correlation functions (CCF) for each order that are then combined to obtain a mean-weighted CCF. This mean-weighted CCF is then used to generate the radial velocities.
For HD26965 we found 483 useful public Doppler measurements between October 27th 2003 and December 5th 2013 available from ESO HARPS archive. The DRS pipeline and further post-reduction analysis produced 437 radial velocity measurements with a mean error of $\sigma$=0.43 m s$^{-1}$, and yielded a set of 65 binned radial velocities with a mean uncertainty of $\sigma_{\rm BIN}$=0.42 m s$^{-1}$. HARPS vacuum enclosure was opened in 2015 as part of an upgrade on the fibers. We refer to the pre-upgrade data as HARPS OLD.
We include 82 post-upgrade HARPS velocities between September 9th 2015 and March 27th 2016. This post-upgrade data is labeled HARPS NEW.
For all the analyses, the unbinned data from each instrument is used and is treated separately with their corresponding independent velocity offset and noise (jitter) properties.
In the case of the spectrographs equipped with an I$_{2}$ cell (HIRES, PFS, CHIRON), the reported velocities are the weighted mean of the velocities of the individual chunks while the uncertainties correspond to the standard deviation of all the chunk velocities about that mean.
For HARPS, where the observations were carried out using simultaneous Thorium exposures, the RV uncertainty is provided by the DRS and it is estimated directly from a Gaussian fit to the CCF \citep{Bouchy2001}.
{The 1,111 radial velocity measurements are shown from Table \ref{tab:firstset} to Table \ref{tab:lastset}.
\begin{figure*}\label{fig:pergram}
\centering
\vspace{-1.cm}
\includegraphics[scale=0.65]{F2.pdf}
\caption{{\it Top:} Generalized Lomb-Scargle periodogram of the unbinned combined velocities. The highest power is found at a period of 42.43 days. The vertical red arrows mark the position of the stellar rotation period and the period found in the time series. The dotted line shows the 0.1\% significance level, determined by 1,000 bootstrap resamplings. {\it Bottom:} Periodogram of sampling (window function) for the combined data.}
\end{figure*}
\section{Periodogram Analysis}\label{sec:pergram}
We started to examine the radial velocity data by using the traditional periodogram a\-na\-lysis approach to look for any periodicities embedded in the data. We used the generalized version \citep{Zechmeister2009} of the Lomb-Scargle periodogram (\citealt{Lomb1976, Scargle1982}, hereafter GLS) where we set up a minimum period of 1 day and a maximum period of 10,000 days for the search, with 80,000 trial periods evenly spaced in the frequency domain.
Figure \ref{fig:pergram} (top panel) shows the GLS periodogram of the combined radial velocities of HIRES, PFS, CHIRON and HARPS. The velocities have been mean subtracted. A maxima at 42 days (marked with a red arrow) clearly exceeds the power threshold of 0.1\% significance level. There are also two power maxima close to the 0.1\% significance threshold at $\sim$38 days and at $\sim$360 days. In the bottom panel of Figure \ref{fig:pergram} we show the periodogram of the sampling (window function) of the combined radial velocities. The secondary power spectrum peak at 360 days found in the periodogram of the velocities is also present here, and therefore can be attributed to the frequency of sampling. However, the peak at 38 days is not present and therefore further investigation is required to determine the origin of this possible signal, which we discuss below as being due to the rotation period of the star.
\section{Bayesian Analysis}\label{sec:bayes}
In addition to the traditional periodogram analysis we have performed a Bayesian analysis to search for periodic signals embedded in the data.
We modeled the radial velocities of HD26965 following the statistical model defined in \citet{Tuomi2014} and also applied in \citet{JenkinsTuomi2014} where we include the following elements:
\begin{itemize}[]
\item[] $1)$ A function describing a $k$-Keplerian planet model
\item[] $2)$ A linear trend term
\item[] $3)$ A red-noise model consisting of a $p$-th order moving average - MA($p$) - model with an exponential smoothing
\item[] $4)$ Linear correlations with the stellar activity indicators
\end{itemize}
We write the statistical model as follows
\begin{eqnarray}\label{eq:bayesmodel}
y_{i,j} = \gamma_{j} + \dot{\gamma}t_{i} + f_{k}(t_{i}) + \epsilon_{i,j} + \sum_{n=1}^{q_{j}} c_{n,j} \xi_{n,i,j} \nonumber\\
+\,\sum_{l=1}^{p}\phi_{j,l}\text{ exp }{\left\{ \frac{t_{i-l} - t_{i}}{\tau_{j}} \right\}} r_{i-l,j}
\end{eqnarray}
\noindent where $y_{i,j}$ corresponds to the observation at time $t_{i}$ for the $j$-th instrument, $\gamma_{j}$ is the velocity offset for the $j$-th dataset, $\dot{\gamma}$ is a linear trend term, and $r_{i,j}$ denotes the residuals after subtracting the model from the measurement. The function $f_{k}$ is a superposition of $k$-Keplerian signals,
\begin{equation}\label{eq:keplerian}
f_{k} (t_{i}) =\sum_{m=1}^{k} K_{m} [\, \text{cos}( \omega_{m}+ \nu_{m}(t_{i})) + e_{m} \text{cos}(\omega_{m}) ]\,
\end{equation}
where $K_{m}$ is the velocity semi-amplitude, $\omega_{m}$ is the longitude of pericenter, $\nu_{m}$ is the true anomaly and $e_{m}$ is the eccentricity. $\nu_{m}$ is also a function of the orbital period and the mean anomaly $M_{0,m}$. Hence, $f_{k}$ is fully described by $K_{m}$, $\omega_{m}$, $e_{m}$, $M_{0,m}$ and $P_{m}$ {\bf, $m\in\{1,...,k\}$}.\newline
The white noise term is denoted by the additive random variable $\epsilon_{i,j}$. We assume that there is an excess white noise in each data set with a variance of $\sigma_{j}$ such that
\begin{equation}
\epsilon_{i,j} \sim \mathcal{N}(0,\sigma^{2}_{i}+\sigma^{2}_{j})
\end{equation}
\noindent with $\sigma_{i}$ the uncertainty associated with the measurement $y_{i,j}$ and $\sigma_{j}$ is the excess white noise or jitter for the $j$-th dataset, that is treated as a free parameter in the model.
The remaining terms define the rest of the noise model, including the red-noise component: the first term with parameters $c_{n,j}$ describes the linear correlations with $q$ stellar activity indicators $\xi_{n,i,j}$ for the $n$-different instruments. The second term is the MA($p$) component with smoothing over a timescale $\tau_{j}=4$ days and $\phi_{j,l}$ with a value between -1 and 1 to quantify the correlation between measurements. The smoothing timescale is set to 4 days for simplicity \citep{Tuomi2013b}. We assume the noise is correlated in this timescale although with higher cadence smaller timescales would likely be more appropriate \citep{Tuomi2013b, Feng2016}.
\subsection{Posterior Samplings and Signal Detection}
To estimate the posterior probability of the parameters in the model given the observed data we use Bayes rule that states
\begin{equation}\label{eq:bayesrule}
P(\theta\, | \,y) = \displaystyle \frac{ P(y\,| \,\theta)\, P(\theta) }{\,\int P(y \,| \,\theta)\,P(\theta)\, d\theta}
\end{equation}
\noindent where $ P(y\,| \,\theta)$ is the probability density of the measurements given the parameters (likelihood function) and $P(\theta)$ corresponds to the prior, i.e., what is known about a given parameter and its constraints before making the measurement. The denominator in equation \ref{eq:bayesrule} is a normalizing constant such that the posterior must integrate to unity over the parameter space.
In our model we chose the priors for the orbital and instrumental parameters as listed in Table \ref{table:priors}.
\begin{table}\label{table:priors}
\center
\caption{Prior selection for the parameters}
\begin{tabular}{lccc}
\hline\hline
\multicolumn{1}{l}{Parameter}& \multicolumn{1}{c}{Prior Type}&\multicolumn{1}{c}{Range}\\
\hline
Semi-amplitude&Uniform& $K\in [\,0,K_{\rm max} ]\,$ \\
Period& Jeffrey's&$P\in [\, 1, 2P_{\rm obs}]\,$\\
Eccentricity& $\mathcal{N}(0,\sigma_{e})$ &$e \in [0,1)$\\
Long. of Peric.& Uniform &$\omega \in [\, 0,2\pi]\,$\\
Mean Anomaly& Uniform &$M_{0} \in [\, 0,2\pi]\,$\\
Jitter& Uniform & $\sigma_{J} \in [0, K_{\rm max}] $\\
Smoothing time scale & Constant& $\tau_{j}$=4\\
\hline
\vspace{0.2cm}
\end{tabular}
\end{table}
In order to investigate the signal initially found with GLS periodogram of the combined radial velocities we use our Bayesian detection method where we sample the parameter space using the Delayed-Rejection Adaptive-Metropolis (DRAM) algorithm \citep{Haario2006} based on the Adaptive-Metropolis (AM) algorithm \citep{Haario2001}, applied in \citet{Tuomi2014a} and \citet{JenkinsTuomi2014}. DRAM and AM are both methods for improving the efficiency of the Metropolis-Hasting algorithm \citep{Metropolis1953, Hastings1970}. The idea behind using DRAM is that when the posterior of a parameter is multimodal, such as the orbital period in the case of Keplerian fits to radial velocity datasets, and a new state for the chain is rejected (see full details in \citealt{Tuomi2014}), a new proposed state is drawn centered on the last one. Up to three rejections are allowed before that part of the posterior is finally discarded as a region of low probability. This has the benefit of sampling more heavily the posterior phase space, at the cost of a longer run-time.
Tempered samplings are also performed when searching for signals. We include a $\beta$ parameter following \citet{Tuomi2014}, such as $\beta \in (0,1)$, meaning we use $P(\theta\, | \,y)^{\beta}$ instead of the standard posterior probability density, $P(\theta\, | \,y)$. This way we can define the ``temperature'' of the chain simply as $T = 1/\beta$ and so a ``hot'' chain is defined when $T>1$ and a ``cold'' chain is where $T=1$. When $T>1$ the relative height of the maxima in the posterior probability density are decreased to prevent the chains from getting stuck in regions of high probability, allowing them to visit the entire period parameter space. The typical length of a chain is set to be between $10^{6} - 10^{7}$ for the search run and $10^{6}$ for the initial burn-in period.
We performed a first run for a zero-planet model to determine the observational baseline, and the instrumental noise and stellar noise parameters for each set of radial velocities. We then searched for a signal in the radial velocity data considering a 1-planet model. The search was initially done by setting the temperature for the chain hot enough to let the chain explore the entire parameter space. This is especially helpful when the parameter space is highly multimodal. Our tolerance threshold for the acceptance rate is based on the optimal acceptance rate of the Metropolis-Hastings algorithm which is $\sim$0.234 \citep{Roberts1997}. A lower threshold for the chain to be accepted was set to 10\%, so hot chains with lower acceptance rates were discarded. From these runs we found a strong signal was present with a period of 42 days. The signal identified from the maximum of the posterior probability density distribution is shown in Figure \ref{fig:posterior}, left panel.
We repeated this process by adding additional signals to the model, but we found no more statistically significant periods in the distribution of the posterior probability densities.
Finally, to constrain the detected signal, we performed parameter estimations via the AM algorithm by setting a cold chain ($\beta$=1) with the parameters initially set as a small ball around the parameters found previously by the hot chain run with DRAM. This gave rise to the posterior histograms shown in Figure \ref{fig:param_dist}, where the period, amplitude, and minimum mass distributions show nice Gaussian forms centered on their respective values, and the eccentricity distribution is consistent with zero.
Table \ref{tab:system} summarizes the final set of values for the parameters from our analysis.
\begin{table}
\center
\caption{Solutions for HD26965. Final set of orbital and instrumental parameters. 1$\sigma$ errors.}
\label{tab:system}
\begin{tabular}{lcc}
\hline\hline
\multicolumn{1}{l}{Parameter}& \multicolumn{1}{c}{ \hspace{2cm} }&\multicolumn{1}{c}{HD26965
~$b$} \\ \hline
$P$ (days) & &42.364 $\pm$ 0.015 \\
$K$ (m s$^{-1}$) && 1.59 $\pm$ 0.15 \\
$e$ & &0.017$\pm$ 0.046 \\
$\omega$ (rad) && 0.31 $\pm$ 1.93 \\
$M_{0}$ (rad) & & 4.92 $\pm$ 1.92 \\%\hline
$a$ (AU) & &0.215 $\pm$ 0.008 \\
$m \sin i$ (M$_{\oplus}$) & & 6.91 $\pm$ 0.79 \\%\hline
$\gamma_{\rm PFS}$ (m s$^{-1}$)& & 0.13 $\pm$ 0.80 \\
$\gamma_{\rm HIRES}$ (m s$^{-1}$)& & 0.50 $\pm$ 0.57 \\
$\gamma_{\rm CHIRON}$ (m s$^{-1}$)& & 0.43 $\pm$ 0.79 \\
$\gamma_{\rm HARPS, old}$ (m s$^{-1}$)& & 0.17 $\pm$ 0.50 \\
$\gamma_{\rm HARPS,new}$ (m s$^{-1}$)& & 0.45 $\pm$ 0.78 \\
$\dot{\gamma}$ (m s$^{-1}\,$year$^{-1}$) & & -0.031 $\pm$ 0.037 \\\hline
$\sigma_{\rm PFS }$ (m s$^{-1}$)& & 1.54 $\pm$ 0.20 \\
$\sigma_{\rm HIRES}$ (m s$^{-1}$)& & 2.38 $\pm$ 0.15 \\
$\sigma_{\rm CHIRON}$ (m s$^{-1}$)& & 1.78 $\pm$ 0.15 \\
$\sigma_{\rm HARPS, old}$ (m s$^{-1}$)& & 1.11 $\pm$ 0.05 \\
$\sigma_{\rm HARPS,new}$ (m s$^{-1}$)& & 0.69 $\pm$ 0.07 \\
$\phi_{\rm PFS}$& & 0.82 $\pm$ 0.10 \\
$\phi_{\rm HIRES}$& & 0.61 $\pm$ 0.07 \\
$\phi_{\rm CHIRON}$& & 0.62 $\pm$ 0.06 \\
$\phi_{\rm HARPS, old}$ && 0.81 $\pm$ 0.04 \\
$\phi_{\rm HARPS,new}$ && 0.90 $\pm$ 0.08 \\
c$_{\rm S \, PFS}$ (m s$^{-1}$) && 61.1$\pm$ 14.0 \\
c$_{\rm S \, HIRES}$ (m s$^{-1}$) & & 53.1 $\pm$ 15.0 \\
c$_{\rm BIS \, HARPS, old}$ && 0.086 $\pm$ 0.046 \\
c$_{\rm FWHM \, HARPS, old}$ && 1.8 $\pm$ 4.5 \\
c$_{\rm S \, HARPS, old}$ (m s$^{-1}$) && 1.6 $\pm$ 2.6\\
c$_{\rm H_{\alpha} \, HARPS, old}$ (m s$^{-1}$) & &-12.1 $\pm$ 8.9 \\
c$_{\rm He I \rm \, HARPS, old}$ (m s$^{-1}$) & & -76 $\pm$ 34\\
c$_{\rm BIS \, HARPS,new}$ && 0.27 $\pm$ 0.11 \\
c$_{\rm FWHM \, HARPS,new}$ && 0.089 $\pm$ 0.018 \\
c$_{\rm S \, HARPS,new}$ (m s$^{-1}$) && 105 $\pm$ 36\\
c$_{\rm H_{\alpha} \, HARPS,new}$ (m s$^{-1}$) & &76 $\pm$ 125 \\
c$_{\rm He I \rm \, HARPS,new}$ (m s$^{-1}$) & & 23 $\pm$ 141\\
\hline
\end{tabular}
\vspace{.5cm}
\end{table}
\subsection{Model Selection}
It is important to define a robust methodology that allows us to compare the results for two given models in order to address the statistical significance of one model with respect to the other.
The probability of a model $\mathcal{M}$, containing the best-fit parameters for the observed data $y$, is given by
\begin{equation}
P( \mathcal{M}\, | \,y) = \frac{P(y \, | \, \mathcal{M}) P(\mathcal{M})}{\sum^{k}_{i=1} P(y \, | \, \mathcal{M}) P(\mathcal{M})}
\end{equation}
In particular, we want to know if the model containing one planet is more probable than a zero-planet model and so on for the $k$-planet model with respect to a $k-1$-keplerian model. To solve this, we compute the probability of a given model by using the corresponding value of the Bayesian Information Criterion (BIC). A complete and detailed discussion can be found in \citet{TuomiJones2012} and \citet{Feng2016}. To compare the $\mathcal{M}_{k}$ model with a previous $\mathcal{M}_{k-1}$ model, we simply compute the logarithm of the Bayes factor, ln $B_{k,k-1}$, defined via
\begin{equation}\label{eq:bayesfactor}
{\rm ln}\, B_{k,k-1} = {\rm ln}\, P(y | \mathcal{M}_{k}) - {\rm ln}\, P(y | \mathcal{M}_{k-1})
\end{equation}
Furthermore, the model containing the best-fit parameters that support the signal has to fulfill the detection criteria described in \citet{Tuomi2012}. It must hold that
\begin{equation}
P(y | \mathcal{M}_{k}) = s P(y| \mathcal{M}_{k-1})
\end{equation}
\noindent where $s>10^{4}$. Hence using the Bayes factor defined in equation \ref{eq:bayesfactor}, we require that the $\mathcal{M}_{k}$ model describing the $k$-keplerian signal has to be more statistically probable than the $\mathcal{M}_{k-1}$ model associated with the $k-1$-keplerian signal. Following the conservative threshold from \citet{Tuomi2014}, we define that the evidence ratio should be
\begin{equation}
{\rm ln}\, B_{k,k-1} > 9.2
\end{equation}
which translates posterior odds of 10,000:1 that the $k$ model is selected over the $k$-1 model, in order to satisfy our detection criteria.
Table \ref{tab:bfactors} shows the Bayes factors for $k=0,1,2$-planet models with and without activity correlation terms.
\begin{table}
\center
\caption{Logarithm of Bayes factors comparing a $k$=0, $k$=1 and $k$=2 Keplerian models with and without activity correlations}
\label{tab:bfactors}
\begin{tabular}{lcc}
\hline\hline
\multicolumn{1}{c}{\bf Bayes Factor}&\multicolumn{1}{c}{\bf Activity}&\multicolumn{1}{c}{\bf No Activity}\\
\multicolumn{1}{l}{\bf ln \boldmath$B_{k,k-1}$}&\multicolumn{1}{c}{\bf Correlations}&\multicolumn{1}{c}{\bf Correlations}\\ \hline
ln $B_{1,0}$&43.38 & 35.44\\
ln $B_{2,1}$&0.61 & 2.96\\
\hline
\end{tabular}
\vspace{.5cm}
\end{table}
\begin{figure*}\label{fig:posterior}
\centering
\includegraphics[scale=0.34,angle=-90]{F3.pdf}
\hspace{0.1cm}\vspace{1cm}\includegraphics[scale=0.33,angle=-90]{F4.pdf}
\caption{{\it Left:} Posterior probability densities as output from our Bayesian code for a 1-planet model. {\it Right:} Phased-folded radial velocity curve.}
\end{figure*}
\begin{figure*}\label{fig:param_dist}
\centering
\includegraphics[scale=0.25,angle=-90]{F5.pdf}\includegraphics[scale=0.25,angle=-90]{F6.pdf}
\vspace{0.5cm}
\includegraphics[scale=0.25,angle=-90]{F7.pdf}\includegraphics[scale=0.25,angle=-90]{F8.pdf}
\caption{{\it Left to Right, Top to Bottom:} Final distribution of period, semi-amplitude, minimum planetary mass and eccentricity resulting from a cold-chain Adaptive Metropolis run. The numbers at the top of each figure correspond to the mode, mean, variance, skewness and kurtosis, respectively. The solid line represents a Gaussian curve with same mean and variance.}
\end{figure*}
Figure \ref{fig:thresholds} shows the minimum mass detection thresholds for additional planets orbiting around HD26965. The green-filled area represents the liquid-water habitable zone estimated according to \citet{Kopparapu2013b, Kopparapu2013a}. The thresholds are calculated following the methods in \citet{Tuomi2014}. From this figure we can say that planets with minimum masses in excess of Neptune in the habitable zone can be ruled out meaning if there are HZ planets orbiting HD26965, they would likely be super-Earths or smaller. The red circle represents the planet candidate, barely, but significantly above the detection threshold.
\begin{figure}\label{fig:thresholds}
\centering
\includegraphics[scale=0.33,angle=-90]{F9.pdf}
\caption{Minimum mass detection thresholds for additional pla\-nets orbiting around HD26965 for periods between 1 and 10,000 days. The green-filled area highlights the habitable zone for this K dwarf.}
\end{figure}
\subsection{Signal injection}
\begin{figure}[ht]
\label{fig:injec_pergram}
\centering
\includegraphics[scale=0.6]{F10.pdf}
\caption{{\it Top:} Generalized Lomb-Scargle periodogram for each dataset where a 1-planet model with the best-fit orbital parameters has been injected into the original measurements. Red arrows mark the candidate period found in the original time series. The dotted lines show the 0.1\% significance level, determined by 1,000 bootstrap resamplings.}
\end{figure}
As an additional test to investigate if the signal was supported for each instrument we performed a signal injection on the individual datasets. We use the best-fit parameters (i.e. $P$, $K$, $\omega$, $M_{0}$, e) of the putative signal by using a Keplerian function described in equation \ref{eq:keplerian}. The hypothesis is the following: if the 42-day signal is injected in a given dataset and we run our usual Bayesian analysis we should, in principle, easily detect it. If the signal is indeed present in the dataset we should recover $\sim$twice the best-fit radial velocity semi-amplitude as it has been boosted by the injection. On the other hand, if the recovered velocity semi-amplitude is significantly lower than our best-fit values, that would suggest the actual data is not supported by the instrument, or in other words, the precision of the instrument plus the current number of observations do not allow the signal to be detected. When boosting our signal, we recover the candidate period for HARPS OLD, HIRES and CHIRON datasets. For CHIRON, however, the expected peak at 42-days in the GLS is not unique, although it is above the 0.1\% significance level, as can be seen from the periodograms shown in Figure \ref{fig:injec_pergram}. In the case of PFS data, we did not recover the candidate period. Instead, we found a strong power at $\sim$5 days. This could be caused by the sparse sampling and lower number of observations available from this instrument.
\section{Stellar Activity and RV correlations}\label{sec:act_corr}
To investigate the nature of the detected 42 day signal, we perform a similar analysis as in \citet{Santos2014} on the activity indices available from each instrument. First we searched for periodicities present in the activity indices themselves, again using the GLS, and we show these results in Figure \ref{fig:act_per}. There are no statistically significant peaks associated with the 42 day signal we detect in the radial velocities. However, the periodogram of the HARPS S-indices shows an emerging peak at ~38 days, which is very close to the signal we detected in the radial velocities. Interestingly, this was the period found for the rotation of the star from previous analysis of Ca \sc ii \rm lines \citet{Saar1997} which also agrees with the period inferred from ROSAT measurements (37.1 days; \citealt{Pizzolato2003}).
Figure \ref{fig:correlations} shows the correlations between radial velocity and activity indicators for HARPS, PFS and HIRES. The combined S-indices we show have been mean subtracted and then combined together. We have computed the Pearson Rank test coefficients to determine the correlation between these quantities. Results are listed in Table \ref{tab:pearsonr} where we also list the uncertainties associated with each coefficient. To calculate these uncertainties we ran 10,000 bootstraps and created a distribution of $r$ coefficients for every activity index, where the standard deviation of the distribution gave us a measurement of the uncertainty on the coefficients. We note that the correlations are not significant within these uncertainties, given the standard statistical limits for claiming a weak ($r<$0.5), a moderate (0.5 $\le r \le$ 0.7), and a strong correlation ($r >$0.7), therefore we can conclude that the stellar activity indicators do not argue against a Doppler origin for the signal, yet the correlations indicate we must consider them in our full statistical model. Indeed, the correlations suggest there is a weak impact of the stellar noise on the velocities, and we confirm this since the probability of our statistical model is higher when we include these correlations, compared to when we exclude them (see Table \ref{tab:bfactors}).
\begin{figure}\label{fig:act_per}
\centering
\includegraphics[scale=0.6]{F11.pdf}
\caption{Generalized Lomb-Scargle periodogram of the activity indicators available from the different spectrographs. From top to bottom: BIS$_{\rm HARPS}$, CCF FWHM$_{\rm HARPS}$, S$_{\rm HARPS}$, H$_{\alpha\, \rm HARPS}$, He \sc i\rm$_{\,\rm HARPS}$, S$_{\rm PFS}$, S$_{\rm HIRES}$ and S$_{\rm COMBINED}$. The arrows mark the position of the signal found at 42.37 days in the radial velocity series and the reported stellar rotation period of 37.1 days from \citet{Saar1997}. The dotted lines show the 0.1\% significance level, determined by 1,000 bootstrap resamplings. There are no statistically significant power in the activity indicators matching the radial velocity period.}
\end{figure}
\begin{figure*}\label{fig:correlations}
\centering
\includegraphics[scale=0.75]{F12.pdf}
\caption{Radial velocity correlations with respect to the seven different activity indicators: BIS$_{\rm HARPS}$, CCF FWHM$_{\rm HARPS}$, S$_{\rm HARPS}$, H$_{\alpha\, \rm HARPS}$, He \sc i \rm$_{\rm HARPS}$, S$_{\rm PFS}$, S$_{\rm HIRES}$. We also include the combined S-indices. The dotted lines mark the 1:1 relationships.}
\end{figure*}
\begin{table}
\center
\caption{Pearson Rank test coefficients. Correlation between activity indicators and radial velocities}
\label{tab:pearsonr}
\begin{tabular}{lcc}
\hline\hline
\multicolumn{1}{l}{Activity Indicator}& \multicolumn{1}{c}{\hspace{1cm} }&\multicolumn{1}{c}{$r$}\\ \hline
BIS$_{\rm HARPS}$& & -0.02 $\pm$ 0.04 \\
FWHM$_{\rm HARPS}$& & 0.74 $\pm$ 0.04 \\
S$_{\rm HARPS}$ && 0.14 $\pm$ 0.04 \\
H$_{\alpha \,\rm HARPS}$ && 0.05 $\pm$ 0.04 \\
He I$_{\,\rm HARPS}$& & -0.04 $\pm$ 0.04 \\
S$_{\rm PFS}$ && 0.48 $\pm$ 0.12 \\
S$_{\rm HIRES}$ && 0.44 $\pm$ 0.07 \\
S$_{\rm COMBINED}$ &&0.23 $\pm$ 0.03 \\
\hline
\end{tabular}
\vspace{.5cm}
\end{table}
\section{Testing variability and stability of the period and amplitude}\label{sec:variability}
Following a similar approach as in \citet{JenkinsTuomi2014}, we tested the variability and stability of the signal of our candidate. For this analysis we only considered the HIRES and HARPS OLD datasets, since both of them have a fairly continuous sampling of Doppler measurements along the $\sim$16 years of observational baseline.
The measurements include a total of 662 unbinned velocities, and we chose JD$_{s}$=2454600 as the point to split the data, since this was close to the center of the time baseline of the observations and also produced a well balance between HARPS and HIRES data (i.e., not biased to an instrument in particular). The data prior to JD$_{s}$ contained 408 data points and the dataset after the split point contained 254 measurements.
We performed the Bayesian analysis on these 2 subsets of velocities independently, running cold chains to constrain the orbital parameters of a 1-Keplerian model. We found the signal is detected with values in agreement within uncertainties for the two baselines tested, as well as for the full data set described above. This shows us that the signal is not varying in time and thus the period and amplitude of our planetary candidate is stable over the tested observational baseline, another strong argument against an activity origin since activity processes should be quasi-static, varying over a few rotation periods of the star.
\begin{figure*}\label{fig:photometry}
\centering
\includegraphics[scale=0.56]{F13.pdf}
\caption{{\it Top:} Photometric measurements from ASAS. Grey circles show the complete set of photometry while black-filled circles are those that meet the criteria as robust points described in the text. {\it Middle:} Generalized Lomb-Scargle periodogram of the ASAS photometry. Red arrows mark the position of the planetary candidate signal at 42.37 days and the stellar rotation period reported by \citet{Saar1997} of 38.7 days. No significant powers near the period - or an integer multiple of it - are found in the periodogram. The highest power is seen at 515 days. {\it Bottom:} Periodogram of the residuals after removing the 515 day period signal.}
\end{figure*}
\section{ASAS Photometry}\label{sec:photo}
To complement the analysis we gathered photometric data available from the All-Sky Automated
Survey (ASAS) Catalog \citep{Asas1997} to investigate if any periodic signal could be seen in light curves, particularly the rotational period of the star. As mentioned above, the literature values were reported by \citet{Noyes1984} and \citet{Saar1997}, where they found $P_{\rm rot}\sim$37.10 days for this old star. We show the ASAS photometric measurements in Figure \ref{fig:photometry}. From the five different apertures available, we selected aperture 1 as its MAD\footnote{Median Absolute Deviation = median($\abs{x_{i} - {\rm median}(x)})$} value of 0.219 mag was the smallest. The mean uncertainty in the V-band photometry is $\sigma_{\rm ASAS}$= 0.036 mag. Grey circles correspond to the entire photometry set of 568 useful points acquired from 2000 to 2009. However, we excluded the data with poor quality (those not marked ``A'' or ``B'' in the catalog) and also those measurements that deviated more than 3-$\sigma$ with respect to the mean value of the time series. The highest quality data (316 points) are shown as black circles in the top panel of Figure \ref{fig:photometry}.
The bottom panel in Figure \ref{fig:photometry} again shows the GLS for the ASAS photometry. We sample the period space starting at a minimum period of 1 day and up to 10,000 days, performing 80,000 period samples. Considering just the data before JD$=$2452300 tends to favor peaks with higher power towards high frequencies (periods $\sim$1 day) but without any significant period (or an integer multiple) near the period associated with the 42 day signal of the reported planetary candidate. We also ran the periodogram analysis on the full photometric dataset with no significant periods found. Following the relations in \citet{Hatzes2002}, we found that a filling factor of $f$=0.15 would be required to induce the RV amplitude of 1.6 m s$^{-1}$ of the signal found in the combined data. If we consider the spots on the surface of the star to be opaque, for the sake of simplicity, the ratio between the stellar flux and the flux considering spots covering 0.15\% of the surface of the star would be 0.9985. This means, the loss of light due to spots on the stellar surface can be translated into a $\Delta \, m$= 1.64 mmag. Given the precision of the ASAS photometry for this star, we conclude it is insufficient to be informative.
\section{Mount Wilson HK measurements}\label{sec:hk}
Given that we find some moderate correlations between the spectral activity indicators and the radial velocities, we supplemented our activity analysis by studying the original Ca \sc ii \rm H\&K data from the Mount Wilson Observatory HK Project \citep{Wilson1978}. The Project data are publicly available from the NSO archive\footnote{http://www.nso.edu/node/1335} and include more than 2,000 stars observed from 1966 to 1995.
There are 1,155 HK observations for HD26965 from JD=2439787.8 to 2449771.7. The processed data do not include associated uncertainties to the calibrated S-values. According to \citet{Duncan1991}, the uncertainties in the Mount Wilson S-values can be calculated using the weights, $W$, included in the data that are derived from the photon counts of the measurements. The uncertainty in the S-index measurements is simply defined as $\sigma_{S}= S \,(\sqrt{W})^{-1}$. We applied this formula to the reported weights to provide proper uncertainties for the measurements of this star. All HK values for HD26965 can be found in Table \ref{tab:hk_data}.
\begin{figure}\label{fig:hk}
\centering
\includegraphics[scale=0.57]{F14.pdf}
\caption{{\it Top:} Mount Wilson S-values for HD26965. Second, Third \& Bottom Panels: Periodogram of the Mount Wilson S-values, periodogram of residuals after removing $\sim$4100 days and periodogram of residuals after removing $\sim$715 days, respectively. Red and black dashed lines mark the position of the possible rotation periods of $\sim$37 and $43$ days, respectively. }
\end{figure}
We proceeded to run the same periodogram analysis as for the radial velocities and the photometric time-series from ASAS.
A clear long-term variability of $\sim$4,100 days can be identified (see top and second top panels in Figure \ref{fig:hk}), providing evidence for a long-period magnetic activity cycle, similar to the long-period solar cycle. After removing this signal by modeling it with a sinusoidal function, a second period of $\sim$715 days is found in the periodogram (second bottom panel in the figure), likely representing another, shorter period magnetic cycle. Finally, the bottom panel in the figure shows the periodogram of the Mount Wilson S-values after removing the 4,100 and 715 day period signals from the data. In this residual periodogram a signal of $\sim$42.3 days remains in the data. This is most likely the value reported in \citet{Baliunas1996}.
The peak is clearly not unique, casting some doubt on its reality, but given that it matches the detected signal in our radial velocity data sets, and rotation periods are known to be quasi-period due to differing spot patterns, changing stellar activity levels, and differential rotation, we must entertain the real possibility that this is actually the rotational period of the star, and not the 38 day period that we found in the measured spectroscopic activity indices. If this is the reality, then HD26965 represents a case where most of the current suite of tests that we employ to detect planets using radial velocity analyses, fail to remove the noise introduced by the rotation of the star, meaning that we now require better methods to be employed on stars where there are clear correlations between the radial velocities and various activity indicators if we want to detect planets that induce amplitudes at the $\sim1$ m s$^{-1}$ level.
\section{Conclusions}\label{sec:conc}
Through the application of generalized Lomb-Scargle periodograms and tempered MCMC samplings, we conclude that there is a strong periodic signal in the radial velocities of the quiescent and slow rotating K dwarf HD26965. If interpreted as the Doppler signal induced on the star by an orbiting planet, our best solution explains these variations by the presence of a low-mass, super-Earth planetary candidate that has a minimum mass of 6.92$\pm$0.79 M$_{\oplus}$ orbiting the host star with a period of 42.364$\pm$0.015 days and at a distance of 0.215$\pm$0.008 AU.
Analysis of V-band photometry from ASAS does not show any significant periodic signal. However, since the amplitude of the signal is small, the precision of the data is not sufficient to detect the signal within the noise of the photometry.
Analysis of the stellar activity indicators does not show statistical evidence supporting a chromospheric origin for the periodic variations in the radial velocities, although we have found correlations between the radial velocities and the activity indices from the different spectrographs. However, when we analyze the independently acquired chromospheric calcium S-indices from the Mt. Wilson HK project, and after removing two long period activity cycles, we find evidence for the rotation of the star closely matching the period of the radial velocity detected signal.
Regarding this last point, we note that although it is important to properly include activity correlations into any global model of radial velocity data, which when done for this data set we find a higher statistical probability for the given Keplerian model supporting the planetary signal, if there are statistically significant correlations between activity indicators and the velocity measurements, then additional external activity indicators should be acquired, where possible. Also, moving away from linear correlation models between current activity indices and the radial velocities may be necessary, particularly if the data suggests more complex models, such as quadratics, might be favored. In any case it is clear that the inclusion of multiple sources of external data that also rule out possible magnetic cycles and rotation periods as the source of any radial velocity signal, can help to maintain the lowest false-positive rate for any given Doppler survey.
In summary, despite all the evidence favoring a Doppler signal present in this radial velocity data set, the methods described in this paper do not seem to be able to disentangle weak planetary signals from residual photospheric noise, at least when the orbital periods are close to the rotation period of the star and there are correlations present between the velocities and the measured activity indicators.
\acknowledgments
MRD acknowledges the support of CONICYT-PFCHA/Doctorado Nacional-21140646, Chile. JSJ acknowledges support by Fondecyt grant 1161218 and partial support by CATA-Basal (PB06, CONICYT). MGS is supported by CONICYT-PFCHA/Doctorado Nacional-21141037, Chile. The authors wish to recognize and acknowledge the very significant cultural role and reverence that the summit of Mauna Kea has always had within the indigenous Hawaiian community. We are most fortunate to have the opportunity to conduct observations from this mountain. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France.
\newpage
\begin{table*}
\center
\caption{HIRES Radial Velocities of HD26965}\label{tab:firstset}
\label{tab:hiresrv}
\begin{tabular}{lccc}
\hline\hline
\multicolumn{1}{l}{BJD}& \multicolumn{1}{c}{RV}&\multicolumn{1}{c}{$\sigma$ RV}& \multicolumn{1}{c}{S}\\
\multicolumn{1}{l}{}& \multicolumn{1}{c}{(m s$^{-1}$)}&\multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(dex)}\\ \hline
2452235.83300& -1.415& 1.2524& 0.1792\\
2452236.85549& -1.805& 1.3973& 0.1887\\
2452237.89810& 2.040& 1.3150 &0.1679\\
2452307.73757& -2.992& 1.4854& 0.1785\\
2452536.99956& -1.836& 1.4599& 0.1609\\
2452601.99297& -4.451& 1.2644& 0.1812\\
2452856.13402& 1.361& 1.5298& 0.1630\\
2452856.13536& -3.705& 1.3752& 0.1605\\
...&...&...&...\\
\hline
\end{tabular}
\tablecomments{This table is published in its entirety in the machine-readable format. A portion is shown here for guidance regarding its form and content.}
\end{table*}
\begin{table*}
\center
\caption{PFS Radial Velocities of HD26965}
\label{tab:pfsrv}
\begin{tabular}{lccc}
\hline\hline
\multicolumn{1}{l}{BJD}& \multicolumn{1}{c}{RV}&\multicolumn{1}{c}{$\sigma$ RV}& \multicolumn{1}{c}{S}\\
\multicolumn{1}{l}{}& \multicolumn{1}{c}{(m s$^{-1}$)}&\multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(dex)}\\ \hline
2455852.81626& 0.0445& 1.1825& 0.1541\\
2455852.81753& -2.5663& 1.1740& 0.1590\\
2455852.81876&-4.0511& 1.0451& 0.1595\\
2455852.82000& -2.2046& 1.1274& 0.1606\\
2456175.89728& -0.5003& 0.9877& 0.1511\\
2456285.67699& -3.6936& 0.8562& 0.1526\\
2456285.67840& -2.8543& 0.8220& 0.1530\\
2456285.67978& -3.6332& 0.8654& 0.1529\\
... &...&...&...\\
\hline
\end{tabular}
\tablecomments{This table is published in its entirety in the machine-readable format. A portion is shown here for guidance regarding its form and content.}
\end{table*}
\vspace{-0.5cm}
\begin{table*}
\center
\caption{CHIRON Radial Velocities of HD26965}
\label{tab:chironrv}
\begin{tabular}{lcc}
\hline\hline
\multicolumn{1}{l}{BJD}& \multicolumn{1}{c}{RV}&\multicolumn{1}{c}{$\sigma$ RV}\\
\multicolumn{1}{l}{}& \multicolumn{1}{c}{(m s$^{-1}$)}&\multicolumn{1}{c}{(m s$^{-1}$)}\\ \hline
2456941.80711& 0.7180& 1.5175\\
2456941.81054& 1.9459& 1.4766\\
2456941.81463& 0.9758& 1.6280\\
2456942.79556& 2.0399& 1.4729\\
2456942.79920& 0.6670& 1.5552\\
2456942.80289& 0.7219& 1.4767\\
2456943.75367& 1.8081& 1.5570\\
2456943.75739& 1.0934& 1.7136\\
...&...&...\\
\hline
\end{tabular}
\tablecomments{This table is published in its entirety in the machine-readable format. A portion is shown here for guidance regarding its form and content.}
\end{table*}
\vspace{-0.5cm}
\begin{table*}
\centering
\caption{HARPS OLD Radial Velocities of HD26965}\label{tab:lastset}
\begin{tabular}{lccccccc}
\hline\hline
\multicolumn{1}{l}{BJD}& \multicolumn{1}{c}{RV}&\multicolumn{1}{c}{$\sigma$ RV}& \multicolumn{1}{c}{BIS}& \multicolumn{1}{c}{FWHM}& \multicolumn{1}{c}{S}& \multicolumn{1}{c}{H$_{\alpha}$}& \multicolumn{1}{c}{He \sc i}\\
\multicolumn{1}{l}{}& \multicolumn{1}{c}{(m s$^{-1}$)}&\multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(dex)}& \multicolumn{1}{c}{(dex)}&\multicolumn{1}{c}{(dex)}\\ \hline
2452939.80613& -0.434& 0.519&1.254& 5896.904& 0.1753& 0.1161& 0.5067\\
2452939.80685& 0.231& 0.512&1.279& 5897.505& 0.1742& 0.1163& 0.507\\
2452939.80756& -0.367& 0.572& 1.387& 5901.748& 0.1747& 0.1163& 0.505\\
2452939.80827& 0.176& 0.524& 1.720& 5898.172& 0.1759& 0.1153& 0.5061\\
2452939.80899& 2.305& 0.954& 3.537& 5899.064& 0.1726& 0.1152& 0.5039\\
2452939.80969& 1.191& 0.612& 1.192& 5899.194& 0.1722& 0.1163& 0.5075\\
2452940.76906& -3.630& 0.378& 0.184& 5898.594& 0.175& 0.1157& 0.5076\\
2452945.76432& -3.375& 0.353& 2.281& 5896.436& 0.171& 0.1153& 0.5067\\
...&...&...&...&...&...&...&...\\
\hline
\end{tabular}
\tablecomments{This table is published in its entirety in the machine-readable format. A portion is shown here for guidance regarding its form and content.}
\end{table*}
\begin{table*}
\centering
\caption{HARPS NEW Radial Velocities of HD26965}\label{tab:lastset}
\begin{tabular}{lccccccc}
\hline\hline
\multicolumn{1}{l}{BJD}& \multicolumn{1}{c}{RV}&\multicolumn{1}{c}{$\sigma$ RV}& \multicolumn{1}{c}{BIS}& \multicolumn{1}{c}{FWHM}& \multicolumn{1}{c}{S}& \multicolumn{1}{c}{H$_{\alpha}$}& \multicolumn{1}{c}{He \sc i}\\
\multicolumn{1}{l}{}& \multicolumn{1}{c}{(m s$^{-1}$)}&\multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(m s$^{-1}$)}& \multicolumn{1}{c}{(dex)}& \multicolumn{1}{c}{(dex)}&\multicolumn{1}{c}{(dex)}\\ \hline
2457274.86039 & -0.1517 & 0.1828 &18.5315 &5951.1924 & 0.1777 & 0.1170 &0.5055\\
2457274.86304 &-0.0462 &0.1852 &18.4779 &5951.7017 & 0.1768 & 0.1171 &0.5059\\
2457274.86566 & 0.4185 &0.2041 &18.6366 &5951.8066 & 0.1777 & 0.1174 &0.5059\\
2457274.86845 &0.3169 &0.2026 &18.1923 &5951.7324 & 0.1778 & 0.1173 & 0.5051\\
2457277.84809 & 0.4107 &0.2189 &18.9974 &5950.3081 & 0.1763 & 0.1179 & 0.5056\\
2457277.85019 & 0.1997 &0.2209 & 18.6706 &5950.1743 & 0.1777 &0.1176 & 0.5049\\
...&...&...&...&...&...&...&...\\
\hline
\end{tabular}
\tablecomments{This table is published in its entirety in the machine-readable format. A portion is shown here for guidance regarding its form and content.}
\end{table*}
\begin{table*}
\center
\caption{ASAS Photometry of HD26965}\label{tab:asas_data}
\begin{tabular}{lcccccccccccc}
\hline\hline
\multicolumn{1}{l}{HJD -2450000} & \multicolumn{1}{c}{MAG 4}&\multicolumn{1}{c}{MAG 0}& \multicolumn{1}{c}{MAG 1}& \multicolumn{1}{c}{MAG 2}& \multicolumn{1}{c}{MAG 3}& \multicolumn{1}{c}{MER 4}& \multicolumn{1}{c}{MER 0}& \multicolumn{1}{c}{MER 1}&\multicolumn{1}{c}{MER 2}& \multicolumn{1}{c}{MER 3}& \multicolumn{1}{c}{GRADE}& \multicolumn{1}{c}{FRAME}\\ \hline
1953.56936 & 29.999 & 29.999 & 29.999& 29.999& 29.999 & 0.028 &0.063& 0.050& 0.034& 0.029 & C &9642 \\
2172.77457 & 5.375 & 5.451 &5.324& 5.361 & 5.374 & 0.033 &0.054& 0.035 &0.027& 0.029 &A &32848 \\
2206.76432 &29.999 & 29.999 &29.999& 29.999 &29.999 & 0.030 &0.084& 0.057 &0.037& 0.031 &C &37541 \\
2227.69007 & 5.274 & 4.831 &4.866 & 5.055 &5.187 & 0.030 &0.046 &0.041 &0.031 &0.032 &A &39837 \\
2230.68849 & 5.235 & 4.866 &4.942 & 5.091 &5.179 & 0.038 &0.107 &0.090 &0.063 &0.050 &D &40329 \\
2234.67705 & 5.263 & 4.766 &4.821 & 5.023 &5.175 & 0.032 &0.046 &0.040 &0.032 &0.035 &A &40805\\
2236.67316 &29.999 & 29.999 &29.999& 29.999 &29.999 & 0.034& 0.046& 0.039 &0.031& 0.036 & C &41118 \\
2501.90399 &29.999 & 29.999 &29.999& 29.999 &29.999 & 0.033& 0.044& 0.044 &0.035& 0.038 & C &16126 \\
2529.79004 & 4.598 & 3.968 &4.145 & 4.378 &4.523 & 0.044 &0.060 &0.052 &0.043 &0.046 &B &19144 \\
2549.77699 & 4.867 & 5.224 &4.976 & 4.923 &4.875 & 0.028 &0.060 &0.056 &0.040 &0.033 &B &20837 \\
2553.76674 & 4.727 & 4.377 &4.445 & 4.585 &4.681 & 0.035 &0.049 &0.052 &0.043 &0.041 &B &21379 \\
2558.77118 &29.999 & 29.999 &29.999& 29.999 &29.999 & 0.028& 0.052& 0.039& 0.028& 0.027 & C& 22103 \\
2655.59412 & 4.934 & 6.289 &5.861 & 5.440 &5.139 & 0.029 &0.036 &0.037 &0.027& 0.029 &A &35213 \\
2954.75032 & 4.916 & 5.227 &5.076 & 4.981 &4.927 & 0.032 &0.032 &0.037 &0.028& 0.029 &A &79986 \\
... & ... & ... & ... & ... & ... & ... & ... & ... & ... & ... & ... & ... \\ \hline
\end{tabular}
\tablecomments{This table is published in its entirety in the machine-readable format. A portion is shown here for guidance regarding its form and content.}
\end{table*}
\begin{table*}
\center
\caption{Mount Wilson HK Project measurements of HD26965}\label{tab:hk_data}
\begin{tabular}{lcc}
\hline\hline
\multicolumn{1}{l}{S$_{\rm MW}$} & \multicolumn{1}{c}{JD - 2444000}&\multicolumn{1}{c}{W} \\ \hline
0.223 & -523.2 & 15956.8 \\
0.207 & -522.2 & 16877.2 \\
0.234 & -457.2 & 7264.7 \\
0.206 & -219.2 & 17620.2 \\
0.218 & -175.2 & 17201.9 \\
0.213 & -172.2 & 17475.1 \\
0.219 & -107.2 & 34054.6 \\
0.205 & 172.8 & 8716.2 \\
0.227 & 190.8 & 17363.2 \\
0.221 & 200.8 & 10277.2 \\
0.214 & 224.8 & 8639.2 \\
0.205 & 260.8 & 8593.6 \\
0.217 & 278.8 & 8700.7 \\
0.194 & 482.9939 & 3409.7 \\
... & ... & ... \\
\hline
\end{tabular}
\tablecomments{This table is published in its entirety in the machine-readable format. A portion is shown here for guidance regarding its form and content.}
\end{table*}
\bibliographystyle{aa}
|
\section{Introduction}
The properties of galaxies we observe at the present epoch are the result
of a variety of physical mechanisms acting throughout their assembly history.
Galaxy evolution leads to a bimodal colour distribution that features
a red and a blue branch in the colour-stellar mass diagram
\citep[e.g. ][]{Baldry04}.
Interestingly, each colour branch is also associated to a different
galaxy morphology: early type galaxies are located in the
well defined red sequence, whereas late-type galaxies populate
the blue cloud \citep{conselice2006}.
The fraction of galaxies in the red sequence increases as a function of
local environmental density \citep{baldry2006},
especially at high stellar masses.
However, the environmental
dependence of the morphology fractions is milder than that of colour
fractions for low-masses \citep{Bamford09}.
Outliers of these two broad colour branches include galaxies in a so-called
intermediate green valley. These galaxies are interpreted as experiencing a
decrease of their star-formation rate (SFR)
\citep{Schawinski14,Powell17,Bait17,Coenda18}.
Star formation (SF) quenching is
higher for more massive galaxies,
and increases at fixed stellar mass with environmental density (given either
by local density or halo mass) as becomes
evident from the strong correlations of the fractions of quiescent galaxies with stellar mass, environment and halo-centric radius,
found both locally \citep{Wetzel12}
and at high redshift
\citep{Peng10, Muzzin12, Lin14, Jian17, Kawinwanichakij17}.
Besides, the way in which red and quiescent fractions vary with
environment and stellar mass are different for central and satellite galaxies
\citep{weinmann2006, Wetzel12, Kovac14, Darvish17, Smethurst17, Wang17}.
Central and satellite galaxies of a given stellar mass are subject
to very different
processes. Central galaxies are believed to reside at the centres of
their host dark matter (DM) haloes. They can receive new gas via cooling flows,
and they cannibalize satellite galaxies whose orbits in the group
potential decay due to dynamical friction.
Dry mergers (both minor and major)
can explain the shallower slope of the massive end of the red sequence
\citep{Jimenez11}. On the other hand, major mergers contribute
to produce massive, passive galaxies \citep{vanderWel09}, doubling the
mass of the massive brightest cluster galaxies since $z\sim 1$
\citep[e.g. ][]{Shankar15}.
While orbiting in the main
group, satellite galaxies are also affected by
galaxy encounters that may result in morphological transformations
\citep{Kannan15}. However, the key processes that are believed to quench
the SF in satellite galaxies are triggered by
gas disruption due to tidal stripping \citep[TS, ][]{Merritt83}
and ram pressure stripping (RPS, \citealt{gg72, Abadi99}).
During TS, the material is pulled
from a galaxy by the global tidal field of the host halo
or a larger neighbour galaxy, the former being
more relevant than the latter in low-velocity close encounters taking place
in galaxy groups \citep{Villalobos14}.
Tidal shocks at the pericentres of the orbit of satellite galaxies
induce dynamical
instabilities and impulsive tidal heating of the stellar distribution
that are particularly relevant for dwarf galaxies in environments
similar to our Local Group \citep{Kazantzidis11}.
Besides, the impact of TS (from negligible
effect to complete galaxy disruption) also depends on
the satellite morphology \citep{Chang13}.
RPS is a result of the ram-pressure (RP) exerted by the hot diffuse gas
of a group
or cluster of galaxies
on satellites moving through it at velocities that could
be close to supersonic.
This physical process may account for the gradual
removal of both the hot gas halo and the cold gas disc
\citep{BoselliGavazzi06,Jaffe13,Poggianti17}, giving place
to `jellyfish' galaxies in cases of extreme RPS \citep{Bellhouse17}
where the trailing stripped tails show signs of shock heating and gas
compression.
Hydrodynamical simulations show that, on one hand, RP contributes
to SF quenching through gas stripping and, on the other,
can also produce a temporary and moderate enhancement of SF
(more likely to occur in Milky Way-type disc galaxies)
as a result of gas compression that takes place
at pericentre passage
\citep{Bekki14, YozinBekki15, Steinhauser16, Ruggiero17}.
The instantaneous complete removal
of the hot diffuse gas halo of a galaxy after its
infall into a larger halo,
that takes place
regardless of which physical mechanism may actually be responsible
for it,
has been standard ingredient in SAMs.
This effect is generally referred to as
strangulation
\citep*{larson80,balogh2000}.
Although this simple prescription for halo stripping can
account for the presence of passive galaxies in clusters at high
redshift, it results in fractions of red satellite galaxies which are
higher than those observed in groups of galaxies
\citep[e.g.][]{baldry2006, weinmann2006, kimm2009};
this excess of red satellites constitutes the satellite overquenching problem.
Based on these results, there is
now general consensus that the instantaneous stripping of the hot gas
of satellites due to shock-heating
is a crude approximation
of the process. This is additionally supported by the observational evidence
that large fractions of near-IR-bright, early-type galaxies in groups
\citep{jeltema2008} and also in clusters \citep{sun2007, Wagner17}
have extended
X-ray emission, indicating that they retain significant hot gas haloes
even in these dense environments.
Studies based on smoothed particle hydrodynamics (SPH) simulations
have found that the hot gas haloes of
satellites are not stripped instantly \citep{mccarthy2008}, although
RPS of halo gas is much more effective than that of disc gas
\citep{bekki2009}.
Moreover, the effect of RPS on the hot corona seems to be
even milder than expected because of
turbulence in the interface with the intracluster medium (ICM)
and fuelling of gas from this
component to the satellite hot gas halo \citep{Quilis17}.
This mounting evidence motivated several authors
to include prescriptions
for a gradual removal of hot halo gas in SAMs,
with the inherent limitations of this numerical technique.
Some of them only consider the effect of RPS
\citep[e.g. ][]{Kang08, font2008, weinmann2010, GonzalezPerez14}
while others also take into account the
influence of TS \citep{guo11, kimm2011, Henriques17, Stevens17}.
If hot gas removal proceeds
gradually by the action of RPS and/or TS, the
time-scale of SF quenching
becomes longer and the
discrepancies in the fraction of red/quiescent galaxies
between models and observations are alleviated
although some disagreements remain.
The prescriptions adopted are generally based on
the analytic formulation of \citet{mccarthy2008}.
The effect of RPS on the cold gas of galaxy discs
has been implemented only in few SAMs
\citep{on2003,lanzoni2005, bdl2008, tecce10, luo16, Stevens17}.
The combination of suppressed cooling flows due to
strangulation and RPS of cold disc gas as in \citet{tecce10}
worsens the problem of excess of
red satellites. The presence of hot gas halo during gradual
removal acts as a shield that helps to prevent a too efficient action of
RPS on the cold gas phase. Although this is taken into account
by \citet{Stevens17}, the condition used
seems to still produce larger amounts of removed cold gas than
expected.
The star formation history of galaxies and, consequently,
the fractions of star forming and passive galaxies, are
strongly affected by supernova (SN) feedback.
\citet{hirschmann16} tested the impact of different
schemes for stellar feedback
and gas recycling using the semi-analytic model \textsc{Gaea}.
They find that a model characterized by
a strong ejection coupled to
a mass-loading that is explicitly dependent
on redshift
is reasonably successful
in reproducing the observed `anti-hierarchical' trends
in galaxy mass assembly, with
a delayed
metal enrichment and a realistic present-day stellar and
gaseous metallicity.
They adopt a redshift dependence of
the mass-loading factor (ratio between the gas outflow rate
from galaxies
and the star formation rate) similar to the one provided by \citet{muratov15}, who analyse the galaxy-scale gaseous outflows in the cosmological hydrodynamical
zoom FIRE (Feedback
in Realistic Environments) simulations.
In this work, we present
an updated version of the semi-analytic model
\textsc{sag}~\citep[acronym for Semi-Analytic Galaxies,][]{cora2006, lcp08, orsi14,
munnozarancibia2015, Gargiulo15},
which includes an improved treatment of
environmental effects
through the implementation of
the gradual removal of hot gas in satellites by RPS and TS,
allowing also the action of these processes on the cold gas disc,
under certain conditions; the modelling of
supernova (SN) feedback is also modified.
This version of \textsc{sag}~was used to generate one of the galaxy catalogues of the
\textsc{MultiDark Galaxies} project \citep{knebe17},
which is based on the Planck cosmology $1\,h^{-1} \,{\rm Gpc}$
\textsc{MultiDark} simulation MDPL2
\citep{Klypin16}. This catalogue is publicly available%
\footnote{\url{http://dx.doi.org/10.17876/cosmosim/mdpl2/007}}
in the \textsc{CosmoSim} database%
\footnote{\url{https://www.cosmosim.org}}
together with those generated
by the semi-analytic models \textsc{Sage} \citep{Croton16}
and \textsc{Galacticus} \citep{Benson12}.
This paper presents a detailed description
of the modifications introduced,
emphasizing the advantages with respect to previous works.
Details of
the MDPL2 simulation and a brief summary of the basic
aspects of \textsc{sag}~are given in Section~\ref{sec:model}.
Improvements introduced regarding
environmental effects and SN feedback are explained
in Sections~\ref{sec:env} and~\ref{sec:snfeedFIRE}, respectively.
The method applied to tune the free parameters of the model
and the comparison of model results
with observational constraints imposed to calibrate the
model are presented in Section~\ref{sec:calibration}.
Section~\ref{sec:SAG-predictions}
includes an analysis of general
galaxy properties predicted by \textsc{sag}~\citep[complementing the discussion already presented in ][]{knebe17},
and of those related with SF quenching:
dependence of the fraction of quenched galaxies as a function of stellar mass,
halo mass and halo-centric distance.
We show that the relations involving the fraction of passive galaxies
are better reproduced by modifying a parameter introduced in
the modelling of SN feedback which has impact on the SF history of galaxies.
For this variant of the model, we analyse the role of environmental
processes in Section~\ref{sec:ts-rps}, focussing on the stripped mass and atomic gas content.
In Section~\ref{sec:SAMcomp}, we compare details of our model and its main results with other SAMs.
In Section~\ref{sec:conclu}, we present the summary and conclusions of this work.
Appendix~\ref{ap:rpscalc} describes the estimation
of the hot gas stripping radius due to RP.
\section{Galaxy formation model}
\label{sec:model}
Our semi-analytic model of galaxy formation and evolution
\textsc{sag}~uses DM haloes extracted from a cosmological
DM simulation and their corresponding merger trees as the
basic inputs to construct the galaxy population.
We use the \textsc{MultiDark} simulation MDPL2,
which is part of the \textsc{CosmoSim} database.
\subsection{MDPL2 simulation}
\label{sec:mdpl2sim}
MDPL2 simulation follows the evolution of $3840^3$ particles within a box of
side-length $1\,h^{-1}\,{\rm Gpc}$,
with a mass resolution
$m_\textrm{p} = 1.5 \times 10^{9}\, h^{-1}\, \textrm{M}_{\odot}$ per
DM~particle. This simulation is
an analogue of the MDPL simulation described in \cite{Klypin16},
but ran with a different realisation of the initial conditions.
It is consistent with
a flat $\Lambda$CDM model characterized by Planck cosmological parameters:
$\Omega_{\rm m}$~=~0.307, $\Omega_\Lambda$~=~0.693,
$\Omega_{\rm B}$~=~0.048, $n_{\rm s}$~=~0.96
and $H_0$~=~100~$h^{-1}$~km~s$^{-1}$~Mpc$^{-1}$, where $h$~=~0.678
\citep{Planck2013}.
DM haloes have been identified with the
\textsc{Rockstar} halo finder
\citep{Behroozi_rockstar}, and merger trees were constructed with
\textsc{ConsistentTrees} \citep{Behroozi_ctrees}.
Overdensities with at least $N_\text{min}$~=~20~DM
particles were considered in the detection by the halo finder.
Each halo is characterized by measuring the physical properties
defined by the particle distribution, assuming spherical overdensity
approximation.
The virial mass is defined as the mass enclosed by a sphere of radius
$r_\text{vir}$, so that the mean density reaches a constant factor $\Delta=200$ times
the critical density of the Universe $\rho_\textrm{c}$, i.e.
\begin{equation}
M_\textrm{vir}(<r_\textrm{vir}) = \Delta \rho_\textrm{c}
\frac{4 \pi}{3} r_\textrm{vir}^3.
\label{eq:Mvir}
\end{equation}
Moreover, the virial velocity of each halo is defined in terms of these
properties as $V_\textrm{vir}= \sqrt{G M_\textrm{vir}/r_\textrm{vir}}$,
where $G$ is the gravitational constant.
The detected DM haloes can be over the background density or lie
within another DM haloes. To differentiate them, henceforth the
former will be referred to as \textsl{main host} haloes, whereas the latter as
subhaloes.
The calculation of the physical properties of subhaloes
considers only
the bound particles of the substructure identified by the halo finder.
\subsection[]{Semi-analytic model of galaxy formation SAG}
\label{sec:sag}
The version of our semi-analytic model of galaxy formation and evolution
\textsc{sag}~presented here is a further
development of the model described by \citet{cora2006},
which is based on \citet{springel2001},
and later improved in \citet{lcp08}, \citet{tecce10},
\citet{orsi14}, \citet{munnozarancibia2015} and
\citet{Gargiulo15}.
\textsc{sag}~includes the effects of radiative cooling of hot gas,
star formation, feedback from SN explosions,
chemical enrichment with a scheme that
tracks several chemical elements contributed by
different sources (stellar winds and supernovae Type Ia and II)
taking into account the lifetime of progenitors
\citep{cora2006},
growth of supermassive black holes (BHs) in galaxy centres and the consequent
AGN feedback, and starbursts triggered by disc
instabilities or galaxy mergers \citep{lcp08}.
These starbursts contribute to the formation of a bulge component,
which sizes are estimated as described in \citet{munnozarancibia2015};
the cold gas that has been transferred to the bulge is gradually
consumed, thus starbursts are
characterised by a time-scale
\citep{Gargiulo15}
instead of being instantaneous.
\citet{tecce10} added the effect of RPS on the cold
disc gas of satellite galaxies
implementing the criterion
from \citet{gg72}.
Two important new features have been added to \textsc{sag}~as described in the present work.
One is related to the inclusion of an explicit redshift dependence
in the estimation of the reheated and ejected mass as a result
of SN feedback,
guided by one of the ejective feedback schemes implemented in the \textsc{Gaea} semi-analytic model \citep{hirschmann16}.
This redshift dependence becomes crucial to reproduce
observed galaxy properties at high redshift, as we will show in
Sections~\ref{sec:SAG-const} and \ref{sec:SAG-predictions}.
The second major improvement
replaces the strangulation scheme
that removes the hot gas from satellites instantaneously
(standard practice in SAMs until recently)
by a gradual removal of the halo gas by the
action of TS and/or RPS (gradual starvation%
\footnote{
The process
of removal of the hot gas haloes of satellite galaxies is usually
indistinctly called `strangulation' or `starvation', regardless of which
physical mechanism may actually be responsible for it.
In this study, we will use the term `strangulation' to refer to the
instantaneous removal of hot gas from satellites, and we
will use `gradual starvation' to refer to the gradual removal of halo gas by any process.}).
The features that the current version of
\textsc{sag}~inherits from its predecessors are described in \citet{Gargiulo15},
except for those related to the RPS of cold
gas \citep{tecce10},
which will be described in Section~\ref{sec:env} along with the other
environmental processes introduced in this work.
The new SN feedback scheme is described in Section~\ref{sec:snfeedFIRE}.
Furthermore, the model of radio mode AGN feedback, in which
black holes generate jets and bubbles as they accrete gas from
hot atmospheres suppressing gas cooling, has been replaced by that
described in
\citet[][see their eq. S24 in the supplementary material]{henriques_mcmc_2015}.
This means that
the dependences with the fraction of hot gas with respect to the virial mass
and with the virial velocity of the halo (introduced by \citealt{lcp08})
are substituted by a dependence
with the hot gas mass. The aim of this change is to make AGN feedback more
efficient at late times.
The semi-analytic model assigns
one galaxy to each new detected halo in the simulation, and it
follows the halo merger trees to compute the evolution of the galaxy
properties.
Each considered system of haloes is constituted by only one
group/cluster central
galaxy, the one associated to the main host halo,
so that the other galaxies act as satellites.
When two haloes merge, the smaller one loses mass
to tides as it orbits within the larger structure, until
the satellite
is no longer identified by the halo finder.
We assume that the galaxy contained within
this disappeared subhalo
survives until it eventually merges with the central galaxy of its
host halo. During this temporary stage, these galaxies are called
\textsl{orphan} satellite galaxies.
\section{Environmental processes}
\label{sec:env}
As satellite galaxies orbit within their main host haloes,~DM and
the baryonic components in their
own subhaloes
are affected by numerous environmental processes.
We model mass stripping produced by tidal forces and RP.
TS will act on every component of the galaxy (stars, gas and~DM)
whereas~RPS is a hydrodynamical process which will only affect the gas.
Thus, a two-stage model for gas removal is
now included in \textsc{sag}. TS and RPS both act first on the hot gas halo: whichever
effect is stronger determines the amount of gas stripped.
Once a significant fraction of the hot halo is
removed, RPS can start affecting the cold gas disc, following
\citet{tecce10}.
The combination of the RPS of the cold gas phase
with the gradual starvation scheme
applied to the hot gas
constitutes an improvement
of our model
with respect to previous works that ignore the RPS of the cold gas
\citep[e.g. ][]{guo11, kimm2011, GonzalezPerez14}. Recently, this effect
has been taken into account by \citet{luo16} and \citet{Stevens17};
while the former authors do not consider the hot gas as a shield
that regulates the action of RP on the cold gas, the latter
take into account this possibility using a different criterion
than the one adopted in \textsc{sag}.
\subsection{Estimation of RP}
\label{sec:RPfit}
There have been numerous studies using hydrodynamic simulations
focused on the effects of RPS on cold gas discs
\citep[e.g.][]{Abadi99, RoedigerBrueggen07, Tonnesen07}
and hot gas haloes \citep{mccarthy2008, Quilis17}.
RP is defined as the product of the ICM density at the location
of the satellite galaxy and the square of its velocity relative to the ICM.
The importance of the local variations of the ambient gas makes
analytical calculations of the RP effect slightly inaccurate,
particularly for the resulting
size of the gas disc.
\citet{Tonnesen08}
show that these variations can produce
differences in the pressure of more than one order of magnitude at fixed
cluster-centric distance.
As demonstrated by \citet{tecce10} from the analysis
of adiabatic hydrodynamic simulations of galaxy clusters,
the values of RP present a radial profile mainly determined
by the dependence on cluster-centric distance of the ICM
\footnote{We refer to the hot gas associated with a
main host halo using the term `intracluster medium', even for host halo
masses smaller than that of a galaxy cluster.}
density.
Values of RP of the order of $10^{-12}\,h^2\,{\rm dyn}\,{\rm cm}^{-2}$,
$10^{-11}\,h^2\,{\rm dyn}\,{\rm cm}^{-2}$ and
$10^{-10}\,h^2\,{\rm dyn}\,{\rm cm}^{-2}$
are considered weak, medium and strong, respectively
\citep{RoedigerBrueggen06}.
At $z=0$,
mean RP values range from weak to medium
for clusters masses of $\approx 10^{14}\,{\rm M}_{\odot}$ and
$\approx 10^{15}\,{\rm M}_{\odot}$, respectively.
At $z\approx 2$, most of
satellite galaxies within progenitors of currently massive clusters are already
experiencing medium-level RP, and at $z \lesssim 0.5$
a significant fraction of galaxies ($\approx 20$ per cent)
suffer strong levels of RP \citep[see fig. 5 of ][]{tecce10}.
This evolution
depends mainly on the build-up of the
ICM density over time.
We take into account all these effects through
the implementation in \textsc{sag}~of fitting formulae that estimate the RP
experienced by galaxies in haloes of different
mass as a function of halo-centric distance and redshift,
which are derived from a procedure similar
to that described by \citet{tecce11}.
The new fit uses
alternative formulae that capture more adequately the RP behaviour,
as described in Vega-Mart\'inez et al. (in preparation)
using the self-consistent information provided by \citet{tecce10}
from the analysis of hydrodynamic simulations of galaxy clusters.
This procedure, called the `gas-particles' method in \citet{tecce10},
automatically takes into account any local variation
of the density and/or the velocity field
which may result from the dynamics of the gas in the hydrodynamic simulation.
The fitting formulae generated from these data allow
to include RP effects into SAMs applied
on dark matter-only simulations which lack gas physics, as the one used in this
work.
This way of estimating RP values
constitutes
the main advantage of our treatment of RPS.
\subsection{Orbits of orphan galaxies}
\label{sec:orbits}
For those satellites
with identified host subhaloes, the effects of dynamical
friction and TS are provided self-consistently by tracking the subhalo
evolution in the base simulation.
This approach breaks down at the subhalo resolution
limit, when the galaxy host by a subhalo that is no longer identified
becomes an orphan satellite.
In previous versions of \textsc{sag}, positions and velocities of orphans
were assigned using the current position and velocity of the
most-bound DM particle of the galaxy's host subhalo at the time
it was last identified \citep{lcp08, tecce10}, or
assuming a circular orbit with decaying radial distance
estimated from the dynamical friction and
initial halo-centric distance and velocity
determined by the virial radius and velocity of the host (sub)halo,
respectively
\citep{Gargiulo15}.
In any case, the merging criterion applied to orphans compares
the time elapsed since the time of infall
with a dynamical friction time-scale; when the former is larger than the latter,
an orphan galaxy is considered merged with a central galaxy regardless of the
relative distance between them.
In the current version
of \textsc{sag}, the phase-space of orphans is obtained from the integration of
the orbits of subhaloes that will give place to an orphan galaxy
once \textsc{sag}~is applied
to the merger trees extracted from the underlying
simulation,
as described in detail by
Vega-Mart\'inez et al. (in preparation).
When a subhalo is not longer tracked,
the last known position, velocity and virial mass are taken as
initial conditions to integrate its orbit numerically.
The orbital evolution
is consistent with the potential well of the
host halo, and takes into account mass loss by TS and dynamical friction effects,
following some aspects of the works by
\citet{gan2010} and \citet{kimm2011}.
The integration is carried out until
a merger event occurs, which is defined
when the halo-centric distance
becomes smaller than $10$ per cent of the virial radius of the host halo,
taken as an estimation of the radius of a central galaxy,
since in this stage previous to the application of \textsc{sag}~we do not have
information on the properties of galaxies that will populate these haloes.
This merging criterion allows to reach general good agreement with the redshift
evolution of the merger rate of galaxies obtained by considering
the merger time-scale given by \citet{jiang2008}, which is inferred from
a high-resolution cosmological hydro/N-body simulation with star formation.
\subsection{Assumptions for the action of RPS and TS}
We assume that the time-scale for RPS and TS
of any of the baryonic components is given
by~$t_\text{dyn,orb}=2\,\pi\,{\omega}^{-1}$, where $\omega$ is the angular
velocity of the satellite, according to \citet{Zentner05}.
In each interval~$\Delta T$ between simulation snapshots,
the RPS and/or TS effect removes
a fraction~$\Delta T/t_\text{dyn,orb}$ of gas and stars outside the
corresponding
stripping radius,~$r_\text{s}$,
if this fraction is less than unity; otherwise, the whole
estimated stripped mass,
$M_\text{strip}$,
is removed.
In the following, we describe the way in which the stripping
radius $r_\text{s}$ is derived for hot gas, cold gas disc,
disc stars and bulge stars,
referring to this quantity as $r_\text{s,hot}$, $R_\text{s,cold}$,
$R_\text{s,disc}$ and $r_\text{s,bulge}$, respectively.
\subsection[]{Gradual stripping of the hot gas haloes of satellites}
\label{sec:hotgasstrip}
The new version of \textsc{sag}~presented in this paper allows
galaxies to keep their hot gas haloes when they become
satellites. These haloes are then gradually stripped by the action of~RPS
and/or~TS, and as long as they
survive they can replenish the satellite's cold gas via gas cooling.
This is the
scheme we call gradual starvation.
In this new scenario, the mass of hot gas $M_\text{hot}$ available for cooling
to the central galaxy of each main host halo
is calculated at the beginning of each
simulation snapshot as
\begin{equation}
\begin{split}
M_\text{hot} = & f_{\rm b} M_\text{vir} - M_{\star,\text{cen}} -
M_\text{cold,cen} - M_\text{BH,cen} \\
& - \sum_{i=1}^{N_\text{sat}} \left( M_{\star,i} +
M_{\text{cold,}i} + M_{\text{BH,}i} + M_{\text{hot,}i} \right),
\end{split}
\label{eq:hotgas}
\end{equation}
where $f_{\rm b}$~=~0.1569 is the universal baryon fraction \citep{Planck2013}
and $M_\star$, $M_\text{cold}$~and
$M_\text{BH}$~are, respectively, the total masses of stars, of cold gas and
of the central black hole
(BH) of each of the galaxies contained within a given main host halo.
Central and satellite galaxies are considered separately in order
to discount the total mass in the remaining hot gas haloes of
the $N_\text{sat}$ satellite galaxies within the same main host halo.
Our prescription for~RPS of the hot gas halo is based on the
model from \citet{font2008}, which
uses the criterion for~RPS for a spherical distribution of gas determined by
\citet{mccarthy2008} from the results of hydrodynamic simulations. The gas beyond a
satellite-centric radius~$r_\text{sat}$ will be removed if the
value $P_{\rm ram}$ of RP, which is given by the fitting formulae
described in Section~\ref{sec:RPfit}, meets the
condition
\begin{equation}\label{eq:rpshot}
P_{\rm ram} > \alpha_\text{RP} \frac{G
M_\text{sat}(r_\text{sat}) \rho_\text{hot}
(r_\text{sat})}{r_\text{sat}},
\end{equation}
where $\rho_\text{hot}$ is the density of the satellite's hot gas halo,
and $\alpha_\text{RP}$ is a geometrical constant of order unity chosen to match the results of~hydrodynamic simulations.
\citet{mccarthy2008} find that their simulations support $\alpha_\text{RP}=2$.
They also test the effect of values as large as $\alpha_\text{RP}=10$, which
produce milder effects. Since they do not consider gas cooling that can
also reduce the hot gas content,
we adopt an intermediate value of $\alpha_\text{RP}=5$ in our model.
The total mass $M_\text{sat}$ of a satellite is
\begin{equation}\label{eq:msat}
\begin{split}
M_\text{sat}(r_\text{sat}) = & M_\star + M_\text{cold}\\
&+ 4\pi \int_0^{r_\text{sat}} \left[ \rho_\text{hot}(r) +
\rho_\text{DM}(r) \right] r^2 dr,
\end{split}
\end{equation}
assuming that $r_\text{sat}$ is large enough to contain all the stars and cold
gas.
All density profiles are represented using
isothermal spheres, i.e.,
$\rho_\text{hot} = {M_\text{hot}}/({4\pi\, r_\text{hot}\, r^2})$,
where $r_\text{hot}$ is the radius that contains all of $M_\text{hot}$.
This radius
initially adopts the value of the subhalo
virial radius, $r_\text{vir}$; in the case of orphan satellites, $r_\text{vir}$~preserves
the value corresponding to the last time the subhalo was identified.
Combining equations~\eqref{eq:rpshot} and~\eqref{eq:msat} one can
numerically solve for the hot gas stripping radius due to~RP,
$r_\text{s,hot}^\text{RPS}$ (see Appendix~\ref{ap:rpscalc}).
To determine the halo gas loss of satellite galaxies via TS
we assume that the hot gas distributes
parallel to the~DM.
The bounding radius for the DM,~$r_\text{DM}$,
is given by
$r_\text{vir}$.
This radius gives the tidal radius determined by TS, that is
$r_\text{s,hot}^\text{TS}=r_\text{DM}$, and
is compared with the hot gas stripping radius due to~RP,
$r_\text{s,hot}^\text{RPS}$.
The smaller of the
two will be the stripping radius~$r_\text{s,hot}$~which will contain the remaining
hot gas mass; all hot gas beyond that radius can be stripped.
Thus, the value of~$r_\text{hot}$~is updated such that
$r_\text{hot}^\text{new}=r_\text{s,hot}$.
Once the hot gas outside the new stripping radius is removed,
we assume that the
remaining gas quickly redistributes its mass and restores an isothermal profile,
but truncated at~$r_\text{hot}$,
as in~\citet{font2008} and \citet{kimm2011}.
The pressure of the ambient~ICM could act to confine the satellite's hot
gas, preventing any outflow from going beyond the stripping radius
\citep{mj2010}.
If in a subsequent simulation time step the value
obtained for~$r_\text{hot}$~is larger than the previous one,
$r_\text{hot}$~is not updated and no gas is lost in that time step.
In previous versions of \textsc{sag}, based on the strangulation scheme,
all feedback processes occurring in satellite
galaxies were assumed to transfer cold gas and its associated metals from the
satellite into the hot phase of the central galaxy of their main host
halo (i.e. the ICM). In our new implementation, we assume
that the hot gas halo of a satellite is depleted
when
the hot gas mass
drops below a fraction $f_\text{hot,sat}$ of the baryonic mass of the
satellite,
specified in
Section~\ref{sec:parameters}.
Before this situation takes place,
all feedback processes will transport gas
and metals to the hot phase of the same galaxy,
proportionally to its content of
hot gas; the remaining reheated mass and associated metals are transferred
to the corresponding central galaxy.
If the latter is a central galaxy of a subhalo (i.e. it is also a satellite)
and its
hot gas halo also drops below $f_\text{hot,sat}$ times the baryon fraction,
then the reheated gas and metals of the orphan satellite
are transferred to the main host halo.
In any given simulation time step,
all those galaxies, either central or satellite,
that still have a hot gas reservoir will proceed to cool gas.
Gas cooling rates are calculated using the simple model presented in
\citet{springel2001}, but considering
the total radiated power per
chemical element given by \citet{foster2012}.
\subsection[]{Ram pressure and tidal stripping of cold gas disc}
\label{sec_RPS_TS_cold}
Ram pressure exerted by the intra-group/cluster medium can also affect
the cold gas disc of satellite galaxies. Its action could
be regulated by
the presence of hot gas halo in the gradual starvation scheme.
Hot gas haloes in satellite galaxies will be replenished by SNe feedback as
long as there is some star formation and the ratio between
the hot gas halo and the baryonic mass of the galaxy is larger than
the fraction $f_\text{hot,sat}$
(see section~\ref{sec:hotgasstrip}). At a certain point, this ratio will
become smaller than $f_\text{hot,sat}$, and the hot gas halo will be gradually
reduced by gas cooling and/or stripping processes (RPS, TS) reaching very
low values.
We assume that for as long as this ratio is larger than $0.1$, the
hot gas halo shields the cold gas
disc from the action of~RPS.
Once this condition is not fulfilled anymore,
the ambient~RP starts affecting the cold gas.
This threshold has been chosen small enough to allow the role of the hot gas as a
shield for a sufficiently long time; the action of RPS on the cold disc gas
becomes too effective
without any restriction of this kind.
We consider the model for~RPS of cold gas disc introduced in \citet{tecce10},
which is
based on the simple criterion proposed by \citet{gg72}. The cold gas of the
galactic disc located at a galactocentric radius~$R$ will be stripped away if
the~RP exerted by the ambient medium on the galaxy exceeds the restoring force
per unit area due to the gravity of the disc,
\begin{equation}\label{eq:rps}
P_{\rm ram} > 2\pi G \Sigma_\text{disc}(R)
\Sigma_\text{cold}(R).
\end{equation}
Here $\Sigma_\text{disc}$, $\Sigma_\text{cold}$ are the surface densities of the
galactic disc (stars plus cold gas) and of the cold gas disc, respectively.
The discs of stars and gas are modelled by
an exponential surface density profile given by
$\Sigma(R) = \Sigma_0 \exp(-R/R_\text{d})$
where $\Sigma_0$ is the central surface density and~$R_\text{d}$ is the scale length
of the
disc. This scale length is estimated as
$R_\text{d}=(\lambda/\sqrt{2})R_\text{vir}$ \citep{mmw98},
where $\lambda$ is the spin parameter of the DM halo in which
the galaxy resides.
Starting from
condition~\eqref{eq:rps} it can be shown that~RPS will remove from a galaxy all
cold gas beyond a stripping radius~$R_\text{s,cold}$ given by
\begin{equation}\label{eq:rstrip}
R_\text{s,cold}^\text{RPS} = -0.5 R_\text{d} \ln \left( \frac{P_{\rm ram}}{2\pi G \Sigma_\text{0,disc}
\Sigma_\text{0,cold}} \right),
\end{equation}
where $\Sigma_\text{0,disc}$ and $\Sigma_\text{0,cold}$ are the central surface
densities of the stellar disc and of the cold gas disc, respectively. We assume
for simplicity that both disc components have the same scale length.
To account for the effect of tides, at each simulation
snapshot
the value for~$R_\text{s,cold}^\text{RPS}$ obtained from equation~\eqref{eq:rstrip} is compared
to~$r_\text{DM}$, which determines the stripping radius due to TS,
$R_\text{s,cold}^\text{TS}$;
if the former is smaller than the latter we
set~$R_\text{s,cold}=R_\text{s,cold}^\text{RPS}$.
If the resulting~$R_\text{s,cold}$
is smaller than the current value of the cold gas disc radius,~$R_\text{cold}$, all the gas beyond~$R_\text{s,cold}$
is stripped away.
The stripped gas is added to the hot gas component of the central
galaxy (either of a main halo or a subhalo).
If the latter is
central galaxy of a subhalo already stripped of halo gas, the stripped gas goes
to the ICM.
After a stripping event produced by RPS and/or TS, the
remaining disc gas is assumed to form an exponential disc
truncated at a
radius~$R_\text{cold}^\text{new}=R_\text{s,cold}$
and with a new scale length defined as
$R_\text{d,cold}^\text{new}=R_\text{cold}^\text{new}/7$
(assuming that $99$ per cent of the cold gas disc is contained within $7*R_\text{d,cold}$).
\subsection[]{Tidal stripping of stars}
\label{sec:TSstars}
Unlike~RPS,~TS may also affect the stellar components of a satellite galaxy. To
process the stellar stripping we consider the disc and bulge separately. In the
case of the disc, we compare the current value
of~$r_\text{DM}$, with the size of the
stellar disc, $R_{\rm disc}$.
If $R_{\rm disc} > r_\text{DM}$, we assume that the galaxy loses
all stars beyond~$r_\text{DM}$~and that the remaining disc stars still form an
exponential disc,
truncated at $R_{\text{s},\text{disc}} = r_\text{DM}$,
but with the stellar mass redistributed with a new scale length $R_{\text{d},\text{disc}} = R_{\text{s},\text{disc}}/7$, as it is considered for the cold gas disc.
The scale length of the galaxy disc composed by cold gas and stars is
then defined from the mass-weighted scale length of their respective
components as $R_\text{d}=(M_\text{cold}\,R_\text{s,cold}+M_{\text{disc}}\,R_{\text{s},\text{disc}})/(M_\text{cold}+M_{\text{disc}})$.
If a bulge component is present, to evaluate the TS of its stars we assume that
they are distributed according to a \citet{hernquist1990} profile,
\begin{equation}
\rho(r) = \frac{M_\text{bulge}}{2\pi}\, \frac{a_\text{b}}{r}\, \frac{1}{(r + a_\text{b})^3}
\end{equation}
where $a_\text{b}$ is a scale length related to the bulge half-mass radius by $r_\text{b,h}=
(1 + \sqrt{2}) a_\text{b}$. To calculate~$r_\text{b,h}$
we follow the procedure outlined in
\citet{cole2000}
and adapted to our model in \citet{munnozarancibia2015}.
If the radius that contains the $99$ per cent of the bulge mass, given by
$r_\text{bulge}=198.5*a_\text{b}$,
satisfies the condition $r_\text{bulge} > r_\text{s,bulge}$
with $r_\text{s,bulge}=r_\text{DM}$,
we assume that the stars in the bulge
beyond
$r_\text{s,bulge}$
are stripped and that the remaining bulge stars are
redistributed still following a Hernquist profile truncated at
$r_\text{s,bulge}$, with a new scale length
defined as $a_\text{b}^\text{new}=r_\text{s,bulge}/198.5$.
The stars removed by~TS
are assumed to orbit freely within the main host halo.
For practical purposes they are assigned to a new `stellar halo' component of
the corresponding central galaxy.
As these halo stars evolve and die, they inject
gas and metals directly into
the ICM.
\section{Supernovae feedback}
\label{sec:snfeedFIRE}
In previous versions of~\textsc{sag},
the amount of reheated mass produced by the SNe arising
in each star forming event
is assumed to be
\begin{equation}
\Delta M_{\rm reheated} = \frac{4}{3} \epsilon {\frac{\eta E_\text{SN}}{V_{\rm vir}^2}} \Delta M_{\star},
\label{eq:feedbackSN}
\end{equation}
where $\eta$ is the number of
SNe generated from the stellar population of mass $\Delta M_\star$ formed,
$E_\text{SN}=10^{51}\,{\rm erg}$ is the energy released
by a SN, $V_{\rm vir}$ is the virial velocity of the host (sub)halo, which
is a measure of its potential well, and $\epsilon$ is the
SNe feedback
efficiency, i.e.,
a free parameter that controls the amount of cold gas reheated by the
energy generated by SNe.
The number of SNe
depends on the initial mass function (IMF) adopted and for core collapse supernovae (SNe CC) is
estimated as
\begin{equation}
\eta = \frac{\int^{\infty}_8 \phi(m)\,\,{\rm d}m}{\int_0^{\infty} \phi(m)\, m\,\,{\rm d}m},
\end{equation}
\noindent where $m$ is
the stellar mass and $\phi(m)$ is the IMF.
This quantity
is constant for an universal IMF, like the Chabrier IMF
\citep{Chabrier03}
adopted here.
The energy and metals generated by SNe and stellar winds produced
by progenitors in different mass ranges are released with different time-scales
depending on the lifetime of the progenitors, as given by \citet{pm93}.
In order to satisfy observational constraints at high redshifts, we found it
necessary to modify this model of SNe feedback.
\citet{hirschmann16} find that the observed trends in galaxy assembly
are reproduced by replacing
the `fiducial' feedback scheme implemented in \textsc{Gaea}
model
with a parametrization of the mass-loading factor similar to
the one inferred by
\citet{muratov15} from the analysis of the FIRE simulations.
We use this information
to modify the estimation of the reheated mass (eq.~\ref{eq:feedbackSN})
by simply adding new factors that take into account the dependence on redshift and an additional modulation with virial velocity.
Thus, the feedback scheme in the
current version of~\textsc{sag}~produces a reheated gas mass given by
\begin{equation}
\Delta M_{\rm reheated} = \frac{4}{3} \epsilon {\frac{\eta E_\text{SN}}{V_{\rm vir}^2}} \,(1+z)^{\beta}\,\left(\frac{V_{\rm vir}}{60\,{\rm km}\,{\rm s}^{-1}}\right)^{\alpha_\text{F}}\Delta M_{\star},
\label{eq:feedfire}
\end{equation}
\noindent where
the exponent $\alpha_\text{F}$ takes the values $-3.2$ and $-1.0$
for virial velocities
smaller and larger than $60\,{\rm km}\,{\rm s}^{-1}$, respectively.
Although the fit provided by \citet{muratov15}
indicates that the value of $\beta$ is 1.3,
we consider this exponent to be a free parameter of
the model~\textsc{sag}, allowing it to absorb
aspects of physical processes that are not properly captured by the model
and helping to reproduce the observational constraints imposed,
detailed in Section~\ref{sec:parameters}.
In section~\ref{sec:SAG-const}, we show the results of
this procedure and also discuss the impact of assigning to
$\beta$
the value
found by \citet{muratov15}.
The reheated gas is transferred from the cold to the hot phase,
subsequently returning to the cold phase through gas cooling
taking place in both central and satellite galaxies.
However, to avoid an excess of stellar mass at high redshifts,
some of the hot gas must be ejected out of the halo reducing
the hot gas reservoir available for gas cooling \citep{guo11, henriques13,
hirschmann16}. Hence, we also consider the energy conservation argument
presented by \citet{guo11} to calculate the ejected hot gas mass
\begin{equation}
\Delta M_{\rm ejected}= \frac{\Delta E_\text{SN} - 0.5\,\Delta M_{\rm reheated}\,V_{\rm vir}^2}{0.5\,V_{\rm vir}^2},
\label{eq:EnergyCons}
\end{equation}
\noindent where $\Delta E_\text{SN}$ is the energy injected by massive stars
which we model in a way similar to the modified reheated mass,
as also done by \citet{hirschmann16},
\begin{equation}
\Delta E_{\rm SN} = \frac{4}{3} \epsilon_\text{ejec} {\frac{\eta E_\text{SN}}{V_{\rm vir}^2}} \,(1+z)^{\beta}\,\left(\frac{V_{\rm vir}}{60\,{\rm km}\,{\rm s}^{-1}}\right)^{\alpha_{\rm F}}\Delta M_{\star}\,0.5\,V_\text{SN}^2.
\label{eq:ejecfire}
\end{equation}
\noindent Here, $\epsilon_\text{ejec}$ is the corresponding efficiency,
considered as another free parameter of the model,
and $0.5\,V_\text{SN}^2$ is
the mean kinetic energy of SN ejecta per
unit mass of stars formed. Following the analysis of \citet{muratov15},
we adopt the fit for 95th percentile wind velocity as a function of the
virial velocity of the halo
(see their fig. 8 and eq. 10), such that
$V_\text{SN}=1.9\,V_\text{vir}^{1.1}$.
In the gradual starvation scheme implemented in~\textsc{sag}, galaxies
keep their hot gas halo when they become satellites. These
haloes are reduced by gas cooling and environmental effects (RPS, TS),
but they also can be reconstructed by the injection of reheated gas
or the reincorporation of ejected gas.
The ejected gas mass is assumed to
be re-incorporated back onto the (sub)halo from which it was expelled
within a time-scale that depends on the inverse of (sub)halo mass, $M_{\rm vir}$,
as assumed by \citet{henriques13} in order to reproduce the
observed evolution of the galaxy stellar mass function.
Thus, the reincorporated mass is given by
\begin{equation}
\Delta M_{\rm reinc}= \gamma\,\Delta M_\text{ejected}\,\frac{M_{\rm vir}}{10^{10}\,{\rm M}_{\odot}},
\label{eq:reinc}
\end{equation}
\noindent where the parameter $\gamma$ regulates the efficiency of the process
and is also a free parameter of \textsc{sag}~model.
In all the above parametrizations, we have added the energy injected by
SNe Ia. We adopt the single degenerate
model in which a SN Ia occurs by carbon deflagration in C--O
white dwarfs in binary systems whose components have masses
between $0.8$ and $8\,{\rm M}_{\odot}$ \citep{gr83}.
We implement the formalism presented by \citet{Lia2002},
choosing a fixed fraction of binary systems, $A_\text{bin}=0.05$.
As for SNe CC, we also consider the lifetime of the progenitors.
SNe CC and SNe Ia, together with low- and intermediate-mass stars,
contribute with metals that pollute the cold and hot gas, affecting
the cooling rates and regulating also in this way the subsequent events
of star formation.
Details on the chemical model implemented in~\textsc{sag}~are given in \citet{cora2006},
with the latest updates on chemical yields presented in \citet{Gargiulo15}.
\section{Calibration of the model SAG}
\label{sec:calibration}
The physical processes included in \textsc{sag}~model, as occurs for all SAMs,
involve several free parameters which regulate the modelling of these
processes.
Proper values of these free parameters are found by imposing constraints
given by observed galaxy properties that the SAM should satisfy.
\subsection{Calibration process}
\label{sec:parameters}
The calibration process of \textsc{sag}~model is performed by implementing the Particle
Swarm Optimization (PSO) technique presented in \citet{ruiz2015}.
This numerical tool allows us to tune the free parameters via the exploration of
the \textsc{sag}~parameter space by random walks of a set of `particles' (the
\textit{swarm}, equivalent to the chains in Monte Carlo Markov techniques)
that share information between them, thus determining the evolution of the
exploration from both their individual and collective experience.
By comparing the model results against a given set of observables, the PSO
method yields a set of best-fitting values for the free parameters.
For this current version of ~\textsc{sag}, we consider nine free parameters for the
calibration process.
The corresponding equations that involve these free parameters and a brief
description of the related physical processes can be found in \citet{ruiz2015},
except those related to the new SN feedback and ejection/reincorporation scheme
implemented in the current version of~\textsc{sag}, presented in
Section~\ref{sec:snfeedFIRE}.
The free parameters are
the star formation efficiency ($\alpha$), the
efficiency of SN feedback from stars formed in both the disc and the bulge
($\epsilon$, equation~\ref{eq:feedfire}), the efficiency of ejection of gas from the
hot phase ($\epsilon_\text{ejec}$, equation~\ref{eq:ejecfire}) and of its
reincorporation ($\gamma$, equation~\ref{eq:reinc}), the exponent that regulates the
redshift evolution of the mass-loading factor of the reheated and ejected mass
($\beta$, equations~\ref{eq:feedfire} and~\ref{eq:ejecfire}), the growth
of super massive BHs and efficiency of AGN feedback ($f_\text{BH}$ and
$\kappa_\text{AGN}$, respectively), the factor involved in the distance scale
of perturbation to trigger disc instability events ($f_{\rm pert}$), and the
fraction that determine the destination of the reheated cold gas
($f_\text{hot,sat}$) introduced in Section~\ref{sec:hotgasstrip}.
The set of observables used for the calibration process is the stellar mass
functions (SMF) at $z=0$ and $z=2$, the star formation rate distribution
function (SFRF), the fraction of mass in cold gas as a function of stellar mass
(CGMF) and the relation between bulge mass and the mass of the central
supermassive BH (BHB).
For the SMFs, we adopt the compilation data used by \citet{henriques_mcmc_2015},
which for $z=0$ is a combination of the SMF of the
SDSS
from \citet{baldry08} and \citet{li_smf_2009}, and of the Galaxy And
Mass Assembly (GAMA) from \citet{baldry_smf_2012}, while for $z=2$ is a
combination of the data of the Cosmic Evolution Survey (COSMOS) from
\citet{dominguezsanchez11}, the Ultra Deep Survey (UltraVISTA) from
\citet{muzzin13} and \citet{ilbert13}, and the FourStar Galaxy Evolution Survey
(ZFOURGE) from \citet{tomczak14}.
For the SFRF, which is the number density of galaxies in a certain SFR
interval, we use data from a flux-limited sample of galaxies observed with the
\textsl{Herschel} satellite which gives the total (IR+UV) instantaneous SFR for the
redshift interval $z \in [0.0,0.3]$, as presented in \citet{gruppioni15}.
This observable is compared with the \textsc{sag}~output at $z=0.14$, which is the closest
output of the model to $z=0.15$, the mean value of the redshift range of observations.
For the CGMF, we adopt observational data from \citet{boselli14}, which is
based on a volume limited sample, within the range
$\rm{log}(M_{\star}[{\rm M}_{\odot}]) \in [9.15,10.52]$ in stellar mass.
Finally, for the BHB relation, we combine the datasets from
\cite{mcconnell_bhb_2013} and \cite{kormendy_bhb_2013}.
This particular set of constraints and the observational data adopted were
defined as a common set for calibration during the Cosmic CARNage
workshop
aimed to compare different calibrated galaxy formation models based on the same
cosmological DM-only simulation \citep{Knebe18}.
This comparison project was initiated in the
nIFTy Cosmology workshop \citep{knebe15}.
Due to the large size of the simulation, the computing time needed to run the model several times becomes quickly prohibitively long. Hence,
the calibration is carried out by running \textsc{sag}~with the
merger trees extracted from a smaller box of the whole MDPL2 simulation that
constitute a representative sample of all the merger trees contained in the
$1\,h^{-1}\,{\rm Gpc}$ side-length box.
We divide the full volume of the MDPL2 simulation in
$9^3$ parallelepiped sub-boxes, each of them with roughly the same number density of
haloes of the whole volume.
Despite each of these samples constitutes only $\sim 0.137$ per cent of the
whole simulation, they are large enough to be considered themselves simulations
from which a statistically significant set of merger trees can be extracted for
the further application of the~\textsc{sag}~model.
Therefore, we compared the halo mass function for all these subboxes
with the one obtained for the whole simulation, choosing the most similar
one simply by visual inspection, selecting for this case a subbox with an
effective length side of $111.19\,h^{-1}\,{\rm Mpc}$.
\begin{table}
\centering
\begin{tabular}{l r r}
\hline
Parameter & Best-fitting value\\
\hline
$\alpha$ & 0.04 \\
$\epsilon$ & 0.33 \\
$\epsilon_{\rm ejec}$ & 0.022 \\
$f_{\rm BH}$ & 0.06 \\
$\kappa_{\rm AGN}$ & 3.02 $\times$ 10$^{-5}$ \\
$f_{\rm pert}$ & 14.56 \\
$\gamma$ & 0.055 \\
$f_{\rm hot,sat}$ & 0.277 \\
$\beta$ & 1.99 \\
\hline
\end{tabular}
\caption{Best-fitting values of the free parameters of \textsc{sag}~model obtained with the~PSO technique. This set of values is
obtained from the application of \textsc{sag}~to the merger trees
of the subbox selected from the MDPL2 simulation.
}
\label{table1}
\end{table}
The values of the parameters obtained from the calibration of the model to
match the aforementioned set of observables are shown in Table~\ref{table1}.
\subsection[]{Functions and relations used as constraints}
\label{sec:SAG-const}
The new features included in~\textsc{sag}~allow the model
to reach a good match with the observational constraints
imposed to calibrate the model, i.e.,
the SMF at $z=0$
and $z=2$, the SFRF at $z=0.14$, and the CGMF and BHB and relations at $z=0$,
as
we detail below.
In addition to the results of the calibrated model,
we also evaluate the impact on these functions and relations of fixing
the value of the parameter $\beta$
involved in the redshift dependence of the reheated and ejected mass.
For this alternative model, we adopt the value given
by the fit of \citet{muratov15}, i.e. $\beta=1.3$,
leaving the values of the remaining free parameters of \textsc{sag}~unchanged.
We refer to this slightly modified version of the model as \textsc{sag$_{\beta1.3}$}.
It is worth noting that the small value of the parameter $\gamma$ that regulates the amount of ejected gas that is reincorporated makes the results of the model insensitive to the fate of the reincorporated gas, i.e. subhalo from which it was ejected or main host halo where the satellite resides. We tested the latter possibility and found no change in the galaxy properties shown in this work.
\subsubsection[]{Stellar mass functions}
\begin{figure}
\centering
\includegraphics{Fig1.pdf}
\caption{
Stellar mass functions used as constraints to calibrate the \textsc{sag}~model.
Results obtained from the calibrated model \textsc{sag}~with the best-fitting
values given in Table~\ref{table1} are represented by red solid lines.
{\it Top panel}:
Stellar mass function at
$z=0$.
{\it Bottom panel}:
Stellar mass function at $z=2$.
Observational data
compiled by \citet{henriques_mcmc_2015} is represented by filled
circles with error bars.
The agreement between model and observations is rather good, except
for the excess at the high-mass end at $z=0$.
This excess becomes less evident when comparing with
estimations compiled and analysed
by \citet[][grey shaded area]{Bernardi17}, which include data from
\citet{Chen12} and \citet{Mendel14}.
In both panels, results obtained from model \textsc{sag$_{\beta1.3}$}~are shown by a blue dashed line.
The SMF at $z=0$ is not significantly affected by the change in the value of the
parameter $\beta$,
whereas this gives rise to an excess in the low-mass end of the SMF at $z=2$.
}
\label{fig:SAGconstraints_SMF}
\end{figure}
In Fig.~\ref{fig:SAGconstraints_SMF}
both the local SMF (top panel) and the SMF at $z=2$ (bottom panel)
generated by the model are shown, where
the break and low-mass end given by observations
are reproduced.
This is a due to the new implementation of SN feedback.
The classical estimation of the reheated mass
considered in previous versions of \textsc{sag}~(equation~\ref{eq:feedbackSN})
allows a good match of the SMF only at $z=0$.
The addition of ejection acting on the hot phase
and the redshift dependence of both reheated and ejected mass
(equations~\ref{eq:feedfire} and \ref{eq:ejecfire})
become key ingredients to avoid an excess of star formation
overproducing stellar mass at high redshifts.
This good agreement is obtained by a stronger redshift dependence than
the one found by \citet{muratov15}, whose parametrization is characterized
by a parameter $\beta=1.3$.
From Table~\ref{table1}, which shows the best-fitting values
of the free parameters of \textsc{sag},
we can see that the exponent that regulates this
redshift dependence
takes a value
$\beta=1.99$.
On the other hand, model \textsc{sag$_{\beta1.3}$}~produces an
excess in the low-mass end of the
SMF at $z = 2$, as shown by the dashed line
in Fig.~\ref{fig:SAGconstraints_SMF}. This is
due to the higher levels of
SFR this model features at higher redshifts, as shown and
discussed in Section~\ref{sec:SAG-pred-sfr}. However,
the SMF at $z = 0$ is not significantly affected by
the change in the value of the parameter $\beta$.
None of the improvements included in \textsc{sag}~leads to a better fit of the massive
end of the SMF at $z=0$, which shows an excess of galaxies with stellar mass
$M_{\star} \gtrsim 2\times 10^{11}\,{\rm M}_\odot$.
This excess
is not avoided by the implementation
of the more locally efficient radio mode AGN feedback
\citep{henriques_mcmc_2015}.
In any case, AGN feedback allows to recover the break of the SMF
which characteristic stellar mass highly depends
on the feedback efficiency regulated by the two free parameters
involved in the AGN feedback ($\kappa_{\rm AGN}$ and $f_{\rm BH}$).
However, the slope
of the high-mass end of the SMF
is insensitive to the AGN feedback efficiency,
as we have checked by manually exploring values of the parameters
related with this process around the best-fitting values.
It is important to take into account the uncertainties in the
observational data used as constraints,
considered in the comparison plot of the SMF at $z=0$.
As specified in Section~\ref{sec:parameters}, the compilation made by
\citet{henriques_mcmc_2015}
includes data from \citet{li_smf_2009} for $z=0$.
As discussed by \citet{Bernardi17}, this particular set of data gives
the lowest values of the comoving number density of galaxies for
stellar masses $\gtrsim 10^{11}\,{\rm M}_\odot$ in comparison with
other results presented in the literature because of
the use of inappropriate algorithms for estimating the observed flux
of the most luminous galaxies and
inadequate assumptions about the stellar population modelling.
While systematic effects on the SMF
related with photometry account for differences
of only 0.1 dex, systematics arising from different treatments of the
stellar population can be as large as $0.5$ dex.
The estimations compiled and analysed by \citet{Bernardi17},
which include data from \citet{Chen12} and \citet{Mendel14},
are represented by a grey shaded area.
When comparing with these sets of data,
the excess in the comoving number density of massive galaxies
predicted by \textsc{sag}~becomes less pronounced.
\subsubsection[]{Star formation rate function}
\begin{figure}
\centering
\includegraphics{Fig2.pdf}
\caption{
Star formation rate distribution function at $z=0.14$
used as constraint to calibrate the \textsc{sag}~model.
Results obtained from the calibrated model \textsc{sag}~with the best-fitting values
given in Table~\ref{table1}, and from the \textsc{sag$_{\beta1.3}$}~model are shown by red solid
and blue dashed lines, respectively.
The latter produces a decrease in the number density of high SFR values.
Model results are compared with observations by \citet{gruppioni15}
represented by filled circles with error bars.
}
\label{fig:SAGconstraints_SFRF}
\end{figure}
In addition to the SMF at different redshifts,
the SFRF at $z=0.14$
helps to recover the observed
evolution of SFR and mass growth in
galaxies.
Fig.~\ref{fig:SAGconstraints_SFRF}
presents
the SFRF at $z = 0.14$ for the galaxy population
generated by the calibrated model \textsc{sag}~(solid line).
The agreement with the
observations by \citet{gruppioni15}
is very good for intermediate
values of the SFR, in the range
$1\,{\rm M}_{\odot}\,\text{yr}^{-1} \lesssim {\rm SFR} \lesssim 20\,{\rm M}_{\odot}\,\text{yr}^{-1}$,
being slightly underpredicted (still within the dispersion of observed data)
for lower values. \textsc{sag}~gives higher number density
of galaxies for $\text{SFR} \approx 60\,{\rm M}_{\odot}\,\text{yr}^{-1}$,
reaching again a good agreement for the highest SFR bin
($\approx 200\,{\rm M}_{\odot}\,\text{yr}^{-1}$).
This particular behaviour of our model is quite similar to the predictions
at low redshifts provided by other SAMs considered
by \citet{gruppioni15} in a direct comparison with their data.
The agreement shown here for SFR above $100\,{\rm M}_{\odot}\,\text{yr}^{-1}$
is better
than the one reached by the `FIRE feedback model' analysed
by \citet{hirschmann16}
from which the current feedback scheme implemented in \textsc{sag}~is inspired.
This better agreement could have been achieved simply because the
SFRF was used to calibrate the model.
However, it is not always possible
to recover the complete behaviour of an observational constraint,
as it is the case of the massive end of the SMF at $z=0$.
This tension that emerges during the calibration procedure denotes
the need of improvement in the physical processes modelled and/or
inconsistencies between different sets of observed galaxy properties.
On the other hand,
model \textsc{sag$_{\beta1.3}$}~produces a decrease in the number density of high SFR values,
as shown by the dashed line.
\subsubsection{Cold gas mass fraction}
\begin{figure}
\centering
\includegraphics{Fig3.pdf}
\caption{
Fraction of mass in cold gas as a function of stellar mass at $z=0$
used as constraints to calibrate the \textsc{sag}~model.
Results obtained from the calibrated model \textsc{sag}~with the best-fitting
values given in Table~\ref{table1} are represented by red solid lines
and reddish filled contours.
Red thin solid lines denote the standard deviation around the mean
(thick red solid line).
In both cases,
the contour levels are:
$[0.01, 0.19, 0.26, 0.38, 0.68, 0.95, 0.997]$ in terms of the maximum
number density of points of each sample.
CGMF is almost unaffected by adopting $\beta=1.3$,
as shown by results from the \textsc{sag$_{\beta1.3}$}~model represented by a blue dashed line
and empty dotted contours.
This relation is compared
with observations by \citet[][filled circles with error bars]{boselli14}.
}
\label{fig:SAGconstraints_CG}
\end{figure}
The fact that galaxy growth and star formation quenching
are well captured by \textsc{sag}~becomes evident in the cold gas content
of model galaxies (Fig.~\ref{fig:SAGconstraints_CG}).
The CGMF obtained from \textsc{sag}~also behaves as expected within
the stellar mass range considered for the calibration of the model
(${\rm log}(M_{\star}[{\rm M}_{\odot}])\in [9.15,10.52]$).
Their mean values (thick solid line)
are in good agreement with observational data from
\citet[][filled circles with error bars]{boselli14};
thin solid lines represent the standard deviation around the mean and
the shaded contours show the distribution of model values.
For high stellar masses (${\rm log}(M_{\star}[{\rm M}_{\odot}]) > 10.52$),
\textsc{sag}~predicts lower fractions of cold gas with respect to those
expected from a simple extrapolation of the
trend found at lower masses.
However, these values are still contained within the range
of cold gas fractions inferred
from the atomic and molecular gas fractions presented by
\citet[][see their fig.~3]{Saintonge16}, obtained from data of
the Arecibo Legacy Fast ALFA (ALFALFA),
the GALEX Arecibo SDSS Survey (GASS) and the CO Legacy Database for GASS
(COLD GASS) surveys.
The same trend is obtained with model \textsc{sag$_{\beta1.3}$}.
\subsubsection[]{Black hole-bulge mass relation}
\begin{figure}
\centering
\includegraphics{Fig4.pdf}
\caption{
Relation between black hole and bulge mass at $z=0$
used as constraints to calibrate the \textsc{sag}~model.
Results obtained from the calibrated model \textsc{sag}~with the best-fitting
values given in Table~\ref{table1} are represented by reddish filled contours, while
results obtained from the \textsc{sag$_{\beta1.3}$}~model are shown by empty dotted contours.
In both cases, the
contour levels are the same as those of Fig.~\ref{fig:SAGconstraints_CG}.
The BHB relation is almost no affected when
adopting $\beta=1.3$.
Observational data are taken from
\citet[][black circles]{kormendy_bhb_2013}
and
\citet[][black triangles]{mcconnell_bhb_2013}.
}
\label{fig:SAGconstraints_BHB}
\end{figure}
The BHB relation displayed by the calibrated \textsc{sag}~model
(Fig.~\ref{fig:SAGconstraints_BHB}, filled contours) is in good general agreement
with the observational trend denoted by data from
\citet{mcconnell_bhb_2013} and \citet{kormendy_bhb_2013}.
However, black holes
seem to have higher masses
than expected for large bulge masses,
but still within
the allowed ranges defined by
the dispersions of the observational data.
This trend is preserved in
model \textsc{sag$_{\beta1.3}$}~(empty dotted contours).
This is a result of the restriction imposed by both the high mass end of the
SMF at $z=0$ and the BHB relation.
During the calibration process,
the parameter that controls the growth of super massive BHs, $f_{\rm BH}$,
takes a high value in order to make the AGN feedback as effective as possible.
Due to the existing degeneration between the $f_{\rm BH}$ and $\kappa_{\rm AGN}$
parameters, a similar behaviour of the high-mass end of the SMF at $z=0$ can be achieved by
taking lower values of $f_{\rm BH}$ at the expense of increasing the feedback
efficiency, $\kappa_{\rm AGN}$.
Nevertheless, large values of $f_{\rm BH}$ are preferred
as a direct consequence of the statistical test used in the calibration process.
The chi-square calculated for each constraint
\citep[][see their eq. 27]{ruiz2015} gives heavier
weights to the mass bulge ranges where the bulk of the galaxies are located,
which decreases the statistical significance of the high-mass galaxies in
the selection of the best-fitting parameters.
\subsubsection[]{Origin of the excess in the high-mass end of the SMF at $z=0$}
\label{sec:originSMF}
The AGN feedback together with
additional modifications
introduced in the model, that is,
the redshift dependence of the reheated and ejected material
and the scalings adopted for the reincorporation of the latter
(Equations~\ref{eq:feedfire},~\ref{eq:ejecfire} and~\ref{eq:reinc}),
allow to recover downsizing
in the star formation rates, as shown later in Fig.~\ref{fig:SFRDvsz}.
Hence, there is not an excess of SF in high-mass galaxies
that could be responsible for the excess of the high-mass end of the
SMF at $z=0$.
A test made with the semi-analytic model
\textsc{sage}~discussed in \citet{knebe17}
shows that
stars added to an intra-cluster component as a result of tidal disruption
of satellite galaxies would enhance the massive end of the SMF if
they would have merged with the central galaxy.
In \textsc{sag}, satellite galaxies are not tidally disrupted when becoming orphans as in \textsc{sage}~but suffer TS.
The stripped mass is added to the stellar halo of the corresponding central galaxy (see Section~\ref{sec:TSstars}) and
represents the intra-cluster stars. The stellar mass density of this component is
$\sim 2$ and $\sim3$ orders of magnitude lower than
the one characteristic of the stellar mass of the whole galaxy population, for redshifts $z=0$ and $2$, respectively.
This means that the excess of the high-mass end of the SMF at $z=0$ could be attributed to inefficient stripping or disruption of satellites by effects of tides in the model.
In order to evaluate the efficiency of TS in our model,
we estimate the fraction of mass in stars contained in the stellar
halo of central spiral galaxies and compare them with results
from \citet{Merritt16}.
They estimate the stellar mass in excess of a
disc+bulge beyond 5 half-mass radius of eight spiral galaxies in the Dragonfly Nearby Galaxies Survey (DNGS)
and obtain an average halo fraction of $0.009 \pm 0.005$.
In our model, spiral galaxies are those characterized by a ratio between stellar bulge and total stellar mass
${\rm B/T} < 0.85$; this cut allows to reproduce the observed morphological distribution \citep{conselice2006}.
Median values of the halo fractions (estimated with respect to the total stellar mass of central galaxies, including the stellar halo) increase with stellar mass and are comprised within the range
$\approx 3\times 10^{-5} - 7\times 10^{-4}$
for stellar masses $M_{\star} \approx 10^{10} - 10^{11} \,{\rm M}_{\odot}$.
These median values are at least
one order of magnitude smaller than the average halo fraction obtained
by \citet{Merritt16}.
Although
the difference between the methods used to obtain the mass of the stellar halo in the observations and in the model might contribute to such a discrepancy, it seems quite plausible that TS is not efficient enough in our model, as hinted by the excess in the high-mass end
of the SMF at $z=0$.
In other words, the mass of merging satellites might be higher than expected because they do not suffer enough TS.
\begin{figure}
\centering
\includegraphics{Fig5.pdf}
\caption{
Conditional stellar mass function for central (solid line) and satellite (dashed line) galaxies in model \textsc{sag}. Galaxies are selected according to the mass of the main host halo they reside in as indicated in the legend of the different panels. Model results are compared with data from the group catalogues constructed by
\citep{Yang09b} from SDSS DR4; central galaxies are identified by filled circles and satellites are represented by open squares.
}
\label{fig:CSMFz0}
\end{figure}
Another possible cause of the excess in the high-mass end of the SMF at $z=0$ might be
related to
dry mergers at $z\lesssim 2$ of low-mass galaxies with massive ones.
In this context, it is worth mentioning a caveat regarding the
merger condition adopted for orphan galaxies. As we briefly explain
in Section~\ref{sec:orbits}, it is based on a criterion that demand
an assumption about the radius of the central galaxy (assumed to be
$10$ per cent
of the virial radius of the host halo).
The drawback of this approximate estimation
is that it
can be very large for cluster-size haloes. Therefore, the
integrated orbit can easily take the orphan satellite
within the sphere determined by those radii, thus merging with its central.
In order to test this possibility, we estimate the conditional stellar mass functions for central and satellite galaxies within different ranges of halo mass in \textsc{sag}, which are shown in Fig.~\ref{fig:CSMFz0}. Following \citet{Kang14}, they are compared with data from SDSS DR4 \citep{Yang09b}.
There are less model satellites than observed
in all the halo mass ranges analysed.
This deficiency is more pronounced in more massive haloes
($\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [13.5, 13.8], [14.1, 14.4]$), and
it is accompanied by a shift to higher stellar masses of the population of central galaxies.
Hence, both an inefficient TS and over-merging of satellite galaxies are responsible for the excess
of high-mass galaxies found in the SMF at $z=0$.
These drawbacks in the model are mainly related to aspects of the integration of orphan galaxies, like the treatment of TS and the merging criterion, which will be revisited (Vega-Mart\'inez et al. in preparation).
\section{Model predictions}
\label{sec:SAG-predictions}
We present galaxy properties predicted by the model,
that are not used as constraints to calibrate it.
We first focus on the evolution of the
cosmic star formation rate density (SFRD) and
of the specific star formation rate (sSFR).
The former is given by
the volume-averaged sum of SF of all galaxies at any
given time, and the
latter is
defined as the ratio between the SFR
and the stellar mass of the galaxy.
We connect the trends of the sSFR of satellites and centrals
with the stellar mass content of galaxies populating
DM haloes of different mass through
the stellar-to-halo mass relation (SHM).
According to the value of the sSFR, we classify galaxies as active and passive.
We analyse the dependences of the fraction of passive galaxies
with stellar mass, halo mass and halo-centric distance.
\subsection[]{Star formation rate density}
\label{sec:SAG-pred-sfr}
\begin{figure}
\centering
\includegraphics{Fig6.pdf}
\caption{
Evolution of the cosmic star formation rate density for all galaxies
(red thick solid line)
compared with observational data compiled by \citet{behroozi13c}.
Contribution from galaxies lying in different mass ranges at $z=0$ are
represented by thin black lines of different style,
as indicated in the legend.
Downsizing in stellar mass assembly becomes evident.
Intermediate-mass galaxies
$\log(M_{\star} [{\rm M}_{\odot}]) \in [10.1,10.5]$
are the main responsible
of the excess in the SFRD at $z=0$.
The global SFRD obtained from the calibrated model \textsc{sag}~but
adopting $\beta=1.3$ (model \textsc{sag$_{\beta1.3}$}) is
added for comparison, being represented by a blue thick dashed line.
Model \textsc{sag$_{\beta1.3}$}~is characterized by higher (lower) levels of SFR at high (low)
redshifts with respect to the calibrated model \textsc{sag}, in better agreement
with observations.
}
\label{fig:SFRDvsz}
\end{figure}
The evolution of the cosmic SFRD predicted by \textsc{sag}~is presented in
Fig.~\ref{fig:SFRDvsz} (thick solid line) and is
compared with data compiled by
\citet{behroozi13c}.
Although model results show consistency with observations for $z<4$,
they under-predict the SFRD at higher redshifts. Besides,
the decline in the SFRD towards low redshifts is less pronounced in the model than in observations, thus resulting in an excess of SFRD at $z=0$. The peak and normalization of the SFRD vary significantly among different galaxy formation models
\citep[e.g. ][]{guo16}.
The low SFRD predicted at high redshifts is
a consequence of the redshift dependence of the reheated and ejected gas fractions in
the new feedback scheme implemented in \textsc{sag}~(Equations~\ref{eq:feedfire} and~\ref{eq:ejecfire}),
an effect that
becomes even stronger in our model because of the preferred high value
of $\beta$ in the calibration process, as discussed
in Section~\ref{sec:SAG-const}. The lower value
suggested by the fit found by \citet{muratov15}, e.g. $\beta=1.3$,
allows to reconcile model predictions with observed SFRD at high redshifts,
as shown in Fig.~\ref{fig:SFRDvsz}
by model \textsc{sag$_{\beta1.3}$}~(thick dashed line).
This better match takes place at
the expense of increasing
the number
density of low-mass galaxies
at $z=2$ (dashed line in bottom panel of Fig.~\ref{fig:SAGconstraints_SMF}).
Therefore, it is not possible to satisfy both observational constraints
simultaneously.
This tension between the observed
redshift
evolution of the SFRD and
$z=2$ galactic SMF has been also
noted by \citet{hirschmann16}
from their analysis of the evolution of the correlation of sSFR with
the stellar mass (main sequence of star-forming galaxies, see their fig.~9);
they find that ejective models that
successfully reproduce the measured
evolution of SMF
tend to under-estimate the sSFR at
high redshifts.
In the same line, the recent work of \citet{Rodrigues17}, based on the exploration of the
parameter space of a version of the \textsc{GALFORM} semi-analytic model,
shows that the evolution of the SMF
is recovered by a particular set of parameters
that favours larger SN feedback efficiency at higher redshifts, although
the rise of the SFRD with redshift in the range $0.5 <z < 1$ is not as steep as observed.
\begin{figure*}
\centering
\includegraphics{Fig7.pdf}
\caption{Specific star formation rate as a function of stellar mass
for all galaxies in the calibrated model \textsc{sag}~(filled contours, considering the same
levels as the Fig.~\ref{fig:SAGconstraints_CG})
at different redshifts ($z=0$, $2.2$ and $3.5$).
Mean values of star-forming galaxies
(${\rm sSFR}>10^{-10.7}\,{\rm yr}^{-1}$)
are estimated for central and satellite galaxies
(red solid thick and thin lines,
respectively); error bars represent $1\sigma$-standard deviation around the
mean.
They are compared with redshift dependent fits to
observational data compiled by
\citet{behroozi13c}
(filled circles with error bars).
The gradual removal of hot gas in satellites through RPS and TS allows
satellite galaxies
to have a behaviour similar as central ones.
Results of model \textsc{sag$_{\beta1.3}$}~for both central and satellite galaxies
are represented by blue dashed thick and thin lines, respectively.
}
\label{fig:sSFRmstarz}
\end{figure*}
The difficulty
in satisfying simultaneously the observed evolution of the SFRD and the SMF at high redshift
also emerges in the analysis of observational results.
It is well known that
the star formation history inferred from the observed evolution of the stellar mass differs from instantaneous indicators of star formation, being the former $\approx 0.6$ dex smaller than the latter at $z=3$ \citep{Wilkins2008}.
Thus, the integrated star formation history implies a local
stellar mass density in excess of that measured.
According to recent studies \citep{Madau2014}, this discrepancy in the
stellar mass density can be $\approx 0.2$ dex for $z\lesssim 3$ depending on the data considered,
being smaller than previously estimated.
In any case,
these
discrepancies can arise for
several reasons such as inaccurate dust extinction corrections, underestimation
of stellar masses due to the outshining of old stellar populations in star
forming galaxies, or the evolution of the integrated stellar IMF in galaxies.
Both the star formation history implied by instantaneous indicators, which are typically dominated by very massive stars, and the star formation history inferred from the evolution of
the average stellar mass density are affected by the IMF assumed. The discrepancies between them might be mitigated when considering an evolving IMF that is
top heavy at high redshifts \citep{Wilkins2008}.
The stronger suppression of SF at
high redshifts is evident in galaxies of all masses
being more pronounced in low-mass galaxies, thus
allowing to reach the antihierarchical assembly of stellar mass.
This becomes evident from the evolution of the
galactic SMF presented by
\citet[][see their fig.~1]{hirschmann16};
for ejective feedback models,
high-mass galaxies
($M_{\star} \gtrsim 10^{10}\,{\rm M_{\odot}}$) are already
in place
at $z\approx 2-3$, while the population of low-mass galaxies keeps
increasing towards lower redshifts, in agreement with the trend denoted by
observational measurements.
Downsizing in galaxy assembly is also produced by \textsc{sag}, as
can be appreciated from the contribution
to the global SFRD of galaxies in different mass ranges selected at $z=0$,
as indicated by thin black lines of different style in Fig.~\ref{fig:SFRDvsz}.
The peak of the SFRD of galaxies with stellar mass
$M_{\star}> 10^{11}\,{\rm M}_{\odot}$
is located at $z \approx 2$, while intermediate-mass galaxies
$(\log(M_\star [{\rm M}_\odot]) \in [10.5,10.9])$ reach the
peak at $z \approx 1$. Lower-mass galaxies are still in the regime
of increasing SFRD while approaching to $z=0$.
Downsizing is also
present in model \textsc{sag$_{\beta1.3}$}~but with a change in the SF history of galaxies
which start to form stars earlier.
As expected, this general
trend is consistent with the predictions of ejective feedback models in
\citet[][see bottom row of their fig.~4]{hirschmann16}.
Galaxies with stellar masses $M_{\star}\lesssim 3\times 10^{10}\,{\rm M}_{\odot}$
are responsible for the excess of the local SFRD with respect to observations,
especially those in the mass range
$\log(M_\star [{\rm M}_\odot]) \in [10.1,10.5]$ that dominate
the contribution to the SFRD at $z \lesssim 0.3$.
\citet{hirschmann16} note this problem for the most massive galaxies
($M_{\star} \sim 10^{12}\,{\rm M}_{\odot}$) and attribute this failure
in reproducing the SFRD at $z \lesssim 0.5$ to a
radio-mode AGN feedback not efficient enough in suppressing SF in these
galaxies. Although this could be a possible
explanation for galaxies within the stellar mass range
$\log(M_\star [{\rm M}_\odot]) \in [10.1,10.5]$ in our model,
the restrictions imposed to calibrate \textsc{sag}~with the PSO method
could also favour this excess.
Both the high-mass end of the SMF at $z=0$ and the local BHB relation
help finding appropriate values of the free parameters involved
in the AGN feedback scheme.
However, the SFRF at $z=0.14$ also plays a role.
We have shown that the tuned parameters allow to achieve a
highly satisfactory agreement with the data presented by \citet{gruppioni15} (see Fig.~\ref{fig:SAGconstraints_SFRF}).
However,
the IR+UV SFRD estimated from these data
through the integration of
the best-fitting modified Schechter function
to the IR+UV SFRFs down to $\log{\rm (SFR)}=-1.5$
is higher than the optical SFRD presented by \citet{behroozi13c}
at low redshift ($z<0.5$), which is reflected in
the excess shown by the model in Fig.~\ref{fig:SFRDvsz}.
We evaluate the impact of the constraint used in the calibration process
by imposing the
evolution of the SFRD upto $z=2$ as an observational restriction
in place of
the SFRF at $z=0.14$.
We consider the values for the SFR at several redshifts within
that range taken from the
fit to observational data presented in \citet{behroozi13c}.
We find that the excess of the SFRD at $z=0$ is not avoided even with
this restriction.
This aspect deserves more investigation, in order to disentangle
what particular combination of physical processes is causing this effect,
or what process
is still not well captured by the modelling. The analysis done in
Section~\ref{sec:originSMF}
points to TS,
tidal disruption and over-merging as processes
that deserve special attention.
\begin{figure}
\centering
\includegraphics{Fig8.pdf}
\caption{
Stellar mass as a function of DM halo mass for galaxies generated by both the
calibrated model \textsc{sag}~and model \textsc{sag$_{\beta1.3}$}~at $z=0$
compared with the parametrization given by \citet{Moster10}.
{\it Top panel}: Relation for central galaxies of main host haloes
represented by
a coloured contour map,
considering the same contour levels as those of Fig.~\ref{fig:SAGconstraints_BHB}.
Mean values of the stellar mass for each bin of halo mass
for models \textsc{sag}~and \textsc{sag$_{\beta1.3}$}~are depicted by red squares and blue triangles, respectively;
vertical error bars are $1\sigma$-standard deviation around the mean.
{\it Bottom panel}: Same as top panel but for
satellite galaxies within DM subhaloes.
Differences between this relation for central and satellite galaxies translate
into the different normalization and shape of their respective main sequences
(see Fig.~\ref{fig:sSFRmstarz}).
}
\label{fig:SHMratio}
\end{figure}
\subsection[]{Specific star formation rate: main sequence}
\label{sec:SAG-pred-ssfr-mainseq}
Fig.~\ref{fig:sSFRmstarz} shows the relation
between the sSFR and stellar mass of
galaxies generated by
the calibrated model \textsc{sag}~at different
redshifts ($z=0$, $0.7$, $2.2$ and $3.5$)
by the density map.
At $z=0$ (top-left panel),
this distribution resembles very much the one presented
by \citet{Salim07}, both in shape and normalization.
They estimate the SFR of
50,000 optically selected galaxies in the local Universe
from gas-rich dwarfs to massive ellipticals
using ultraviolet and optical data from the
{\it Galaxy Evolution Explorer} (GALEX) and
SDSS Data Release 4, respectively.
Comparing with their fig. 15, it is evident that the dependence of
sSFR with stellar mass for model galaxies is shallower than observed.
This can be better quantified by considering the mean values of
sSFR for different stellar mass bins for star-forming galaxies.
The tight correlation followed by them is known as the main sequence of galaxies.
Following \citet{Brown17}, who investigate environment driven gas depletion
in satellite galaxies using a sample from SDSS, we classify galaxies as
star-forming when they have
${\rm sSFR}>10^{-10.7}\,{\rm yr^{-1}} $.
This limit allows a better separation between
active and passive galaxies in our model than
the cut
${\rm sSFR}=10^{-11}\,{\rm yr^{-1}}$,
commonly used in the
literature \citep[e.g.][W12 hereafter]{Wetzel12}.
Although this is
inferred from the bimodality
that emerges for massive galaxies within groups and clusters
(see Section~\ref{sec:SAG-pred-ssfr-distr}),
we apply this cut to galaxies in all mass ranges, following W12. This criterion also comprises the star-forming
sample of galaxies selected by \citet[][see their fig. 17]{Salim07}
through optical emission lines
and the BPT diagram.
Mean values of sSFR of star-forming galaxies are shown for
central and satellite
galaxies separately.
These main sequences are compared with
redshift dependent fits
to observational data
collected by
\citet{behroozi13c},
all corrected to a Chabrier IMF
(see their table $8$ in appendix F). The data set,
specified in their table 5, includes the
sSRF measured by \citet{Salim07} described above.
The general agreement is rather good at all redshifts, especially
regarding the normalization.
Reproducing the right normalization of this correlation has been
challenging;
galaxy formation models have failed in reproducing this feature
underpredicting it, specially at high redshifts \citep{Daddi07, Weinmann12, Xie17}.
The similarity between
the main sequence of central and satellite galaxies
for any of the redshifts considered highlights the importance
of modelling environmental effects through RPS and TS, with the consequent
gradual removal of hot halo gas.
The strangulation scheme considered in previous versions of \textsc{sag}~quenches
star formation in satellite
galaxies too early, leading to values of SFR one order of magnitude lower than
those characterising central galaxies, thus generating a
complete separate main sequence for satellites.
Satellite galaxies in the current model lie on a main sequence
that is systematically lower than the one traced by centrals at all redshifts,
but differences are within the $1\sigma$-standard deviation around the mean.
Although the normalization is quite good for both central and
satellite galaxies,
the model predicts flatter sequences than those inferred from observations.
Mean values of sSFR of central galaxies
decrease for stellar masses $M_{\star}\gtrsim 10^{10}\,{\rm M}_{\odot}$,
being this trend more pronounced for centrals than for satellite galaxies;
the main sequence for satellites remains almost flat at $z=3.5$.
This leads to
more similar values of mean sSFR for these two galaxy populations
at high stellar masses.
For low-mass galaxies ($M_{\star} \lesssim 10^{10}\,{\rm M}_{\odot}$),
the agreement between \textsc{sag}~predictions and observational fits
is better at $z=0$ than at higher redshifts,
while the opposite situation occurs for high-mass galaxies. The
higher values of sSFR at $z=0$ for galaxies with stellar masses
within the range
$\log(M_\star [{\rm M}_{\odot}]) \in [10.1,10.5]$
are consistent
with the fact that these galaxies are the main responsible of
the excess of the cosmic SFRD at $z=0$
(see Fig.~\ref{fig:SFRDvsz}).
Both the normalization and shape of the main sequence of central
and satellite galaxies are affected when considering
model \textsc{sag$_{\beta1.3}$}, as it is shown in Fig.~\ref{fig:sSFRmstarz}
by dashed thick and thin lines, respectively.
Central and satellite galaxies from model \textsc{sag$_{\beta1.3}$}~achieve smaller sSFR
values than those characterizing the same galaxy types generated by model
\textsc{sag}~in all the redshifts considered,
being these differences larger for smaller stellar masses.
They originate in the large amount of stellar mass acquired by low
mass galaxies at high redshift in model \textsc{sag$_{\beta1.3}$}~(see
bottom panel of Fig.~\ref{fig:SAGconstraints_SMF})
because the SFRD in this model is higher than in \textsc{sag}~for $z\gtrsim 1$ (see Fig.~\ref{fig:SFRDvsz}). Besides, the SFRD in \textsc{sag$_{\beta1.3}$}~is slightly lower at $z=0$,
a pattern that is common to galaxies
within different $z=0$ stellar mass ranges,
contributing to reduce the sSFR.
Moreover, the decrease of sSFR with decreasing stellar mass
for $M_{\star} \lesssim 10^{10}\,{\rm M}_{\odot}$, which
is a trend opposite to the one typical of observational data, becomes more evident in model \textsc{sag$_{\beta1.3}$}.
This particular behaviour is similar to the one obtained
by \citet[][see their fig.~11]{Xie17}
who use an updated version of the
semi-analytic model GAEA \citep{hirschmann16}
considering the same ejective
model of SN feedback
in which
the new feedback scheme implemented
in our model is inspired; the similarity with results from model \textsc{sag$_{\beta1.3}$}~is
not surprising since they adopt a parameter $\beta=1.25$.
\cite{Xie17} argue that central galaxies with underestimated
SFRs are responsible for the decreasing trend of the sSFR
for low stellar masses.
In fact, from Fig.~\ref{fig:sSFRmstarz}, we can clearly see that
both central and satellite galaxies contribute to the
decrease of sSFR with decreasing stellar mass.
The responsible of this trend is the
additional modulation with the virial velocity introduced in the estimation of reheated and ejected mass (eqs.~\ref{eq:feedfire} and \ref{eq:ejecfire}), following the broken power law suggested by \citet{muratov15}.
The exponent $\alpha_\text{F}=-3.2$ adopted for low-mass galaxies ($V_{\rm vir} < 60\,{\rm km}\,{\rm s}^{-1}$) makes the SN feedback too strong for them, which then
have too low SFR. We have verified this by running
a variant of model \textsc{sag$_{\beta1.3}$} in which the slope $\alpha_{\rm F}$ is fixed to -1 (the value corresponding to $V_{\rm vir} > 60\,{\rm km}\,{\rm s}^{-1}$), for any virial velocity.
In this case, the main sequence of central galaxies become flatter at low stellar masses, while satellites present a mild trend in which sSFR increases monotonically with decreasing stellar mass.
The differences in normalization
and shape of the main sequence of central and satellite galaxies
are not explained by the particular treatment of
environmental effects introduced in \textsc{sag}. Avoiding the action of RPS on the hot
and cold gas phases only produces a higher number of active satellite galaxies,
without changing the mean values of the sSFR of star-forming ones.
In order to understand the origin of these differences,
we examine the relationship between stellar mass and DM halo mass.
We consider the mass of the main host haloes for central galaxies
and the subhaloes masses for satellite galaxies,
excluding orphans.
Fig.~\ref{fig:SHMratio} shows such relations,
representing galaxies for the calibrated model \textsc{sag}~with a coloured contour map
on top of which
the mean values of the stellar mass for each bin of halo mass
(depicted by squares) are superimposed.
Mean values of model \textsc{sag$_{\beta1.3}$}~are also added (represented by triangles),
showing that the SHM relation is not appreciably affected by
a change in the parameter $\beta$.
Mean values are compared with
the fitting function for the stellar-to-halo mass relation at $z=0$
derived by \citet{Moster10}; such parametrization was estimated
by requiring that the observed SMF obtained
from SDSS DR3 \citep{Panter04} is reproduced when
populating with galaxies the haloes and subhaloes in an {\em N}-body simulation.
Note that this comparison is not fair for satellite galaxies,
but it is done with the purpose of highlighting the differences
in the relation for centrals and satellites.
The general trend is very well reproduced by model central galaxies.
For halo mass $M_{\rm vir}\lesssim 10^{12}\,{\rm M}_{\odot}$, differences
between centrals and satellites become more evident.
For a given stellar mass, central galaxies inhabit more massive haloes, while
satellites prefer less massive ones, being the dispersion of the
relation (vertical error bars) larger for smaller stellar masses.
Systematic differences in the halo virial mass between main host
and satellite haloes translate into differences
in the hot halo mass
and associated
cooling rates, which impact on the SF activity leaving
different imprints in the main sequence of central and satellite galaxies.
\begin{figure*}
\includegraphics[width=2.0\columnwidth]{Fig9.pdf}
\caption{Specific star formation rate distributions for galaxies
with stellar masses in different mass bins at $z=0$:
$\log(M_{\star} [{\rm M}_{\odot}])\in[9.7, 10.1]$ (left panel),
$\log(M_{\star} [{\rm M}_{\odot}])\in[10.1, 10.5]$ (middle panel) and
$\log(M_{\star} [{\rm M}_{\odot}])\in[10.5, 11.0]$ (right panel).
Coloured thin lines of different style represent satellite galaxies
residing in main host DM haloes within different mass ranges,
as indicated in the legend.
Central galaxies are identified with black thick solid lines; all central
galaxies are considered, without any restriction regarding
their main host DM haloes.
A bimodal distribution
emerges only for the most massive galaxies considered, with a break at
${\rm sSFR}=10^{-10.7}\,{\rm yr}^{-1}$, a value
in good agreement with the one adopted by \citet{Brown17} and
slightly higher than the one identified from the observational
catalogue constructed by W12.
}
\label{fig:sSFRhisto-mstar-mhalo}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.99\columnwidth]{Fig10a.pdf}
\includegraphics[width=0.99\columnwidth]{Fig10b.pdf}
\caption{Fraction of quenched galaxies
(${\rm sSFR} < 10^{-10.7}\,{\rm yr}^{-1}$) in \textsc{sag}~model at $z=0$ as
a function of stellar mass (top panel) and main host halo mass (bottom panel).
Satellite galaxies are binned according to their main host
halo or stellar mass, and are represented by coloured lines of different style,
as indicated in the legend. Central galaxies,
represented by a thin black dashed line, are only included in the top panel
without making any distinction to the mass of the halo they reside in.
Error bars show the $68$ per cent bayesian confidence interval estimated
following
\citet{Cameron11}; they are hardly visible for the satellite population.
These fractions are compared to
those obtained by W12 (different symbols associated to different
line styles).
The calibrated \textsc{sag}~model underpredicts the fraction of quenched galaxies.
Only the most massive galaxies, those in the stellar mass range
$10.9 < {\text {\rm log} (M_{\star} [{\rm M}_{\odot}])} < 11.3$
within DM host haloes with masses $M_{\rm halo}\gtrsim 10^{14}\,{\rm M}_{\odot}$
achieve a quenched fraction similar to the observed one.
This lack of general agreement is related to
the underpredicted SFRD at high redshifts,
as discussed in Section~\ref{sec:SAG-pred-sfr}.
The value $\beta=1.99$
that regulates the redshift dependence of the reheated and
ejected mass in \textsc{sag}~model is too high and produces
a shift of the SF activity to lower redshifts not leaving enough
time for quenching.
}
\label{fig:fq_mstar_mhalo-erebos}
\end{figure}
\begin{figure}
\centering
\includegraphics{Fig11a.pdf}
\includegraphics{Fig11b.pdf}
\caption{Same as Fig.~\ref{fig:fq_mstar_mhalo-erebos}
for galaxies generated with
model \textsc{sag$_{\beta1.3}$}.
The general agreement with observations is rather good except for the
underprediction of the quenched fraction for galaxies within
the mass range $\log (M_{\star} [{\rm M}_{\odot}]) \in [10.1, 10.5]$.
}
\label{fig:fq_mstar_mhalo}
\end{figure}
\subsection[]{Specific star formation rate: distribution}
\label{sec:SAG-pred-ssfr-distr}
We analyse the distribution of sSFR
for central and satellite galaxies within different stellar mass ranges,
discriminating the latter
according to the mass of their main host
DM haloes\footnote{From hereafter,
all the sampling of satellite galaxies is done according to the masses of
their main host haloes, even if it is simply referred as \textsl{halo}.}.
We select galaxies within the same stellar and halo mass ranges as
W12, who construct galaxy group catalogues
using a group-finding algorithm based on the one presented
by \citet{Yang05}.
W12 derive stellar masses and sSFR
for galaxies
within groups and clusters from the spectroscopic information
provided by the SDSS Data Release 7; the median redshift of the sample is
$z = 0.045$.
Left, middle and right panels of
Fig.~\ref{fig:sSFRhisto-mstar-mhalo} show, respectively, the sSFR
distributions\footnote{Probability density function at the bin, normalized
such that the integral over the range is equal to unity.}
of galaxies with stellar masses within the ranges
$\log(M_{\star} [{\rm M}_{\odot}]) \in [9.7,10.1]$,
$\log(M_{\star} [{\rm M}_{\odot}]) \in [10.1,10.5]$ and
$\log(M_{\star} [{\rm M}_{\odot}]) \in [10.5,11.0]$.
Satellite galaxies are discriminated according to the
mass of their main host haloes.
In all cases, both central and satellite galaxies have a well
distinguished peak at high values of sSFR
(${\rm sSFR} > 10^{-10}\,{\rm yr}^{-1}$).
A bimodal distribution
emerges only for the most massive
galaxies considered.
The break of the bimodality takes place at
${\rm sSFR}=10^{-10.7}\,{\rm yr}^{-1}$, a value
in good agreement with the one adopted by \citet{Brown17} and
slightly higher than the one identified from the observational
catalogue constructed by W12.
While the sSFR distribution of galaxies in their observational sample
exhibits a clear bimodality
at all stellar masses, regardless of their classification in centrals or
satellites (see their fig. 1),
the authors clarify that the strong sharpness of the peak near
${\rm sSFR}=10^{-12}\,{\rm yr}^{-1}$ is a result of an artifact in spectral reductions.
They also emphasize that the distributions exhibit
a tail to much lower sSFR, like the sSFR distributions
of our model galaxies do, which are in better agreement with the behaviour
shown by \citet{Salim07}, as it was described in
Section~\ref{sec:SAG-pred-ssfr-mainseq}.
Morever, \citet{Kauffmann14} concludes
that such a distribution is expected for galaxies characterized by a bursty mode of star formation with times of formation between $1$ and $10\,{\rm Gyrs}$. This is consistent with the results we obtain from a detailed
analysis of quenching time-scales of satellites \citep[][submitted]{Cora18b}.
Consistent with the discussion also presented in Section~\ref{sec:SAG-pred-ssfr-mainseq}
when analysing the main sequence of central and satellite galaxies at $z=0$,
the high sSFR peak of the distributions of satellite galaxies
is slightly shifted towards lower sSFR values than for centrals, while
such a trend is not evident in the observational sample.
Despite these differences between our results and those of
W12, there are several similarities in relevant aspects of
the distributions.
As in the observational sample, more massive central galaxies
have larger probability of having low sSFR at the expense of a reduction
in the peak height at high sSFR. However, satellites only show such a trend
for those cases where the bimodality becomes evident in the model.
Both in the simulated and observed sample, there are
more satellites with low sSFR than centrals while the opposite occurs for
high sSFR, regardless of the stellar mass bin considered.
Besides, the number of satellite galaxies with low(high) sSFR
increases(decreases) systematically as we consider more massive
DM haloes.
These general results remain valid for galaxies generated by model \textsc{sag$_{\beta1.3}$}.
A quantitative comparison with observational results is attained through
the estimation of the fraction of quenched galaxies for different stellar
mass and halo mass ranges.
\subsection[]{Quenched fractions of central and satellite galaxies from the model SAG}
\label{sec:SAG-pred-fq}
From a given population of model galaxies
selected according to their stellar mass and/or main host halo mass,
we estimate
the quenched fraction by considering those galaxies
that are passive at a given time of interest, that is, those that satisfy
the criterion
${\rm sSFR} < 10^{-10.7}\,{\rm yr}^{-1}$,
as emerges from the distribution of sSFR analysed in the previous section.
We compare model results
with those obtained by W12.
Although these authors consider a galaxy as passive when its
${\rm sSFR} < 10^{-11}\,{\rm yr}^{-1}$, we
prefer to estimate the quenched fraction
taking into account the separation between
active and passive galaxies that emerge from the model at the expense of
accepting a shift in the sSFR cut with respect to the one
proposed by W12.
Stellar mass and main host halo mass dependence of the quenched
fraction at $z=0$ ($fq_{\rm z0}$) are shown
in the left and right panels
of Fig.~\ref{fig:fq_mstar_mhalo-erebos}, respectively.
In both cases, the same mass bins as those chosen by W12
are used.
Model results are identified with lines.
Error bars show the $68$ per cent bayesian confidence interval estimated
following
\citet{Cameron11}; their method is applied to estimate errors of fractions
in all subsequent plots.
The overwhelmingly numerous galaxy population obtained with the MDPL2 simulation makes a very good statistics in these results, so most of these error bars are too small to be distinguished.
For the stellar mass dependence, we present the quenched fraction
of both central (thin dashed line) and satellite (different
line styles) galaxies;
the latter are grouped according to the halo mass they reside in.
For the halo mass dependence, only satellites within different
mass ranges are considered.
Symbols depict the results of W12 (see their fig. 3(a) and 3(b)).
From this comparison, we can see that the
model underpredicts the fraction of quenched galaxies.
Only the most massive galaxies, those in the stellar mass range
$\log(M_\star [\textrm{M}_\odot]) \in [10.9, 11.3]$,
residing in haloes with masses
$M_\textrm{halo}\gtrsim 10^{14}\,\textrm{M}_\odot$
are characterized by a quenched fraction similar to the observed one.
This drawback of the model is the result of the high
value of the parameter $\beta$ achieved during the calibration
process when trying to satisfy the constraint imposed by the
SMF at $z=2$, giving raise to an underprediction of the SFRD at
high redshifts, as discussed in Section~\ref{sec:SAG-pred-sfr}.
Thus, the star formation activity is shifted to later epochs having less time
to be quenched.
We have also shown that fixing $\beta$ in the lower value suggested
by \citet{muratov15}
allows to obtain a SFRD in better agreement with observations at high
redshift,
at the expense of having an excess in the number density of low-mass galaxies
at $z=2$.
Results on the quenched fractions are in favour of choosing a model
that prioritises the right
evolution of the SFRD.
In the following, we show that quenched fractions are considerably
improved for galaxies generated from model \textsc{sag$_{\beta1.3}$}.
\subsection[]{Quenched fractions of central and satellite galaxies from the model SAG$_{\beta 1.3}$}
\label{sec:fq-ms-mh-r}
Fig.~\ref{fig:fq_mstar_mhalo} shows the stellar and main host halo mass
dependence of the quenched fraction of galaxies generated by model \textsc{sag$_{\beta1.3}$}.
The fractions of quenched galaxies at $z=0$
are larger than those obtained with the calibrated model \textsc{sag}~characterized by the parameter
$\beta=1.99$.
The general agreement with values inferred by W12 is rather good, reproducing
the observed trends
for both central and satellite galaxies. This fact highlights the importance of an adequate efficiency of SN feedback at high redshifts in determining the passive fraction of galaxies.
Despite this general good agreement, some differences in a certain range of
stellar mass for satellites still remain.
Namely, the quenched fractions of satellite galaxies with stellar masses
$\log (M_{\star} [{\rm M}_{\odot}]) \in [10.1, 10.5]$ are clearly
under-predicted for
any halo mass considered.
This mass range is responsible of the excess of the SFRD at $z=0$
(see Fig.~\ref{fig:SFRDvsz}).
However,
satellites do not contribute significantly to the SFRD excess.
This under-prediction is caused by the lack of satellite galaxies
in this stellar mass range as a result of over-merging, as it is
evident from the depletion in the conditional stellar mass function
of satellites at $\log (M_{\star} [h^{-2}\,{\rm M}_{\odot}]) \approx 10$ (see Fig.~\ref{fig:CSMFz0}).
As noted by W12, both central and satellite galaxies
are more likely to be quenched as their stellar masses increase,
being this dependence stronger for the former and mainly
produced by self-regulating processes such as AGN and SN feedback,
i.e. mass quenching
\citep[e.g.][]{Peng10,Henriques17}.
Instead, satellites have a milder dependence of the
quenched fraction on stellar mass, and its increment is more
gradual for satellites hosted by more massive haloes, as a result
of the larger effect of environmental processes on satellites
with lower stellar masses.
Hence, for a given stellar mass, the fraction of passive satellites
increases with increasing halo mass, being the dependence
on halo mass much milder for higher mass galaxies.
This is more clearly shown in
the right panel of Fig.~\ref{fig:fq_mstar_mhalo}, where
we can also appreciate that,
for a given halo mass, more massive satellites
are more likely to be quenched.
\begin{figure}
\centering
\includegraphics{Fig12a.pdf}
\includegraphics{Fig12b.pdf}
\caption{Fraction of quenched satellite galaxies
(${\rm sSFR} < 10^{-10.7}\,{\rm yr}^{-1}$) in model \textsc{sag$_{\beta1.3}$}~at $z=0$ as
a function of the projected halo-centric distance normalized with
the virial radius of the main host halo.
Satellites are binned according to their
main host halo mass (top panel) and stellar mass (bottom panel).
Model results are shown by different lines;
error bars show the $68$ per cent Bayesian confidence interval estimated
following \citet{Cameron11}.
Only satellites within the stellar mass range
$ \log (M_{\star} [{\rm M}_{\odot}]) \in [9.7, 10.5]$
are considered in the top panel and with main host halo mass
$\log (M_{\rm halo} [{\rm M}_{\odot}]) > 12.5$ in the bottom one.
Such selection is done following W12 for comparison purposes.
Their data are represented by different symbols according to
the different stellar or halo mass ranges.
}
\label{fig:fq_r}
\end{figure}
Fig.~\ref{fig:fq_r} shows how the fraction of $z=0$ quenched satellite galaxies
in model \textsc{sag$_{\beta1.3}$}~varies with the projected halo-centric distance,
normalized by the
virial radius of the main host DM halo in which the galaxies reside.
Satellites are grouped according to their main host halo mass (top panel)
or stellar
mass (bottom panel), and their quenched fractions are depicted
by different line styles. In the first case,
only galaxies within the stellar mass
range $\log (M_{\star} [{\rm M}_{\odot}]) \in [9.7, 10.5]$ are considered.
In the second case, all galaxies within haloes with masses
$M_{\rm halo} > 3.16 \times 10^{12}\,{\rm M}_{\odot}$ are taken into account.
These selections
allow to make a direct comparison with the quenched
fractions of the sample analysed
by W12, represented by different
symbols according to the stellar or halo mass considered.
In general terms,
the radial gradient of the observed quenched fraction is well recovered
by the model for both the stellar mass and halo mass selected sets.
When the halo mass is varied (top panel), the radial profiles are a bit steeper
than observed.
The model gives a good match of $fq_{\rm z0}$
for $R_{\rm proj}/r_{\rm vir}\lesssim 0.2$
and $R_{\rm proj}/r_{\rm vir}\gtrsim 0.5$ for galaxies
within haloes of mass
$\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [12.3, 13.2]$
and
$\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [13.2, 14.1]$, respectively.
Quenched fractions for the highest halo mass bin
($\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [14.1, 15.0]$)
are below W12 results for all halo-centric distances, and
similar to those corresponding to the immediately smaller halo mass
range.
This result is consistent with the rather flat behaviour of the
predicted quenched
fractions as a function of halo mass for masses
$M_{\rm halo} \gtrsim 4 \times 10^{13}\,{\rm M}_{\odot}$,
shown in the bottom panel of
Fig.~\ref{fig:fq_mstar_mhalo}.
The discrepancies found
when the halo mass is varied
are reflected in the mismatch between model
predictions and values inferred from observations
for the two lowest stellar mass bins
considered in the bottom panel of Fig.~\ref{fig:fq_r}.
As we can see,
predicted values of $fq_{\rm z0}$ are comprised within the range delimited
by the corresponding observed quenched fractions,
with galaxies within the stellar mass range
$\log (M_{\star} [{\rm M}_{\odot}]) \in [10.1, 10.5]$
only achieving a good agreement near the cluster centre;
this highlights once again the inability of the model
in making adequate predictions
for this particular mass range
because of over-merger,
as already noted.
This might have affected the fraction of quenched low-mass satellites,
since most of surviving orphans are those on less eccentric orbits
that keep far from the central galaxy
avoiding the merging region and preventing orphans from experiencing
medium to strong-level RP.
\begin{figure*}
\centering
\includegraphics{Fig13.pdf}
\caption{Mean values of the
fraction of hot gas mass
stripped by RP at each snapshot of the simulation
as a function of redshift for model \textsc{sag$_{\beta1.3}$}.
All satellites (passive and active) at $z=0$ are included.
They are grouped according to their local
stellar mass (different line styles)
and main host halo mass (different panels), as indicated in the legends.
For any galaxy mass and
and halo mass,
the fraction of hot gas mass stripped by RP
increases with decreasing redshift.
}
\label{fig:stripfrac_hot}
\end{figure*}
\begin{figure}
\centering
\includegraphics{Fig14.pdf}
\caption{Mean values of the
fraction of accumulated stripped hot gas by RP as a function of stellar mass for model \textsc{sag$_{\beta1.3}$}.
All satellites (passive and active) at $z=0$ are included.
They are grouped according to their local
main host halo mass (different line styles), as indicated in the legend.
}
\label{fig:stripfrac_hot_acc}
\end{figure}
\begin{figure}
\centering
\includegraphics{Fig15.pdf}
\caption{Mean values of the fraction of accumulated stripped cold gas by RP as a function of stellar mass for satellites in model \textsc{sag$_{\beta1.3}$}~which interstellar medium
is not longer shielded by the hot gas halo.
They are grouped according to the mass of their local
main host haloes (different line styles), as indicated in the legend.}
\label{fig:stripfrac_cold}
\end{figure}
On the other hand,
the predicted quenched fractions of
satellites within the two more massive stellar mass bins
($\log (M_{\star} [{\rm M}_{\odot}]) > 10.5$)
follow the observed radial profiles very well.
Since SF in more massive galaxies
is mainly suppressed by mass quenching processes at both
low and high redshifts \citep{Lin14, Kawinwanichakij17},
both the radial and halo mass dependence of the fractions of quenched massive galaxies
predicted by model
\textsc{sag$_{\beta1.3}$}~support the modelling
of those mass quenching processes that are particularly relevant
for galaxies with stellar masses
$\log (M_{\star} [{\rm M}_{\odot}]) > 10.5$, such as
AGN feedback and disc instabilities.
While the former reduces the amount of cooled gas,
the latter produces starbursts leading to a rapid exhaustion
of the cold gas reservoir.
These self-regulating physical processes
explain the SF quenching in massive central galaxies
but are still active as quenching mechanisms
even when galaxies become satellites, as demonstrated by \citet{Peng12},
playing a dominant role with respect to environmental effects.
This does not mean than environmental processes are irrelevant for high-mass galaxies. Indeed, these galaxies suffer the effects of RPS but they are smaller than those experienced by low-mass satellites, as shown in the next Section.
\section[]{Role of environment: Stripped mass and atomic gas content}
\label{sec:ts-rps}
In order to understand the role of environment on
satellite galaxies of different mass,
we show in Fig.~\ref{fig:stripfrac_hot}
mean values of the fraction of hot gas mass
stripped by RP, $f_{\rm strip,hot}$, at each snapshot of the simulation
as a function of redshift, for model \textsc{sag$_{\beta1.3}$}. This quantity is estimated considering
the ratio between the stripped mass and the mass of hot gas just before
being stripped;
the latter is given by the sum of the stripped mass and the
mass of hot gas at the redshift considered.
Satellites are grouped according to their current stellar mass
(different lines) and the virial mass of their current main host
DM haloes (different panels).
As expected, RPS is more efficient in removing hot gas from galaxies
with smaller stellar masses
as demonstrated from both the analysis of observational data
\citep[e.g.][]{Fillingham16} and hydrodynamical simulations
\citep{Bahe15}.
For any galaxy mass and
and halo mass,
the fraction of hot gas mass stripped by RP
increases with decreasing redshift.
This is explained
by the mass growth of the host DM halo and the consequent
higher densities in the ICM, and high redshift infalling
galaxies that have more time to approach the cluster core,
thus suffering stronger RP
(\citealt{Jaffe15, Jaffe16}, Vega-Mart\'inez et al., in prep.).
Local values of the removed fraction at each stripping episode
are quite similar for any halo mass,
being of the order of $\approx 0.03-0.05$, with lower values corresponding
to lower mass haloes.
The dependence of this fraction with halo
mass becomes more evident at higher redshifts;
the effect of RP is more pronounced in clusters
of $M_{\rm vir} \approx 10^{15}\,{\rm M}_\odot$ since $z=1.5$ than in less massive ones,
consistent with the distribution of RP values presented by
\citet[][see their fig. 5]{tecce10}.
Fig.~\ref{fig:stripfrac_hot_acc} shows the mean values of the cumulative stripped hot gas fraction
as a function of stellar mass of satellite galaxies binned according to
the mass of the main host halo they inhabit. This fraction is estimated as the ratio between the accumulated stripped hot gas mass since the first
RP stripping event suffered by the satellite and the sum of its current hot gas mass and the accumulated
stripped hot gas.
Again, we see that the mass stripped by RP is larger for less massive satellites residing in more massive halos. For satellites within low-mass haloes
($\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [12.3, 13.2]$), the cumulative stripped hot gas fractions
decrease from $\approx 50$ per cent for $M_{\star} \approx 3 \times 10^{9}\,{\rm M}_\odot$ to $\approx 20$ per cent for $M_{\star} \approx 10^{11}\,{\rm M}_\odot$. For high-mass haloes
($\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [14.1, 15.]$), these fractions increase to $\approx 70$ and $\approx 40$ per cent, respectively.
In our model, both RP and TS act on the hot gas halo, but in general the former process gives rise to a smaller stripping radius than the latter, indicating that RPS dominates TS.
Thus, TS
is considered as a secondary effect
\citep{mccarthy2008, font2008, Bahe15}.
This is also consistent with inferences from observational results,
like the efficient SF quenching detected in
Virgo cluster
\citep{Boselli16}.
We demonstrate
that low-mass satellites are the galaxies mainly affected by the
way in which environmental processes regulate the content of the hot gas
reservoir
by making a test in which
satellites are allowed to keep their hot gas halo at infall
but neither RPS nor TS are activated, so that the reduction of hot gas
is a result of ejection and/or gas cooling.
Only low-mass satellites
($\log (M_{\star} [{\rm M}_{\odot}]) \in [9.5, 10.5]$)
experience a reduction in the quenched fraction which is as high as
$\approx 0.15$ for galaxies
residing within the most
massive haloes
($\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [14.1, 15.]$).
However, the quenched fractions of more massive satellites remain
unchanged.
The lack of significant change in the fraction of quenched massive satellites
when their hot gas reservoir suffers all type of possible effects
(strangulation, gradual removal, suppression of environmental processes)
is a strong evidence
that environmental processes do not play a significant role
on the SF activity in massive satellites
although these galaxies do suffer them as shown
in Figs.~\ref{fig:stripfrac_hot} and \ref{fig:stripfrac_hot_acc}.
This was already deduced from
the good match between model results and observations for the radial
distribution of the fraction of quenched massive galaxies
(see Fig.~\ref{fig:fq_r}).
\begin{figure}
\centering
\includegraphics{Fig16.pdf}
\caption{
Atomic hydrogen gas fraction as a function of stellar mass for satellite
galaxies in model \textsc{sag$_{\beta1.3}$}~within different ranges of main host halo mass,
as indicated in the legend.
Different lines
and associated shaded areas represent the corresponding median values
and $10$ and $90$ percentiles, respectively.
Median values are compared with the
average atomic gas fraction presented by
\citet{Brown17}, who consider
HI data from ALFALFA survey (different symbols).
The match between model results and observations is rather good for
satellites with stellar mass larger than $10^{10}\,{\rm M}_{\odot}$
residing in haloes more massive than
$\approx 10^{12}\, {\rm M}_{\odot}$,
consistent with the good behaviour of the radial dependence of
the fraction of quenched massive satellites.
}
\label{fig:mHI-fraction-halo-bins}
\end{figure}
The cold gas reservoir of satellite galaxies can also be reduced by the
action of RPS,
as inferred from images of highly asymmetric HI distribution
well within the stellar disc in cluster spirals
\citep[e.g.][]{Abramson11,Kenney15, Bellhouse17}.
Fig.~\ref{fig:stripfrac_cold} shows the mean values of the cumulative stripped cold gas fraction as a function of stellar mass of satellite galaxies in model \textsc{sag$_{\beta1.3}$}~which interstellar medium
is not longer shielded by the hot gas halo. They are grouped according to the mass of their main host haloes.
The dependences of these
fractions with stellar mass
and halo mass are quite similar to those
characterising the stripped hot gas
(Fig.~\ref{fig:stripfrac_hot}), although the values
of the fractions of cumulative stripped cold gas are one order of magnitude smaller
than for the hot gas.
The small fraction of cold gas removed by RPS
has no impact on the fractions of quenched galaxies,
which remain unchanged if RPS of cold gas is suppressed.
However, this slight reduction of the cold gas content allows to obtain
a better agreement with the observed
relation of the atomic hydrogen (HI) gas content
as a function of stellar mass
than in the case where RPS of the cold disc is deactivated.
This good agreement is shown in Fig.~\ref{fig:mHI-fraction-halo-bins}
for satellites
within main host haloes of different mass in model \textsc{sag$_{\beta1.3}$}.
HI mass is estimated
assuming that total cold gas is the sum of the contribution
of molecular (H2) and atomic hydrogen content and $30$ per cent helium
\citep[][eq. 2]{boselli14}. The molecular hydrogen content
is estimated adopting the scaling relation of
molecular gas-to-stellar mass ratio \citep[][table 3]{boselli14}, in which
the coefficients correspond to molecular hydrogen masses
estimated from CO intensities using the H-band luminosity-dependent
CO-to-H2 conversion factor.
We compare against measurements
by \citet{Brown17} from a spectral stacking technique
applied to a multi-wavelength sample of satellite galaxies
selected
from the SDSS,
with HI data from ALFALFA survey,
with stellar masses larger than $10^{9}\,{\rm M}_{\odot}$ (different
symbols correspond to average values for different bins of halo mass).
The match between model results and observations is good for
satellites with stellar mass $M_{\star}\gtrsim 10^{10}\,{\rm M}_{\odot}$
residing in haloes more massive than
$\approx 10^{12}\, {\rm M}_{\odot}$,
consistent with the good behaviour of the radial dependence of
the quenched fraction for this galaxy population, giving additional
support to the modelling of mass quenching processes, as discussed in
section~\ref{sec:fq-ms-mh-r}.
For smaller stellar masses,
the mean HI-to-stellar mass fraction is underestimated by
\textsc{sag$_{\beta1.3}$}~regardless of the host halo mass.
This is not related to the modelling of RPS
since this
general trend persists even when the action of RPS on the cold gas is
deactivated,
in which case there is a systematic increase of this fraction
by $\approx 0.25$ dex for all stellar masses.
Hence, the low-mass content of atomic gas might be explained by
levels of star formation higher than expected in low-mass
galaxies as becomes evident from the excess of the cosmic SFRD at $z=0$ (see
Fig.~\ref{fig:SFRDvsz}).
Results from \textsc{sag$_{\beta1.3}$}~regarding the levels of current SF and atomic gas content
call for a more refined treatment
of quiescent SF, like
distinguishing the neutral and molecular phases
of the cold gas as \citep[e.g. ][]{Lagos11a}.
Molecular gas is not significantly affected
by RPS as arises both from observations
\citep[e.g][]{Abramson11} and hydrodynamical simulations \citep{Tonnesen09};
this cold gas phase could be highly perturbed by high-speed tidal interactions
\citep{Scott15}.
\section{Comparison with other SAMs}
\label{sec:SAMcomp}
In a recent work, \citet{Henriques17} study
environmental and mass quenching
using the model described in \citet{henriques_mcmc_2015},
which is an updated version of the Munich semi-analytic model
{\textsc {l-galaxies}}
\citep{guo11}.
Since we adopt the prescription
used by \citet{Henriques17} to model AGN feedback, the discrepancies between
the predicted fractions of quenched galaxies from {\textsc {l-galaxies}} and
those inferred from observations seem to originate in
the modelling of other baryonic (e.g. SN feedback, SF, recycling) and environmental
processes, which affect the complex baryon cycle.
The new features implemented in \textsc{sag}, including the new SN feedback scheme and
treatment of environmental processes,
differ considerably
with respect to the corresponding implementation in {\textsc {l-galaxies}}.
Regarding gas disruption processes, the most relevant difference is that
\textsc{sag}~allows gradual removal of hot gas
reservoir by TS and RPS in both types of satellite galaxies (those
that keep their DM subhaloes and orphans), while
in {\textsc {l-galaxies}}
orphan galaxies
lose immediately their hot gas reservoir by TS, as
this process is considered responsible of the disruption of their DM
subhaloes (in general, hot gas is assumed to be tidally stripped
at the same rate as dark matter).
Their assumption might be too strong, not taking into account
the fact that the mass resolution of the simulation prevents subhalo detection
in some cases.
Therefore, the assumption adopted in \textsc{l-galaxies}
prevents orphans from the gradual removal of hot gas through RPS.
The action of RPS is only allowed
in satellites that keep their substructures.
As described in \citet{guo11}, the stripping radius is estimated
from the balance of self-gravity and ram pressure.
After the stripping event, the remaining hot gas expands out to the
subhalo~$r_\text{vir}$~recorded at infall.
\citet{guo11} argue that feedback could provide the required energy
to redistribute the gas in this manner, which would require outflow velocities
comparable to the circular velocity of the host group. This assumption
on the distribution of the remaining hot gas results in an increased efficiency
of~RPS compared to our procedure, since there is always hot gas
beyond the current stripping radius.
Besides, the halo becomes increasingly more diffuse which leads to reduced cooling
rates. Therefore, their model results in
an increased fraction of passive low-mass satellites
\citep{guo11,henriques13,Hirschmann14}.
With the aim of reducing the excess
of passive satellites predicted by \textsc{l-galaxies},
\citet{henriques_mcmc_2015} define a minimum threshold halo mass
($10^{14}\,{\rm M}_{\odot}$) for RP to be effective as
an stripping mechanism of the hot gas
halo of satellites that keep their substructure,
below which the action of RP is suppressed.
However, despite of this additional condition, discrepancies between predicted and
observed fractions of quenched galaxies still remain.
Their condition for the action of RP
not only differs from the one we adopt (equation~\ref{eq:rpshot}),
which action is self-regulated without any restriction,
but it also considers that the orbital
velocity of the satellite is approximated by the virial
circular velocity of the main halo. This aspect is improved in our
implementation by taking the information of the actual orbits of satellites,
including those of orphans.
Furthermore, model \textsc{l-galaxies}
does not include RPS of the cold gas disc.
\citet{GonzalezPerez14} present a
variant of \textsc{galform} semi-analytic model with
a molecular-based SF law that also assumes
gradual stripping of the hot gas by the action of RP
adopting the same parametrization
for the gravitational restoring as in \textsc{sag}~
(see equation~\ref{eq:rpshot}) but with the value $\alpha_{\rm RP}=2$
suggested by \citet{mccarthy2008}. The main difference with
our implementation is that RP is set to its maximum value for the whole
orbit of the satellite galaxy since
it is
calculated at its pericentre. Therefore, the stripped hot gas is overestimated
leading to lower fractions of atomic gas than those inferred from observational
data \citep[see fig. 7b of][]{Brown17}.
This disagreement would become even larger if RPS of cold gas were included
in their model.
The impact of both cold gas and hot gas stripping
on
the HI fractions of local galaxies
is investigated by \citet{Stevens17} using an updated version of the
\textsc{dark sage}~semi-analytic model,
which evolves the one-dimensional structure of
galactic discs in annuli of fixed specific angular momentum.
Quiescent SF depends on the molecular gas fraction of each annulus;
they consider
both metallicity- and pressure-based prescriptions for
determining the ratio of H2.
The action of RP on both hot gas and cold gas phases
is regulated by conditions similar to those applied in our model \textsc{sag}, that is,
those given by \citet{mccarthy2008} and \citet{gg72},
respectively. An additional important similarity to note is that
\textsc{dark sage}~also considers
that the hot gas halo protects the cold gas disc from the
action of RP. However, the condition adopted to allow RPS
differs from ours, since
RP is able to remove cold gas from the disc
when the total
baryonic mass of the galaxy (cold gas and stars) exceeds the
mass of hot gas halo.
\citet{Stevens17} show that
the satellite HI fractions
are underpredicted by their full
model even when it has been calibrated using the
observed HI fraction of galaxies as a constraint with the greatest weight.
These fractions
become closer to the observational data when cold gas stripping is suppressed.
Among the possible explanations of these results offered by the authors,
there is one directly related to the limitations of SAMs, namely,
the fact that the hot gas of satellite
galaxies is not replenished by accretion from the ICM.
The effect of RPS of cold gas is also taken into account by \citet{luo16} in a branch of \textsc{l-galaxies} \citep{Fu13} which also
assumes galaxy discs divided into multiple rings, and a SFR dependent on the local surface density of the
molecular gas. This model is unable to reproduce the observed trends of quenched fractions of satellites and centrals. The number of low-mass passive galaxies are over-predicted despite gradual stripping of the hot gas is considered. The action of RPS of the cold gas increase the discrepancies with observational data. However, it
is difficult to quantify the effect of this environmental process because the passive fraction of central galaxies is under-predicted for any stellar mass hinting to problems not related with environment.
Even with the aforementioned drawback affecting the hot gas halo of satellites
and a simple modelling of galaxy discs
and quiescent star formation, our model \textsc{sag$_{\beta1.3}$}~is able to achieve better
agreement with observed atomic gas content and quiescent fractions
than other current SAMs.
\section{Summary and conclusions}
\label{sec:conclu}
The updated version of our semi-analytic
model of galaxy formation \textsc{sag}~has been used to
generate one of the galaxy catalogues of the
\textsc{MultiDark Galaxies} project \citep{knebe17},
which is based on the Planck cosmology $1\,h^{-1} \,{\rm Gpc}$
\textsc{MultiDark} simulation MDPL2. The model has been calibrated using
the PSO technique \citep{ruiz2015} combined with a particular set of
constraints and the observational
data defined in \citet{Knebe18}. This catalogue is publicly available
in the \textsc{CosmoSim} database.
We also consider a second galaxy population generated by a slightly different
version of the model, characterized by a change in one of the
parameters involved in the modelling of SN feedback (model \textsc{sag$_{\beta1.3}$}).
We analyse the properties of the galaxies in these two catalogues,
both of star forming and quiescent galaxies, splitting the population
in centrals and satellites, and also sampling them according to the mass
of the main host haloes they inhabit.
The detailed analysis performed allowed us to evaluate the impact of
the improved treatment of environmental effects implemented in the model,
and the modification introduced in the SN feedback scheme, given by
the addition of an explicit
redshift dependence in the estimation of the reheated and ejected mass.
We implement
the gradual removal of hot gas in satellites from RPS and TS,
allowing also the action of these processes on the cold gas disc
under certain conditions.
This is an improvement of our model
with respect to previous studies that ignore the RPS of the cold gas
\citep{guo11, kimm2011, GonzalezPerez14, Henriques17}.
The advantage of our implementation of RPS
with respect to previous works \citep[e.g. ][]{Henriques17, Stevens17}
resides in the use of a
fitting formulae for RP
experienced by galaxies in haloes of different mass as a function of
halo-centric distance and redshift
instead of analytic estimations.
The effect
of TS is considered as an additional mechanism contributing to the
removal of the hot and cold gas. It can also affect the stellar components.
Another important improvement with respect to other
SAMs is the integration of orphan satellites
(Vega-Mart\'inez et al., in preparation).
In the following, we summarize the main results of this work
obtained from the analysis of the two galaxy catalogues generated with
the calibrated model \textsc{sag}~and model \textsc{sag$_{\beta1.3}$}:
\begin{itemize}
\item SN feedback plays an important role in
the star formation history of galaxies.
In order to avoid the excess of the faint end of the SMF
at $z=2$ and to give rise to downsizing in the
stellar mass assembly, it is necessary to reduce the availability of cold
gas for star formation at high redshifts.
Thus, both the reheating of the cold gas phase and the ejection of part
of the hot gas reservoir, with the possibility of being reincorporated later,
must be stronger at earlier epochs.
This has been demonstrated by the inclusion of an explicit redshift dependence
of both the reheated and ejected mass in the new feedback scheme
implemented in \textsc{sag}. Results are sensitive to the value of the power-law slope of such
dependence, denoted by the parameter $\beta$ in our model.
When $\beta=1.99$, as emerged from the calibration process in which the
free parameters are restricted by imposing the SMF at $z=2$ as constraint,
the predicted SFRD becomes too low with respect to observational data
at high redshifts. When fixing that parameter in the value
suggested by the fit of \citet{muratov15} obtained from the FIRE hydrodynamical
simulations ($\beta=1.3$),
the trend of SFRD is reconciled with the observed
one (Fig.~\ref{fig:SFRDvsz}) at the expense of an excess in the faint end
of the SMF at $z=2$ (Fig.~\ref{fig:SAGconstraints_SMF}).
\item
The gradual removal of hot halo gas through a
robust treatment of environmental effects, as the one implemented in our model,
becomes crucial to produce main sequences that
evolve consistently with observational data for both
central and satellite galaxies (Fig~\ref{fig:sSFRmstarz}).
Cooling flows keep replenishing the cold gas reservoir of satellites
after first infall during gradual starvation, delaying the beginning of the quenching of SF.
The differences in normalization and shape of the main
sequence of central and satellite galaxies that still remain
arise because, for a given stellar mass,
central
galaxies inhabit more massive haloes than satellites
(Fig.~\ref{fig:SHMratio}), which impact on the mass of the hot halo
and associated cooling rates.
The fact that galaxies start to form stars earlier in model \textsc{sag$_{\beta1.3}$}~is
reflected in
the lower normalization of the main sequence of star-forming galaxies
achieved by this model with
respect to model \textsc{sag}.
In both models, the main sequences
of central and satellite galaxies are characterized by a decreasing
trend of sSFR with decreasing
stellar mass for $M_{\star} \lesssim 10^{10}\,{\rm M}_{\odot}$. The strong dependence with the virial velocity for low-mass galaxies introduced in the estimation of the reheated and ejected mass is responsible for this effect.
\item
The higher efficiency of SN feedback at high
redshifts, required to recover the downsizing in stellar mass
assembly, must be characterized by a
mild redshift dependence of the reheated and ejected mass,
as in model \textsc{sag$_{\beta1.3}$}, in order to achieve the
expected behaviour of the fractions of quenched galaxies
(those with
$sSFR < 10^{10.7}\,{\rm yr}^{-1}$)
as a function of stellar mass, halo mass and the halo-centric distances
(Figs.~\ref{fig:fq_mstar_mhalo} and~\ref{fig:fq_r}).
The right fractions
of quenched massive galaxies
($\log(M_{\star}[{\rm M}_{\odot}]) > 10.5$)
as a function of both halo mass
and halo-centric distance
support the modelling of those
self-regulating physical processes
that are particularly relevant for massive galaxies
(AGN feedback; disk instabilities), which are still active as quenching
mechanisms even when galaxies become satellites,
dominating over environmental quenching;
the effect exerted by the latter is milder in high-mass galaxies than in low-mass ones.
\item
RPS plays a dominant role among the environmental
processes considered in our model and contributes to regulate
adequately the mass of the hot gas halo and the cold
gas disc (when it is not longer shielded by the hot halo),
being more efficient in removing
gas from galaxies with smaller stellar masses, as expected
(Figs.~\ref{fig:stripfrac_hot}, ~\ref{fig:stripfrac_hot_acc} and~\ref{fig:stripfrac_cold}).
For any galaxy mass, the fraction of
hot and cold gas mass stripped by RP increases with decreasing redshift,
as a result of higher densities achieved by the ICM,
associated to the mass growth of the main host DM haloes, and
the fact
that high redshift infalling galaxies have more time to approach
the cluster core where they suffer strong RP.
The total stripped mass
is higher
in more massive clusters
($M_{\rm vir} \approx 10^{15}\,{\rm M}_\odot$)
than in less massive ones, consistent with the distribution of RP values
inherent to groups and clusters \citep{tecce10}.
The cumulative stripped hot gas fraction increases
from $\approx 20$ to $\approx 50$ per cent for satellites with $M_{\star} \approx 3 \times 10^{9}\,{\rm M}_\odot$ and $M_{\star} \approx 10^{11}\,{\rm M}_\odot$, respectively, residing in low-mass haloes
($\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [12.3, 13.2]$). For high-mass haloes
($\log (M_{\rm halo} [{\rm M}_{\odot}]) \in [14,1 15.]$), these fractions increase to $\approx 40$ and $\approx 70$ per cent, respectively.
\item The cold gas disc is less affected by RP than the hot gas,
with
cumulative stripped cold gas fractions
being one order of magnitude smaller for the former than for the latter.
Thus, RPS on the cold gas does not affect the fraction of quenched galaxies
but it definitely contributes to reach
the right atomic hydrogen (HI) gas content
for satellites with different stellar mass,
especially for more massive ones
($M_{\star} \gtrsim 10^{10}\,{\rm M}_{\odot}$)
residing in haloes with masses
$M_{\rm vir} \gtrsim 10^{12}\, {\rm M}_{\odot}$
(Fig.~\ref{fig:mHI-fraction-halo-bins}).
Distinguishing the neutral and
molecular phases of the cold gas could help to
reproduce the observed levels of HI gas content of low-mass satellites.
\end{itemize}
Our results highlight the impact of specific aspects of mass and environmental quenching on galaxy evolution.
The stronger effect of SN feedback at higher
redshifts and the physics of environmental processes
captured by the RP fitting formulae
combined with the orbital evolution of orphan galaxies
make our model \textsc{sag$_{\beta1.3}$}~able to generate a galaxy population
with fractions of gas and
levels of SF that allow to achieve
better
agreement with observed atomic gas content and quiescent fractions
than other current SAMs.
These attainments of our model
are reached even though it does not include a
detailed treatment of galaxy discs
as in {\textsc {dark sage}} \citep{Stevens17}, or a molecular-based SF law as {\textsc {galform}}
\citep{GonzalezPerez14}.
Differences still found for particular ranges of stellar and halo mass
give hints for further improvement of the model.
\section*{Acknowledgements}
The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V.
(www.gauss-centre.eu) and the Partnership for Advanced Supercomputing in Europe
(PRACE, www.prace-ri.eu) for funding the \textsc{MultiDark} simulation project
by providing computing time on the GCS Supercomputer SuperMUC at Leibniz
Supercomputing Centre (LRZ, www.lrz.de). The MDPL2 simulation has been performed
under grant pr87yi. Our collaboration has been supported by the DFG grant GO 563/24-1.
This work was done in part using the Geryon computer at the
Center for Astro-Engineering UC, part of the BASAL PFB-06, which received
additional funding from QUIMAL 130008 and Fondequip AIC-57 for upgrades.
We thank the referee for useful comments and suggestions that have contributed to improve this work.
We acknowledge Tom\'as Tecce for his valuable contribution in
initial stages of the development of the current version of \textsc{sag}~and Mario Abadi for useful discussion.
SAC acknowledges funding from {\it Consejo Nacional de Investigaciones
Cient\'{\i}ficas y T\'ecnicas} (CONICET, PIP-0387), {\it Agencia Nacional
de Promoci\'on Cient\'ifica y Tecnol\'ogica} (ANPCyT, PICT-2013-0317), and {\it Universidad Nacional de La Plata} (G11-124),
Argentina.
CVM, TH, FC and IDG acknowledge CONICET, Argentina, for their supporting fellowships.
ANR acknowledges funding from ANPCyT (PICT-2014-2862) and from {\it Secretar\'ia de Ciencia y Tecnolog\'ia
de la Universidad Nacional de C\'ordoba} (PID 30720450100484).
AO acknowledges support from project AYA2015-66211-C2-2 of the
Spanish {\it Ministerio de Economia, Industria y Competitividad}.
AMMA acknowledges support from CONICYT-PCHA/Doctorado Nacional 2011-21110870, BASAL PFB-06 and FONDECYT grant 3160776.
GY acknowledges financial support from the {\it Ministerio de Econom\'ia y
Competitividad} and the {\it Fondo Europeo de Desarrollo Regional}
(MINECO/FEDER, UE) in Spain through grant AYA2015-63810-P.
\bibliographystyle{mnras}
|
\section{Introduction}
Multiple imputation (MI) \citep{Rubin1987} is a simple but powerful method for dealing with missing data. MI as originally conceived proceeds in two stages: A data disseminator creates a small number of completed datasets by filling in the missing values with samples from an imputation model. Analysts compute their estimates in each completed dataset and combine them using simple rules to get pooled estimates and standard errors that incorporate the additional variability due to the missing data. %
MI was originally developed for settings in which statistical agencies or other data disseminators provide multiply imputed databases to distinct end-users. There are a number of benefits to MI in this setting: The disseminator can support approximately valid inference for a wide range of potential analyses with a small set of imputations, and the burden of dealing with the missing data is on the imputer rather than the analyst. All analyses conducted on the publicly available files can be based on the same set of imputations, ensuring that differences in results are not due to the handling of missing data.
With the introduction of easy-to-use software to generate imputations and combine estimates it has become increasingly common for users to create their own imputations prior to analysis. The set of methods available to generate imputations has also grown substantially, from simple parametric models and resampling methods to iterative classification and regression tree-based algorithms and flexible Bayesian nonparametric models. There are several textbook treatments of multiple imputation (e.g. \cite{Rubin1987,LittleRubin200209,van2012flexible,CarpenterKenward201302}) but fewer recent reviews of the variety of methods available to create multiply imputed files.
This paper provides a review of MI, with a focus on methods for generating imputations and the theoretical results and empirical evidence available to guide the selection and critique of imputation procedures. We restrict attention to methods for imputing item missing data (imputing the subset of values that are missing for an incomplete observation) in settings with independent observations. Much of the discussion also applies to other data structures, and to problems other than item missing data where MI has proven useful (see \cite{reiter:raghu:07} for some examples of other uses for multiple imputation).
The paper proceeds as follows: Section~\ref{sec:how} briefly reviews the mechanics of multiple imputation for a scalar estimand. Section~\ref{sec:when} reviews the conditions under which the usual MI rules give valid inference. Section~\ref{sec:implications} summarizes the practical implications of the theoretical results, particularly for choosing a method for generating imputations. Section~\ref{sec:singlegen} reviews methods for imputing a single variable subject to missingness. Section~\ref{sec:multiplegen} reviews methods for imputing several variables. Section~\ref{sec:choosing} discusses some of the considerations for choosing an imputation model. Section~\ref{sec:conclusion} concludes with discussion and directions for future work.
\section{Multiple imputation: How Does it Work?}\label{sec:how}
Let $Y_i = (Y_{i1}, Y_{i2}, \dots Y_{ip})$ denote a $p-$dimensional vector of values corresponding to the $i^{th}$ unit and $R_i = (R_{i1}, R_{i2}, \dots R_{ip})$ be a vector of indicator variables representing the response pattern, where $R_{ij} = 1$ if $Y_{ij}$ is observed and is zero otherwise. We will use lowercase letters to distinguish fixed values from random variables, and denote the realized values in a particular dataset with a tilde (e.g., $R_i$ is a random vector, $r_i$ is a particular value that might be taken by $R_i$, and $\tilde r_i$ is the observed response pattern for unit $i$ observed in a particular dataset).
Let $R = \{R_i : 1\leq i\leq n\}$ with $r$ and $\tilde r$ defined similarly.
The observed and missing values from a dataset of size $n$ with response pattern $R$ are denoted $\Yobs(R) = \{Y_{ij} : r_{ij}=1,\ 1\leq j\leq p,\ 1\leq i\leq n\}$ and $\Ymis(R) = \{Y_{ij} : r_{ij}=0,\ 1\leq j\leq p,\ 1\leq i\leq n\}$, respectively.
Where the explicit dependence on the response pattern is a distraction we will drop the functional notation and simply refer to $\Ymis$ and $\Yobs$.
We assume throughout that the missing data are {\em missing at random (MAR) } \citep{Rubin1987}, that is,
\begin{align}
\Pr( R=\tilde r\mid \Yobs(\tilde r)=\tilde y_{obs}&, \Ymis(\tilde r)=y_{mis}, \phi)
\end{align}
takes the same value for all $y_{mis}$ and $\phi$, where $\phi$ parameterizes our model of the response mechanism (the distribution of $(R\mid Y)$). Under MAR we do not need to explicitly model the response process to impute the missing data. \citep[Result 2.3]{Rubin1987}. MI may be used for missing data that are not MAR provided we explicitly model the response mechanism or make other identifying assumptions (see \cite{Rubin2003-zl} for related discussion and examples of MI for non-MAR missing data).
\subsection{Multiple imputation for a scalar estimand}\label{sec:ignorable-mi}
Let $Q$ be an estimand of interest, which may be a function of complete data in a finite population or a model parameter. Let $\hat Q(Y)$ be an estimator of $Q$ with sampling variance $U$ estimated by $\hat U(Y)$; where there is no ambiguity we refer to these as $\hat Q$ and $\hat U$.
In order to fix ideas we focus on scalar $Q$. Inference for vector $Q$ is similar in spirit; see \cite[Chapter 3]{Rubin1987}, also \cite[Chapter 4, Section 3]{Schafer1997} or the review in \cite[Section 2.1]{reiter:raghu:07}.
Assume $\Ymis^{(1)}, \Ymis^{(2)}, \dots, \Ymis^{(M)}$ are $M$ imputations for $\Ymis$. %
Define $\hat Q^{(m)}= \hat Q(\Yobs, \Ymis^{(m)})$, the estimator computed using the $m^{th}$ completed dataset (with $\hat U^{(m)}$ defined similarly), and
\begin{gather}
\bar Q_M = \sum_{m=1}^M \frac{\hat Q^{(m)}}{M},\quad
\bar U_M = \sum_{m=1}^M \frac{\hat U^{(m)}}{M},\quad
B_M = \sum_{m=1}^M \frac{(\hat Q^{(m)} - \bar Q_M)^2}{M-1}\label{eq:miB}.
\end{gather}
These statistics form the basis for inference under MI: $\bar Q_M$ averages the estimate computed in each imputed dataset to obtain an estimate of $Q$. The variance estimator of $\bar Q_M$ has an ANOVA style decomposition:
\begin{equation}
T_M = \bar U_M + \left(1+\frac{1}{M}\right)B_M,
\end{equation}
where $\bar U_M$ is an estimate of the variance of $\hat Q$ if we had the complete data (``within-imputation'' variance), and $B_M$ estimates the excess variance due to the missing values (``between-imputation'' variance). The factor $(1+1/M)$ is a bias adjustment for small $M$, as explained in \cite[][Chapter 3.3]{Rubin1987}.
MI was originally derived under Bayesian considerations. The Bayesian derivation of MI begins with the identities
\begin{align}
P(Q\mid \Yobs) &= \int P(Q\mid \Ymis, \Yobs)P(\Ymis\mid \Yobs)\,d\Ymis\label{eq:mippred}\\
\E(Q\mid \Yobs) &= \E(\E(Q\mid \Ymis, \Yobs)\mid \Yobs)\label{eq:mibayesQ}\\
\Var(Q\mid \Yobs) &=
\E(\Var(Q\mid \Ymis, \Yobs)\mid \Yobs)
\nonumber\\
&\quad + \Var(\E(Q\mid \Ymis, \Yobs)\mid \Yobs)\label{eq:mibayesT}
\end{align}
When imputations are generated from $P(\Ymis\mid\Yobs)$, the MI statistics are Monte Carlo estimates of the relevant quantities:
\begin{align}
\bar Q_M&\approx \E(\E(Q\mid \Ymis, \Yobs)\mid \Yobs) = \E(Q\mid \Yobs)\\
\bar U_M&\approx \E(\Var(Q\mid \Ymis, \Yobs)\mid \Yobs),\\
(1+1/M)B_M&\approx \Var(\E(Q\mid \Ymis, \Yobs)\mid \Yobs)\\
T_M &\approx \Var(\E(Q\mid \Yobs)).
\end{align}
\cite{Rubin1987} proposed constructing confidence intervals for $Q$ based on an asymptotic normal approximation to the posterior distribution \eqref{eq:mippred}: Taking $M$ to infinity, $(\bar Q_\infty-Q)\sim N(0, T_\infty)$ approximately in large samples. In large samples with finite $M$ interval estimation for $Q$ proceeds using a reference $t-$distribution for $\bar Q_M$: $(\bar Q_M-Q)\sim t_{\nu_M}(0, T_M)$. \cite{Rubin1987} computed an approximate value for $\nu_M$ using a moment matching argument, obtaining $\nu_M = (M-1)\left(1 + {1/r_M} \right)^2$ where $r_M = (1+1/M)B_M/{\bar U_M}$ is a measure of the relative increase in variance due to nonresponse. \cite{Barnard1999-ke} proposed an alternative degrees of freedom estimate with better behavior in moderate samples, suggesting it for general use. See \cite{reiter:raghu:07} for a review of combining rules for more general estimands.
\section{Multiple Imputation: When Does it Work?}\label{sec:when}
In this section we give a high-level review of some of the justifications for using MI and the estimators given above. Special consideration is given to results that can inform the selection of an imputation model.
\subsection{Bayesian (in)validity under MI}\label{sec:bayes-valid}
Since the MI estimators were derived under Bayesian arguments we might hope that MI yields valid Bayesian inference. In general it does not. Suppose the analyst has specified a Bayesian model as $P_A(Y, Q) = P_A(Y\mid Q)P_A(Q)$. The analyst's inference is based on the posterior distribution
\begin{align}
P_A(Q\mid \Yobs) &= \int P_A(Q\mid \Ymis, \Yobs)P_A(\Ymis\mid \Yobs)\,d\Ymis. \label{eq:QAnalystpost}
\end{align}
Now suppose the imputer has generated imputations according to $\Ymis^{(m)}\sim P_I(\Ymis\mid \Yobs)$. On computing $\hat Q(\Yobs, \Ymis^{(m)})$ the analyst has a draw from the hybrid model
\begin{align}
P_H(Q\mid \Yobs) &= \int P_A(Q\mid \Ymis, \Yobs)P_I(\Ymis\mid \Yobs)\,d\Ymis\label{eq:QHybridpost}
\end{align}
If $P_A(\Ymis\mid\Yobs) = P_I(\Ymis\mid\Yobs)$, then MI delivers the analyst's posterior inference in the sense that $\hat Q^{(m)}$ is a draw from \eqref{eq:QAnalystpost}. If the posterior distribution for $Q$ is approximately normal and $M$ is not too small the MI statistics will give a reasonable approximation to the posterior.
However, in practice the imputer and the analyst will likely have different models for $(\Ymis\mid \Yobs)$.
Even if one analyst should happen to share the same model as the imputer, the next analyst may have a different set of beliefs encoded in their model, resulting in $P_{A'}(\Ymis\mid\Yobs) \neq P_A(\Ymis\mid\Yobs)$. In this case the imputer cannot deliver valid Bayesian inference to both analysts with a single set of imputations.
Since Bayesian validity is generally unattainable (and good repeated sampling behavior is desirable in its own right), MI is usually evaluated based on its frequentist properties. The remaining subsections explore conditions under which MI yields valid frequentist inference.
\subsection{Frequentist Validity: Conditions on complete data inference}
We will follow \cite{Rubin1996} and assume that the complete data inference is at least {\em confidence valid}, meaning that a nominal $100(1-\alpha)\%$ confidence interval has actual coverage at least $100(1-\alpha)\%$. (The stronger condition of {\em randomization validity} requires that the nominal and actual coverage rates agree.) We also assume that the sampling distribution of $\hat Q$ is normal, so that valid confidence intervals can be obtained from $\hat Q$ and $\hat U$. In this case confidence validity requires that
\begin{align}
\E(\hat Q) &= Q\label{eq:comdata1}\\
\E(\hat U) &\geq \Var(\hat Q)\label{eq:comdata2},
\end{align}
where the expectation and variance are over repeated sampling. Randomization validity obtains when $\E(\hat U) = \Var(\hat Q)$.
We depart slightly from \cite{Rubin1996,Rubin1987} in omitting any conditioning on fixed values in a finite population.
In practice normality and \eqref{eq:comdata1}-\eqref{eq:comdata2} may only hold asymptotically, or when particular modeling assumptions are correct. Whether this is plausible for a particular analysis will depend on the nature of $\hat Q$. For our purposes we will assume that any necessary conditions for confidence validity with completely observed data are satisfied, since our primary consideration is the impact of missingness and imputation. Of course, if the complete data inference is not valid it would be unreasonable to expect MI or any other missing data procedure to remedy the issue.
\subsection{Proper imputation for valid inference}
Chapter 4, Section 4.2 in \cite{Rubin1987} outlines conditions under which MI inferences are randomization or confidence valid when $M=\infty$. Imputations satisfying these conditions for a particular estimand $Q$ and posited response mechanism are known as {\em proper} imputations. Proper imputation coupled with valid complete data inference yields valid MI inference \citep[Result 4.1]{Rubin1987}. It is important to remember that imputations are only proper with respect to a particular estimand $Q$ and a posited response mechanism.
We focus on three essential conditions necessary for an imputation procedure to be proper for an estimand $Q$. (The other conditions are somewhat technical and generally not the source of improper imputations and invalid inference in practice.)
\subsubsection{Three essential conditions for proper imputation.}\label{sec:essential}
\cite{Rubin1996} distilled the formal definition of proper imputation given in \cite[Section 4.2]{Rubin1987} into three conditions that generally ensure imputations are proper. They concern the behavior of the MI statistics under repeated realizations of the response mechanism, holding the sample values $Y$ fixed (that is, under repeated sampling from $P(R\mid Y)$). The first two conditions require that $\bar Q_{\infty}$ and $\bar U_{\infty}$ be approximately unbiased for $\hat Q$ and $\hat U$:
\begin{align}
\E(\bar Q_\infty \mid Y) &\approx \hat Q(Y)\label{eq:prop1} \\
\E(\bar U_\infty \mid Y) &\approx \hat U(Y)\label{eq:prop2},
\end{align}
where the expectations are with respect to $P(R\mid Y)$.
Naturally \eqref{eq:prop1}-\eqref{eq:prop2} will hold if $P(\Ymis\mid\Yobs)$ is correctly specified by the imputer. However, imputations made under misspecified models can still satisfy \eqref{eq:prop1}-\eqref{eq:prop2} so long as they broadly capture the features of the predictive distribution that are relevant for computing $Q$ and $U$ and the proportion of missing data is not extreme. To see this more clearly we can write %
\begin{align}
\E(\bar Q_\infty \mid Y) &= \sum_{m=1}^\infty \E\left( \hat Q(Y_{obs}(R), Y^{(m)}_{mis}(R))\mid Y\right).
\end{align}
With no missing data the expectations inside the sum are all $\hat Q(Y)$. With modest amounts of missing data, the imputed values need to be sufficiently poor to overwhelm the influence of the observed data in computing $Q$. (What constitutes ``sufficiently poor'' naturally depends on $Q$.) Similar logic applies to $\bar U_\infty$.
The third condition for proper imputation is more subtle: It requires that the between-imputation variability $B_\infty$ be approximately unbiased for the variance of $\bar Q_\infty$:
\begin{equation}
\E(B_\infty\mid Y)\approx \Var(\bar Q_\infty\mid Y).\label{eq:properB}
\end{equation}
Satisfying this condition generally requires that we account for uncertainty {\em in the imputation model itself} (or equivalently uncertainty in the parameters indexing a model class), since the observed data used to estimate the model, $\Yobs(R)$, varies over samples from the response mechanism. (Recall that the variance in \eqref{eq:properB} is with respect to $P(R\mid Y)$.)
Many seemingly reasonable stochastic imputation procedures fail to be proper because they do not satisfy \eqref{eq:properB}; these include imputing from a model by plugging in the MLE or drawing imputations from the empirical distribution of observed cases \citep[][Ch. 4]{Rubin1987}. Accounting for uncertainty in the imputation model can be achieved (or approximated) in a variety of ways, such as sampling the parameters indexing a particular model class from their posterior under a Bayesian model or through small adjustments to the bootstrap (as described in Section~\ref{sec:hotdeck}). See Section~\ref{sec:uncertainimp} for further discussion.
\subsection{Congeniality and confidence validity}\label{sec:congenial}
It is well-known that the MI estimate $T_\infty$ can be inconsistent for certain choices of $Q$ \citep{Wang1998-wp, Robins2000-wp,Kim2002-cy, Nielsen2003-er, Kim2006-ea}. The bias is typically positive and tends to have limited influence on coverage rates for common estimands when the amount of missingness is not extreme \citep{Rubin2003-zl}. \cite{Rubin1996} reviewed early examples of inconsistency and gave sufficient conditions for MI inference to be confidence proper (i.e., for $T_\infty$ to conservatively estimate $Var(\bar Q_\infty)$); they are similar to the conditions in Section~\ref{sec:essential}, averaged over repeated sampling of $Y$ in addition to the response mechanism. %
\cite{Meng1994} introduced the concept of {\em congeniality} for understanding the inconsistency of the MI variance estimate. Roughly, an analysis procedure is { congenial} to an imputation model $P_I(\Ymis\mid\Yobs)$ if we can take the complete data analysis and embed it into a Bayesian model $P_A(Y\mid Q)P_A(Q)$ such that
\begin{enumerate}
\item Its posterior $P_A(Q\mid Y)$ recapitulates the desired analysis in the sense that
\begin{gather}
\E_A(Q\mid Y) = \hat Q(Y),\quad \Var_A(Q\mid Y) = \hat U(Y).
\end{gather}
\item It matches the imputation model, i.e.,
\begin{equation}
P_A(\Ymis\mid \Yobs) = P_I(\Ymis\mid \Yobs).
\end{equation}
\end{enumerate}
Under congeniality, MI delivers samples from $P_A(Q\mid \Yobs)$ (Section~\ref{sec:bayes-valid}), which we have constructed to yield confidence valid inference. Unless the analyst is the imputer, congeniality is less a condition we should try to satisfy than one we should try to fail gracefully -- uncongeniality is generally ``the rule not the exception'' \citep{Xie2017congenial}, for the same reasons discussed in Section~\ref{sec:bayes-valid}.
\cite{Xie2017congenial} revisited the behavior of MI inferences under uncongeniality and provided a host of new results. At a high level their findings affirm and generalize common rules of thumb originating with \cite{Meng1994}: Even if the ``true'' model is nested within the imputer's and the analyst's models (e.g., if the imputation model includes both relevant and irrelevant covariates in an otherwise correctly specified regression model for the missing data), standard MI inference may be invalid. However, if the analyst's procedure is {\em self-efficient} (meaning essentially that their estimator cannot be improved by ignoring relevant data \citep{Meng1994,Meng2003-qm}), then:
\begin{enumerate}
\item When the imputer's model is more saturated than the analyst's, the usual MI inference is confidence valid and generally robust.
\item When the imputer's model is less saturated than the analyst's, confidence validity is not guaranteed.
\end{enumerate}
It is generally safer to conduct an uncongenial analyses under (1) than under (2), since conservative inferences will obtain. \cite{Xie2017congenial} also provide remarkably simple and broadly applicable (if somewhat exacting) alternative variance estimates that are valid under uncongeniality: Use $T^*_M=2T_M$ for a vector $Q$, or sum and square the standard errors for a univariate $Q$: $T^*_M = (\sqrt{U_M} + \sqrt{B_M})^2 + (1/M)B_M$.
Like most strong theoretical results, \cite{Xie2017congenial}'s results depend on a number of assumptions. One of these assumptions is that the true model (``God's model'') is nested within the imputation model class. In his discussion of the paper, \cite{Reiter2017congenial} notes that ``[I]n my experience, very low coverage rates in MI confidence intervals arise more often from the imputation procedure generating bias in [$\bar Q_\infty$] than from bias in the MI variance estimator,'' often due to rote application of default imputation procedures. This has been in part a shared experience (\cite{Murray2016-cy}), motivating the focus of this review on the specification of imputation models.
\section{Practical Implications of theoretical results for Imputation Modeling}\label{sec:implications}
The theoretical results summarized above suggest a number of practical considerations for generating imputations. These are reviewed below; for more detailed discussion and examples, see e.g. \cite{Rubin1987,Little1988,Rubin1996,van2012flexible}. Throughout this section and the rest of the paper we will continue to refer to procedures that generate imputations as ``imputation models'', regardless of whether they are completely specified probability models.
\subsection{Imputations should reflect uncertainty about missing values and about the imputation model.}\label{sec:uncertainimp}
The goal in multiple imputation is to account for uncertainty due to the missing values in subsequent inference. This is a different objective than estimating or predicting the missing values, which could generally be achieved via simpler means. The situation in MI is similar to the more familiar task of constructing valid predictive intervals with a regression model, where we need to account for uncertainty in the unobserved response as well as uncertainty in the regression fit.
Suppose we have a single variable subject to missingness, to be imputed using a regression model. If we were only concerned with reconstructing the missing values, we would just impute the fitted values. This would clearly lead to invalid MI inferences. Instead, MI propagates the intrinsic uncertainty about the missing values via some stochastic mechanism, for example, by adding a randomly generated residual to the regression prediction. However, to achieve at least approximately proper imputations we also need to account for uncertainty about the imputation model itself -- that is, uncertainty in the fitted values of the regression model.
Methods that do not appropriately reflect both sources of uncertainty tend to violate \eqref{eq:properB} and underestimate the between-imputation variance, yielding standard errors that are too small and anti-conservative inferences \citep{Rubin1987,Rubin1996}.
Bayesian imputation procedures provide a natural mechanism to account for model uncertainty. Imputations are generated from
\begin{equation}
P(\Ymis\mid\Yobs) = \int P(\Ymis\mid \theta, \Yobs)P(\theta\mid\Yobs)d\theta.
\end{equation}
where $\theta$ is a parameter indexing a model for $Y$ (or a model for $\Ymis$ given $\Yobs$). To see how model uncertainty propagates, observe that imputations can be sampled compositionally: For $1\leq m\leq M$, first draw a value $\theta^{(m)}\sim P(\theta\mid\Yobs)$ and then sample $\Ymis^{(m)}\sim P(\Ymis\mid \theta^{(m)}, \Yobs)$. Model uncertainty is represented by $P(\theta\mid\Yobs)$, and the intrinsic uncertainty about the missing values is represented by $P(\Ymis\mid \theta, \Yobs)$.
Approximations to full Bayesian inference have also proven useful: \cite{Rubin1986-ge}'s approximate Bayesian bootstrap for proper hot deck imputation is one early example (Section~\ref{sec:hotdeck}). Chapter 10 of \cite{LittleRubin200209} reviews several others.
Of course, Bayesian modeling is not magic -- if $\theta$ indexes a class of misspecified models then we should expect our imputations and inferences to suffer, at least for estimands that are sensitive to this misspecification. For example, when $\Ymis$ contains variables with significant skew a multivariate normal imputation model would likely yield approximately valid inference for marginal means but invalid inference for some marginal quantiles, since \eqref{eq:prop1} can be violated when $Q$ is an extreme quantile.
From a coverage perspective, model misspecification becomes increasingly consequential in large samples where the complete data standard errors are small and $P(\theta\mid\Yobs)$ will tend to concentrate on the parameters of the ``best'' misspecified model. Even small biases due to misspecification in the imputation model can become large relative to the pooled standard errors. Enlarging the imputation model class $P(Y \mid \theta)$ via non- and semiparametric Bayesian modeling can guard against misspecification and also mitigate the artificial certainty implied by fixing a regular parametric model and only considering uncertainty in its parameters. Section~\ref{sec:npbayesimp} explores recent promising developments in this area.
\subsection{Imputation models should generally include as many variables as possible.}
There are multiple reasons for entertaining the largest possible imputation model: The missing at random assumption tends to be more tenable as more completely-observed variables are added to the imputation model. In addition, if variables predictive of the missing values are left out of the imputation model but used to compute $Q$ or $U$, then the imputations will be improper -- the imputed values will be incorrectly independent of the omitted variables, leading to bias over repeated imputations (violations of \eqref{eq:prop1} or \eqref{eq:prop2}) \citep{Rubin1996}. In this case the analysis and imputation models are uncongenial in the ``wrong'' way -- the imputer's model is less-saturated than the analysis model. In sum, the cost of excluding a relevant variable (invalid inference) is often greater than the cost of including an irrelevant variable (roughly, additional variance). This is particularly relevant when the analyst and imputer are not the same, and the imputations must support many unspecified analyses. Even when the imputer and the analyst are the same it would be useful to generate one set of imputations that can support the usual process of iterative model building and refinement, rather than generating a new set of imputations for each analysis model that is considered. See \cite{Collins2001-bv} and \cite{Schafer2003-hh} for further discussion of the tradeoffs involved.
These points are particularly relevant for design variables in complex surveys. Design-based estimators will typically use stratum and cluster information to compute $U$. \cite{Reiter2006-za} show empirically that failing to account for an informative sampling design can lead to invalid inference. They suggest including indicator variables for strata and cluster membership in the imputation model, or including stratum fixed effects and cluster random effects in imputation models. It may be useful to include estimated response propensities or final adjusted survey weights (sampling weights with e.g. calibration and post-stratification adjustments) as well, especially if complete design information is not available to the imputer \citep{Rubin1996}. %
\subsection{Imputation models should be as flexible as possible.}
Finally, imputation models should try to ``track the data'' \citep{Rubin1996} by modeling relevant features of the joint distribution of the missing values. Loosely, a feature of the joint distribution is relevant if it is a possible target of inference itself, or more generally if it yields a more accurate predictive distribution for the missing data. Interactions, nonlinearities, and non-standard distributional forms are all potentially relevant features.
As \cite{Meng1994} succinctly put it, ``Sensible imputation models should not only use all available information to increase predictive power, but should also be as general and objective as practical in order to accommodate a potentially large number of different data analyses.'' We would add that where possible, imputation models should have some capacity to {\em adapt} to unanticipated features of the data (such as interactions, nonlinearities, and complex distributions), especially when the imputer has limited time and resources to spend on iteratively improving the imputation model. %
\section{Generating imputations for a single variable}\label{sec:singlegen}
We begin by cataloging some of the more common approaches to generating imputations for a single variable subject to missingness, conditional on other fully observed variables. In the next section we consider how these can be extended to generate imputations for several variables.
\subsection{Regression Modeling}
Imputation by sampling from univariate regression models is conceptually straightforward. Generalized linear models and extensions to deal with complications such as zero-inflation and truncation are popular options; these are not reviewed in depth here but see e.g. \cite{VanBuuren1999}, \cite{Raghunathan2001}, \cite{su2011multiple}, or \cite{van2012flexible} (Chapter 3). These methods are quite common in practice, but since most readers will be familiar and they are well-reviewed elsewhere we will not enumerate them here.
To generate proper imputations some method should be used to account for parameter uncertainty -- simple strategies like sampling from the regression model with parameters fixed at the observed data MLE are generally improper.
Posterior sampling under a non- or weakly informative prior tends to be proper when the model fits well. Prior distributions can also ease problems like separation in logistic regression and apply helpful regularization in conditional models with many variables in the conditioning set \citep{su2011multiple}.
\subsection{Hot Deck/Nearest Neighbor Methods.}\label{sec:hotdeck}
The hot deck and other nearest-neighbor methods \citep{Chen2000-gy,Andridge2010-hu} begin by defining a distance metric between cases in terms of the observed covariates. Imputations for a missing value are borrowed from a nearby completely observed case (the ``donor''). These methods tend to be simpler to implement than fully specified regression models and often make fewer assumptions. However, these methods are far from assumption free -- the choice of distance metric, the definition of the donor pool, and how to sample from the donor pool all influence the quality of imputations.
The hot deck \citep{Andridge2010-hu} defines distance via cross-classifications of fully observed variables which determine adjustment cells. Missing values are imputed by sampling with replacement from the pool of donors within the same cell. This strategy ensures that all imputations are plausible values, which is an appealing feature relative to regression imputation. Complications arise when there are many fully observed variables to incorporate into the cross-classification or when the sample size is low, leading to many small or empty adjustment cells.
MI with the hot deck is also known to be improper for simple estimands like a population mean \citep{Rubin1986-ge}. The hot deck effectively assumes that the distribution of missing values within an adjustment cell is {\em exactly} the empirical distribution of the observed values within that cell, which leads to $B$ having downward bias (due to ignoring uncertainty in the implicit imputation model). \cite{Rubin1986-ge} propose a simple modification that makes the hot deck proper, based on an approximation to the Bayesian bootstrap \citep{Rubin1981-os}. Instead of sampling the $n_m$ missing values from the empirical distribution of the $n_o$ observed values within an adjustment cell, the approximate Bayesian bootstrap (ABB) first samples a set of $n_o$ values with replacement from the observed data and then samples $n_m$ imputed values with replacement from this set. This simple adjustment yields proper imputations for the population mean of the adjustment cell \citep{Rubin1986-ge}. (See also \cite{Kim2002-cy} for a more accurate variance estimate in small samples.)
Predictive mean matching (PMM) \citep{Little1988} instead measures the distance between cases by the distance between their predicted means for the variable subject to missingness (traditionally estimated using a linear regression, although in principle any method could be used to make the prediction). PMM generalizes the hot deck, which is a special case of PMM using saturated models with categorical predictors. By avoiding the discretization and making some assumptions about the relationships between the predictors and the response (such as linearity) PMM can handle more variables than the hot deck, but may be sensitive to the predictive model specification.
To define the donor pool \cite{Heitjan1991-jn} proposed sampling from a window of $k$ nearby potential donors in PMM in the hope of making the method approximately proper. The donor's value may be imputed, or its residual can be added to the predicted mean of the missing value to generate an imputation. \cite{Schenker1996-bp} found these two approaches to perform similarly in simulations; the former will always impute a previously realized value, which may be desirable. See \cite{Vink2014-yr} for an approach to semi-continuous variables. \cite{Morris2014-vx} compared newer developments and current implementations of these techniques, cautioning in particular against the imputation of a single nearest neighbor (which appears to be common in software implementations of PMM) as it is improper.
\begin{figure}[ht]
\begin{center}
\begin{tikzpicture}[
scale=0.9,
node/.style={%
draw,
rectangle,
},
node2/.style={%
draw,
circle,
},
]
\node [node] (A) {$y_1<0.9$};
\path (A) ++(-135:\nodeDist) node [node2] (B) {$A_1$};
\path (A) ++(-45:\nodeDist) node [node] (C) {$y_2<0.4$};
\path (C) ++(-135:\nodeDist) node [node2] (D) {$A_2$};
\path (C) ++(-45:\nodeDist) node [node2] (E) {$A_3$};
\draw (A) -- (B) node [left,pos=0.25] {no}(A);
\draw (A) -- (C) node [right,pos=0.25] {yes}(A);
\draw (C) -- (D) node [left,pos=0.25] {no}(A);
\draw (C) -- (E) node [right,pos=0.25] {yes}(A);
\end{tikzpicture}
\hspace{0.05\linewidth}
\begin{tikzpicture}[scale=3.3]
\draw [thick, -] (0,1) -- (0,0) -- (1,0) -- (1,1)--(0,1);
\draw [thin, -] (0.8, 1) -- (0.8, 0);
\draw [thin, -] (0.0, 0.4) -- (0.8, 0.4);
\node at (-0.1,0.4) {0.4};
\node at (0.8,-0.1) {0.9};
\node at (0.5,-0.2) {$y_1$};
\node at (-0.3,0.5) {$y_2$};
\node at (0.9,0.5) {$A_1$};
\node at (0.4,0.7) {$A_2$};
\node at (0.4,0.2) {$A_3$};
\end{tikzpicture}
\end{center}
\caption{(Left) An example CART tree, with internal nodes labeled by their splitting rules and terminal nodes given labels $A_h$. (Right) The corresponding partition of $(Y_1, Y_2)$.}
\label{fig:treestep}
\end{figure}
PMM and the hot deck can be made more adaptive using recursive partitioning.
\cite{reiter2005using} and \cite{Burgette2010-dq} proposed imputation via classification and regression trees (CART, \cite{Breiman1984-jg}). A tree is grown using fully observed data to predict the variable subject to missingness. Then each incomplete case is assigned to its corresponding leaf, and an imputation is sampled from donors within in the same leaf. The imputer can control the size of the donor pool by growing the tree down to a specified minimum leaf size. This is a special case of PMM using CART to generate predictions; we could also think of it as an adaptive hot deck that leverages the most predictive variables and balances the size of the adjustment cells. Figure~\ref{fig:treestep} shows an example tree grown on two variables $(Y_1, Y_2)$ to impute a third ($Y_3$), along with the corresponding partition which forms the adjustment cells.
\cite{reiter2005using} and \cite{Burgette2010-dq} drew ABB samples from within the leaves in an effort to generate proper imputations. \cite{van2012flexible} (Algorithm 3.6) suggested also accounting for uncertainty in the tree itself by growing it on a different bootstrap sample for each imputed dataset. \cite{Doove2014-na} proposed imputation by growing a random forest (an ensemble of trees) \citep{breiman2001random} of size $k$ by bootstrapping the complete cases and (optionally) sub-sampling the variables, as in traditional applications of random forests. An imputed value is generated by sampling from the $k$ trees and then following the procedure to generate a CART imputation. \cite{Shah2014-ul} proposed fitting a random forest, estimating its predictive error variance, and generating imputations as the random forest prediction plus a normally distributed residual.
Limited results exist comparing these different recursive partitioning methods, and there is similarly limited guidance as to how they should be tuned. But they can be fast and effective imputation engines, particularly for large sets of categorical variables that take a relatively limited set of levels (see e.g. \cite{Akande2017-gq}).
\section{Generating Imputations for Multiple Variables}\label{sec:multiplegen}
There are two basic strategies for imputing multivariate missing data: Jointly modeling the variables subject to missingness, or specifying a collection of univariate conditional imputation models that condition on all the other variables (this approach goes under various names including sequential regression multivariate imputation \citep{Raghunathan2001} and multiple imputation by chained equations \citep{VanBuuren1999}, but we will use ``fully conditional specification'' (FCS) as in \cite{VanBuuren2006}). Joint models can be further classified into ``simultaneous'' approaches that define a multivariate distribution $f(Y)$ directly or ``sequential'' approaches that build up a multivariate distribution using a ladder of conditional distributions, where the model for each variable conditions only on those earlier in the sequence. Appendix~\ref{sec:software} has pointers to software implementations of many methods described in this section.
To describe the different approaches we need some new notation: Let $\Yobsj$ and $\Ymisj$ denote the set of observed and missing values for the $j^{th}$ variable. Let $\Yimp$ denote an imputed dataset, and $\Yimpj$ denote a set of imputations for $\Ymisj$. We will use the subscript $(-j)$ to denote the same quantities for all but the $j^{th}$ variable.
\subsection{Joint specification: Simultaneous approaches}\label{sec:jointsim}
Early simultaneous joint modeling approaches were based on the multivariate normal (MVN) or $t$ distribution; these are reviewed in \cite{Schafer1997} and \cite{LittleRubin200209}. For high dimensional continuous observations low-rank structure can be imposed on the covariance matrix \citep{Audigier2016-va}. Various authors have proposed imputing categorical data under a misspecified MVN model, either leaving the continuous imputations for discrete variables as-is or rounding them based on some thresholds \citep{Horton2003-ak,Bernaards2007-ae}. This is naturally more complicated when the discrete variables are not ordinal, particularly if they take many levels. Additionally, end users may not trust imputations from a data disseminator if the imputed data appear invalid. Therefore it is often preferable to use models that are appropriate for the types of variables at hand.
For small numbers of strictly discrete variables a simple multinomial model may be feasible. However, with a large number of discrete variables it is impossible to fit saturated multinomial models and further restrictions are necessary. Options include log-linear models \citep{Schafer1997}, latent class models \citep{Vermunt2008,Gebregziabher2010,Vidotto2015-zq}, or multiple correspondence analysis \citep{Audigier2017-ye} (which is closely related to a certain class of multivariate logit models \citep{Fithian2017-nt}). %
Joint models for mixed continuous and categorical data are also available. For the remainder of Section~\ref{sec:jointsim}, suppose we have collected the continuous variables into a vector $Y$ and the discrete variables into another vector $X$.
The
general location model (GLOM) %
\citep{Olkin1961,Little1985,Schafer1997} %
assumes that $(Y\mid X=x)\sim
N(\mu_x, \Sigma_x)$ and $X\sim \pi$.
(\cite{Liu1998} generalized the $(Y\mid X)$ model to the larger class of elliptically symmetric distributions.)
The number of parameters in this saturated model grows rapidly with the sample space of $X$, so imputers typically impose further constraints. Examples include common covariance structure ($\Sigma_x\equiv \Sigma$ for all $x$), removing higher-order effects from the conditional means by specifying
$\mu_x = D(x)B$ for a matrix of regression coefficients $B$ and design
vector $D(x)$, and imposing log-linear constraints on $\pi$ to rule our higher-order interactions in the marginal model for $X$.
\subsubsection{Mixtures and Nonparametric Bayesian Models. }\label{sec:npbayesimp} Even without additional parameter constraints, most parametric joint models make restrictive assumptions. Mixture models provide a simple and expressive way to enrich a parametric model class. For example, latent class models for categorical data are mixtures of independence models (log-linear models with only main effects) which have proven useful in multiple imputation \citep[e.g.,][]{Vermunt2008, Gebregziabher2010}. Mixtures of multivariate normal distributions can model complex features of joint continuous distributions \citep{Bohning2007,Elliott2007}.
Several Bayesian nonparametric models have recently been proposed for multiple imputation. Most of these are based on infinite mixture models or their truncated approximations (but see \cite{Paddock2002-os} for an early exception based on Polya trees, and also the sequential regression approach in \cite{Xu2016-xy}). Relative to parametric Bayesian approaches these models are appealing for their ability to grow in complexity with increasing sample size. Under some circumstances this can allow the model to capture unanticipated structure like interactions and nonlinear relationships or nonstandard distributions, reflecting these in the imputed values.
Recall that we have separated the data into vectors of categorical variables $X$ and continuous variables $Y$. For imputing multivariate categorical data, \cite{Si2013} adopt a truncated version of the Dirichlet process mixture of product multinomials (DP-MPMN) proposed
by \cite{Dunson2009}. This is a latent class model with a large number of classes (say $\kx$) and a particular prior over the class distribution.
Suppose the $j^{th}$ categorical variable takes (possibly unordered) values indexed by ${1,2,\dots,d_j}$ and
let $\Hxi \in \{1, \dots, \kx\}$ be a latent mixture component index for observation $i$. Let $\Pr(X_{ij} = x_{ij} \mid \Hxi = \hx) = \psi_{\hx x_{ij}}^{(j)}$. The DP-MPMN model assumes that
\begin{gather}
\Pr(\Hxi=\hx) = \byx{\phi_\hx}\label{eq:dx0}\\
\Pr(X_i = x_i \mid \Hxi = \hx, \Psi) = \prod_{j=1}^{p}\psi_{\hx x_{ij}}^{(j)}\label{eq:dx},
\end{gather}
so that the elements of $X$ are conditionally independent given the latent class membership.
The prior on $\phi$ is a truncated version of the stick-breaking construction for the Dirichlet process (DP) \citep{Sethuraman1994}, introduced in \cite{Ishwaran2001} to simplify Gibbs sampling in DP mixture models:
\begin{gather}
\byx{\phi_\hx} = {\xi_\hx}\prod_{l<\hx}(1-{\xi_l}),\quad
\{{\xi_\hx}\}_{\hx=1}^{\kx} \iid Beta(1, {\alpha}),\quad
{\xi_{\kx}}\equiv 1.\label{eq:Hip}
\end{gather}
The model is completed with prior distributions on $\Psi$ and $\alpha$ (see \cite{Si2013} for a complete specification). \cite{Manrique-Vallier2014-ct,Manrique-VallierSMTrunc} extended this model to assign zero probability to impossible values of $X$, such as cells that are logically impossible (pregnant men or children collecting retirement benefits) or necessarily empty due to skip patterns. \cite{Manrique-Vallier2016-zr} introduced a variant of this model for edit-imputation that simultaneously accounts for missing values and observed values that are logically impossible but present due to measurement error. \cite{Hu2017-mc} extended this model to nested data structures (i.e., hierarchical structures like individuals nested within households) in the presence of structural zeros.
For imputing continuous data \cite{Kim2014-kz} suggested a truncated DP mixture of multivariate normal distributions. Let $\Hyi$ be the mixture component index for record $i$. This model assumes that
\begin{gather}
\Pr(\Hyi=\hy) = \byy{\phi_{\hy}}\label{eq:ymodel0}\\
\left(Y_i\mid \Hyi=\hy, -%
\right) \sim N(\mu_\hy, \Sigma_\hy),\label{eq:ymodel}
\end{gather}
with a prior on $\byy{\phi_{\hy}}$ defined via a stick-breaking process similar to \eqref{eq:Hip}. \cite{Kim2014-kz} modified the model in \eqref{eq:ymodel} to constrain the support of $Y$ to a set $\mathcal{A}$ with bounds determined by a set of linear inequalities, so that $\Pr(Y\not\in \mathcal{A})=0$ under the prior. \cite{Kim2015-eq} extended this approach to simultaneous edit-imputation, generating imputed values for observations outside of $\mathcal{A}$ via a measurement error model.
\cite{Murray2016-cy} built a hierarchical mixture model for mixed continuous and categorical observations by combining the models in \eqref{eq:dx0}-\eqref{eq:dx} and \eqref{eq:ymodel0}-\eqref{eq:ymodel}, with two important adjustments. First, \eqref{eq:ymodel} is modified to include a regression on $X$ with component-specific coefficients:
\begin{equation}
(Y_i\mid X_i = x_i, \Hyi=\hy,-) \sim N(D(x_i)B_\hy, \Sigma_\hy).\label{eq:ymodelwithx}\\
\end{equation}
By default the design matrix $D(x_i)$ encodes main effects. Allowing the component means to depend on $X$ greatly reduces the number of mixture components necessary to capture $X-Y$ relationships. Second, the mixture component indices in each model are given a hierarchical prior introduced by \cite{Banerjee2013}:
\begin{gather}
\Pr(\Hxi=\hx, \Hyi=\hy \mid Z_i=z)
= \byx{\phi_{z\hx}}\byy{\phi_{z\hy}}\label{eq:hmodel}\\
\Pr(Z_i=z) = \lambda_z,\label{eq:zmodel}
\end{gather}
Here $\lambda_z$ is assigned a stick-breaking prior, Each pair $\byx{\phi_z} = \left(\byx{\phi_{z1}},\dots,\byx{\phi_{z\kx}}\right)'$ and
$\byy{\phi_z} = \left(\byy{\phi_{z1}}, \dots,\byy{\phi_{z\ky}}\right)'$ are probability vectors also assigned independent truncated stick breaking priors.
This is a ``mixture of mixtures'' model; marginalizing over the latent variables the joint density is
\begin{equation}
f(X_i, Y_i) = \sum_{z=1}^{\kz} \lambda_z\left(\sum_{\hy=1}^\ky \byy{\phi_{z\hy}}N(Y_i; D(X_i)B_{\hy}, \Sigma_{\hy}) \right)\left(\sum_{\hx=1}^\kx \byx{\phi_{z\hx}}\prod_{j=1}^{p}\psi_{\hx X_{ij}}^{(j)}\right)\label{eq:xy}.
\end{equation}
Each mixture component is itself composed of two mixture models, one for $(Y\mid X)$ and one for $X$. These lower-level mixtures share some parameters ($B, \Sigma,$ and $\Psi$), enforcing a degree of parsimony.
\cite{DeYoreo2016} used a similar hierarchical mixture model constructed based on different considerations, splitting the variables into sets based on their type (ordinal or nominal) and high or low rates of missing values. An expressive model class is specified for the variables with high rates of missing values, and a simpler model class is utilized for variables with low rates of missingness. Ordinal variables are explicitly modeled as such by thresholding mixtures similar to \eqref{eq:ymodelwithx}.
Further extensions, combinations, and enhancements of these models are possible. Despite their complexity, all of these models have been shown to perform well for MI with real, complicated data and little or no tuning.
\subsection{Fully Conditional Specification}
FCS avoids explicit joint probability models by specifying a collection of univariate conditional imputation models instead \citep{VanBuuren1999,Raghunathan2001}. Each univariate model typically conditions on all the remaining variables. In FCS the missing values are imputed by iteratively sampling from these conditional models:
\begin{enumerate}
\item Begin by filling in $\Ymis$ with plausible values to generate an initial completed dataset, stored in $\Yimp$ %
\item For $1\leq j\leq p$, use a univariate imputation method to sample new imputed values for $\Ymisj$ from a distribution $P(\Ymisj \mid \Yobsj, \Yimpnotj)$, and store them in $\Yimpj$.
\item Iterate the previous step until apparent convergence and return the final value of $\Yimp$
\end{enumerate}
This process is repeated $M$ times, saving the returned value as one of the $M$ imputations. Any of the univariate imputation methods in the previous section could be used. This lends FCS some flexibility relative to the joint-simultaneous approaches described above.
But this flexibility comes at a cost: Even if each $g_j$ is a completely specified probability model, taken together they often do not correspond to a proper joint distribution for $Y$ \citep{Arnold1989-wz,Arnold2001-pf}. A set of full conditional distributions that do not correspond to any joint distribution is said to be {\em incompatible}. Simple adjustments like adding polynomial terms or interactions to univariate regression models can induce incompatibility \citep{Liu2014-md}.
While the algorithm above looks like a standard Gibbs sampler, if the conditional models are incompatible the behavior of the FCS imputation algorithm is unclear: The imputations from the FCS algorithm given above may converge to a unique limiting distribution, or fail to converge to any unique limiting distribution, or converge to different distributions depending on the initial values and/or order of the updates.
\cite{Li2012} give examples of incompatible FCS models with fixed parameters whose imputations either diverge or converge to different stationary distributions depending on the order of their updates. This phenomenon seems to be rare in real data, and \cite{Zhu2015-fn} note that estimating rather than fixing parameters ameliorates at least some of the problems in \cite{Li2012}'s examples.
There are some limited convergence results available when the fully conditional specification comprises univariate Bayesian regression models. \cite{Liu2014-md} study an iterative FCS imputation procedure that uses a set of Bayesian regression models $g_j(Y_{ij} \mid , \Ynotj, \theta_j)$ with prior distributions $\pi_j(\theta_j)$. With a slight abuse of notation, define
\begin{align}
g_j(\Yobsj\mid \Yimpnotj, \theta_j) &= \prod_{i=1}^n g_j(Y_{ij}\mid \Yimpj, \theta_j)^{R_{ij}}\label{eq:gj1}\\
g_j(\Yimpj\mid \Yobsj, \Yimpnotj, \theta_j) &= \prod_{i=1}^n g_j(Y_{ij}\mid \Yimpnotj, \theta_j)^{1-R_{ij}} .\label{eq:gj2}
\end{align}
Algorithm~\ref{alg:fcs} gives one iteration of an iterative FCS sampler under these models.
\begin{algorithm}
\caption{Iterative FCS Sampler from \cite{Liu2014-md}}\label{alg:fcs}
\begin{algorithmic}
\item[] For $1\leq j\leq p$,
\begin{enumerate}
\item Sample $\theta_j\sim \pi_j(\theta_j\mid \Yobsj, \Yimpnotj)\propto g_j(\Yobsj\mid \Yimpnotj, \theta_j)\pi_j(\theta_j)$
\item Sample $\Yimpj\sim g_j(\Yimpj\mid \Yobsj, \Yimpnotj, \theta_j)$
\end{enumerate}
\end{algorithmic}
\end{algorithm}
We can compare this approach to a proper MCMC algorithm under a joint model. Specifically we consider a collapsed Gibbs sampler \citep{Liu1994-ri} that targets $P(\Ymis\mid\Yobs) = \int P(\Ymis, \theta\mid \Yobs)d\theta$ directly, by jointly sampling $(\Ymisj, \theta\mid \Yobsj, \Yimpnotj)$ at each step. It is impractical to use directly, but it is helpful to make comparisons with Algorithm~\ref{alg:fcs}.
Let the joint model be given by $f(Y_i\mid \theta)$, with full conditionals $f_j(Y_{ij} \mid \Ynotj, \theta)$ and joint prior distribution $\pi(\theta)$ (where $\theta=(\theta_1,\theta_2,\dots,\theta_p)$). Define $f_j(\Yobsj\mid \Yimpnotj, \theta)$ and $f_j(\Yimpj\mid \Yobsj, \Yimpnotj, \theta)$ as in equations \eqref{eq:gj1}-\eqref{eq:gj2}. Algorithm~\ref{alg:collapsedgibbs} gives one iteration of the collapsed Gibbs sampler.
\begin{algorithm}
\caption{Collapsed Gibbs Sampler for a Joint Model}\label{alg:collapsedgibbs}
\begin{algorithmic}
\item[] For $1\leq j\leq p$,
\begin{enumerate}
\item Sample $\theta\sim \pi(\theta\mid \Yobsj, \Yimpnotj)\propto f_j(\Yobsj\mid \Yimpnotj, \theta)\pi(\theta)$
\item Sample $\Yimpj\sim f(\Ymisj \mid, \Yobsj, \Yimpnotj, \theta)$
\end{enumerate}
\end{algorithmic}
\end{algorithm}
Under some regularity conditions the two algorithms are equivalent in finite samples if we can write $\pi(\theta) = \pi_j(\theta_j)\pi_{(-j)}(\theta_1,\theta_2,\dots,\theta_{j-1}, \theta_{j+1},\dots,\theta_p)$ for any $j$ and the set of $g_j$'s are compatible and correspond to the full conditionals of $f$ \citep{Hughes2014-au}. This is sufficient to ensure that the conditional distributions in both steps of each algorithm agree.
If $\pi(\theta) \neq \pi_j(\theta_j)\pi_{(-j)}(\theta_1,\theta_2,\dots,\theta_{j-1}, \theta_{j+1},\dots,\theta_p)$ for some $j$ but the conditional models are compatible and correspond to the full conditionals of $f$, the two algorithms agree as $n\rightarrow\infty$ provided the FCS algorithm has a unique stationary distribution \citep{Liu2014-md}. Intuitively, in this case the data in $Y^{(-j)}$ influence $\theta_j$ indirectly through the other parameters, but the FCS algorithm ignores this information. Asymptotically the priors become irrelevant in regular parametric models, but in finite samples inference based on the FCS imputations may be inefficient in this regime \citep{Seaman2016-mz}.
Finally, \cite{Liu2014-md} show that if the FCS algorithm uses an inconsistent set of models but has a unique stationary distribution then MI estimates computed using imputations from Algorithm~\ref{alg:fcs} are consistent provided that the following conditions hold:
\begin{enumerate}
\item The collection of conditional models are incompatible, but become compatible with a joint model $f$ after constraining $\theta$.
\item The model class defined by $f$ contains the true distribution that generated the data.
\end{enumerate}
These are rather restrictive; verifying a unique stationary distribution is challenging, as is checking condition 1 above. It also seems unlikely that condition 2 will hold exactly for the simple parametric models in common use.
\cite{Zhu2015-fn} provide some further convergence results for FCS algorithms where each observation is missing at most one value, but without assuming a unique stationary distribution for the FCS chain.
\subsection{Joint specifications: Sequential approach}
Sequential approaches to imputation modeling fix a permutation of $1,2,\dots,p$ and build up a joint distribution from a series of univariate models. For example, if the variables are already in the desired order we would have
\begin{equation}
f(Y) = f_1(Y_1)f_2(Y_2\mid Y_1)f_3(Y_3\mid Y_2, Y_1)\dots f_p(Y_p\mid Y_{p-1},\dots,Y_1).
\end{equation}
Examples of this approach include \citep{ibrahimlips, ibrahim, ibrahim:chen:lip:herr,Lee2016-wc,Xu2016-xy}, among others.
Provided that each $f_j$ is a proper univariate probability model, a sequential specification always defines a coherent joint model, unlike FCS approaches. However, different orderings will generally lead to different joint distributions and potentially different fits. Heuristics have been proposed for selecting the order, for example ordering variables by their types (e.g. \cite{ibrahim}) or percentage of missing values (e.g., \cite{rubin1990efficiently}).
The latter is particularly well-motivated when the missing data are monotone (when there is an ordering such that $R_{ij}=0\Rightarrow R_{ij'}=0$ for $j'>j$. ). If the missing data are not exactly monotone one can identify a permutation that is nearly monotone and use FCS or delete observed values to ``monotonize'' the missing data pattern, so that proper sequential techniques can be used for the majority of missing values (as in \cite{Rubin2003-uf} and extended in \cite{Li2014-lx}).
Another consideration in joint-sequential modeling is that variables
early in the sequence may have complex distributions because they are marginalized over many related covariates. For example, Figure~\ref{fig:sipp} shows the joint distribution of householder earnings and age, conditional on whether the householder has any children living in the same household (the data are from complete cases in wave one of the Survey of Income and Program Participation's 2008 panel). The distributions are quite complicated, and it would be difficult to capture them well with simple parametric regression models in any order.
\begin{figure}
{\includegraphics[width=.8\linewidth]{fig/sipp-age-earn-ownkid}
}
\caption{Joint distribution of householder age and log total earnings, stratified on whether the household includes one of the householder's own children, using the population \cite{Murray2016-cy} constructed from complete cases in the first wave of the Survey of Income and Program Participation's 2008 panel.}
\label{fig:sipp}
\end{figure}
%
\section{Choosing and assessing an imputation strategy}\label{sec:choosing}
\subsection{Comparing FCS and Joint approaches}
FCS and joint approaches have competing strengths. FCS models are relatively simple to implement and widely available in software, especially compared to joint-sequential approaches. Joint-simultaneous models including the multivariate normal, log-linear models, and the GLOM are also easy to set up and widely available, but inflexible in practice even relative to simple FCS procedures (e.g. \cite{VanBuuren2007, Stuart2009,He2010,Drechsler2010,kropko2014multiple}).
More sophisticated joint models can be challenging to implement, although this is changing -- many of the nonparametric Bayesian methods have publicly available implementations (Appendix~\ref{sec:software}). However, even with a good implementation the nonparametric Bayesian models are generally more computationally expensive than simpler joint models (especially those based on low-rank methods, e.g. \cite{Audigier2016-va,Audigier2017-ye}) or FCS methods. Joint-sequential approaches currently take more effort to set up, but they inherit many of the positive features of FCS and joint-simultaneous approaches (univariate models that are readily assessed and modified but also consistent with joint models).
The convergence properties of FCS in general settings is still mostly an open question. The behavior of FCS algorithms under non- or quasi-Bayesian imputation procedures like PMM is entirely an open question. While the lack of a coherent joint distribution does undermine the theoretical justifications for MI inference detailed in \cite{Rubin1987}, experience with FCS in simulations and real applications does not seem to suggest that either lack of convergence or compatibility with a joint model are necessarily overriding concerns.
In fact, under the current theoretical results ensuring that the imputations generated by FCS converge to the imputations under a proper joint model requires using restrictive (implicit) joint models and there is strong empirical evidence that these joint models can be too simple to perform well with realistic data (e.g. \cite{Murray2016-cy,Akande2017-gq}). Therefore at this point it would probably be a mistake to choose the models in an FCS imputation routine to try to ensure convergence; it seems much more important to use flexible, adaptive imputation models wherever possible, whether using a joint or FCS imputation strategy.
Imputers who do choose to use FCS should use flexible univariate models wherever possible and take care to assess apparent convergence of the algorithm, for example by computing traces of pooled estimates or other statistics and using standard MCMC diagnostics \citep[Chpater 11]{gelman2013bayesian}. It may also be helpful to examine the results of many independent runs of the algorithm with different initializations and to use random scans over the $p$ variables to try to identify any convergence issues and mitigate possible order dependence.
\subsection{Practical considerations derived from MI theory}
We can also compare methods on the practical considerations derived from theoretical results as summarized in Section~\ref{sec:implications}:
\subsubsection{Accounting for uncertainty.} Most of the methods reviewed above include some mechanism for reflecting imputation model uncertainty. Bayesian or approximately Bayesian methods (including the approximate Bayesian bootstrap) do this naturally, whether part of a joint modeling or FCS imputation routine. Their behavior is not well understood in the FCS setting, however. Tree-based methods seem promising for some applications, but more work is required to find parameter settings and resampling strategies that make them reliably proper.
\subsubsection{ Include as many variables as possible.} Joint-sequential models may be easier to fit than FCS with many covariates, since all but one univariate model will include fewer than $p$ predictors. Simultaneous joint models somewhat lag behind sequential and FCS approaches here. This is particularly true with mixed data types and many fully observed covariates - most of these models are not easily adapted to condition on additional covariates, so fully observed variables must be included as additional variables in the joint model. Modeling fully observed variables instead of conditioning on them can waste ``degrees of freedom'' and lead to poorer model fit for the conditional distribution of the missing data. Carefully constructed models can help \citep{DeYoreo2016}, but seem to only go so far.
\subsubsection{ Use flexible imputation models.} Non- and semiparametric methods (Bayesian and otherwise, such as sequential tree-based methods) are flexible in their ability to capture certain unanticipated features of the data. Empirically these methods can outperform existing default MI procedures in simulations, particularly when the simulations are not built around simple parametric models themselves. More of these realistic evaluations are needed, as discussed in Sections~\ref{sec:empiricalcomp} and ~\ref{sec:conclusion}.
However, with flexible imputation models it can be challenging to manually adjust the imputation model to incorporate prior information or address model misfit. Incorporating meaningful prior information into nonparametric Bayesian imputation models is challenging but not impossible; see e.g. \cite{Schifeling2016-zs} for a strategy to include prior information in DP-MPMN models. While iterative imputation model refinement and assessment is ideal, it is not always possible. Empirical evidence suggests that flexible imputation models are much better as defaults than simple parametric models or PMM using linear models.
\subsection{Empirical comparisons between methods}\label{sec:empiricalcomp}
Empirical comparisons of several different imputation models on realistic datasets are relatively rare. Most papers introducing a new imputation model evaluate it using synthetic data generated from a researcher-specified multivariate probability model. The new imputation model is typically compared to a small number of competitors. These simulation studies can be informative -- for example, both \cite{Burgette2010-dq} and \cite{Doove2014-na} found evidence that imputations for continuous values generated via recursive partitioning can preserve interactions but underestimate main effects. However, models that are easy to simulate from and present in a paper will naturally be gross simplifications of the distribution of data in real populations. %
Simulations based on repeated sampling from realistic populations can be more informative. In these studies a population is compiled from existing data. Random samples are taken from these populations and values are ``blanked out'' via a known stochastic nonresponse mechanism. Each of the resulting incomplete datasets are multiply imputed and used to compute a range of estimates and confidence intervals, assessing the bias, coverage and efficiency of the MI estimates under the imputation model. Since the missing values are known, these can all be compared against the frequentist operating characteristics of the complete data procedure without appeal to asymptotic theory or other approximations. While the results are specific to a particular population and a set of estimands, this framework is much closer to reality than fully synthetic examples.
There are several recent examples of this kind of evaluation: \cite{Akande2017-gq} compared FCS with CART, the DP-MPMN model described in \ref{sec:npbayesimp},
and a default application of FCS with main effects multinomial logistic regression in a large repeated-sampling study of imputation using categorical data from the American Community Survey. The DP-MPMN imputations tended to yield better coverage than FCS-CART overall, but had much worse coverage for a small number of estimands. \cite{Manrique-VallierSMTrunc} also demonstrated the utility of accounting for structural zeros in this model with a population constructed from publicly available data from the U.S. Census. A default version of \cite{Murray2016-cy}'s joint model for mixed data types outperformed FCS using the default settings in R's \texttt{mice} package %
\citep{VanBuuren2011} in a large repeated-sampling study with data from the Survey of Income and Program Participation. Evidence suggested that misspecification bias was primarily to blame for FCS's poor performance. %
\subsection{Imputation model diagnostics}
A more obvious way to choose between imputation models is by fitting multiple and choosing the one that appears to fit the data best. Checking the fit of imputation models is challenging, but some approaches have been proposed. For methods that employ univariate regressions, imputers can examine standard diagnostics for those models \citep{Abayomi2008-if,su2011multiple}. \cite{Abayomi2008-if} suggested other diagnostic plots comparing imputed and observed values, primarily comparing marginal and bivariate distributions. Under MAR the distribution of missing values may be different than the distribution of observed values; \cite{Bondarenko2016-dj} used estimated response propensities to adjust for this and make diagnostic plots more comparable. \cite{He2012-ac} proposed posterior predictive checks, comparing the distribution of estimands computed on the multiply imputed datasets to the distribution of those estimands computed on entirely synthetic datasets generated by the imputation method (see also \cite{Nguyen2015-jy}). These checks require the imputer to choose relevant estimands and generate many samples from posterior predictive distributions, which can be computationally expensive.
\section{Conclusion}\label{sec:conclusion}
Over thirty years after Rubin's extensive treatment of MI \citep{Rubin1987}, experience with the method has cemented its reputation as a principled and practical solution to missing data problems. MI remains an active and fertile research area. While the behavior of the MI estimates have been the subject of intense scrutiny, relatively little is known about the comparative merits of various imputation models that have been proposed in recent years. Considerations based on theoretical findings suggest the use of more flexible imputation models where possible. Empirical evidence also suggests that simple defaults (MVN/log-linear models, or default FCS imputation using simple imputation models such as PMM with linear mean functions or regression models including only main effects) should be avoided, or at least carefully scrutinized.
Nonparametric Bayesian methods for generating imputations have recently emerged as a promising technique for generating imputations. In addition to new model development, more work is needed on scalable posterior computation with these models. In addition, the heuristic justification for why Bayesian MI ``tends to be proper'' is based on the asymptotic behavior of parametric Bayesian models \citep{Rubin1987}. It would be interesting to revisit this argument from the perspective of Bayesian nonparametric models, where the asymptotics are more involved (see \cite{Rousseau2016-dw} for a recent review). For example, can semiparametric Bernstein von-Mises results be derived for likely targets of MI inference under Bayesian nonparametric models used for imputation?
Joint-sequential approaches appear understudied and underutilized in the literature, perhaps because they currently require more intervention to set up. More research is needed on the implications of choosing different permutations of the variables in joint-sequential approaches. Further development of algorithmic approaches for selecting good joint-sequential variable orderings in the same vein as \cite{Li2014-lx} would also be welcome. There remains considerable work to be done in characterizing the behavior of FCS approaches to generating imputations; while some theoretical results exist, they are limited in scope and do not address some of the most effective variants of these algorithms (including PMM and CART).
More empirical comparisons of imputation methods and models are also needed. The field would benefit greatly from a repository of ready-to-use synthetic populations constructed from real data files. A common set of samples from these populations complete with missing values already generated would allow for easy comparisons across methods. A forward-thinking statistical agency could kickstart this repository, providing a public good (and possibly improving the state of their own missing data imputation routines) by sponsoring an imputation challenge in the spirit of a Kaggle competition.
The applications of MI have grown far beyond imputing item missing data in public use files: MI is used with synthetic data for disclosure limitation \citep{Rubin1993-pj,Reiter2002-lt, Raghunathan2003-yi}, to adjust for measurement error \citep{Cole2006-uf, Blackwell2015-eu}, and to perform statistical matching/data fusion \citep{Rassler2004-pb, Reiter2012-oa,Fosdick2016-fi}. In these new settings the amount of missing data can be much greater than typical applications of MI for item missing data, and imputation model development, selection, and assessment is even more consequential. We expect that new models and methods for multiple imputation will be an active research area for the foreseeable future.
\bibliographystyle{./imsart-nameyear}
|
\section{Introduction}
\label{sec:intro}
The rapid advance of mobile edge computing (MEC) has been the last mile of enabling a shared, low-latency computational environment for multi-vendor mobile edge applications. MEC performs computing offloading, data storage, caching and processing, request distribution and service delivery from the mobile edge to end users \cite{shi2016edgevision}. Applications with low latency tolerance, such as augmented reality (AR), video streaming, and online gaming, can deploy their services on the edge hosts at a cost, to achieve lower latency and better user experience \cite{ec-etsi-hu15}.
As the market gets mature, there will be multiple 5G service providers (SPs) provisioning MEC services to cover the same area: bigger wholesale players will invest in infrastructure to actually build mobile edge base stations, while there will also be mobile virtual network operators (MVNOs) renting resources from the former. These SPs can collaborate with each other in several ways for better utilization of the resources at the edge: virtual SPs have to place mobile edge (ME) applications on one of the rented edge hosts, preferably with lower cost, regardless of SPs. On the other hand, MEC base stations from different SPs can share resources with each other to process bursting requests.
\begin{figure*}[t]
\centering
\includegraphics[width=.9\textwidth]{intro.pdf}
\caption{\textbf{A MEC scenario in a certain service area. There are 3 ME base stations from 3 different SPs: SP-A, SP-B, and SP-C. They serve there own users within the service area. For resource sharing and optimization purpose, the base stations are also connected to each other.}}
\label{fig:intro}
\end{figure*}
For encouraging SPs to enroll their eligible MEC base stations and hosts in resource sharing, it is common to give incentives to SPs for contributing their resources of the hosts for hosting edge applications. Following the changing demand of end users, certain types of edge applications need to be deployed on, migrated to, or removed from an edge host, in order to meet the service requirement. By deploying the edge applications at the right places, the edge application provider will save costs, while providing high-quality service with low latency to the end users. Meanwhile, the edge host will collect incentives for its resources effectively used.
Clearly, the edge computing framework needs a placement service to dynamically check the user needs and the available edge hosts, and determine the placement or removal of edge applications. In datacenters, virtual machine (VM) placement has been well investigated, mainly with the focus of more efficient resource utilization and lower operational expense (OPEX). However, the collaboration of multiple SPs and mobile edge applications vendors are posing new challenges for ME application placement from the following aspects:
\begin{itemize}
\item A placement model has to make transparent and consistent selections of the best host for each request for edge computing resources. Moreover, the model has to take into consideration that a mobile edge application may require multiple services chained together at the edge.
\item A trusted party is required to determine the best place for application deployment. When an edge application is deployed on a mobile edge host, the application vendor needs to pay for the usage of the host. The placement algorithm has to avoid affiliation to either SP to ensure a neutral decision is made strictly according to the resource and the cost. It may create conflicts of interest to put any SP involved into the position of making placement decisions: placing mobile edge applications onto the SP's own hosts would bring revenue for renting their resources.
\item The application placement service needs to be steadily available. Both the mobile edge hosting service providers and the mobile edge application providers can constantly change. The placement service provider must remain in service regardless of the joining or quitting of vendors.
\end{itemize}
The challenges above urge a comprehensive solution uniting all SPs and their edge hosts without bias. In this paper, we present an architecture combined with its algorithm, namely \textsf{EdgeChain}, to create a decentralized placement service for mobile edge application that does not require trust to any party, i.e., trustless placement service. Compared with current placement solutions, \textsf{EdgeChain} has the following contributions:
\begin{itemize}
\item A cost model is presented as a stochastic programming problem, factoring in the pricing of edge hosts, latency, and service chaining.
\item We develop a heuristic placement algorithm based on the proposed cost model with the consideration of efficiency for running by the blockchain.
\item We introduce blockchain technologies to the MEC resource orchestration framework with two considerations: the first is to store the global resource availability, allocation, and consumption information that helps our algorithm make optimized decisions based on the global resource information. The second consideration is to have a decentralized public ledge for ensuring the neutrality of the placement decisions.
\item The \textsf{EdgeChain} framework is presented to run our algorithm for making placement decisions. In our design, SPs and mobile edge application vendors participate in the maintenance of the blockchain. An \textsf{EdgeChain} client is embedded in the network function virtualization (NFV) framework to determine the placement based on the existing information on the blockchain. To our best knowledge, this is the first work that leverages blockchain to coordinate SPs for MEC application placement.
\item Simulation results of our placement algorithm show its effectiveness in mobile edge host resource sharing among SPs. We also implement the \textsf{EdgeChain} by leveraging VeChain \cite{vechain}, an enterprise-level blockchain-as-a-service framework derived from Ethereum \cite{wood2014ethereum}.
\end{itemize}
We divide the contents into the following sections. The related work is illustrated in Section \ref{sec:relatedwork}. Section \ref{sec:problem} formulates the problem. Section \ref{sec:algorithm} proposes the heuristic \textsf{EdgeChain} placement algorithm based on the problem formulation. Then the simulation results are shown in Section \ref{sec:numerical}. Section \ref{sec:conclusion} concludes the paper.
\section{Related Work}
\label{sec:relatedwork}
The research directions in network service chaining (NSC) were discussed in \cite{john2013sfc-direction}. For security considerations, the authors highlighted the difficulty of bringing short-lived network services to targeted users in a single subscriber network by using the current security schemes. The potential security problems in SFC were stated in RFC7498 \cite{quinn2015problem}, including service overlay security, trusted classification policy, and secure SFC encapsulation. We investigated a placement problem in MEC with the consideration of application availability in \cite{zhu2017availability}.
Xiong et al. proposed a pricing strategy for offloading the blockchain's resource-consuming proof-of-work tasks to edge computing nodes \cite{xiong2017edge}. A two-stage Stackelberg game model was presented with both the edge computing service provider and the miners involved. A hierarchical distributed control system was built using Hyperledger Fabric blockchain \cite{stanciu2017control}. The hosting locations of cloud and fog of blockchain were compared in \cite{samaniego2017baas} for IoT networks with the conclusion that fog nodes were better as network latency was the dominant factor.
Nakamoto introduced the concept of blockchain and implemented Bitcoin \cite{nakamoto2008bitcoin}, a decentralized cryptocurrency that first resolved the double spending problem. Blockchains are based on Merkle trees \cite{Bayer1993} to efficiently allow multiple documents to be saved together in a block. As a decentralized public ledger, blockchains can serve beyond cryptocurrencies. Ethereum \cite{wood2014ethereum} used blockchain to store smart contracts that support building virtually any decentralized application.
\begin{table}[tb]
\centering
\caption{Parties involved in a MEC placement scenario}
\begin{tabularx}{.48\textwidth}{p{.07\textwidth}X}
\toprule %
\textbf{Party} & \textbf{Description} \\
\toprule %
\textit{Users} & Subscribers of applications and services over 5G networks with MEC enabled. \\
\textit{MECSPs} & MEC service providers, who deliver MEC hosting services that can run \textit{MEApps} at the network edge, close to end users. Examples include telecommunication companies like Rogers and Telus in Canada. \\
\textit{MEAVs} & Mobile edge application vendors, who provide \textit{MEApps} and services to end users. For instance, a company selling AR services. \\
\textit{MEApps} & MEApps stand for mobile edge applications provided by MEAVs. \\
\textit{MEHosts} & Servers that belong to different MECSPs to provide hosting service of \textit{MEApps}. \\
\textit{HostLinks} & Network links between hosts, regardless of which MECSP they belong to. \\
\textit{AppLinks} & When \textit{MEApps} are chained together, virtual links will be established for data transmissions traveling through the chain. \\
\bottomrule
\end{tabularx}
\label{table:parties}
\end{table}
\begin{table}[tb]
\centering
\caption{Notations Used in Problem Formulation}
\begin{tabularx}{.48\textwidth}{lX}
\toprule %
\textbf{Notation} & \textbf{Description} \\
\toprule %
$s, \mathbb{V}_s, \mathbb{L}_s, v, l$ & $s$ is a service chain. $\mathbb{V}_s$ is the set of all \textit{MEApps} in $s$. $\mathbb{L}_s$ is the set of all \textit{AppLinks} in $s$. A \textit{MEApp} in $s$ is denoted by $v \in \mathbb{V}_s$, and an \textit{AppLink} between two \textit{MEApps} in $s$ is denoted by $l \in \mathbb{L}_s$. \\
$\mathbb{H}, \mathbb{E}, h, e, \mathbb{V}_h$ & $\mathbb{H}$ is the set of all \textit{MEHosts}. $\mathbb{E}$ is the set of all \textit{HostLinks}. A \textit{MEHost} is denoted by $h \in \mathbb{H}$, and a link between two \textit{MEHosts} is denoted by $e \in \mathbb{E}$. $\mathbb{V}_h$ is the set of all \textit{MEApps} placed on $h$. \\
$u, m, c_s$ & $u$ is an end user. $m$ is a MECSP. $c_s$ is the cost of deploying $s$. \\
$c_v, c_{v h_{i}, v' h_{j}}$ & $c_v$ is the cost of deploying $v$. $c_{v h_{i}, v' h_{j}}$ is the cost of the \textit{AppLink} between $v$ on $h_i$ and $v'$ on $h_j$. \\
$n_s, P_m, h_m$ & $n_s$ is the total number of users requesting $s$. $P_m$ is a random variable denoting the percentage of the users of MECSP $m$. $h_m$ is an edge host of $m$. \\
$\gamma_m, \delta_m$ & $\gamma_m$ is the unit price of serving $m$'s own subscribers. $\delta_m$ is the extra charge for $m$ serving users of other MECSPs. \\
$C_v, M_v$ & CPU and memory requirement of the \textit{MEApp} $v$. \\
$B(e_{ij}), \zeta_{e_{ij}}$ & $B(e_{ij})$ is the total bandwidth capacity of \textit{HostLink} $e_{ij}$. $\zeta_{e_{ij}}$ is the unit price of the bandwidth of $e_{ij}$.\\
$B_V(e_{ij})$ & $B_V(e_{ij})$ is the total bandwidth used by \textit{MEApps} deployed on $h_i$ and $h_j$. \\
$B(v_{h_i}, v_{h_j})$ & Bandwidth used between \textit{MEHosts} $h_i$ and $h_j$. \\
$C_h, M_h$ & CPU and memory capacity of the \textit{MEHost} $h$. \\
$t_{e_{ij}}, t_s, T_s$ & $t_{e_{ij}}$ is the latency incurred on \textit{HostLink} $e_{ij}$. $t_s$ is the latency of the service chain $s$. $T_s$ is the max latency allowed by $s$. \\
\bottomrule
\end{tabularx}
\label{table:notations}
\end{table}
\section{Problem Formulation}
\label{sec:problem}
We first list all parties involved in a MEC placement scenario in Table \ref{table:parties}. The problem is formulated from a \textit{MEAV}'s point of view: \textit{MEApps} are direct consumers of the computing resources in the MEC environment, because a \textit{MEAV} needs to pay \textit{MECSPs} for hosting its applications in order to serve their \textit{users} and meet the latency requirement. Each \textit{MEApp} is equivalent to a virtual machine (VM) deployed on a \textit{MEHost}. \textit{MEApps} provided by different \textit{MEAV} can be combined as a service chain to provide comprehensive services. A service chain may span multiple \textit{MEAVs}. In this case, revenues generated by the service chain can be distributed according to the usage of each \textit{MEApp} on the service chain. For instance, a full-fledged AR service can load real-time navigation information from an online map application, while it can also load promotions of a shopping mall nearby from the mall's application. The navigation data is collected by the online map application, and the shopping mall application gets paid if the user "clicks" the links of the promotions.
The notations used in formulating the problem is shown in Table \ref{table:notations}. Define a chained service $s$ as a forwarding graph \cite{brown2015service} $G_s = (\mathbb{V}_s, \mathbb{L}_s)$, where $\mathbb{V}_s$ is the set of all \textit{MEApps} contributing to the service, and $\mathbb{L}_s$ is the set of all \textit{AppLinks} connecting applications together. A \textit{MEApp} is denoted by $v \in \mathbb{V}_s$, and an \textit{AppLink} between two \textit{MEApps} is denoted by $l \in \mathbb{L}_s$.
The chained service is deployed on a graph of connected \textit{MEHosts} $G_h = (\mathbb{H}, \mathbb{E})$, where $\mathbb{H}$ is the set of all \textit{MEHosts} owned by various \textit{MECSPs} and $\mathbb{E}$ is the set of all \textit{HostLinks}. A \textit{MEHost} is denoted by $h \in \mathbb{H}$, and a \textit{HostLink} between two \textit{MEHosts} is denoted by $e \in \mathbb{E}$. The \textit{HostLinks} can be either physical or virtual links with fixed capacities and latencies.
Suppose in a certain service area, there are $n_s$ \textit{users} from various \textit{MECSPs} requesting the same chained service $s$ from a \textit{MEAV}. We use $m$ to denote a \textit{MECSP} and $h_m$ for a \textit{MEHost} that belongs to $m$. Define an assigning function $x_{v h_m}$, whose value is $1$ if VM $v$ is assigned to Host $h_m$, 0 otherwise.
\begin{equation}
x_{v h_m} \triangleq \begin{cases}
1, & v \text{ is deployed on }h_m; \\
0, & \text{otherwise}.
\end{cases}
\end{equation}
Define a binary indicator of an \textit{AppLink} between two chained \textit{MEApps} in $s$, denoted by $L(v_{h_i}, v_{h_j})$, such that
\begin{equation}
\label{eqn:linkbin}
L(v, v') \triangleq \begin{cases}
1, & \text{$l\in \mathbb{L}_s$ exists between } v \text{ and } v'; \\
0, & \text{otherwise}.
\end{cases}
\end{equation}
Also, we use $e_{ij}$ to represent the \textit{HostLink} between $h_i$ and $h_j$. The cost of deploying $s$ is the sum of the cost of deploying each \textit{MEApp} $v$ of the service and the cost of the traffic between each two adjacent \textit{MEApps} in the service chain. It can be shown by
\begin{equation}
\begin{aligned}
c_s = & \sum_{h_m \in H} \sum_{v \in \mathbb{V}_s} c_{v h_m} x_{v h_m} \\
& + \sum_{h_i, h_j \in H}\sum_{v, v' \in \mathbb{V}_s} c_{v h_{i}, v' h_{j}} x_{v h_i} x_{v' h_j} L(v, v'),
\end{aligned}
\end{equation}
where $c_s$ represents the cost of deploying $s$ and $c_{v h_m}$ is for the cost of a \text{MEApp} $v$ deployed on a \text{MEHost} $h_m$. We assume that the pricing scheme for the same \textit{MECSP} is the same across all of its hosts. For a \textit{MEHost} $h_m$, define its basic unit resource price, which is the unit price of serving its own subscribers, as $\gamma_m$. When $h_m$ is serving users of other \textit{MECSPs}, it charges a premium of $\delta_m$ for its unit resource, as the return for doing courtesy for its partners. Therefore, the shared unit resource price of $h_m$ can be represented by $(\gamma_m + \delta_m)$. Define $C_{h_m}$ and $M_{h_m}$ to be the capacity of vCPU and memory provided by $h_m$. Define $C_v$ and $M_v$ as the vCPU and memory consumed by $v$. Define $P_m$ to be the random variable for percentage of the \textit{users} using the service chain $s$ via networks of the \textit{MECSP} $m$. Depending on the numbers of active users for each \textit{MECSP}, the total cost for the \textit{MEAV} to place its \textit{MEApp} $v$ onto a host of $m$ is the cost incurred by \textit{users} of $m$ plus the cost by \textit{users} of other \textit{MECSPs}:
\begin{equation}
\begin{aligned}
c_{v h_m} = & n_s (C_v + M_v) P_m \gamma_m \\
& + n_s (C_v + M_v) (1 - P_m) (\gamma_m + \delta_m) \\
= & n_s (C_v + M_v) \left[ P_m \gamma_m + (1 - P_m) (\gamma_m + \delta_m) \right] \\
= & n_s (C_v + M_v) \left[ \gamma_m + (1 - P_m) \delta_m \right].
\end{aligned}
\end{equation}
When a request from a \textit{user} for a service chain arrives, the blockchain would know the \textit{MECSP} from which the \textit{user} subscribes. For the same placement decision, the value $c_s$ can significantly differ over changing distribution of \textit{users}. An example can be two \textit{MECSPs} $m_1$ and $m_2$, each with one host $h_{m_1}$ and $h_{m_2}$. If all \textit{users} are subscribers of $m_1$ and all \textit{MEApps} are placed on \textit{MEHosts} of $m_1$, then the cost payable by the \textit{MEAV}s would be lower than if all \textit{users} were subscribers of $m_2$.
\begin{figure}[htb]
\centering
\includegraphics[width=.48\textwidth]{sfc-example.pdf}
\caption{\textbf{An example of ME service chains. There are 3 ME service chains sharing the services provided by 8 \textit{MEApps}.}}
\label{fig:sfc-example}
\end{figure}
\subsection{HostLink Unit Price}
When it comes to the cost modeling of a link between two \textit{MEHosts}, link availability is an important part for the service consistency of the \textit{MEApps}. If a heavily used \textit{HostLink} is down, consequences can be catastrophic: even if all individual \textit{MEApps} are running, the traffic would not be able to flow through between one or more pairs of \textit{MEApps} and the service chain would not be functional. For each \textit{HostLink} $e_{ij}$, there can be one or more \textit{AppLinks} sharing its bandwidth. \textit{HostLink} outages require migrating the \textit{MEApps} if they cannot be fixed in time. Therefore, \textit{HostLink} availability has significant influence on possible \textit{MEApp} migrations and the potential costs incurred.
The link unit price of a \textit{HostLink} $e_{ij}$, denoted by $\zeta_{e_{ij}}$, is then defined to describe how much to use the \text{HostLink} $e_{ij}$. The following two parameters will determine $\zeta_{e_{ij}}$.
The first parameter is $L(v_{h_i}, v_{h_j})$ as defined in Eqn. (\ref{eqn:linkbin}). The more \textit{AppLinks} a \textit{HostLink} carries, the more vital and expensive it becomes. The reason behind this ranking parameter is the potential consequence of migration: failure of a \textit{HostLink} used by many VMs would lead to massive migration of all \textit{MEApps} connected by that \textit{HostLink}, which would be more disruptive to the service chain.
The other parameter $B_V(e_{ij})$ is the total bandwidth consumed by traffic between \textit{MEApps} on the two hosts. It is selected because larger bandwidth usages would cause challenges at the time of migration: it can be hard to find another link with enough capacity.
\begin{equation}
B_V(e_{ij}) \triangleq \left[ \sum_{v_{h_i}, v_{h_j}, h_i \neq h_j} B(v_{h_i}, v_{h_j}) \right].
\end{equation}
Combining the two parameters, we define the unit price $\zeta_{e_{ij}}$ of a \textit{HostLink} $e_{ij}$, as the factor of the number of \textit{AppLinks} between two hosts times the factor of traffic flowing through these links:
\begin{equation}
\zeta_{e_{ij}} = \frac{\left[ \sum_{v_{h_i}, v_{h_j}, h_i \neq h_j} L(v_{h_i}, v_{h_j}) \right]}{N_{e_{ij}}} \frac{B_V(e_{ij})}{B(e_{ij})},
\end{equation}
where $N_{e_{ij}}$ is the maximum number of virtual links possible on $e_{ij}$. Therefore, $\zeta_{e_{ij}} \in [0, 1]$. The value of $\zeta_{e_{ij}}$ will rise to mark up a link's importance given it is either occupied by more pairs of VMs, or there is more traffic assigned to $e_{ij}$, or both. The cost of any two \textit{MEApps} is then the sum of the cost serving users that belong to the \textit{MECSPs} owning $h_i$ and $h_j$ and the cost serving other users timed by the price factor $\kappa_{e_{ij}}$:
\begin{equation}
\begin{aligned}
c_{v h_{i}, v' h_{j}} = & n_s \zeta_{e_{ij}} (P_{m_{h_i}} + P_{m_{h_j}})\kappa_{m_{h_i} m_{h_j}} \\
& + n_s \zeta_{e_{ij}} (1 - P_{m_{h_i}} - P_{m_{h_j}}) (\kappa_{m_{h_i} m_{h_j}} + \sigma_{m_{h_i} m_{h_j}}) \\
= & n_s \zeta_{e_{ij}} [ (P_{m_{h_i}} + P_{m_{h_j}}) \kappa_{m_{h_i} m_{h_j}} \\
& + (1 - P_{m_{h_i}} - P_{m_{h_j}}) (\kappa_{m_{h_i} m_{h_j}} + \sigma_{m_{h_i} m_{h_j}}) ].
\end{aligned}
\end{equation}
\subsection{HostLink latency}
Define the latency of the link $e_{ij}$ to be $t_{e_{ij}}$. For a service chain $s$, the total latency $t_s$ is then
\begin{equation}
t_s = \sum_{h_i, h_j \in \mathbb{H}} \sum_{v_{h_i}, v_{h_j} \in \mathbb{V}_s} L(v_{h_i}, v_{h_j}) x_{v h_i} x_{v h_j} t_{e_{ij}}.
\end{equation}
In the equation above, $t_{e_{ij}}$ is a constant depending on the particular $e_{ij}$. If $h_i = h_j$, then we consider the latency to be 0, since no actual \textit{HostLink} is used for data transmission between the two \textit{MEApps}. Define the maximum latency allowed for the service chain $s$ is $T_s$. Then there must be $t_s \leq T_s$ to meet the latency requirement.
\subsection{Stochastic Programming Formulation}
The problem is formulated as a stochastic programming optimization. Define $\mathbb{V}_h$ as the set of all \textit{MEApps} deployed on the \textit{MEHost} $h$. The objective is to minimize the total cost of the service chain $s$ to provide service with the lowest cost to the end user. As discussed in Section \ref{sec:problem}, the optimization is to minimize the costs on \textit{MEHosts} and \textit{HostLinks} for all \textit{MEApps} of $s$.
\noindent{\textit{Minimize}}
\begin{equation}
\label{eqn:objective}
\begin{aligned}
c_s = &\sum_{h_m \in \mathbb{H}} \sum_{v \in \mathbb{V}_s} c_{v h_m} x_{v h_m} \\
&+ \sum_{h_i, h_j \in \mathbb{H}}\sum_{v, v' \in \mathbb{V}_s} c_{v h_{i}, v' h_{j}} x_{v h_i} x_{v' h_j} L(v, v') \\
= &\sum_{h_m \in \mathbb{H}} \sum_{v \in \mathbb{V}_s} x_{v h_m} n_s (C_v + M_v) \left[ \gamma_m + (1 - P_m) \delta_m \right] \\
& + \sum_{h_i, h_j \in \mathbb{H}}\sum_{v, v' \in \mathbb{V}_s} n_s \zeta_{e_{ij}} [ (P_{m_{h_i}} + P_{m_{h_j}}) \kappa_{m_{h_i} m_{h_j}} \\
& + (1 - P_{m_{h_i}} - P_{m_{h_j}}) (\kappa_{m_{h_i} m_{h_j}} + \sigma_{m_{h_i} m_{h_j}}) ] L(v, v'),
\end{aligned}
\end{equation}
\begin{equation}
\label{eqn:constraint1}
\begin{aligned}
& \mathit{w.r.t.} \qquad x_{v h_m}, \\
& \mathit{s.t.} \qquad B(e_{ij}) \geq \sum_{v_{h_i}, v_{h_j}, h_i \neq h_j} B(v_{h_i}, v_{h_j}),
\end{aligned}
\end{equation}
\begin{equation}
\label{eqn:constraint2}
C_h \geq \sum_{v \in \mathbb{V}_h} C_v,
\end{equation}
\begin{equation}
\label{eqn:constraint3}
M_h \geq \sum_{v \in \mathbb{V}_h} M_v,
\end{equation}
\begin{equation}
\label{eqn:constraint4}
\sum_{h_i, h_j \in \mathbb{H}} \sum_{v_{h_i}, v_{h_j} \in \mathbb{V}_s} L(v_{h_i}, v_{h_j}) x_{v h_i} x_{v h_j} t_{e_{ij}} \leq T_s.
\end{equation}
\subsection*{\textbf{Remarks}}
\begin{itemize}
\item Function (\ref{eqn:objective}) is the objective function. It minimizes the cost of all \textit{MEApps} and \textit{AppLinks} by using less hosts, while not exhausting them.
\item Constraint (\ref{eqn:constraint1}) is the \textit{HostLink} bandwidth capacity bounds between each two hosts. Traffic transmitted between any two hosts $h_i$ and $h_j$ must not exceed the corresponding bandwidth capacity $B(e_{ij})$.
\item Constraints (\ref{eqn:constraint2}) and (\ref{eqn:constraint3}) are the CPU and memory capacity bounds for each \textit{MEHost}. The CPU and memory used by \textit{MEApps} coordinating with each other and by intra-host communications must not exceed $C_h$ and $M_h$.
\item Constraint (\ref{eqn:constraint4}) is the latency requirement of the service chain $s$ to ensure that the total latency of $s$ must not exceed the maximum latency allowed $T_s$.
\end{itemize}
\section{The \textsf{EdgeChain} Placement Algorithm}
\label{sec:algorithm}
The formulation presented in the previous section is a stochastic programming problem. Problems of this type been proved to be NP-hard \cite{gaivoronski2011knapsack}. It may not be computationally feasible when attempting to solve it in large scale. To apply our model to real-world scenarios, we design a heuristic algorithm called \textsf{EdgeChain} to achieve suboptimal results by applying a hybrid strategy of best-fit and first-fit decreasing algorithm. The pseudo code of the algorithm is shown in Algorithm \ref{algo:edgechain}.
\begin{algorithm}[htb]
\DontPrintSemicolon
\SetKwData{HostList}{host\_list}
\SetKwData{CurrMEApp}{app}
\SetKwData{RemainingCPU}{cpu\_left}
\SetKwData{RemainingMem}{mem\_left}
\SetKwData{Latency}{latency}
\SetKwData{MaxLatency}{max\_latency}
\SetKwData{And}{{\bf and}}\SetKwData{Or}{{\bf or}}
\KwData{\HostList: list of candidate MEHosts}
\KwData{\CurrMEApp: requested \textit{MEApp} to be placed, including its max latency allowed, stored in \Latency}
\KwData{\MaxLatency: max latency allowed for the service chain}
\KwResult{The best \textit{MEHost} in \HostList to place \CurrMEApp, or \textsf{none} if no valid host is found}
\Begin{
sort by percentage of \textit{users} of the service chain descending \;
\If{multiple \textit{MEHosts} found} {
sort \HostList by the locations of \CurrMEApp's last-hop \textit{MEApps} \;
\If{still multiple \textit{MEHosts} found} {
sort by the latency of the \textit{HostLinks} to the previous \textit{MEApps} in the service chain ascending \;
}
}
\For{$h \in \HostList$} {
$\Latency \leftarrow $ all latencies added together if \CurrMEApp placed on $h$ \;
\If{\Latency $\leq$ \MaxLatency} {
\RemainingCPU $\leftarrow$ calculate remaining vCPU by $C_h$ and $C_v$ of each \textit{MEApp} placed on $h$ \;
\RemainingMem $\leftarrow$ calculate remaining memory by $M_h$ and $M_v$ of each \textit{MEApp} placed on $h$ \;
\If{$\RemainingCPU \geq 0$ \And $\RemainingMem \geq 0$} {
\Return $h$\;
}
}
}
\Return \textsf{none}\;
}
\caption{\textsf{EdgeChain} Placement Algorithm}
\label{algo:edgechain}
\end{algorithm}
\subsection{Processing Order and selection of MEHosts}
The \textsf{EdgePlace} algorithm runs on each mining node based on the Ethereum platform. The algorithm retrieves its input information from the blockchain, as all transactions and updates are recorded on the blockchain. The \textsf{EdgePlace} algorithm will select the \textit{MEHosts} following the steps below.
\subsubsection{Users}
Sort all \textit{MEHosts} by the percentage of \textit{users} of the service chain. For each \textit{MEApp} on the service chain, consider which \textit{MECSP} has most users using it. Then \textit{MEHosts} with the same \textit{MECSP} will have higher ranks to deploy this \textit{MEApp}. Since all \textit{MEHosts} of the same \textit{MECSP} have the same unit resource cost, the \textit{MEApp} can be placed on any of the \textit{MEHosts} that belongs to the best \textit{MECSP}, to avoid the situation that too many \textit{MEApps} are concentrated on one \textit{MEHost}.
\subsubsection{Last-hop MEApp}
For \textit{MEHosts} given higher priority in the previous step, sort by the locations of last-hop \textit{MEApps}. \textit{MEHosts} hosting the previous-hop \textit{MEApps} will be considered first. This step is to reduce the traffic cost between different \textit{MECSPs}.
\subsubsection{Latency}
For \textit{MEHosts} given higher priority in the previous step, sort by the latency of the \textit{HostLinks} to the previous \textit{MEApps} in the service chain. \textit{MEHosts} with lower latency will be considered first.
After the list of candidate \textit{MEHosts} are sorted according to the steps above, the algorithm iterates the list and pick the first valid \textit{MEHost} that has enough resources to place the \textit{MEApp}, as well as meeting the latency requirement of the service chain.
\section{EdgeChain Design and Implementation}
In this section, we introduce the design and implementation of \textsf{EdgeChain}, a blockchain-based system that integrates with the existing MEC architecture for \textit{MECSPs} and the scheduler of \textit{MEAV}. There are mainly two reasons the blockchain is used in the system:
\begin{itemize}
\item The blockchain acts as a public ledger that stores all useful information and transactions made during the placement process. Exposure of the information would help the placement algorithm make optimized decisions considering the global resource demand and allocation. The blockchain enables such centralized resource information, in a decentralized implementation.
\item As a public ledger applying proof-of-work verifications, the blockchain makes it nearly impossible to tamper the history stored in the blockchain. The \textsf{EdgeChain} algorithm will be downloaded by all mining nodes and they will execute the same algorithm with the same input. The placement result will only be accepted by the system if majority of the mining nodes reach agreement on the output. This will ensure the neutrality of the placement decisions.
\end{itemize}
The system takes requests to place \textit{MEApps} from \textit{MEAVs}, and the placement algorithm runs as the smart contract on the blockchain to select the best \textit{MEHost} from all candidates. The NFV orchestrator of the related \textit{MECSP} receives and enforces the placement decision, while posting the transaction onto the blockchain for recording. While this paper is written, the blockchain is implemented based on VeChain \cite{vechain}, an enterprise-level blockchain-as-a-service framework derived from Ethereum.
\begin{figure}[htb]
\centering
\includegraphics[width=.48\textwidth]{data-structure.pdf}
\caption{\textbf{Data entities and their relationship used by and stored in EdgeChain, including \textit{MECSPs}, \textit{MEHosts}, \textit{HostLinks}, \textit{SvcChains}, \textit{MEApps}, and \textit{AppLinks}.}}
\label{fig:data-structure}
\end{figure}
\subsection{Data Entities}
As Fig. \ref{fig:data-structure} shows, there exist 6 types of data entities on the blockchain and they are related to each other to represent the status of \textit{MEHosts} and placement decision of running \textit{MEApps}. The descriptions of these data entities are illustrated below. Each data entity record has a unique Ethereum address for other to locate it on the blockchain. All types of data entities can be created, updated and deleted, while the blockchain will keep the audit trail of every change.
\subsubsection{MECSP}
When a \textit{MECSP} record is registered to \textsf{\textsf{EdgeChain}}, a record of this \textit{MECSP} is added with the Ethereum addresses pointing to the records of all its eligible \textit{MEHosts} and \textit{HostLinks}. A \textit{MECSP} record is updated whenever there is change to any \textit{MEHost} or \textit{HostLink}.
\subsubsection{MEHost}
A \textit{MEHost} record registers under an existing \textit{MECSP} to the blockchain. In a record, the vCPU and memory capabilities can be found, along with the Ethereum addresses pointing to the records of all \textit{MEApps} placed onto it.
\subsubsection{HostLink}
Similar to \textit{MEHosts}, a \textit{HostLink} is under a registered \textit{MECSP}, which contains the two \textit{MEHosts} it connects, and the bandwidth of the \textit{HostLink}.
\subsubsection{SvcChain}
A service chain is registered by a \textit{user} to the blockchain to reflect the resource consumption of a chained service, including that from \textit{MEApps} and the corresponding \textit{AppLinks}. The service chain can have \textit{MEApps} from multiple \textit{MEAVs}.
\subsubsection{MEApp}
A \textit{MEAV} will submit a record of a \textit{MEApp} whenever it needs to spin up one. A record stores the vCPU, memory usage of the \textit{MEApp}.
\subsubsection{AppLink}
\textit{AppLinks} describe chained relationship between two \textit{MEApps}. The source and destination \textit{MEApps} are stored in an \textit{AppLink} record, as well as network bandwidth requirement of this link.
\begin{figure}[htb]
\centering
\includegraphics[width=.48\textwidth]{workflow.pdf}
\caption{\textbf{Typical work flow of EdgeChain. \textit{MECSPs}, \textit{MEAVs}, and mining nodes participate in the process. Steps of the work flow are marked by circled numbers and alphabets with details documented in Section \ref{sec:workflow}.}}
\label{fig:workflow}
\end{figure}
\subsection{EdgeChain Work Flow}
\label{sec:workflow}
A typical \textsf{EdgeChain} work flow can be demonstrated by Fig. \ref{fig:workflow}, where there are three parties participating in the entire process: \textit{MECSPs}, \textit{MEAVs}, and mining nodes. We use circled numbers and alphabets to define the work flow in sequence.
\circled{1} A \textit{user} requests a service chain from the blockchain. Such requests will be sent to the blockchain every time a user requests a service chain.
\circled{2} The request for the service chain is recorded. When the request is synced to the mining nodes, it will be broken into requests for \textit{MEApps}. The mining nodes will run the logic to break down the service chain creation request. Then the requests for \textit{MEApps} are propagated to all corresponding \textit{MEAVs}.
\circled{3} Based on its user demand, the \textit{MEApp} Scheduler decides to create a new instance of \text{MEApp} and pass the request to the Ethereum client of the \textit{MEAV}.
\circled{4} The Ethereum client running the \textsf{EdgeChain} service sends the request to the blockchain, creating records for the request of placing a new \textit{MEApp}.
\circled{5} The request of creating a new \textit{MEApp} arrives at a MECSP through its Ethereum client.
\circled{6} For every \textit{MECSP}, the Ethereum client requests the NFV Orchestrator (NFVO) to call the \textsf{EdgeChain} placement algorithm downloaded to the resource manager for the decision of the placement. This will ensure that the placement algorithm be executed by different parties for verifying the results. The placement result returned by the next step will only be accepted if majority of the parties return the same placement result.
\circled{7} The NFVO calls the \textsf{EdgeChain} placement algorithm for the placement decision. Note that the decision can be a hash representing any \textit{MEHost} within the entire MEC network. If the result points to a \textit{MEHost} which does not belong to the current \textit{MECSP}, then no actual placement will be done. Instead, only the result along with the algorithm's hash will be returned to the Ethereum client for verification.
\circled{8} If the result points to a \textit{MEHost} of the current \textit{MECSP}, then the NFVO will sends the request to place the \textit{MEApp} to the VNF Manager (VNFM). Also, a transaction shown in Fig. \ref{fig:state} will be posted to the blockchain to record that placement actually occurs.
\circled{9} The VNFM sends the request to the NFV Infrastructure (NFVI) deploy the \textit{MEAPP} onto the target \textit{MEHost}.
\circled{T} The mining nodes periodically perform the mining process to verify the blockchain, as well as earning Ethers for requesting placement services. Meanwhile, the resource manager periodically synchronizes with the NFVI for the up-to-date resource usage and availability, and then posts the updated information to the blockchain.
\begin{figure}[htb]
\centering
\includegraphics[width=.48\textwidth]{state.pdf}
\caption{\textbf{A placement transaction in EdgeChain. A state transition happens upon a transaction. As this figure shows, \textit{MEApp-4} is to be placed with the requirement of 2 vCPUs and 4096 MB of memory. The input of the EdgeChain placement algorithm is the current state of the two \textit{MEHosts}. The result is to place \textit{MEApp-4} onto \textit{MEHost-2}. After the transaction is accepted, the resources taken by \textit{MEApp-4} are deducted from the remaining resources of \textit{MEHost-2}.}}
\label{fig:state}
\end{figure}
\section {Numerical Results}
\label{sec:numerical}
In this section, we illustrate the numerical results of the MEC placement cost changes based on varying mobile edge application user cases using CloudSim \cite{cloudsim}. To clearly demonstrate the focused trends, the following assumptions are made to simplify the modeling of the problem without losing generality. We first discuss the placement results output by the \textsf{EdgeChain} algorithm for the same service chain on the same set of \textit{MEHosts}.
\begin{enumerate}
\item The unit costs of the CPU and memory of all hosts for the same \textit{MECSP} are the same.
\item Costs of network bandwidth for all links follow the same unit price.
\item One mobile edge application includes the same type of VMs with the same CPU, memory and network bandwidth requirements.
\item A request from the user will be processed by one VM, while the VM may communicate with other VMs to exchange information.
\end{enumerate}
\subsection{Parameters}
With the assumptions above, we choose parameters for our placement model to evaluate the performance and the facts under different circumstances. First, we choose a MEC service scenario of 3 \textit{MECSPs} $m_1$, $m_2$, and $m_3$, each with 3 \textit{MEHosts}, where $h_1$, $h_2$, $h_3$ belong to $m_1$, $h_4$, $h_5$, $h_6$ belong to $m_2$, and $h_7$, $h_8$, $h_9$ belong to $m_3$.
Three identical requested service chain, each with 5 \textit{MEApps} is to be placed. The \textit{MEApps} of each service chain are denoted by $v_1$, $v_2$, $v_3$, $v_4$, and $v_5$. The service chain starts from $v_1$ and ends at $v_5$: $v_1 \rightarrow v_2 \rightarrow v_3 \rightarrow v_4 \rightarrow v_5$. We assume that all \textit{MEApps} have the same CPU, memory and bandwidth requirements, which are shown in Table \ref{table:params}, along with other parameters.
\begin{table}[htb]
\caption{Parameters for the MEC scenario}
\begin{tabularx}{.48\textwidth}{XX|XX}
\toprule %
\textbf{Parameter} & \textbf{Value} & \textbf{Parameter} & \textbf{Value} \\
\toprule %
$C_v$ & $2$ vCPUs & $M_v$ & $2048$ MB \\
$C_h$ & $64$ vCPUs & $M_h$ & $65536$ MB \\
$\gamma_{m_1}$ & 1.0 & $\delta_{m_1}$ & 0.2 \\
$\kappa_{m_1}$ & 1.0 & $\sigma_{m_1}$ & 0.2 \\
$\gamma_{m_2}$ & 0.8 & $\delta_{m_2}$ & 0.5 \\
$\kappa_{m_2}$ & 0.8 & $\sigma_{m_2}$ & 0.5 \\
$\gamma_{m_3}$ & 1.2 & $\delta_{m_3}$ & 0.3 \\
$\kappa_{m_3}$ & 1.2 & $\sigma_{m_3}$ & 0.3 \\
$n_s$ & 100 users & $P_m$ & \textit{var} \\
$B(e_{ij})$ & 10000 Mbps & $B(v, v')$ & 30 Mbps \\
$t_{e_{ij}}$ & 15 ms & $T_s$ & 50 ms \\
\bottomrule
\end{tabularx}
\label{table:params}
\end{table}
\begin{figure}[t!]
\centering
\begin{subfigure}
\centering
\includegraphics[width=2.5in, trim={1in 2.58in 1in 2.58in}]{place1.pdf}
{\\(a) $P_{m_1} = 0.5$ and $P_{m_2} = P_{m_3} = 0.25$.}
\end{subfigure}
~
\begin{subfigure}
\centering
\includegraphics[width=2.5in, trim={1in 2.58in 1in 2.58in}]{place2.pdf}
{\\(b) $P_{m_1} = P_{m_2} =0.25$ and $P_{m_3} = 0.5$.}
\end{subfigure}
\caption{\textbf{Placement results of 3 service chains consisting of 15 \textit{MEApps} in all. The value of $\delta_{m_1}$ changes from 0.1 to 0.6. Figure (a) shows the placement decision when $P_{m_1} = 0.5$ and $P_{m_2} = P_{m_3} = 0.25$. In comparison, Figure (b) shows the placement decision when $P_{m_1} = P_{m_1} = 0.25$ and $P_{m_3} = 0.5$.}}
\label{fig:place}
\end{figure}
\subsection{Placement trends with changing unit resource premium}
The placement decision changes by the increase of $\delta_{m_1}$ under different user distributions are shown in Fig. \ref{fig:place}, where $\delta_{m_1}$, the unit resource premium payable to the \textit{MECSP} for hosting \textit{MEApps} for others, increases from 0.1 to 0.6. For comparison, in Fig. \ref{fig:place}(a), most users are from $m_1$. There is $P_{m_1} = 0.5$ and $P_{m_2} = P_{m_3} = 0.25$. Meanwhile, in Fig. \ref{fig:place}(b), most users subscribe services from $m_3$ as $P_{m_1} = P_{m_1} = 0.25$ and $P_{m_3} = 0.5$.
From the results of the two scenarios, we learn that the \textit{MEHosts} with lower combination of unit resource base price ($\gamma_m$) and unit resource premiums ($\delta_m$) will be selected first. The \textit{MEHosts} of the \textit{MECSP} will have more weight upon consideration if there are more \textit{users} from that \textit{MECSP}.
\begin{figure}[tb]
\centering
\includegraphics[width=.48\textwidth, trim={0.5in 4in 0.5in 4in}]{user_dist.pdf}
\caption{\textbf{Numbers of \textit{MEApps} placed on the 3 \textit{MEHost} with different percentages of \textit{users} in the network. Users of $m_1$ increase from 0\% to 100\%, while those of $m_2$ decrease from 100\% to 0\%. There is no user for $m_3$.}}
\label{fig:user-dist}
\end{figure}
\subsection{Placement trends with changing user distribution}
To further demonstrate the impact from the distribution of the \textit{users}, we simulate various scenarios with different percentages of \textit{users} for $m_1$ and $m_2$, while there is no user for $m_3$. Users of $m_1$ increase from 0\% to 100\%, while those of $m_2$ decrease from 100\% to 0\%.
The results have shown the trends of \textit{MEApps} migrating to \textit{MEHosts} owned by the \textit{MECSP} that has more active users to avoid premiums charged by other \textit{MECSPs}. However, resource sharing still takes place ($m_3$ hosting \textit{MEApps} for $m_1$ and $m_2$) when needed for better latency results and service quality.
\section{Conclusions}
\label{sec:conclusion}
In this paper, we have presented the architecture and the algorithms for mobile edge applications placement for multiple mobile edge computing service providers, leveraging the blockchain-based system called \textsf{EdgeChain}. Future work will be considering multiple service chains initiated by multiple users, to achieve lower overall costs for the entire system.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The recent impressive advances of machine learning and in particular deep learning methods in solving several challenging tasks, came along with a vulnerability that questions the security of several applications. In fact, as shown in ~\citet{Szegedy2013}, small and visually imperceptible perturbations of images, called \emph{adversarial examples}, can lead to highly confident misclassification. Moreover, these
adversarial examples can generalize across different networks ~\citep{papernot2016practical, Universal_adv}, which gives the \emph{adversary} the ability to fool artificial systems without any prior knowledge on their architecture and makes models deployed for real-world applications an easy target of adversarial attacks.
A growing body of work is now devoted to designing adversarial attacks, quantifying the robustness of classifiers and building effective defense mechanisms ~\citep{Fawzi2015, Kurakin2016, papernot2016distillation, DeepFool, Ensemble_adv}. A standard approach known as \emph{adversarial training}, suggested in \citet{Goodfellow2015_adv}, consists in augmenting the training set with its adversarial version. More recent methods \citep{mixup} try to encourage linear behavior between training examples, thus building some robustness to adversarial examples.
In this paper, we suggest a new approach to building some resistance to adversarial attacks by training a discriminator to distinguish latent representations of real samples from the ones of adversarial examples. The discriminator gets intermediate features of a classifier as input. At the same time, the classifier is trained not only to correctly classify the training data but also to fool the discriminator when fed an adversarial example. This enforces invariance of the hidden representation with respect to the nature of the input i.e whether it is a real sample or an adversarial examples. The classifier ends up filtering out the adversarial \emph{noise} at the level of the hidden representation, leading the adversarial example to be classified as its corresponding real sample.
We perform experiments on the MNIST datasets to validate the proof of concept. Our results suggest that our procedure compares well to standard adversarial training, and that combining the two methods may further improve robustness.
\section{Related work}
The idea of augmenting a network with a discriminator to enforce hidden representation invariance is at the heart of the work \citet{Ganin2016domainadaptation} on domain adaptation, where the network learns features that adapt to different domains for the same task.
This idea has been exploited in several recent works.
\citet{metzen2017detecting} uses this kind of architecture in the context of {\it detection}: the discriminator, trained separately from the classifier, is merely used as detector of adversarial attacks.
By contrast, the role of the discriminator in our approach is to teach the classifier to build robust features, in a procedure similar in spirit to the GAN framework ~\citep{Goodfellow2014GAN}.
Fader nets \citep{lample2017fader} use an adversarial loss for an encoder to extract the invariant patterns that are orthogonal to some image attributes. This helped disentangling these attributes factors and allowed controlling them in a generative model setting.
\citet{nuisances} takes the view from information theory to a regularizing term added to the classifier loss, which filters out the information of a given class of `nuisances', i.e variations of the inputs that are irrelevant to the task. This regularizer takes the form of mutual information, which can be estimated by means of a discriminator. Our work can be viewed as a generalization of this proposal to the case of adversarial perturbations.
\section{Approach}
\subsection{Adversarial perturbations}
Consider a classifier $C_\theta \colon \mathcal{X} \rightarrow \mathcal{Y}$
trained by minimizing the average of a given loss function $\mathcal{L}_\theta(x, y)$ over labeled examples $(x, y)$ in a training dataset.
An \emph{adversary} willing to fool the classifier may design input perturbations $\bar{x} = x+\delta(x)$ so as to maximize the loss:
\begin{equation} \label{adv}
\delta(x) = \argmax_{\delta \in \mathcal{C}} \mathcal{L}_\theta(x+\delta, y)
\end{equation}
where $\mathcal{C}$ encodes a set of constraints controlling the amplitude of the perturbations -- e.g small enough to remain quasi-imperceptible to a human observer. As shown in ~\citet{Szegedy2013}, deep neural networks are highly unstable under such adversarial perturbations.
The simplest methods to generating perturbations defined in (\ref{adv}) are first order approaches, which linearly expand the loss
$\mathcal{L}_\theta(x+\delta, y) \simeq \mathcal{L}_\theta(x, y) + \langle \delta, \nabla_x \mathcal{L}_\theta(x, y)\rangle$. If the constraints $\mathcal{C}$ take the form $\|\delta\|_{\infty} \leq \epsilon$ for some $\epsilon >0$,
Equ. (\ref{adv}) yields the so-called fast gradient method for generating adversarial examples ~\citep{Goodfellow2015_adv}:
\begin{equation} \label{sign_grad}
x^{adv} = x + \epsilon \mbox{sign} \left(\nabla_x \mathcal{L}_ \theta(x, y)\right)
\end{equation}
Other generation methods have been proposed, which include ~\citet{Kurakin2016, DeepFool}. In our preliminary experiments, we use the procedure (\ref{sign_grad}) to generate adversarial examples {\it on-the-fly} during training.
\subsection{Augmented network}
Assuming the classifier has $L$ layers, we choose a layer $\ell$, which effectively splits the network into an encoder part $E$ (layers $1, \cdots \ell$) and a residual classification part $R$ (layers $\ell, \ell+1, \cdots L$). We supplement the network with a discriminator $D$ branched off from $\ell$. If $\ell$ has $k$ units, $D \colon \mathbb{R}^k \rightarrow [0,1]$ takes as input the feature at $\ell$ and produces an output interpreted as the probability of the feature coming from a real (as opposed to adversarial) input.
Given an input image $x \in \mathbb{R}^d$ we thus have the following notation:
\begin{itemize}
\item $z = E(x)$ is the feature vector at layer $\ell$
\item $R(z) = R(E(x)) = C_\theta(x)$ is the output of the original classifier
\item $D(z)$ is the output of the discriminator, and the probability that $z$ is the representation of a sample of the real data.
\end{itemize}
The architecture is shown in Fig \ref{fig:model_}.
\begin{figure}[h]
\centering
\includegraphics[width=9cm]{model_.pdf}
\caption{Model architecture}
\label{fig:model_}
\end{figure}
\subsection{Adversarially Augmented Training}
At each iteration, we forward-pass a mini-batch of real data $(x, y)$, then generate and forward pass the corresponding mini-batch of adversarial examples $(x^{adv}, y)$.
Classifier and discriminator are trained simultaneously.
The discriminator is trained as a binary classifier where $z$ has tag $t=1$
when $z = E(x)$ and $t=0$ when $z=E(x^{adv})$. The classifier is trained to classify each real input correctly (which involves both encoder and residual) and to fool the discriminator (which involves only the encoder). Training the classifier on the discriminator's response should enforce an invariance across real samples and their adversarial versions at the level of the latent representation $z$.
This can be combined with standard adversarial training for the correct classification of each adversarial input.
Let $\theta_{enc}, \theta_{res}$ and $\theta_{disc}$ the parameters of the encoder, the residual classifier and the discriminator respectively, and $\theta=(\theta_{enc}, \theta_{res})$. We use the following losses:
{\bf Classification Loss}. For each labeled sample $(x, y)$, the classifier outputs class probabilities $P_\theta(\cdot |x)$. We use the cross-entropy loss for the classification task.
This loss can be extended to include classical adversarial training \citep{Goodfellow2015_adv}:
$$ \mathcal{L}_{\theta}(x,x^{adv}, y) = - \alpha \log P_\theta(y|x) - (1-\alpha)\log P_\theta(y|x^{adv}),$$
where $\alpha \in [0, 1]$ is a convex mixing parameter (we used $\alpha = 0.5$)
{\bf Discriminator Loss}. For each feature $z$ with $t$, the discriminator outputs the probability $D(z)=P_{\theta_{disc}}(t=1|z)$. We use binary cross-entropy to train the discriminator to distinguish the real features from the adversarial ones:
$$
\mathcal{L}_{\theta_{disc}}(z, t) = - \log P_{\theta_{disc}}(t|z)
$$
{\bf Encoder Adversarial Loss}. We use the cross-entropy loss to train (the encoder part of) the classifier on the discriminator response, so as to make the adversarial representation be classified as real by the discriminator:
$$
\mathcal{L}^{adv}_{\theta_{enc}}(x^{adv}) = -\beta \log D(E(x^{adv})),
$$
where $\beta > 0$ is a factor controlling the importance of the adversarial loss i.e how much we enforce invariance (we used $\beta = 1$)
\section{Experiments -- Proof of concept}
{\bf Setup.}
The reported experiments are based on a 3-layer feed-forward network using leaky ReLU as activations. The number of hidden units of its hidden layers are respectively 512, 256, 128. The discriminator is fed the output of the second hidden layer. The adversarial examples were generated using $\epsilon = 0.25$.
The discriminator is a simple shallow network with 128 hidden units, a ReLU activation and a 0.5-dropout.
{\bf Training.}
At each iteration, we update the discriminator, and the classifier with adversarially augmented loss. One can try multiple updates of the discriminator for one update of the classifier as suggested in \citet{Goodfellow2014GAN}, but we noticed that a single update gave already the expected training behavior.
Figure \ref{fig:model_curves} gives the classification accuracies reached by the classifier and the discriminator for real and adversarial examples in the training and validation dataset, as a function of the training epochs.
\begin{figure}[h]
\centering
\includegraphics[width=13cm]{final_wkp_curves.pdf}
\caption{Mnist Adversarially Augmented Training}
\label{fig:model_curves}
\end{figure}
Table \ref{tab:mnist_test} gives the test performance of the same model trained with 4 different approaches: simple training of the classifier, standard adversarial training (SAT), adversarially augmented training (A2T) and finally the adversarially augmented adversarial training (A3T).
\begin{table}[h]
\centering
\begin{tabular}{ r || c | c }
& Accuracy on real & Accuracy on adversarial \\ \hline
Simple training & 98.18\% & 25.37\% \\
AT ($\alpha=\frac12, \beta=0$) & 98.89\% & 94.45\% \\
A2T ($\alpha=\beta=1$) \hspace{3mm} & 97.77\% & 89.74\%\\
A3T ($\alpha=\frac12, \beta=1$) & 98.72\% & {\bf 96.10\%} \\
\hline
\end{tabular}
\vspace{5mm}
\caption{Test accuracies for MNIST on real and adversarial samples ($\epsilon=0.1$, 10 runs)}
\label{tab:mnist_test}
\end{table}
\vspace*{-0.7cm}
\section{Discussion}
We proposed a new approach to enforce invariance of a classifier's features with respect to adversarial perturbations. We proved the efficacy of the method with some proof of concept experiments.
More work needs to be done to test the sensitivity of the model to the hyperparameters and the adversarial generation method. A crucial point will also be to test its scalability to larger datasets and architectures.
The existence of adversarial examples suggests that neural networks, despite their impressive generalization abilities in perceptual tasks, do not learn the true underlying concepts defining the correct labels. One could hope that a procedure seeking feature invariance would push the network to not only ignore adversarial perturbations but also to focus more on relevant patterns in real data. We conjecture that our approach is one step in that direction.
\bibliographystyle{apalike}
|
\section{Introduction}
The standard model (SM) of particle physics, while enjoying tremendous
success as an accurate description of nature, has many parameters
whose values look mysterious from a theoretical perspective. Why are
the Higgs mass and the energy scale of the cosmological constant so
small compared to the Planck scale? Why is $\theta_{QCD}$ so small?
What is the origin of the hierarchy of fermion masses? Such questions
have inspired many efforts to go beyond the standard model. Following
the discovery of the Higgs boson, there is a new item, dubbed ``Higgs
near-criticality,'' on the list: why is the Higgs self-coupling
$\lambda$ (in conjunction with the top quark Yukawa coupling $y_t$)
so close to the critical value beyond which the Higgs potential
becomes unstable at high scales? The situation is illustrated in
fig.\ \ref{fig:stability} \cite{STAB}, which shows the
regions of stability, metastability and instability of our vacuum, in
the $\lambda$-$y_t$ plane, with the small ellipse of the measured
values falling in the narrow region of metastability. In the
metastability (instability) region the vacuum lifetime is longer
(shorter) than the age of the Universe.
\begin{figure}[b]
\hspace{-0.4cm}
\centerline{
\includegraphics[width=0.92\hsize]{HNCcoup}}
\caption{Regions of the $\lambda$-$y_t$ plane leading to
stability, metastability or instability of the Higgs potential at high
scales (at NNLO accuracy \cite{STAB}). In the region labeled ``Nonperturbativity'' $\lambda$ becomes strong below the Planck scale. The couplings are defined at the electroweak scale.}
\label{fig:stability}
\end{figure}
The answer could of course be that it is a coincidence: for fixed
$y_t$, the quartic coupling is $0.01$ below the stability boundary
($0.03$ above the instability line),
which is a tuning of only 8\% (23\%) relative to its actual value. On the
other hand if $\lambda$ could {\it a priori} have taken any value between
zero and $4\pi$, this becomes a tuning of $0.08\%$ ($0.2\%$), more in accord with
the visual impression from fig.~\ref{fig:stability}. This is
predicated on the assumption that there is no new physics coupled to the Higgs at high scales (up to the Planck scale) that might shift the stability boundaries relative to where they are shown. Nevertheless since there is an
anthropic reason for $\lambda$ to avoid the instability region, it
is tempting to construct a scenario where this explains the
coincidence.
While anthropic reasoning is eschewed by many physicists, if
there is a landscape of vacuum states in which anthropically sensitive
parameters are sampled, it seems difficult to dismiss. For example
the very large number of flux compactifications in string theory
\cite{Douglas:2003um,Ashok:2003gk} make it plausible that our universe
is part of a much larger multiverse \cite{LindeEternal}. A solution of the cosmological
constant ($\Lambda$) problem was proposed in which $\Lambda$ is
finely scanned by these flux vacua \cite{BP}, yielding
values consistent with anthropic bounds \cite{Weinberg:1987dv}.
Coleman's wormhole mechanism \cite{Coleman:1988tj} is another example
of a multiverse in which the most likely value of $\Lambda$ is small
(in fact vanishing).
In this context, Rubakov and Shaposhnikov argued \cite{Rubakov:1989pn}
that the observed values of physical constants might generically be
close to the boundaries of the anthropically allowed regions. If the
probability distribution is such that the most likely value
of a parameter is anthropically forbidden, then the most likely
observed value would be close to the boundary, since there are no
observers on the forbidden side. The near-criticality of the Higgs
potential looks like a possible example of this phenomenon.
The anthropic necessity of Higgs stability is an old observation
that was used to put a lower bound on the Higgs mass (or an upper
bound on the heaviest quark mass) as early
as 1979 \cite{Stab0}. Improved predictions using
higher orders in the loop expansion were subsequently made
\cite{STAB,StabMore}. An indication of
how delicate the tuning is for near-criticality is provided by
the comparison of such predictions at different levels of precision \cite{ELattice}:
at LO our vacuum would be deep in the instability region, at NLO
in the middle of the metastable one and at NNLO very close to
the stability boundary.
Of particular relevance for our work, the implications of Higgs
stability within a landscape of vacua with $\lambda$ scanning were
studied in ref.\ \cite{Hall}, assuming conditions just like those
suggested by ref.\ \cite{Rubakov:1989pn} for the underlying
probability distribution $P(\lambda)$, namely that it is maximized
in the unstable region of small $\lambda$. In that work, a
model-independent analysis was done, where no particular model of
the landscape was proposed; rather a reasonable functional form for
$P(\lambda)$ was assumed, which led to predictions for the Higgs mass prior
to its measurement.
We think it is worthwhile to revisit the question of Higgs
near-criticality within a specific model of the landscape, since
such a study may reveal nontrivial challenges to the overall
consistency of such a picture, that may be shared by other
possible examples. At the same time we introduce a new kind of
landscape that is particularly simple and amenable to calculations,
namely the vacuum states provided by the minima of the potential of an axion field (whose detailed properties are very
different from those of the QCD axion).
We are inspired by a string-theory-motivated construction, axion
monodromy, previously used for inflation
\cite{AxMono} and by the relaxion mechanism
\cite{relaxion}, used for solving the weak scale hierarchy problem. In
contrast to these applications however, we wish to avoid classical
evolution of the axion $a$ during cosmological evolution. Instead,
the universe is assumed to split into causally disconnected domains
where $a$ sits in different local minima of its potential. These
vacuum states were populated by quantum fluctuations of $a$ during a period of inflation, are essentially stable against tunneling once formed, and
so realize a tractable example of a multiverse. The probability distribution
is calculable in terms of the axion potential, given certain
assumptions about the cosmological evolution that we will specify.
\section{Landscape for $\bma{\lambda}$}
The field $a$ has a potential of the form
\begin{eqnarray}
V(a) &=& \overline V(a)
- \Lambda_b^4\cos(a/f) \ ,
\label{Va}
\end{eqnarray}
where $\overline V$ denotes the part of the potential that can be approximated as non-oscillatory on a field range large compared to $2\pi f$.
As the field $a$ has an axionic origin ($a$ is a pseudo-Goldstone
boson, like a phase field), it originally enjoys a shift
symmetry $a\rightarrow a +c$ that is broken by the potential (\ref{Va}).
The term $\Lambda_b^4\cos(a/f) $ breaks the shift symmetry down to a discrete subgroup $a\rightarrow a +2\pi f$, while $\overline V(a)$ breaks the
shift symmetry completely (at least in the range we consider; see below). It is then natural to expect these breaking terms to be
much smaller than the typical mass scale or cutoff of the theory that
we will call
$\Lambda$. In a string theoretic UV completion, $\Lambda$ could be
the string scale.
We assume then $\Lambda_b \lesssim \Lambda$ and take
$\overline V(a) = \Lambda^4\ V(\eta a /\Lambda)$, with $\eta\ll 1$.
For our purposes
it will suffice to keep the linear term of this function:
\begin{equation}
\overline V(a) = \eta \Lambda^3 a +\dots
\label{smoothV}
\end{equation}
This linear term should accurately describe
the potential $\overline V(a)$
in a typical field region. Without loss of generality (by doing
a shift in the field), we can
take this typical region to be in the vicinity of $a=0$ and
we can also take $\eta>0$ so that $\overline V(a)$ is a growing function of $a$.
A concrete example for $\overline V$ that arises in certain string theory
compactifications \cite{AxMono} is
\begin{equation}
\overline V(a) = M^4\sqrt{1 + a^2/F^2}\simeq M^4 a/ F,
\label{monod}
\end{equation}
where the linear approximation is valid in the region where
$a\gg F$. Here, $M$ and $F$ are generically at the string scale, but
if the axion arises from a warped throat, then $M$ can be
parametrically suppressed by a warp factor, which may be
exponentially small.
Another example is the clockwork axion \cite{clockwork}, with $\overline V(a)=\epsilon \Lambda^4 \cos(a/F)$, and $F=N f \gg f$,
which hierarchy can be arranged in a natural way. In this setting, the field range is compact, $2\pi F$, but we are interested in a patch $\Delta a$ with $2\pi f \ll \Delta a \ll 2\pi F$, and there
we can expand $\overline V(a)$ as in (\ref{smoothV}) around some
typical value $a_0$, obtaining $\eta = -(\epsilon \Lambda/F) \sin(a_0/F)$.
Let the minima of the potential (\ref{Va}) be labeled by an integer $n$,
such that $a_n \simeq 2\pi n f$.
A basic condition for having a landscape is that
$\overline V$ must be sufficiently flat
so that it does not destroy the local minima of the oscillatory
part. This requires
\begin{equation}
\overline V{}'(a) =
\eta\,\Lambda^3
\lesssim {\Lambda_b^4\over f}\ ,
\label{cond1}
\end{equation}
which, if satisfied, would naively imply infinitely many local minima.
In realistic string constructions however, there is back-reaction from
large windings, so that the actual number of minima is limited to
$N \lesssim 10-100$, beyond which the above description breaks down, and possibly an extra dimension decompactifies \cite{Liam}.
In clock-work constructions the number of vacua is also finite as the field range is compact.
We assume that in addition to $\overline V$, there is a coupling
of $a$ to the Higgs potential:
\begin{equation}
V_h = \Big(-\mu_h^2 + c_h \eta\, a \Lambda \Big) |H|^2 +
\left(\lambda+ c_\lambda\, \frac{\eta \, a}{\Lambda}\right) |H|^4\ .\label{Vah}
\end{equation}
Such couplings also break the shift symmetry and so we assign a factor $\eta$ to them.
The $a$-terms in (\ref{Vah}) could be regarded as arising from a generalization of eq.\ (\ref{monod})
by taking $M^4 \to M^4 + {\cal O}(\Lambda^2|H|^2, |H|^4)$ or
from expanded $\cos(a/F)$ potentials in the clockwork realization.
In the landscape of vacua of the $a$ field,
where $\langle a\rangle =
a_n \simeq 2\pi n f$, this shifts the bare values ({\it i.e.,} the values at the UV scale $\Lambda$) of the Higgs
parameters to
\begin{eqnarray}
\label{scanning}
\mu_n^2 &=& \mu_h^2 - n\, c_h \eta\,(2\pi f)\Lambda\,,\nonumber\\
\lambda_n &=& \lambda + n\, c_\lambda\eta \frac{2\pi
f}{\Lambda}
\equiv \lambda + n\,\delta\lambda\, .
\label{dleq}
\end{eqnarray}
Here we assume that some other mechanism solves the
weak scale hierarchy problem ({\it e.g.} a relaxion mechanism \cite{relaxion}) so that $\mu_n$ is of electroweak size and focus on the shift in the Higgs coupling. For reasons detailed below we also choose $c_\lambda>0$. Likewise we must assume there is another
mechanism for solving the cosmological constant problem, since the
vacuum energy varies between $a$-vacua due to the nonperiodic part of the potential $\overline V$.
We consider three possible scenarios, each associated to one of the
three critical boundaries shown in fig.\ \ref{fig:zoom}; these are
the boundaries of instability and metastability at zero temperature,
and the boundary of high-temperature instability that depends upon the
assumed reheating temperature (dashed lines). Our mechanism explains
why we would observe $(\lambda,y_t)$ to be near (and to the right of) one of these boundaries.
The characteristics of the
three categories are summarized in table \ref{tab}.
Fig.~\ref{fig:zoom} shows trajectories of successive vacua that
exemplify each case. Which one of the three is actually realized
depends upon cosmological parameters, as we will discuss
in more detail in the next section.
{\renewcommand{\arraystretch}{1.5}
\begin{table}
\centering
\small
\begin{tabular}{| c | c | c | c |}
\hline
& (1) & (2) & (3) \\ \hline \hline
Boundary & $T=0$ Instablity & $T_R$ Instability & Stability \\ \hline
Vacuum & Quantum & Thermal & Inflationary \vspace*{-3pt}\\
Selection & $T=0$ decay & decay & decay \\ \hline
$\delta \lambda$ & $\sim 0.05$ & $\sim 0.02$ & $\ll 0.01$ \\ \hline
$M_t$/GeV & $173.34\pm 2.28$ & $ 173.34^{+1.34}_{-2.28}$ & $\simeq 171$
\\ \hline
\end{tabular}
\caption{\it Characteristics of the three cases we consider in the text,
{regarding} the critical boundary, vacuum selection mechanism, step in $\lambda$ needed and range of top mass (inside the experimental $3\sigma$ band) required.}
\label{tab}
\end{table}
}
In case (1) we end close to the instability boundary and the probability to live in vacua beyond that boundary is depleted by $T=0$ decay,
in which the Higgs vacuum has a lifetime that is shorter than the age of the Universe. To explain why a
point lying in the experimentally allowed ellipse at $y_t\simeq 0.95$
corresponds to the most probable anthropically allowed vacuum, we need $\delta\lambda\sim 0.05$, the approximate width of the metastable
region.\footnote{This number can be estimated as follows. The
vacuum decay rate per unit volume is $\Gamma\sim h_t^4
e^{-8\pi^2/(3|\lambda(h_t)|)}$, where $h_t$ is the preferred value for tunneling. The decay probability is $\Gamma$ times the 4D volume of our past light-cone
$\sim (e^{140}/m_P)^4$. Decay probabilities of order one require
$\lambda(h_t)\sim -0.05$ and this number is confirmed by a more
sophisticated calculation (see {\it e.g.} \cite{EEGIRS}). Thus the
metastable region is approximately $\lambda(h_t)\in\{-0.05,0\}$. This
translates to the region shown in Fig.\ \ref{fig:stability} after running the
couplings down to the weak scale.}
Scenario (1) could take place for any value of the top mass, within the
experimentally preferred region, which we take to be the 3$\sigma$
range $M_t=173.34\pm 3\times 0.67$ GeV \cite{top}.
In case (2) we end in a vacuum near the instability boundary for decay by thermal fluctuations with a high reheating temperature,
that reduce the region of metastability. As concrete examples
we illustrate the cases of $T_R=10^{14}$
and $T_R=10^{16}$ GeV. The boundary of the reduced region is
shown as the dashed lines in Fig.~\ref{fig:zoom} (see refs.\
\cite{Tdec,Higgstory}), and a possible trajectory illustrating this
case is shown along $y_t\simeq 0.934$. A smaller step size
$\delta\lambda \sim 0.02$ is suggested for naturally explaining the
distance of the SM point from the dashed boundary. This mechanism,
for such large $T_{\rm R}$ favors the lower range of the top mass,
with $M_t\simeq 173.34^{+1.34}_{-2.28}$ GeV.
In case (3) we end very close to the stability boundary beyond which
the Higgs vacuum is unstable against decay during inflation, for
sufficiently large values of ${\cal H}_I\sqrt{N_e}$. This case is
illustrated by the trajectory passing through the bottom of the
experimental ellipse. Here the most probable state would be
the one closest to the boundary in the absolute stability region, and
it would require a very small step size $\delta\lambda$ to be
naturally close to the experimental ellipse. Although this
possibility is currently disfavored, it is not excluded and provides
another possible regime for explaining near-critical
stability, if the top mass is very close to its lowest
$3\sigma$ value, $M_t\simeq 171$ GeV.
\begin{figure}[t]
\hspace{-0.4cm}
\centerline{
\includegraphics[width=0.92\hsize]{VacuaTracks}}
\caption{Zoom-in of Fig.~\ref{fig:stability} showing also the instability lines for thermal vacuum decay with $T_R=10^{14-16}$ GeV (red dashed lines). Trajectories of $a$-vacua are shown
(surviving ones in black, doomed ones in white) for the three cases discussed in the text. We use $M_h=125.09\pm 0.24$ GeV \cite{mh} and $M_t=173.34\pm 0.67$ GeV \cite{top} at 1-$\sigma$ for the experimental ellipses.}
\label{fig:zoom}
\end{figure}
Once $\delta\lambda$ is fixed, (\ref{dleq}) can be used to
to eliminate the unknown parameter $\eta$ in terms of $f$ and
$\delta\lambda$. We introduce the ratio $\dh \equiv
\delta\lambda/0.05$ [which is of order unity in case (1)] to allow for the
possibility of any of the three cases. Hence
\begin{equation}
c_\lambda \eta = 0.05\,\delta_\lambda\, \frac{\Lambda}{2\pi f}\ .
\label{eta}
\end{equation}
\section{Probability distribution of vacua}
\label{prob_sec}
A key ingredient of our scenario is the process by which
the vacua get populated by quantum fluctuations during inflation, and
the resulting probability distribution function $P(t,a_n)$ for the different
vacua. It is governed by the Fokker-Planck
equation
\begin{equation}
{\partial P\over dt} = {\partial\over\partial a}
\left[{V'(a)\over 3 {\cal H}_I}P + {{\cal H}_I^3\over 8\pi^2}{\partial P\over\partial a}
\right]\ ,
\label{FPeq}
\end{equation}
(see for example refs.\ \cite{Linde,EGR,Zurek}) where ${\cal H}_I$ is the Hubble
parameter during inflation. We take the inflationary
contribution to the energy density to be much larger than $V(a)$ and
consider ${\cal H}_I$ to be approximately constant.
Then the
stationary solution to (\ref{FPeq}) is\footnote{If $a$ contributes significantly
to the energy density, the stationary solution is
$P(a) \sim \exp \{ 24\pi^2 m_P^4/[V_I+V(a)]\}$,
where $V_I$ is the inflaton field potential and $m_P$ the
reduced Planck mass.
An expansion for small $V(a)/V_I$ reproduces Eq.~(\ref{Peq}).}
\begin{equation}
P(a) \sim e^{ -8\pi^2 V(a)/3 {\cal H}_I^4}\ .
\label{Peq}
\end{equation}
We assume for the moment that this
stationary solution (\ref{Peq}) is reached and determines the relative probabilities of the different vacua (disregarding for now the possible decays along the Higgs direction). The necessary
conditions to justify this assumption will be discussed below. We do not care about the
normalization of $P(a)$ as we are only interested in relative probabilities between different vacua.
At the local minima of the potential we have $V(a_n)\simeq \overline V(a_n)$,
neglecting the uninteresting constant contribution $-\Lambda_b^4$ and
taking $v\ll \Lambda$, where $v=246$ GeV is the Higgs vacuum
expectation value, with $v^2/2=\langle |H|^2\rangle$. With our
convention $ \eta>0$, the underlying landscape probability
distribution prefers the more negative values of $n$, which reduce
$\overline V(a_n)$. By choosing $c_\lambda\eta>0$ we then favor
negative $\lambda_n$ in eq.\ (\ref{scanning}) and unstable Higgs
potentials are preferred within the landscape.
In order to have significant variation of $P(a_n)$
near the instability boundary, the exponent of (\ref{Peq}) should
change by $O(1)$ between neighboring vacua. The ratio of the
probabilities of the second and first anthropically allowed vacua, relative to
the anthropic boundary, is given by
\begin{equation}
-\ln {P_2\over P_1} = {8\pi^2\Delta\overline V\over 3
{\cal H}_I^4} \ \stackrel{>}{{}_\sim}\ {\cal O}(1) \ ,
\label{dPeq}
\end{equation}
where
\begin{equation}
\Delta \overline V = \eta (2\pi f) \Lambda^3 = 0.05\,
\Lambda_r^4\ .
\label{dveq}
\end{equation}
In the last step we removed $\eta$ by using (\ref{eta})
and introduce the quantity $\Lambda_r$ (that appears repeatedly) as
\begin{equation}
\Lambda_r\equiv \left({\dh}/{c_\lambda}\right)^{1/4}\Lambda\ .
\end{equation}
Condition (\ref{dPeq}) will lead to the most likely anthropically allowed vacuum being
the one closest to the critical boundary in question. It
imposes a maximum value of the Hubble rate during inflation:
${\cal H}_I^4 \lesssim (8\pi^2\Delta\overline V/ 3)$.
On the other hand, the derivation of the Fokker-Planck equation
from the stochastic approach to tunneling \cite{Linde}
assumes that ${\cal H}_I > m_a$, the mass of the $a$
field. It is possible that this is only a sufficient and not
a necessary condition \cite{Linde2}, but if we respect it
[along with (\ref{dPeq})] then
${\cal H}_I$ should be in the interval\footnote{Here we account for
the displacement away from the minimum of the cosine potential due to
the linear term, using $V'=0$ to eliminate $\cos(a/f)$ in $m_a^2 =
V''$, and (\ref{dveq}) to reexpress $\Delta \overline V$.}
\begin{equation}
m_a=\frac{\Lambda_b^2}{f}\left[1-\left(0.3\frac{\Lambda_r}{\Lambda_b}\right)^8\right]^{1/4} \lesssim {\cal H}_I
\lesssim 1.07\,\Lambda_r\ .
\label{HIrange}
\end{equation}
The upper limit is plotted in Figure~\ref{fig:bounds} with the
label ``$H_I$ range.''\ \
Information on the lower limit, which varies from point to point
in the plane, is
conveyed by the dashed lines; {\it e.g.,} on the line labeled
``$m_a/\Lambda_r=0.25$,'' the interval for ${\cal H}_I/\Lambda_r$ is
(0.25,\,1.07). On the other hand, Eq.\ (\ref{cond1}), required to
guarantee the existence of a landscape of $a$-vacua (which coincides with
the requirement $m_a>0$), gives the limit
\begin{equation}
\Lambda_b > 0.3\Lambda_r\, ,
\label{Lbrmin}
\end{equation}
which is also plotted in Figure~\ref{fig:bounds} and labeled
``Landscape.''
If we also insist that the inflaton potential dominates
over the $a$ potential, then ${\cal
H}^2_I\gtrsim 2\pi\eta\Lambda^3 N f/3 m_P^2$, where we have assumed
that $a = 2\pi N f$ in the vicinity of our $a$-vacuum.
Using (\ref{dveq}) to eliminate $\eta f$
and combining with the upper limit
in (\ref{HIrange}) we find
\begin{equation}
\label{Lambda_lim}
\frac{\Lambda_r}{m_P}\stackrel{<}{{}_\sim} \frac{8.4}{\sqrt{N}} \ ,
\end{equation}
which is not very constraining ({\it e.g.} if $N\lesssim 100$ or {$\Lambda_r\ll \Lambda$}).
\begin{figure}[t]
\hspace{-0.4cm}
\centerline{
\includegraphics[width=0.92\hsize]{ParamSpace}}
\caption{Excluded (shaded) regions in the
plane of $\Lambda_b/\Lambda_r$
versus $f/\Lambda_r$. ``Landscape'' region violates
condition (\ref{cond1}); ``Long Inflation'' region violates
condition (\ref{noeternal}); ``Stability'' denotes the vacuum
stability bound (\ref{T0rate}); ``$H_I$ range'' curve
denotes the limit beyond which the interval of allowed
inflationary Hubble
rates ${\cal H}_I$ from (\ref{HIrange}) vanishes.
The dashed diagonal lines indicate the lowest allowed value of
${\cal H}_I/\Lambda_r$ in this allowed range
as the axion mass is varied (see text). The bounds corresponding to
``$T_R < \Lambda_b$'' depend upon $\Lambda_r/m_P$ and are shown for
two values of that ratio.
}
\label{fig:bounds}
\end{figure}
\section{Vacuum stability}
\label{vac_stab}
For our own $a$-vacuum to be habitable, it must not
decay too quickly through tunneling to neighboring axionic vacua
(not to be confused with the possible {decay along the Higgs direction}).
This might occur during inflation, after reheating, when
the effect of finite temperature is important, or at late times
when we can consider $T$ to be zero.
At zero temperature, the criterion for vacuum stability becomes
\begin{equation}
A e^{-S_4} \lesssim {\cal H}_0^4
\label{T0rate}
\end{equation}
where ${\cal H}_0$ is the present Hubble constant ($\sim e^{-140}m_{P}$ in Planckian units). $S_4$ is the 4D Euclidean action for critical bubbles corresponding to
transitions between neighboring vacua \cite{Coleman:1977py}.
In (\ref{T0rate}), the prefactor $A = (S_4/2\pi)^2 J$, with $J$ being a ratio of functional
determinants with dimensions of [mass]$^4$.
The $J$ factor is
difficult to compute, but is expected to be of order $\Lambda_b^4$
or $f^4$, always smaller than $\Lambda^4$ and $m_{P}^4$,
so it is conservative to require $S_4\gtrsim 560$ as a condition for vacuum stability.
We numerically compute the bounce solution and resulting $S_4$ and
plot this stability condition, labeled ``Stability,'' in Figure~\ref{fig:bounds}.
An analytic formulation of the stability criterion can be obtained
using the thin-wall approximation \cite{Coleman:1977py}, in which the
4D action is
\begin{equation}
S_{4,tw} \simeq {27\pi^2\over 2}\, {\sigma^4\over \Delta V^3}
\ ,
\label{S4eq}
\end{equation}
depending upon the bubble wall tension
\begin{equation}
\sigma \simeq \int_{0}^{2\pi f} da\,
\sqrt{2 \Lambda_b^4[1-\cos(a/f)]} = 8 \Lambda_b^2 f\ ,
\end{equation}
and the potential difference between neighboring vacua as
given by (\ref{dveq}). By numerical calculation of the actual
tunneling action, we find that this approximation is not very good
in the region of parameter space of interest; however by comparing the
exact and approximate results it is possible to correct for this.
The relevant parameter determining how well the thin-wall
approximation works is $\Lambda_b/\Lambda_r$,\footnote{{By the rescalings $\hat a = a/f$ and $x =
r \Lambda_b^2/f$, we can write
$S_4 = 2\pi^2 (f/\Lambda)^4\int dx\,x^3\left[\sfrac12 \hat a'^2 +
(0.3\Lambda_r/\Lambda_b)^4\hat a - \cos \hat a\right]$, using
(\ref{dveq}). The thin-wall approximation breaks down as the
coefficient of the linear term becomes large.}}\ \ and we find
that the fractional error in the action can be accurately fit to the formula
\begin{equation}
1-{S_4\over S_{4,tw}} \simeq 7.1\times
10^{-5}\left(\Lambda_b\over\Lambda_r\right)^{-7.845}
\label{twfit}
\end{equation}
where $S_4$ is the full numerical value. This function is shown in
fig.\ \ref{fig:tw}.
\begin{figure}[t]
\hspace{-0.4cm}
\centerline{
\includegraphics[width=0.92\hsize]{tw-error}}
\caption{Fractional error in the thin-wall approximation
for the 4D tunneling action, as a function of $\Lambda_b/\Lambda_r$.
The analytic fit (\ref{twfit}) as well as the numerical results are
shown.}
\label{fig:tw}
\end{figure}
In the case of vacuum transitions due to thermal excitation over the
barrier, one should estimate the 3D action for critical bubbles,
taking also into account the thermal corrections to the potential.
This is not a straightforward task: it depends on possible couplings
of $a$ to other sectors of the theory and is limited to temperatures
well below the critical temperature $T_c$ above which the dynamics
responsible for the nonperturbative generation of the barriers in the
axion potential become ineffective, but this is unspecified in our scenario.
If the reheating temperature $T_R$ is above $T_c$ one expects the
effective temperature-dependent barrier height
$\Lambda_b^4(T)$ to start falling as a power of $T$ \cite{Preskill:1982cy}.
Given the level of uncertainty on
$T_c$, we content ourselves with imposing the condition that $T_R <
T_c\sim \Lambda_b$, as a
rough estimate for $T_c$.
To obtain $T_R$ we use the relation for the Hubble parameter
during radiation domination ${\cal H}_R = 0.33\,\sqrt{g_*}\,
T_R^2/m_P$. Assuming instant reheating we have
${\cal H}_R ={\cal H}_I$ with ${\cal H}_I$ respecting
(\ref{HIrange}), which translates into the range
\begin{equation}
0.54 \frac{\sqrt{m_a m_P}}{\Lambda_r}
<\frac{T_R}{\Lambda_r} \left(\frac{g_*}{g_*^{\rm SM}}\right)^{1/4}< 0.56 \sqrt{\frac{m_P}{\Lambda_r}}\ ,
\end{equation}
with $g_*^{\rm SM}=106.75$. We exclude a point in parameter space if the lower limit of this range is bigger than $\Lambda_b/\Lambda_r$.
The resulting limit is shown in Fig.~\ref{fig:bounds}, labeled $T_R<\Lambda_b$,
for two representative values of $\Lambda_r/m_P =0.5,\,1$.
In cases (1) and (2) we must also consider the possibility of vacuum
decay along the Higgs field direction, since we end up in the
metastable region with respect to such decays. Metastability here
means that quantum fluctuations at zero temperature are slow on the
time scale ${\cal H}_0^{-1}$, and it does not take into account the
possibility that tunneling was triggered at an earlier time by
inflation. In fact during inflation, if $\cal{H}_I$ is higher than
the instability scale, the Higgs field can be pushed over the barrier
that separates the electroweak vacuum from the unstable region of
field space \cite{EGR,Zurek,Higgstory}, and this leads to an
upper bound on ${\cal H}_I\sqrt{N_e}$, where $N_e$ is the number of
e-folds. As discussed in the next section, this kind of bound
can be generically violated in our framework if a very long period of
inflation is needed to guarantee that the stationary solution to the
Fokker-Planck equation is reached. In fact, this is the vacuum
selection mechanism in case (3).
For cases (1) and (2) we then have to forbid such decays during inflation. A simple way of circumventing this danger is to have a nonminimal coupling $\xi |H|^2R$ between the
Higgs field and the Ricci scalar $R$ \cite{EGR}. During inflation,
$R=-12 {\cal H}_I^2$, and this provides a contribution
$12 \xi {\cal H}_I^2$ to the squared Higgs mass, that stabilizes the potential or suppresses Higgs fluctuations altogether (for $\xi>3/16$), relaxing the
bound on ${\cal H}_I\sqrt{N_e}$ \cite{Higgstory}. Subsequent to
inflation, during preheating the induced Higgs mass term oscillates
along with the inflaton, and this can cause parametric resonant
production of Higgses, whose associated classical field can probe the
instability region again \cite{preheatdec,Dani} and trigger vacuum
decay. To avoid this, it is sufficient to have $\xi$ in the range
$(0.06-4)$ \cite{Dani}, which we assume to be the case for scenarios (1) and (2).
\section{Initial conditions}
We have assumed that the stationary solution of the Fokker-Planck
equation was achieved during inflation. Here we consider
how long a period of inflation would be required to achieve this, starting from some
different initial condition, for example that $P(a)$ was peaked around
the true vacuum state. The barriers between neighboring vacua
must be large enough to prevent tunneling at late times,
while the scale of
inflation must be sufficiently low so that
$P(a)$ is not too flat, eq.\ (\ref{dPeq}). Both of these tend to
slow the
time evolution of $P$.
It is instructive to consider a toy model consisting of a double-well
potential $V(\phi)$ with just two vacuum states, separated by a
barrier height $V_b$ that is large compared to the energy difference
between the two vacua. The system is initially
sharply
localized in one of the vacua, $\phi_1$, and allowed to evolve in
time according to the
Fokker-Planck equation. By a combination of numerical
and analytical methods one discovers two relevant time scales,
hierarchically different. The shorter one, $\tau_1\simeq 3{\cal
H}_I/[2V''(\phi_1)]$, is associated with the
spread of $P$ until it reaches an approximately Gaussian shape around
the starting vacuum, $P(\phi)\simeq
\exp[-(\phi-\phi_1)^2/(2\sigma_1^2)]$, with $\sigma_1^2=3 {\cal
H}_I^4/[8\pi^2V''(\phi_1)]$. This solution is valid for small displacements and
is quasi-stationary. The long time scale, $\tau_t$, is associated with
the probability leakage to the second vacuum at $\phi_2$, through the
top of the barrier, at $\phi_t$. The associated rate,
$\Gamma=1/\tau_t$, is
\begin{equation}
\Gamma \sim {{\cal H}_I^3\over 16 \pi^2 \sigma_1 \sigma_t}\,
e^{-8\pi^2 V_b/3 {\cal H}_I^4}\ ,
\end{equation}
where $\sigma_t^2 = 3 {\cal H}_I^4/(8\pi^2 |V''(\phi_t)|)$.
Applying this estimate to our scenario, we see that
to avoid an exponentially long period of inflation, one
needs ${\cal H}_I^4 \gtrsim 8\pi^2 V_b/3$, while
condition (\ref{dPeq})
implies ${\cal H}_I^4 \lesssim 8\pi^2\Delta\overline V/3$. Using
$V_b =\Lambda_b^4$ and $\Delta\overline V$ from (\ref{dveq}),
the combined conditions
require
\begin{equation}
\Lambda_b/\Lambda_r< 0.47\ .
\label{noeternal}
\end{equation}
Hence it is possible to
satisfy all the criteria without having a very long period of inflation.
However, a more generic situation is to admit a prior period of
eternal inflation, which would automatically justify the stationary
solution since then an arbitrarily long period of evolution
could occur prior to the final stage of observable inflation. Two common situations
can admit eternal inflation. First, inflation could be chaotic during
the primordial stage, with the inflaton displaced high enough on its
potential so that upward quantum fluctuations can dominate over the
classical downhill evolution \cite{LindeEternal}. Second, the inflaton (not necessarily
the same inflaton that is responsible for the final stage of inflation) could be
trapped in a false vacuum with an exponentially long lifetime, the
exponential of the tunneling action
\cite{LinVil}. Either case allows us to relax
the requirement (\ref{noeternal}).
\section{Summary and conclusions}
We have presented a concrete realization of a mechanism to explain the
near-criticality of the SM Higgs quartic coupling $\lambda$. It uses
an axion-like field $a$ with a potential that develops a large number
of non-degenerate vacua in which $\lambda$ takes different values,
effectively scanning, due to a coupling of the Higgs to $a$. The vacua
are assumed to be populated during inflation with probabilities that
depend exponentially on the ratio $V(a)/{\cal H}_I^4$. By
appropriately choosing the sign of the overall slope of $V(a)$, vacua
with increasingly negative values of $\lambda$ are favored. The
conditional probability for a particular vacuum state given that it is
compatible with observers, is zero if it undergoes catastrophic
decay of the Higgs vacuum. Thus the most likely anthropically allowed
states are those that are close to a critical line in the plane of
$\lambda$ and $y_t$.
We discussed three different scenarios,
summarized in Table~1 and illustrated by Fig.~\ref{fig:zoom}.
They require different cosmological histories and parameters
for the potential of the $a$ field,
and they depend upon the precise value of the top quark mass.
In case (1), vacua beyond the instability line are depleted by quantum
tunneling, which is faster than the age of the universe. In case (2),
that requires a large reheating temperature, thermal fluctuations over
the Higgs barrier remove vacua beyond the thermal instability line.
In case (3), which requires a high inflationary Hubble rate or a
large number of e-folds, Higgs fluctuations induced during inflation
trigger vacuum decay along the unstable Higgs direction, effectively
selecting vacua with stable Higgs potentials.
While the mechanism we have discussed offers an explanation for the
intriguing near-criticality of the Higgs quartic coupling, it does not
address the hierarchy problem. It would be quite interesting to find
a mechanism that could address both issues simultaneously, especially
given the fact that similar mechanisms ({\it e.g.} relaxions) offer
potential solutions to the hierarchy problem.
It is perhaps disappointing that this scenario does not make
positive predictions for new physics at experimentally accessible
energies.
Since the only new field, the axion,
has a mass \{typically much larger than the electroweak scale, there are no manifestations at low
energy. Instead, we predict an {\it absence} of new physics coupling to the
Higgs field at low scales, to the extent that such couplings would
move the critical lines of stability away from their standard model
values. On the other hand, we think it is interesting that despite
the lack of low-energy experimental tests, the mechanism is highly
constrained by considerations of theoretical and cosmological
consistency. It shows that the mere existence of a landscape is not
sufficient for a successful anthropic explanation of tuning problems.
Our results further indicate that the new physics scale should
generically be very high (not far below the string or Planck scale) to
make the vacua of the landscape stable against tunneling both during
inflation and at late times, and that a prior period of eternal
inflation is strongly motivated.
{\bf Acknowledgments.} J.M.C. thanks A.\ Linde, L.\ McAllister and
M.\ Trott for helpful discussions, and the CERN Theory Department
and Niels Bohr International Academy for hospitality
while this work was in progress, which was also supported by the Natural
Sciences and Engineering Research Council of Canada. The work of J.R.E. has been partly
supported by the ERC grant 669668 -- NEO-NAT -- ERC-AdG-2014, the
Spanish Ministry MINECO under grants 2016-78022-P and
FPA2014-55613-P, the Severo Ochoa excellence program of MINECO (grant
SEV-2016-0588) and by the Generalitat grant 2014-SGR-1450.
\bibliographystyle{apsrev}
|
\section{Introduction}
Advancements in developing high-performance hardware platforms like GPUs have been a significant enabler for shifting machine learning models, such as neural networks, from rather theoretical concepts to practical solutions to a wide variety of problems.
%
For example, different applications available on smartphones use machine learning models in order to perform services such as categorizing photos based on faces present in the picture through face recognition, digital assistance through speech recognition and synthesis, suggesting similar applications through recommendation systems, and so forth.
%
While existing hardware platforms like GPUs are capable of performing these tasks, designing more energy-efficient and high-performance hardware is crucial in order to allow pervasive deployment of machine learning models across different platforms, from data centers to smartphones and the Internet of Things (IoT) devices.
%
Even tough there has been a substantial effort to design accelerators or use alternative methods for efficient deployment of machine learning models such as convolutional neural networks \cite{lin2018fft}, the training phase of these models has been overlooked.
%
One of the difficulties in designing hardware that is capable of training is that the training phase is typically much more complicated and computationally expensive compared to inference.
%
We believe that designing high-performance and/or energy-efficient hardware for training is of high importance due to several reasons.
%
First, transferring users' data to remote servers puts the users' privacy in danger.
%
Second, communication latency may affect latency-critical applications like online control systems.
%
Third, adapting to a changing environment requires updating model parameters frequently and may be costly, especially in bandwidth-limited devices.
%
And lastly, the energy consumption of wireless modules that need to send/receive data to/from cloud is relatively high \cite{altamimi2015energy}.
%
A suitable hardware for training machine learning models should operate with high performance, be scalable, and be able to train various models by exploiting resource sharing and real-time reconfiguration.
%
Dimensionality reduction, which is the target machine learning model in this paper, has several advantages.
%
First, it removes redundant information from the set of input features, which typically improves the performance of machine learning models.
%
Features that are highly correlated or are closely dependent do not carry much additional information and only make the model more complicated and computationally expensive.
%
Second, transforming features into a lower-dimensional space is more suitable from a hardware design point of view since it leads to a less complex design, less resource consumption, lower memory usage, and so on.
%
This work presents a hardware-friendly algorithm for high-performance, scalable, and reconfigurable training and deployment of dimensionality reduction models.
%
The main focus of this work is to deal with scalability issue of existing hardware implementations for dimensionality reduction while it also considers reconfigurability in order to use the same hardware for various dimensionality reduction algorithms.
%
The rest of this paper is organized as follows.
%
Section~\ref{sec:related} reviews prior work in designing hardware for dimensionality reduction.
%
Section~\ref{sec:prelim} provides some background information about the effect of dimensionality reduction on other machine learning models, explains specific algorithms for dimensionality reduction, and discusses the scalability issue of existing implementations.
%
Section~\ref{sec:implementation} presents our proposed algorithm for dimensionality reduction as well as its hardware implementation.
%
Section~\ref{sec:experiments} demonstrates experimental results and finally, Section~\ref{sec:conclusion} concludes the paper.
%
\section{Related Work} \label{sec:related}
One of the most successful attempts at designing an algorithm for scalable dimensionality reduction is random projection.
%
Random projection is based on the Johnson-Lindenstrauss lemma \cite{johnson1984extensions} and is much simpler than other distance-preserving algorithms such as PCA (Principal Component Analysis).
%
Random projection has been applied to various applications and it has been shown that its quality of results are comparable to other algorithms \cite{bingham2001random, fern2003random, dasgupta2000experiments}.
%
Recently, Fox~\textit{et al.} \cite{fox2016random} implemented random projection in hardware using a simple algorithm that only requires addition and subtraction.
%
The shortcoming of the random projection though, is that it only deals with mixture of Gaussian variables.
%
In other words, it only considers second-order statistics when transforming data points to a lower-dimensional space.
%
In order to consider higher-order statistics (HOS), another class of algorithms known as ICA (Independent Component Analysis) \cite{comon1994independent} is used.
%
Among different algorithms that implement ICA, EASI (Equivariant Adaptive Separation via Independence) \cite{cardoso1996equivariant} has been one of the most suitable ones from a hardware implementation standpoint because it only requires addition and multiplication.
%
EASI includes both training and inference of a dimensionality reduction model that implements ICA.
%
Meyer-Baese \textit{et al.} \cite{meyer2015independent} implement EASI in hardware, but their work has a few shortcomings.
%
First, the clock frequency and throughput are very low.
%
Second, the clock frequency decreases by increasing the number of input or output dimensions.
%
This is a serious problem, especially given the high dimensionality of datasets in existing machine learning problems.
%
Nazemi \textit{et al.} \cite{nazemi2017high} try to address these issues by defining a new approximation to stochastic gradient descent algorithm that is suitable for hardware implementation.
%
Their implementation of EASI using the aforementioned approximation increases the clock frequency by one order of magnitude compared to \cite{meyer2015independent} and keeps the clock frequency independent of input and output dimensions.
%
However, \cite{nazemi2017high} suffers from poor scalability in that its hardware implementation for four input dimensions and two output dimensions consume more than one third of the digital signal processing (DSP) blocks on their target FPGA platform.
%
In general, the major problem with PCA and ICA is their high hardware complexity in terms of adders and multipliers, which limits their scalability to larger dimensions.
%
\section{Preliminaries} \label{sec:prelim}
This section demonstrates the effect of dimensionality reduction on the accuracy of other machine learning models and provides background information about some of the dimensionality reduction models and scalability issue of existing hardware implementations.
%
Additionally, it provides a short mathematical description of the EASI algorithm that is used later in Section~\ref{sec:implementation} for justifying the proposed algorithm.
%
\subsection{Effect of Dimensionality Reduction on Accuracy}
One of the major advantages of dimensionality reduction is that it decreases both computation and storage complexity.
%
When a dimensionality reduction algorithm is applied, the number of input features is reduced and therefore, a smaller amount of memory is required to store the input features.
%
Additionally, the machine learning model that follows the dimensionality reduction model needs to deal with a lower number of input features, which in turn makes that model less computationally expensive.
%
Fig.~\ref{fig:dim-accuracy} compares the classification accuracy for various datasets, different dimensionality reduction algorithms, and different number of input features.
%
For all these datasets, an artificial neural network with two hidden layers is trained in order to perform classification.
%
Fig.~\ref{fig:mnist} demonstrates the effect of reducing dimensionality of input features on the classification accuracy for images in the MNIST dataset \cite{lecun2010mnist}.
%
The MNIST database of handwritten digits includes 70,000 samples where each sample is a 28x28 image (784 pixels total) and the objective is to classify each sample into one of ten classes 0-9.
%
It can be observed that reducing the number of input features to about 100 (\char`\~ 8x reduction) using random projection and bilinear transform does not affect the classification accuracy.
%
In this dataset, PCA and ICA can achieve even higher degrees of reduction (\char`\~ 16x) without any noticeable accuracy degradation.
%
Similarly, Fig.~\ref{fig:har} shows the effect of dimensionality reduction on HAR dataset \cite{anguita2013public}.
%
This dataset uses the accelerometer and gyroscope embedded in a smartphone to measure a group of volunteers' activities over a period of time and the objective is to classify each sample into one of six classes that determines a volunteer's activity.
%
The original number of input features for each sample is 561.
%
It can be seen that ICA and random projection outperform the other two methods and can achieve about 6x reduction in input features without significantly affecting the classification accuracy.
%
Bilinear transform does not perform well in this dataset and the classification accuracy is below 60\%.
%
Lastly, Fig.~\ref{fig:ads} demonstrates the effect of dimensionality reduction on Ads dataset \cite{kushmerick1999learning}.
%
This dataset represents a set of possible advertisements on Internet pages where each sample has 1558 input features, which include the geometry of the image, phrases occurring in the URL, the anchor text, words occurring near the anchor text, etc., and the objective is to determine whether a sample is an advertisement or not.
%
It is observed that reducing the number of input features to five (\char`\~ 300x reduction) does not affect the classification accuracy.
%
This observation corroborates the results presented in prior work on this dataset.
%
It can be concluded that dimensionality reduction models typically perform well for a variety of datasets, including the ones that deal with images, time series, and/or natural language.
%
However, various dimensionality reduction algorithms perform differently on these datasets.
%
This is another reason why a piece of hardware that is capable of implementing different dimensionality reduction algorithms is superior.
%
\begin{figure}[t]
\centering
\subfloat[MNIST Dataset - Image] {
\includegraphics[width=0.8\columnwidth]{figs/mnist.pdf}
\label{fig:mnist}
}
\subfloat[HAR Dataset - Time Series] {
\includegraphics[width=0.8\columnwidth]{figs/har.pdf}
\label{fig:har}
}
\subfloat[Ads Dataset - Natural Language and Image] {
\includegraphics[width=0.8\columnwidth]{figs/ads.pdf}
\label{fig:ads}
}
\caption{Classification accuracy for various datasets, different dimensionality reduction algorithms, and different number of input features.}
\label{fig:dim-accuracy}
\end{figure}
\subsection{Random Projection}
In a random projection, the lower-dimensional input features are found by multiplying the original input features by a randomly generated matrix $R$:
%
\begin{equation}
{\mathbf{v}}_{n \times 1} = R_{n \times m}{\mathbf{x}}_{m \times 1} \quad m \geq n
\label{eq:random-projection}
\end{equation}
%
where $\mathbf{x}$ is a column vector of input features, $R$ is the randomly generated matrix, $\mathbf{v}$ is a column vector of features in the lower-dimensional space, $m$ is the dimensionality of input features, and $n$ is the dimensionality of features in the lower-dimensional space.
%
The elements of $R$ (i.e. $r_{ij}$) are often sampled from a Gaussian distribution, however, there are other proposed distributions such as the ones introduced in \cite{achlioptas2001database, li2006very} that are more suitable for hardware implementation.
%
In this work, we use the distribution that is described in \cite{fox2016random}:
%
\begin{equation*}
r_{ij} =
\begin{cases}
1, & \text{with probability} \quad 1/(2n) \\
0, & \text{with probability} \quad 1 - 1/n \\
-1, & \text{with probability} \quad 1/(2n)
\end{cases}
\label{eq:distribution}
\end{equation*}
%
The advantage of this distribution is that is replaces all multiplications with addition and subtraction and therefore, reduces the hardware cost.
%
One of the major advantages of random projection is that the model does not need to be trained based on input data.
%
As a result, the $R$ matrix can be computed offline without having any information about upcoming input features.
%
\subsection{PCA Whitening}
PCA whitening is an important preprocessing step in many machine learning algorithms.
%
The objective of PCA whitening is to transform data to a lower-dimensional space such that features are less correlated with each other and all features have the same variance (typically unit variance).
%
This can be written as
%
\begin{equation}
{\mathbf{z}}_{n \times 1} = W_{n \times m}{\mathbf{x}}_{m \times 1} \quad m \geq n
\label{eq:pca-whitening}
\end{equation}
%
where $\mathbf{x}$ is a column vector of input features, $W$ is the whitening matrix, $\mathbf{z}$ is a column vector of whitened features, $m$ is the dimensionality of input features, and $n$ is the dimensionality of features in the lower-dimensional space.
\subsection{EASI Algorithm}
ICA can be defined as a generative model in which input features are modeled as linear combinations of some independent components:
%
\begin{equation*}
{\mathbf{x}}_{m \times 1} = A_{m \times n}{\mathbf{s}}_{n \times 1} \quad m \geq n
\end{equation*}
%
where $\mathbf{x}$ is a column vector of input features, $A$ is the mixing matrix, $\mathbf{s}$ is a column vector of random independent components, $m$ is the dimensionality of input features, and $n$ is the dimensionality of independent components in the lower-dimensional space.
%
The objective of ICA is to estimate the mixing matrix and independent components without having any prior information about them.
%
The estimation can be achieved by applying a whitening matrix followed by an orthogonal transformation, i.e. rotation, of intermediate input features.
%
This process is illustrated in Fig.~\ref{fig:whitening-rotation}.
The whitening step can be written as
%
\begin{equation*}
{\mathbf{z}}_{n \times 1} = W_{n \times m}{\mathbf{x}}_{m \times 1} \quad m \geq n
\end{equation*}
%
where $\mathbf{x}$ is a column vector of input features, $W$ is the whitening matrix, and $\mathbf{z}$ is a column vector of whitened features.
%
By definition, $W$ is a whitening matrix if $\mathbf{z}$ is spatially white, that is
%
\begin{equation*}
\Sigma_{{\mathbf{z}}_{n \times n}} = {\mathrm{E}}[{\mathbf{z}}{\mathbf{z}}^T] = I_{n \times n}
\end{equation*}
%
where $\Sigma_{{\mathbf{z}}}$ is the covariance matrix of ${\mathbf{z}}$, and ${\mathrm{E}}$ is the expectation operator.
%
\begin{figure*}[t]
\centering
\subfloat[] {
\includegraphics[width=0.3\textwidth]{figs/1.pdf}
\label{fig:transformed}
} \quad
\subfloat[] {
\includegraphics[width=0.3\textwidth]{figs/2.pdf}
\label{fig:whitened}
} \quad
\subfloat[] {
\includegraphics[width=0.3\textwidth]{figs/3.pdf}
\label{fig:rotated}
}
\caption{Finding independent components by whitening (b) followed by a rotation (c). The whitened features have the same variance and therefore, an appropriate rotation can find the original independent components.}
\label{fig:whitening-rotation}
\end{figure*}
One of the methods for finding the whitening matrix is to minimize the the Kullback-Leibler divergence \cite{kullback1951information} between $\Sigma_{{\mathbf{z}}}$ and $I$.
%
The adaptive updating algorithm for $W$ that minimizes Kullback-Leibler divergence can be written as
%
\begin{equation}
W_{k + 1} = W_{k} - \mu_{k}[{\mathbf{z}_{k}}{\mathbf{z}_{k}}^T - I] W_{k}
\label{eq:whitening}
\end{equation}
%
in which $k$ is the iteration index and $\mu$ is the learning rate.
%
The learning rate does not necessarily need to change across iterations, i.e. $\mu_{k} = \mu$.
The goal of EASI is to find a separation matrix that provides an estimate of independent components without having any prior information about independent components, $\mathbf{s}$, or the mixing matrix, $A$.
%
This can be written as
%
\begin{equation}
{\mathbf{y}}_{n \times 1} = B_{n \times m}{\mathbf{x}}_{m \times 1}
\label{eq:output-features}
\end{equation}
%
where $\mathbf{y}$ is a column vector of estimates of independent components and $B$ is the separation matrix.
%
The separation matrix can be found by applying the whitening matrix followed by a rotation, i.e.
%
\begin{equation*}
B_{n \times m} = U_{n \times n} W_{n \times m}
\end{equation*}
%
where $U$ is an orthogonal matrix.
%
An adaptive updating algorithm that keeps $U$ an orthogonal matrix can be found by
%
\begin{equation}
U_{k + 1} = U_{k} - \mu_{k}[{\mathbf{g}}({\mathbf{y}_{k}}){\mathbf{y}_{k}}^T - {\mathbf{y}_{k}}{\mathbf{g}}({\mathbf{y}_{k}})^T] U_{k}
\label{eq:rotation}
\end{equation}
%
where ${\mathbf{g}}(.)$ is a nonlinear function that introduces HOS into the problem.
%
A global adaptive updating algorithm for the separation matrix can be found by $B_{k + 1} = U_{k + 1} W_{k + 1}$ and plugging in $W_{k + 1}$ and $U_{k + 1}$ from Eq.~\ref{eq:whitening} and Eq.~\ref{eq:rotation}, respectively.
%
By neglecting the $\mu^{2}$ term, the adaptive updating algorithm for $B$ can be written as
%
\begin{equation}
B_{k + 1} = B_{k} - \mu_{k}[{{\mathbf{y}_{k}}{\mathbf{y}_{k}}^T - I + \mathbf{g}}({\mathbf{y}_{k}}){\mathbf{y}_{k}}^T - {\mathbf{y}_{k}}{\mathbf{g}}({\mathbf{y}_{k}})^T] B_{k}
\label{eq:separation}
\end{equation}
%
Eq.~\ref{eq:separation} is known as the EASI algorithm for independent component analysis.
\subsection{Scalability Problem}
Although the hardware implementation of EASI algorithm that is presented in \cite{nazemi2017high} increases the clock frequency by an order of magnitude compared to its prior work, the design suffers from poor scalability.
%
Figure~\ref{fig:easi-smbgd} depicts different stages of hardware implementation of EASI algorithm based on \cite{nazemi2017high}.
%
The algorithm consists of five high-level stages where each stage is responsible for one of the steps explained in Algorithm~\ref{alg:easi-smbgd}.
%
By calculating the number of adders and multipliers required for implementing each stage, one can observe that the hardware complexity of both adder and multiplier units is $\mathcal{O}(mn^2)$.
%
This is obviously not a scalable algorithm and its hardware implementation will occupy the resources available on an FPGA very quickly.
\begin{figure*}[bt]
\centering
\includegraphics[width=0.8\textwidth]{figs/EASI_SMBGD.jpg}
\caption{Hardware implementation of the EASI algorithm.}
\label{fig:easi-smbgd}
\end{figure*}
\begin{algorithm}[tb]
\caption{EASI with Modified Update Rule}
\label{alg:easi-smbgd}
\begin{algorithmic}[1]
\Require
\Statex $\mathbf{x}$: \text{input features}
\Ensure
\Statex $B$: \text{separation matrix}
\Statex $\mathbf{y}$: \text{input features in the lower-dimensional space}
\Repeat
\State Update $\mathbf{y}$ according to Eq.~\ref{eq:output-features}
\State Apply cubic nonlinearity to $\mathbf{y}$
\State Calculate ${{\mathbf{y}_{k}}{\mathbf{y}_{k}}^T - I + \mathbf{g}}({\mathbf{y}_{k}}){\mathbf{y}_{k}}^T - {\mathbf{y}_{k}}{\mathbf{g}}({\mathbf{y}_{k}})^T$
\State Update relative gradient
\State Update separation matrix according to Eq.~\ref{eq:separation}
\Until{convergence}
\end{algorithmic}
\end{algorithm}
\section{Proposed Solution} \label{sec:implementation}
By looking at Eq.~\ref{eq:separation} carefully, we observe that the ${\mathbf{y}_{k}}{\mathbf{y}_{k}}^T - I$ term is in fact responsible for dealing with second-order statistics while the $\mathbf{g}({\mathbf{y}_{k}}){\mathbf{y}_{k}}^T - {\mathbf{y}_{k}}{\mathbf{g}}({\mathbf{y}_{k}})^T$ term deals with higher-order statistics.
%
Adding this to what was explained in Fig.~\ref{fig:whitening-rotation} and the fact that random projection is a suitable algorithm for dealing with second-order statistics, we propose the following solution for implementing the EASI algorithm more efficiently.
%
Initially, input features are fed to a random projection module that preserves the second-order distance among these features, but reduces the dimensionality to an intermediate value $p$.
%
After that, a modified datapath for the EASI algorithm that bypasses the ${\mathbf{y}_{k}}{\mathbf{y}_{k}}^T - I$ term applies a rotation to intermediate input features in order to find features that are independent and have a dimensionality $n$.
%
This process is repeated until the model is trained and can be used later for inference.
%
The major advantage of this solution is that the number of inputs to the EASI module will be decreased because of the dimensionality reduction that is performed by the random projection module.
%
This in turn reduces the number of adders and multipliers in the EASI module linearly due to the linear dependency between the hardware complexity and number of input dimensions.
%
Given the fact that the hardware implementation of random projection has a low overhead, this enables dealing with higher number of input dimensions at the cost of slightly increasing latency.
%
The increase in latency is due to the fact that EASI applies whitening and rotation in parallel, but the proposed solution applies whitening and rotation sequentially.
%
However, the asymptotic latency of random projection is negligible compared to EASI and can be ignored.
%
By comparing Eq.~\ref{eq:pca-whitening} with Eq.~\ref{eq:output-features} and Eq.~\ref{eq:whitening} with Eq.~\ref{eq:separation}, we observe that the algorithm's flow for implementing PCA whitening and EASI are the same.
%
In both algorithms, not only the basic operations like matrix-vector multiplication, matrix-matrix multiplication, and update rule are the same, but also the dimensionality of matrices and vectors are the same.
%
As a result, a hardware that implements EASI for ICA can be used for implementing PCA whitening as well.
%
The only difference is that EASI has an additional term $\mathbf{g}({\mathbf{y}_{k}}){\mathbf{y}_{k}}^T - {\mathbf{y}_{k}}{\mathbf{g}}({\mathbf{y}_{k}})^T$ that needs to be bypassed for PCA whitening simply by using a multiplexer.
%
This allows real-time reconfigurability by issuing proper control signals for each algorithm and therefore, enables using the same hardware for both PCA whitening and ICA.
%
In conclusion, our hardware implementation will comprise of a random projection module followed by an EASI module.
%
The hardware can be used to perform random projection, PCA whitening, ICA, or a combination of random projection with the other two algorithms.
%
This not only achieves implementing different dimensionality reduction algorithms on the same piece of hardware, but also allows dealing with higher number of dimensions.
%
\section{Experimental Results And Discussion} \label{sec:experiments}
\subsection{Dataset}
In order to demonstrate the potentials of proposed solution, we use the Waveform Database Generator (Version 2) Dataset \cite{breiman1984classification}, which is publicly available on UCI Machine Learning Repository \cite{Lichman:2013}.
%
The dimensionality of input features is 40 where all features are noisy and the latter 19 are pure noise with a zero mean and variance of one.
%
The input features are all real numbers and there are no missing values in the dataset.
%
There are three classes of waves and the output classes represent combinations of two out of three of these base waves.
%
The number of samples is 5000 where we use the first 4000 for training of our models and the remaining 1000 for testing.
%
The objective is to classify samples into the three output classes.
%
In this work, we remove the latter eight input features and therefore, reduce the number of features that are pure noise to 13.
%
As a result, the total number of input features will be 32.
%
\subsection{Machine Learning Model}
Our machine learning model consists of a dimensionality reduction module followed by an artificial neural network with two hidden layers and 64 neurons per each layer.
%
For dimensionality reduction, we use different algorithms such as EASI, random projection, or a combination of both.
%
For training the model, we first train the dimensionality reduction model in an unsupervised manner and reduce the dimensionality of input features.
%
After that, we train the neural network using features in the reduced space.
%
Finally, we use the dimensionality reduction model to decrease the dimensionality of test data and use the neural network for classification.
%
\subsection{Results}
Table~\ref{table:software-results} compares the classification accuracy for different dimensionality reduction models and various number of intermediate and output features.
%
It is observed that in configurations where the number of output features is the same, applying EASI independently or using random projection followed by EASI result in almost the same classification accuracy.
%
However, as we will show later, the amount of hardware resources required for the latter is substantially smaller.
%
\begin{table}[t]
\centering
\caption{Classification accuracy for different models and various number of intermediate and output features.}
\begin{tabular}{c c c c c c c}
$m$ & Algorithm 1 & $p$ & Algorithm 2 & $n$ & Accuracy (\%) \\
\midrule
32 & -- & -- & EASI & 16 & 84.6 \\
32 & Random Projection & 24 & EASI & 16 & 84.5 \\
\midrule
32 & -- & -- & EASI & 8 & 80.9 \\
32 & Random Projection & 16 & EASI & 8 & 80.8
\end{tabular}
\label{table:software-results}
\end{table}
Table~\ref{table:hardware-results} summarizes the amount of resources required for implementing models where the number of input dimensions is 32 and the number of output dimensions is 8, after successful synthesis on FPGA.
%
In both implementations, 32-bit floating-point variables and operations are used.
%
The target FPGA is part of Arria 10 family which includes 427,200 adaptive logic modules (ALMs), 55,562,240 bits of block RAM, and 1518 DSP blocks.
%
It can be observed that the number of digital signal processors (DSPs), adaptive logic modules (ALMs), and bits required to store values in registers is reduced by a factor of two in the second scenario.
%
In general, it is expected that the amount of savings will be proportional to $m / p$.
%
As a result, using the random projection module to decrease the intermediate dimensionality further will lead to a more efficient hardware implementation.
%
However, this typically affects the classification accuracy of different models.
%
Therefore, the designer needs to trade off the hardware cost and accuracy in order to find a desirable point for number of intermediate features that achieves a relatively high classification accuracy, but reduces the hardware cost as much as possible.
%
Note that the number of resources presented in Table~\ref{table:hardware-results} are more than the capacity of the target FPGA board and these numbers demonstrate the projected amount of required resources.
%
Note that the pipelined implementation allows all algorithms to operate at the same clock frequency.
%
As a result, using random projection followed by EASI does not lead to a lower frequency of operation, but slightly increases the latency.
%
On our target FPGA, the post-place and route frequency of operation is 106.64\si{\mega\hertz}.
%
\begin{table}[t]
\centering
\caption{Comparison of hardware cost between EASI and random projection followed by EASI.}
\begin{tabular}{c c c c c c}
Input & Intermediate & Output & DSPs & ALMs & Registers \\
\midrule
32 & -- & 8 & 4052 & 38122 & 138368 \\
32 & 16 & 8 & 2212 & 70031 & 75392
\end{tabular}
\label{table:hardware-results}
\end{table}
\balance
\section{Conclusion} \label{sec:conclusion}
In this work, we presented a hardware-friendly algorithm for improving the scalability of existing dimensionality reduction models.
%
Additionally, we presented a reconfigurable hardware implementation that is capable of performing random projection, PCA whitening, and ICA through the EASI algorithm.
%
The part of hardware implementation that improves scalability divides the whitening and rotation tasks between the random projection and EASI modules, respectively.
%
This allows improving the hardware cost by a linear factor which is proportional to the ratio of the number of input features to intermediate features.
%
Our experimental results show a 2x hardware cost reduction for a specific dataset, without affecting the classification accuracy by more than 0.1\%.
%
\section*{Acknowledgements}
This research was sponsored in part by contracts from DARPA's Microsystems Technology Office and the National Science Foundation.
%
\bibliographystyle{IEEEtran}
|
\section{Introduction}
A great deal of work has been devoted in the last years to
determine the role of mass and environmental quenching
on the properties of central and satellite galaxies.
Mass quenching\footnote{Although the term ``secular quenching'' describes more accurately the meaning of what we refer to as ``mass quenching'', we adopt the latter following the convention in the literature.} refers to any internal process
determined by the galaxy stellar mass.
These
self-regulating processes, such as feedback from
active galactic nuclei (AGN)
\citep[e.g. ][]{Fabian12, Beckmann17}
and stellar feedback
\citep[e.g. ][]{Hopkins14, Chan18},
can affect both central and satellite galaxies, and
are thought to be responsible
for the dependence of galaxy properties on stellar
mass.
Environmental quenching, on the other hand,
corresponds to the decline of star formation (SF) activity of satellite galaxies due to their accretion into a
massive dark matter halo. It involves environment-dependent physical processes,
like ram pressure stripping \citep[RPS, ][]{gg72}, tidal stripping
\citep[TS, ][]{Merritt83},
thermal evaporation \citep{CowieSongaila77}, turbulent viscous
stripping \citep{Nulsen82} and
galaxy harassment \citep{Moore96}.
The dependence of the efficiency of mass quenching on environment and of environmental quenching on stellar mass is somewhat controversial.
Mass and environmental quenching have been reported to be independent of
each other \citep[e.g.][]{Peng10, Quadri12, Muzzin12, Lin14, GaborDave15}, while
other studies conclude the opposite
\citep[e.g.][]{Darvish16, Kawinwanichakij17}.
\citet{Darvish16} find
that environmental quenching is more efficient for more massive galaxies
($M_{\star} \gtrsim 5 \times 10^{10}\,{\rm M}_{\odot}$) at $z \lesssim 1$,
whereas mass quenching gains relevance at $z \gtrsim 1$ and is
more efficient in higher density environments.
\citet{Kawinwanichakij17}
extend the analysis to lower mass galaxies
demonstrating that the presence of nearly all quiescent
low-mass galaxies
($\approx 3 \times 10^{9} - 10^{10}\,{\rm M}_{\odot}$) can be explained by
environmental quenching.
At fixed stellar mass, the star formation rate density (SFRD)
declines with decreasing redshift more abruptly for galaxies in clusters than in the field
\citep{Guglielmo15}, highlighting the importance of environmental quenching, especially since $z\sim 1.5$ \citep{nantais17}.
On-going programmes like GOGREEN
(Gemini Observations of Galaxies in Rich Early Environments)
will contribute
towards determining
the role of environment in the evolution of low-mass galaxies \citep{Balogh17}.
Galaxy interactions and mergers might explain the scenario in which environmental and mass quenching are not separable, i.e. in the regime of very dense environments and very massive galaxies \citep{Darvish16}.
Massive galaxies
($M_{\star}\sim 10^{10} - 10^{11.5}\,M_{\odot}$) have
typically experienced
around one
major merger since $z = 1$ \citep{Xu2012a}.
Likewise, ellipticals in groups achieve their spheroidal morphology through
major mergers that take place at early epochs ($z \gtrsim 1$)
when they inhabit the progenitor haloes of these systems, and
their SF truncation occurs later when they fall into a galaxy group \citep{Feldmann11}.
Massive inflows
of gas in major mergers of star-forming galaxies trigger starbursts and feed the central supermassive black holes (BHs)
(\citealt[e.g.][]{Hopkins08, Khabiboulline14}); while starbursts consume most of the nuclear gas,
feedback from supernovae (SNe) and AGN expels
the residual gas.
Galaxy formation models have also explored
the relative importance of different quenching mechanisms
using both hydrodynamical simulations
\citep[e.g. ][]{Bahe15, Taylor17, Dave17}
and semi-analytic models of galaxy formation
\citep[SAMs; e.g.][]{delucia2012, Henriques17, Stevens17}.
Both approaches predict that quenching begins once
the cooling of gas becomes inefficient
\citep[e.g.][]{Schawinski14, Peng15}.
Thus, the cold gas disc is no longer replenished by cooling flows
and is consumed within a few Gyr. This results in a decline of the star formation rate
(SFR) until the galaxy becomes passive.
Such scenario corresponds to the second stage in the so-called \textsl{delay-then-rapid}
quenching scenario
proposed by \citet[][W13, hereafter]{Wetzel13}, in which the onset
of quenching since the first infall
can take $\approx 4\,{\rm Gyr}$ for
low-mass
galaxies
($\approx 10^{10}\,{\rm M_{\odot}}$)
to $\approx 2\,{\rm Gyr}$ for
high-mass
ones ($\approx 10^{11}\,{\rm M_{\odot}}$). Once the SF quenching begins,
this takes place in a short time-scale ($\lesssim 1\,{\rm Gyr}$).
A crucial aspect in determining
the gas cooling efficiency and, thus, the
quenching time-scales of satellite galaxies is the modelling of the hot gas removal in galaxies.
The updated version of the semi-analytic model
of galaxy formation \textsc{sag}~(acronym for Semi-Analytic Galaxies) presented in \citet[][hereafter Paper I]{Cora18a} incorporates a gradual removal of hot gas of satellites by the action of RPS and TS. RPS can also affect the cold gas disc. This improved treatment of
environmental effects combined with a modified modelling of
SN feedback has been successful in reproducing
several galaxy properties at both low and high redshifts.
In particular, the predicted fractions of quenched galaxies
as a function of stellar mass, halo mass and halo-centric distances
are in good agreement with \citet{Wetzel12}. To achieve this result, a power-law slope of the redshift
dependence of the reheated and ejected mass is used following that from the
zoom FIRE (Feedback
in Realistic Environments) hydrodynamical simulations
\citep{muratov15}. In this variant of the model, referred to as \textsc{sag$_{\beta1.3}$}~in Paper I, the values of the rest of the free parameters obtained from the calibration process remain unchanged.
In this second paper of a series, we explore further the relevance
of mass quenching and environmental quenching on currently passive satellite
galaxies, and determine the time-scales involved in the quenching process.
This paper is organized as follows.
Section~\ref{sec:model}
presents the galaxy formation model that combines a cosmological dark matter (DM) simulation with the \textsc{sag$_{\beta1.3}$}~model; we describe those aspect of our SAM that are relevant to this work,
summarising the main conclusions of Paper I.
In Section~\ref{sec:fq-z0-zinfall}, we
analyse the
fraction of
quenched galaxies
at the time of first infall, and the relative role of environmental and mass
quenching after their first infall.
In Section~\ref{sec:tq}, we discuss the quenching time of $z=0$ passive satellites that are star-forming at first infall.
In Section~\ref{sec:quenchscenario}, we propose a delay-then-fade quenching scenario, and
identify the physical processes associated with its two phases.
We present a summary of the main results and
our conclusions in Section~\ref{sec:conclu}.
\section{Galaxy formation model}
\label{sec:model}
The galaxy formation model combines our semi-analytic model of galaxy formation and evolution
\textsc{sag}~and the cosmological DM \textsc{MultiDark} simulation MDPL2,
which is part of the \textsc{CosmoSim} database\footnote{\url{https://www.cosmosim.org}}.
The MDPL2 simulation follows the evolution of $3840^3$ particles within a box of side-length $1\,h^{-1}\,{\rm Gpc}$,
with a mass resolution
$m_\textrm{p} = 1.5 \times 10^{9}\, h^{-1}\, \textrm{M}_{\odot}$ per
DM~particle \citep{Klypin16}.
It is consistent with
a flat $\Lambda$CDM model characterised by Planck cosmological parameters:
$\Omega_{\rm m}$~=~0.307, $\Omega_\Lambda$~=~0.693,
$\Omega_{\rm B}$~=~0.048, $n_{\rm s}$~=~0.96
and $H_0$~=~100~$h^{-1}$~km~s$^{-1}$~Mpc$^{-1}$, where $h$~=~0.678
\citep{Planck2013}.
The simulation provides the DM haloes and their corresponding merger trees required by \textsc{sag}~to generate the galaxy population and track the evolution of galaxy properties.
DM haloes have been identified with the
\textsc{Rockstar} halo finder
\citep{Behroozi_rockstar}, and merger trees were constructed with
\textsc{ConsistentTrees} \citep{Behroozi_ctrees}.
DM haloes detected over the background density are referred to
as \textsl{main host} haloes, whereas those lying
within another DM haloes are subhaloes.
\textsc{sag}~assigns
one galaxy to each new detected halo in the simulation.
Thus, each system of haloes contains a central galaxy associated to the main host halo and satellite galaxies.
Those galaxies that are assigned to DM haloes that are no longer
identified by the halo finder after the merger of two haloes
are called orphan satellites.
The model \textsc{sag}~originates from the semi-analytic model presented
in \citet{springel2001} and was further modified as described
in \citet{cora2006},
\citet{lcp08}, \citet{tecce10},
\citet{orsi14}, \citet{munnozarancibia2015} and
\citet{Gargiulo15}.
\textsc{sag}~includes the effects of radiative cooling of hot gas,
star formation, feedback from SN explosions and AGN, and starbursts triggered by disc
instabilities and/or galaxy mergers, and features
a chemical enrichment model that
tracks several chemical elements contributed by
different sources (stellar winds and SNe
Type Ia and II )
taking into account the lifetime of progenitors.
This model has been improved by the implementation of a proper treatment
of environmental effects on satellite galaxies. The strangulation scheme,
in which the hot gas halo is removed instantly when the galaxy becomes a satellite, is replaced by gradual starvation produced by the combined action of RPS and TS.
Therefore, gas cooling
takes place in both central and satellite galaxies.
The cold gas disc is also subject to
RPS
when it is not longer shielded by the hot gas halo;
it can also suffer the effects of TS.
Stellar mass loss as a result of TS is also taken into account.
The advantage of our implementation with respect to previous SAMs
\citep{GonzalezPerez14, Henriques17, Stevens17} resides in the integration of the
orbits of orphan galaxies
according
to the potential well of the
host halo, taking into account mass loss by TS and dynamical friction effects,
and in the use of
fitting formulae to estimate ram pressure (RP) as a function of halo mass,
halo-centric distance and redshift
(Vega-Mart\'inez et al., in preparation).
The combination of these two improvements
provides values of RP consistent with the position and velocity
of satellite galaxies.
The latest version of \textsc{sag}~also considers a modification in the SN feedback scheme that includes an explicit redshift dependence
of the reheated and ejected mass,
and in the rate of BH growth during gas
cooling involved in the AGN feedback.
A detailed description of the updated version of \textsc{sag}~is presented in Paper I.
We summarize here the implementations related to the new SN feedback scheme, the modification of AGN feedback and the treatment of environmental effects on the hot and cold gas phases
of satellite galaxies.
\subsection{SN feedback}
SNe are associated to
star-forming
events and reheat part of the cold gas
disc. As a result, the reheated gas is transferred to the hot phase.
The reheated mass is given by
\begin{equation}
\Delta M_{\rm reheated} = \frac{4}{3} \epsilon {\frac{\eta E_\text{SN}}{V_{\rm vir}^2}} \,(1+z)^{\beta}\,\left(\frac{V_{\rm vir}}{60\,{\rm km}\,{\rm s}^{-1}}\right)^{\alpha_\text{F}}\Delta M_{\star},
\label{eq:feedfire}
\end{equation}
where $\eta$ is the number of
SNe generated from the stellar population of mass $\Delta M_\star$ formed,
$E_\text{SN}=10^{51}\,{\rm erg}$ is the energy released
by a single SN, and $V_{\rm vir}$ is the virial velocity of the host (sub)halo.
The exponent $\alpha_\text{F}$ takes the values $-3.2$ and $-1.0$
for virial velocities
smaller and larger than $60\,{\rm km}\,{\rm s}^{-1}$, respectively. This additional modulation with virial velocity as well as the explicit redshift dependence are included following \citet{hirschmann16}, based on the results from
the zoom FIRE hydrodynamical simulations of \citet{muratov15}.
The
SN feedback
efficiency, $\epsilon$,
and the power-law slope of the redshift dependence, $
\beta$, are
free parameters that control the amount of cold gas reheated by the
energy generated by SNe as a function of time.
The number of SNe,
$\eta$, depends on the initial mass function (IMF);
we adopt the Chabrier IMF
\citep{Chabrier03}.
During the SN feedback process, some hot gas can also be ejected to an external reservoir. Based on the energy conservation argument
presented by \citet{guo11}, the ejected hot gas mass
is given by
\begin{equation}
\Delta M_{\rm ejected}= \frac{\Delta E_\text{SN} - 0.5\,\Delta M_{\rm reheated}\,V_{\rm vir}^2}{0.5\,V_{\rm vir}^2},
\label{eq:EnergyCons}
\end{equation}
\noindent where $\Delta E_\text{SN}$ is the energy injected by massive stars, estimated as the reheated mass but with the addition of the
mean kinetic energy of SN ejecta per
unit mass of stars formed, $0.5\,V_\text{SN}^2$, i.e.
\begin{equation}
\Delta E_{\rm SN} = \frac{4}{3} \epsilon_\text{ejec} {\frac{\eta E_\text{SN}}{V_{\rm vir}^2}} \,(1+z)^{\beta}\,\left(\frac{V_{\rm vir}}{60\,{\rm km}\,{\rm s}^{-1}}\right)^{\alpha_{\rm F}}\Delta M_{\star}\,0.5\,V_\text{SN}^2.
\label{eq:ejecfire}
\end{equation}
\noindent The wind velocity is given by $V_\text{SN}=1.9\,V_\text{vir}^{1.1}$ \citep{muratov15}.
The efficiency of ejection, $\epsilon_\text{ejec}$, is another free parameter of the model.
The ejected gas mass is assumed to
be re-incorporated back onto the (sub)halo from which it was expelled and is modelled following \citet{henriques13}.
Thus, the reincorporated mass is given by
\begin{equation}
\Delta M_{\rm reinc}= \gamma\,\Delta M_\text{ejected}\,\frac{M_{\rm vir}}{10^{10}\,{\rm M}_{\odot}},
\label{eq:reinc}
\end{equation}
\noindent where the parameter $\gamma$ is a free parameter that regulates the efficiency of the process.
\subsection{AGN feedback}
AGN feedback reduces gas cooling in large haloes, preventing them
from forming stars at late times.
This process is triggered by
gas accretion events onto super-massive BHs.
AGN feedback has been included in \textsc{sag}~by \citet{lcp08} following \citet{croton2006},
but the modelling of the BH growth was subsequently modified
(\citealt{ruiz2015}, Paper I).
BHs grow via
gas flows to the galactic core
triggered by disc instabilities or galaxy
mergers. BHs are assumed
to merge instantaneously when a merger occurs.
The resulting BH grows through cold gas accretion
following
\begin{equation}
\Delta M_{\rm BH} = f_{\rm BH} \frac{{M_{\rm sat}}}{M_{\rm cen}} \frac{M_{\rm cold,sat}
+ M_{\rm cold,cen}}{(1 + 280 \, {\rm km}\,{\rm s}^{-1} / V_{\rm vir})^2},
\end{equation}
where $M_{\rm cen}$ and $M_{\rm sat}$ are the masses of the merging central and
satellite galaxies, and $M_{\rm cold, cen}$ and $M_{\rm cold,sat}$ are their
corresponding cold gas masses.
The fraction of cold gas accreted onto the central
super-massive BH, $f_{\rm BH}$, is a free parameter of the model.
BHs can also grow during gas cooling processes taking place once a
static hot gas halo has formed around the central galaxy. Following \citet{henriques_mcmc_2015}, the BH growth rate is given by
\begin{equation}
\label{eq:BHgrowth-gascool}
\dot{M}_{\rm BH} =
\kappa_{\rm AGN} \frac{M_{\rm BH}}{10^8 \,{\rm M}_{\odot}}
\frac{M_{\rm hot}}{10^{11} \,{\rm M}_{\odot}},
\end{equation}
where $M_{\rm hot}$ and
$M_{\rm BH}$ are
the hot gas and BH masses, respectively.
The efficiency of cold gas accretion is given by the free parameter $\kappa_{\rm AGN}$.
AGN feedback reduces the amount of gas that cools by
\begin{equation}
\dot{M}_{\rm cool}^{\prime}= \dot{M}_{\rm cool}-\frac{L_{\rm BH}}{{\rm V}_{\rm vir}^2/2},
\label{eq:cool}
\end{equation}
where the BH luminosity, ${L_{\rm BH}}$, i.e. the
mechanical heating generated by the BH accretion, is given by
\begin{equation}
L_{\rm BH}= \eta \, \dot{M}_{\rm BH} \, c^2,
\label{eq:lumbh}
\end{equation}
where $\eta=0.1$ is the standard efficiency of energy
production that occurs in the vicinity of the event horizon, and $c$ is the speed of light.
\subsection{Gradual removal of the hot gas halo of satellite galaxies}
The hot gas halo of satellite galaxies is gradually stripped by the action of~RPS
and/or~TS.
The gas beyond a
satellite-centric radius~$r_\text{sat}$ will be removed by RPS if the
pressure term $P_{\rm ram}$ satisfies the
condition derived from the hydrodynamic simulations
by \citet{mccarthy2008}
\begin{equation}\label{eq:rpshot}
P_{\rm ram} > \alpha_\text{RP} \frac{G
M_\text{sat}(r_\text{sat}) \rho_\text{hot}
(r_\text{sat})}{r_\text{sat}}.
\end{equation}
\noindent Here, $\rho_\text{hot}$ is the density profile of the hot gas halo for an
isothermal sphere, i.e.,
$\rho_\text{hot} = {M_\text{hot}}/({4\pi\, r_\text{hot}\, r^2})$,
where $r_\text{hot}$ is
the radius containing $M_\text{hot}$.
This radius
initially adopts the value of the subhalo
virial radius, $r_\text{vir}$. In the case of orphan satellites, $r_\text{vir}$~preserves
the value corresponding to the last time the subhalo was identified.
For the geometrical constant $\alpha_\text{RP}$ we adopt $\alpha_\text{RP}=5$, which is an intermediate value within the range considered by \citet{mccarthy2008}.
The total mass $M_\text{sat}$ of a satellite is
\begin{equation}\label{eq:msat}
\begin{split}
M_\text{sat}(r_\text{sat}) = & M_\star + M_\text{cold}\\
&+ 4\pi \int_0^{r_\text{sat}} \left[ \rho_\text{hot}(r) +
\rho_\text{DM}(r) \right] r^2 dr,
\end{split}
\end{equation}
assuming that $r_\text{sat}$ is large enough to contain all the stars and cold
gas.
We also assume that the DM
is distributed in
an isothermal sphere density profile.
The hot gas stripping radius due to~RP,
$r_\text{s,hot}^\text{RPS}$ can be obtained by solving numerically the combined equations ~\eqref{eq:rpshot} and~\eqref{eq:msat} (see Appendix A in Paper I).
This radius is compared with the
tidal radius determined by TS,
that is,
$r_\text{s,hot}^\text{TS}=r_\text{DM}$,
where the bounding radius for the DM,~$r_\text{DM}$,
is given by $r_\text{vir}$; we are assuming that
the hot gas distributes
parallel to the~DM.
The smaller of
$r_\text{s,hot}^\text{RPS}$ and $r_\text{s,hot}^\text{TS}$
is the stripping radius~$r_\text{s,hot}$.
Thus, the value of~$r_\text{hot}$~is updated such that
$r_\text{hot}^\text{new}=r_\text{s,hot}$.
All hot gas beyond that radius is stripped.
The
remaining gas is redistributed restoring an isothermal profile,
but truncated at~$r_\text{hot}$,
as in~\citet{font2008} and \citet{kimm2011}.
\subsection[]{Ram pressure and tidal stripping of cold gas disc}
\label{sec_RPS_TS_cold}
RPS of cold gas is modelled according to the simple criterion proposed by \citet{gg72}, which was implemented in \textsc{sag}~by \citet{tecce10}. The cold gas of the
galactic disc located at a galactocentric radius~$R$ is stripped away when
the~RP exerted on the galaxy by the intragroup/intracluster medium exceeds the restoring force
per unit area due to the gravity of the disc, that is
\begin{equation}\label{eq:rps}
P_{\rm ram} > 2\pi G \,\Sigma_\text{disc}(R)\,
\Sigma_\text{cold}(R),
\end{equation}
where $\Sigma_\text{disc}$, $\Sigma_\text{cold}$ are the surface densities of the
galactic disc (stars plus cold gas) and of the cold gas disc, respectively.
The discs of stars and gas are modelled by
an exponential surface density profile given by
$\Sigma(R) = \Sigma_0 \exp(-R/R_\text{d})$
where $\Sigma_0$ is the central surface density and~$R_\text{d}$ is the scale length
of the
disc. This scale length is estimated as
$R_\text{d}=(\lambda/\sqrt{2})R_\text{vir}$ \citep{mmw98},
where $\lambda$ is the spin parameter of the DM halo in which
the galaxy resides.
The stripping radius~$R_\text{s,cold}^\text{RPS}$, beyond which all cold gas is removed, is derived from Eq.~\eqref{eq:rps} and is given by
\begin{equation}\label{eq:rstrip}
R_\text{s,cold}^\text{RPS} = -0.5 R_\text{d} \ln \left( \frac{P_{\rm ram}}{2\pi G \,\Sigma_\text{0,disc}\,
\Sigma_\text{0,cold}} \right),
\end{equation}
where $\Sigma_\text{0,disc}$ and $\Sigma_\text{0,cold}$ are the central surface
densities of the stellar disc and of the cold gas disc, respectively. For simplicity, both disc components are considered to have the same scale length.
We have checked that the precise distribution of the stellar content has a negligible impact on our results.
We assume that the hot gas halo of satellite galaxies shields the cold gas
disc from the action of~RPS. Thus, the ambient~RP starts affecting the cold gas once
the hot gas halo has been
reduced to very low values by gas cooling and/or stripping processes (RPS, TS), that is, when its mass becomes less than $10$ percent the baryonic mass of the galaxy.
This threshold has been chosen small enough to allow the role of the hot gas as a
shield for a sufficiently long time, otherwise RPS on the cold disc gas
becomes too effective.
The stripping radius due to TS,
$R_\text{s,cold}^\text{TS}$, is determined by the bounding radius of the DM halo, ~$r_\text{DM}$.
At each snapshot of the simulation, $R_\text{s,cold}^\text{TS}$ is compared with
the stripping radius determined by RPS, $R_\text{s,cold}^\text{RPS}$ (equation~\eqref{eq:rstrip}), in order
to account for the effect of tides.
The smaller of these two radii determines the stripping radius $R_\text{s,cold}$ and all the gas beyond this radius
is stripped away.
The stripped gas is added
either to the hot gas component of
the central galaxy (which represents the intragroup/intracluster medium) or to the hot gas halo of satellite galaxies
(an orphan galaxy could in turn orbit
a satellite galaxy).
If the hot gas phase is no longer present in the latter case, the stripped cold gas is transferred
to the hot gas contained in the main halo.
After a stripping event produced by RPS and/or TS, the
remaining disc gas is redistributed following an exponential surface
density profile truncated at $R_\text{cold}^\text{new}=R_\text{s,cold}$,
with a new scale length defined as
$R_\text{d,cold}^\text{new}=R_\text{cold}^\text{new}/7$
(assuming that $99$ percent of the cold gas disc is contained within $7*R_\text{d,cold}$).
\subsection{Galaxy properties from the SAG$_{\beta 1.3}$ model}
The free parameters of
the \textsc{sag}~model
were tuned using the Particle Swarm
Optimization Technique (PSO) by \citet{ruiz2015} adopting a set of five observational constraints defined in
\citet{Knebe18}. Namely, these are the stellar mass
functions (SMF) at $z=0$ and $z=2$
(compilation data used by \citealt{henriques_mcmc_2015}), the star formation rate (SFR) distribution
function for the
redshift interval $z \in [0.0,0.3]$ \citep{gruppioni15},
the fraction of mass in cold gas as a function of stellar mass
\citep{boselli14},
and the relation between bulge mass and the mass of the central
supermassive black hole, taken from \citet{mcconnell_bhb_2013} and \citet{kormendy_bhb_2013}.
During the calibration process, the power-law slope $\beta$ that
characterises the redshift dependence included in the modified SN feedback scheme
(eqs.~\ref{eq:feedfire} and \ref{eq:ejecfire})
was considered a free parameter of the model.
As discussed in Paper I, the fit to the SMF at $z=2$ results in $\beta=1.99$, which is above the value obtained from the zoom FIRE hydrodynamical simulations of
\citet{muratov15}, $\beta=1.3$.
As a result, the predicted SFRD
becomes too low with respect to observational data
at high redshifts. A better agreement between model predictions and
observational data is obtained by fixing $\beta=1.3$ while keeping
the rest of the parameter values from the calibration process.
This variant of the model is referred to as the \textsc{sag$_{\beta1.3}$}~model in Paper I. As a result of this choice for $\beta$, the model predicts a higher number density of low-mass galaxies in the SMF at $z=2$.
We classify galaxies as
star-forming
or passive according to their
specific star formation rate density (sSFR).
Passive galaxies are those with
${\rm sSFR} < 10^{-10.7}\,{\rm yr}^{-1}$. This cut is obtained
from the distribution of sSFR of our model galaxies
(see fig.~9 in Paper I).
The underpredicted SFRD at high redshifts from the calibrated model (characterised by $\beta=1.99$) results in
fewer
passive galaxies at $z=0$ (see fig.~10 in Paper I). This occurs due to
the bulk of SF being shifted to lower redshifts, meaning that
the galaxies lack sufficient time to quench their SF.
Only the most massive galaxies
($\log(M_{\star}[{\rm M}_{\odot}])\in [10.9, 11.3]$),
within DM host haloes with masses
$M_{\rm halo}\gtrsim 10^{14}\,{\rm M}_{\odot}$,
achieve quenched fractions similar to those presented by \citet{Wetzel12}.
By using a milder redshift dependence of the reheated and ejected mass, as characterised by \textsc{sag$_{\beta1.3}$}, we achieve the
expected behaviour of the fractions of quenched galaxies
as a function of stellar mass, halo mass and the halo-centric distances
(see figs.~11 and 12 in Paper I).
The analysis of the galaxy catalogue generated by the \textsc{sag$_{\beta1.3}$}~model
allows us to conclude that RPS plays a dominant role
among the environmental processes considered in our model, being more effective for lower mass galaxies
residing in more massive haloes
and at lower redshifts
(see figs.~13, 14 and 15 in Paper I), and
adequately regulates
the mass of the hot gas halo and the cold gas disc.
While the gradual starvation of the hot
gas reservoir is a key ingredient
to determine the right SF history of satellite galaxies,
the mild effect suffered by the cold gas disc improves the atomic gas content of galaxies when compared to observations (see fig.~16 in Paper I).
In view of these results, the current work is based on the analysis of
the properties of galaxies generated by the \textsc{sag$_{\beta1.3}$}~model.
\section{Quenched fractions:
dependence on
time at first infall}
\label{sec:fq-z0-zinfall}
Both mass and environmental quenching can affect the galaxy
after being accreted but their relative role depends
on both stellar mass and host halo mass
and cannot be estimated from the fraction of
quenched galaxies at $z=0$.
We
explore this issue
by
making a deeper analysis of further information
provided by the \textsc{sag$_{\beta1.3}$}~model. This involves the
dependence
of quenched fractions
and stellar mass growth
on
the time of first infall,
defined as the moment in which the galaxy
becomes a satellite for the first time; there is strong
evidence that environmental effects
produce SF quenching
from
first infall (W13).
We denote the corresponding redshift as $z_{\rm infall}$. \\
\subsection{$z=0$ quenched fractions}
\label{sec:fqz0}
Fig.~\ref{fig:fqz0_zinfall} shows the fraction of
quenched satellites at $z=0$, $fq_{\rm z0}$,
as a function of $z_{\rm infall}$.
Satellites are binned
according to their present-day stellar mass as indicated in the legend.
We consider those galaxies identified as satellites within main host haloes of mass
$M_{\rm halo} \geq 10^{12.3}\,{\rm M}_{\odot}$,
following the selection criterion adopted in \citet{Wetzel12}.
For a given stellar mass range,
$fq_{\rm z0}$ increases for satellites with higher values of $z_{\rm infall}$, as expected.
Thus, galaxies that have been satellites for longer have had more time
to experience
environmental quenching.
Values of $fq_{\rm z0}$ are larger
for larger stellar masses, at any $z_{\rm infall}$.
However, the dependence of $fq_{\rm z0}$
on
$z_{\rm infall}$
is more pronounced for less
massive satellites. This trend becomes gradually milder for more
massive galaxies.
For a given stellar mass range,
the average values of the quenched fractions are
consistent with the results
of the stellar and halo mass dependence of $fq_{\rm z0}$
shown in
fig.~11
of Paper I.
The relations $fq_{\rm z0}$ vs. $z_{\rm infall}$ are separated by a major gap
between the two lowest and two highest stellar mass bins
which is produced by the under-prediction of passive galaxies in the mass
range ${\log} (M_{\star} [{\rm M}_{\odot}]) \in [10.1, 10.5]$,
discussed in Paper I.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig1.pdf}
\caption{
Fraction of
satellite galaxies that are quenched at $z=0$, $fq_{\rm z0}$,
as a function of the redshift at infall, $z_{\rm infall}$.
All galaxies identified as satellites within main host haloes of present-day mass
${\rm log} (M_{\rm halo} [{\rm M}_{\odot}]) \geq 12.3$
are included.
Different lines represent the values of $fq_{\rm z0}$
for galaxies in different
local stellar mass ranges,
as indicated in the legend.
Error bars show the $68$ percent bayesian confidence interval estimated
following \citet{Cameron11};
they are only visible for the highest stellar mass bin.
}
\label{fig:fqz0_zinfall}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=2.0\columnwidth]{Fig2.pdf}
\caption{
Fraction of
satellite galaxies that are quenched at first infall, $fq_{\rm infall}$,
as a function of the redshift at infall, $z_{\rm infall}$.
Different lines represent the values of $fq_{\rm infall}$
for satellites in different
present-day stellar mass ranges, as indicated in the legend.
They are grouped according to the mass of their main host haloes
(different panels).
Error bars show the $68$ percent bayesian confidence interval estimated
following \citet{Cameron11};
they are only visible for the highest stellar mass bin.
}
\label{fig:fqzinfall_zinfall}
\end{figure*}
Low-mass
satellites characterised by
$z_{\rm infall} \approx 0$ are mostly
star-forming
galaxies
($fq_{\rm z0} \approx 0.0 - 0.1$). This is because SF quenching as a result of
environmental effects is negligible for galaxies that have been recently
accreted, and these galaxies are too small to suffer the consequences
of mass quenching
while they were centrals.
However,
more than $50$ percent of
high-mass
satellites
accreted at the present epoch are passive.
Massive galaxies
have thus
suffered mass quenching processes as centrals prior to infall
(e.g. AGN feedback
that reduces the gas cooling rates;
starbursts
triggered by both mergers and disk instabilities
that exhaust the cold gas reservoir)
as already found by
\citet{vandenBosch08}.
Hence, environmental
processes can be considered
as the dominant quenching mechanism in low-mass
galaxies consistent with the conclusions of \citet{Wetzel12}.
The fraction of quenched galaxies at the present epoch reach
similarly high
values
($\sim 0.8 - 0.95$)
for all galaxies regardless of their stellar mass when they have been
satellites for more than $\approx 8\, {\rm Gyr}$ ($z_{\rm infall} \gtrsim 1$).
The time elapsed
is enough for the maximum possible
effect of those processes affecting satellites to take place.
Both the nature of these processes
and their relative role (if there is more than one at play, e.g. environmental
quenching and/or mass quenching)
can vary for galaxies with different masses.
\citet{Lin14} investigate the relation
between environmental quenching and mass quenching efficiencies
with stellar mass for galaxies in
a large optically selected sample of field and group
galaxies obtained from the Pan-STARRS1 Medium-Deep
Survey, which span a redshift range $0.2 <z<0.8$.
For lower redshifts ($0.2<z<0.5$), they find that SF in more massive galaxies
($M_{\star} \gtrsim 1-2 \times 10^{10}\,M_{\odot}$)
is mainly suppressed by mass quenching processes
while environmental quenching gains relevance for lower mass galaxies.
Mass quenching keeps the dominant role for massive galaxies
in the higher redshifts probed ($0.5 < z < 0.8$).
This range overlaps with the one considered by
\citet{Kawinwanichakij17},
who find similar results from
the FourStar Galaxy Evolution (ZFOURGE) survey.
\subsection{Quenched fractions at first infall}
\label{sec:fqinfall}
We study the fraction of satellite galaxies selected at $z=0$ that are quenched at the time of first infall, $fq_{\rm infall}$.
This is
estimated from the SFR and stellar mass at $z_{\rm infall}$.
The relation between $fq_{\rm infall}$ and $z_{\rm infall}$
is presented in
Fig.~\ref{fig:fqzinfall_zinfall},
grouping satellite galaxies according to their local stellar mass (different lines)
and the mass
of their main host haloes
(different panels).
Low-mass
satellites
have not suffered mass quenching before becoming satellites,
since $fq_{\rm infall} \approx 0$ for any value of $z_{\rm infall}$. This is consistent with the previous analysis of $fq_{\rm z0}$ (Fig.~\ref{fig:fqz0_zinfall}). Intermediate mass satellites
($M_{\star} [{\rm M}_{\odot}] \in [10^{10.5}, 10^{10.9}]$)
accreted at $z_{\rm infall} \gtrsim 1.5$
also have $fq_{\rm infall} \approx 0$,
being almost all still
star-forming
by that time. The most massive satellites considered in this analysis
($M_{\star} [{\rm M}_{\odot}] \in [10^{10.9}, 10^{11.3}]$)
have $fq_{\rm infall}\approx 0 - 0.3$
for these early accretion times, with the larger fractions corresponding to those satellites
residing within massive main host haloes
($M_{\rm vir} [{\rm M}_{\odot}] \in [10^{14.1}, 10^{15.}]$).
As $z_{\rm infall}$ decreases, $fq_{\rm infall}$ increases
monotonically
for galaxies in the two more massive stellar-mass bins,
These galaxies experience mass quenching processes as centrals prior to infall, which becomes more efficient due to stellar mass growth.
Combining the results of the fractions $fq_{\rm z0}$ and $fq_{\rm infall}$
for a given stellar mass, we see that for
$M_{\star} \gtrsim 10^{10.9} \,{\rm M}_{\odot}$
almost
all satellite galaxies that are
star-forming
at time of first infall are quenched by $z=0$,
regardless of $z_{\rm infall}$.
The fraction of
low-mass
satellites
($M_{\star} \lesssim 10^{10.1} \,{\rm M}_{\odot}$)
that were
star-forming
at infall and
quenched after infall is directly given by $fq_{\rm z0}$
because galaxies in this mass range are all
star-forming
at infall
($fq_{\rm zinfall}\approx 0$)
regardless of $z_{\rm infall}$.
If satellites are not discriminated according
to $z_{\rm infall}$ then we find that $fq_{\rm z0}\approx 0.46$ for low-mass satellites residing
within main host haloes of mass
$M_{\rm halo} \geq 10^{12.3}\,{\rm M}_{\odot}$.
Thus, we arise to
conclusions similar to those obtained
by W13 from the analysis of the results presented in their fig. 7.
That is,
half of
the
satellites with low stellar mass
that were
star-forming
at the time of first infall have been
quenched by the present, while
essentially all massive satellites
that were
initially
star-forming
have been quenched.
The dependence of the fractions
$fq_{\rm infall}$
on $z_{\rm infall}$ for different stellar mass ranges
only changes slightly for galaxies within haloes
of mass
$M_{\rm halo} \gtrsim 10^{13.2}\,{\rm M}_{\odot}$,
as can be seen from comparing the middle and right panels of
Fig.~\ref{fig:fqzinfall_zinfall}.
This lack of strong dependence of the quenched fractions
on main host halo mass
for high-mass
haloes
is consistent with the behaviour that emerged from the analysis
of the radial dependence of the quenched fraction presented in Paper I (see fig.
12).
The major difference
appears
for galaxies
in the two highest stellar mass bins
that have been accreted by
smaller mass haloes, where
$fq_{\rm infall}$ decreases by $\sim 0.15$,
regardless of $z_{\rm infall}$,
with respect to those
fractions
typical of more massive haloes.
This is the result
of selecting galaxies by the same present-day stellar mass combined
with the different stellar-mass growth rate
experienced by
galaxies
of a given $z=0$ stellar mass
that are
located in different environments \citep{Guglielmo15};
cluster galaxies form their stars sooner than those in the field,
giving rise to a more pronounced slope of the decline of the cosmic
SFR
density with redshift in clusters.
Thus,
galaxies selected within a given stellar-mass range at $z=0$ that have been accreted by lower mass haloes have smaller
stellar mass at a given time of infall
\citep[see fig. 12 in][]{behroozi13c},
which implies
lower efficiency of mass quenching prior to infall with the
consequent lower values of $fq_{\rm infall}$.
\subsection{Implications of the fraction of currently passive satellites already quenched at $z_{\rm infall}$}
\label{sec:implications}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig3.pdf}
\caption{Fraction of
$z=0$ quenched satellite galaxies that are
quenched at first infall as a function of present-day stellar mass for galaxies
residing within main host haloes of different mass (different line styles).
Error bars are obtained from the propagation of errors of the quotient.
The fractions $fq_{\rm infall}/fq_{\rm z0}$ increase with stellar mass, meaning
that more massive galaxies are more likely to be quenched
as centrals through mass quenching processes.
For intermediate mass satellites,
the ratio $fq_{\rm infall}/fq_{\rm z0}$
is larger for lower mass haloes.
}
\label{fig:fq-at-infall-z0-ms-mh}
\end{figure}
The fraction of currently quiescent satellites
that are already quenched at first infall
can be estimated
by combining
the information provided by the
$z=0$
quenched fraction
($fq_{\rm z0}$, Fig.~\ref{fig:fqz0_zinfall}) and the quenched fraction
at first infall ($fq_{\rm infall}$, Fig.~\ref{fig:fqzinfall_zinfall}).
Thus, this fraction is calculated simply
by the ratio between
$fq_{\rm zinfall}$
and $fq_{\rm z0}$.
Fig.~\ref{fig:fq-at-infall-z0-ms-mh} shows the ratio
$fq_{\rm infall}/fq_{\rm z0}$ as a function of stellar mass in bins of
present-day main host halo mass.
This ratio increases with stellar mass, meaning
that more massive galaxies are more likely to be quenched
as centrals through mass quenching processes, consistent with
the conclusion inferred from
Fig.~\ref{fig:fqzinfall_zinfall}.
At fixed stellar mass (within the range
$M_{\star} [{\rm M}_{\odot}] \in [10^{10.1}, 10^{10.5}]$),
this ratio is larger for satellites residing in lower mass haloes.
The dependence of
the ratio $fq_{\rm infall}/fq_{\rm z0}$
on stellar mass and halo
mass is shown in Fig.~\ref{fig:fq-at-infall-z0-ms-mh}. Our model predictions are
similar to W13
(see their fig. 10a), but the halo-mass dependence is less pronounced for \textsc{sag}~model galaxies.
Their results are obtained from
the comparison of
galaxy group/cluster catalogues
from the Sloan Digital Sky Survey (SDSS) Data Release 7 with
a galaxy catalogue
constructed from a high-resolution
cosmological simulation
by applying the subhalo abundance matching (SHAM) technique
\citep{Conroy06, Vale06}.
From
the trend with halo mass,
W13
conclude that quenching prior to first infall is more important in
galaxies accreted by
lower mass host haloes,
finding this fact reasonable considering that
they
fell in more recently.
However,
such a trend can be interpreted in a different way if
we take into account that,
for quenched satellites
at a fixed present-day stellar mass and epoch,
$fq_{\rm infall}$ is lower in less massive haloes because of the lower stellar mass at the time of first infall.
As a consequence of this, mass quenching is less effective prior to infall for galaxies accreted by less massive haloes, as we discuss in Section~\ref{sec:fqinfall}.
The ratio $fq_{\rm infall}/fq_{\rm z0}$ becomes larger
for galaxies residing in lower mass haloes because
$fq_{\rm z0}$ is also smaller for them (although not shown in Fig.~\ref{fig:fqz0_zinfall}).
This is a result of the milder effect of RPS (found to be the dominant environmental process among those considered in \textsc{sag}) exerted by
lower mass haloes
\citep[also see figs.~14 and 15 of][]{tecce10}.
Hence, from the analysis of model results,
we can assert that only considering the fraction
$fq_{\rm infall}/fq_{\rm z0}$ in the way presented in
Fig.~\ref{fig:fq-at-infall-z0-ms-mh}
is misleading:
there are many variables playing an important role in the
process of SF quenching.
With the additional information provided by
the fractions $fq_{\rm z0}$ and $fq_{\rm infall}$, separately,
we conclude that
mass quenching prior to infall is less efficient for galaxies accreted by
less massive haloes, and quenching after infall is also less efficient in smaller
haloes.
\subsection{Relative role of environmental and mass quenching after first infall}
\label{sec:mass_quench_after}
In order to disentangle the role of environmental
and mass quenching
after first infall for satellites
with different present-day stellar mass,
we compare
their evolution to a control sample of central galaxies that never become satellites.
We consider satellites residing in main host haloes with
present-day mass $M_{\rm halo} \geq 12.3 \,{\rm M}_{\odot}$.
They are selected according to their $z=0$ stellar mass, keeping those that
have been accreted at $z_{\rm infall}=1$ and were
star-forming
at that time.
The control sample of
star-forming
central galaxies at $z=1$ is built with the requirements that they
have the same stellar mass as satellites at infall and reside within main host haloes with masses
in the range of subhalo masses that characterise the satellite galaxies at the
moment of accretion.
These stellar and subhalo mass bins
are defined by
tracking satellites back in time
till $z=1$ and
calculating the
$10$th and $90$th
percentiles of these mass distributions.
This information is detailed in the legends of
Figs.~\ref{fig:sSFRevolSatCen} and~\ref{fig:lumBHevolSatCen} that show the evolution of the sSFR and the BH luminosity (given by eq.~\ref{eq:lumbh}), respectively, for
both galaxy populations.
Lines represent median values and the corresponding dashed areas depict the $10$th and $90$th percentiles.
The upper panel corresponds to galaxy populations determined by
satellites with present-day stellar mass within the two less massive ranges considered in Fig.~\ref{fig:fqz0_zinfall} and
Fig.~\ref{fig:fqzinfall_zinfall}, i.e. $M_{\star} [{\rm M}_{\odot}] \in [10^{9.7}, 10^{10.5}]$, while the
bottom panel presents those determined by satellites with $z=0$ stellar mass within the two more massive ranges considered, i.e.
$M_{\star} [{\rm M}_{\odot}] \in [10^{10.5}, 10^{11.3}]$.
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{Fig4.pdf}
\caption{
Comparison of the evolution of the sSFR from
$z=1$ of current satellite galaxies that have been accreted at $z_{\rm infall}=1$ and are
star-forming
at that time (blue dashed line) with that of a control sample built with current central galaxies that are also
star-forming
at $z=1$ (red solid line).
Lines represent median values and the corresponding
shaded
areas depict the $10$th and $90$th percentiles.
Satellites residing in main host haloes with
present-day mass $M_{\rm halo} \geq 12.3 \,{\rm M}_{\odot}$ are considered. They are selected according to their $z=0$ stellar mass and are tracked back till $z=1$ to define the ranges of stellar mass and subhalo mass
that characterise them at the time of first infall.
By
calculating
the
$10$th and $90$th
percentiles of the mass distributions, we define the $z=1$ stellar mass and halo mass bins
that are considered to select the population of central galaxies, as indicated in the legend.
This selection guarantees that satellites and centrals have similar initial stellar mass and (sub)halo mass at $z=1$.
{\em Top panel:} Galaxy populations determined by
satellites with present-day stellar mass within the two less massive ranges considered in Fig.~\ref{fig:fqz0_zinfall} and
Fig.~\ref{fig:fqzinfall_zinfall}, i.e. $M_{\star} [{\rm M}_{\odot}]) \in [10^{9.7}, 10^{10.5}]$.
{\em Bottom panel:}
Galaxy populations determined by satellites with
$z=0$
stellar mass within the two more massive ranges considered, i.e.
$M_{\star} [{\rm M}_{\odot}]) \in [10^{10.5}, 10^{11.3}]$.
}
\label{fig:sSFRevolSatCen}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.0\columnwidth]{Fig5.pdf}
\caption{
Same as Fig.~\ref{fig:sSFRevolSatCen} but for the evolution of the BH luminosity.
}
\label{fig:lumBHevolSatCen}
\end{figure}
The sSFR evolution of low-mass satellites is clearly distinct from that of low-mass centrals (top panel of Fig.~\ref{fig:sSFRevolSatCen});
the decline of the sSFR towards lower redshifts is more pronounced for the former than for the latter.
This means that low-mass satellites become passive mainly because of the action of environmental processes,
reinforcing the conclusion inferred from the analysis of
Figs.~\ref{fig:fqz0_zinfall} and
~\ref{fig:fqzinfall_zinfall}.
The effect of TS on the DM subhaloes of satellites and the larger effects of RPS suffered by low-mass satellites (see Fig. 14 in Paper I) contribute to reduce the mass of their hot gas halo with the consequent reduction of the BH luminosity (top panel
of Fig.~\ref{fig:lumBHevolSatCen}), which is directly related to the efficiency of AGN feedback.
The hot halo mass of centrals keeps growing towards lower redshifts as a result of cosmological gas accretion during halo-mass growth. Thus, according to equation~\ref{eq:BHgrowth-gascool}, the BH-growth rate through gas cooling also increases giving place to higher BH luminosity at lower redshifts. However, for low-mass centrals, these luminosities are not high enough to considerably reduce gas cooling ($L_{\rm BH}\sim 10^{48}\,{\rm erg}\,{\rm s}^{-1}$ at $z=0$). Consequently, this galaxy population remains mostly
star-forming
at the present, as indicated by the solid line and associated shaded area in the top panel of Fig.~\ref{fig:sSFRevolSatCen}.
High-mass central galaxies are characterised by BH luminosities that are between $\sim 2$ (at $z=0$) and $\sim 3$ (at $z=1$) orders of magnitude larger than for low-mass ones (Fig.~\ref{fig:lumBHevolSatCen}). Thus, AGN feedback becomes an efficient mass quenching process for more massive centrals,
as evident
from the abrupt decline of
the sSFR shown in the bottom panel
of Fig.~\ref{fig:sSFRevolSatCen}.
The stellar mass above which AGN feedback leads to SF quenching ($M_{\star} \gtrsim 10^{10.5}\,{\rm M}_{\odot}$) and the halo mass in which they reside ($M_{\rm halo}\gtrsim 10^{12}\, {\rm M}_{\odot}$)
are in agreement with the redshift-independent characteristic halo and stellar mass scales discussed by \citet{Henriques18}.
The evolution of the sSFR and BH
luminosity
of high-mass central galaxies
are also closely followed
by high-mass satellites,
although with a lower normalization.
Here, environmental processes contribute to the decline of star formation.
These results demonstrate that
mass quenching keeps
playing a major role in the
decline
of SF after high-mass galaxies become satellites.
\section{Satellite quenching times}
\label{sec:tq}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig6.pdf}
\caption{Mean values of quenching time-scale, $t_{\rm q}$,
for satellite galaxies that are quenched at $z=0$ and
were
star-forming
at infall
as a function of their stellar mass.
Different lines depict satellites residing within main host haloes of different
present-day mass as indicated in the legend;
shaded areas represent the corresponding $1\,\sigma$ standard deviation around
the mean.
}
\label{fig:tq-mstar-mhalo}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=2.0\columnwidth]{Fig7.pdf}
\caption{
Probability density function of quenching times, $t_{\rm q}$,
for the satellite population grouped in three stellar mass ranges
(different panels). Different style lines correspond to galaxies
with different times of first infall which are comprised within a range of $\approx 1\,{\rm Gyr}$ around
four selected values of $z_{\rm infall}$
($z=0.25, 0.5, 0.8$ and $1$) as indicated in the legends,
which correspond to redshift ranges
$z \in [0.2,0.3]$, $[0.45,0.6]$, $[0.7,0.9]$, $[0.95,1.2]$.
All galaxies within main host haloes of present-day mass
$M_{\rm halo} \geq 10^{12.3}\,{\rm M}_{\odot}$
are considered.
The most massive satellites are characterised by the lowest quenching
times ($t_{\rm q} \approx 0.05 - 2 \, {\rm Gyr}$).
The distribution of $t_{\rm q}$
peaks at $\approx 2 \, {\rm Gyr}$ for both satellites of intermediate mass and the lowest mass satellites that have been accreted
more recently ($z_{\rm infall}\in [0.2-0.3]$).
Those
low-mass
satellites that have been accreted earlier
are characterised by quenching times comprised within a broad range ($t_{\rm q} \approx 1.5 - 6 \, {\rm Gyr}$).
}
\label{fig:tq}
\end{figure*}
Now we study the quenching times of those
$z=0$
passive satellites that
are
star-forming
at time of first infall, i.e. those that quench as satellites
by the contribution of both environmental and mass quenching.
We estimate the quenching time, $t_{\rm q}$,
considering
the period of time elapsed since
$z_{\rm infall}$ and
the moment in which the galaxy becomes passive,
identifying the snapshot
in the simulation in which
the galaxy sSFR falls below the threshold
$10^{-10.7}\,{\rm yr}^{-1}$.
This quenching time is defined
in a similar way as in W13.
We present the relation between $t_{\rm q}$ and stellar mass
for satellites in main host haloes of different present-day mass in
Fig.~\ref{fig:tq-mstar-mhalo}.
The trends found are quite similar
to those shown by W13,
with
low-mass
galaxies
characterised by higher quenching times
than
more massive ones,
and there being
almost no secondary dependence on host halo mass.
Mean values of $t_{\rm q}$ predicted by our model are
$\approx 4-5\,{\rm Gyr}$ for $M_{\star} \approx 10^{10} \,{\rm M}_{\odot}$
and $\approx 2-3\,{\rm Gyr}$ for $M_{\star} \approx 10^{11} \,{\rm M}_{\odot}$.
The scatter around these mean values is quite large, showing that
low-mass
galaxies can achieve
quenching times as long as $6\,{\rm Gyr}$, consistent with W13.
The stellar mass dependence of the quenching time could be subject to bias if we consider
the anticorrelation between the sSFR and stellar mass for galaxies
in the star-forming main sequence.
By defining a quenching time as the time it takes to drop below an absolute value of sSFR, high-mass satellites would take less time to become passive because they have less distance to cover in sSFR space.
However, the main sequence of star-forming satellites at different redshifts in our model
are rather flat (see fig. 7 in Paper I, thin dashed lines). Therefore,
the stellar mass dependence of the quenching time that we find is a robust result.
In order to confirm this,
we have made the test of estimating the quenching time for galaxies defined as passive when they satisfy the condition
${\rm sSFR} < 10^{-12}\,{\rm yr}^{-1}$, that is, dropping the threshold adopted to define a galaxy as passive.
In this case, quenching times
for low-mass satellites become longer
by $\approx 0.8\,{\rm Gyr}$ but, contrary to expectations, quenching times for high-mass galaxies remain unaffected by the lower threshold, making the stellar-mass dependence of the quenching time even more pronounced. This
is explained by the
evolution of the sSFR for low- and high-mass satellites shown in
Fig.~\ref{fig:sSFRevolSatCen}.
Low-mass satellites quench their SF in a more gradual way
than high-mass
satellites. The latter are abruptly quenched reaching
very low values of sSFR,
so they are able
to satisfy the more stringent condition imposed for being passive
in the same short quenching time as
required when the original threshold was applied.
Regarding the halo mass independence, W13 argue that it reflects the fact that
$t_{\rm q}$ is measured
since time of first infall, which involves preprocessing in smaller haloes;
the fraction of quiescent satellites that start quenching in another halo
varies from $\approx 0.15$ to $\approx 0.5$ depending on current stellar mass and host halo mass
(see also \citealt{delucia2012}).
However, this lack of sensitivity of quenching times
to main host halo mass is also found by \citet{Oman16} even when
they isolate environmental effects of the most recent host halo (within the range
$10^{13}-10^{15}\,{\rm M}_{\odot}$)
by measuring these time-scales since galaxies cross $2.5\,$$r_\text{vir}$~
of the current host (according to their definition of time of infall),
thus avoiding the time invested in pre-processing within other subhaloes.
\citet{Oman16} find that quenching occurs after a delay time $\approx 3.5-5\,{\rm Gyr}$ since infall
with quenching time-scales slightly shorter for higher mass galaxies.
Although the stellar mass dependence is weaker than the one found by W13,
time-scales obtained from these two works are of the same order of magnitude
once corrections for different definitions of infalling time are taken into account
\citep{Oman16}.
The general good agreement of \textsc{sag$_{\beta1.3}$}~quenching times with values and
trends found by W13 and \citet{Oman16} is encouraging
and supports other related predictions of our model.
Fig.~\ref{fig:tq} shows the probability density functions
\footnote{Probability density function at the bin, normalized
such that the integral over the range is equal to unity.}
of $t_{\rm q}$,
for satellites in
three of the
stellar mass ranges analysed previously
(different panels),
whose
redshifts of first infall are comprised within a range of
$1\,{\rm Gyr}$ around
$z_{\rm infall}=0.25, 0.5, 0.8$ and $1$ (different line styles).
Since there is no halo mass dependence of $t_{\rm q}$,
all galaxies within main host haloes of present-day mass
$M_{\rm halo} \geq 10^{12.3}\,{\rm M}_{\odot}$
are considered.
The most massive satellites are characterised by the lowest quenching
times, with values of $t_{\rm q} \approx 0.05 - 2 \, {\rm Gyr}$.
Satellites of intermediate mass have a distribution of $t_{\rm q}$
that peaks at $\approx 2 \, {\rm Gyr}$.
This is also the case for the lowest mass satellites that have been accreted
more recently ($z_{\rm infall}\in [0.2-0.3]$).
At first glance one could think that the latter result is biased by
construction,
since we are considering a set of galaxies that are
star-forming
at infall
and have quenched
at $z=0$, so their time-scale for quenching is restricted.
However, the fact that we do
find satellites satisfying all these criteria means that there is an efficient
mechanism acting at low redshifts that is able to quench SF of satellites.
The quenching times of those low-mass satellites that have been accreted earlier cover a broad range from $\approx 1.5 \, {\rm Gyr}$ to $\approx 6 \, {\rm Gyr}$, with a slight preference for the longer times.
The combination of the previous analysis
with the information on the relative proportion of galaxies accreted at different times allows us to explain the anti-correlation of the average values of $t_{\rm q}$ with stellar mass (Fig.~\ref{fig:tq-mstar-mhalo}).
For a given stellar mass range, the mean values of
$t_{\rm q}$
are estimated considering satellites with all possible values of $z_{\rm infall}$.
The majority of
$z=0$
passive satellites in the two highest
stellar mass bins
($M_{\star} > 10^{10.1} \,{\rm M}_{\odot}$)
have
been accreted at low and intermediate redshifts
($z_{\rm infall}\lesssim 0.5$),
whereas
the bulk of the current lowest mass galaxies considered here
have become satellites much earlier.
Thus, early accreted low-mass
galaxies characterised by long quenching times outnumber the recently accreted ones, which reach shorter quenching times. Consequently,
the mean values of $t_{\rm q}$ become higher for less massive satellites.
\subsection{Discussion}
\label{sec:discussion}
Environmental effects such as~RPS and~TS
depend strongly on the orbit of the satellite
\citep[see e.g.][]{vollmer2001}. A galaxy on a more radial orbit, after passing
through the denser regions of the halo centre, will experience a strong~RP ($10^{-10}\,h^2\,{\rm dyn}\,{\rm cm}^{-2}$)
earlier than a galaxy which is initially in a slowly decaying circular orbit; it will also experience alternating periods of strong and weak~RP
($10^{-12}\,h^2\,{\rm dyn}\,{\rm cm}^{-2}$),
as its orbit
within the halo takes it from the dense central region to the lower-density
outskirts and back again. Another galaxy which is instead in a slowly decaying
circular orbit will experience an ever increasing~RP as it sinks towards denser
regions \citep{bdl2008, tecce10}. In the case of~TS, the mass loss also
depends strongly on orbital circularity
\citep[e.g.][]{tb2001,tb2004,zb2003,gan2010}, with most of the subhalo mass
being lost during pericentric passages.
The long quenching time-scales $t_{\rm q}$
of
low-mass
satellites can be attributed to long dynamical friction time-scales
that keep
low-mass
satellites in the outskirts of their main host haloes where RPS
is less efficient.
Effectively, from the analysis of cosmological simulations,
\citet{Quilis17} show that massive satellites
($M_{\star} > 10^{10}\,M_{\odot}$) are found at short halo-centric distances
at low redshifts ($z\lesssim 0.5$),
whereas
smaller systems are mainly located in the external
regions. This dichotomy in the stellar mass dependence of the average radial position is less pronounced at higher redshifts, with all galaxies being
quite evenly distributed throughout the groups/clusters.
It is worth noticing that low-mass satellites accreted at $z_{\rm infall} \gtrsim 0.5$ can have any quenching time within the
range $\approx 1 - 6$ Gyr (Fig.~\ref{fig:tq}). This result indicates that,
depending on their conditions at infall,
many satellites are likely to have more eccentric orbits that bring them close
to the halo centre characterised by higher values of RP and quench in short time-scales.
An important aspect to take into account on top of galaxy dynamics
is the natural halo mass growth which is accompanied by an increase in the
efficiency of RPS,
giving place to faster quenching at lower redshifts.
This might account for the short quenching times
($t_{\rm q}\approx 1-2\,{\rm Gyrs}$)
of high-mass satellites and recently accreted
low-mass
ones, as well as for the long quenching times of
those satellites accreted at earlier times.
\subsection{Comparison with other works}
The values and the stellar mass dependence of the quenching times $t_{\rm q}$ obtained from our model are consistent with those estimated by W13. This trend with stellar mass
is also roughly recovered
by \citet{Hirschmann14} who infer these time-scales from the requirement that
their predicted quiescent fractions become consistent with observations,
since the semi-analytic model they use \citep{guo11}
significantly overestimates the quiescent fraction of satellites and underestimate
that of centrals for the stellar mass and densities considered.
From the analysis performed on SDSS and
3D-HST/CANDELS data,
\citet{Fossati17} estimate quenching time-scales of passive satellites selected according to the their position in the
rest-frame UVJ color-color diagram.
Their definition of quenching time is similar to the one adopted in our work, i.e. the time elapsed since first infall till the satellite becomes passive.
These times were obtained from mock
catalogues constructed from a semi-analytic model \citep{henriques_2015} that match the number density and
redshift uncertainty of observed galaxies.
For quenched satellites at $z=0$ residing in low-mass haloes ($M_{\rm halo} < 10^{13}\,{\rm M}_{\odot}$), they
find quenching times of the order of $\approx 7 \,{\rm Gyr}$. The quenching times are lower ($\approx 5 \,{\rm Gyr}$) for satellites of more massive haloes, which the authors attribute to very massive haloes included
in the SDSS dataset that exert larger environmental effects.
Besides, for a given halo mass bin, these times have almost no dependence on stellar mass (see their fig.~21).
Our results, obtained for a population of
$z=0$
quenched satellite galaxies,
are not consistent with their findings at $z=0$
since the quenching times for our model galaxies show a much weaker dependence on halo mass, despite the fact that our halo-mass sample contains high mass clusters, and a clear trend with stellar mass.
The different approaches used to derive quenching time-scales make the identification of the source of discrepancies difficult.
On the other hand,
the values of $t_{\rm q}$ obtained from our model and the stellar-mass and halo-mass trends they follow are consistent with the quenching times
obtained
for high redshift galaxies in the 3D-HST/CANDELS sample. For $z\sim 0.7-1.5$, \citet{Fossati17} find values of the order of $\approx 4-5 \,{\rm Gyr}$ for low mass galaxies
and $\lesssim 2\,{\rm Gyr}$ for
the most massive ones, with negligible dependence on halo mass within the considered range ($M_{\rm halo}\lesssim 10^{14}\,{\rm M}_{\odot}$).
The latter values are larger than the
quenching time-scales found for clusters galaxies with $M_{\star} \gtrsim 10^{10.5}\,{\rm M}_{\odot}$ by \citet{Foltz18}, which are $\approx 1.5\, {\rm Gyr}$ and $\approx 1.24\,{\rm Gyr}$
at $z\sim 1$ and $z\sim 1.5$, respectively;
they classify galaxies as passive or
star-forming
according to
dust-corrected rest-frame colours derived from spectral energy distribution
fitting assuming the
delay-then-rapid quenching scenario (W13) to constraint the quenching time-scales.
A fair comparison with results from these works requires samples of model quenched satellites selected at high redshifts, which will be addressed in a near future.
\section{The delay-then-fade quenching scenario}
\label{sec:quenchscenario}
From the requirement of producing the correct sSFR distribution,
W13 propose the delay-then-rapid quenching scenario,
in which the times for the onset of SF quenching depend on the satellite mass
($t_{\rm q,delay} \approx 2 - 4\,{\rm Gyr}$) and, once started,
the quenching takes place rapidly
with SFR declining exponentially
with an e-folding time
$\tau_{\rm Q,fade} \approx 0.8 \, {\rm Gyr}$ for $M_{\star} \approx 10^{9.7} \,{\rm M}_{\odot}$ and
$\tau_{\rm Q,fade} \approx 0.2 \, {\rm Gyr}$ for $M_{\star} \approx 10^{11.3} \,{\rm M}_{\odot}$.
In the delay phase of this two-stage model for satellite's SFR evolution, the SFR of an accreted star-forming galaxy is assumed to diminish gradually after infall in a similar way to the SFR of a central galaxy of the same stellar mass as the satellite at infall.
The quenching time-scales inferred by \citet{Oman16}
also involve
a delay time until
the onset of quenching after infall ($\approx 3.5-5\,{\rm Gyr}$) and a short time-scale for the fading of SF
($\lesssim 2\, {\rm Gyr}$). The latter was estimated by considering the time required by their galaxies to reach the fraction of passive satellites observed within galaxy clusters.
From Fig.~\ref{fig:sSFRevolSatCen}, we can see that the evolution of the sSFR of satellite galaxies in our model deviates
from that of centrals some time after infall, and the decline of the sSFR of satellites becomes faster. This behaviour seems to be consistent with a two-stage model for the SFR evolution of satellites. However, it is not clear which physical processes are
in action
in each of the two phases of SF quenching.
The first stage in which the SFR of satellites evolves similarly to that of
centrals (`delay phase') seems to be associated with the period of time required by the hot gas cooling to become inefficient after infall \citep[e.g.][]{Schawinski14}.
This transition experienced by the gas cooling efficiency might define the beginning of
the second stage in which the SFR of satellites declines more abruptly than that of centrals (`fading phase') because the cold gas reservoir
is no longer replenished by gas cooling from the hot halo. This phase might involve cold gas disc consumption through SF and/or removal through SN
feedback and/or RPS, ending when the galaxy becomes passive.
In order to identify the physical mechanisms that play a relevant role in each of the two phases of the SFR decline,
we analyse the mass content of the hot and cold gas reservoirs of $z=0$ passive satellites that are star-forming at infall considering different moments along their lifetime.
Fig.~\ref{fig:fhot} shows the mean values of the fraction of hot gas with respect to the total baryonic mass, $f_{\rm hot}$,
at infall (dashed-dotted line), at the moment of quenching, i.e, when their sSFR drops below the threshold adopted ($10^{-10.7}\,{\rm yr}^{-1}$; dashed line), and at $z=0$ (solid line),
as a function of their $z=0$ stellar mass.
At time of first infall, $f_{\rm hot}\approx 0.8$ for any $z=0$ stellar mass; this fraction reflects the condition of central galaxies just prior to infall.
When high-mass satellites become passive, this fraction remains quite high ($f_{\rm hot} \approx 0.7$). For low-mass satellites, this fraction becomes lower ($f_{\rm hot} \approx 0.4$) but still consistent with the presence of a hot halo; the corresponding $1\sigma$ scatter indicates that most of low-mass passive satellites have more than $20$ percent of hot gas at the moment of SF quenching.
We find that only $\approx 15$ and
$\approx 5$ percent of satellites with stellar mass $M_{\star} \approx 10^{10} \,{\rm M}_{\odot}$ and
$M_{\star} \approx 10^{11} \,{\rm M}_{\odot}$, respectively, quench their star formation after the hot gas has been reduced to less than $10$ percent of the total baryonic mass. Most of them have experienced at least one event of RPS of the cold gas before becoming passive.
These satellites
are galaxies that have been accreted much earlier than the rest of the satellite population.
The mean values of their redshift of first infall
vary from $z_{\rm infall} \approx 0.86$ to $\approx 0.94$, depending on the $z=0$ stellar mass range considered, while the corresponding values for those galaxies that quench their SF before hot gas depletion are comprised within the range $z_{\rm infall} \approx 0.52-0.78$. Therefore, the quenching times of this particular set of satellites
are longer ($t_{\rm q}\approx 5-6\,{\rm Gyr}$) than those typical of the bulk of the satellite population (see Figs.~\ref{fig:tq-mstar-mhalo} and~\ref{fig:tq}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig8.pdf}
\caption{
Mean fraction of hot gas with respect to the total baryonic mass at different moments along the lifetime of $z=0$ passive satellites that are active at infall as a function of their $z=0$ stellar mass. These moments are: time of first infall ($z_{\rm infall}$; dashed dotted blue line), moment of SF quenching ($z_{\rm q}$; dashed green line), and the present epoch (z=0). The corresponding dashed areas denote the $1\sigma$ standard deviation.
}
\label{fig:fhot}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{Fig9.pdf}
\caption{
Mean fraction of cold gas with respect to the sum of cold gas and stellar mass at different moments along the lifetime of $z=0$ passive satellites that are active at infall as a function of their $z=0$ stellar mass. These moments are: time of first infall ($z_{\rm infall}$; dashed dotted blue line), moment of SF quenching ($z_{\rm q}$; dashed green line), and the present epoch (z=0). The corresponding dashed areas denote the $1\sigma$ standard deviation.
}
\label{fig:fcold}
\end{figure}
One important implication of the aforementioned results is that
the complete removal of the hot gas halo is not a requirement neither for the onset of SF quenching (beginning of a faster decline of the SFR) nor for the SF quenching itself.
Gas cooling, which is linked to the properties of the hot gas halo, namely, mass, metallicity and density profile \citep[e.g.][]{springel2001},
may become inefficient even when a substantial amount of hot gas is still available.
The hot gas halo kept by galaxies after infall is mildly reduced by RPS for high-mass satellites (see fig.~14 in Paper I), but gas cooling is considerably suppressed by the action of AGN feedback (equation~\ref{eq:cool}). This reduction of the gas cooling efficiency for more massive star-forming galaxies becomes evident in
parsec-scale hydrodynamical simulations \citep{Armillotta16}.
This is consistent with the dominance of mass quenching over environmental quenching in the SFR decline of high-mass satellites (see Figs.~\ref{fig:sSFRevolSatCen}~and \ref{fig:lumBHevolSatCen}).
In low-mass satellites,
gas cooling
may become inefficient once the hot gas reservoir has been sufficiently reduced, although not necessarily depleted, by the stronger effect of RPS and gas cooling itself.
Further reduction of the hot gas halo after SF quenching is the result of the action of RPS that contributes to determine the fractions of hot gas achieved at $z=0$ ($f_{\rm hot} \approx 0.2$ for $M_{\star} \approx 10^{10}\,M_{\odot}$ and $f_{\rm hot} \approx 0.6$ for $M_{\star} \approx 10^{11}\,M_{\odot}$, Fig.\ref{fig:fhot}). Clearly, this effect dominates over any possible replenishment with gas reheated by SN feedback.
Note that satellites do not accrete hot gas cosmologically in our model, a common feature in SAMs \citep[e.g.][]{henriques_2015,Stevens17}.
Fig.~\ref{fig:fhot} shows that the bulk of the $z=0$ passive satellite population experiences SF quenching that is not associated to the action of RPS of the cold gas disc,
which in our model takes place only when this component is not longer protected by the hot gas halo.
This result is supported by qualitatively similar conclusions obtained from
cosmological chemodynamical simulations \citep{Kawata08} and the
semi-analytic model \textsc{dark sage} \citep{Stevens17}. This SAM
also includes the effect of RPS on both the cold-gas disc and the hot-gas halo, with the former being shielded by the presence of the latter, although the corresponding implementations vary with respect to those in our model.
Fig.~\ref{fig:fcold} shows
the mean values of the fraction of cold gas with respect to the sum of cold gas and stellar mass, $f_{\rm cold}=M_{\rm cold}/(M_{\rm cold}+M_{\star})$, for the same satellite population and stages as in Fig.~\ref{fig:fhot}. Mean values at time of first infall vary from $f_{\rm cold}\approx 0.6$ for $M_{\star} \approx 10^{10}\,M_{\odot}$ to $f_{\rm cold} \approx 0.4$ for $M_{\star} \approx 10^{11}\,M_{\odot}$; the 1-$\sigma$ dispersion is larger for more massive satellites, which indicates that some high-mass galaxies are already gas-poor when they are accreted, with levels of SF that are likely close to the limit of quenching.
This is a clear sign of the action of mass-quenching processes on high-mass galaxies prior to infall (see Fig.~\ref{fig:fqzinfall_zinfall}).
This also explains the lower values of the quenching times, $t_{\rm q}$, of more massive satellites (see Figs.~\ref{fig:tq-mstar-mhalo}~and \ref{fig:tq}). All satellites, regardless of their $z=0$ stellar mass, have low cold gas fractions by the time they become passive ($f_{\rm cold} \approx 0.1-0.2$); their cold gas
has been gradually consumed by star formation and/or removed through outflows produced by SN feedback.
The fraction $f_{\rm cold}$
continues decreasing beyond
the quenching time $t_{\rm q}$
up to the present epoch by the remaining modest star formation taking place
and/or the action of
RPS on the cold gas disc if it is no
longer shielded by the hot gas halo.
Since gas cooling efficiency plays a key role in the SFR decline, we estimate the rates of gas cooling at specific moments, following the analysis applied to the hot and cold gas fractions (see Figs.~\ref{fig:fhot} and~\ref{fig:fcold}).
Satellites of any $z=0$ stellar mass have, on average, gas cooling rates $\lesssim 5\,{\rm M}_{\odot}\,{\rm yr}^{-1}$ when becoming passive.
This result means that there is not a sharp cut-off in the cold gas supply.
Instead, the gas cooling rate becomes progressively lower departing from values within the range $\approx 30-40\,{\rm M}_{\odot}\,{\rm yr}^{-1}$ when star-forming galaxies are accreted.
From the estimation of the evolution of median gas cooling rate for the same satellite population considered in Fig.~\ref{fig:sSFRevolSatCen}, we can see (not shown here)
a transition to a faster reduction of the gas cooling rate at $\approx 20\,{\rm M}_{\odot}\,{\rm yr}^{-1}$, which is connected with the beginning of the fading phase where the SFR decline is faster.
Thus, the results of our model indicate that the fading phase begins when the gas cooling rate has been reduced by $\approx 50$ percent with respect to its value at infall.
Our results support a two-stage quenching scenario for satellite galaxies, where the length of time of the delay and fading phases are comprised within the time $t_{\rm q}$, which characterises the whole quenching process.
We estimate the length of time of the fading phase, $t_{\rm q,fade}$, as
the period of time
elapsed since the gas cooling rate is reduced to half the value it has at infall
until the satellite becomes passive.
We find that $t_{\rm q,fade}\approx 1.5\,{\rm Gyr}$ for satellites of any $z=0$ stellar mass, with a very mild trend to be shorter for more massive galaxies.
This result is representative of a large percentage of the satellite galaxies in our sample. The cooling-rate-based condition of transition from the delay phase to the fading one is satisfied by
$\approx 90$ percent of satellites with stellar mass $M_{\star} \approx 10^{10.5} \,{\rm M}_{\odot}$. This percentage decreases for satellites with both lower and higher stellar masses, and reaches values as low as $\approx 60$ and $\approx 40$ percent, respectively. Those satellites that do not fulfil this requirement are galaxies that already have very low cooling rates when they are accreted ($\approx 5\,{\rm M}_{\odot}\,{\rm yr}^{-1}$).
For those satellites that go through both the delay and fading phase, we estimate the length of time of the delay phase, $t_{\rm q,delay}$, from the difference between the mean values of
$t_{\rm q}$ (Fig.~\ref{fig:tq-mstar-mhalo}) and $t_{\rm q,fade}$ for each stellar mass range.
The delay time varies from $t_{\rm q,delay}\approx 3\,{\rm Gyr}$ for low-mass satellites to $t_{\rm q,delay}\approx 1\,{\rm Gyr}$ for high-mass ones.
Taking into account the scatter in the estimation of these times,
we find that the values and stellar mass dependence of $t_{\rm q,delay}$ are in agreement with those obtained by W13. Mean values of $t_{\rm q,fade}$ are also consistent with the e-folding time over which the SFR
fades, $\tau_{\rm Q,fade}$, estimated by W13,
although our model does not predict statistical significant lower values for more massive satellites.
We can conclude that the SF quenching of $z=0$ passive satellites can be described by a two-stage quenching scenario, characterised by a delay and a fading phase. The fading time is largely independent of $z=0$ stellar mass. It is shorter than the delay time for low-mass galaxies,
whereas the opposite situation occurs for high-mass ones. Therefore, we consider that the term `delay-then-rapid' proposed by W13 to dub the SF quenching scenario is not representative of the SF history of $z=0$ passive satellites with high stellar mass. Thus, we prefer to use the more
inclusive
term {\em delay-then-fade} to fairly describe all the possible situations.
\subsection{Discussion}
Based on the physical processes implemented in the \textsc{sag}~model (gas cooling, star formation, SN and AGN feedback, RPS of the hot and cold gas phases),
we have identified the relative impact of mass and environmental quenching on the SF history of $z=0$ passive satellites. We find that the SF quenching of satellites is well described by a delay-then-fade quenching scenario.
The rate of gas cooling from the hot halo plays a decisive role in the beginning of the fading phase. It is mainly determined by an internal process (AGN feedback) in high-mass satellites, and by an environmental process (RPS of the hot gas halo) in low-mass ones.
The cut-off of the cold gas replenishment
by gas cooling,
regardless of which physical
process may actually be responsible for it,
is named `strangulation'
\citep[e.g.][]{Peng15}.
It is important to introduce a note of caution regarding the meaning attributed to this term in different works in the literature.
The works that have introduced this concept \citep{larson80,balogh2000,balogh00b} consider that strangulation also involves the stripping of the hot gas halo of satellite galaxies, a convention followed in subsequent works \citep[e.g.][]{Kawata08}. The evidences that the hot gas removal takes place in a few Gyrs after accretion \citep{Rasmussen06,Vijayaraghavan15} justify the gradual starvation scenario implemented in \textsc{sag}~(see Sec.~\ref{sec:model}). However, in previous versions of our model, as well as of other SAMs \citep{kauffmann93,weinmann2006b, croton2006}, the modelling of the strangulation mechanism has been oversimplified by assuming that the hot gas halo is removed instantly when a galaxy becomes a satellite. Thus, in the context of SAMs, the term `strangulation' implies this crude formulation of the process.
\citet{Peng15} propose that strangulation is the primary mechanism for shutting down star formation.
They
find that the
mass-dependent metallicity difference between quiescent
and star-forming galaxies in SDSS
can be very well reproduced by a close-box model that assumes that
galaxies become passive after the cold gas supply is halted.
However, the results from our model suggest that strangulation,
in the general sense used by \citet{Peng15},
is a very strict condition to define the beginning of the SF fading because it is not necessary a complete suppression of gas cooling but only a more pronounced reduction of the cooling rates.
Besides, even assuming strangulation, they obtain
a typical time-scale for SF quenching of $\approx 4\,{\rm Gyr}$ (largely independent of stellar mass),
which is longer by $2.5\,{\rm Gyr}$ than our predictions for $t_{\rm q,fade}$.
This disagreement points to the fact that a simple close-box model
is not adequate to capture all the complex physical processes, both internal
and environmental, that affect the star formation activity of a galaxy.
Although this objective is better accomplished by \textsc{sag}, it is worth noticing that our model is not able to recover the observed metallicity difference between star forming and quiescent galaxies.
We find an increasing trend of stellar metallicity with stellar
mass in both cases, but not a significant difference between
the metallicities of these two populations at a given stellar mass.
This could be attributed to the fact that we discriminate galaxies in passive
and
star-forming
according to their sSFR instead of using their colours, as
in \citet{Peng15}. Therefore, many satellites classified as
star-forming
might be close
to be quenched and their metallicities would be
pretty similar to those of passive
galaxies. This issue deserves a deeper analysis.
The results of our model are in
line with
the scenario of `overconsumption' proposed
by \citet{McGee14}, based on the analysis of satellite quenching
times at a range of redshifts derived from observations. In this scenario, SF quenching of satellite galaxies is driven by secular outflows
once cosmological accretion of gas is halted, and
satellites become passive long before orbit-based gas stripping, such as RPS of the hot halo or of the cold gas disc, can have some impact on the quenching process.
The predictions of \textsc{sag$_{\beta1.3}$}~agree partially with the later aspect.
Although our model supports the fact that RPS of the cold gas plays
a secondary role or even has negligible
effect on the SF quenching, the RPS
on the hot gas halo
does play an important role since
this process contributes to reduce the hot gas reservoir of low-mass satellites which drives the decline of the gas cooling rate.
In the context of their overconsumption model, \citet{McGee14} find that
a constant mass-loading of the
wind can reproduce the evolution of the quenching time-scales of the delay phase as compiled from
the literature (including the local value given by W13,
which is of the same order of magnitude as the one inferred from \textsc{sag$_{\beta1.3}$}).
\citet{Oman16} show that
this simple model is in conflict with
the stellar mass dependence of the quenching time-scale. This conclusion is supported by our model,
which is able to reproduce such a dependence based on a
new feedback scheme
that involves a
redshift dependent mass-loading factor
(see eq.~\ref{eq:feedfire}),
a feature that has demonstrated to be crucial in reproducing several
observed galaxy properties (Paper I; \citealt{Collacchioni18}).
\section{Summary and conclusions}
\label{sec:conclu}
We have analysed a galaxy catalogue generated by applying
the updated version of our semi-analytic model of galaxy
formation
\textsc{sag}, described in detail in Paper I, on the cosmological
\textsc{MultiDark} $1\,h^{-1}\,{\rm Gpc}$ MDPL2 simulation with the aim
of contributing to our
understanding of the relative role of environmental and mass quenching
processes on satellite galaxies of different present-day stellar mass
hosted by DM haloes of different mass. We also estimate the
quenching time-scales involved.
The latest improvements implemented in \textsc{sag}~include
a robust model of environmental effects through the action of RPS and TS
coupled to the integration of the orbits of orphan satellites,
and a higher efficiency of SN feedback allowed by an explicit
redshift dependence of the reheated and ejected mass.
In Paper I, we have demonstrated that a variant of the model referred to as
\textsc{sag$_{\beta1.3}$}~allows us to
achieve good agreement with observational results for
several galaxy properties at both low and high redshifts.
In particular, the agreement of the predictions of \textsc{sag$_{\beta1.3}$}~with the observed fraction of currently passive satellites as a function of stellar mass, halo mass and halo-centric distance makes this model suitable to carry out the present work.
We analyse subsamples of galaxies
selected according to their stellar mass, main host halo mass and
time of first infall. Our main conclusions can be summarised as follows:
\begin{itemize}
\item
From the analysis of the relative importance of mass quenching and environmental quenching for local quenched satellites of different stellar mass, we find
$M_{\star} \approx 10^{10.5} \,{\rm M}_{\odot}$ to be the mass scale where mass quenching becomes important. This is also a characteristic mass scale for quenching in central galaxies \citep{Henriques18}.
Environmental processes,
on the other hand, dominate the SF quenching of
low-mass
satellite galaxies
($M_{\star} \lesssim 10^{10.1} \,{\rm M}_{\odot}$).
This picture is
consistent with
the results of
previous works
(\citealt{vandenBosch08}, \citealt{Peng10}, W13, \citealt{Lin14}, \citealt{Kawinwanichakij17}, \citealt{CochraneBest18}).
These conclusions
are inferred
from the following results:
\begin{itemize}
\item
Galaxies of any stellar mass that have been satellites for more than $\approx 8\, {\rm Gyr}$ ($z_{\rm infall} \gtrsim 1$) are characterised by
similarly
high
values of
$z=0$
quenched fractions
($fq_{\rm z0} \approx 0.8 - 0.95$).
Such high quenched fractions suggest that the time elapsed since first infall is enough for
the combined action of
mass and environmental processes to fully quench
satellite
galaxies.
\item
Low-mass
satellites
($M_{\star} \lesssim 10^{10.1} \,{\rm M}_{\odot}$)
have not suffered mass quenching while being centrals
as evidenced by the null values of the
corresponding
quenched fraction at time of first infall ($fq_{\rm infall}$;
Fig.~\ref{fig:fqzinfall_zinfall}).
Their
low mass prevents them from being quenched by self-regulating processes
such as AGN feedback or disc instabilities.
The same is true after infall, i.e. the quenched fractions at $z=0$ result
solely from environmental processes.
Values of $fq_{\rm z0}$ are larger for galaxies that
have been accreted earlier (higher $z_{\rm infall}$)
because they have been satellites for longer periods of time, thus being affected by environmental quenching mechanisms for longer
(Fig.~\ref{fig:fqz0_zinfall}).
\item High-mass satellites ($M_{\star} \gtrsim 10^{10.5} \,{\rm M}_{\odot}$) are more
likely to be quenched
prior to infall.
At a given redshift of first infall,
the fractions of quenched satellites $fq_{\rm infall}$ are higher for more massive galaxies (Fig.~\ref{fig:fqzinfall_zinfall}).
At early accretion epochs ($z_{\rm infall}\approx 1.5$),
$\approx 30$ percent of galaxies with local stellar mass
$M_{\star} \gtrsim 10^{10.9} \,{\rm M}_{\odot}$
accreted by high-mass
haloes ($M_{\rm vir}[{\rm M}_{\odot}] \in [10^{14.1}, 10^{15.}]$) are quenched.
The fractions $fq_{\rm infall}$ increase for lower values of $z_{\rm infall}$,
which may result from the combination of the stellar mass growth of a galaxy prior to infall
and the time elapsed under the action of mass quenching processes while being central.
\item Mass quenching plays a major role in the
decline of the SF for high-mass galaxies after infall.
Within the same stellar and host halo mass
range
defined at a given time of first infall,
satellites and centrals follow a similar evolution of the sSFR and BH luminosity, directly related with the efficiency of AGN feedback
(Figs.~\ref{fig:sSFRevolSatCen} and~\ref{fig:lumBHevolSatCen}).
These quantities are slightly smaller for satellites as a consequence of the additional effect of
RPS, which dominates among the
environmental processes included
in our model and exerts milder effects on more massive galaxies (Paper I).
\end{itemize}
\item
For a given
$z=0$
stellar mass, SF quenching mechanisms are less efficient in galaxies
accreted by lower mass haloes
both prior to and after their first infall.
The stellar mass at infall is smaller for satellites of less massive haloes
as a consequence of the different stellar-mass growth rates that characterise galaxies of the same $z=0$ stellar mass in different environments.
Therefore, at a given time, galaxies accreted by lower mass haloes experience milder mass quenching
prior to first infall (lower values of $fq_{\rm infall}$) than a population with the same $z=0$ stellar mass infalling in more massive haloes (Fig.~\ref{fig:fqzinfall_zinfall}).
Moreover, mass quenching after infall is also reduced, which combined with
the milder environmental effects exerted by lower mass haloes produces lower values of $fq_{\rm z0}$.
Both fractions $fq_{\rm z0}$ and $fq_{\rm infall}$ are reduced for less massive haloes at any stellar mass. However,
$fq_{\rm z0}$ is more strongly reduced, and thus the ratio
$fq_{\rm infall}/fq_{\rm z0}$, which gives the fraction of
galaxies quenched at $z=0$
that are already quenched at first infall, takes
larger values
for galaxies residing in host haloes of lower mass
(Fig.~\ref{fig:fq-at-infall-z0-ms-mh}).
The interpretation of this trend may lead to a conclusion opposite to the one drawn from our analysis, i.e.
that quenching prior to infall is more important in less massive
haloes, as the one discussed in W13.
\\
\item The quenching times of
$z=0$
passive satellites that were
star-forming
at first infall are anti-correlated with their present-day stellar mass (Fig.~\ref{fig:tq-mstar-mhalo}), consistent with results from W13:
the average values of $t_{\rm q}$ are
$\approx 4-5\,{\rm Gyr}$ for $M_{\star} \approx 10^{10} \,{\rm M}_{\odot}$
and $\approx 2-3\,{\rm Gyr}$ for $M_{\star} \approx 10^{11} \,{\rm M}_{\odot}$.
These average values of $t_{\rm q}$ are estimated including satellites with all possible values of $z_{\rm infall}$.
The anti-correlation with $z=0$ stellar mass arises because
early accreted
low-mass
galaxies
achieve quenching times as long as $\approx 6\,{\rm Gyr}$ (Fig.~\ref{fig:tq}), and these galaxies
outnumber the recently accreted ones, characterised by shorter quenching times.
\\
\item
Overall, we can characterise the SF quenching process of
$z=0$ passive satellites as consisting of two stages. During the first one, which is measured from
the moment that a
star-forming
galaxy becomes a satellite,
the gradual decline of the SFR resembles that of centrals of the same stellar mass as the satellites at infall. High levels of SF are sustained by high rates of gas cooling that only experience a mild reduction during this stage; their values at infall are
$\approx 30-40\,{\rm M}_{\odot}\,{\rm yr}^{-1}$.
When the cooling rate reaches half its value at infall ($\approx 20\,{\rm M}_{\odot}\,{\rm yr}^{-1}$, on average),
there is a transition from a slow to a rapid decline of the gas cooling rate, as suggested by the evolution on the sSFR in Fig.~\ref{fig:sSFRevolSatCen},
which denotes the end of the delay phase.
This process takes place in a delay time
that ranges from
$t_{\rm q,delay} \approx 3\,{\rm Gyr}$ for low-mass
satellites ($M_{\star} \lesssim 10^{10}\,{\rm M}_{\odot}$)
to $\approx 1\,{\rm Gyr}$ for
high-mass ones ($M_{\star} \approx 10^{11}\,{\rm M}_{\odot}$).
In the second stage,
the SFR declines faster until the satellite becomes passive (see the behaviour of the sSFR in Fig.~\ref{fig:sSFRevolSatCen}).
SF fades because the cold gas
supply is reduced at a faster rate (reaching values as low as
$\approx 5\,{\rm M}_{\odot}\,{\rm yr}^{-1}$ by the time the satellites are quenched) rather than being halted at the beginning of this fading phase as assumed in the strangulation quenching scenario \citep{Peng15}.
The cold gas disc is gradually consumed mainly through SF and/or removal through SN feedback. This process occurs in a fading time
$t_{\rm q,fade} \approx 1.5\,{\rm Gyr}$, regardless of stellar mass.
Our model is only consistent with the delay-then-rapid quenching scenario proposed by W13 for low-mass satellites. Since the delay time is shorter for more massive satellites and the fading time is largely independent of stellar mass, we find that the SF history of all $z=0$ passive satellites is better described by a {\em delay-then-fade} quenching scenario.
\\
\item Environmental processes have an important role during the delay for
the onset of SF quenching for low-mass satellites.
The gradual removal of the hot gas halo through RPS
is directly influenced by the
orbital evolution of galaxies and the mass growth of DM haloes.
The relevance of this process diminishes for high-mass satellites
which keep a large fraction of the hot gas reservoir by the time they become passive ($f_{\rm hot}\gtrsim 0.6$ ). These fractions are smaller for low-mass galaxies (mean values of $f_{\rm hot} \approx 0.4$; fractions can be as low as $\approx 0.2$ considering the scatter) but still consistent with the presence of a hot halo (Fig.~\ref{fig:fhot}). The larger reduction of the hot gas mass in less massive satellites yields
to the decrease of their gas cooling rates. For high-mass satellites, this decrease occurs as a consequence of AGN feedback.
RPS of the cold gas does not play any role in the fading of SF. It only acts on those satellites that have lost their protective hot gas halo, and contributes towards reducing even more the cold gas fraction after quenching (Fig.~\ref{fig:fcold}).
\end{itemize}
It is worth noting that either mass quenching or starbursts triggered by mergers could be responsible for
the SF quenching of galaxies while being centrals. The relative importance of these processes will be examined in
another work. Furthermore, we plan to extend the analysis presented here to high redshifts. The dependence of
the quenching time-scales on redshift has been discussed recently in several observational studies \citep{Balogh16,Fossati17,Foltz18}. Comparison with
results provided by them will allow both to test our model and to help to our understanding of the SF quenching over cosmic time.
\section*{Acknowledgements}
We thank the referee for the constructive report that improved the quality of the manuscript.
The authors gratefully acknowledge the Gauss Centre for Supercomputing e.V.
(www.gauss-centre.eu) and the Partnership for Advanced Supercomputing in Europe
(PRACE, www.prace-ri.eu) for funding the \textsc{MultiDark} simulation project
by providing computing time on the GCS Supercomputer SuperMUC at Leibniz
Supercomputing Centre (LRZ, www.lrz.de). The MDPL2 simulation has been performed
under grant pr87yi.
This work was done in part using the Geryon computer at the
Center for Astro-Engineering UC, part of the BASAL PFB-06, which received
additional funding from QUIMAL 130008 and Fondequip AIC-57 for upgrades.
SAC acknowledges funding from {\it Consejo Nacional de Investigaciones
Cient\'{\i}ficas y T\'ecnicas} (CONICET, PIP-0387), {\it Agencia Nacional
de Promoci\'on Cient\'ifica y Tecnol\'ogica} (ANPCyT, PICT-2013-0317), and {\it Universidad Nacional de La Plata} (G11-124),
Argentina.
TH and CVM acknowledge CONICET, Argentina, for their supporting fellowships.
AO acknowledges support from project AYA2015-66211-C2-2 of the Spanish Ministerio de Econom\'ia, Industria y Competitividad.
This project has received funding from the European Union's Horizon 2020 Research and Innovation Programme under the Marie Sklodowska-Curie grant agreement No 734374.
\bibliographystyle{mnras}
|
\section{Introduction}
The abundant production of top pairs at the Large Hadron Collider~(LHC)
provides an opportunity for detailed studies of top-quark properties, for
tests of the Standard Model~(SM) in the top sector, and for measurements of fundamental
parameters such as the top-quark mass. With the Higgs boson mass now known
with high precision, the $W$-boson and top-quark masses have become strongly
correlated, and an accurate determination of both would lead to a SM test of
unprecedented precision~\cite{Patrignani:2016xqp-EWreview, Baak:2014ora}.
The present value of the indirect top-mass determination from electroweak
precision data~($176.7\pm 2.1$~GeV, see~\cite{Patrignani:2016xqp-EWreview})
is in slight tension, at the $1.6\,\sigma$ level, with the direct
measurements. The latest combination of the Tevatron and the LHC
results~\cite{ATLAS:2014wva} yields~$173.34\pm 0.76$~GeV, but more recent
results favour even smaller values, close to $172.5$~GeV,
see~\cite{Aaboud:2016igd, Khachatryan:2015hba, CMS-PAS-TOP-17-007,
ATLAS-CONF-2017-071}. Recent reviews of top-mass measurements by the ATLAS
and CMS collaborations can be found in Refs.~\cite{Pearson:2017jck}
and~\cite{Castro:2017yxe}.
It has been shown that in the Standard Model as is (i.e.~assuming no new physics
effects up to the Planck scale), the vacuum is stable if the top mass, $\mt$,
is below $171$~GeV (i.e.~very close to its present value), metastable up to
$176$~GeV, and unstable above this value~\cite{Degrassi:2012ry,
Buttazzo:2013uya, Andreassen:2017rzq, Chigusa:2017dux}. The current value
is safely below the instability region. However, it should not be forgotten
that the absence of new physics up to the Planck scale is a very strong
assumption. The only conclusion we can draw from these results is that there
is no indication of new physics below the Planck scale coming from the
requirement of vacuum stability. On the other hand, the fact that the Higgs
boson quartic coupling almost vanishes at the Planck scale may have some deep
meaning that we are as yet unable to unveil.
Besides the issues related to electroweak tests and the stability of the
vacuum, the question on how precisely we can measure the top mass at hadron
colliders also has its own significance, related to our understanding of QCD
and collider physics. In view of the large abundance of top-pair production
at the LHC, it is likely that precise measurements will be performed with
very different methods, and that comparing them will give us confidence in
our ability to handle hadron-collider physics problems.
Top-mass measurements are generally performed by fitting $\mt$-dependent
kinematic distributions to Monte Carlo predictions. The most precise ones,
generally called \emph{direct measurements}, rely upon the full or partial
reconstruction of the system of the top-decay products.
The ATLAS and CMS measurements of Refs.~\cite{Aaboud:2016igd}
and~\cite{Khachatryan:2015hba}, yielding the value $172.84 \pm 0.34~{\rm
(stat)} \pm 0.61$~(syst)~GeV and $\mt=172.44 \pm 0.13~{\rm (stat)} \pm
0.47$~(syst)~GeV respectively, fall into this broad category.
The top mass cannot be defined in terms of the mass distribution of the
system of its decay products: since the top quark is a coloured object, no
final-state particle system can be unambiguously associated with it. On the
other hand, the top mass is certainly related to the mass distribution of the
system of objects arising from top decay, i.e.~hard leptons,
neutrinos and hard, $b$-flavoured hadronic jets. The mass distribution of
this system can be computed and measured, and the top mass enters this
computation as a parameter. By extracting its value from a fit to the
measured distributions, we are unavoidably affected by theoretical errors
that must be carefully assessed. In particular, these errors will depend upon
the accuracy of the modelling of these distributions.
The absence of a ``particle truth level'' for the top-decay products has led
to speculations that the top mass cannot be extracted reliably in the direct
measurements. The extracted mass is unavoidably a parameter in the
theoretical calculation or in the Monte Carlo generator that is used to
compute the relevant distributions. It has thus been argued that, because of
this, and since shower Monte Carlo~(SMC) models are accurate at leading
order~(LO) only, the extracted mass cannot be identified with a theoretically
well-defined mass, such as the pole mass or the \MSB{} mass (that differ
among each other only at the NLO level and beyond).
In the present work, we use NLO-accurate generators, so that the previously
mentioned objection does not actually apply. Moreover, it can be argued that,
in the narrow width approximation and at the perturbative level, the mass
implemented in Monte Carlo generators corresponds to the pole
mass~\cite{Nason:2017cxd} even if we do not use NLO-accurate generators.
It was also argued in Ref.~\cite{Hoang:2014oea} that the Monte Carlo mass
parameter differs from the top pole mass by an amount of the order of a
typical hadronic scale, that was there quantified to be near 1~GeV. It was
further argued that this difference is, in fact, intrinsic in the uncertainty
with which the pole mass can even be defined, because of the presence of a
renormalon in the relation of the pole to the \MSB~mass~\cite{Bigi:1994em,
Beneke:1994sw}.
Recent studies~\cite{Beneke:2016cbu, Hoang:2017btd} have shown that the
renormalon ambiguity in the top-mass definition is not as large as previously
anticipated, being in fact well below the current experimental
error.\footnote{In fact, values in this range were obtained much earlier in
Refs.~\cite{Pineda:2001zq, Bali:2013pla}, mostly in a bottom physics
context, but since the renormalon ambiguity does not depend upon the heavy
quark mass, they also apply to top.} The fact remains, however, that
non-perturbative corrections to top-mass observables (not necessarily related
to the mass renormalon) are present, can affect a top-mass determination, and
are likely to be parametrically of the order of a typical hadronic scale. We
believe, however, that this does not justify the introduction of a ``Monte
Carlo mass'' concept, since it is unlikely that non-perturbative effects,
affecting top-mass observables, can be parametrized as a universal shift of
the top-mass parameter. The real question to answer is whether these
non-perturbative effects are of the order of 100 MeV, 1 GeV, or more. While
a top-mass determination from threshold production at an $e^+e^-$ collider would
be free of such uncertainties~\cite{Beneke:2015kwa, Simon:2016htt},\footnote{
This is also the case for a top-mass determination based upon the spectrum
of $\gamma\gamma$ production near the $t\bar{t}$
threshold~\cite{Kawabata:2016aya}, that however is likely to be
statistically limited, even at the high luminosity LHC.} at hadron
colliders, non-perturbative effects of this order are likely to affect, to
some extent, most top-mass observables that have been proposed so
far.\footnote{For a recent discussion of all these issues see
Ref.~\cite{Nason:2017cxd}.}
The theoretical problems raised upon the top-quark mass measurement issues
have induced several theorists to study and propose alternative methods. The
total cross section for $t\bar{t}$ production is sensitive to the top mass,
and has been computed up to the NNLO order in QCD~\cite{Czakon:2013goa}, and
can be used to extract a top mass value~\cite{Khachatryan:2016mqs,
Aad:2014kva, Langenfeld:2009wd}.
In Ref.~\cite{Alioli:2013mxa}, observables related to the $t{\bar t}+{\rm jet}$
kinematics are considered. The authors of Ref.~\cite{Kawabataa:2014osa}
presented a method based upon the charged-lepton energy spectrum, that is not
sensitive to top production kinematics, but only to top decay,
arguing that, since this has been computed at NNLO accuracy~\cite{Gao:2012ja,
Brucherseifer:2013iv}, a very accurate measurement may be achieved. Some
authors have advocated the use of boosted top jets (see
Ref.~\cite{Hoang:2017kmk} and references therein). In
Ref.~\cite{Agashe:2016bok}, the authors make use of the \bjet{} energy peak position,
that is claimed to have a reduced sensitivity to production dynamics. In
Ref.~\cite{Frixione:2014ala}, the use of lowest Mellin moments of lepton
kinematic distributions is discussed. In the leptonic channel, it is also
possible to use distributions based on the ``stransverse'' mass
variable~\cite{Sirunyan:2017idq}, which generalizes the concept of transverse
mass for a system with two identical decay branches~\cite{Lester:1999tx,
Barr:2009jv}.
Some of these methods have in fact been exploited~\cite{CMS-PAS-TOP-13-006,
CMS-PAS-TOP-15-002, Aad:2015waa, Sirunyan:2017idq, Aaboud:2017ujq} to yield
alternative determinations of $\mt$. It turns out, however, that the direct
methods yield smaller errors at the moment, and it is likely that alternative
methods, when reaching the same precision level, will face similar
theoretical problems.
\subsection{Goals of this work}
In this work, we exploit the availability of the new
\POWHEGBOX{}~\cite{Nason:2004rx, Frixione:2007vw, Alioli:2010xd} generators
for top-pair production, i.e.~the \ttbnlodec{}~\cite{Campbell:2014kua} and
\bbfourl{}~\cite{Jezo:2016ujg} ones, in order to perform a theoretical study
of uncertainties in the top-mass determination. In particular, we are in a
position to assess whether NLO corrections in top decay, that are implemented
in both the \ttbnlodec{} and \bbfourl{} generators, and finite width effects,
non-resonant contributions and interference of radiation generated in
production and decay, that are implemented in \bbfourl{}, can lead to sizeable
corrections to the extracted value of the top mass. Since the \hvq{}
generator~\cite{Frixione:2007nw}, that implements NLO corrections only in
production, is widely used by the experimental collaborations in top-mass
analyses, we are particularly interested in comparing it with the new
generators, and in assessing to what extent it is compatible with
them.\footnote{The \hvq{} and \ttbnlodec{} generators can be found under the
{\tt User-Processes-V2} directory of the \VTWO{} repository in the {\tt
hvq} and {\tt ttb\_NLO\_dec} directories, respectively. The \bbfourl{}
generator can be found under the {\tt User-Processes-RES/b\_bbar\_4l}
directory of the \RES{} code. Detailed instructions are found at
\url{powhegbox.mib.infn.it}.} We will consider variations in the scales,
parton distribution functions~(PDFs) and the jet radius parameter to better
assess the level of compatibility of the different generators.
We are especially interested in effects that can be important in the
top-mass determination performed in direct measurements. Thus, the main focus
of our work is upon the mass of a reconstructed top, that we define as a
system comprising a hard lepton, a hard neutrino and a hard $b$ jet. We will
assume that we have access to the particle truth level, i.e.~that we can also
access the flavour of the $b$ jet, and the neutrino momentum and
flavour. We are first of all interested in understanding to what extent the
mass peak of the reconstructed top depends upon the chosen NLO+PS
generator. This would be evidence that the new features introduced in the
most recent generators are mandatory for an accurate mass extraction.
We will also consider the inclusion of detector effects in the form of a
smearing function applied to our results. Although this procedure is quite
crude, it gives a rough indication of whether the overall description of the
process, also outside of the reconstructed resonance peak, affects the
measurement.
Besides studying different NLO+PS generators, we have also attempted to give
a first assessment of ambiguities associated with shower and non-perturbative
effects, by interfacing our NLO+PS generators to two shower Monte Carlo
programs: \PythiaEightPtwo{}~\cite{Sjostrand:2014zea} and
\HerwigSevenPone{}~\cite{Bahr:2008pv, Bellm:2015jjp}. Our work focuses upon
NLO+PS and shower matching. We thus did not consider further variations of
parameters and options within the same parton shower, nor variations on the
observables aimed at reducing the dependence upon those.\footnote{An
interesting example of work along this direction can be found in
Refs.~\cite{Wicke:2008iz} and~\cite{Sjostrand:2013cya}, where the impact of
the colour reconnection model on top-mass measurement is analyzed. In
Ref.~\cite{Andreassen:2017ugs}, a study is performed to determine whether
the use of jet-grooming techniques in top-mass measurement can reduce the
Monte Carlo tune dependence.}
We have also considered two alternative proposals for top-mass measurements:
the position of the peak in the $b$-jet energy~\cite{Agashe:2016bok} and the
leptonic observables of Ref.~\cite{Frixione:2014ala}. The first proposal is
an example of a hadronic observable that should be relatively insensitive to
the production mechanism, but may be strongly affected by NLO corrections in
decay. The second proposal is an example of observables that depend only
upon the lepton kinematics, and that also depend upon production dynamics,
thus stronger sensitivity to scale variations and PDFs may be expected. It is
also generally assumed that leptonic observables should be insensitive to the
$b$-jet modeling. One should remember, however, that jet dynamics affects
lepton momenta via recoil effects, so it is interesting to study whether
there is any ground to this assumption.
The impact of NLO corrections in decays and finite-width effects were also
considered in Ref.~\cite{Heinrich:2017bqp} for a number of top-mass related
observables, and in Ref.~\cite{Bevilacqua:2017ipv} for the method relying
upon the $t\bar{t}j$ final state. Here we are more interested in observables
related to direct measurements, that are not considered there. Furthermore,
we focus our studies upon the differences with respect to the widely-used
\hvq{} generator.
\subsection{Preamble}
The study presented in this work was triggered by the availability of new
NLO+PS generators describing top decay with increasing accuracy. As such, its
initial aim was to determine whether and to what extent these new generators,
and the associated new effects that they implement, may impact present
top-mass measurements. As we will see, had we limited ourselves to the study
of the NLO+PS generators interfaced to \PythiaEightPtwo{}, we would have
found a fairly consistent picture and a rather simple answer to this
question.
Since another modern shower generator that can be interfaced to our NLO+PS
calculation is available, namely \HerwigSevenPone{}, we have developed an
appropriate interface to it, and have also carried out our study using it as
our shower model. Our results with \HerwigSevenPone{} turn out to be quite
different from the \PythiaEightPtwo{} ones, to the point of drastically
altering the conclusions of our study. In fact, variations in the extracted
top mass values due to switching between \PythiaEightPtwo{} and
\HerwigSevenPone{} prevail over all variations that can be obtained within
\PythiaEightPtwo{} by switching among different NLO+PS generators, or by
varying scales and matching parameters within them. Moreover, the comparison
of the various NLO+PS generators, when using \HerwigSevenPone{}, does not
display the same degree of consistency that we find within
\PythiaEightPtwo{}. If, as it seems, the differences found between
\PythiaEightPtwo{} and \HerwigSevenPone{} are due to the different shower
models (the former being a dipole shower, and the latter an angular-ordered
one), the very minimal message that can be drawn from our work is that, in
order to assess a meaningful theoretical error in top-mass measurements, the
use of different shower models, associated with different NLO+PS generators,
is mandatory.
Our results are collected in tables and figures that are presented and
discussed by giving all details that are necessary to reproduce them. We
present a large number of results that show the effect of changing parameter
settings and matching methods in the NLO+PS calculations, some of which are
very technical. Since this may obscure the logical development of our work,
we have written our Summary (Sec.~\ref{sec:Summary}) in such a way that the
main logical developments and findings are presented in a concise way. In
fact, the summary section can be read independently of the rest of the paper,
and may be used to navigate the reader through the rest of the material.
\subsection{Outline}
The paper is organized as follows. In Sec.~\ref{sec:generators} we briefly
review the features of the \hvq{}, \ttbnlodec{} and \bbfourl{} generators.
We also discuss the interfaces to the parton-shower programs
\PythiaEightPtwo{} and \HerwigSevenPone{}.
In Sec.~\ref{sec:pheno}, we detail the setup employed for the
phenomenological studies presented in the subsequent sections.
In Sec.~\ref{sec:reconstructedpeak}, we perform a generic study of the
differences of our generators focusing upon the mass distribution of the $W\,
b$-jet system. The aim of this section is to show how this distribution is
affected by the different components of the generators by examining results
at the Born level, after the inclusion of NLO corrections, after the parton
shower, and at the hadron level.
In Sec.~\ref{sec:methodology} we describe how we relate the computed
value of our observables to the corresponding value of the top mass that
would be extracted in a measurement.
In Sec.~\ref{sec:mwbj} we consider as our top-mass sensitive observable the
peak position in the mass distribution of the reconstructed top, defined as
the mass of the system comprising the hardest lepton and neutrino, and the
jet with the hardest $b$-flavoured hadron, all of them with the appropriate
flavour to match a $t$ or a $\bar{t}$. We study its dependence upon the
NLO+PS generator being used, the scale choices, the PDFs, the value of $\as$
and the jet radius parameter. Furthermore, we present and compare results
obtained with the two shower Monte Carlo generators \PythiaEightPtwo{} and
\HerwigSevenPone{}.
We repeat these studies for the peak of the $b$-jet energy
spectrum~\cite{Agashe:2016bok} in Sec.~\ref{sec:Ebjet}, and for the leptonic
observables~\cite{Frixione:2014ala} in Sec.~\ref{sec:LepObs}.
In Sec.~\ref{sec:Summary} we summarize our results, and in
Sec.~\ref{sec:Conc} we present our conclusions. In the appendices we give
some technical details.
\section{NLO+PS generators}
\label{sec:generators}
In this section we summarize the features of the \POWHEGBOX{} generators used
in the present work, i.e.~the \hvq{}, the \ttNLOdec{} and the \bbfourl{}
generators.
The \hvq{} program~\cite{Frixione:2007nw} was the first top-pair production
generator implemented in \POWHEG{}. It uses on-shell matrix elements for NLO
production of $t\bar{t}$ pairs. Off-shell effects and top decays, including
spin correlations, are introduced in an approximate way, according to the
method presented in Ref.~\cite{Frixione:2007zp}. Radiation in decays is
fully handled by the parton-shower generators. The ones that we consider,
\PythiaEightPtwo{} and \HerwigSevenPone{}, implement internally matrix-element
corrections for top decays, with \HerwigSevenPone{} also optionally
including a \POWHEG{}-style hardest-radiation generation. In these cases,
their accuracy in the description of top decays is, for our purposes,
equivalent to the NLO level.
The \ttNLOdec{} code~\cite{Campbell:2014kua} implements full spin
correlations and NLO corrections in production and decay in the narrow-width
approximation. Off-shell effects are implemented via a reweighting method,
such that the LO cross section includes them exactly. As such, it also contains
contributions of associated top-quark and $W$-boson production at LO. It
does not include, however, interference of radiation generated in production
and decay.
In \ttbnlodec{} the \POWHEG{} method has been adapted to deal with radiation
in resonance decays. Radiation is generated according to the \POWHEG{}
Sudakov form factor both for the production and for all resonance decays that
involve coloured partons. This feature also offers the opportunity to modify
the standard \POWHEG{} single-radiation approach. Rather than picking the
hardest radiation from one of all possible origins (i.e.~production and
resonance decays), the \POWHEGBOX{} can generate simultaneously the hardest
radiation in production and in each resonance decay. The LH events
generated in this way can thus carry more radiated partons, one for
production and one for each resonance. Multiple-radiation events have to be
completed by a shower Monte Carlo program, that has to generate radiation
from each origin without exceeding the hardness of the corresponding
\POWHEG{} one, thus requiring an interface that goes beyond the simple Les
Houches standard~\cite{Boos:2001cv}.
A general procedure for dealing with decaying resonances that can radiate by
strong interactions has been introduced and implemented in a fully general
and automatic way in a new version of the \POWHEGBOX{} code, the
\RES~\cite{Jezo:2015aia}. This framework allows for the treatment of
off-shell effects, non-resonant subprocesses including full interference, and
for the treatment of interference of radiation generated in production and
resonance decay.\footnote{A related approach within the \MCatNLO{} framework
has been presented in Ref.~\cite{Frederix:2016rdc}.} In
Ref.~\cite{Jezo:2016ujg} an automated interface of the \RES{} code to the
\Openloops{}~\cite{Cascioli:2011va} matrix-element generator has been
developed and used to build the \bbfourl{} generator, that implements the
process $pp\to b\bar{b}\,\fourl$, including all QCD NLO corrections in the
4-flavour scheme, i.e.~accounting for finite $b$-mass effects. So, double-top,
single-top and non-resonant\footnote{By non-resonant we mean processes that
do not contain an intermediate top quark, e.g.~$pp \to b\,\bar{b} \,Z\to
b\,\bar{b} \,W^+\, W^- \to b\, \bar{b}\,\fourl $.} diagrams are all
included with full spin-correlation effects, radiation in production and
decays, and their interference.
As for the \ttNLOdec{} generator, \bbfourl{} can generate LH events including
simultaneous radiation from the production process and from the top and
anti-top decaying resonances. It thus requires a non-standard interface to
parton-shower Monte Carlo programs, as for the case of the \ttNLOdec{}
generator.
\subsection{Interface to shower generators}
According to the \POWHEG{} method, the PS program must complete the event
only with radiation softer than the \POWHEG{} generated one. In the standard Les
Houches Interface for User Processes~(LHIUP)~\cite{Boos:2001cv}, each
generated event has a hardness parameter associated with it, called {\tt
scalup}. This parameter is set in \POWHEG{} to the relative transverse
momentum of the generated radiation and each emission attached by the parton
shower must have a $\ensuremath{p_{\sss T}}\xspace$ smaller than its value. The LHIUP treats all
emissions on an equal footing, and has no provision for handling radiation
from decaying resonances. This drives a standard PS to allow showering to
start from scales of the order of the resonance mass.
\subsubsection{Generic method}
\label{sec:genericmethod}
References~\cite{Campbell:2014kua} and~\cite{Jezo:2016ujg} introduce a
generic method for interfacing \POWHEG{} processes that include radiation in
decaying resonances with PS generators. According to this method, shower
radiation from the resonance is left unrestricted, and a veto is applied
\emph{a posteriori}: if any radiation in the decaying resonance shower is
harder than the \POWHEG{} generated one, the event is discarded, and the same
LH event is showered again. We also stress that the standard PS
implementations conventionally preserve the mass of the resonance, as long as
the resonance decay products, including eventually the radiation in decay,
have the resonance as mother particle in the LH event record.
The hardness of the radiation associated with the decaying top ($t\to
W\,b\,g$) in \POWHEG{} is given by
\begin{equation}
\label{eq:bg_splitting}
t=2\,\frac{E_g}{E_b} \,p_g\cdot p_b = 2\,E_g^2 \(1-\beta_b\cos\theta_{bg}\),
\end{equation}
where $p_{g/b}$ and $E_{g/b}$ are the four momentum and energy of the gluon
and of the bottom quark, $\beta_b$ is the velocity of the bottom quark and
$\theta_{bg}$ is the angle between the bottom and gluon momenta, all
evaluated in the top rest frame. This hardness definition is internally used
to define the corresponding Sudakov form factor. The same should be also
used to limit the transverse momentum generated by the PS in the resonance
decay.
The practical implementation of the veto procedure depends on whether we are
using a dipole, as in \PythiaEightPtwo{}, or an angular-ordered shower, as in
\HerwigSevenPone{}. If we are using a dipole ($\ensuremath{p_{\sss T}}\xspace$-ordered) shower, it is
sufficient to check the first shower-generated emission from the bottom quark
and (if present at the LH level) from the gluon arising in top decay. The
hardness $t_b$ of the shower-generated emission from the bottom is evaluated
using eq.~(\ref{eq:bg_splitting}), while the one from the gluon is taken to
be
\begin{equation}
\label{eq:gg_splitting}
t_g=2\,E_1^2\,E_2^2\,\frac{(1-\cos\theta_{12})}{(E_1+E_2)^2}\,,
\end{equation}
where $E_{1,2}$ are the energies of the two gluons arising from the
splitting, and $\theta_{12}$ is the angle between them. Both $t_g$ and $t_b$
are computed in the top frame. If they are smaller than $t$, the event is
accepted, otherwise it is showered again.
In the case of angular-ordered showers, as in \HerwigSevenPone{}, it is not
enough to examine the first emission, because the hardest radiation may take
place later. As shown in Ref.~\cite{Nason:2004rx}, in the leading
logarithmic approximation, the hardest emission in an angular-ordered shower
can be always found by following either the quark line in a $q \to q g$
splitting, or the most energetic line in a $g \to gg$ splitting. Thus, when
inspecting the sequence of splittings, in order to find the hardest
radiation, if the parton that generates the shower is a fermion (in our case,
the $b/\bar{b}$ quark), we simply follow the fermionic line; in case of a
gluon splitting, we follow the most energetic gluon. We go on until either
the shower ends, or we reach a $g\to q \bar{q}$ splitting. Since this last
process is not soft-singular, configurations with the hardest emission
arising after it are suppressed.
\subsubsection{Standalone implementations in \PythiaEightPtwo{}}
\label{sec:PY8_different_showers}
The \PythiaEightPtwo{} generator provides facilities for implementing the
above-described method to internally veto radiation in resonance decays. We
prepared two implementations, each based on a different facility, and now we
describe them in turn.
\begin{enumerate}
\item
At every radiation generated by \PythiaEightPtwo{}, a function is called internally
using the \verb!UserHooks! facility. The function inspects the radiation
kinematics. If the radiation comes from top decays, it computes its
transverse momentum, according to eqs.~(\ref{eq:bg_splitting})
and~(\ref{eq:gg_splitting}). If the transverse momentum is larger than the
one of the radiation generated by \POWHEG{} in the resonance decay, the
emission is vetoed, and \PythiaEightPtwo{} tries to generate another splitting.
The process is repeated until an acceptable splitting is generated. This
behaviour is achieved by implementing the method
\begin{verbatim}
UserHooks::doVetoFSREmission,
\end{verbatim}
whose description can be found in the \PythiaEightPtwo{} manual~\cite{pythiamanual}.
It is activated by setting the \PythiaEightPtwo{} flag
\begin{verbatim}
POWHEG:bb4l:FSREmission:veto = on.
\end{verbatim}
\item
The {\tt UserHooks} facility also allows us to set the initial scale of
final-state shower evolution (for the shower arising from the decaying
resonances) equal to the transverse momentum of the top radiation in decay.
This is achieved using the method
\begin{verbatim}
UserHooks::scaleResonance,
\end{verbatim}
and is activated by setting the \PythiaEightPtwo{} flag
\begin{verbatim}
POWHEG:bb4l:ScaleResonance:veto = on.
\end{verbatim}
This method has the disadvantage of relying upon the assumption that the
hardness definition used by \PythiaEightPtwo{} is compatible with the \POWHEG{} one.
\end{enumerate}
Both methods are implemented in the file
\begin{verbatim}
PowhegHooksBB4L.h
\end{verbatim}
in the \bbfourl{} subprocess directory.
We have chosen implementation 1 as our default, and compared it with the
other implementations in order to validate it and estimate matching
uncertainties.
\subsubsection{Standalone implementations in \HerwigSevenPone{}}
\label{sec:HW7_different_showers}
Also in the case of \HerwigSevenPone{} we have prepared two implementations that
use the MC internal facilities to perform the veto:
\begin{enumerate}
\item
After the whole time-like shower has been developed, but before hadronization
has been carried out, the showers from the $b$ and from the \POWHEG{}
radiated gluon in top decay are examined. In the case of the $b$, the quark
line is followed, and the transverse momentum of the radiation is computed
(in the top frame) according to eq.~(\ref{eq:bg_splitting}). In the case of
the gluon, the hardest line is followed, and the transverse momentum of the
radiation is computed according to eq.~(\ref{eq:gg_splitting}). If a
radiation is found with transverse momentum harder than the \POWHEG{}
generated one, the full event is reshowered, starting from the same LH event.
The corresponding method is called
\begin{verbatim}
FullShowerVeto::vetoShower,
\end{verbatim}
and we have implemented it in the files
\begin{verbatim}
bb4lFullShowerVeto.h, bb4lFullShowerVeto.cc.
\end{verbatim}
\item
We veto each radiation in resonance decay if its transverse momentum is
harder than the \POWHEG{} generated one. In this case, \HerwigSevenPone{} tries again
to generate radiation starting from the (angular ordering) hardness parameter
of the vetoed one. As in \PythiaEightPtwo{} second method, we have to rely in this
case upon the \HerwigSevenPone{} definition of the radiation transverse momentum. The
corresponding method is called
\begin{verbatim}
ShowerVeto::vetoTimeLike
\end{verbatim}
and we implemented it in the files
\begin{verbatim}
bb4lShowerVeto.h, bb4lShowerVeto.cc.
\end{verbatim}
\end{enumerate}
We will adopt implementation~2 as our \HerwigSevenPone{} default, and compare
with the other one in order to validate it, and also in order to get an
indication of the size of matching uncertainties.
\section{Phenomenological analysis setup}
\label{sec:pheno}
We simulate the process $p\,p \to b\, \bar{b}\,\fourl$, which is available in
all three generators. It is dominated by top-pair production, with a smaller
contribution of $Wt$ topologies. For the observables we consider, the decay
of one of the two top quarks is mostly irrelevant, so that our result will
also hold for semileptonic decays.
In the \hvq{} and \ttbnlodec{} generators we renormalize the top mass in the
pole-mass scheme, while in the \bbfourl{} one we adopt the complex mass
scheme~\cite{Jezo:2016ujg}, with the complex mass defined as
$\sqrt{\mt^2-i\,\mt \,\Gamma_t}$.
We perform our simulations for a center-of-mass energy of $\sqrt{s}=8$~TeV.
We have used the {\tt MSTW2008nlo68cl} PDF set~\cite{Martin:2009iq} and we
have chosen as central renormalization and factorization scale ($\ensuremath{\mu_{\sss \mathrm{R}}}$ and
$\ensuremath{\mu_{\sss \mathrm{F}}}$) the quantity $\mu$, defined, following Ref.~\cite{Jezo:2016ujg}, as
the geometric average of the transverse masses of the top and anti-top
\begin{equation}
\label{eq:centralscale}
\mu= \sqrt[4]{\left(E^2_t -p_{z,t}^2\right)\left(E^2_{\bar{t}}
-p_{z,\bar{t}}^2\right)}\,,
\end{equation}
where the top and anti-top energies $E_{t/\bar{t}}$ and longitudinal momenta
$p_{z,t/\bar{t}}$ are evaluated at the underlying-Born level.
In the \bbfourl{} case, there is a tiny component of the cross section given
by the topology
\begin{equation}
pp\to Z g \to (W^+ \to e^+ \nu_e) (W^- \to \mu^-
\bar{\nu}_\mu) (g \to b \bar{b}).
\end{equation}
In this case we define $\mu$ as
\begin{equation}
\label{eq:centralscaleZ}
\mu= \frac{\sqrt{p_{\sss Z}^2}}{2}\,,
\end{equation}
where $p_{\sss Z}=p_{\mu^-}+p_{\bar{\nu}_\mu} + p_{e^+} +p_{\nu_{\sss e}}$.
This case is however very rare and unlikely to have any significance.
The parameter {\tt hdamp} controls the separation of remnants
(see \writeApp\ref{app:remnants}) in the production of $t\bar{t}$ pairs with large
transverse momentum. We set it to the value of the top mass.
\subsection{Physics objects}
\label{sec:physicsObjects}
In our simulations we make the $B$ hadrons stable, in order to simplify the
definitions of $b$ jets. Jets are reconstructed using the
Fastjet~\cite{Cacciari:2011ma} implementation of the anti-$k_{\rm\sss T}$
algorithm~\cite{Cacciari:2008gp} with $R=0.5$. We denote as $B$~(${\bar B}$)
the hardest (i.e.~largest \pT{}) $b$~($\bar{b}$) flavoured hadron. The
$b$~($\bar{b}$) jet is the jet that contains the hardest
$B$~($\bar{B}$).\footnote{Note that this notation is the opposite of what is
commonly adopted for $B$ mesons, where $B$ refers to the meson containing
the $\bar{b}$ quark.} It will be indicated as $j_B$~($j_{\bar{B}}$). We
discard events where the $b$ jet and $\bar{b}$ jet coincide. The hardest
$e^+$~($\mu^-$) and the hardest $\nu_e$~($\bar{\nu}_{\mu}$) are paired to
reconstruct the $W^+$~($W^-$). The reconstructed top~(anti-top) quark is
identified with the corresponding $W^+j_B$ ($W^-j_{\bar{B}}$) pair. In the
following we will refer to the mass of this system as \mwbj{}.
We require the two $b$ jets to have
\begin{equation}
\pT>30\mbox{~GeV}\,, \qquad |\eta|<2.5\,.
\end{equation}
These cuts suppress the single-top topologies.
The hardest $e^+$ and the hardest $\mu^-$ must satisfy
\begin{equation}
\pT>20\mbox{~GeV}\,, \qquad |\eta|<2.4\,.
\end{equation}
\subsection{Generated sample}
For each generator under study, we have produced three samples of events,
each sample computed with a top mass of 169.5, 172.5 and 175.5~GeV,
respectively, with the corresponding decay width computed at NLO. Using the
reweighting feature of the \POWHEGBOX{}, we have computed the event weights
obtained by varying the parton distribution functions and the renormalization
and factorization scales, for a total of 12 weights (see
Secs.~\ref{sec:fac_ren_scales} and~\ref{sec:PDF_dependence} for more
details).
In the reweighting procedure, only the inclusive \POWHEG{} cross section is
recomputed. The Sudakov form factor is not recomputed, so that the radiated
partons retain the same kinematics. For this reason, the change of the
renormalization and factorization scales do not affect the emission of
radiation.
Thus, in order to investigate the sensitivity of the result on the intensity
of radiation, where we are particularly concerned with emissions from the
final-state $b$ quarks, we have also generated samples with the
{\tt NPDF30\_nlo\_as115} and {\tt NNPDF30\_nlo\_as121}, with $\as(\mZ)=0.115$
and $\as(\mZ)=0.121$ respectively, for each generator,
for the central value of the top
mass, i.e.~172.5~GeV. The number of events for each generated sample,
together with an indicative computational time, are reported in
Tab.~\ref{tab:samples}.
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{ \begin{tabular}{l|c|c||c|c||c|c||c|c||c|c|}
\cline{2-11}
&\multicolumn{10}{|c|}{Generated samples} \\
\cline{2-11}
&\multicolumn{6}{|c||}{$m_t$~[GeV]} & \multicolumn{4}{|c|}{$\as(\mZ)$} \\
\cline{2-11}
& \multicolumn{2}{|c||}{$172.5$} & \multicolumn{2}{|c||}{$169.5$} & \multicolumn{2}{|c||}{$175.5$} &
\multicolumn{2}{|c||}{$0.115$} & \multicolumn{2}{|c|}{$0.121$} \\
\cline{2-11}
& \phantom{\Big|}\# events & time &\# events & time & \# events & time & \# events & time & \# events & time \\
\cline{1-11}
\multicolumn{1}{|c|}{\hvq{}} & \phantom{\Big|}12M & 10~h & 3M & 2.5~h & 3M & 2.5~h & 12M & 9~h & 12M & 9~h \\
\cline{1-11}
\multicolumn{1}{|c|}{\ttbnlodec{}} & \phantom{\Big|}12M & 46~d & 3M & 11.5~d & 3M & 11.5~d & 12M & 25~d & 12M & 25~d \\
\cline{1-11}
\multicolumn{1}{|c|}{\bbfourl{}} & \phantom{\Big|}20M & 4600~d & 1.7M & 390~d & 1.7M & 390~d & 3M & 64~d & 3M & 64~d \\
\cline{1-11}
\end{tabular}
}
\caption{Number of events and total CPU time of the generated samples. The
samples used for the $\as$ variations were obtained in a relatively
smaller time, since in this case only the central weight was
computed. This leads to a difference that can be sizeable, depending upon
the complexity of the virtual corrections.}
\label{tab:samples}
\end{table*}
\section{Anatomy of the reconstructed top mass distribution at NLO+PS}
\label{sec:reconstructedpeak}
In this section we investigate the impact of individual ingredients in a
typical NLO+PS calculation on the kinematic distribution of the reconstructed
top mass \mwbj{}. On the perturbative side, we examine the impact of the
different level of accuracy in the treatment of top production and decay
provided by the three generators we are considering, and the impact of
parton-shower effects. On the non-perturbative side, we illustrate the
effect of including hadronization and underlying event in the simulation.
\subsection{Les Houches event level comparison of the generators}
We begin by comparing the three generators at the Les Houches event~(LHE) level.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/hvq_LHE_NLO_LO}
\caption{${d\sigma}/{d \mwbj}$ distribution at LO~(blue) and at NLO~(red)
obtained with the \hvq{} generator, normalized to 1 in the displayed range. In the
bottom panel the ratio with the LO prediction is shown. }
\label{fig:hvqLHE}
\end{figure}
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/bb4l_LHE_NLO_LO}
\caption{${d\sigma}/{d \mwbj}$ distribution at LO~(blue) and at NLO~(red)
obtained with the \bbfourl{} generator, normalized to 1 in the displayed range. In
the bottom panel the ratio with the LO prediction is
shown. }
\label{fig:bb4lLHE}
\end{figure}
In Figs.~\ref{fig:hvqLHE} and~\ref{fig:bb4lLHE} we compare \mwbj{},
normalized to 1 in the displayed range, at LO and NLO accuracy using the
\hvq{} and the \bbfourl{} generators respectively. The \hvq{} generator
includes NLO corrections only in the production process. Thus the \mwbj{}
distributions at LO and NLO are very similar. On the other hand, in the case
of the \bbfourl{} generator~(Fig.~\ref{fig:bb4lLHE}), we observe large
differences below the peak region. These differences are easily interpreted
as due to radiation outside the \bjet{} cone in the top-decay process.
The \ttNLOdec{} generator allows us to specify whether NLO accuracy is
required both in production and decay (default behaviour), or just in
production (by using the \pwgopt{nlowhich 1} option). In
Fig.~\ref{fig:ttdecLHE}
\begin{figure}[tb]
\centering \includegraphics[width=\wfigsing]{figures/ttdec_LHE_NLO_LO}
\caption{${d\sigma}/{d \mwbj}$ distribution with NLO accuracy in
production and decay~(red), only in production~(green) and with LO
accuracy~(blue) obtained with the \ttNLOdec{} generator, normalized to 1 in the
displayed range. In the bottom panel the ratio with the LO prediction is
shown.}
\label{fig:ttdecLHE}
\end{figure}
we compare the two options.
We see that our previous observation is confirmed: the impact of NLO
corrections in production leads to a roughly constant $K$-factor, while the
radiation from top decay affects the shape of the distribution below the peak
region.
A remaining important difference between the \hvq{} and the other two
generators has to do with the way the distribution of the top virtuality is
modeled. The \bbfourl{} and \ttNLOdec{} generators are guaranteed to yield
the correct virtuality distribution at the NLO and LO level, respectively.
This is not the case for the \hvq{} generator, where the resonance structure is
recovered by a reweighting procedure that does not guarantee LO accuracy.
\begin{figure}
\centering \includegraphics[width=\wfigsing]{figures/hvq_ttdec_bb4l_LHE_LO-lhe}
\caption{${d\sigma}/{d \mwbj}$ distribution at LO obtained with
\bbfourl{}~(red), \ttNLOdec{}~(blue) and \hvq{}~(green), normalized to 1
in the displayed range. In the bottom panel the ratio with the
\bbfourl{} prediction is shown.}
\label{fig:allLO}
\end{figure}
This is illustrated in Fig.~\ref{fig:allLO}, where we see that a
non-negligible (although not dramatic) difference in shape is present also at
the LO level between the \hvq{} and the other two generators.
\subsection{Shower effects}
We now examine how the shower, i.e.~the radiation beyond the hardest one,
affects our distributions. First of all, we anticipate an important effect in
\hvq{}, since in this case radiation in decay is fully generated by the
shower. We thus expect a raise of the low mass tail in the \mwbj{}
distribution, comparable in size to the one observed in the \bbfourl{} and
\ttNLOdec{} generators at the LHE level. Conversely, in the \bbfourl{} and
\ttNLOdec{} cases, we expect smaller shower corrections, since the hardest
radiation in decay is already included at the LHE level.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigdoub]{figures/hvq_LHE_NLO_PS} \fignewline
\includegraphics[width=\wfigdoub]{figures/bb4l_LHE_NLO_PS}
\caption{${d\sigma}/{d \mwbj}$ distribution obtained with \hvq{}~(left pane) and
\bbfourl{}~(right pane) at the NLO LHE level~(green), and at NLO+shower (in
red \PythiaEightPtwo{} and in blue \HerwigSevenPone{}), normalized to 1 in
the displayed range. In the bottom panel the ratio with the NLO LHE is
shown.}
\label{fig:NLO-PS}
\end{figure}
This is illustrated in Fig.~\ref{fig:NLO-PS}, where we clearly see that in the
\hvq{} case there is an important increase of the cross section below the
peak. On the other hand, in the \bbfourl{} case this increase is minor or
even absent, depending upon which shower program is used. In both cases, we
see an enhancement in the region above the peak. This is attributed to shower
radiation that is captured by the \bjet{} cone. We observe that, after
shower, the \hvq{} result becomes qualitatively very similar to the
\bbfourl{} one, as shown in Fig.~\ref{fig:bb4l+hvq-PS}.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/hvq_bb4l_PY8_showerOnly}%
\caption{${d\sigma}/{d \mwbj}$ distribution, normalized to 1 in the displayed range,
obtained with \bbfourl{}~(red) and \hvq{}~(blue) at the NLO+PS level using
\PythiaEightPtwo{}.}
\label{fig:bb4l+hvq-PS}
\end{figure}
The inclusion of the shower in \ttNLOdec{} leads to effects
similar to those observed in \bbfourl{}.
\subsection{Hadronization and underlying events}
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigdoub]{figures/hvq_PY8_shower_mpi_had}\fignewline
\includegraphics[width=\wfigdoub]{figures/hvq_HW7_shower_mpi_had}
\caption{${d\sigma}/{d \mwbj}$ distribution obtained with \hvq{} interfaced with
\PythiaEightPtwo{}~(left panel) and \HerwigSevenPone{}~(right panel). In
green, the NLO+PS results; in red, hadronization effects are included; in
blue, NLO+PS with multi-parton interactions~(MPI); and in black, with
hadronization and MPI effects. The curves are normalized using the NLO+PS
cross section in the displayed range.}
\label{fig:hvq-PS-NP}
\end{figure}
In Fig.~\ref{fig:hvq-PS-NP} we show the effect of hadronization and
multi-parton interactions~(MPI), as modeled by \PythiaEightPtwo{} and
\HerwigSevenPone{}, when interfaced to the \hvq{} generator. We can see the
large effect of the hadronization on the final distribution. This effect is
also considerably different between \PythiaEightPtwo{} and \HerwigSevenPone{}. There
are two main features that emerge in these plots. First of all, as expected,
the MPI raise the tail of the distributions above the peak. In fact,
MPI-generated particles are deposited in the \bjet{} cone, thus increasing
the \bjet{} energy. Hadronization widens the peak for both
generators. However, in the \PythiaEightPtwo{} case, we also observe a clear
enhancement of the low mass region, that is not as evident in the
\HerwigSevenPone{} case. In the combined effect of hadronization and MPI,
\HerwigSevenPone{} has a wider peak. On the other hand, the high tail enhancement
seems similar in the two generators.
We remark that the different mechanisms that lead to an increased cross
section above and below the top peak depend on the jet radius parameter
$R$. By increasing (or decreasing) $R$, the peak position is shifted to the left
(or right). Furthermore, differences in the implementation of radiation from
the resonances, the hadronization model and the underlying events can also
shift the peak, leading eventually to a displacement of the extracted top
mass, that should be carefully assessed.
\section{Methodology}
\label{sec:methodology}
In the following sections we will examine various sources of theoretical
errors in the top-mass extraction, focusing upon three classes of
observables: the reconstructed mass peak, the peak of the \bjet{} energy
spectrum~\cite{Agashe:2016bok}, and the leptonic observables of
Ref.~\cite{Frixione:2014ala}.
The reconstructed mass observable bears a nearly direct relation with the top
mass. If two generators with the same $\mt$ input parameter yield a
reconstructed mass peak position that differ by a certain amount, we can be
sure that if they are used to extract the top mass they will yield results
that differ by roughly the same amount in the opposite direction. Of course, this is
not the case for other observables. In general, for an observable $O$
sensitive to the top mass, we will have
\begin{equation}
O = O_c + B \(\mt-\mtc\) + {\cal O}\left(\(\mt-\mtc\)^2\right),
\label{eq:linearfitfunc}
\end{equation}
where $\mt$ is the input mass parameter in the generator, and
$\mtc=172.5$~GeV is our reference central value for the top mass. $O_c$ and
$B$ differ for different generators or generator setups. Given an
experimental result for $O$, $O_{\rm exp}$, the extracted mass value is
\begin{equation}
\mt = \mtc+\frac{O_{\rm exp}-O_{c}}{B}\,.
\end{equation}
By changing the generator setup, $O_{\rm c}$ and $B$ will assume the
values $O_{\rm c}'$ and $B'$, and will yield a different extracted mass $m'_t$.
We will thus have
\begin{equation}
{m'_t-\mt} = \frac{O_{\rm c}-O_{\rm c}'}{B} + \(O_{\rm exp}-O_{\rm c}'\)\frac{B-B'}{BB'}\,.
\end{equation}
The second term is parametrically smaller, of one order higher in the
deviation between the two generators, if we assume that at least one of them
yields a $\mt$ value sufficiently close to $\mtc$. We thus have
\begin{equation}
\label{eq:delta_mt}
m'_t-\mt \approx \frac{O_{\rm c}-O_{\rm c}'}{B}\,.
\end{equation}
In practice, in the following, we will compute the $B$ parameter
using the \hvq{} generator, that is the fastest one. We also
checked that using the other generators for this purpose yields
results that differ by at most 10\%{}.
\section{Reconstructed top mass distribution $\boldsymbol{\mwbj}$}
\label{sec:mwbj}
The peak of the reconstructed mass \mwbj{}, defined in
Sec.~\ref{sec:physicsObjects}, is a representative of all the direct
measurement methods. Our simplifying assumptions, that the $b$ jets are
unambiguously identified and the neutrinos are fully reconstructed, including
their sign, lead to an ideal resolution on the top peak that is not
realistic. We thus compute these distributions also introducing a smearing
that mimics the experimental systematics. This very crude approach allows us
to concentrate more on theoretical issues rather then experimental ones. For
example, if by using two different generators (or the same generator with
different settings) we find differences in the extracted mass using our ideal
\mwbj{} observable, we would be forced to conclude that there is an
irreducible theoretical error (i.e.~an error that cannot be reduced by
increasing the experimental accuracy) on the mass measurement. The same
problem in case of the smeared distribution should instead be considered less
severe, since the corresponding error may be reduced if the experimental
resolution is improved.
We remark that also ``irreducible'' errors (according to the definition given
above) may in fact be reduced in practice. This is the case if one of the
generators at hand does not fit satisfactorily measurable distributions
related to top production. As an example, a generator may not fit reasonably
the profile of the $b$ jet, and we may be forced to change the allowed range
for the parameters that control it, possibly reducing the error.
In the following, we will compare our three generators interfaced to
\PythiaEightPtwo{}, and consider scale variation effects and PDF dependence. In
order to investigate the sensitivity to the intensity of radiation from the
$b$ quark, we also consider different values of $\as$ as input. We will then
investigate the \HerwigSevenPone{} and \PythiaEightPtwo{}
differences.\footnote{Unless specified otherwise, \PythiaEightPtwo{} and
\HerwigSevenPone{} are setup to run in full hadron mode including shower,
hadronization and multi-parton interactions.}
It is quite obvious that the coefficient $B$ of eq.~(\ref{eq:linearfitfunc})
should be very near 1 for the \mwbj{} observable. The values for the $B$
coefficients that we have obtained with the three generators showered with
\PythiaEightPtwo{}, by a linear fit of the $\mt$ dependence of the \mwbj{}
distribution, are collected in Tab.~\ref{tab:mwbj_B_values}, and confirm our
expectation.
\begin{table}[tb]
\centering
{ \begin{tabular}{l|c|c|}
\cline{2-3}
$\phantom{\Big|}$ & $B$, no smearing & $B$, smearing \\
\cline{1-3}
\multicolumn{1}{ |c|}{\hvq{}} & $\phantom{\Big|}\Bfrommwbjhvq \pm \Berrfrommwbjhvq$ & $\Bfrommwbjsmearhvq \pm \Berrfrommwbjsmearhvq$ \\
\cline{1-3}
\multicolumn{1}{ |c|}{\ttbnlodec{}} & $\phantom{\Big|} \Bfrommwbjttdec \pm \Berrfrommwbjttdec$ & $\Bfrommwbjsmearttdec \pm \Berrfrommwbjsmearttdec$ \\
\cline{1-3}
\multicolumn{1}{ |c|}{\bbfourl{}} & $\phantom{\Big|} \Bfrommwbjbbfourl \pm \Berrfrommwbjbbfourl$ & $\Bfrommwbjsmearbbfourl \pm \Berrfrommwbjsmearbbfourl$ \\
\cline{1-3}
\end{tabular}}
\caption{Values for the $B$ coefficients of eq.~(\ref{eq:linearfitfunc})
for the \mwbj{} peak position, for the non-smeared and smeared
distributions~(see Sec.~\ref{sec:mwbj-cmp} for details), obtained with
the \hvq{}, \ttbnlodec{} and \bbfourl{} generators showered with
\PythiaEightPtwo{}.}
\label{tab:mwbj_B_values}
\end{table}
\subsection{Comparison among the different NLO+PS generators}
\label{sec:mwbj-cmp}
We begin by showing comparisons of our three generators, interfaced with
\PythiaEightPtwo{}, for our reference top-mass value of 172.5~GeV.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/bb4l_ttdec_PY8_mwbj_nosmear}
\caption{${d\sigma}/{d \mwbj}$ distribution obtained with the \bbfourl{} and \ttNLOdec{}
generators interfaced with \PythiaEightPtwo{}, for $\mt=172.5$~GeV.}
\label{fig:MassPeaks-py8-bb4l-ttb}
\end{figure}
We show in Fig.~\ref{fig:MassPeaks-py8-bb4l-ttb} the \mwbj{} distribution for
the \bbfourl{} and \ttNLOdec{} generators. We see that the two generators
yield a very similar shape. We have extracted the position of the maximum by
fitting the distribution with a skewed Lorentzian function of the form
\begin{equation}
y(\mwbj)=\frac{b[1+d(\mwbj-a)]}{(\mwbj-a)^2+c^2}+e\,.
\end{equation}
The peak $\mwbjmax$ is defined by
\begin{equation}
\frac{ d \, y(\mwbj)}{ d \mwbj} \Big|_{\mwbj=\, \mwbjmax}
= 0\,.
\end{equation}
The fitting procedure is described in \writeApp\ref{app:fit}.
As we can see from Fig.~\ref{fig:MassPeaks-py8-bb4l-ttb}, the \bbfourl{} and
\ttNLOdec{} results are very close to each other. We take this as an
indication that interference effects in radiation and other off-shell
effects, that are included in \bbfourl{} but not in \ttNLOdec{}, have a very
minor impact on the peak position, at least if we consider a measurement with
an ideal resolution.
In order to mimic experimental resolution effects, we smear our distribution
with a Gaussian of width $\sigma=15$~GeV (that is the typical experimental
resolution on the reconstructed top mass)
\begin{equation}
f_{\rm smeared}(x) = \mathcal{N} \int dy \, f(y)\,
\exp\left(-\frac{(y-x)^2}{2\sigma^2}\right)\,,
\end{equation}
where $\mathcal{N}$ is a normalization constant.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/bb4l_ttdec_PY8_mwbj_smear}
\caption{Smeared ${d\sigma}/{d \mwbj}$ distribution obtained with the \bbfourl{}
and \ttNLOdec{} generators interfaced with \PythiaEightPtwo{}, for
$\mt=172.5$~GeV.}
\label{fig:MassPeaks-py8-bb4l-ttb-smeared}
\end{figure}
The results, obtained with the same fitting procedure, are shown in
Fig.~\ref{fig:MassPeaks-py8-bb4l-ttb-smeared}. Smearing effects are such
that more importance is given to the region away from the peak, where there
are larger differences between the two generators, leading to a difference in
the peak position of \diffttdecbbfourl~MeV.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/bb4l_hvq_PY8_mwbj_nosmear}
\caption{${d\sigma}/{d \mwbj}$ distribution obtained with the
\bbfourl{} and \hvq{} generators
interfaced with \PythiaEightPtwo{}, for
$\mt=172.5$~GeV.}
\label{fig:MassPeaks-py8-bb4l-hvq}
\end{figure}
In Figs.~\ref{fig:MassPeaks-py8-bb4l-hvq} and
\ref{fig:MassPeaks-py8-bb4l-hvq-smeared},
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/bb4l_hvq_PY8_mwbj_smear}
\caption{Smeared ${d\sigma}/{d \mwbj}$ distribution obtained with the
\bbfourl{} and \hvq{} generators interfaced with \PythiaEightPtwo{}, for
$\mt=172.5$~GeV.}
\label{fig:MassPeaks-py8-bb4l-hvq-smeared}
\end{figure}
we compare the \bbfourl{} and the \hvq{} generators in the non-smeared and
smeared case respectively. We see a negligible difference in the peak
position in the non-smeared case, while, in the smeared case, the \hvq{}
generator differs from \bbfourl{} by \diffhvqbbfourl~MeV, similar in
magnitude to the case of \ttbnlodec{}, but with opposite sign.
These findings are summarized in
Tab.~\ref{tab:mwbj_showerOnly},
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/cmp_bb4l_table.tex}}
\caption{Differences in the $\mwbj{}$ peak position for $\mt$=172.5~GeV for
\ttbnlodec{} and \hvq{} with respect to \bbfourl{}, showered with
\PythiaEightPtwo{}, at the NLO+PS level and at the full hadron level.}
\label{tab:mwbj_showerOnly}
\end{table*}
where we also include results obtained at the shower level.
We notice that \hvq{}, in spite of the fact that it does not implement NLO
corrections in top decay, yields results and distributions that are quite
close to those of the most accurate \bbfourl{} generator. This is due to the
fact that \PythiaEightPtwo{} includes matrix-element corrections~(MEC) in top decay by
default, and MEC are equivalent, up to an irrelevant normalization factor, to
next-to-leading order corrections in decay. This observation is confirmed by
examining, in Tab.~\ref{tab:mwbj_MEC},
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/MEC_table}}
\caption{$\mwbj{}$ peak position for $\mt$=172.5~GeV obtained with the three
different generators, showered with \PythiaEightPtwo{}+MEC~(default). We
also show the differences between \PythiaEightPtwo{}+MEC and
\PythiaEightPtwo{} without MEC. }
\label{tab:mwbj_MEC}
\end{table*}
the impact of the MEC setting on our predictions. When MEC are switched off,
we see a considerable shift, near 1~GeV, in the \hvq{} result for the peak
position in the smeared distribution, and a very minor one in the \bbfourl{}
and \ttbnlodec{} generators, that include the hardest emission off $b$
quarks. Thus, we conclude that the MEC in \PythiaEightPtwo{} do a decent job in
simulating top decay as far as the \mwbj{} distribution is concerned. The
remaining uncertainty of roughly \diffaverage~MeV in the case of both \hvq{}
and \ttbnlodec{} generators, pulling in opposite directions, is likely due to
the approximate treatment of off-shell effects.
\subsubsection{Renormalization- and factorization-scale dependence}
\label{sec:fac_ren_scales}
In this section, we study the dependence of our results on the
renormalization and factorization scales ($\ensuremath{\mu_{\sss \mathrm{R}}}$ and $\ensuremath{\mu_{\sss \mathrm{F}}}$), that gives an
indication of the size of higher-orders corrections. We varied $\ensuremath{\mu_{\sss \mathrm{R}}}$ and
$\ensuremath{\mu_{\sss \mathrm{F}}}$ around the central scale $\mu$ defined in eqs.~\eqref{eq:centralscale}
and~(\ref{eq:centralscaleZ}) according to
\begin{equation}
\ensuremath{\mu_{\sss \mathrm{R}}} = \ensuremath{K_{\sss \mathrm{R}}}\, \mu \, , \quad \ensuremath{\mu_{\sss \mathrm{F}}} = \ensuremath{K_{\sss \mathrm{F}}} \, \mu \, ,
\end{equation}
where $(\ensuremath{K_{\sss \mathrm{R}}},\ensuremath{K_{\sss \mathrm{F}}})$ are varied over the following combinations
\begin{equation} \label{eq:scalechoices}
\bigg\{ (1,1), \, (2,2),\, \left( \frac{1}{2}, \frac{1}{2} \right),
\, (1,2), \, \left( 1, \frac{1}{2} \right), \, (2,1),
\, \left( \frac{1}{2},1 \right) \bigg\}.
\end{equation}
We take $\ensuremath{K_{\sss \mathrm{R}}}=\ensuremath{K_{\sss \mathrm{F}}}=1$ as our central prediction. We find that for \bbfourl{}
there is a non-negligible scale dependence, that in the smeared case yields a
theoretical uncertainty of~${}^{+\varscalemax}_{-\varscalemin}$~MeV. For
\ttbnlodec{} and \hvq{} this uncertainty is smaller than
\varscaleothers~MeV. This is due to the fact that, in the last two
generators, the NLO corrections are performed for on-shell tops, and the top
width is subsequently generated with a smearing procedure. Thus, NLO
corrections remain constant around the top peak, leading to a constant scale
dependence. This leads to an underestimate of scale uncertainties in
\ttbnlodec{} and \hvq{}.
\subsubsection{PDF set dependence}
\label{sec:PDF_dependence}
We evaluated the dependence from the PDFs
by considering the central member of the following PDF sets:
\begin{itemize}
\item {\tt MSTW2008nlo68cl ($\as(\mZ)=0.120179$)} (default)~\cite{Martin:2009iq},
\item {\tt PDF4LHC15\_nlo\_30\_pdfas ($\as(\mZ)=0.118$)}~\cite{Butterworth:2015oua}\,,
\item {\tt CT14nlo ($\as(\mZ)=0.118$)}~\cite{Dulat:2015mca}\,,
\item {\tt MMHT2014nlo68cl ($\as(\mZ)=0.120$)}~\cite{Harland-Lang:2014zoa}\,,
\item {\tt NNPDF30\_nlo\_as\_0118 ($\as(\mZ)=0.118$)}~\cite{Ball:2014uwa}\,.
\end{itemize}
We generated the events by using the {\tt MSTW2008nlo68cl} set, and obtained
all other predictions using the internal reweighting facility of the
\POWHEGBOX{}. We find that the corresponding differences in the \mwbj{} peak
position are typically below \varPDF~MeV and the variations are very similar
for all the NLO+PS generators.
We also generated a sample using the central parton-distribution function of
the {\tt PDF4LHC15\_nlo\_30\_pdfas} set, and, by reweighting, all its
members, within the \hvq{} generator. In this case, our error is given by the
sum in quadrature of all deviations. We get a variation of
\pdferrorhvqnosmear~MeV in the non-smeared case, and \pdferrorhvqsmear~MeV for
the smeared distribution. We find that the variation band obtained in this
way contains the central value results for the different PDF sets that we
have considered. It thus makes sense to use this procedure for the estimate
of PDF uncertainties. On the other hand, reweighting for the 30 members of
the set in the \bbfourl{} case is quite time consuming, since the virtual
corrections are recomputed for each weight. We thus assume that the PDF
uncertainties computed in the \hvq{} case are also valid for the \bbfourl{}
and \ttbnlodec{} cases, since the dependence on the PDF is mostly due to the
implementation of the production processes, and all our generators describe
it at NLO accuracy, and since we have previously observed that by reweighting
to several PDF sets we get very similar variations for all generators.
In general, PDF uncertainties are rather small. This is probably due to the
fact that, in order to shift the position of the peak, some differences must
be present in the modeling of final-state radiation~(FSR). These differences
may arise from differences in $\as$. However, reweighting in \POWHEG{} only
affects the inclusive cross section, and not the radiation, and thus
final-state radiation is not modified by these changes.
\subsubsection{Strong-coupling dependence}
\label{sec:as_dependence}
In \POWHEGBOX{} the scale used to generate the emissions is the transverse
momentum of the radiation (with respect to the emitter). At the moment,
facilities to study uncertainties due to variations of this scheme are not
available. On the other hand, these uncertainties would lead to a different
radiation pattern around the $b$ jet, that can in turn have a non-negligible
effect on the reconstructed mass.
The simplest way at our disposal for studying the sensitivity of the
reconstructed mass to the intensity of radiation from the $b$ quark is by
varying the value of $\as$. To this end we use the {\tt NNPDF30\_nlo\_as115}
and {\tt NNPDF30\_nlo\_as121} sets, where $\as(\mZ)$=0.115 and
$\as(\mZ)$=0.121, respectively. As stated earlier, we cannot use the
\POWHEG{} reweighting facility in order to study this effect, and thus we
generated two dedicated samples (see Tab.~\ref{tab:samples}).
We found that the extracted peak positions in the smeared \mwbj{}
distributions for the two extreme values of $\as$ differ by
\deltaalphasbbfourl~MeV for the \bbfourl{} generator, by
\deltaalphasttdec~MeV for the \ttbnlodec{} generator and by
\deltaalphashvq~MeV for \hvq{}. The small $\as$-sensitivity in the \hvq{}
case is expected, since, in this case, radiation in decays is handled by the
shower, and thus should be studied by varying shower parameters. In the
\bbfourl{} and \ttbnlodec{} case, the variation is very similar, since they
both include NLO radiation in decay, and the direction of the variation is as
expected, i.e.~the peak position is larger for the smaller $\as$ value, due
to the reduced loss of energy outside the jet cone. Differences in the case
of non-smeared distributions are in all cases not larger than
\deltaalphasunsmear~MeV.
We can estimate the typical scale of radiation in top decay as being of the
order of 30~GeV, i.e.~one-half of the typical $b$ energy in the top rest
frame. The ratio of the upper to lower $\as(\mZ)$ values that we have
considered is 1.052, and it becomes 1.06 at a scale of $30$~GeV. On the other
hand, a scale variation of a factor of two above and below $30$~GeV yields a
variation in $\as$ of about 26\%. This can be taken as a rough indication
that a standard scale variation would yield to a variation in the peak
position that is more than a factor four larger than the one obtained by
varying $\as$.
\subsubsection{Matching uncertainties}
The {\tt FSREmission} veto procedure (i.e.~implementation 1 of
Sec.~\ref{sec:PY8_different_showers}) represents the most accurate way to
perform the vetoed shower on the \POWHEGBOX{} generated events, because it
uses the \POWHEG{} definition of transverse momentum rather than the
\PythiaEightPtwo{} one. The {\tt ScaleResonance} procedure (i.e.~method 2) introduces
a mismatch (see Sec.~\ref{sec:PY8_different_showers}) that we take as an
indication of the size of the matching uncertainties. The extracted peak
position for the \bbfourl{} and \ttbnlodec{} with the two matching procedures
are summarized in Tab.~\ref{tab:mass_extraction-matching}.
\begin{table*}[tb]
\centering
{\input{./tables/matching_variations_table.tex}}
\caption{$\mwbj{}$ peak position for $\mt$=172.5~GeV obtained with the
\bbfourl{} and \ttbnlodec{} generators, showered with \PythiaEightPtwo{},
for the {\tt ScaleResonance}~({\tt SR}) veto procedure. The differences
with {\tt FSREmission}~({\tt FSR}), that is our default, are also shown.}
\label{tab:mass_extraction-matching}
\end{table*}
We can see that these differences are roughly 20~MeV in \bbfourl{} for both
the no-smearing and smearing case, and in \ttbnlodec{} they are a few MeV for
the no-smearing case, and 20~MeV with smearing. When using the generic veto
method of Sec.~\ref{sec:genericmethod} we find differences of comparable
size.
\subsubsection[Summary of scale, PDF and $\as$ variations]{Summary of scale, PDF and $\boldsymbol{\as}$ variations}%
\label{sec:mwbjSummary}
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/th_error_table.tex}}
\caption{Theoretical uncertainties associated with the $\mwbj{}$ peak
position extraction for $\mt$=172.5~GeV for the three different generators,
showered with \PythiaEightPtwo{}. The PDF uncertainty on the \bbfourl{} and
\ttbnlodec{} generators is assumed to be equal to the \hvq{} one, as
explained in Sec.~\ref{sec:PDF_dependence}.}
\label{tab:mass_extraction-errors}
\end{table*}
In Tab.~\ref{tab:mass_extraction-errors} we summarize the uncertainties due
to scale, PDF and strong-coupling variations, connected with the extraction
of the \mwbj{} peak position, for the input mass $\mt=172.5$~GeV, for all the
generators showered with \PythiaEightPtwo{}.
The upper (lower) error due to scale variation reported in the table is
obtained by taking the maximum (minimum) position of the \mwbj{} peak for
each of the seven scales choices of eq.~(\ref{eq:scalechoices}),
minus the one obtained for the central scale.
In the PDF case, as discussed in Sec.~\ref{sec:PDF_dependence}, we compute the
PDF uncertainties only for the \hvq{} generator, and assume that they are the
same for \bbfourl{} and \ttbnlodec{}.
We consider a symmetrized strong-coupling dependence uncertainty, whose
expression is given by
\begin{equation}
\delta { \mwbj\left(\as(\mZ)\right)} = \pm \frac{\left| \mwbj(0.115) -\mwbj(0.121)\right|}{2}\,.
\end{equation}
We stress that these variations have only an indicative meaning. In a
realistic analysis, experimental constraints may reduce these uncertainties.
We also stress that these are not the only theoretical uncertainties. Others
may be obtained by varying Monte Carlo parameters. Here we focus specifically
on those uncertainties that are associated with the NLO+PS generators.
As we have already discussed, the use of the \hvq{} and the \ttbnlodec{}
generators would lead to a negligible bias in the \mwbj{} distribution if we
were able to measure it without any resolution effects. However, if we
introduce a smearing to mimic them, the description of the region away from
the peak plays an important role, and the \hvq{} and \ttbnlodec{} generators
yield predictions for the mass peak position that are shifted by roughly
\diffaverage~MeV in the downward and upward direction respectively with
respect to \bbfourl{}.
We also notice that the \bbfourl{} generator is the most affected by
theoretical uncertainties. In particular, the \ttbnlodec{} and \hvq{}
generators have an unrealistically small scale dependence of the peak shape,
due to the way in which off-shell effects are approximately described. The
\ttbnlodec{} generator displays a non-negligible sensitivity only to the
strong-coupling constant. The theoretical errors that we have studied here
lead to very small effects for the \hvq{} generator, since it does not
include radiative corrections in the top decay. On the other hand, the \hvq{}
generator is bound to be more sensitive to variation of parameters in
\PythiaEightPtwo{}, that in this case fully controls the radiation from the
$b$ quark.
\subsubsection{Radius dependence}
In this section we investigate the stability of the previous results with
respect to the choice of the jet radius. The results are summarized in
Tab.~\ref{tab:mass_extraction-radius}.
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/radius_table.tex}
}
\caption{ \mwbj{} peak position obtained with the \bbfourl{} generator for
three choices of the jet radius. The differences with the \ttbnlodec{} and
the \hvq{} generators are also shown.}
\label{tab:mass_extraction-radius}
\end{table*}
For the distributions without smearing, the differences between the three
generators are small and decrease as $R$ increases. For the smeared
distributions, the differences between \ttbnlodec{} and \bbfourl{} decrease
as the radius increases, while the difference between the \hvq{} and the
\bbfourl{} generator increases.
The small differences in $R$ dependence among the three generators in the
non-smeared cases can be understood if we consider that differences in the $b$
radiation do not affect much the peak position in the non-smeared
distribution, but rather they affect the strength of the tail on the left
side of the peak. On the other hand, the peak position is affected by
radiation in production and by the underlying-event structure, that is very
similar in the three generators.
It should be noticed that the difference between the displacements of the
\ttbnlodec{} and \hvq{} with respect to \bbfourl{} is less than
\diffdiffRttdec~MeV and \diffdiffRhvq~MeV, respectively, below the current
statistical precision of top-mass measurements. Thus, the good agreement
found among the three generators persists also for different $R$ values.
\subsection{Comparison with \HerwigSevenPone{}}
In order to assess uncertainties due to the showering program, in this
section we compare the results obtained using \HerwigSevenPone{} and
\PythiaEightPtwo{}.
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/cmp_generators_shower.tex}}
\caption{$\mwbj{}$ peak position for $\mt$=172.5~GeV obtained with the three
different generators, showered with \HerwigSevenPone{}~({\HerwigSevenPlot}). The
differences with \PythiaEightPtwo{}~({\PythiaEightPlot}) are also shown.}
\label{tab:mass_extraction-shower}
\end{table*}
In Tab.~\ref{tab:mass_extraction-shower} we compare the \mwbj{} peak position
extracted for the input mass $\mt = 172.5$~GeV using the three generators
showered with \PythiaEightPtwo{} and \HerwigSevenPone{}. For the \hvq{}
generator, the differences are of the order of \pyminushwhvq~MeV for both the
smeared and non-smeared case, but with opposite signs. In the smeared case,
both the \ttbnlodec{} and \bbfourl{} generators yield much larger
differences, of more than 1~GeV.
In Tab.~\ref{tab:mass_extraction-shower-showerOnly}
\begin{table*}[tb]
\centering
{\input{./tables/cmp_generators_shower_showerOnly.tex}}
\caption{Differences between \PythiaEightPtwo{} and \HerwigSevenPone{} in the
extracted $\mwbj{}$ peak position for $\mt$=172.5~GeV obtained with the
three different generators, at the NLO+PS level (PS only) and including
also the underlying events, the multi-parton interactions and the
hadronization~(full).}
\label{tab:mass_extraction-shower-showerOnly}
\end{table*}
we report the differences between the \HerwigSevenPone{} and \PythiaEightPtwo{}
predictions for all the generators, at the NLO+PS level and at the full
hadron level. We notice that at the NLO+PS level and without smearing, the
differences between the two parton-shower programs are negligible. For the
smeared distributions, at both the NLO+PS and full level, the differences are
roughly 1~GeV for the \bbfourl{} and the \ttbnlodec{} generator. For \hvq{}
the differences are considerably smaller, although not quite
negligible. Furthermore, accidental compensation effects seem to emerge in
this case if we compare the peak displacement in the distributions with and
without smearing.
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/radius_table_shower.tex}}
\caption{ Differences in the \mwbj{} peak position obtained matching the
three generators with \PythiaEightPtwo{} and \HerwigSevenPone{}, for three
choices of the jet radius.}
\label{tab:mass_extraction-shower-radius}
\end{table*}
\begin{figure*}[tb]
\centering
\includegraphics[width=\wfigdoub]{figures/bb4l_PY8_HW7_mwbj_nosmear_showerOnly}
\includegraphics[width=\wfigdoub]{figures/bb4l_PY8_HW7_mwbj_nosmear}
\caption{${d\sigma}/{d \mwbj}$ distribution obtained by showering the
\bbfourl{} results with \PythiaEightPtwo{} and \HerwigSevenPone{},
at parton-shower level~(left) and with hadronization and underlying events~(right).}
\label{fig:mwbjshapespyh7}
\end{figure*}
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/bb4l_PY8_HW7_mwbj_smear}
\caption{Smeared ${d\sigma}/{d \mwbj}$ distribution obtained by matching
the \bbfourl{} generator with \PythiaEightPtwo{} and \HerwigSevenPone{}.}
\label{fig:mwbjshapespyh7smeared}
\end{figure}
The origin of these large differences are better understood by looking at the
differential cross sections plotted in Figs.~\ref{fig:mwbjshapespyh7}
and~\ref{fig:mwbjshapespyh7smeared}. In Fig.~\ref{fig:mwbjshapespyh7} we
plot the results for the non-smeared case, at the NLO+PS level~(left) and at
the full hadron level~(right): while the peak position is nearly the same for
both \PythiaEightPtwo{} and \HerwigSevenPone{}, the shape of the curves is very
different around the peak, leading to a different mass peak position when
smearing is applied, as displayed in Fig.~\ref{fig:mwbjshapespyh7smeared}.
We notice that in this last case we see a difference in shape also after
smearing. This suggests that at least one of the two generators may not
describe the data fairly.
Since we observe such large differences in the value of $\mwbjmax$ in
\HerwigSevenPone{} and \PythiaEightPtwo{}, we have also studied whether sizeable
differences are also present in the $\mwbjmax$ dependence upon the jet radius
$R$. The results are shown in Tab.~\ref{tab:mass_extraction-shower-radius},
and displayed in Fig.~\ref{fig:R_mwbj_py8-hw7}.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/R_mwbj_py8_hw7}
\caption{Differences of $\mwbjmax$ between the \PythiaEightPtwo{} and the
\HerwigSevenPone{} showers, for the three generators, as a function of the
jet radius.}
\label{fig:R_mwbj_py8-hw7}
\end{figure}
In the case of the \bbfourl{} generator, the difference between
\PythiaEightPtwo{} and \HerwigSevenPone{} goes from \pyminushwRfour{} to
\pyminushwRsix~MeV. Thus, assuming for instance that \PythiaEightPtwo{} fits
the data perfectly, i.e.~that it extracts the same value of the mass by
fitting the $\mwbjmax$ values obtained with the three different values of
$R$, \HerwigSevenPone{} would extract at $R=0.6$ a mass value that is larger
by \pyminushwdeltaRfoursix{}~MeV from the one extracted at $R=0.4$. We
stress that the differences in the $R$ behaviour of $\mwbjmax$ may have the same
origin as the difference in the reconstructed mass value, since both effects
may be related to the amount of energy that enters the jet cone, and it is
not unlikely that, by tuning one of the two generators in such a way that they
both have the same $R$ dependence, their difference in $\mwbjmax$ would
also be reduced.\footnote{Similarly, one could fit appropriate calibration
observables associated to the $b$-jet structure, along the lines of
Ref.~\cite{Corcella:2017rpt}.} It is unlikely, however that this would lead
to a much improved agreement, since the difference in slope is much less
pronounced than the difference in absolute value.
\subsubsection{Alternative matching prescriptions in \HerwigSevenPone{}}
We have examined several variations in the \HerwigSevenPone{} settings, and in
the interface between \POWHEG{} and \HerwigSevenPone{}, in order to understand
whether the \HerwigSevenPone{} results are reasonably stable, or depend upon our
particular settings.
\subsubsection*{MEC and \POWHEG{} options in \HerwigSevenPone{}}
\HerwigSevenPone{} applies matrix-element corrections by default, but it also
offers the possibility to switch them off. In addition, it allows to
optionally replace the MEC with its internal \POWHEG{} method, when
available, to achieve NLO accuracy in top decays.\footnote{These options are
activated by the instructions\fignewline {\tt set
ShowerHandler:HardEmission None} or \fignewline {\tt set
ShowerHandler:HardEmission POWHEG}, respectively.} We have verified
that, as expected, switching off the matrix-element corrections does not
significantly affect the \bbfourl{} and \ttbnlodec{} results. In the case of
the \hvq{} generator, we can compare the default case, where MEC is on, with
the cases where \POWHEG{} replaces MEC, and with the case where neither MEC
nor \POWHEG{} is implemented. The results are shown in
Tab.~\ref{tab:mass_extraction-hvq-mec-powheg-herwig}.
\begin{table}[tb]
\centering
{\input{./tables/MEC_POWHEG_hw7_hvq_table.tex}}
\caption{Differences in the $\mwbj{}$ peak position for the \hvq{} generator
showered with \HerwigSevenPone{}, with MEC switched off (no MEC) or using the
\HerwigSevenPone{} \POWHEG{} option, with respect to our default setting, that
has MEC switched on.}
\label{tab:mass_extraction-hvq-mec-powheg-herwig}
\end{table}
We notice that the inclusion of MEC enhances by more than 1.3~GeV the peak
position of the smeared distribution. A similar result was found in
\PythiaEightPtwo{} (see Tab.~\ref{tab:mwbj_MEC}), where the difference was
slightly less than 1~GeV. The difference between the \POWHEG{} and MEC
results is much below the 1~GeV level but not negligible. This fact is hard
to understand, since the \POWHEG{} and MEC procedures should only differ by a
normalization factor.
We have seen previously that the three NLO+PS generators interfaced to
\PythiaEightPtwo{} yield fairly consistent results for the reconstructed top
mass peak. The same consistency is not found when they are interfaced to
\HerwigSevenPone{}. However, the best agreement is found when the internal
\POWHEG{} option for top decay is activated in \HerwigSevenPone{}, as can be
seen in Tab.~\ref{tab:cmp_herwig_table}.
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/cmp_herwig_table.tex}}
\caption{Differences of \hvq{} and \ttbnlodec{} with respect to \bbfourl{},
all showered with \HerwigSevenPone{}. The result obtained using the
\HerwigSevenPone{} internal \POWHEG{} implementation of top decay, rather
than MEC, labelled as \hvq+PWG, is also shown.}
\label{tab:cmp_herwig_table}
\end{table*}
The difference between the \POWHEG{} and MEC or \POWHEG{} \HerwigSevenPone{}
results is puzzling, since they have the same formal accuracy. We will
comment about this issue later on.
\subsubsection*{Alternative veto procedures in \HerwigSevenPone{}}
\label{sec:HW7_different_showers_results}
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/matching_variations_hw7_table.tex}}
\caption{$\mwbj{}$ peak position for $\mt$=172.5~GeV for \bbfourl{} and
\ttbnlodec{} showered with \HerwigSevenPone{} using the {\tt
FullShowerVeto}~({\tt FSV}) procedure. The differences with {\tt
ShowerVeto}~({\tt SV}), that represents our default, are also shown. }
\label{tab:mass_extraction-matching-hw7}
\end{table*}
As discussed in Sec.~\ref{sec:HW7_different_showers}, \HerwigSevenPone{}
offers two different classes that implement the veto procedure: the {\tt
ShowerVeto}, our default one, and the {\tt FullShowerVeto} class. The
corresponding results are summarized in
Tab.~\ref{tab:mass_extraction-matching-hw7}. For both the \bbfourl{} and the
\ttbnlodec{} the two procedures lead to a 200~MeV difference in the peak
position for the smeared distributions. The origin of such difference is not
fully clear to us. In part it may be ascribed to the fact that when using the
{\tt ShowerVeto} class we mix two different definitions of transverse
momentum (the \HerwigSevenPone{} and the \POWHEG{} one), and in part may be
due to the fact that in the {\tt FullShowerVeto} class the vetoing is done on
the basis of the shower structure after reshuffling has been applied. We
have also checked that the generic procedure of Sec.~\ref{sec:genericmethod},
although much slower, leads to results that are statistically compatible with
the {\tt FullShowerVeto} method.
\subsubsection*{Truncated showers}
It was shown in Ref.~\cite{Nason:2004rx} that, when interfacing a \POWHEG{}
generator to an angular-ordered shower, in order to compensate for the
mismatch between the angular-ordered scale and the \POWHEG{} hardness, that is
taken equal to the relative transverse momentum in radiation, one should
supply appropriate truncated showers. None of our vetoing algorithms take
them into account, but it turns out that \HerwigSevenPone{} provides
facilities to change the settings of the initial showering scale according to
the method introduced in Ref.~\cite{Schofield:2011zi}, that, in our case, are
equivalent to the inclusion of truncated showers (see \writeApp\ref{app:TS}). This is
done by inserting the following instructions in the \HerwigSevenPone{} input file:
\begin{equation}
\label{eq:TSsettings}
\begin{split}
& \mbox{\tt
set PartnerFinder:PartnerMethod Maximum} \\
& \mbox{\tt set
PartnerFinder:ScaleChoice Different}.
\end{split}
\end{equation}
The effects of these settings for the \bbfourl{} and \ttbnlodec{} generators
are shown in Tab.~\ref{tab:mass-extraction-truncated-shower-hw7}.
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/hw7_truncated_shower_table.tex}}
\caption{$\mwbj{}$ peak position for $\mt$=172.5~GeV obtained with the
\bbfourl{} and \ttbnlodec{} generators showered with \HerwigSevenPone{},
with the settings of eq.~(\ref{eq:TSsettings}) (labelled as TS). The
differences with the default results are also shown.}
\label{tab:mass-extraction-truncated-shower-hw7}
\end{table*}
The inclusion of the truncated shower does not introduce dramatic changes in
the peak position: in fact the differences are negligible in the
distributions without smearing, and are roughly \hwTSsmearbbfourlttdec~MeV
when smearing is applied. It should be noticed that these settings slightly
increase the difference with respect to the results obtained with
\PythiaEightPtwo{}.
\section{The energy of the $\boldsymbol{b}$ jet}
\label{sec:Ebjet}
In Ref.~\cite{Agashe:2016bok} it was proposed to extract $\mt$ using the peak
of the energy spectrum of the $b$ jet.
At leading order, the $b$ jet consists of the $b$ quark alone, and its energy
in the top rest frame, neglecting top-width effects, is fixed and given by
\begin{equation}
\Ebjmax=\frac{\mt^2-m_W^2+m_b^2}{2\,\mt}\,,
\label{eq:ebjlo}
\end{equation}
i.e.~the spectrum is a delta function in the energy. In the laboratory
frame, because of the variable boost that affects the top, the delta function
is smeared into a wider distribution, but it can be shown that its peak
position remains at $\Ebjmax$. On the basis of this observation we are led to
assume that also after the inclusion of off-shell effects, radiative and
non-perturbative corrections, the relation between $\Ebjmax$ and the top
pole-mass $\mt$ should be largely insensitive to production dynamics.
We performed a study of the $\Ebjmax$ observable along the same lines adopted
for \mwbj{} in the previous section. If the range of variations of the top
mass around a given central value $\mtc$ is small enough, a linear relation
between $\Ebjmax$ and the top mass must hold, so that we can write
\begin{equation}
\label{eq:B-for-ebj}
\Ebjmax(\mt)= \Ebjmax(\mtc) +B\,(\mt-\mtc)+\mathcal{O}(\mt-\mtc)^2.
\end{equation}
It was suggested in Ref.~\cite{CMS-PAS-TOP-15-002} that the $\Ebj$
distribution $\mathrm{d} \sigma/\mathrm{d} \Ebj$ is better fitted in terms of $\log
\Ebj$. Thus, in order to extract the peak position, we fitted the energy
distribution with a fourth order polynomial
\begin{equation}
y=a+b(x-x^{\rm max})^2+c(x-x^{\rm max})^3+d(x-x^{\rm max})^4\,,
\end{equation}
where $x=\log\Ebj$.
The parameter $B$ of eq.~\eqref{eq:B-for-ebj}, extracted from a linear fit of
the three \Ebjmax{} values corresponding to the three different values of
$\mt$ that we have considered (see Tab.~\ref{tab:samples}) using the \hvq{}
generator showered by \PythiaEightPtwo{}, was found to be
\begin{equation}
\label{eq:B_Ebj}
B= \BfromEbjhvq \pm \BerrfromEbjhvq \, ,
\end{equation}
compatible with the expected value of 0.5 from
eq.~\eqref{eq:ebjlo}.\footnote{When using the \bbfourl{} generator we obtain
$B= \BfromEbjbbfourl \pm \BerrfromEbjbbfourl$, while with the \ttbnlodec{}
one, we get $B= \BfromEbjttdec \pm \BerrfromEbjttdec$. When using
\HerwigSevenPone{} instead of \PythiaEightPtwo{}, we find values compatible with the given
ones within 10\%{}.}
\subsection{Comparison among different NLO+PS generators}
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/bb4l_ttdec_hvq_PY8_Ebj}
\caption{Logarithmic energy distribution obtained with the three generators interfaced
to \PythiaEightPtwo, together with their polynomial fit, in the range
displayed in the figure. The value of $\Ebjmax$ for each generator is
also reported.}
\label{fig:Ebj_bb4l_ttdec_hvq}
\end{figure}
In Fig.~\ref{fig:Ebj_bb4l_ttdec_hvq} we plot the logarithmic energy
distribution for the three generators interfaced to \PythiaEight, together
with their polynomial fit. The extracted \Ebj{} peaks from the \bbfourl{}
and the \ttbnlodec{} generators are compatible within the statistical
errors. On the other hand, the \hvq{} generator yields a prediction which is
roughly \Ebjbbfourlmhvq{} $\pm$ \Ebjbbfourlmhvqerr~MeV smaller than the
\bbfourl{} one. We thus observe that the jet modeling implemented by
\PythiaEightPtwo{} with MEC seems to yield slightly less energetic jets. An
effect going in the same direction was also observed for the \mwbj{}
observable (see Tab.~\ref{tab:mass_extraction-errors}, the first column of
the results with smearing), although to a smaller extent.
\begin{table}[tb]
\centering
{\input{./tables/MEC_table_Ebj}}
\caption{\Ebj{} peak position obtained with the three generators showered
with \PythiaEightPtwo{}. The differences between the peak positions extracted
by switching on and off the matrix-element corrections are also shown.}
\label{tab:Ebj_MEC}
\end{table}
In Tab.~\ref{tab:Ebj_MEC} we have collected the values of $\Ebjmax$ computed
with MEC, and the differences between the results with and without MEC. We
notice that the MEC setting has little impact in the \bbfourl{} and
\ttbnlodec{} cases. On the other hand, in the \hvq{} case the absence of MEC
would have lead to an $\Ebjmax$ value about 2~GeV smaller than with MEC. We
take this as another indication that the implementation of radiation in top
decay using MEC leads to results that are much closer to the NLO+PS ones.
In Tab.~\ref{tab:Ebj_extraction-errors}
\begin{table*}[tb]
\centering
{\input{./tables/th_error_table_Ebj}}
\caption{Theoretical uncertainties for the $\Ebj${} peak position
obtained with the three generators showered with \PythiaEightPtwo{}.
The last column reports the statistical uncertainty of our results.}
\label{tab:Ebj_extraction-errors}
\end{table*}
we summarize our results together with the scale, PDF and $\as$
uncertainties, that are extracted with a procedure analogous to one described
for the \mwbj{} observable. We also report the corresponding
statistical errors of our results. We see that scale and PDF variations have
negligible impact on our observable, the only important change being
associated with the choice of the NLO+PS generator.
We notice that our errors on scale and PDF variations are much smaller than
our statistical errors. On the other hand, these variations are performed by
reweighting techniques, that, because of correlations, lead to errors in the
differences that are much smaller than the error on the individual term. In
view of the small size of these variations, we do not attempt to perform a
better estimate of their error. On the other hand, the variation of $\as$ do
not benefit from this cancellation, and are all below the statistical
uncertainties.
As previously done for \mwbj{}, we have also investigated the dependence of
the \bjet{} peak positions on the jet radius. The results are summarized in
Tab.~\ref{tab:Ebj_cmp_bb4l_allradii}.
\begin{table*}[h!tb]
\centering
{\input{./tables/cmp_bb4l_Ebj}}
\caption{\Ebj{} peak position obtained with the \bbfourl{} generator showered
with \PythiaEightPtwo{}, for three choices of the jet radius. The differences
with the \ttbnlodec{} and the \hvq{} generators are also shown.}
\label{tab:Ebj_cmp_bb4l_allradii}
\end{table*}
While we observe a marked change in \Ebjmax{}, that grows by $\diffEbjcml$
and $\diffEbjumc$~GeV when going from $R=0.4$ to $0.5$ and from $0.5$ to
$0.6$ respectively, \ttbnlodec{} and \hvq{} differ by \bbfourl{} by much
smaller amounts. It is not clear whether such small differences could be
discriminated experimentally.
According to eqs.~(\ref{eq:delta_mt}) and~(\ref{eq:B_Ebj}), the uncertainties
that affect the value of the extracted top mass are nearly twice the
uncertainties on the \bjet{} energy. Considering the difference for $R=0.5$
between \hvq{} and \bbfourl{} in Tab.~\ref{tab:Ebj_cmp_bb4l_allradii}, we see
that, by using \hvq{} instead of \bbfourl{}, the extracted top mass would be
roughly 900~MeV larger. This should be compared with the corresponding
difference of about~150~MeV, that is shown in
Tab.~\ref{tab:mass_extraction-radius}, for the smeared \mwbj{} case.
As before, we have checked the sensitivity of our result to variations in the
matching procedure in \PythiaEightPtwo{}, by studying the difference between {\tt
ScaleResonance} and {\tt FSREmission} options. The differences turn out to
be of the order of the statistical error.
\subsection{Comparison with \HerwigSevenPone{}}
In this section, we study the dependence of our results on the shower MC
program, comparing \HerwigSevenPone{} and \PythiaEightPtwo{} predictions. We extract
the differences in the \Ebjmax{} position for three values of the jet radius:
$R=0.4$, 0.5 and 0.6. The results are summarized in
Tab.~\ref{tab:Ebj_py8-hw7}, where we also show the results at the
PS-only level,
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/cmp_shower_showerOnly_Ebj}}
\caption{Differences in the \Ebj{} peak position between the \PythiaEightPtwo{}
and the \HerwigSevenPone{} showers applied to the three generators for three
choices of the jet radius. The results at the NLO+PS level (PS only) are
also shown.}
\label{tab:Ebj_py8-hw7}
\end{table*}
and in Fig.~\ref{fig:R_Ebj_py8-hw7}.
\begin{figure}[tb]
\centering
\includegraphics[width=\wfigsing]{figures/R_Ebj_py8_hw7}
\caption{Differences of $\Ebjmax$ between the \PythiaEightPtwo{} and the
\HerwigSevenPone{} showers, for the three generators, as a function of the
jet radius.}
\label{fig:R_Ebj_py8-hw7}
\end{figure}
From Tab.~\ref{tab:Ebj_py8-hw7} we clearly see that the \bbfourl{} and the
\ttbnlodec{} generators display larger discrepancies. For example, for the
central value $R=0.5$, we would get $\Delta \Ebjmax{}\approx 2$~GeV, that
roughly corresponds to $\Delta \mt=-4$~GeV. In the case of the \hvq{}
generator the difference is near 1~GeV, implying that the extracted mass
using \hvq{}+\HerwigSevenPone{} would be 2~GeV bigger than the one obtained with
\hvq{}+\PythiaEightPtwo{}.
We find that the differences between \HerwigSevenPone{} and \PythiaEightPtwo{}
increases for larger jet radii. Furthermore, by looking at
Fig.~\ref{fig:R_Ebj_py8-hw7}, we notice that the \bbfourl{} generator
displays a different $R$ dependence, as we have already observed from
Tab.~\ref{tab:Ebj_cmp_bb4l_allradii}. Figure~\ref{fig:R_Ebj_py8-hw7}
indicates that \bbfourl{} and \ttbnlodec{} are in better agreement for larger
values of the jet radius. This was also observed for the peak of the \mwbj{}
smeared distribution~(Tab.~\ref{tab:mass_extraction-radius}).
We notice that, as in the case of the reconstructed mass peak, the
predominant contribution to the difference arises at the parton shower level.
As for the previous cases, we have examined the variations due to a different
choice of the matching scheme in \HerwigSevenPone{}, that we found to be below
the 200~MeV level, and thus negligible in the present context.
\section{Leptonic observables}
\label{sec:LepObs}
In this section, we investigate the extraction of the top mass from the
leptonic observables introduced in Ref.~\cite{Frixione:2014ala}. This method
has been recently studied by the ATLAS collaboration in
Ref.~\cite{ATLAS-CONF-2017-044}.
Following Ref.~\cite{Frixione:2014ala}, we consider the subsequent five observables
\begin{equation}
\begin{array}{l l l }
O_1 = \pT(\ell^+), & O_2= \pT(\ell^+\ell^-), & O_3= m(\ell^+\ell^-), \\[2mm]
O_4= E(\ell^+\ell^-),\quad & O_5= \pT(\ell^+)+\pT(\ell^-), &
\end{array}
\nonumber
\end{equation}
i.e.~the transverse momentum of the positive charged lepton, and the
transverse momentum, the invariant mass, the energy and the scalar sum of the
transverse momenta of the lepton pair. We compute the average value of the
first three Mellin moments for each of the above mentioned observables,
$\langle (O_i)^j\rangle$, with $i=1,\dots,5$ and $j=1,2,3$. We assume that,
if we do not vary too much the range of the top mass, we can write the linear
relation
\begin{equation}
\langle (O_i)^j \rangle =O_{\rm c}^{(ij)} + B^{(ij)} \lq \(\mt\)^j- \(\mtc\)^j \rq.
\label{eq:leptonicObs}
\end{equation}
For ease of notation, we will refer to $O_{\rm c}^{(ij)}$ and $B^{(ij)}$ as
$O_{\rm c}$ and $B$ in the following.
Their determination will be discussed later.
We choose as reference sample the one generated with \bbfourl{} matched with
\PythiaEightPtwo{}, using $\mtc=172.5$~GeV as input mass and the central
choices for the PDF and scales. We indicate the values of the observables
computed with this generator as ${O}^{b\bar{b}4\ell}$, and with $O_{\rm c}'$
the values of the observable computed either with an alternative generator or
with different generator settings, but using as input parameter the same
reference mass. The mass value that we would extract from the events of the
reference sample using the new generator is then given by
\begin{equation}
\label{eq:mtprime}
\mt' = \left[\(\mtc\)^j -\frac{O'_{\rm c}-{O}^{b\bar{b}4\ell }_{\rm c}}{B} \right]^{1/j}\,.
\end{equation}
\subsection{Comparison among NLO+PS generators}
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/th_error_table_py8_lept.tex}}
\caption{The average values of each leptonic observable computed with
\bbfourl{}, \ttbnlodec{} and \hvq{}, showered with \PythiaEightPtwo{},
for $\mt$=172.5~GeV, and their variations with respect to \bbfourl{} are
shown in the first two columns. The differences with respect to their
corresponding central values due to scale and PDF variations are also
shown in columns three and four. Their $\as$ uncertainties, computed as
described in Sec.~\ref{sec:as_dependence} are displayed in column five.
The statistical errors are also reported, except for the scale and PDF
variations, where they have been estimated to be below 13\%{} of the
quoted values.}
\label{tab:Olep-summary-py}
\end{table*}
\begin{table*}[tb]
\centering
\resizebox{\textwidth}{!}
{\input{./tables/th_error_table_hw7_lept.tex}}
\caption{As in Tab.~\ref{tab:Olep-summary-py} but for \HerwigSevenPone{}.}
\label{tab:Olep-summary-hw}
\end{table*}
We begin by showing in Tabs.~\ref{tab:Olep-summary-py}
and~\ref{tab:Olep-summary-hw} the average values of the leptonic observables
computed with our three NLO+PS generators interfaced with \PythiaEightPtwo{}
and \HerwigSevenPone{}. We show the central values, the differences with
respect to \bbfourl{}, and the upper and lower results induced by scale, PDF
and $\as$ variations.
The scale and PDF variations are performed by reweighting. As a consequence
of that, the associated error is much smaller than the statistical error on
the cross section. In order to estimate it, we have divided our sample of
events in ten sub-samples, computed the observables for each sub-sample, and
carried out a straightforward statistical analysis on the ten sets of
results. We found errors that never exceed the quoted value by more than
13\%.
For the PDF variation, we have verified that differences due to variations in
our reference PDF sets (see Sec.~\ref{sec:PDF_dependence}) are very similar
among the different generators. On the other hand, a full error study using
the {\tt PDF4LHC15\_nlo\_30\_pdfas} set was only performed with the \hvq{}
generator, and the associated errors exceed by far the variation band that we
obtain with our reference sets. Thus, also in this case we quote the PDF
variations only for \hvq{}, implying that a very similar variation should
also be present for the others. It is clear from the tables that the PDF
uncertainties are dominant for several observables, and scale variations are
also sizeable.
The large variations in the $\as$ column are not always conclusive because of
the large statistical errors (in parentheses), due to the fact that we cannot
perform this variation by reweighting. However, unlike for the $\mwbj$ case,
here the PDF dependence is not small, and thus we cannot conclude that the
$\as$ variation probes mainly the sensitivity to the intensity of radiation
in decay, since when we vary $\as$ we change also the PDF set.
It is instead useful to look at the effect of MEC on the leptonic
observables, displayed in Tab.~\ref{tab:leptobs_MEC}.
\begin{table*}[tb]
\centering
{\input{./tables/MEC_table_lept.tex}}
\caption{Impact of MEC in \PythiaEightPtwo{} on the leptonic observables for the
different NLO+PS generators.}
\label{tab:leptobs_MEC}
\end{table*}
We observe that in the \bbfourl{} and \ttbnlodec{} case the effect of MEC is
compatible with the statistical uncertainty. In the \hvq{} case we find
instead sizeable effects. This is expected, since large-angle radiation from
the $b$ quark, by subtracting energy to the whole $Wb$ system, affects
significantly also leptonic observables.
In Ref.~\cite{Frixione:2014ala} it was observed that the observables
$\ensuremath{p_{\sss T}}\xspace(\ell^+\ell^-)$ and $m(\ell^+\ell^-)$ had larger errors due to a stronger
sensitivity to radiative corrections, and were more sensitive to
spin-correlation effects. We see a confirmation of this observations in
their larger errors due to scale variation, and in the fact that for \hvq{}
their central value is shifted with respect to the \bbfourl{} and
\ttbnlodec{} generators, that treat spin correlations in a better way.
\begin{table}[tb]
\centering
{\input{./tables/th_error_table-params_lept.tex}}
\caption{Extracted $B$ coefficients for the three different generators
showered with \PythiaEightPtwo{}.}
\label{tab:Bcoeffs-lept}
\end{table}
In Tab.~\ref{tab:Bcoeffs-lept} we show the extracted values of the $B$
coefficients for the first Mellin moment of each observable. The $B$ values
corresponding to the different generators are compatible within the
statistical errors. We thus choose the values computed with the \hvq{}
generator, that have the smallest error. According to
eq.~(\ref{eq:mtprime}), we can translate a variation in an observable into a
variation of the extracted mass, that for the first Mellin moment is simply
obtained applying a $-1/B$ factor. The results are illustrated in
Tab.~\ref{tab:mass_average_lept}.
\begin{table*}[tb]
\begin{center}
\resizebox{\textwidth}{!}
{\input{./tables/extracted_mass_average_lept.tex}}
\caption{Extracted mass in GeV for all the generators, showered with
\PythiaEightPtwo{} and \HerwigSevenPone{}, corresponding to the different leptonic
observables, using as reference sample the \bbfourl{} one generated with
$\mt=172.5$~GeV and showered with \PythiaEightPtwo{}. The quoted errors are
obtained by summing in quadrature the scale, PDF and the statistical
errors. The weighted average is also shown, for all the observables and
considering only their first Mellin moment.}
\label{tab:mass_average_lept}
\end{center}
\end{table*}
The errors shown have been obtained by summing in quadrature the statistical
error and the scale and PDF uncertainties. We have not included the $\as$
variation in the error in order to avoid overcounting, since, in the present
case, is likely to be largely dominated by the change in the associated PDF.
The overall errors on the last two lines of Tab.~\ref{tab:mass_average_lept}
are obtained with the same procedure adopted in Ref.~\cite{Frixione:2014ala}
to account for correlations among the different observables. We do not see
excessive differences among our three generators showered with the same
Monte Carlo generator, while the differences between the \PythiaEightPtwo{} and
\HerwigSevenPone{} results are considerably large. This is also the case for the
\hvq{} generator, that has a much simpler interface to both \PythiaEightPtwo{}
and \HerwigSevenPone{}.
As we did for \mwbjmax{} and \Ebjmax{}, also in the present case we have
computed the leptonic observables without including hadronization effects,
i.e.~at parton-shower only level, in order to determine whether the
differences between \PythiaEightPtwo{} and \HerwigSevenPone{} are due to
the shower or to the hadronization. Our findings are summarized in
Tab.~\ref{tab:lept-PSonly}.
\begin{table}[ht]
\centering
{\input{./tables/py8_minus_hw7-lept.tex}}
\caption{Differences between the \PythiaEightPtwo{} and \HerwigSevenPone{} results
for the leptonic observables, at full hadron level and at parton-level
only.}
\label{tab:lept-PSonly}
\end{table}
Most of the differences already arise at the shower level. We also remark
that, within the same SMC generator, they are not large, yielding differences
in the extracted top mass of the same size as the statistical errors.
We observe in Tab.~\ref{tab:mass_average_lept} that the inclusion of higher
moments of the leptonic observables does not modify appreciably the results
from the first moments. This is a consequence of the large error on the
higher moments, and of the strong correlations among different moments.
The results in Tab.~\ref{tab:mass_average_lept} are also summarized in
Fig.~\ref{fig:leptObs},
\begin{figure}
\centering
\includegraphics[width=\wfigsingmulti]{figures/leptons_mass_extraction_bb4l}\\
\includegraphics[width=\wfigsingmulti]{figures/leptons_mass_extraction_ttdec}\\
\includegraphics[width=\wfigsingmulti]{figures/leptons_mass_extraction_hvq}
\caption{ Extracted mass for the three generators matched with
\PythiaEightPtwo{}~(red) and \HerwigSevenPone{}~(blue) using the first three
Mellin moments of the five leptonic observables. The horizontal band
represents the weighted average of the results, and the black
horizontal line corresponds to $\mt=172.5$~GeV, which is the top
mass value used in the \bbfourl{}+\PythiaEightPtwo{}
reference sample.}
\label{fig:leptObs}
\end{figure}
where the discrepancy between \PythiaEightPtwo{} and \HerwigSevenPone{} and the
mutual consistency of the different observables can be immediately
appreciated.
As for the previous observables, we have studied the effect of changing the
matching scheme, by switching between our two alternative matching schemes
with \PythiaEightPtwo{} and \HerwigSevenPone{}, and by considering the settings of
eq.~(\ref{eq:TSsettings}) in \HerwigSevenPone{}. In both cases we find results
that are consistent within statistical errors.
\section{Summary}
\label{sec:Summary}
In this work we have compared generators of increasing accuracy for the
production and decay of $t\bar{t}$ pairs considering observables suitable for
the measurement of the top mass. The generators that we have considered are:
\begin{itemize}
\item The \hvq{} generator~\cite{Frixione:2007nw}, that implements NLO
corrections in production for on-shell top quarks, and includes
finite-width effects and spin correlations only in an approximate way, by
smearing the on-shell kinematics with Breit-Wigner forms of appropriate
width, and by generating the angular distribution of the decay products
according to the associated tree-level matrix
elements~\cite{Frixione:2007zp}.
\item The \ttbnlodec{} generator~\cite{Campbell:2014kua}, that implements NLO
corrections in production and decay in the narrow-width approximation. Spin
correlations are included at NLO accuracy. Finite width effects are
implemented by reweighting the NLO results using the tree-level matrix
elements for the associated Born-level process, including however all
finite width non-resonant and interference effects at the Born level for
the given final state.
\item The \bbfourl{} generator~\cite{Jezo:2016ujg}, that uses the full matrix
elements for the production of the given final state, including all
non-resonant diagrams and interference effects. This includes interference
of QCD radiation in production and decay.
\end{itemize}
The main focus of our work has been the study of the mass distribution of a
particle-level reconstructed top, consisting of a lepton-neutrino pair and a
$b$-quark jet with the appropriate flavour. The peak position of the mass of
this system is our observable, that is loosely related to the top mass. We
considered its distributions both at the particle level, and by assuming that
experimental inaccuracies can be summarized by a simple smearing with a
resolution function, a Gaussian with a width of 15~GeV, which is the typical
resolution achieved on the top mass by the LHC collaborations. This
observable is an oversimplified version of the mass observables that are used
in direct top-mass measurements, that are the methods that lead to the most
precise mass determinations.
We have found a very consistent picture in the comparison of our three
generators when they are interfaced to \PythiaEightPtwo{}, and thus we begin
by summarizing our results for this case. We first recall \emph{what we
expect} from such comparison. When comparing the \hvq{} and the
\ttNLOdec{} generators, we should remember that the latter has certainly
better accuracy in the description of spin correlations, since it implements
them correctly both at the leading and at the NLO level. However, we do not
expect spin correlations to play an important role in the reconstructed top
mass. As a further point, the \ttNLOdec{} generator implements NLO
corrections in decay. In the \hvq{} generator, the decay is handled by the
shower, where, by default, \PythiaEightPtwo{} includes matrix-element
corrections (MEC). These differ formally from a full NLO correction only by a
normalization factor, that amounts to the NLO correction to the top
width. Thus, as long as the MEC are switched on, we do not expect large
differences between \hvq{} and \ttNLOdec{}. As far as the comparison between
\ttNLOdec{} and \bbfourl{}, we expect the difference to be given by NLO
off-shell effects, and by interference of radiation in production and decay,
since these effects are not implemented in \ttNLOdec{}. This comparison is
particularly interesting, since the interference between production and decay
can be considered as a ``perturbative precursor'' of colour reconnection
effects.
The results of these comparisons can be summarized as follows:
\begin{itemize}
\item The \ttNLOdec{} and the \bbfourl{} generators yield very similar
results for most of the observables that we have considered, implying that
NLO off-shell effects and interference between production and decay are
modest.
\item
As far as \mwbjmax{} (the peak of the reconstructed mass distribution) is
concerned, the \ttNLOdec{} and the \hvq{} generators yield very similar
results, confirming the fact that the MEC implementation in
\PythiaEightPtwo{} has an effect very similar to the \POWHEG{}
implementation of NLO corrections in decay in the \ttNLOdec{}. We have also
observed that, if we switch off the MEC, the agreement between the two
generators is spoiled. More quantitatively, we find that the spread in the
peak of the reconstructed mass at the particle level among the three NLO+PS
generators is never above 30~MeV. On the other hand, if resolution effects
are accounted for with our smearing procedure, we find that the \hvq{}
result is 147~MeV smaller, and the \ttbnlodec{} result 140~MeV larger than
the \bbfourl{} one. These values are safely below currently quoted errors
for the top-mass measurements with direct methods.
If we switch off the MEC in \PythiaEightPtwo{}, we find that the peak position
at the particle level in the \hvq{} case is displaced by 61~MeV, while, if
smearing effects are included, the shift is of $916$~MeV, a rather large
value, that can however be disregarded as being due to the poor accuracy of
the collinear approximation in $b$ radiation when MEC corrections are off.
\item
The jet-energy peak seems to be more sensitive to the modeling of radiation
from the $b$ quark. In fact, while the \ttNLOdec{} and the \bbfourl{}
results are quite consistent with each other, with the peak positions
differing by less than 200~MeV, the \hvq{} result differs from them by more
than 500~MeV. This would correspond to a difference in the extracted mass
of the top quark roughly equal to twice that amount. On the other hand, if
the MEC in \hvq{} are switched off, the shift in the $b$-jet energy peak is
more than 1.9~GeV. This leads us to conclude that the impact of modeling of
$b$ radiation on the $b$-jet peak is much stronger than in the
reconstructed top mass peak. We stress, however, that the difference
between \hvq{} (with MEC on) and the other two generators is safely below
the errors quoted in current measurements~\cite{CMS-PAS-TOP-15-002}.
\item
For the leptonic observables, we generally see a reasonable agreement
between the different generators. The largest differences are found in the
\hvq{} case, for the $\ensuremath{p_{\sss T}}\xspace(\ell^+\ell^-)$ and $m(\ell^+\ell^-)$, larger than
500~MeV with respect to the other two. In Ref.~\cite{Frixione:2014ala} it
was noticed that these observables had larger errors due to a stronger
sensitivity to radiative corrections, and to spin-correlation effects, that
are modelled incorrectly by \hvq{}.
\end{itemize}
Several sources of possible uncertainties have been explored in order to
check the reliability of these conclusions. First of all, two different
matching procedures for interfacing the \ttNLOdec{} and \bbfourl{} generators
to \PythiaEightPtwo{} have been implemented. For example, for the
reconstructed mass peak, we have checked that switching between them leads to
differences below 20~MeV for both generators. The effect of scales,
$\alpha_s$ and PDF uncertainties have also been examined, and were found to
yield very modest variations in the reconstructed mass peak. It was found, in
particular, that scale variations lead to a negligible peak displacement
(below 7~MeV) in the \ttNLOdec{} and \hvq{} case, while the effect is of
${}^{+86}_{-53}$~MeV for \bbfourl{}. The lack of scale dependence in the
\hvq{} and \ttNLOdec{} is easily understood as being due to the fact that the
peak shape is obtained by smearing an on-shell distribution with a
Breit-Wigner form, that does not depend upon any scale, and it suggests that,
in order to get realistic scale-variation errors, the most accurate
\bbfourl{} generator should be used. We have also computed results at the
shower level, excluding the effects of hadronization and multi-parton
interactions, in order to see if the consistent picture found at the hadron
level is also supported by the parton-level results, and we have found that
this is indeed the case.
We have thus seen that the overall picture of the comparison of our three
NLO+PS generators within the framework of the \PythiaEightPtwo{} shower is quite
simple and consistent. For the most precise observable, i.e.~the peak of the
reconstructed mass distribution, it leads to the conclusions that the use of
the most accurate generator may lead to a shift in the measured mass of at
most 150~MeV, which is well below the present uncertainties quoted by the
experimental collaborations.
Our study with \HerwigSevenPone{} instead reveals several problems. We can
summarize our findings as follows:
\begin{itemize}
\item The results obtained with \HerwigSevenPone{} differ substantially from
those obtained with \PythiaEightPtwo{}. In particular, the peak of the
reconstructed mass distribution at the particle level is shifted by -66 and
-39~MeV in the \bbfourl{} and \ttNLOdec{} cases, and by +235 MeV in the
\hvq{} case. When the experimental resolution is accounted for, using our
smearing procedure, the shift raises to -1091 and -1179~MeV in the
\bbfourl{} and \ttNLOdec{} cases, and to -251 MeV in the \hvq{} case.
\item The results obtained within the \HerwigSevenPone{} framework display large
differences between the \hvq{} generator with respect to \bbfourl{} and
\ttNLOdec{} ones. In particular, while the \ttNLOdec{} result exceeds the
\bbfourl{} one only by about 50~MeV in both the particle level and smeared
cases, \hvq{} exceeds \bbfourl{} by 311~MeV at particle level, and by
693~MeV after smearing.
\end{itemize}
These results are quite alarming. The shifts reach values that are
considerably larger than current experimental uncertainties.
In the \hvq{} case, which is the NLO+PS generator currently used for top-mass
studies by the experimental collaborations, the difference in the mass-peak
position between \HerwigSevenPone{} and \PythiaEightPtwo{}, for the smeared
distribution, is -251~MeV, uncomfortably large but still below current
errors. One would then be tempted to conclude that the large shifts may be
linked to some problems concerning the new generators. However, we also
notice that the same difference is +235~MeV when no smearing is applied, so
it is about as large in magnitude but with the opposite sign. This indicates
that the shape of the reconstructed mass distribution is considerably
different in the two shower models. Lastly, if we use the internal \POWHEG{}
implementation of top decay (rather than the MEC) in \HerwigSevenPone{}, the
difference with respect to \PythiaEightPtwo{} raises to 607~MeV. Thus, we
conclude that in the \hvq{} case the smaller difference between
\HerwigSevenPone{} and \PythiaEightPtwo{} is accidental, and is subject to
considerable variations depending upon the settings.
Also in this case we checked whether the MEC yield an improved agreement
between the \hvq{} and the other two generators, as was observed for
\PythiaEightPtwo{}. We find that, by switching off MEC, the
\hvq{}+\HerwigSevenPone{} result decreases by 307~MeV at particle level, and
by 1371~MeV in the smeared case. These effects are qualitatively similar to
what was observed in \PythiaEightPtwo{}. However, in the present case, when
MEC are switched off, the \hvq{} result exceeds the \bbfourl{} one by a
negligible amount at the particle level, and is lower than the \bbfourl{} one
by 678~MeV in the smeared case.
The discrepancy between \hvq{} and the other two generators is mitigated if,
instead of the MEC procedure, the internal \POWHEG{} option of
\HerwigSevenPone{} for top decay is used. In this case, the discrepancy
between \hvq{} and \bbfourl{} is reduced to 244~MeV with no smearing, and to
337~MeV with smearing. We thus see that the consistency of the three NLO+PS
generators interfaced to \HerwigSevenPone{} is not optimal as in
\PythiaEightPtwo{}. It is however acceptable if the internal \POWHEG{}
feature is used rather than MEC in \HerwigSevenPone{}.
We have performed several studies to determine the origin of the difference
between \PythiaEightPtwo{} and \HerwigSevenPone{}, and to check whether it
could be attributed to some problem in our matching procedure. They can be
summarized as follows:
\begin{itemize}
\item
We have shown that the difference is mostly due to the shower model, since
it is already largely present at the parton level.
\item
We have considered the $R$ dependence of the \HerwigSevenPone{} result. It
differs from the one in \PythiaEightPtwo{}, leading to the hope that both
generators may not represent the same set of data well, and tuning them may
reduce their differences. However, we have also noticed that the
difference in slope is much smaller than the difference in size.
\item
We have already mentioned that we have also compared results by making use
of the internal \POWHEG{} implementation of top decay in
\HerwigSevenPone{}, rather than using MEC. We have found non-negligible
differences in this case.
\item
We have implemented alternative veto procedure in the matching of
\HerwigSevenPone{} with the NLO+PS generators. We found differences of the
order of 200~MeV, not large enough to cover the discrepancy with
\PythiaEightPtwo{}.
\item
When interfacing \POWHEG{} generators to angular-ordered showers, in order
to maintain the double-logarithmic accuracy of the shower, one should
introduce the so called ``truncated showers''~\cite{Nason:2004rx}. One
could then worry that the lack of truncated showers is at the origin of the
discrepancies that we found. Fortunately, \HerwigSevenPone{} offers some
optional settings that are equivalent to the introduction of truncated
showers. We found that these options lead to a shift of only 200~MeV in the
peak position.
\end{itemize}
In summary, we found no indication that the discrepancy with
\PythiaEightPtwo{} is due to the specific matching procedure and general
settings that we have used in \HerwigSevenPone{}.
When comparing \HerwigSevenPone{} and \PythiaEightPtwo{} in the computation
of the $b$-jet energy peak, we have found even larger differences: when using
\bbfourl{} and \ttbnlodec{}, the shifts are of the order of 2~GeV, while for
\hvq{} the shift is around 1~GeV. They correspond to differences in the
extracted mass of around 4~GeV in the first two cases, and 2~GeV in the last
one. This is not surprising, in view of the stronger sensitivity of the
$b$-jet peak to the shower model.
Finally, when considering leptonic observables, we find again large
differences between \HerwigSevenPone{} and \PythiaEightPtwo{}. Most
differences already arise at the shower level. Notice that this is in
contrast with the naive view that leptonic observables should be less
dependent upon QCD radiation effects and jet modeling. The comparison between
\HerwigSevenPone{} and \PythiaEightPtwo{} for leptonic observables can by
appreciated by looking at Fig.~\ref{fig:leptObs}, representing the value of
the extracted top mass from a sample generated with \bbfourl{} interfaced to
\PythiaEightPtwo{}.
\section{Conclusions}
\label{sec:Conc}
We focus our conclusions on the results obtained for the reconstructed mass
peak, since the issues that we have found there apply to the direct top mass
measurements, that are the most precise. The experimental collaborations
extensively use the \hvq{} generator for this kind of analyses, and since new
generators of higher accuracy, the \ttbnlodec{} and the \bbfourl{} ones, have
become available, we have addressed the question of whether the physics
effects not included in \hvq{} may lead to inaccuracies in the top-mass
determination. The answer to this question is quite simple and clear when our
generators are interfaced to \PythiaEightPtwo{}. The differences that we find
are large enough to justify the use of the most accurate generators, but not
large enough to drastically overturn the conclusions of current measurements.
Notice that, since the \hvq{} generator does not include NLO corrections in
decays, we might have expected a very different modeling of the $b$-jet in
\hvq{} with respect to the other two generators, leading to important shifts
in the extracted top mass value. It turns out, however, that the
\PythiaEightPtwo{} handling of top decay in \hvq{}, improved with the
matrix-element corrections, does in practice achieve NLO accuracy up to an
irrelevant normalization factor.
This nicely consistent picture does not hold anymore if we use
\HerwigSevenPone{} as shower generator. In particular, it seems that the MEC
implemented in \HerwigSevenPone{} do not have the same effect as the handling of
radiation in decay of our modern NLO+PS generators, leading to values of the
extracted top mass that can differ up to about 700~MeV. Furthermore,
interfacing our most accurate NLO+PS generator (the \bbfourl{} one)
to \HerwigSevenPone{} leads to an extracted top mass of up to 1.2~GeV
smaller with respect to the corresponding result with \PythiaEightPtwo{}.
At this point we have two options:
\begin{itemize}
\item
Dismiss the \HerwigSevenPone{}
results, on the ground that its MEC handling of top decay does not match
our modern generators.
\item
Consider the \HerwigSevenPone{} result as a variation to be included
as theoretical error.
\end{itemize}
We believe that the first option is not soundly motivated. In fact, the
implementation of MEC in \PythiaEightPtwo{} is also \emph{technically} very close
to what \POWHEG{} does. The hardest radiation is essentially generated in the
same way, and in both cases the subsequent radiation is generated with a
lower transverse momentum. Thus the good agreement between the two is not
surprising. The case of \HerwigSevenPone{} is completely different, since in
angular-ordered showers the hardest radiation is not necessarily the
first~\cite{Seymour:1994df}. It is thus quite possible that the differences
we found when \HerwigSevenPone{} handles the decay with MEC, with respect to the
case when \POWHEG{} does, are due to the fact that the two procedures,
although \emph{formally} equivalent (i.e.~both leading to NLO accuracy) are
\emph{technically} different. In this last case, their difference should be
attributed to uncontrolled higher-order effects, and should thus be
considered as a theoretical uncertainty.
A further question that this work raises is whether we should consider the
variation between the \PythiaEightPtwo{} and the \HerwigSevenPone{} programs
as an error that should be added to current top-mass measurements. By doing
so, current errors, that are of the order of 500-600~MeV, would become larger
than 1~GeV. We believe that our crude modeling of the measurement process
does not allow us to draw this conclusion. The analysis procedures used in
direct measurements are much more complex, and involve adequate tuning of the
MC parameters and jet-energy calibration using hadronic $W$ decays in the
same top events. It is not unlikely that these procedures could lead to an
increased consistency between the \PythiaEightPtwo{} and \HerwigSevenPone{}
results. However, in view of what we have found in our study, it is difficult
to trust the theoretical errors currently given in the top quark mass
determination if alternative NLO+PS and shower generators combinations are
not considered.
\section*{Acknowledgments}
We would like to thank Peter Richardson for important suggestions on the
settings of \HerwigSevenPone{} and on our \POWHEG{}-\HerwigSevenPone{}
interfaces, and Stefan Prestel for help with the \PythiaEightPtwo{} code and
for suggestions on the implementation of our \POWHEG{}-\PythiaEightPtwo{}
interface. We also like to thank Simon Pl\"atzer and Stephen Webster for
help and suggestions on our \HerwigSevenPone{} input files, and for
illustrating us aspects of the \HerwigSevenPone{} code. Finally we thanks
Roberto Franceschini, Alexander Mitov and Stefano Pozzorini for suggestions
on the manuscript, and Mike Seymour, Peter Skands, T\"orbjorn Sj\"ostrand,
Michelangelo Mangano, Andreas Papaefstathiou, and Andrzej Siodmok for useful
discussions.
We acknowledge the CINECA and the Regione Lombardia award under the LISA
initiative 2016-2018, for the availability of high performance computing
resources and support.
C.O.~wishes to thank Maurizio Cremonesi for the help in running the code on
the CINECA resources, under the project LISA PWHG-RES. The research of
T.J.~was supported by the Swiss National Science Foundation~(SNF) under
contracts BSCGI0-157722 and CRSII2-160814.
|
\section{Introduction}
In this work we return to the research programme initiated in
\cite{Bastos1,Bastos2}. There we considered the Weyl-Wigner
formulation of quantum mechanics based on a canonical deformation
of the Heisenberg algebra. More specifically, if $z=(x,p) \in
\mathbb{R}^{2n}$ denotes a phase space coordinate of a particle,
then the standard classical Poisson bracket is replaced by the
following deformation
\begin{equation}
\left\{z_{\alpha}, z_{\beta} \right\} = \Omega_{\alpha, \beta}, \hspace{1 cm} 1 \le \alpha, \beta \le 2n,
\label{eqIntroduction1}
\end{equation}
where $(\Omega_{\alpha, \beta})_{1 \le \alpha, \beta \le 2n}$ are the entries of the matrix
\begin{equation}
{\bf \Omega}= \left(
\begin{array}{l l}
{\bf \Theta} & {\bf I}\\
-{\bf I} & {\bf \Upsilon}
\end{array}
\right).
\label{eqNCQM1}
\end{equation}
Here ${\bf I}$ is the $n \times n$ identity matrix and ${\bf
\Theta}=\left(\theta_{ij}\right)_{1\le i,j \le n}, {\bf \Upsilon}=
\left(\eta_{ij}\right)_{1\le i,j \le n}$ are real constant
skew-symmetric $n \times n$ matrices, which measure the extra
noncommutativity in the configuration and momentum spaces,
respectively. One assumes in general that this symplectic form
does not depart appreciably from the standard one \cite{Carroll}:
\begin{equation}
|\theta_{ij} \eta_{kl}| \ll 1,
\label{eqNCQM2}
\end{equation}
for all $i,j,k,l=1, \cdots, n$. Upon quantization, one obtains a deformation of the Heisenberg algebra.
Noncommutative deformations such as this one and others appear in
manifold contexts. They may emerge (i) in the various attempts to
quantize gravity, such as string theory \cite{Seiberg},
noncommutative geometry \cite{Borowiec,Connes1,Madore} and loop
quantum gravity \cite{Rovelli}; (ii) as a means to regularize
quantum field theories \cite{Douglas,Szabo}; (iii) as the result
of quantizing quantum systems with second class constraints
\cite{Henneaux,Nakamura1,Nakamura2}; (iv) in order to stabilize
unstable algebras \cite{Vilela}, (v) as the low-energy physics of
quasicrystals \cite{Monreal}, and quite simply (vi) as a model for
quantum systems under the influence of an external magnetic field
\cite{Delduc}. Regardless of the context and motivation, deforming
the Heisenberg algebra in the realm of non-relativistic quantum
mechanics \cite{Demetrian,Gamboa,Nair} has led to many interesting
and surprising results such as the thermodynamic stability of
black holes \cite{Bastos3}, regularization of black hole
singularities \cite{Bastos4,Bastos5,Nicolini}, modifications of
quantum cosmology scenarios \cite{Obregon,Malekolkalami},
violation of uncertainty principles \cite{Bastos7,Bolonek},
generation of entanglement due to noncommutativity \cite{Bastos6}.
These deformations have also been used in the context of the
Landau problem \cite{Duval,Horvathy} and the quantum Hall effect
\cite{Bellissard}.
We shall follow \cite{Bastos1,Bastos2} and address quantum
mechanics on non-standard symplectic spaces in the framework of
the Weyl-Wigner formulation. As explained in \cite{Bastos2} there
are various aspects of this formulation which make it (in certain
circumstances) more appealing than the ordinary formulation in
terms of operators acting on a Hilbert space. In the Weyl-Wigner
representation one does not have to choose between a position or a
momentum representation of a state, as the Wigner function
provides a joint distribution of both. It is no surprise that the
uncertainty principle and non-locality take their toll on the
Wigner function in the form of a lack of positivity
\cite{Ellinas,Hudson,Soto}. Nevertheless, Wigner functions have
positive marginals and permit an evaluation of expectation values
of observables with the suggestive formula
\begin{equation}
<\widehat{A}> = \int_{\mathbb{R}^{2n}} a(z) W \rho (z) dz,
\label{eqIntroduction2}
\end{equation}
where $\widehat{A}$ is a self-adjoint operator, $a$ is its Weyl
symbol and $W \rho$ is the Wigner function associated with some
density matrix $\rho$. Moreover, the Weyl-Wigner framework is the
only phase space formulation which is symplectically invariant
\cite{Dias5,Folland,Gosson1,Gosson2,Gosson3,Gosson4,Leray,Wong}.
More precisely, if $W \rho (z)$ is a Wigner function and ${\bf S}$
is some symplectic matrix, then $W \rho({\bf S}z)$ is again a
Wigner function. This makes it a privileged framework to study the
semiclassical limit of quantum mechanics
\cite{Littlejohn,Martinez,Zworski}. The symplectic flavour of this
approach manifests itself equally in the dynamics of a quantum
mechanical system. The time evolution is governed by a quantum
counterpart of the Poisson bracket - called Moyal bracket - which
is a deformation of the Poisson bracket
\cite{Bayen,Fedosov1,Fedosov2,Kontsevich,Moyal,Tosiek,Wilde}. From
this perspective one sometimes calls this phase space formulation
of quantum mechanics {\it deformation quantization}.
The Weyl-Wigner formulation is also particularly useful in the
context of quantum information with continuous variables
\cite{Braunstein,Simon,Werner}.
For the quantization of systems with non-standard symplectic
spaces, the advantage of the Weyl-Wigner formulation becomes more
emphatic. The reason is that the Hilbert space for quantum
mechanics on non-standard symplectic spaces is still $L^2
(\mathbb{R}^n)$. So in the operator formulation, there is no
distinction between states for different symplectic structures. On
the other hand, in deformation quantization we proved in
\cite{Bastos1,Bastos2} that the set of states (Wigner functions)
in noncommutative quantum mechanics differs from the set of
ordinary Wigner functions. We have used this fact as a criterion
for assessing when a transition from noncommutative to ordinary
quantum mechanics has taken place \cite{Dias6}.
In \cite{Bastos2} we considered the set $\mathcal{F}^C$ of
ordinary Wigner functions, the set $\mathcal{F}^{NC}$ of Wigner
functions for the noncommutative symplectic structure
(\ref{eqIntroduction1},\ref{eqNCQM1}), and the set $\mathcal{L}$
of positive (classical) probability densities. We proved that in
{\it dimension $n=2$} any pair of these sets have non-empty
intersections and none of them contains the other. The main
purpose of the present work is to generalize this result to
arbitrary dimension $n$ and to the sets $\mathcal{F}^{\omega_1},
\mathcal{F}^{\omega_2}$ of Wigner functions defined on the
symplectic vector spaces $(\mathbb{R}^{2n}; \omega_1),
(\mathbb{R}^{2n}; \omega_2)$ with symplectic forms $\omega_2 \ne
\pm \omega_1$. Notice that the sympectic forms are completely
arbitrary (not necessarily of the form (\ref{eqNCQM1})), albeit
constant.
As a byproduct we extend a number of definitions and results
concerning ordinary Wigner functions to Wigner functions on
arbitrary, nonstandard symplectic spaces. Moreover, as an
application of our results, we solve the so-called {\it
representability problem} in the case of linear coordinate
transformations. This terminology was coined in
\cite{Cohen,Loughlin} in the context of signal processing for pure
states. Here we use the following definition for an arbitrary
(pure or mixed) state. A given real and measurable phase-space
function $F(z)$ is said to be {\it representable} if there exists
a density matrix $\rho$ such that $F$ is the Wigner function on
some symplectic phase-space associated with $\rho$, i.e. $F=W
\rho$. Here we consider linear coordinate transformations of the
form
\begin{equation}
W \rho (z) \mapsto (\mathcal{U}_{{\bf M}} W \rho) (z) = | \det M| W \rho ({\bf M} z),
\label{eqIntroduction3}
\end{equation}
with ${\bf M} \in Gl(2n; \mathbb{R})$. We then address the problem of the representability of $\mathcal{U}_{{\bf M}} W \rho$.
Our results may also be potentially interesting for quantum
information of continuous variable systems. Indeed, the partial
transpose - a linear coordinate transformation which has as sole
effect the reversal of the momentum of one subsystem - is neither
symplectic nor anti-symplectic \cite{Simon,Werner}. Hence, it maps
a Wigner function on the standard symplectic vector space to a
Wigner function on a non-standard one. This transformation is used
as a criterion to assess whether a state is separable or
entangled.
Here is a summary of the main results of this work.
\vspace{0.3 cm} \noindent {\bf 1)} We define the notion of
Narcowich-Wigner spectrum adapted to an arbitrary symplectic form
(Definition \ref{DefinitionQuantumConditions2}) and prove its main
properties (Theorem \ref{TheoremQuantumConditions5}, Theorem
\ref{TheoremPropertiesNWSpectra}).
\vspace{0.3 cm} \noindent {\bf 2)} We introduce the symplectic
spectrum for an arbitrary symplectic form (Definition
\ref{DefinitionOmegaSymplecticSpectrum1}). We prove a
Williamson-type theorem for non-standard symplectic spaces
(Theorem \ref{TheoremOmegaWilliamsonTheorem1}) and show the
usefulness of symplectic spectra for generalized uncertainty
principles (Theorem \ref{TheoremOmegaRSUP1}). In Theorem
\ref{TheoremIdenticalSpectra1} we show that we can always find a
positive-definite matrix which has distinct symplectic spectra
relative to two different symplectic forms. This result is
fine-tuned for the smallest symplectic eigenvalue in Theorem
\ref{TheoremUnboudedQuotient1}.
\vspace{0.3 cm} \noindent {\bf 3)} We solve the {\it
representability problem} for linear coordinate transformations in
Theorems \ref{TheoremSymplecticCovariance} and
\ref{TheoremRepresentabilityProblem}. In Theorem
\ref{TheoremSymplecticCovariance} we prove that $\mathcal{U}_{{\bf
M}} W \rho$ is representable for any Wigner function $ W \rho$ if
and only if ${\bf M}$ is symplectic or antisymplectic. Another new
result is Theorem \ref{TheoremRepresentabilityProblem}, which
states that if ${\bf M}$ is neither symplectic nor antisymplectic,
then there exists a Wigner function $W \rho$ such that
$\mathcal{U}_{{\bf M}} W \rho$ is again a Wigner function and $W
\rho^{\prime}$ for which $\mathcal{U}_{{\bf M}} W \rho^{\prime}$
is not a Wigner function.
\vspace{0.3 cm} \noindent {\bf 4)} Theorem \ref{TheoremSets1} is
one of the main results of this work and it shows that in the
phase-space formulation different quantizations (i.e. on different
symplectic spaces) yield different sets of states. However, there
are always positive distributions and states that are common to
different quantizations. These results are valid in arbitrary
dimension and for arbitrary flat sympletic vector spaces.
\section*{Notation}
We denote by $\mathbb{P}_{k \times k} (\mathbb{K})$ the convex
cone of real $(\mathbb{K}=\mathbb{R})$ (or complex
$(\mathbb{K}=\mathbb{C})$) symmetric (hermitean) positive
(semi-definite) $k \times k$ matrices, and we write $\mathbb{P}_{k
\times k}^{\times} (\mathbb{K})$ if they are positive-definite.
The set of real skew-symmetric $k \times k$ matrices is $A (k;
\mathbb{R})$. The space of Schwartz test functions is denoted by
$\mathcal{S}(\mathbb{R}^n)$ and its dual
$\mathcal{S}^{\prime}(\mathbb{R}^n)$ are the tempered
distributions. Latin indices $i,j,k, \cdots$, take values in the
set $\left\{1, 2, \cdots, n \right\}$, whereas Greek indices
$\alpha , \beta, \gamma, \cdots $ are phase space indices which
take values in $\left\{1,2, \cdots, 2n \right\}$. The inner
product in $L^2(\mathbb{R}^n)$ is given by
\begin{equation}
<\psi| \phi> = \int_{\mathbb{R}^n} \overline{\psi(x)} \phi (x) dx.
\label{eqNotation1}
\end{equation}
We may at times write $ <\psi| \phi>_{L^2(\mathbb{R}^n)}$ and
$||\psi||_{L^2(\mathbb{R}^n)}$ for the corresponding norm if we
want to emphasize that the functions are defined on the
$n$-dimensional Euclidean space $\mathbb{R}^n$.
The Fourier-Plancherel transform for a function $f \in L^1(\mathbb{R}^n) \cap L^2(\mathbb{R}^n)$ is defined by
\begin{equation}
(\mathcal{F}f) (\omega)= \int_{\mathbb{R}^n} f(x) e^{-i x \cdot \omega} dx
\label{eqNotation2}
\end{equation}
and its inverse is
\begin{equation}
(\mathcal{F}^{-1}f) (x)= \frac{1}{(2 \pi)^n} \int_{\mathbb{R}^n} f(\omega) e^{i x \cdot \omega} d \omega.
\label{eqNotation3}
\end{equation}
In this work, we choose units such that Planck's constant is $h= 2 \pi \hbar= 2 \pi$.
\section{Symplectic vector spaces}
Let $V$ be some real $2n$-dimensional vector space. A symplectic form on $V$ is a skew-symmetric non-degenerate bilinear form, that is a map $\omega:V \times V \to \mathbb{R}$ such that
\begin{equation}
\omega (\alpha x + \beta y,z) =\alpha \omega (x,z) + \beta \omega(y,z), \hspace{1 cm} \omega (x,y) = - \omega(y,x),
\label{eqsymplecticform1}
\end{equation}
for all $x,y,z \in V$ and $\alpha, \beta \in \mathbb{R}$, and
\begin{equation}
\omega (x,y) = 0 , \mbox{ for all } y \in V
\label{eqsymplecticform2}
\end{equation}
if and only if $x=0$.
The symplectic group $Sp (V; \omega)$ is the group of linear automorphisms $\phi: V \to V$ such that
\begin{equation}
\omega \left(\phi (z) , \phi (z^{\prime}) \right) = \omega (z, z^{\prime}),
\label{eqsymplecticform2.1}
\end{equation}
for all $z,z^{\prime} \in V$. An automorphism $\phi$ is said to be anti-symplectic if
\begin{equation}
\omega \left(\phi (z) , \phi (z^{\prime}) \right) = - \omega (z, z^{\prime}),
\label{eqsymplecticform2.2}
\end{equation}
for all $z,z^{\prime} \in V$.
The archetypal symplectic vector space is $V=\mathbb{R}^{2n}$ endowed with the so-called {\it standard symplectic form}. As usual we write $z= (x,p) \in V= \mathbb{R}^n \times (\mathbb{R}^n )^{\ast} \simeq \mathbb{R}^{2n}$, where $x \in \mathbb{R}^n $ is interpreted as the particle's position and $p \in (\mathbb{R}^n)^{\ast} $ as the particle's momentum belonging to the cotangent bundle. The standard symplectic form reads
\begin{equation}
\sigma (z,z^{\prime}) = z \cdot {\bf J^{-1}} z^{\prime} = p \cdot x^{\prime} - x \cdot p^{\prime},
\label{eqsymplecticform3}
\end{equation}
where $z=(x,p)$, $z^{\prime}=(x^{\prime},p^{\prime})$ and ${\bf J}= - {\bf J^{-1}}=-{\bf J^T}$ is the $2n \times 2n$ standard symplectic matrix
\begin{equation}
{\bf J}= \left(
\begin{array}{c c}
{\bf 0} & {\bf I}\\
- {\bf I} & {\bf 0}
\end{array}
\right).
\label{eqsymplecticform4}
\end{equation}
The symplectic group $Sp(\mathbb{R}^{2n}; \sigma) \equiv Sp(n; \sigma)$ is then the set of matrices ${\bf P}$ such that
\begin{equation}
{\bf P} {\bf J} {\bf P^T} = {\bf J}.
\label{eqsymplecticform4.1}
\end{equation}
We remark that if ${\bf P} \in Sp (n; \sigma)$, then also ${\bf P^{-1}}, {\bf P^T} \in Sp (n; \sigma)$.
A matrix ${\bf A}$ is $\sigma$-anti-symplectic \cite{Dias5} if
\begin{equation}
{\bf A} {\bf J} {\bf A^T} = - {\bf J}.
\label{eqsymplecticform4.2}
\end{equation}
All $\sigma$-anti-symplectic matrices ${\bf A}$ can be written as
\begin{equation}
{\bf A} = {\bf R}{\bf P},
\label{eqsymplecticform4.3}
\end{equation}
where ${\bf P} \in Sp (n; \sigma)$ and
\begin{equation}
{\bf R} =\left(
\begin{array}{c c}
{\bf I} & {\bf 0}\\
{\bf 0} & - {\bf I}
\end{array}
\right).
\label{eqsymplecticform4.4}
\end{equation}
Thus a $\sigma$-anti-symplectic transformation is a $\sigma$-symplectic transformation followed by a reflection of the particle's momentum.
More generally, for any other symplectic form on $\mathbb{R}^{2n}$, there exists a real skew-symmetric matrix ${\bf \Omega} \in Gl(2n)$ such that
\begin{equation}
\omega (z,z^{\prime}) = z \cdot {\bf \Omega^{-1}} z^{\prime}, \hspace{1 cm} z,z^{\prime} \in \mathbb{R}^{2n}.
\label{eqsymplecticform5}
\end{equation}
A well known theorem in symplectic geometry \cite{Cannas,Gosson1}
states that all symplectic vector spaces of equal dimension are
symplectomorphic. In other words, if $(V, \omega)$ and
$(V^{\prime}, \omega^{\prime})$ have the same dimension, then
there exists a linear isomorphism $\phi: V \to V^{\prime}$ such
that
\begin{equation}
\phi^{\ast} \omega^{\prime} = \omega.
\label{eqsymplecticform6}
\end{equation}
Here $\phi^{\ast}$ denotes the pull-back of $\phi$. In particular, $(\mathbb{R}^{2n}, \sigma)$ is symplectomorphic with $(\mathbb{R}^{2n}, \omega)$ for any symplectic form (\ref{eqsymplecticform5}). So if $\phi:(\mathbb{R}^{2n}, \sigma) \to (\mathbb{R}^{2n}, \omega)$ is given by
\begin{equation}
\phi(z) = {\bf S} z,
\label{eqsymplecticform7}
\end{equation}
for some ${\bf S} \in Gl(2n)$, then we have from (\ref{eqsymplecticform6}) that
\begin{equation}
\phi^{\ast} \omega = \sigma \Leftrightarrow \omega({\bf S} z, {\bf S} z^{\prime}) = \sigma (z,z^{\prime}),
\label{eqsymplecticform8}
\end{equation}
for all $z, z^{\prime} \in \mathbb{R}^{2n}$. And thus
\begin{equation}
{\bf \Omega} = {\bf S} {\bf J} {\bf S^T}.
\label{eqsymplecticform9}
\end{equation}
The matrix ${\bf S}$ in (\ref{eqsymplecticform9}) is not unique. Indeed, we have
\begin{lemma}\label{LemmaDarbouxmap1}
Suppose that ${\bf S}$ is a solution of (\ref{eqsymplecticform9}). Then ${\bf S^{\prime}}$ is also a solution if and only if $ {\bf S^{\prime -1} S}, {\bf S^{-1} S^{\prime}} \in Sp(n; \sigma)$.
\end{lemma}
\begin{proof}
We have ${\bf \Omega} = {\bf S} {\bf J} {\bf S^T}={\bf S^{\prime}} {\bf J} {\bf S^{\prime T}}$ if and only if ${\bf S^{\prime -1} S}{\bf J} ({\bf S^{\prime -1} S})^T = {\bf S^{-1} S^{\prime}} {\bf J} ({\bf S^{-1} S^{\prime}})^T = {\bf J}$.
\end{proof}
We denote by $\mathcal{D} (n;\omega)$ the set of all real $2n \times 2n$ matrices satisfying (\ref{eqsymplecticform9}) and call them {\it Darboux matrices}. The corresponding symplectic automorphism (\ref{eqsymplecticform7}) is called a {\it Darboux map}.
A matrix ${\bf M}$ is $\omega$-symplectic if
\begin{equation}
{\bf M} {\bf \Omega} {\bf M^T}= {\bf \Omega},
\label{eqomegasymplectic1}
\end{equation}
and it is $\omega$-anti-symplectic if
\begin{equation}
{\bf M} {\bf \Omega} {\bf M^T}= - {\bf \Omega}.
\label{eqomegasymplectic2}
\end{equation}
Any $\omega$-symplectic (resp. $\omega$-anti-symplectic) matrix ${\bf M}$ is of the form
\begin{equation}
{\bf M}={\bf S}{\bf P} {\bf S^{-1}},
\label{eqomegasymplectic3}
\end{equation}
where ${\bf S} \in \mathcal{D}(n; \omega)$ and ${\bf P}$ is a $\sigma$-symplectic (resp. $\sigma$-anti-symplectic) matrix.
For future reference we consider the following Proposition.
\begin{proposition}\label{PropositionSymplecticInequality}
Let $\eta $ be an arbitrary symplectic form on $\mathbb{R}^{2n}$ such that
\begin{equation}
|\eta (z,z^{\prime})| \le |\sigma (z,z^{\prime})|,
\label{eqSymplecticInequality1}
\end{equation}
for all $z,z^{\prime} \in \mathbb{R}^{2n}$. Then, there exists a constant $0 < |a| \le 1$ such that
\begin{equation}
\eta = a \sigma.
\label{eqSymplecticInequality2}
\end{equation}
\end{proposition}
\begin{proof}
Let $\eta (z,z^{\prime}) = z^T {\bf \Sigma^{-1}} z^{\prime}$ for $z,z^{\prime} \in \mathbb{R}^{2n}$ and ${\bf \Sigma} \in Gl(2n; \mathbb{R})$. Let $\left\{\delta_{\alpha} \right\}_{1 \le \alpha \le 2n}$ denote the canonical basis of $\mathbb{R}^{2n}$. If we set $z=\delta_{\alpha}$ and $z^{\prime} = \delta_{\beta}$ in (\ref{eqSymplecticInequality1}), we obtain:
\begin{equation}
|\Sigma_{\alpha \beta}^{-1}| \le |J_{\alpha \beta}|,
\label{eqSymplecticInequality3}
\end{equation}
for all $1 \le \alpha, \beta \le 2n$. So $\Sigma_{\alpha \beta}^{-1}=0$, whenever $J_{\alpha \beta}=0$. And so ${\bf \Sigma^{-1}}$ must be of the form
\begin{equation}
{\bf \Sigma^{-1}} = \left(
\begin{array}{c c}
{\bf 0} & {\bf D}\\
- {\bf D} & {\bf 0}
\end{array}
\right),
\label{eqSymplecticInequality4}
\end{equation}
where ${\bf D} = diag (\mu_1, \cdots, \mu_n)$. Next set $z=\delta_i -\delta_j$ and $z^{\prime} = \delta_{n+i}+ \delta_{n+j}$, for $1 \le i < j \le n$. A simple calculation shows that
\begin{equation}
\sigma (z,z^{\prime})=0, \hspace{1 cm} \eta (z,z^{\prime})= \mu_i - \mu_j.
\label{eqSymplecticInequality5}
\end{equation}
But from (\ref{eqSymplecticInequality1}) it follows that $\mu_i = \mu_j$ for all $1 \le i < j \le n$. And so there exists $a \ne 0$ such that (\ref{eqSymplecticInequality2}) holds. But again from (\ref{eqSymplecticInequality1}), we must have $0 < |a| \le 1$.
\end{proof}
\section{Weyl Quantization}
\subsection{Weyl quantization on $(\mathbb{R}^{2n}; \sigma)$}
One of the basic ingredients for the quantization of a classical system defined on the standard symplectic vector space $(\mathbb{R}^{2n}; \sigma)$ is the Heisenberg-Weyl algebra:
\begin{equation}
\left[\widehat{Z}_{\alpha}, \widehat{Z}_{\beta} \right] = i J_{\alpha \beta} \widehat{I}, \hspace{0.5 cm} \left[\widehat{Z}_{\alpha}, \widehat{I} \right] = 0, \hspace{1 cm} \alpha, \beta =1, \cdots, 2n,
\label{eqHeisenbergWeylAlgebra1}
\end{equation}
where $\widehat{Z} = (\widehat{q}, \widehat{p})$ are the quantum-mechanical counterparts of the classical phase-space variables $z=(x,p)$ and $\widehat{I}$ is the identity operator. Their action on functions in their (densely defined) domains in $L^2 (\mathbb{R}^n)$ is given by
\begin{equation}
(\widehat{I} \psi) (x) = \psi (x), \hspace{1 cm}(\widehat{q}_i \psi) (x) = x_i \psi (x), \hspace{1 cm} (\widehat{p}_j \psi) (x) = -i \partial_j \psi (x),
\label{eqHeisenbergWeylAlgebra2}
\end{equation}
for $i,j=1, \cdots, n$. These operators are not bounded, which prevents the construction of a $C^*$-algebra of observables. A familiar way to circumvent this is to consider the alternative algebra of Heisenberg-Weyl displacement operators:
\begin{equation}
\widehat{D^{\sigma}} (\xi) = e^{ i \sigma (\xi,\widehat{Z})}.
\label{eqHeisenbergWeylOperator1}
\end{equation}
The action of $\widehat{D^{\sigma}} (\xi)$ on $\psi \in \mathcal{S} (\mathbb{R}^n)$ is given by the explicit formula
\begin{equation}
\widehat{D^{\sigma}} (\xi) \psi (x) = e^{ip_0 \cdot x - \frac{i}{2} p_0 \cdot x_0} \psi (x-x_0),
\label{eqHeisenbergWeylOperator2}
\end{equation}
for $\xi = (x_0,p_0)$. This extends to a unitary operator from $L^2 (\mathbb{R}^n)$ to $L^2 (\mathbb{R}^n)$ and to a continuous operator from $\mathcal{S}^{\prime} (\mathbb{R}^n)$ to $\mathcal{S}^{\prime} (\mathbb{R}^n)$.
Their commutation relations are easily established from (\ref{eqHeisenbergWeylOperator2}) or, heuristically, from (\ref{eqHeisenbergWeylAlgebra1},\ref{eqHeisenbergWeylOperator1}) using the Baker-Campbell-Hausdorff formula:
\begin{equation}
\widehat{D^{\sigma}} (\xi)\widehat{D^{\sigma}} (\zeta) = e^{\frac{i}{2} \sigma (\xi, \zeta)} \widehat{D^{\sigma}} (\xi + \zeta) = e^{i \sigma (\xi, \zeta)} \widehat{D^{\sigma}} (\zeta) \widehat{D^{\sigma}} (\xi),
\label{eqCommutationRelationsWeylOps1}
\end{equation}
for $\xi, \zeta \in \mathbb{R}^{2n}$. Operators of the form
\begin{equation}
\widehat{U}^{\sigma} (\xi, \tau) =e^{i \tau} \widehat{D^{\sigma}} (\xi)
\label{eqCommutationRelationsWeylOps2}
\end{equation}
constitute an irreducible unitary representation of the Heisenberg group $\mathbb{H} (n)$ of elements $(\xi, \tau) \in \mathbb{R}^{2n} \times \mathbb{R}$ with group multiplication
\begin{equation}
(\xi, \tau) \cdot ( \xi^{\prime}, \tau^{\prime}) = \left(\xi+ \xi^{\prime},\tau + \tau^{\prime} + \frac{1}{2} \sigma (\xi, \xi^{\prime}) \right).
\label{eqHeisenberggroup1}
\end{equation}
This is called the Schr\"odinger representation and, according to the Stone-von Neumann theorem \cite{Reed}, it is in fact the only unitary irreducible representation of $\mathbb{H} (n)$ on $L^2 (\mathbb{R}^n)$ up to rescalings of Planck's constant.
Given a tempered distribution $a^{\sigma} \in \mathcal{S}^{\prime}
(\mathbb{R}^{2n})$ \cite{Grubb,Hormander}, the Weyl operator with
symbol $a^{\sigma}$ is the Bochner integral
\cite{Dubin,Folland,Gosson1,Pool,Wong}
\begin{equation}
\widehat{A} := \left(\frac{1}{2 \pi} \right)^n \int_{\mathbb{R}^{2n}} \mathcal{F}_{\sigma} a^{\sigma} (z) \widehat{D^{\sigma}} (z) d z.
\label{eqWeylOperator1}
\end{equation}
Here $\mathcal{F}_{\sigma} a$ denotes the symplectic Fourier transform
\begin{equation}
\mathcal{F}_{\sigma} a (z ) := \left(\frac{1}{2 \pi} \right)^n \int_{\mathbb{R}^{2n}} e^{-i \sigma (z,z^{\prime})} a(z^{\prime}) dz^{\prime}.
\label{eqSymplecticFourierTransform1}
\end{equation}
We note that $\mathcal{F}_{\sigma}$ in (\ref{eqSymplecticFourierTransform1}) extends into an involutive automorphism $\mathcal{S}^{\prime} (\mathbb{R}^n)\rightarrow \mathcal{S}^{\prime} (\mathbb{R}^n)$.
The Weyl correspondence, written
$\widehat{A}\overset{\mathrm{Weyl}}{\longleftrightarrow}a^{\sigma}$
or $a^{\sigma} \overset{\mathrm{Weyl}}{\longleftrightarrow}
\widehat{A}$ between a symbol $a^{\sigma} \in \mathcal{S}^{\prime}
(\mathbb{R}^{2n})$ and the associated Weyl operator $\widehat{A}$
is a one-to-one map from $\mathcal{S}^{\prime} (\mathbb{R}^{2n})$
onto the space $\mathcal{L} \left(\mathcal{S}(\mathbb{R}^n),
\mathcal{S}^{\prime} (\mathbb{R}^n) \right)$ of bounded linear
operators $\mathcal{S} (\mathbb{R}^n) \to \mathcal{S}^{\prime}
(\mathbb{R}^n)$. This can be proven using Schwartz's kernel
Theorem and the fact that the Weyl symbol $a^{\sigma}$ of the
operator $\widehat{A}$ is related with the distributional kernel
$K_a$ of that operator by a partial Fourier transform
\cite{Bracken,Gosson1,Wong}:
\begin{equation}
a^{\sigma}(x,p)= \int_{\mathbb{R}^n} e^{-i y \cdot p} K_a \left(x + \frac{y}{2},x - \frac{y}{2} \right) dy,
\label{eqWeylOperator2}
\end{equation}
where $K_a \in \mathcal{S}^{\prime} (\mathbb{R}^{2n})$ and the Fourier transform is defined in the usual distributional sense.
Conversely, by the Fourier inversion theorem, the kernel $K_a$ can be expressed in terms of the symbol $a^{\sigma}$:
\begin{equation}
K_a (x,y) = \left(\frac{1}{2 \pi}\right)^n \int_{\mathbb{R}^n} e^{i p \cdot (x-y)} a^{\sigma} \left( \frac{x+y}{2} , p \right) dp.
\label{eqWeylOperator3}
\end{equation}
A Weyl operator is formally self-adjoint if and only if its symbol $a^{\sigma}$ is real.
Now, let $\psi \in L^2 (\mathbb{R}^n)$ and define the rank-one operator $\widehat{\rho_{\psi}}$
\begin{equation}
\widehat{\rho_{\psi}} \phi = <\psi| \phi> \psi,
\label{eqWeylOperator4}
\end{equation}
for all $\phi \in L^2 (\mathbb{R}^n)$. This is a Hilbert-Schmidt operator with kernel:
\begin{equation}
K_{\rho_{\psi}} (x,y) = \psi (x) \overline{\psi (y)} .
\label{eqWeylOperator4}
\end{equation}
The corresponding Weyl symbol is (cf.(\ref{eqWeylOperator2})):
\begin{equation}
\rho^{\sigma}_{\psi} (x,p) = \int_{\mathbb{R}^n} e^{-i y \cdot p} \psi \left(x+ \frac{y}{2} \right) \overline{\psi \left(x- \frac{y}{2} \right)} dy
\label{eqWeylOperator5}
\end{equation}
which is proportional to the celebrated Wigner function \cite{Wigner}:
\begin{equation}
W^{\sigma} \psi (x,p) := \left(\frac{1}{2 \pi } \right)^n \rho_{\psi}^{\sigma} (x,p).
\label{eqWeylOperator6}
\end{equation}
The multiplicative constant $(2 \pi)^{-n}$ is included to ensure the correct normalization
\begin{equation}
\int_{\mathbb{R}^{2n}} W^{\sigma} \psi (z) dz =1,
\label{eqWeylOperator7}
\end{equation}
whenever $||\psi||_{L^2}=1$.
The operator $\widehat{\rho_{\psi}}$ is the density matrix of the pure state $\psi$. Density matrices provide a unified formulation of pure and mixed states, playing a key role in many different contexts, notably semi-classical limit, decoherence, {\it etc} \cite{Giulini}. Let then $\left\{p_{\alpha} \right\}_{\alpha}$ be some probability distribution
\begin{equation}
0 \le p_{\alpha} \le 1, \hspace{1 cm} \sum_{\alpha} p_{\alpha} =1,
\label{eqWeylOperator8}
\end{equation}
and $\left\{\widehat{\rho}_{\alpha} \right\}_{\alpha}$ a collection of pure state density matrices associated with normalized states $\psi_{\alpha} \in L^2 (\mathbb{R}^n)$. The statistical mixture
\begin{equation}
\widehat{\rho} = \sum_{\alpha} p_{\alpha} \widehat{\rho}_{\alpha}
\label{eqWeylOperator9}
\end{equation}
is called the density matrix of a mixed state. One needs some caution in interpreting eq.(\ref{eqWeylOperator9}). The usual statement that the system is in state $\widehat{\rho}_{\alpha}$ with probability $ p_{\alpha}$ is not very rigorous, because there are (in general infinitely) many different decompositions of a density matrix of the form (\ref{eqWeylOperator9}).
An useful way to tell pure states from mixed states is to compute the so-called {\it purity} of the state. Moyal's identity \cite{Moyal} states that if $\psi , \phi \in L^2 (\mathbb{R}^n)$, then $W^{\sigma} \psi , W^{\sigma} \phi \in L^2 (\mathbb{R}^{2n})$ and that:
\begin{equation}
<W^{\sigma} \psi|W^{\sigma} \phi>_{L^2 (\mathbb{R}^{2n})} =\frac{1}{(2 \pi)^n} \left|<\psi| \phi> _{L^2 (\mathbb{R}^n)} \right|^2.
\label{eqMoyalidentity1}
\end{equation}
In particular:
\begin{equation}
(2 \pi)^n||W^{\sigma} \psi||_{L^2 (\mathbb{R}^{2n})}^2 = ||\psi||_{L^2 (\mathbb{R}^n)}^4 =1.
\label{eqMoyalidentity2}
\end{equation}
On the other hand, if we have a mixed state $W^{\sigma} \rho $, then:
\begin{equation}
(2 \pi)^n||W^{\sigma} \rho||_{L^2 (\mathbb{R}^{2n})}^2 <1.
\label{eqMoyalidentity3}
\end{equation}
The {\it purity} of a state $W^{\sigma} \rho$ is defined by
\begin{equation}
0 < \mathcal{P} \left[W^{\sigma} \rho \right] = (2 \pi)^n || W^{\sigma} \rho||_{L^2 (\mathbb{R}^{2n})}^2 \le 1,
\label{eqMoyalidentity3}
\end{equation}
and we have
\begin{equation}
\begin{array}{l l}
\mathcal{P} \left[W^{\sigma} \rho \right] =1 & \mbox{if } W^{\sigma} \rho \mbox{ is a pure state}\\
& \\
\mathcal{P} \left[W^{\sigma} \rho \right] <1 & \mbox{if } W^{\sigma} \rho \mbox{ is a mixed state}
\end{array}
\label{eqMoyalidentity4}
\end{equation}
It is a well known fact from the theory of compact operators \cite{Reed} that an operator $\widehat{\rho}$ can be written as in (\ref{eqWeylOperator8},\ref{eqWeylOperator9}) if and only if it is a positive and trace-class operator with unit trace. If $\rho(x,y)$ is the corresponding kernel then the associated Wigner function is
\begin{equation}
\begin{array}{c}
W^{\sigma} \rho (x,p) = \left(\frac{1}{2 \pi}\right)^n \int_{\mathbb{R}^n} e^{-i y \cdot p} \rho \left(x+ \frac{y}{2} ,x- \frac{y}{2} \right) dy =\\
\\
= \sum_{\alpha} p_{\alpha} W^{\sigma} \psi_{\alpha} (x,p),
\end{array}
\label{eqWeylOperator10}
\end{equation}
with uniform convergence. If $\widehat{A}\overset{\mathrm{Weyl}}{\longleftrightarrow}a^{\sigma}$ is some self-adjoint operator such that $\widehat{A} \widehat{\rho}$ is trace-class then the expectation value of $\widehat{A}$ in the state $\widehat{\rho}$ can be evaluated according to the following remarkable formula.
\begin{equation}
<\widehat{A}>_{\widehat{\rho}} = Tr (\widehat{A}\widehat{\rho}) = \int_{\mathbb{R}^{2n}} a^{\sigma} (z) W^{\sigma} \rho (z) dz.
\label{eqWeylOperator11}
\end{equation}
Because of this formula and of the fact that Wigner functions have positive marginal distributions, one is tempted to interpret them as joint position and momentum probability densities. However, this is precluded by their lack of positivity as stated in Hudson's Theorem \cite{Ellinas,Hudson,Soto}.
\subsection{Weyl quantization on $(\mathbb{R}^{2n}; \omega)$}
Alternatively, one may choose to quantize the system on a non-standard symplectic vector space $(\mathbb{R}^{2n}; \omega)$ with symplectic form $\omega$ given by (\ref{eqsymplecticform5}) \cite{Bastos1,Bastos2,Chowdhuri1,Chowdhuri2,Dias1,Dias2}.
Upon quantization, the Heisenberg-Weyl algebra (\ref{eqHeisenbergWeylAlgebra1}) is replaced by the algebra
\begin{equation}
\left[\widehat{\Xi}_{\alpha},\widehat{\Xi}_{\beta} \right] = i \Omega_{\alpha \beta} \widehat{I}, \hspace{0.5 cm} \left[\widehat{\Xi}_{\alpha},\widehat{I}\right] = 0, \hspace{1 cm} \alpha, \beta =1, \cdots, 2n.
\label{eqNonstandardWeylalgebra1}
\end{equation}
In the sequel, we shall assume that
\begin{equation}
\det {\bf \Omega} =1 , \hspace{1 cm} \det {\bf S}= \pm 1,
\label{eqNonstandardWeylalgebra1.1}
\end{equation}
for all ${\bf S} \in \mathcal{D}(n; \omega)$. This is because different values of the determinant of ${\Omega}$ simply amount to a rescaling of the variables $\left\{\widehat{\Xi}_{\alpha} \right\}_{1 \le \alpha \le 2n}$ (or, equivalently of a rescaling of Planck's constant).
Upon exponentiation of (\ref{eqNonstandardWeylalgebra1}), we obtain the non-standard Heisenberg-Weyl displacement operators:
\begin{equation}
\widehat{D}^{\omega} (\xi) = e^{i \omega (\xi, \widehat{\Xi})},
\label{eqNonstandardWeylalgebra2}
\end{equation}
satisfying the commutation relations
\begin{equation}
\widehat{D}^{\omega} (\xi ) \widehat{D}^{\omega} (\zeta) = e^{\frac{i}{2} \omega (\xi, \zeta)} \widehat{D}^{\omega} (\xi + \zeta) = e^{i \omega (\xi, \zeta)} \widehat{D}^{\omega} (\zeta) \widehat{D}^{\omega} (\xi) .
\label{eqNonstandardWeylalgebra3}
\end{equation}
As before, the elements of the form
\begin{equation}
\widehat{U}^{\omega} (\xi, \tau) = e^{i \tau} \widehat{D}^{\omega} (\xi ) ,
\label{eqNonstandardWeylalgebra4}
\end{equation}
constitute a unitary irreducible representation of the modified Heisenberg group $\mathbb{H}^{\omega} (n)$ with multiplication
\begin{equation}
(\xi, \tau) \cdot (\xi^{\prime}, \tau^{\prime}) = \left(\xi + \xi^{\prime}, \tau + \tau^{\prime} + \frac{1}{2} \omega (\xi, \xi^{\prime}) \right).
\label{eqNonstandardWeylalgebra5}
\end{equation}
Notice that if ${\bf S} \in \mathcal{D}(n; \omega)$ and $\widehat{Z}$ obey the Heisenberg-Weyl algebra (\ref{eqHeisenbergWeylAlgebra1}), then
\begin{equation}
\widehat{\Xi} = {\bf S}\widehat{Z}
\label{eqNonstandardWeylalgebra6}
\end{equation}
satisfy (\ref{eqNonstandardWeylalgebra1}). Consequently, we may write
\begin{equation}
\widehat{D}^{\omega} (\xi )= e^{i \omega(\xi, \widehat{\Xi})} = e^{i \omega(\xi, {\bf S} \widehat{Z})}=e^{i \sigma({\bf S^{-1}} \xi, \widehat{Z})} = \widehat{D}^{\sigma} ({\bf S^{-1}} \xi ).
\label{eqNonstandardWeylalgebra6}
\end{equation}
From this relation and eq.(\ref{eqCommutationRelationsWeylOps1}), one easily proves the commutation relations (\ref{eqNonstandardWeylalgebra3}).
Also, if we perform the transformation $\xi \mapsto z= {\bf S} \xi$ with ${\bf S} \in \mathcal{D}(n; \omega)$ in (\ref{eqWeylOperator1}) and substitute (\ref{eqNonstandardWeylalgebra6}), we obtain:
\begin{equation}
\widehat{A} = \left(\frac{1}{2 \pi} \right)^n \int_{\mathbb{R}^{2n}} \mathcal{F}_{\sigma} a^{\sigma} ({\bf S^{-1}} \xi) \widehat{D}^{\omega} (\xi) d \xi
\label{eqNonstandardWeylalgebra7}
\end{equation}
Next notice that
\begin{equation}
\begin{array}{c}
\mathcal{F}_{\sigma} a ({\bf S^{-1}} \xi) = \left(\frac{1}{2 \pi} \right)^n \int_{\mathbb{R}^{2n}}e^{-i \sigma ({\bf S^{-1}} \xi, z)} a (z) d z = \\
\\
=\left(\frac{1}{2 \pi} \right)^n \int_{\mathbb{R}^{2n}}e^{-i \omega ( \xi, {\bf S}z)} a (z) d z=\\
\\
= \left(\frac{1}{2 \pi} \right)^n \int_{\mathbb{R}^{2n}} e^{-i \omega ( \xi, \xi^{\prime})} a ({\bf S^{-1}} \xi^{\prime}) d \xi^{\prime} .
\end{array}
\label{eqNonstandardWeylalgebra8}
\end{equation}
This then suggests the following definition of the $\omega$-symplectic Fourier transform \cite{Dias2}:
\begin{equation}
\mathcal{F}_{\omega} a(\xi) := \frac{1}{(2 \pi)^n } \int_{\mathbb{R}^{2n}} e^{-i \omega ( \xi, \xi^{\prime})} a(\xi^{\prime}) d \xi^{\prime},
\label{eqNonstandardWeylalgebra9}
\end{equation}
for any $a \in \mathcal{S}^{\prime} (\mathbb{R}^{2n})$. With this choice of normalization $\mathcal{F}_{\omega}$ is involutive $\mathcal{F}_{\omega}\mathcal{F}_{\omega} a= a$.
Also we define the $\omega$-Weyl symbol of the operator $\widehat{A}$:
\begin{equation}
a^{\omega} (\xi) := a^{\sigma} ({\bf S^{-1}} \xi).
\label{eqNonstandardWeylalgebra10}
\end{equation}
Altogether, from (\ref{eqNonstandardWeylalgebra7}-\ref{eqNonstandardWeylalgebra10}), we obtain:
\begin{equation}
\widehat{A} = \frac{1}{(2 \pi)^n } \int_{\mathbb{R}^{2n}} \mathcal{F}_{\omega} a^{\omega} (\xi) \widehat{D}^{\omega} (\xi) d \xi,
\label{eqNonstandardWeylalgebra11}
\end{equation}
which defines the correspondence principle between an operator $\widehat{A}$ and its $\omega$-Weyl symbol $a^{\omega}$. From eq.(\ref{eqWeylOperator6},\ref{eqWeylOperator11}) it follows that
\begin{equation}
\begin{array}{c}
<\widehat{A} >_{\widehat{\rho} } = \frac{1}{(2 \pi)^n} \int_{\mathbb{R}^{2n}} a^{\sigma} (z) \rho^{\sigma} (z) dz=\\
\\
= \frac{1}{(2 \pi)^n } \int_{\mathbb{R}^{2n}} a^{\sigma} ({\bf S^{-1}} \xi) \rho^{\sigma} ({\bf S^{-1}} \xi ) d \xi=\\
\\
= \frac{1}{(2 \pi)^n } \int_{\mathbb{R}^{2n}} a^{\omega} (\xi) \rho^{\omega} ( \xi) d \xi =\\
\\
= \int_{\mathbb{R}^{2n}} a^{\omega} (\xi ) W^{\omega} \rho (\xi ) d \xi,
\end{array}
\label{eqNonstandardWeylalgebra12}
\end{equation}
where
\begin{equation}
W^{\omega} \rho (\xi) = W^{\sigma} \rho ({\bf S^{-1}} \xi),
\label{eqNonstandardWeylalgebra13}
\end{equation}
is the $\omega$-Wigner function associated with the density matrix $\widehat{\rho}$ \cite{Bastos1,Bastos2,Chowdhuri1,Chowdhuri2,Jiang}.
Regarding the purity of the states on the non-standard symplectic space $(\mathbb{R}^{2n}; \omega)$ we have:
\begin{equation}
\begin{array}{c}
\mathcal{P} \left[ W^{\omega} \rho \right] = (2 \pi)^n \int_{\mathbb{R}^{2n}} | W^{\omega} \rho(\xi)|^2 d \xi =(2 \pi)^n \int_{\mathbb{R}^{2n}} | W^{\sigma} \rho({\bf S^{-1}} \xi)|^2 d \xi = \\
\\
= (2 \pi)^n \int_{\mathbb{R}^{2n}} | W^{\sigma} \rho(z)|^2 d z=\mathcal{P} \left[ W^{\sigma} \rho \right]
\end{array}
\label{eqpurity1}
\end{equation}
And thus, we have as before
\begin{equation}
\begin{array}{l l}
\mathcal{P} \left[W^{\omega} \rho \right] =1 & \mbox{if } W^{\omega} \rho \mbox{ is a pure state}\\
& \\
\mathcal{P} \left[W^{\omega} \rho \right] <1 & \mbox{if } W^{\omega} \rho \mbox{ is a mixed state}
\end{array}
\label{eqpurity2}
\end{equation}
We conclude that the main structures of standard quantum mechanics extend trivially to the non-standard symplectic case. However, there are also significative differences concerning the properties of states. We will address these issues in the next section.
\begin{remark}\label{RemarkPPT}
At this point it is worth going back to the partial transpose transformation mentioned in the introduction in relation to the separability problem for systems with continuous variables.
Suppose that a system is constituted of two subsystems - Alice and Bob - with $n^A $ and $n^B$ degrees of freedom $(n=n^A+n^B)$. Their coordinates are $z^A=(x^A, p^A)$ and $z^B=(x^B, p^B)$. We write the collective coordinate $z=(x^A,x^B, p^A, p^B)$. The partial transpose transformation reverses Bob's momentum \cite{Simon,Werner}:
\begin{equation}
z \mapsto {\bf P} z= \left(
\begin{array}{c c c r}
{\bf I} & {\bf 0} & {\bf 0} & {\bf 0} \\
{\bf 0} & {\bf I} & {\bf 0} & {\bf 0} \\
{\bf 0} & {\bf 0} & {\bf I} & {\bf 0} \\
{\bf 0} & {\bf 0} & {\bf 0} & - {\bf I}
\end{array}
\right) ~ \left(
\begin{array}{c}
x^A\\
x^B\\
p^A\\
p^B
\end{array}
\right) =
\left(
\begin{array}{c}
x^A\\
x^B\\
p^A\\
- p^B
\end{array}
\right)
\label{eqPPT1}
\end{equation}
A straightforward computation reveals that the transformation ${\bf P}={\bf P^{-1}}$ is neither $\sigma$-symplectic nor $\sigma$-anti-symplectic. Consequently, in general, $W^{\sigma} \psi ({\bf P^{-1}}z)$ is not a $\sigma$-Wigner function. Rather, one can think of ${\bf P}$ as a Darboux map for the symplectic form
\begin{equation}
\omega (z, z^{\prime}) = z^T {\bf \Omega^{-1}} z^{\prime},
\label{eqPPT2}
\end{equation}
with
\begin{equation}
{\bf \Omega}= {\bf P} {\bf J} {\bf P^T}.
\label{eqPPT3}
\end{equation}
And hence, $W^{\omega} \psi (z)=W^{\sigma} \psi ({\bf P^{-1}}z)$ is a Wigner function on the non-standard symplectic space with symplectic form (\ref{eqPPT2},\ref{eqPPT3}).
\end{remark}
\section{States in phase-space}
In this section we shall always consider real, normalized and
square-integrable functions defined on the phase-space
$\mathbb{R}^{2n}$. A difficult problem consists in assessing
whether such a function $f$ qualifies as a $\omega$-Wigner
function for some symplectic form $\omega$. In other words, how
can one tell if $f$ is the $\omega$-Wigner function of some
density matrix $\widehat{\rho}$?
It is well known that the answer lies in the following positivity condition \cite{Dias3,Lions}. A real and normalized function $f \in L^2 (\mathbb{R}^{2n})$ is a $\sigma$-Wigner function if and only if
\begin{equation}
\int_{\mathbb{R}^{2n}} f(z) W^{\sigma} \psi (z)dz \ge 0,
\label{eqQuantumConditions1}
\end{equation}
for all $\psi \in L^2 (\mathbb{R}^{n})$. On the other hand, $f$ is a $\omega$-Wigner function if and only if there exist a $\sigma$-Wigner function $f^{\sigma}$ and a matrix ${\bf S} \in \mathcal{D}(n; \omega)$ such that $f(\xi) = f^{\sigma} ({\bf S}^{-1} \xi)$. That happens if and only if
\begin{equation}
\begin{array}{c}
\int_{\mathbb{R}^{2n}} f(\xi) W^{\omega} \psi (\xi)d \xi = \int_{\mathbb{R}^{2n}} f^{\sigma} ({\bf S}^{-1}\xi) W^{\sigma} \psi ({\bf S}^{-1} \xi)d \xi =\\
\\
= \int_{\mathbb{R}^{2n}} f^{\sigma}(z) W^{\sigma} \psi (z)dz \ge 0,
\end{array}
\label{eqQuantumConditions2}
\end{equation}
for all $\psi \in L^2 (\mathbb{R}^{n})$.
The positivity conditions (\ref{eqQuantumConditions1},\ref{eqQuantumConditions2}) are somewhat tautological, as they require the knowledge of the entire set of $\sigma$- or $\omega$-Wigner functions of pure states. There are an alternative set of conditions which are more in the spirit of measure theory and Bochner's Theorem. They were derived by Kastler \cite{Kastler}, and Loupias and Miracle-Sole \cite{Loupias} using the machinery of $C^{\ast}$-algebras and were later synthesized by Narcowich \cite{Narcowich1} in the framework of the so-called Narcowich-Wigner (NW) spectrum. This proved to be a very powerful tool to address convolutions of functions defined on the phase-space as we shall shortly see.
Here is a brief summary of the remainder of this section. In the next subsection, we define the NW spectrum of a function for a given symplectic structure and show how it provides a condition of positivity alternative to (\ref{eqQuantumConditions1}) or (\ref{eqQuantumConditions2}). In Theorem \ref{TheoremQuantumConditions5}, we show how the NW spectrum behaves under the convolution. In Theorem \ref{TheoremPropertiesNWSpectra} we gather the main properties of NW spectra. In subsection 4.2 we consider the uncertainty principle in the Robertson-Schr\"odinger form and the associated concept of symplectic spectrum for a given symplectic form $\omega$. The main results of this subsection are the Williamson Theorem for a non-standard symplectic form (Theorem \ref{TheoremOmegaWilliamsonTheorem1}) and Theorem \ref{TheoremUnboudedQuotient1} which shows that one can find positive-definite matrices with lowest symplectic eigenvalues with respect to two different sympletic strucutes as far apart as one wishes. All these results culminate in Theorem \ref{TheoremSets1} of section 4.3, where we show how the sets of positive (classical) measures and the sets of Wigner functions for two distinct symplectic structures relate to each other. Finally, as an application, we prove in Theorem \ref{TheoremRepresentabilityProblem} how a Wigner function behaves under an arbitrary linear coordinate transformation.
\subsection{Narcowich-Wigner spectra}
\begin{definition}\label{DefinitionQuantumConditions1}
Let ${\bf \Sigma} \in A(2n; \mathbb{R})$ and $\alpha \in \mathbb{R}$. A complex-valued, continuous function $f$ on $\mathbb{R}^{2n}$ that has the property that the matrix $\left[ \left[M_{jk} \right] \right]$ with entries
\begin{equation}
M_{jk} =f(a_j-a_k) e^{\frac{i \alpha}{2} a_k \cdot {\bf \Sigma} a_j}
\label{eqQuantumConditions3}
\end{equation}
is a positive matrix on $\mathbb{C}^N$ for each $N \in \mathbb{N}$ and all $a_1, a_2 , \cdots, a_N \in \mathbb{R}^{2n}$ is called a function of the $(\alpha, {\bf \Sigma})$-positive type. If $\alpha=0$, then we shall simply say that $f$ is of the $0$-positive type.
\end{definition}
A function may be of the $(\alpha, {\bf \Sigma})$-positive type for various values of $\alpha$ and these can be assembled in a set.
\begin{definition}\label{DefinitionQuantumConditions2}
The ${\bf \Sigma}$-Narcowich-Wigner spectrum of a complex-valued, continuous function $f$ on $\mathbb{R}^{2n}$ is the set:
\begin{equation}
\mathcal{W}^{{\bf \Sigma}} (f) := \left\{ \alpha \in \mathbb{R}: ~ f \mbox{ is of the $(\alpha, {\bf \Sigma})$-positive type} \right\}.
\label{eqQuantumConditions4}
\end{equation}
More generally the Narcowich-Wigner spectrum of $f$ is the set:
\end{definition}
\begin{equation}
\mathcal{W} (f) := \left\{ (\alpha ,{\bf \Sigma}) \in \mathbb{R} \times A (2n; \mathbb{R}) : ~ f \mbox{ is of the $(\alpha, {\bf \Sigma})$-positive type} \right\}.
\label{eqQuantumConditions4.1}
\end{equation}
For each $\alpha \in \mathbb{R}$ and ${\bf \Sigma} \in A(2n; \mathbb{R})$ the set of all functions of $(\alpha, {\bf \Sigma})$-positive type is a convex cone.
Before we proceed let us briefly recall Bochner's Theorem.
\begin{theorem}\label{TheoremQuantumConditions3}
{\bf (Bochner)} The set of Fourier transforms of finite, positive measures on $\mathbb{R}^k$ $(k \in \mathbb{N})$ is exactly the cone of functions of $0$-positive type.
\end{theorem}
The KLM (Kastler, Loupias, Miracle-Sole) conditions are a twisted generalization of Bochner's theorem.
\begin{theorem}\label{TheoremQuantumConditions4}
{\bf (Kastler, Loupias, Miracle-Sole)} Let $f$ be a real function on the phase-space $\mathbb{R}^{2n}$. Then $f$ is a $\sigma$-Wigner function if and only if its Fourier transform $\mathcal{F}(f)$ satisfies the $\sigma$-KLM conditions:
\vspace{0.3 cm}
\noindent
(i) $\mathcal{F}(f) (0)=1$;
\vspace{0.3 cm}
\noindent
(ii) $\mathcal{F}(f)$ is of the $(1,{\bf J})$-positive type.
\end{theorem}
The first condition ensures the normalization of $f$, while the second one is equivalent to the positivity condition (\ref{eqQuantumConditions1}). In terms of the Narcowich-Wigner spectrum, condition (ii) means
\begin{equation}
1 \in \mathcal{W}^{{\bf J}} (\mathcal{F}(f)).
\label{eqQuantumConditions5}
\end{equation}
On the other hand, if $f$ is everywhere non-negative, then according to Bochner's theorem
\begin{equation}
0 \in \mathcal{W}^{{\bf \Sigma}} (\mathcal{F}(f)),
\label{eqQuantumConditions6}
\end{equation}
for all ${\bf \Sigma} \in A(2n; \mathbb{R})$.
We now generalize the KLM conditions (Theorem \ref{TheoremQuantumConditions4}) for an arbitrary symplectic form $\omega$.
\begin{theorem}\label{TheoremQuantumConditions5}
Let $f$ be a real function on the phase-space $\mathbb{R}^{2n}$. Then $f$ is a $\omega$-Wigner function if and only if its Fourier transform $\mathcal{F}(f)$ satisfies the $\omega$-KLM conditions:
\vspace{0.3 cm}
\noindent
(i) $\mathcal{F}(f)(0)=1$;
\vspace{0.3 cm}
\noindent
(ii) $\mathcal{F}(f)$ is of the $(1,{\bf \Omega})$-positive type.
\end{theorem}
\begin{proof}
As usual $f$ is a $\omega$-Wigner function if and only if there exists a $\sigma$-Wigner function $f^{\sigma}$ and a matrix ${\bf S} \in \mathcal{D}(n; \omega)$ such that $f(z)=f^{\sigma}({\bf S}^{-1}z)$. After a simple calculation, we conclude that $\mathcal{F}(f)(a) =\mathcal{F}(f^{\sigma})({\bf S}^T a)$. Consequently:
\begin{equation}
\begin{array}{c}
\mathcal{F}(f)(a_j-a_k) e^{\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j} = \mathcal{F}(f^{\sigma}) ({\bf S}^T(a_j-a_k)) e^{\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j}=\\
\\
=\mathcal{F}(f^{\sigma}) (b_j-b_k) e^{\frac{i \alpha}{2} b_k \cdot {\bf S}^{-1} {\bf \Omega} ({\bf S}^T)^{-1} b_j} =\mathcal{F}(f^{\sigma}) (b_j-b_k) e^{\frac{i \alpha}{2} b_k \cdot {\bf J} b_j}
\end{array}
\label{eqQuantumConditions7}
\end{equation}
with $b_j={\bf S}^T a_j$ for $j=1, \cdots, N$. We conclude that $\mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f)) = \mathcal{W}^{{\bf J}} (\mathcal{F}(f^{\sigma}))$ and the result follows.
\end{proof}
\begin{remark}\label{Remark5.1}
It follows from the proof that $(\alpha , {\bf \Omega}) \in \mathcal{W} (\mathcal{F}(f))$ if and only if $(\alpha , {\bf J}) \in \mathcal{W} (\mathcal{F}(f^{\sigma}))$.
\end{remark}
Before we proceed, let us recapitulate the concept of Hamard-Schur (HS) product. Let ${\bf A}=\left[ \left[ A_{ij} \right] \right], {\bf B} =\left[ \left[ B_{ij} \right] \right] \in M (k \times n; \mathbb{C})$. Then the HS product is the map $\diamondsuit: M (k \times n; \mathbb{C}) \times M (k \times n; \mathbb{C}) \to M (k \times n; \mathbb{C})$
\begin{equation}
{\bf A} \diamondsuit {\bf B} = \left[ \left[ A_{ij} B_{ij} \right] \right].
\label{eqQuantumConditions8}
\end{equation}
According to Schur's Theorem, the HS product has the property of being a closed operation in the convex cone $\mathbb{P}_{k \times k} (\mathbb{C})$ of positive $k \times k$ matrices, i.e.:
\begin{equation}
\diamondsuit: \mathbb{P}_{k \times k} (\mathbb{C}) \times \mathbb{P}_{k \times k} (\mathbb{C}) \to \mathbb{P}_{k \times k} (\mathbb{C}).
\label{eqQuantumConditions9}
\end{equation}
\begin{theorem}\label{TheoremQuantumConditions5}
Let $f,g \in L^1 (\mathbb{R}^{2n})$ be real-valued functions on
$\mathbb{R}^{2n}$, such that $\mathcal{F}(f),\mathcal{F}(g)$ are
of the $(\alpha, {\bf \Sigma})$ and of the $(\beta, {\bf
\Upsilon})$-positive type, respectively, and $\det(\alpha {\bf
\Sigma} + \beta {\bf \Upsilon}) \ne 0$. Moreover, define $\gamma
:= \left(\det(\alpha {\bf \Sigma} + \beta {\bf \Upsilon})
\right)^{\frac{1}{2n}}$ to be a real $(2n)$-th root of that
determinant (this is always possible since real antisymmetric
matrices have non-negative determinants; they are equal to the
square of a polynomial of its entries - the Pfaffian). Then the
Fourier transform of the convolution $f \star g$ is of the
$\left(\gamma , \frac{\alpha}{\gamma} {\bf \Sigma} +
\frac{\beta}{\gamma} {\bf \Upsilon}\right)$-positive type. For
later convenience we write:
\begin{equation}
(\alpha, {\bf \Sigma}) \oplus (\beta, {\bf \Upsilon}) = \left(\gamma , \frac{\alpha}{\gamma} {\bf \Sigma} + \frac{\beta}{\gamma} {\bf \Upsilon}\right).
\label{eqQuantumConditions9.1}
\end{equation}
\end{theorem}
\begin{proof}
Since $f,g \in L^1 (\mathbb{R}^{2n})$, the Fourier transform of the convolution amounts to the pointwise product $\mathcal{F}(f \star g) = \mathcal{F}(f) \cdot \mathcal{F}(g) $. Thus:
\begin{equation}
\begin{array}{c}
\mathcal{F}(f \star g) (a_j -a_k) e^{\frac{i \gamma}{2} \left( \frac{\alpha}{\gamma} a_k \cdot {\bf \Sigma} + \frac{\beta}{\gamma} {\bf \Upsilon}\right) a_j } =\\
\\
= \left(\mathcal{F}(f) (a_j -a_k) e^{\frac{i \alpha}{2} a_k \cdot {\bf \Sigma} a_j} \right) \left(\mathcal{F}(g) (a_j -a_k) e^{\frac{i \beta}{2} a_k \cdot {\bf \Upsilon} a_j} \right).
\end{array}
\label{eqQuantumConditions10}
\end{equation}
The last expression is the $(jk)$-th component of the HS product of two positive matrices and the result follows.
\end{proof}
Henceforth, we shall use the notation $\mathcal{G}_{{\bf A}} $ to represent the phase-space Gaussian with covariance matrix ${\bf A}$:
\begin{equation}
\mathcal{G}_{{\bf A}} (z) = \frac{1}{(2 \pi)^n \sqrt{\det {\bf A}}} \exp \left(- \frac{1}{2} z \cdot {\bf A^{-1}} z \right).
\label{eqGaussian1}
\end{equation}
We are tacitly assuming that the expectation values $<\widehat{Z}_i>$ are all equal to zero, something which can easily be achieved by a phase-space translation.
For later convenience we assemble the following properties of Narcowich-Wigner spectra (see \cite{Brocker,Dias4,Narcowich1} for the standard symplectic case):
\begin{theorem}\label{TheoremPropertiesNWSpectra}
Let $\Omega \in A (2n; \mathbb{R})$, $\mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$ denote ${\Omega}$-Narcowich-Wigner spectrum of $\mathcal{F}(f)$ for some function $f \in L^2 (\mathbb{R}^{2n})$ with continuous Fourier transform. Then the following properties hold.
\begin{enumerate}
\item If $\mathcal{F}(f)$ is of $(\alpha,{\bf \Omega})$-positive type for some $\alpha \in \mathbb{R}$ and some ${\bf \Omega}\in A(2n; \mathbb{R})$, then $f$ is a real function.
\item $ \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f^{\vee})) =\mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$, where $f^{\vee } (z)=f(-z)$.
\item $\alpha \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$ if and only if $-\alpha \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$.
\item Let $f_{\lambda}(z)= | \lambda|^{2n} f(\lambda z)$ for $\lambda \in \mathbb{R} \backslash \left\{0 \right\}$. Then $\mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f_{\lambda})) = \frac{1}{\lambda^2} \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$.
\item Let $\det {\bf \Omega}=1$ and $\omega (z,z^{\prime}) = z \cdot {\bf \Omega^{-1}} z^{\prime}$ is the associated symplectic form. For ${\bf M} \in Sl(2n; \mathbb{R})$ define $f_{{\bf M}} (z)=f({\bf M}z)$. Then if ${\bf M}$ is $\omega$-symplectic or $\omega$-anti symplectic, we have $\mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f_{{\bf M}}))=\mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$.
\item Let $\psi\in L^2 (\mathbb{R}^n)$ be some pure state and $\omega(z,z^{\prime})=z^{\prime}\cdot {\bf \Omega^{-1}} z$ an arbitrary symplectic form. If $\psi$ is a Gaussian function, then $ \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(W^{\omega} \psi)) = \left[-1,1 \right]$. Otherwise $ \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(W^{\omega} \psi)) = \left\{-1,1 \right\}$.
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item Suppose that $(\alpha,{\bf \Omega}) \in \mathcal{W}(\mathcal{F}(f))$. Then the matrices with entries $M_{jk}=\mathcal{F}(f)(a_j-a_k) e^{\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j}$ are positive in $\mathbb{C}^N$. For $N=2$, we have that
$$
\left(\begin{array}{c c}
\mathcal{F}(f)(0) & \mathcal{F}(f) (a_1-a_2)e^{\frac{i \alpha}{2} a_2 \cdot {\bf \Omega} a_1}\\
& \\
\mathcal{F}(f) (a_2-a_1)e^{\frac{i \alpha}{2} a_1 \cdot {\bf \Omega} a_2} & \mathcal{F}(f)(0)
\end{array} \right)
$$
is positive. In particular it has to be self-adjoint, and setting $\omega:=a_1-a_2$, we obtain:
$$
\overline{\mathcal{F}(f) (\omega)}= \mathcal{F}(f) (-\omega)
$$
for all $\omega \in \mathbb{R}^{2n}$, where we used the continuity of $\mathcal{F}(f)$. This is equivalent to $f$ being a real function.
\item A simple calculation shows that $\mathcal{F}(f^{\vee})=\left[\mathcal{F}(f) \right]^{\vee}$. It follows that
$$
\mathcal{F}(f^{\vee}) (a_j-a_k)e^{\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j}=\mathcal{F}(f) (a_k-a_j)e^{\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j}=\mathcal{F}(f) (b_j-b_k)e^{\frac{i \alpha}{2} b_k \cdot {\bf \Omega} b_j}
$$
where $b_j=-a_j$, etc. If follows that $(\alpha , {\bf \Omega}) \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f^{\vee}))$ if and only if $(\alpha , {\bf \Omega}) \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$.
\item We have
$$
\begin{array}{c}
\mathcal{F}(f) (a_j-a_k)e^{-\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j} =\left[\mathcal{F}(f) \right]^{\vee} (a_k-a_j)e^{\frac{i \alpha}{2} a_j \cdot {\bf \Omega} a_k}=\\
\\
=\mathcal{F}(f^{\vee}) (a_k-a_j)e^{\frac{i \alpha}{2} a_j \cdot {\bf \Omega} a_k}
\end{array}
$$
and the result follows from 2.
\item Since $\mathcal{F} f_{\lambda} (z) = \mathcal{F} f \left(\frac{z}{\lambda} \right)$, we have that
$$
\mathcal{F} f_{\lambda} (a_j-a_k)e^{-\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j} = \mathcal{F} f \left(\frac{a_j}{\lambda} - \frac{a_k}{\lambda}\right) e^{-\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j}=\mathcal{F} f (b_j-b_k)e^{-\frac{i \alpha \lambda^2}{2} b_k \cdot {\bf \Omega} b_j},
$$
with $a_j= \lambda b_j$. We conclude that $\alpha \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f_{\lambda}))$ if and only if $\alpha \lambda^2 \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$.
\item We have that ${\bf M}{\bf \Omega} {\bf M^T} = \epsilon {\bf \Omega}$ with $\epsilon=+1$ (resp. $\epsilon=-1$) if ${\bf M}$ is $\omega$-symplectic (resp. $\omega$-anti symplectic).
On the other hand, we have $\mathcal{F} f_{{\bf M}} (z)= \mathcal{F} f (({\bf M^{-1}})^T z)$. Thus
$$
\begin{array}{c}
\mathcal{F} f_{{\bf M}} (a_j-a_k)e^{-\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j} = \mathcal{F} f \left(({\bf M^{-1}})^T a_j - ({\bf M^{-1}})^T a_k \right) e^{-\frac{i \alpha}{2} a_k \cdot {\bf \Omega} a_j}=\\
\\
=\mathcal{F} f (b_j-b_k)e^{-\frac{i \alpha }{2} b_k \cdot {\bf M} {\bf \Omega} {\bf M^T }b_j}=\mathcal{F} f (b_j-b_k)e^{-\frac{i \alpha \epsilon }{2} b_k \cdot {\bf \Omega} b_j},
\end{array}
$$
where $a_j={\bf M^T }b_j$. We conclude that $\alpha \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f_{{\bf M}})$ if and only if $\epsilon \alpha \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}(f))$. The rest is a consequence of 3.
\item We proved in \cite{Dias4} that $ \mathcal{W}^{{\bf \sigma}} (\mathcal{F}(W^{\sigma} \psi)) = \left[-1,1 \right]$ if $\psi$ is a Gaussian function, and $ \mathcal{W}^{{\bf \sigma}} (\mathcal{F}(W^{\sigma} \psi)) = \left\{-1,1 \right\}$ otherwise. The rest is a consequence of Remark \ref{Remark5.1}.
\end{enumerate}
\end{proof}
\subsection{Uncertainty principle and symplectic spectra}
Consider again the Gaussian measure (\ref{eqGaussian1}). It is well known that this is the Wigner function associated with some density matrix if and only if \cite{Narcowich1}:
\begin{equation}
{\bf A}+\frac{i}{2} {\bf J} \ge 0,
\label{eqGaussian2}
\end{equation}
that is ${\bf A}+\frac{i}{2} {\bf J} $ is a positive matrix in $\mathbb{C}^{2n}$. This matrix inequality is known in the literature as the Robertson-Schr\"odinger uncertainty principle (RSUP). It is stronger than the more familiar Heisenberg inequality (see below), as it also accounts for the position-momentum correlations. It also has the advantage of being invariant under linear symplectic transformations. Indeed, if $W^{\sigma} \rho (z) $ is the Wigner function of some density matrix $\widehat{\rho}$ and ${\bf P} \in Sp (n; \sigma)$, then $W^{\sigma} \rho ({\bf P} z) $ is the Wigner function $W^{\sigma} \rho_{{\bf P}}$ of some other density matrix $\widehat{\rho}_{{\bf P}}$, related to $\widehat{\rho}$ by a metaplectic transformation \cite{Leray,Shale,Weil}. Accordingly, if ${\bf A_{\rho}}$ and ${\bf A_{\rho_{P}}}$ are the corresponding covariance matrices, then the two are related by the following transformation:
\begin{equation}
{\bf A_{\rho}} = {\bf P} {\bf A_{\rho_{P}}} {\bf P^T},
\label{eqGaussian3.1}
\end{equation}
where we used the fact that $\det {\bf P}=1$ for ${\bf P} \in Sp(n; \sigma)$. Thus:
\begin{equation}
{\bf A_{\rho}} +\frac{i}{2} {\bf J} \ge 0 \Leftrightarrow {\bf A_{\rho_{P}}} +\frac{i}{2} {\bf P^{-1} } {\bf J} {\bf\left( P^{-1} \right)^T} \ge 0 \Leftrightarrow {\bf A_{\rho_{P}}} +\frac{i}{2} {\bf J} \ge 0,
\label{eqGaussian3}
\end{equation}
since ${\bf P^{-1} } \in Sp(n; \sigma)$.
To establish whether a state satisfies the RSUP (\ref{eqGaussian2}), one computes the eigenvalues of the matrix ${\bf A}+\frac{i}{2} {\bf J} $ and verifies if they are all greater or equal to zero. Alternatively, we may verify the RSUP by resorting to the symplectic eigenvalues of ${\bf A}$ and Williamson's Theorem.
First notice that if ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$, then the matrix ${\bf A} {\bf J^{-1}}$ has the same eigenvalues as ${\bf A^{1/2}} {\bf J^{-1}} {\bf A^{1/2}}$ \cite{Williamson}. So the eigenvalues of ${\bf A} {\bf J^{-1}}$ come in pairs $\pm i \lambda_{\sigma, j}({\bf A})$, where $\lambda_{\sigma, j}({\bf A}) >0$, $j=1, \cdots, n$.
\begin{definition}\label{DefinitionWilliamsontheorem1}
Let $\lambda_{\sigma, j}({\bf A})$, $j=1, \cdots, n$ denote the moduli of the eigenvalues of ${\bf A} {\bf J^{-1}}$ written as an increasing sequence
\begin{equation}
0 < \lambda_{\sigma, 1}({\bf A}) \le \lambda_{\sigma, 2}({\bf A}) \le \cdots \le \lambda_{\sigma, n}({\bf A}).
\label{eqGaussian3}
\end{equation}
They are called the $\sigma$-Williamson invariants or $\sigma$-symplectic eigenvalues of ${\bf A}$, and the $n$-tuple
\begin{equation}
Spec_{\sigma} ({\bf A}):= \left( \lambda_{\sigma, 1}({\bf A}) , \lambda_{\sigma, 2}({\bf A}) , \cdots , \lambda_{\sigma, n}({\bf A}) \right)
\label{eqGaussian4}
\end{equation}
is called the $\sigma$-symplectic spectrum of ${\bf A}$.
\end{definition}
The invariance of the $\sigma$-symplectic spectrum of ${\bf A}$ under the transformation ${\bf A} \mapsto {\bf A^{\prime}} ={\bf P} {\bf A} {\bf P^T}$ with ${\bf P} \in Sp(n; \sigma)$ is easily established. Indeed suppose that $\lambda \in Spec_{\sigma} ({\bf A})$. Then
\begin{equation}
\begin{array}{c}
0 = \det ({\bf A}{\bf J^{-1}} \pm i \lambda {\bf I}) = \det ({\bf P^{-1}}{\bf A^{\prime}}{\bf P^{-1 T}}{\bf J^{-1}} \pm i \lambda {\bf I}) =\\
\\
= \det ({\bf P^{-1}}{\bf A^{\prime}}{\bf J^{-1}}{\bf P} \pm i \lambda {\bf I}) = \det ({\bf A^{\prime}}{\bf J^{-1}} \pm i \lambda {\bf I}),
\end{array}
\label{eqGaussian5}
\end{equation}
where we used $ {\bf J} {\bf P^T} = {\bf P^{-1}} {\bf J}$. Thus $Spec_{\sigma} ({\bf A^{\prime}})=Spec_{\sigma} ({\bf A})$.
The following theorem is due to Williamson \cite{Williamson}:
\begin{theorem}\label{Williamsontheorem2}
{\bf (Williamson)} Let $A \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$. There exists a symplectic matrix ${\bf P} \in Sp(n; \sigma)$ such that
\begin{equation}
{\bf P} {\bf A} {\bf P^T} = diag \left(\lambda_{\sigma, 1}({\bf A}) , \cdots , \lambda_{\sigma, n}({\bf A}), \lambda_{\sigma, 1}({\bf A}) , \cdots , \lambda_{\sigma, n}({\bf A}) \right) ,
\label{eqGaussian6}
\end{equation}
where $\lambda_{\sigma, j}({\bf A})$, $j=1, \cdots, n$ are the $\sigma$-Williamson invariants of ${\bf A}$.
\end{theorem}
We are now in a position to reexpress the RSUP in terms of the $\sigma$-symplectic spectrum. Let ${\bf A}$ be the covariance matrix of some Wigner function $W^{\sigma} \rho (z)$ and let ${\bf D}$ denote the diagonal matrix appearing in (\ref{eqGaussian6}). We thus have
\begin{equation}
{\bf A} + \frac{i}{2} {\bf J} \ge 0 \Leftrightarrow {\bf D} + \frac{i}{2} {\bf P} {\bf J} {\bf P^T} \ge 0 \Leftrightarrow {\bf D} + \frac{i}{2} {\bf J} \ge 0 .
\label{eqGaussian7}
\end{equation}
The eigenvalues of the last matrix are easily computed. They are equal to $\lambda_{\sigma, j} ({\bf A}) \pm \frac{1}{2}$, $j=1, \cdots, n$. Thus the last inequality in (\ref{eqGaussian7}) holds if and only if \cite{Narcowich1}
\begin{equation}
\lambda_{\sigma, 1} ({\bf A}) \ge \frac{1}{2}.
\label{eqGaussian8}
\end{equation}
Let now $W^{\sigma} \rho (z)$ be some Wigner function with covariance matrix ${\bf A}$. Let ${\bf P} \in Sp(n; \sigma)$ be the symplectic matrix diagonalizing ${\bf A}$ as in (\ref{eqGaussian6}). As we argued before, $W^{\sigma} \rho ({\bf P^{-1}}z)$ corresponds to another Wigner function $W^{\sigma} \rho_{{\bf P}} (z)$ with covariance matrix ${\bf P} {\bf A} {\bf P^T} = {\bf D}$ (cf.(\ref{eqGaussian3.1},\ref{eqGaussian6})). This means that since the covariance matrix of $W^{\sigma} \rho_{{\bf P}} (z)$ is diagonal, we have
\begin{equation}
<\widehat{q}_j^2>_{\rho_{{\bf P}}} = <\widehat{p}_j^2>_{\rho_{{\bf P}}} = \lambda_{\sigma, j} ({\bf A}), \hspace{1 cm} j=1, \cdots, n
\label{eqGaussian9}
\end{equation}
Thus
\begin{equation}
<\widehat{q}_j^2>_{\rho_{{\bf P}}} <\widehat{p}_j^2>_{\rho_{{\bf P}}} = \left(\lambda_{\sigma, j} ({\bf A}) \right)^2 \ge \frac{1}{4} , \hspace{1 cm} j=1, \cdots, n
\label{eqGaussian10}
\end{equation}
which is the Heisenberg uncertainty principle. Consequently, the lowest value of $\frac{1}{2}$ for each $\sigma$-symplectic eigenvalue (cf.(\ref{eqGaussian8})), corresponds to a minimal uncertainty in some direction in the phase-space. Thus the extremal situation, where
\begin{equation}
\lambda_{\sigma, 1} ({\bf A}) =\lambda_{\sigma, 2} ({\bf A}) = \cdots = \lambda_{\sigma, n} ({\bf A})=\frac{1}{2},
\label{eqGaussian11}
\end{equation}
(which is equivalent to $2 {\bf A} \in Sp(n; \sigma)$) can only be achieved by a Gaussian pure state. This is known in the literature as Littlejohn's Theorem \cite{Bastiaans,Littlejohn}.
For future reference we state and prove the following proposition.
\begin{proposition}\label{PropostionEigenvector}
Let $e, f \in \mathbb{R}^{2n}$ be such that $\sigma (f,e) \ne 0$.
Then there exists a matrix ${\bf A} \in \mathbb{P}^{\times}_{2n
\times 2n} ( \mathbb{R})$ such that $u=e + if$ is an eigenvector
of $i {\bf A} {\bf J^{-1}}$ with eigenvalue equal to
$\lambda_{\sigma,1} ({\bf A})$ (if $\sigma (f,e)>0$) or $-
\lambda_{\sigma,1} ({\bf A})$ (if $\sigma (f,e)<0$), where
$\lambda_{\sigma,1} ({\bf A})$ is the smallest $\sigma$-Williamson
invariant of ${\bf A}$.
\end{proposition}
\begin{proof}
Suppose that $\lambda_1:= \sigma (f,e) >0$. Set $f^{\prime} =
\frac{f}{\sqrt{\lambda_1}}$ and $e^{\prime} =
\frac{e}{\sqrt{\lambda_1}}$. It follows that $\sigma (f^{\prime},
e^{\prime}) =1$. A well known theorem in symplectic geometry
\cite{Cannas,Gosson1} states that we can find a symplectic basis
$\left\{e_i^{\prime},f_j^{\prime} \right\}_{1\le i,j \le n}$ such
that $e_1^{\prime}= e^{\prime}$ and $f_1^{\prime}=f^{\prime}$.
Next choose an arbitrary set of positive numbers
$\lambda_2,\lambda_3, \cdots , \lambda_n$ such that $\lambda_1 \le
\lambda_2 \le \cdots \le \lambda_n$, and define
\begin{equation}
e_i = \sqrt{\lambda_i} e_i^{\prime}, \hspace{1 cm} f_j = \sqrt{\lambda_j} f_j^{\prime}, \hspace{1 cm} 1 \le i,j \le n.
\label{eqEigenvector1}
\end{equation}
Now define $v_i=e_i$, and $v_{i+n}=f_i$ (and similarly $v_i^{\prime}=e_i^{\prime}$, and $v_{i+n}^{\prime}=f_i^{\prime}$), and $\lambda_{i+n}= \lambda_i$ for $i=1, \cdots, n$. We can now rewrite (\ref{eqEigenvector1}) in a more compact manner:
\begin{equation}
v_{\alpha} = \sqrt{\lambda_{\alpha}} v_{\alpha}^{\prime}, \hspace{1 cm} 1 \le \alpha \le 2n .
\label{eqEigenvector2}
\end{equation}
Since $\left\{v_{\alpha}^{\prime} \right\}_{1 \le \alpha \le 2n} $ is a symplectic basis, we have:
\begin{equation}
\sigma (v_{\alpha}, v_{\beta})= \sqrt{\lambda_{\alpha} \lambda_{\beta}} \sigma (v_{\alpha}^{\prime}, v_{\beta}^{\prime}) = \sqrt{\lambda_{\alpha} \lambda_{\beta}} J_{\beta \alpha} = \lambda_{\alpha} J_{\beta \alpha},
\label{eqEigenvector3}
\end{equation}
where we used the fact that $\lambda_{i+n}= \lambda_i$, for $i=1,2, \cdots, n$. In general the basis $\left\{v_{\alpha} \right\}_{1 \le \alpha \le 2n} $ is not orthonormal. However, there exists a matrix ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} ( \mathbb{R})$ such that
\begin{equation}
v_{\alpha}^T {\bf A^{-1}} v_{\beta} = \delta_{\alpha \beta}, \hspace{1 cm} 1 \le \alpha, \beta \le 2n.
\label{eqEigenvector4}
\end{equation}
The matrix ${\bf A}$ defines an inner product in $\mathbb{C}^{2n}$:
\begin{equation}
<z, z^{\prime}>_{{\bf A}} := \overline{z} \cdot {\bf A^{-1}} z^{\prime}, \hspace{1 cm} z, z^{\prime} \in \mathbb{C}^{2n}.
\label{eqEigenvector5}
\end{equation}
We conclude the proof by showing that $(\lambda_1, \lambda_2, \cdots, \lambda_n)$ is the symplectic spectrum of ${\bf A}$ and that ${\bf A} {\bf J^{-1}}(e +i f) = -i \lambda_1 (e +i f)$.
Let $T : \mathbb{R}^{2n}\to \mathbb{R}^{2n}$ be the linear transformation $T(v)={\bf AJ^{-1}} v$, with matrix representation ${\bf AJ^{-1}}$ in the canonical basis. Then
\begin{equation}
<v_{\alpha}, T\left( v_{\beta} \right)>_{{\bf A}} = \sigma (v_{\alpha}, v_{\beta})= \lambda_{\alpha} J_{\beta \alpha}.
\label{eqEigenvector6}
\end{equation}
So the matrix representation ${\bf T}$ of $T$ with respect to the basis $\left\{v_{\alpha} \right\}_{1 \le \alpha \le 2n}$ is
\begin{equation}
{\bf T}= \left(
\begin{array}{c c}
{\bf 0} & - {\bf \Lambda}\\
{\bf \Lambda} & {\bf 0}
\end{array}
\right),
\label{eqEigenvector7}
\end{equation}
where ${\bf \Lambda} = diag (\lambda_1, \cdots, \lambda_n)$. The eigenvalues of ${\bf T}$ (and hence of $T$) are easily computed. They are: $\left\{\pm i \lambda_j\right\}_{j=1, \cdots,n}$ and the associated eigenvectors of ${\bf AJ^{-1}}$ are:
\begin{equation}
{\bf AJ^{-1}} (e_j \pm i f_j) = \mp i \lambda_j (e_j \pm i f_j), \hspace{1 cm} j=1, \cdots, n.
\label{eqEigenvector8}
\end{equation}
In particular ${\bf AJ^{-1}} (e+if) = - i \lambda_1 (e+if)$.
If $\sigma (f,e)<0$, then by following an identical procedure, we would obtain: ${\bf AJ^{-1}} (e+if) = i \lambda_1 (e+if)$.
\end{proof}
Now we turn to the non-standard symplectic space $(\mathbb{R}^{2n}; \omega)$. Suppose that $W^{\omega} \rho (z)$ is some $\omega$-Wigner function and let ${\bf S} \in \mathcal{D} (n; \omega)$ denote an arbitrary Darboux matrix. According to (\ref{eqNonstandardWeylalgebra13})
\begin{equation}
W^{\sigma} \rho (z) := W^{\omega} \rho ({\bf S} z)
\label{eqGaussian12}
\end{equation}
is a $\sigma$-Wigner function. Hence it has to comply with RSUP. Let ${\bf A}^{\sigma}$ and ${\bf A}^{\omega}$ denote the covariance matrices of $W^{\sigma} \rho $ and $W^{\omega} \rho $, respectively. We thus have
\begin{equation}
{\bf A}^{\omega} = {\bf S} {\bf A}^{\sigma} {\bf S^T}.
\label{eqGaussian13}
\end{equation}
Consequently,
\begin{equation}
{\bf A}^{\sigma} + \frac{i}{2} {\bf J} \ge 0 \Leftrightarrow {\bf A}^{\omega} + \frac{i}{2} {\bf S} {\bf J} {\bf S^T }\ge 0 \Leftrightarrow {\bf A}^{\omega} + \frac{i}{2} {\bf \Omega} \ge 0.
\label{eqGaussian14}
\end{equation}
The last inequality is valid irrespective of whatever Darboux matrix ${\bf S}$ we use. We shall call this inequality the $\omega$-RSUP.
Next let ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$, and consider the eigenvalues of the matrix ${\bf A} {\bf \Omega^{-1}}$. As in the standard case ${\bf A J^{-1}}$, these come in pairs $\pm i \lambda_{\omega, j} ({\bf A})$, $\lambda_{\omega, j} ({\bf A}) >0$, $j=1, \cdots, n$.
\begin{definition}\label{DefinitionOmegaSymplecticSpectrum1}
The moduli of the eigenvalues of ${\bf A} {\bf \Omega^{-1}}$ written as an increasing sequence $0 < \lambda_{\omega, 1} ({\bf A}) \le \lambda_{\omega, 2} ({\bf A}) \le \cdots \le \lambda_{\omega, n} ({\bf A})$ will be called the $\omega$-Williamson invariants or the $\omega$-symplectic eigenvalues of ${\bf A}$. Moreover, the $n$-tuple
\begin{equation}
Spec_{\omega} ({\bf A}) := \left( \lambda_{\omega, 1} ({\bf A}) , \lambda_{\omega, 2} ({\bf A}) , \cdots , \lambda_{\omega, n} ({\bf A}) \right)
\label{eqGaussian15}
\end{equation}
is called the $\omega$-symplectic spectrum of ${\bf A}$.
\end{definition}
It is easy to show that the $\omega$-symplectic spectrum of ${\bf A}$ is invariant under the $\omega$-symplectic transformation ${\bf A} \mapsto {\bf P} {\bf A} {\bf P^T}$, with ${\bf P} \in Sp (n; \omega)$. The proof follows the same steps as in (\ref{eqGaussian5}).
\begin{lemma}\label{LemmaRescalings1}
Let ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ and $\mu >0$. The matrix ${\bf B} = \mu {\bf A}$ has $\omega$-symplectic spectrum $Spec_{\omega} ({\bf B}) = \mu Spec_{\omega} ({\bf A})$.
\end{lemma}
\begin{proof}
This is a simple consequence of the following identity:
\begin{equation}
\det ({\bf B} {\bf \Omega}^{-1} - \lambda {\bf I} ) = \mu^{2n} \det \left({\bf A} {\bf \Omega}^{-1} - \frac{\lambda}{\mu} {\bf I} \right),
\label{eqLemmaRescalings1}
\end{equation}
and thus $\lambda \in Spec_{\omega} ({\bf B})$ if and only if $\frac{\lambda}{\mu} \in Spec_{\omega} ({\bf A})$.
\end{proof}
Next, we state and prove the counterpart of Williamson's Theorem on $(\mathbb{R}^{2n}; \omega)$.
\begin{theorem}\label{TheoremOmegaWilliamsonTheorem1}
Let $A \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$. Then there exists
a Darboux matrix ${\bf S} \in \mathcal{D}(n; \omega)$ such that
\begin{equation}
{\bf S^{-1}} {\bf A}{\bf \left(S^{-1}\right)^T} = diag \left(\lambda_{\omega,1} ({\bf A}) , \cdots, \lambda_{\omega,n} ({\bf A})
, \lambda_{\omega,1} ({\bf A}) , \cdots, \lambda_{\omega,n} ({\bf A})\right),
\label{eq12.4}
\end{equation}
where $\left\{\lambda_{\omega,j} ({\bf A})\right\}_{1 \le j \le n}$ are the $\omega$-Williamson invariants of ${\bf A}$.
\end{theorem}
\begin{proof} Let ${\bf S^{\prime}} \in \mathcal{D} (n; \omega)$ be a Darboux matrix. Then ${\bf S^{\prime -1}} {\bf A} ({\bf S^{\prime -1}})^T \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$. By the $\sigma$-Williamson Theorem, there exists ${\bf M} \in Sp(n; \sigma)$ such that
\begin{equation}
{\bf M^T} {\bf S^{\prime -1}} {\bf A} ({\bf S^{\prime -1}})^T {\bf M} = {\bf D},
\label{eq12.5}
\end{equation}
where
\begin{equation}
{\bf D}=diag \left(\lambda_{\sigma,1} ({\bf B}) , \cdots, \lambda_{\sigma,n} ({\bf B})
, \lambda_{\sigma,1} ({\bf B}) , \cdots, \lambda_{\sigma,n} ({\bf B})\right)
\label{eq12.6}
\end{equation}
and $\left(\lambda_{\sigma,1} ({\bf B}) , \cdots, \lambda_{\sigma,n} ({\bf B}) \right)$ is the $\sigma$-symplectic spectrum of ${\bf B} = {\bf S^{\prime -1}} {\bf A} ({\bf S^{\prime -1}})^T $.
Let ${\bf S}= {\bf S^{\prime}} ({\bf M^T})^{-1}$. By Lemma \ref{LemmaDarbouxmap1}, ${\bf S} \in \mathcal{D} (n; \omega)$ is a Darboux matrix. The result then follows if we prove that $\left(\lambda_{\sigma,1} ({\bf B}) , \cdots, \lambda_{\sigma,n} ({\bf B}) \right)$ is the $\omega$-symplectic spectrum of ${\bf A}$. We have that $\lambda \in Spec_{\omega}({\bf A})$ if and only if
\begin{equation}
\begin{array}{c}
0=\det({\bf A} {\bf \Omega^{-1}} \pm i \lambda {\bf I}) = \det ({\bf S} {\bf D} {\bf S^T} \pm i \lambda {\bf \Omega}) = \\
\\
=\det ({\bf D}\pm i \lambda {\bf S^{-1}} {\bf \Omega} ({\bf S^T})^{-1} ) = \det ({\bf D} \pm i \lambda {\bf J}),
\end{array}
\label{eq12.7}
\end{equation}
that is if and only if $\lambda \in Spec_{\sigma}({\bf B})$.
\end{proof}
There is a converse to the previous result.
\begin{theorem}\label{theorem2}
Let ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ and consider a set of positive numbers $0 < \lambda_1 \le \lambda_2 \le \cdots \le \lambda_n$. Then there exists a matrix ${\bf P} \in Gl (2n)$ such that
\begin{equation}
{\bf P^{-1}} {\bf A} {\bf \left(P^{-1}\right)^T} = diag (\lambda_1 , \cdots , \lambda_n, \lambda_1 , \cdots , \lambda_n).
\label{eqTheo15}
\end{equation}
Moreover, the set $Spec_{\delta} ({\bf A})= (\lambda_1 , \cdots , \lambda_n)$ is the $\delta$-symplectic spectrum with respect to the symplectic form
\begin{equation}
\delta (z, z^{\prime}) = z \cdot {\bf \Delta^{-1}} z,
\label{eqTheo16}
\end{equation}
with
\begin{equation}
{\bf \Delta} = {\bf P} {\bf J} {\bf P^T}.
\label{eqTheo17}
\end{equation}
\end{theorem}
\begin{proof}
Let $Spec_{\sigma} ({\bf A})=(\lambda_{\sigma,1} ({\bf A}) , \cdots , \lambda_{\sigma,n} ({\bf A}))$ denote the $\sigma$-symplectic spectrum of ${\bf A}$ with respect to the standard symplectic form $\sigma$. By Williamson's Theorem, there exists a $\sigma$-symplectic matrix ${\bf M} \in Sp(n; \sigma)$ such that
\begin{equation}
{\bf M^T} {\bf A} {\bf M} = diag (\lambda_{\sigma,1} ({\bf A}) , \cdots , \lambda_{\sigma,n} ({\bf A}),\lambda_{\sigma,1} ({\bf A}) , \cdots , \lambda_{\sigma,n} ({\bf A})).
\label{eqTheo18}
\end{equation}
Let ${\bf R}$ be the matrix:
\begin{equation}
{\bf R}= diag \left( \sqrt{\frac{\lambda_1}{\lambda_{\sigma,1} ({\bf A})}}, \cdots, \sqrt{\frac{\lambda_n}{\lambda_{\sigma,n} ({\bf A})}} , \sqrt{\frac{\lambda_1}{\lambda_{\sigma,1} ({\bf A})}}, \cdots, \sqrt{\frac{\lambda_n}{\lambda_{\sigma,n} ({\bf A})}} \right).
\label{eqTheo19}
\end{equation}
We then recover (\ref{eqTheo15}) with ${\bf P}=({\bf R} {\bf M^T})^{-1}$. If we define ${\bf \Delta}$ as in (\ref{eqTheo17}), then ${\bf P} \in \mathcal{D} (n; \delta)$ (cf. (\ref{eqTheo16})). The remaining statement is then a simple consequence of the previous theorem.
\end{proof}
Obviously, the matrix ${\bf P}$ in the previous theorem is not unique. Indeed, in the proof, we used the $\sigma$-symplectic spectrum $Spec_{\sigma} ({\bf A})$ with respect to the standard symplectic form $\sigma$. But we could equally have used the spectrum $Spec_{\omega} ({\bf A})$ with respect to any other symplectic form $\omega$. That would obviously lead to a different matrix ${\bf P}$.
We are now able to restate the $\omega$-RSUP (\ref{eqGaussian14}) in terms of the $\omega$-symplectic spectrum of the covariance matrix ${\bf A}$.
\begin{theorem}\label{TheoremOmegaRSUP1}
Let ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ and let $\lambda_{\omega,1} ({\bf A})$ denote its smallest $\omega$-Williamson invariant. Then it is the covariance matrix of a $\omega$-Wigner function if and only if
\begin{equation}
\lambda_{\omega ,1} ({\bf A}) \ge \frac{1}{2}.
\label{eqOmegaRSUP1}
\end{equation}
\end{theorem}
\begin{proof}
If ${\bf A}$ is the covariance matrix of a $\omega$-Wigner function, then it satisfies the $\omega$-RSUP (\ref{eqGaussian14}). Let ${\bf D} = diag (\lambda_{\omega ,1} ({\bf A}) , \cdots, \lambda_{\omega ,n} ({\bf A}),\lambda_{\omega , 1} ({\bf A}) , \cdots, \lambda_{ \omega , n} ({\bf A}) )$, where $\left\{\lambda_{ \omega , j} ({\bf A}) \right\}_{1 \le j \le n}$ are the $\omega$-Williamson invariants of ${\bf A}$. Thus, from (\ref{eq12.4}) we have
\begin{equation}
{\bf A} + \frac{i}{2} {\bf \Omega} \ge 0 \Leftrightarrow {\bf D}+ \frac{i}{2} {\bf S^{-1}}{\bf \Omega}{\bf S^{-1 T}} \ge 0 \Leftrightarrow {\bf D}+ \frac{i}{2} {\bf J} \ge 0.
\label{eqOmegaRSUP2}
\end{equation}
Again, the eigenvalues of ${\bf D}+ \frac{i}{2} {\bf J} $ are of the form $\lambda_{ \omega ,j} ({\bf A}) \pm \frac{1}{2}$, $j=1, \cdots, n$, and so the last inequality in (\ref{eqOmegaRSUP2}) is equivalent to (\ref{eqOmegaRSUP1}).
Conversely, if ${\bf A}$ verifies (\ref{eqOmegaRSUP1}), then it also satisfies the $\omega$-RSUP and the Gaussian measure with covariance matrix ${\bf A}$ is a Wigner function.
\end{proof}
It is worth mentioning that Narcowich-Wigner spectra of Gaussians are completely determined by the lowest Williamson invariant for every symplectic form. More specifically, we have:
\begin{theorem}\label{TheoremNWspectrumGaussians}
Let $\mathcal{G}_{{\bf A}}$ denote a Gaussian measure in phase space (\ref{eqGaussian1}) and $\omega(z,z^{\prime})= z^{\prime} \cdot {\bf \Omega^{-1}} z$ a symplectic form. Then the following statements are equivalent
\begin{enumerate}
\item $\alpha \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}\mathcal{G}_{{\bf A}})$,
\item ${\bf A} + \frac{i \alpha}{2 }{\bf \Omega} \ge 0$,
\item $\lambda_{\omega,1} ({\bf A}) \ge \frac{|\alpha|}{2}$.
\end{enumerate}
\end{theorem}
\begin{proof}
The equivalence of these statements was proven in \cite{Narcowich1} for the case $\omega=\sigma$. Let $\mathcal{G}_{{\bf A}}^{\sigma} (z)=\mathcal{G}_{{\bf A}} ({\bf S} z) = \mathcal{G}_{{\bf S^{-1} A \left(S^{-1} \right)^T}} (z)$ for ${\bf S} \in \mathcal{D}(n; \omega)$. From Remark \ref{Remark5.1} we have:
\begin{equation}
\begin{array}{c}
\alpha \in \mathcal{W}^{{\bf \Omega}} (\mathcal{F}\mathcal{G}_{{\bf A}}) \Leftrightarrow \alpha \in \mathcal{W}^{{\bf J}} (\mathcal{F}\mathcal{G}_{{\bf S^{-1} A \left(S^{-1} \right)^T}}) \Leftrightarrow \\
\\
{\bf S^{-1} A \left(S^{-1} \right)^T} + \frac{i \alpha}{2} {\bf J} \ge 0 \Leftrightarrow {A } + \frac{i \alpha}{2} {\bf \Omega} \ge 0
\end{array}
\label{eqOmegaRSUP2.1}
\end{equation}
This proves the equivalence of 1 and 2. The equivalence of 2 and 3 was shown in the proof of Theorem \ref{TheoremOmegaRSUP1}.
\end{proof}
\begin{remark}\label{Remark1}
Notice that, if $\sigma \ne \pm \omega$, then in general $Spec_{\sigma} ({\bf A}) \ne Spec_{\omega} ({\bf A})$. Here is an example. For $n=2$, let
\begin{equation}
{\bf A} = \left(
\begin{array}{c c c c}
\alpha & 0 & 0 &0 \\
0 & \alpha & 0 & 0\\
0 & 0 & \beta & 0\\
0 & 0 & 0 & \beta
\end{array}
\right),
\label{eqRemark1}
\end{equation}
and
\begin{equation}
{\bf \Omega} = \left(
\begin{array}{c c c c}
0 & 0 & \theta_1 &0 \\
0 & 0 & 0 & \theta_2\\
- \theta_1 & 0 & 0 & 0\\
0 & -\theta_2 & 0 & 0
\end{array}
\right),
\label{eqRemark2}
\end{equation}
with $\alpha, \beta >0$, and $\theta_1 =\frac{1}{\theta_2} \ge 1$. After a straightforward computation, we conclude that
\begin{equation}
Spec_{\sigma} ({\bf A}) = \left( \sqrt{\alpha \beta} , \sqrt{\alpha \beta} \right),
\label{eqRemark3}
\end{equation}
while
\begin{equation}
Spec_{\omega} ({\bf A}) = \left( \frac{\sqrt{\alpha \beta}}{\theta_1} , \frac{\sqrt{\alpha \beta}}{\theta_2} \right).
\label{eqRemark4}
\end{equation}
So if $\theta_1=\theta_2=1$, then the two spectra coincide, otherwise they differ.
\end{remark}
The example in the previous remark is just a particular instance of Theorem \ref{TheoremIdenticalSpectra1} (see below). But first we recall the following Lemma which was proven in \cite{Dias5}. We denote by $Sp^+ (n; \sigma) = Sp(n; \sigma) \cap \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ the set of real symmetric positive-definite symplectic $2n \times 2n$ matrices.
\begin{lemma}\label{LemmaSymplecticAntiSymplectic1}
Let ${\bf M} \in Gl(2n)$ and assume that ${\bf M}^T {\bf G} {\bf M} \in Sp(n; \sigma)$ for every
\begin{equation}
{\bf G} = \left(
\begin{array}{c c}
{\bf X} & {\bf 0}\\
{\bf 0} & {\bf X}^{-1}
\end{array}
\right) \in Sp^+ (n; \sigma).
\label{eqLemmaSymplecticAntiSymplectic1}
\end{equation}
Then ${\bf M}$ is either $\sigma$-symplectic or $\sigma$-anti-symplectic.
\end{lemma}
\begin{theorem}\label{TheoremIdenticalSpectra1}
Let $\omega_1$, $\omega_2$ denote two symplectic forms on $\mathbb{R}^{2n}$. Then $Spec_{\omega_1} ({\bf A})= Spec_{\omega_2} ({\bf A})$ for all ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ if and only if $\omega_1 = \pm \omega_2$.
\end{theorem}
\begin{proof}
Let us first assume that $\omega_1= \sigma$, $\omega_2= \omega$ and $Spec_{\sigma} ({\bf A})= Spec_{\omega} ({\bf A})$ for all ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$. By the $\sigma$-Williamson Theorem (Theorem \ref{Williamsontheorem2}) and the $\omega$-Williamson Theorem (Theorem \ref{TheoremOmegaWilliamsonTheorem1}), there exist ${\bf P_{A}} \in Sp(n; \sigma)$ and ${\bf S_A} \in \mathcal{D} (n; \omega)$ such that
\begin{equation}
{\bf S_A^{-1} A S_A^{-1 T}} = {\bf P_A A P_A^T}.
\label{eqTheoremIdenticalSpectra1}
\end{equation}
Consequently
\begin{equation}
{\bf A} = {\bf S_A^{\prime} A S_A^{\prime T}},
\label{eqTheoremIdenticalSpectra2}
\end{equation}
where
\begin{equation}
{\bf S_A^{\prime}} = {\bf S_A P_A} \in \mathcal{D} (n; \omega).
\label{eqTheoremIdenticalSpectra3}
\end{equation}
From (\ref{eqTheoremIdenticalSpectra2}) we conclude that there exists ${\bf S_A^{\prime}} \in \mathcal{D} (n; \omega)$ such that:
\begin{equation}
{\bf S_A^{\prime}} = {\bf A} ({\bf S_A^{\prime T}})^{-1} {\bf A}^{-1}.
\label{eqTheoremIdenticalSpectra4}
\end{equation}
Since this holds for all ${\bf A} \in \mathbb{P}_{2n \times 2n}^{\times} (\mathbb{R})$, in particular it is true for ${\bf A} \in Sp^+ (n; \sigma)$. Consequently:
\begin{equation}
\begin{array}{c}
{\bf \Omega} = {\bf S_A^{\prime} J} {\bf S_A^{\prime}}^T = {\bf A} ({\bf S_A^{\prime T}})^{-1} {\bf A}^{-1} {\bf J} ({\bf A}^{-1})^T ({\bf S_A^{\prime }})^{-1} {\bf A} =\\
\\
={\bf A} ({\bf S_A^{\prime T}})^{-1} {\bf J} ({\bf S_A^{\prime }})^{-1} {\bf A}
= - {\bf A} \left( {\bf S_A^{\prime} J} {\bf S_A^{\prime}}^T \right)^{-1} {\bf A } =\\
\\
= -{\bf A}{\bf \Omega}^{-1}{\bf A}
\end{array}
\label{eqTheoremIdenticalSpectra5}
\end{equation}
So, we have proven that
\begin{equation}
{\bf \Omega}^T= {\bf A}{\bf \Omega}^{-1}{\bf A}
\label{eqTheoremIdenticalSpectra6}
\end{equation}
for all ${\bf A} \in Sp^+ (n; \sigma)$.
Next, fix an arbitrary ${\bf S} \in \mathcal{D} (n; \omega)$. From (\ref{eqTheoremIdenticalSpectra5}) and (\ref{eqsymplecticform9}) we have:
\begin{equation}
{\bf S} {\bf J} {\bf S}^T = {\bf \Omega}= - {\bf A} ({\bf S} {\bf J} {\bf S}^T )^{-1} {\bf A} \Leftrightarrow {\bf S}^{-1} {\bf A} ({\bf S}^T)^{-1} {\bf J} \left( {\bf S}^{-1} {\bf A} ({\bf S}^T)^{-1}\right)^T = {\bf J},
\label{eqTheoremIdenticalSpectra7}
\end{equation}
for all ${\bf A} \in Sp^+ (n; \sigma)$. In other words ${\bf S}^{-1} {\bf A} ({\bf S}^T)^{-1} \in Sp^+ (n; \sigma)$ for all ${\bf A} \in Sp^+ (n; \sigma)$. But from Lemma \ref{LemmaSymplecticAntiSymplectic1}, this is possible if and only if ${\bf S}^{-1}$ is $\sigma$-symplectic or $\sigma$-antisymplectic. In the first case, we have $\omega= \sigma$ and in the second case $\omega= - \sigma$.
Next consider arbitrary symplectic forms $\omega_1 (z,z^{\prime})=z \cdot {\bf \Omega_1}^{-1} z^{\prime}$, $\omega_2 (z,z^{\prime})=z \cdot {\bf \Omega_2}^{-1} z^{\prime}$ such that $Spec_{\omega_1} ({\bf A})= Spec_{\omega_2} ({\bf A})$ for all ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$. Now let ${\bf S_1} \in \mathcal{D} (n; \omega_1)$. For any ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$, we have that ${\bf S_1}{\bf A} {\bf S_1}^T\in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$. On the other hand, it is easy to verify that (cf. (\ref{eq12.7}))
\begin{equation}
Spec_{\omega_1} ({\bf S_1}{\bf A} {\bf S_1}^T)= Spec_{\sigma} ({\bf A}),
\label{eqTheoremIdenticalSpectra8}
\end{equation}
and
\begin{equation}
Spec_{\omega_2} ({\bf S_1}{\bf A} {\bf S_1}^T)= Spec_{\omega} ({\bf A}),
\label{eqTheoremIdenticalSpectra9}
\end{equation}
with
\begin{equation}
\omega (z,z^{\prime})=z \cdot {\bf \Omega}^{-1} z^{\prime} \hspace{1 cm} {\bf \Omega} = {\bf S_1}^{-1} {\bf \Omega_2} ({\bf S_1}^{-1})^T.
\label{eqTheoremIdenticalSpectra10}
\end{equation}
We conclude that
\begin{equation}
Spec_{\omega_1} ({\bf S_1}{\bf A} {\bf S_1}^T)= Spec_{\omega_2} ({\bf S_1}{\bf A} {\bf S_1}^T) \Leftrightarrow Spec_{\sigma} ({\bf A}) = Spec_{\omega} ({\bf A}),
\label{eqTheoremIdenticalSpectra11}
\end{equation}
for all ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$. From the previous analysis $\omega = \pm \sigma$ or, equivalently, $\omega_1 = \pm \omega_2$.
\end{proof}
Let us now prove the following refinement of Theorem \ref{TheoremIdenticalSpectra1}. We have proven that for $\omega_2 \ne \pm \omega_1$, we can always find matrices ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ with $Spec_{\omega_1} ({\bf A}) \ne Spec_{\omega_2} ({\bf A})$. We now show that in fact the smallest $\omega_1$- and $\omega_2$-Williamson invariants cannot coincide for all ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$.
\begin{theorem}\label{TheoremUnboudedQuotient1}
Let $\omega_1, \omega_2$ be symplectic forms on $\mathbb{R}^{2n}$ such that $\omega_1 \ne \pm \omega_2$. For any $K>0$ there exists a matrix ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ such that
\begin{equation}
\frac{\lambda_{\omega_1,1} ({\bf A})}{\lambda_{\omega_2,1} ({\bf A})} \ge K.
\label{eqUnboudedQuotient1}
\end{equation}
\end{theorem}
\begin{proof}
Following an argument similar to that in Theorem \ref{TheoremIdenticalSpectra1}, we may assume without loss of generality that $\omega_1 = \sigma$ and set $\omega_2 = \omega$. Suppose that there exists $K>0$ such that
\begin{equation}
\frac{\lambda_{\sigma,1} ( {\bf A})}{\lambda_{\omega,1} ({\bf A})} < K,
\label{eqUnboudedQuotient2}
\end{equation}
for all ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$.
It follows from items 2 and 3 of Theorem \ref{TheoremNWspectrumGaussians} (setting $\alpha = \pm \lambda_{\sigma,1} ({\bf A})$ and ${\bf \Omega}={\bf J}$) that
\begin{equation}
{\bf A} \pm i \lambda_{\sigma,1} ({\bf A}) {\bf J} \ge 0,
\label{eqUnboudedQuotient3}
\end{equation}
while (setting $\alpha = \pm \frac{\lambda_{\sigma,1} ({\bf A})}{K}$, and taking (\ref{eqUnboudedQuotient2}) into account)
\begin{equation}
{\bf A} \pm i \frac{\lambda_{\sigma,1} ({\bf A})}{K} {\bf \Omega} \ge 0.
\label{eqUnboudedQuotient4}
\end{equation}
From the previous inequality, we also have:
\begin{equation}
{\bf A} +i \lambda_{\sigma,1} ({\bf A}) {\bf J} + i\lambda_{\sigma,1} ({\bf A}) \left(\pm \frac{ {\bf \Omega}}{K} - {\bf J} \right) \ge 0.
\label{eqUnboudedQuotient5}
\end{equation}
Let $u_1=e_1 + i f_1$ be an eigenvector of ${\bf A J^{-1}} $ associated with the eigenvalue $-i \lambda_{\sigma,1} ({\bf A})$, i.e.
\begin{equation}
{\bf AJ^{-1}} u_1 = -i \lambda_{\sigma,1} ({\bf A}) u_1.
\label{eqUnboudedQuotient6}
\end{equation}
Multiplying (\ref{eqUnboudedQuotient5}) on the left by $\overline{ u_1}\cdot {\bf J}$, on the right by ${\bf J^T} u_1={\bf J^{-1}} u_1$, and taking into account that $\lambda_{\sigma,1} ({\bf A}) >0$ yields:
\begin{equation}
i \overline{u_1} \cdot {\bf J} \left(\pm \frac{ {\bf \Omega}}{K} - {\bf J} \right) {\bf J^{-1}} u_1 \ge 0.
\label{eqUnboudedQuotient7}
\end{equation}
If we define the symplectic form
\begin{equation}
\eta (z, z^{\prime}):= \frac{1}{K} z \cdot {\bf J \Omega J^T} z^{\prime} ,
\label{eqUnboudedQuotient8}
\end{equation}
for $z,z^{\prime} \in \mathbb{R}^{2n}$, we get from (\ref{eqUnboudedQuotient7}):
\begin{equation}
\pm \eta (f_1,e_1) + \sigma (f_1,e_1) \ge 0.
\label{eqUnboudedQuotient9}
\end{equation}
So the previous equation holds for all vectors $e_1,f_1$ such that $u_1= e_1 + i f_1$ is an eigenvector of $i {\bf A J^{-1}}$ for some matrix ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ with eigenvalue equal to the smallest Williamson invariant of ${\bf A}$. By Proposition \ref{PropostionEigenvector} such a matrix exists for all vectors $e, f \in \mathbb{R}^{2n}$ such that $\sigma (f,e) >0$. So, in fact, we have
\begin{equation}
|\eta (f,e)| \le \sigma (f,e) ,
\label{eqUnboudedQuotient10}
\end{equation}
for all vectors $e,f$ such that $\sigma (f,e) >0$. If $\sigma (f,e) <0$, then (since $\sigma (f,e) = -\sigma (e,f) )$, it follows that
\begin{equation}
|\eta (f,e)| \le \sigma (e,f) .
\label{eqUnboudedQuotient11}
\end{equation}
Altogether, we may write
\begin{equation}
|\eta (f,e)| \le |\sigma (f,e)| ,
\label{eqUnboudedQuotient12}
\end{equation}
whenever $\sigma (f,e) \ne 0$. Suppose that $\sigma (f,e) = 0$. We may regard $f$ and $e$ as elements of some Lagrangian plane \cite{Cannas}. Since Lagrangian planes have no interior points, we can find a sequence $\left\{e_n \right\}_{n \in \mathbb{N}}$ converging to $e$, such that $\sigma (f,e_n) \ne 0$. But from (\ref{eqUnboudedQuotient12}), we have:
\begin{equation}
|\eta (f,e_n)| \le |\sigma (f,e_n)|.
\label{eqUnboudedQuotient12.1}
\end{equation}
However, since symplectic forms are continuous, inequality (\ref{eqUnboudedQuotient12}) must hold for all $e,f \in \mathbb{R}^{2n}$. By Proposition \ref{PropositionSymplecticInequality}, there exists a constant $0< |a | \le 1$ such that
\begin{equation}
\eta = a \sigma .
\label{eqUnboudedQuotient13}
\end{equation}
From (\ref{eqUnboudedQuotient8}), we have:
\begin{equation}
{\bf \Omega} = - a K {\bf J}.
\label{eqUnboudedQuotient14}
\end{equation}
Since the symplectic form $\det ({\bf \Omega})=1$, we conclude that $\omega = \pm \sigma$ and we have a contradiction. Thus (\ref{eqUnboudedQuotient2}) cannot hold.
\end{proof}
\subsection{The landscape of Wigner functions}
As an application of the previous result, we prove the following Theorem on the symplectic covariance of Wigner functions.
\begin{theorem}\label{TheoremSymplecticCovariance}
Let ${\bf M} \in Gl(2n; \mathbb{R})$. Then the operator
\begin{equation}
F (z) \mapsto (\mathcal{U}_{{\bf M}} F)(z) := | \det{\bf M}| F (Mz)
\label{eqSymplecticCovariance1}
\end{equation}
maps (pure or mixed) $\omega$-Wigner functions to $\omega$-Wigner functions if and only if ${\bf M}$ is either $\omega$-symplectic or $\omega$-anti-symplectic.
\end{theorem}
\begin{proof}
One calls a transformation which maps quantum states to quantum states a quantum channel or a quantum dynamical map.
We start by considering the case $\omega =\pm \sigma$. It is a well documented fact that $\mathcal{U}_{{\bf M}}$ is a quantum dynamical map if ${\bf M}$ is a $\sigma$-symplectic or $\sigma$-anti-symplectic matrix \cite{Dias5}.
Conversely, suppose that $\mathcal{U}_{{\bf M}}$ is a quantum dynamical map. Then, in particular it maps every Gaussian $\sigma$-Wigner function $\mathcal{G}_{{\bf A}}$ with covariance matrix ${\bf A}$ to another Gaussian $\sigma$-Wigner function with covariance matrix ${\bf M^{-1}} {\bf A} ({\bf M^{-1}})^T$.
For every ${\bf A} \in \mathbb{P}_{2n \times 2n}^{\times} (\mathbb{R})$, the matrix $\frac{1}{2 \lambda_{\sigma,1} ({\bf A})} {\bf A} $ satisfies the $\sigma$-RSUP. Indeed, acccording to Lemma \ref{LemmaRescalings1}, its smallest Williamson invariant is $\frac{1}{2 \lambda_{\sigma,1} ({\bf A})} \lambda_{\sigma,1} ({\bf A})=\frac{1}{2}$. Hence the matrix $\frac{1}{2 \lambda_{\sigma,1} ({\bf A})} {\bf M^{-1}} {\bf A} ({\bf M^{-1}})^T$ must also satisfy the $\sigma$-RSUP:
\begin{equation}
\frac{1}{2 \lambda_{\sigma,1} ({\bf A})} {\bf M^{-1}} {\bf A} ({\bf M^{-1}})^T + \frac{i}{2} {\bf J} \ge 0 \Leftrightarrow {\bf A} + i \lambda_{\sigma,1} ({\bf A}){\bf M} {\bf J} {\bf M}^T \ge 0.
\label{eqWignerCovariance2}
\end{equation}
Next define ${\bf \Omega} = \gamma^{-1} {\bf M} {\bf J} {\bf M}^T$, where $\gamma = | \det {\bf M}|^{1/n}$. Thus $\omega (z,z^{\prime}) = z^T {\bf \Omega^{-1}} z^{\prime}$ is a normalized symplectic form. From (\ref{eqWignerCovariance2}), we have:
\begin{equation}
{\bf A} + i \mu {\bf \Omega} \ge 0,
\label{eqWignerCovariance3}
\end{equation}
where $\mu = \lambda_{\sigma,1} ({\bf A}) \gamma$. Consequently:
\begin{equation}
\lambda_{\omega,1} ({\bf A}) \ge \mu \Leftrightarrow \frac{\lambda_{\sigma,1} ({\bf A})}{\lambda_{\omega,1} ({\bf A})} \le \frac{1}{\gamma},
\label{eqWignerCovariance4}
\end{equation}
for all ${\bf A} \in \mathbb{P}_{2n \times 2n}^{\times} (\mathbb{R})$. But according to Theorem \ref{TheoremUnboudedQuotient1}, this can happen only if $\omega = \pm \sigma$, or ${\bf \Omega} = \pm {\bf J}$. This means that $\gamma^{-1/2} {\bf M}$ is a $\sigma$-symplectic or a $\sigma$-anti-symplectic matrix. Consequently, if $\mathcal{W}^{{\bf J}} (\mathcal{F}W \rho)$ denotes the Narcowich-Wigner spectrum of some Wigner function $W \rho$, then the Narcowich-Wigner spectrum of $\mathcal{U}_M (W \rho)$ becomes $\mathcal{W}^{{\bf J}} \left( \mathcal{F}\mathcal{U}_M (W \rho)\right)= \gamma^{-1} \mathcal{W}^{{\bf J}} (\mathcal{F}W \rho)$. Let $\psi \in L^2 (\mathbb{R}^n)$ be some non-Gaussian state. Then $\mathcal{W}^{{\bf J}} (\mathcal{F}W \psi) = \left\{ - 1, 1 \right\}$ (see Theorem \ref{TheoremPropertiesNWSpectra}, item 6). It follows that $\mathcal{W}^{{\bf J}} (\mathcal{F} \mathcal{U}_M W \psi) = \left\{ - \frac{1}{\gamma}, \frac{1}{\gamma} \right\}$. And thus, $\mathcal{U}_M W \psi$ is a Wigner function if and only if $\gamma =1$.
Next, assume that $\omega \ne \pm \sigma$. Let $W^{\omega} \rho$ be an arbitrary $\omega$-Wigner function. Then, given a Darboux matrix ${\bf S} \in \mathcal{D}(n; \omega)$, the function
\begin{equation}
W^{\sigma} \rho (z)= W^{\omega} \rho ({\bf S} z)
\label{eqWignerCovariance4.1}
\end{equation}
is a $\sigma$-Wigner function.
Define the matrix ${\bf P}$ by way of
\begin{equation}
{\bf P}:= {\bf S^{-1} M S}.
\label{eqWignerCovariance4.2}
\end{equation}
We then have that $\left(\mathcal{U}_{{\bf M}} W^{\omega} \rho\right) (z)$ is a $\omega$-Wigner function if and only if $\left(\mathcal{U}_{{\bf M}} W^{\omega} \rho \right) ({\bf S} z)$ is a $\sigma$-Wigner function. On the other hand:
\begin{equation}
\begin{array}{c}
\left(\mathcal{U}_{{\bf M}} W^{\omega} \rho \right) ({\bf S} z)= | \det {\bf M} | W^{\omega} \rho ({\bf MS} z) = | \det {\bf M} | W^{\sigma} \rho ({\bf S^{-1} MS} z) =\\
\\
=| \det {\bf P} | W^{\sigma} \rho ({\bf P} z) = \left(\mathcal{U}_{{\bf P}} W^{\sigma} \rho \right) ( z)
\end{array}
\label{eqWignerCovariance4.3}
\end{equation}
Thus, $\mathcal{U}_{{\bf M}} W^{\omega} \rho$ is a $\omega$-Wigner function for every density matrix $\widehat{\rho}$ if and only if $\mathcal{U}_{{\bf P}} W^{\sigma} \rho$ is a $\sigma$-Wigner function for every density matrix $\widehat{\rho}$. This, in turn, happens if and only if ${\bf P}$ is $\sigma$-symplectic or $\sigma$-anti-symplectic. But, from (\ref{eqWignerCovariance4.2}), this is equivalent to ${\bf M} $ being $\omega$-symplectic or $\omega$-anti-symplectic.
\end{proof}
\begin{remark}\label{Remarksymplecticcovariance}
This theorem is a small improvement on the results of \cite{Dias5} for $\omega= \pm \sigma$. Indeed, here we admit {\it a priori} an arbitrary value of $\det {\bf M} (\ne 0)$. In \cite{Dias5} we assumed $\det {\bf M}=1$ {\it ab initio}.
\end{remark}
We now provide a characterization of the sets of classical and quantum states with arbitrary symplectic structure.
Let $\mathcal{L} (\mathbb{R}^{2n})$ denote the set of Liouville measures on $\mathbb{R}^{2n}$. These are the continuous, pointwise non-negative functions normalized to unity. In other words, they are classical probability densities on the phase-space. Moreover, let $\mathcal{F}^{\omega_1} (\mathbb{R}^{2n})$ and $\mathcal{F}^{\omega_2} (\mathbb{R}^{2n})$ be the convex sets of $\omega_1$- and $\omega_2$-Wigner functions for some symplectic forms $\omega_1$, $\omega_2$ such that $\omega_2 \ne \pm \omega_1$. If the dimension is clear, we shall simply write $\mathcal{L},\mathcal{F}^{\omega_1},\mathcal{F}^{\omega_2}$.
In \cite{Bastos2} we proved that the sets
\begin{equation}
\begin{array}{l}
\mathcal{A}_1 = \mathcal{F}^{\omega_1} \backslash \left( \mathcal{F}^{\omega_2} \cup \mathcal{L} \right)\\
\mathcal{A}_2 = \mathcal{F}^{\omega_2} \backslash \left( \mathcal{F}^{\omega_1} \cup \mathcal{L} \right)\\
\mathcal{A}_3 = \mathcal{L} \backslash \left( \mathcal{F}^{\omega_1} \cup \mathcal{F}^{\omega_2} \right)\\
\mathcal{A}_4 = \left( \mathcal{F}^{\omega_1} \cap \mathcal{F}^{\omega_2} \right) \backslash \mathcal{L} \\
\mathcal{A}_5 = \left( \mathcal{F}^{\omega_1} \cap \mathcal{L} \right) \backslash \mathcal{F}^{\omega_2} \\
\mathcal{A}_6 = \left( \mathcal{F}^{\omega_2} \cap \mathcal{L} \right) \backslash \mathcal{F}^{\omega_1}\\
\mathcal{A}_7 = \mathcal{F}^{\omega_1} \cap \mathcal{F}^{\omega_2} \cap \mathcal{L}
\end{array}
\label{eqSets1}
\end{equation}
are all non-empty, when $n=2$, $\omega_1 = \sigma$ and $\omega_2 = \omega$ given by (\ref{eqNCQM1}).
Here, we generalize the result for arbitrary dimension and symplectic forms $\omega_1 \ne \pm \omega_2$.
\begin{figure}
\begin{center}
\vspace{-4cm} \hspace{-4cm}\includegraphics [scale=0.5]{1.pdf}
\caption{Different sets of functions and their intersections.}
\end{center}
\label{diagrama}
\end{figure}
\begin{theorem}\label{TheoremSets1}
Let $\omega_1$, $\omega_2$ be arbitrary symplectic forms on
$\mathbb{R}^{2n}$ with $\omega_1 \ne \pm \omega_2$. Then the sets
$\mathcal{A}_i$, $i=1, \cdots,7$ defined as in (\ref{eqSets1}) are
all non-empty (see Figure $1$).
\end{theorem}
\begin{proof}
As in \cite{Bastos2} we shall suggest ways to construct examples of functions in each of these sets. We start with
\noindent
${\bf \mathcal{A}_3:}$ Consider an arbitrary Gaussian function $\mathcal{G}_{{\bf A}}$ with covariance matrix ${\bf A}$. Clearly $\mathcal{G}_{{\bf A}} \in \mathcal{L} $. Compute the $\omega_1$- and $\omega_2$-symplectic spectra of ${\bf A}$. Let $\lambda_{max,1} ({\bf A})= max \left\{ \lambda_{\omega_1,1} ({\bf A}), \lambda_{\omega_2,1} ({\bf A})\right\}$. If $\lambda_{max,1} ({\bf A})< \frac{1}{2}$, then $\mathcal{G}_{{\bf A}} \notin \mathcal{F}^{\omega_1} \cup \mathcal{F}^{\omega_2}$ and we are done. If $\lambda_{max,1} ({\bf A}) \ge \frac{1}{2}$, then choose some $0 < \mu < \frac{1}{2 \lambda_{max,1}({\bf A})}$. According to Lemma \ref{LemmaRescalings1}, the Gaussian $\mathcal{G}_{\mu {\bf A}}$ is such that $\lambda_{max,1} (\mu {\bf A})< \frac{1}{2}$ and thus $\mathcal{G}_{\mu {\bf A}} \notin \mathcal{F}^{\omega_1} \cup \mathcal{F}^{\omega_2}$.
\vspace{0.3 cm}\noindent
${\bf \mathcal{A}_7:}$ Again consider a Gaussian $\mathcal{G}_{{\bf A}}$ with covariance matrix ${\bf A}$ and let $\lambda_{min,1} ({\bf A})= min \left\{ \lambda_{\omega_1,1} ({\bf A}), \lambda_{\omega_2,1} ({\bf A})\right\}$. If $\lambda_{min,1} ({\bf A}) \ge \frac{1}{2}$, then $\mathcal{G}_{{\bf A}} \in \mathcal{A}_7$. Otherwise, let $\mu \ge \frac{1}{2 \lambda_{min,1} ({\bf A})}$. Then $\lambda_{min,1} (\mu {\bf A})\ge \frac{1}{2}$ and $\mathcal{G}_{\mu {\bf A}} \in \mathcal{A}_7$.
\vspace{0.3 cm}\noindent
${\bf \mathcal{A}_5:}$ According to Theorem \ref{TheoremUnboudedQuotient1}, there exists a Gaussian $\mathcal{G}_{{\bf A}}$ with covariance matrix ${\bf A}$, such that $\lambda_{\omega_1,1} ({\bf A}) > \lambda_{\omega_2,1} ({\bf A})$. Let $\mu$ be such that $2\lambda_{\omega_1,1} ({\bf A}) > \mu^{-1} > 2 \lambda_{\omega_2,1} ({\bf A})$. Then the Gaussian $\mathcal{G}_{\mu {\bf A}}$ is such that $\lambda_{\omega_1,1} (\mu {\bf A}) \ge \frac{1}{2}$, while $\lambda_{\omega_2,1} (\mu {\bf A}) < \frac{1}{2}$. Hence $\mathcal{G}_{\mu {\bf A}} \in \mathcal{A}_5$.
\vspace{0.3 cm}\noindent
${\bf \mathcal{A}_6:}$ The result follows the same steps as in $\mathcal{A}_5$, with the replacement $\omega_1 \leftrightarrow \omega_2$.
\vspace{0.3 cm}\noindent
${\bf \mathcal{A}_1:}$ Let $h \in \mathcal{F}^{\sigma} \backslash \mathcal{L}$, that is a $\sigma$-Wigner function which is not everywhere nonnegative. Then the function $h_1(z) = h({\bf S_1}^{-1}z)$ with ${\bf S_1} \in \mathcal{D}(n; \omega_1)$ is in $\mathcal{F}^{\omega_1} \backslash \mathcal{L}$. Choose $h_5 \in \mathcal{A}_5$. Since $h_5 \notin \mathcal{F}^{\omega_2}$, there exists $g_2 \in \mathcal{F}^{\omega_2} $, such that (cf.(\ref{eqQuantumConditions2})):
\begin{equation}
b := \int_{\mathbb{R}^{2n}} h_5(z) g_2(z)dz < 0.
\label{eqSets2}
\end{equation}
Let also
\begin{equation}
a := \int_{\mathbb{R}^{2n}} h_1(z) g_2(z)dz .
\label{eqSets3}
\end{equation}
First suppose that $a \le 0$. Let $z_0 \in \mathbb{R}^{2n}$ be such that $h_1 (z_0) < 0 < h_5 (z_0)$. Such a $z_0$ can always be found because $h_5 (z) \ge 0$ for all $z$ and is not identically zero. On the other hand if $h_1 (z) \in \mathcal{F}^{\omega_1} \backslash \mathcal{L}$, then $h_1 (z - \zeta) \in \mathcal{F}^{\omega_1} \backslash \mathcal{L}$ for any fixed $\zeta$. So we can always translate $h_1$ so that $h_1 (z_0) < 0$. Next, choose $1>p>0$ such that
\begin{equation}
\frac{p}{1-p} > \frac{h_5(z_0)}{|h_1(z_0)|}.
\label{eqSets4}
\end{equation}
The function
\begin{equation}
f_1(z) = p h_1(z) + (1-p) h_5(z)
\label{eqSets5}
\end{equation}
being a convex combination of $h_1, h_5 \in \mathcal{F}^{\omega_1}$ also belongs to $\mathcal{F}^{\omega_1}$. Moreover, since by (\ref{eqSets4}), $f_1(z_0)<0$, we also have $f_1 \notin \mathcal{L}$. Finally from (\ref{eqSets2},\ref{eqSets3}):
\begin{equation}
\int_{\mathbb{R}^{2n}} f_1(z) g_2(z)dz =p a + (1-p)b <0,
\label{eqSets6}
\end{equation}
and thus $f_1 \notin \mathcal{F}^{\omega_2}$. Altogether, $f_1 \in \mathcal{A}_1$.
Now suppose that we have instead $a > 0$. We start by showing that, by translating $h_1(z) \mapsto T_{\zeta} h_1(z) = h_1(z- \zeta)$ appropriately , we can always find $z_1 \in \mathbb{R}^{2n}$ such that
\begin{equation}
T_{\zeta} h_1 (z_1) < 0< \frac{h_5(z_1)}{|b|} < \frac{|\left(T_{\zeta}h_1 \right)(z_1)|}{a}.
\label{eqSets7}
\end{equation}
Indeed, let $z_2 \in \mathbb{R}^{2n}$ be such that $h_1(z_2) <0$ and $z_1$ such that
\begin{equation}
0< h_5(z_1) < (2 \pi)^n |b h_1(z_2)|.
\label{eqSets8}
\end{equation}
It is always possible to find such a $z_1$ because we may assume, without compromising our argument, that $h_5 \in \mathcal{S} ( \mathbb{R}^{2n})$. If $\zeta = z_1-z_2$, then we have from equation (\ref{eqSets8}), the Cauchy-Schwartz inequality, the purity condition (\ref{eqMoyalidentity4}), and by replacing $h_1$ by $T_{\zeta} h_1$ in (\ref{eqSets3}):
\begin{equation}
\begin{array}{c}
0 < \frac{h_5(z_1)}{|b| } a = \frac{h_5(z_1)}{|b| } \left| \int_{\mathbb{R}^{2n}} T_{\zeta} h_1 (z) g_2 (z) \right| \le\\
\\
\le \frac{h_5(z_1)}{|b| } ||T_{\zeta} h_1 ||_{L^2} ||g_2||_{L^2} \le \frac{h_5(z_1)}{(2 \pi)^n |b| } < |h_1(z_2)| = | T_{\zeta} h_1 (z_1)|,
\end{array}
\label{eqSets9}
\end{equation}
which proves (\ref{eqSets7}) for $ T_{\zeta} h_1 $.
From (\ref{eqSets7}) we can choose $p \in \mathbb{R}$ such that
\begin{equation}
0 < \frac{\alpha}{1+ \alpha} <p< \frac{\beta}{1+ \beta} <1,
\label{eqSets10}
\end{equation}
with
\begin{equation}
0< \alpha =\frac{h_5(z_1)}{|\left(T_{\zeta} h_1 \right)(z_1)|} < \frac{|b|}{a}= \beta.
\label{eqSets10.1}
\end{equation}
This is because the function $x \in \mathbb{R}^+ \mapsto \frac{x}{1+x}$ is strictly increasing.
Next, let $f_1 \in \mathcal{F}^{\omega_1}$ be defined as in (\ref{eqSets5}) with $h_1$ replaced by $T_{\zeta} h_1$. Since
\begin{equation}
f_1(z_1) = -p |\left(T_{\zeta} h_1 \right)(z_1)| + (1-p) h_5(z_1) <0
\label{eqSets11}
\end{equation}
we have $f_1 \notin \mathcal{L}$. Similarly:
\begin{equation}
\int_{\mathbb{R}^{2n}} f_1(z) g_2(z) dz = p a - (1-p) |b| <0,
\label{eqSets12}
\end{equation}
and thus $f_1 \notin \mathcal{F}^{\omega_2}$. Altogether $f_1 \in \mathcal{A}_1$.
\vspace{0.3 cm}\noindent
${\bf \mathcal{A}_2:}$ The result follows the same steps as in $\mathcal{A}_1$, with the replacement $\omega_1 \leftrightarrow \omega_2$.
\vspace{0.3 cm}\noindent
${\bf \mathcal{A}_4:}$ To simplify the argument, we start by noticing that proving that $f_4 \in \mathcal{A}_4$ is equivalent to proving that $g_4 \in \left( \mathcal{F}^{\sigma} \cap \mathcal{F}^{\omega} \right) \backslash \mathcal{L}$ where $g_4(z) = f_4 ({\bf S_1}z)$, $\omega (z,z^{\prime}) = z \cdot {\bf \Omega}^{-1} z^{\prime}$, ${\bf \Omega} = {\bf S_1}^{-1} {\bf \Omega_2} ({\bf S_1}^{-1})^T$ with ${\bf S_1} \in \mathcal{D} (n; \omega_1)$, ${\bf S_2} \in \mathcal{D} (n; \omega_2)$. Indeed, $f_4 \notin \mathcal{L} \Leftrightarrow g_4 \notin \mathcal{L}$. Moreover, $f_4 (z) \in \mathcal{F}^{\omega_1} \Leftrightarrow g(z) =f_4 ({\bf S_1} z) \in \mathcal{F}^{\sigma}$. Finally, $f_4 (z) \in \mathcal{F}^{\omega_2} \Leftrightarrow f_4 ({\bf S_2} z) = g_4 ({\bf S_1}^{-1} {\bf S_2} z) \in \mathcal{F}^{\sigma}$. But since ${\bf S_1}^{-1} {\bf S_2} \in \mathcal{D} (n; \omega)$, this is equivalent to $g_4 \in \mathcal{F}^{\omega}$.
Let us now devise a way to construct a function $g_4 \in \left( \mathcal{F}^{\sigma} \cap \mathcal{F}^{\omega} \right) \backslash \mathcal{L}$. Consider the function
\begin{equation}
\gamma (\alpha)= \det ({\bf J} - \alpha {\bf \Omega} ).
\label{eqSets13}
\end{equation}
We claim that there exists $ \alpha_0 >1$ such that
\begin{equation}
\gamma_0 := \gamma (\alpha_0) >0.
\label{eqSets14}
\end{equation}
Indeed, $\gamma (\alpha)=\det \left({\bf J \Omega^{-1}} - \alpha {\bf I} \right)$. Thus $\gamma$ is the characteristic polynomial of the matrix ${\bf J \Omega^{-1}}$. We thus have:
\begin{equation}
\begin{array}{c}
\gamma (\alpha)=\alpha^{2n} + Tr \left({\bf J \Omega^{-1}} \right) \alpha^{2n-1} + \cdots + \det \left({\bf J \Omega^{-1}} \right) = \\
\\
=\alpha^{2n} + Tr \left({\bf J \Omega^{-1}} \right) \alpha^{2n-1} + \cdots +1.
\end{array}
\label{eqSets14.1}
\end{equation}
Hence $\gamma (\alpha) \simeq \alpha^{2n}$ as $\alpha \to + \infty$, which proves the claim.
Define ${\bf \Sigma_1} = {\bf \Omega}$ and ${\bf \Sigma_2} = \gamma_0^{-\frac{1}{2n}} ({\bf J} - \alpha_0 {\bf \Omega})$ and the corresponding normalized symplectic forms:
\begin{equation}
\eta_1 = \omega, \hspace{1 cm} \eta_2 (z, z^{\prime}) = z \cdot {\bf \Sigma_2^{-1}} z^{\prime}.
\label{eqSets14.1}
\end{equation}
We claim that there exists ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ such that
\begin{equation}
\lambda_{\eta_1,1} ({\bf A}) = \alpha_0 -1,
\label{eqSets15}
\end{equation}
and
\begin{equation}
\lambda_{\eta_2,1} ({\bf A}) \ge \gamma_0^{\frac{1}{2n}}.
\label{eqSets15.1}
\end{equation}
Indeed, from Theorem \ref{TheoremUnboudedQuotient1}, we can find ${\bf A} \in \mathbb{P}^{\times}_{2n \times 2n} (\mathbb{R})$ such that
\begin{equation}
\frac{\lambda_{\eta_2,1} ({\bf A})}{\lambda_{\eta_1,1} ({\bf A})} \ge \frac{\gamma_0^{\frac{1}{2n}}}{\alpha_0 -1}.
\label{eqSets16}
\end{equation}
After a possible rescaling ${\bf A} \mapsto \mu {\bf A}$ $(\mu >0)$, we obtain (\ref{eqSets15}) and (\ref{eqSets15.1}) follows from (\ref{eqSets16}).
Conditions (\ref{eqSets15},\ref{eqSets15.1}) mean that
\begin{equation}
{\bf A} + i (\alpha_0 -1) {\bf \Omega} \ge 0,
\label{eqSets18}
\end{equation}
while
\begin{equation}
{\bf A} + i \alpha_0 {\bf \Omega} \ngeq 0.
\label{eqSets19}
\end{equation}
On the other hand, we also have
\begin{equation}
{\bf A} + i \gamma_0^{\frac{1}{2n}} {\bf \Sigma_2} \ge 0.
\label{eqSets21}
\end{equation}
Next, let $h^{\sigma}$ be a non-Gaussian pure state with Planck constant $\alpha_0$, that is (cf. item 6 of Theorem \ref{TheoremPropertiesNWSpectra})
\begin{equation}
\mathcal{W}^{{\bf J}} (\mathcal{F} h^{\sigma})= \left\{- \alpha_0,\alpha_0 \right\} .
\label{eqSets21.1}
\end{equation}
Given ${\bf S} \in \mathcal{D} (n; \omega)$, let $h^{\omega} (z) = h^{\sigma} ({\bf S^{-1}} z)$. From Remark \ref{Remark5.1}:
\begin{equation}
\left\{- \alpha_0, \alpha_0 \right\} \subset \mathcal{W}^{{\bf \Omega}} (\mathcal{F} h^{\omega}) .
\label{eqSets21.2}
\end{equation}
If $\mathcal{G}_{{\bf A}}$ is a Gaussian with covariance matrix ${\bf A}$, then from (\ref{eqSets18}-\ref{eqSets21}):
\begin{equation}
(\alpha_0 -1, {\bf \Omega}) , (\gamma_0^{\frac{1}{2n}}, {\bf \Sigma_2}) \in \mathcal{W} (\mathcal{F} \mathcal{G}_{{\bf A}}),
\label{eqSets22}
\end{equation}
but
\begin{equation}
(\alpha_0, {\bf \Omega}) \notin \mathcal{W} (\mathcal{F} \mathcal{G}_{{\bf A}}).
\label{eqSets23}
\end{equation}
Given (\ref{eqSets23}), we can always find $h^{\sigma} \notin \mathcal{L}$ such that
\begin{equation}
g_4 = \mathcal{G}_{{\bf A}} \star h^{\omega} \notin \mathcal{L}.
\label{eqSets24}
\end{equation}
Indeed, we have for ${\bf S} \in \mathcal{D} (n; \omega)$:
\begin{equation}
\begin{array}{c}
g_4 ({\bf S} z) =(\mathcal{G}_{{\bf A}} \star h^{\omega}) ({\bf S} z) = \int_{\mathbb{R}^{2n}} \mathcal{G}_{{\bf A}} ({\bf S} z- z^{\prime}) h^{\omega} (z^{\prime} )d z^{\prime} =\\
\\
=\int_{\mathbb{R}^{2n}} \mathcal{G}_{{\bf A}} ({\bf S}( z- z^{\prime \prime })) h^{\omega} ({\bf S} z^{\prime \prime}) d z^{\prime \prime} =\\
\\
=\int_{\mathbb{R}^{2n}} \mathcal{G}_{{\bf S}^{-1}{\bf A}({\bf S}^{-1})^T} ( z- z^{\prime \prime}) h^{\sigma}(z^{\prime \prime}) d z^{\prime \prime}.
\end{array}
\label{eqSets25}
\end{equation}
Since ${\bf S}^{-1}{\bf A}({\bf S}^{-1})^T + i \alpha_0 {\bf J} = {\bf S}^{-1} \left({\bf A} + i\alpha_0 {\bf \Omega} \right) ({\bf S}^{-1})^T$, from (\ref{eqSets19}) this is not a positive matrix. Consequently, the Gaussian $\mathcal{G}_{{\bf S}^{-1}{\bf A}({\bf S}^{-1})^T} ( z)$ is not a $\sigma$-Wigner function with Planck constant $\alpha_0$. It is a well known fact that, under these circumstances \cite{Janssen1,Janssen2}, the function $h^{\sigma} \notin \mathcal{L}$, which is a Wigner function with Planck constant $\alpha_0$ can always be chosen so that $(\mathcal{G}_{{\bf A}} \star h^{\omega}) ({\bf S} z) = (\mathcal{G}_{{\bf S}^{-1}{\bf A}({\bf S}^{-1})^T} \star h^{\sigma}) (z) \notin \mathcal{L}$.
We thus have for the function
\begin{equation}
g_4 (z) = (\mathcal{G}_{{\bf A}} \star h^{\omega} ) (z)
\label{eqSets26}
\end{equation}
that $g_4 \notin \mathcal{L}$.
Next, notice that ${\bf S}^{-1}{\bf A}({\bf S}^{-1})^T + i (\alpha_0-1) {\bf J} \ge 0$, and so $(\alpha_0-1, {\bf J}) \in \mathcal{W} (\mathcal{F}\mathcal{G}_{{\bf S}^{-1}{\bf A}({\bf S}^{-1})^T})$. Since $(1-\alpha_0, {\bf J}) \oplus (\alpha_0, {\bf J}) = (1, {\bf J})$, we conclude from (\ref{eqSets25}) and Theorem \ref{TheoremQuantumConditions5} that $g_4 ({\bf S} z)$ is of the $(1, {\bf J})$-positive type. And thus $g_4 (z) \in \mathcal{F}^{\omega}$.
It remains to prove that $g_4 (z) \in \mathcal{F}^{\sigma}$. Since $(\gamma_0^{\frac{1}{2n}}, {\bf \Sigma_2}) \oplus (\alpha_0, {\bf \Omega}) = (1, {\bf J})$, the result follows again from Theorem \ref{TheoremQuantumConditions5}.
\end{proof}
We conclude this subsection by applying the previous techniques to the following representability problem. This completes the discussion initiated in \cite{Dias5} and addressed also in Theorem \ref{TheoremSymplecticCovariance}.
\begin{theorem}\label{TheoremRepresentabilityProblem}
Let $\omega_1 (z,z^{\prime})= z \cdot {\bf \Omega_1^{-1}} z^{\prime}$ be a symplectic form and suppose that ${\bf M} \in Gl(2n; \mathbb{R})$ is not $\omega_1$-symplectic nor $\omega_1$-anti-symplectic. Define $\omega_2 (z,z^{\prime})= z \cdot {\bf \Omega_2^{-1}} z^{\prime}$ where ${\bf \Omega_2}=\frac{1}{\alpha} {\bf M \Omega_1 M^T}$, with $\alpha = | \det {\bf M}|^{\frac{1}{n}}$. Consider again the linear operator $\mathcal{U}_{{\bf M}}$ as defined in (\ref{eqIntroduction3}). Then there exist $W^{\omega_1} \rho \in \mathcal{F}^{\omega_1} $ such that $\mathcal{U}_{{\bf M}}(W^{\omega_1} \rho) \in \mathcal{F}^{\omega_1} \cap \mathcal{F}^{\omega_2}$ and $W^{\omega_1} \rho^{\prime} \in \mathcal{F}^{\omega_1} $ such that $\mathcal{U}_{{\bf M}}(W^{\omega_1} \rho^{\prime}) \in \mathcal{F}^{\omega_2} \backslash \mathcal{F}^{\omega_1}$.
\end{theorem}
\begin{proof}
We follow the strategy used in the proof of the previous theorem for the cases $\mathcal{A}_5$ and $\mathcal{A}_7$. We just have to be cautious because we do not have necessarily $| \det {\bf M}|=1$.
So let $W^{\omega_1} \rho= \mathcal{G}_{{\bf A}}$ be a Gaussian (\ref{eqGaussian1}) with covariance matrix ${\bf A}$ such that
\begin{equation}
{\bf A} + \frac{i}{2}{\bf \Omega_1} \ge 0.
\label{eqTheoremRepresentabilityProblem1}
\end{equation}
Since
\begin{equation}
\mathcal{U}_{{\bf M}}( \mathcal{G}_{{\bf A}})= \mathcal{G}_{{\bf M^{-1}}{\bf A}({\bf M^{-1}})^T},
\label{eqTheoremRepresentabilityProblem2}
\end{equation}
we conclude that $\mathcal{U}_{{\bf M}}(W^{\omega_1} \rho) \in \mathcal{F}^{\omega_1}$ if and only if
\begin{equation}
{\bf M^{-1}}{\bf A}({\bf M^{-1}})^T + \frac{i}{2}{\bf \Omega_1} \ge 0 \Leftrightarrow {\bf A} + \frac{i \alpha}{2}{\bf \Omega_2} \ge 0.
\label{eqTheoremRepresentabilityProblem3}
\end{equation}
Since ${\bf M}$ is not a $\omega_1$-symplectic or a $\Omega_2$-antisymplectic matrix, then $\omega_2 \ne \pm \omega_1$. The matrix ${\bf A}$ satisfies (\ref{eqTheoremRepresentabilityProblem1}) and (\ref{eqTheoremRepresentabilityProblem3}) if and only if $\lambda_{\omega_1,1}({\bf A}) \ge \frac{1}{2}$ and $\lambda_{\omega_2,1}({\bf A}) \ge \frac{\alpha}{2}$. From the case $\mathcal{A}_7$ of the previous theorem we already know how to construct such a matrix after a possible rescaling.
Regarding the function $W^{\omega_1} \rho^{\prime}$ we assume that it is again a Gaussian $\mathcal{G}_{{\bf B}}$ with covariance matrix ${\bf B}$. It is a $\omega_1$-Wigner function if and only if
\begin{equation}
{\bf B} + \frac{i}{2}{\bf \Omega_1} \ge 0,
\label{eqTheoremRepresentabilityProblem4}
\end{equation}
and $\mathcal{U}_{{\bf M}}(W^{\omega_1} \rho^{\prime}) \in \mathcal{F}^{\omega_2} \backslash \mathcal{F}^{\omega_1}$ provided
\begin{equation}
{\bf B} + \frac{i \alpha}{2}{\bf \Omega_2} \ngeqslant 0.
\label{eqTheoremRepresentabilityProblem5}
\end{equation}
Again, this happens if and only if $\lambda_{\omega_1,1}({\bf B}) \ge \frac{1}{2}$ and $\lambda_{\omega_2,1}({\bf B}) < \frac{\alpha}{2}$. From the case $\mathcal{A}_5$ of the previous theorem, we know that such a matrix exists.
\end{proof}
\section*{Acknowledgements}
We would like to thank Catarina Bastos for providing Figure 1. The
work of N.C. Dias and J.N. Prata is supported by the COST Action
1405 and by the Portuguese Science Foundation (FCT) under the
grant PTDC/MAT-CAL/4334/2014.
|
\section{Introduction}
The oxidation state of an atom describes the number of its valence electrons in a very simple picture. Within this framework, redox reactions are understood as changes of the oxidation state by integer numbers (of charges). In metal-organic complexes, such reactions typically involve a change in the number of bonds to neighbouring atoms. They are particularly important in nature, where they occur, e.g., in the oxygen metabolism by O$_2$ coordination to haemoglobin. Furthermore, they are important in the perspective of spintronics, because an oxidation reaction also involves a change of the spin state.
Surface-anchored metalloporphyrins are model systems, where the interplay between ligand geometry, oxidation state, magnetism and (spin) transport has been studied in great detail down to the single molecule level \cite{Auwarter2015,Gottfried2015,Wende2007,Iacovita2008,Tsukahara2009,Wackerlin2010,Schmaus2011,Wackerlin2012,Hermanns2012,Jacobson2017}. Scanning tunneling microscopy and spectroscopy (STM/STS) has been used to show that additional ligands like H, CO or Cl can induce changes in the oxidation and spin state of the core metal ion~\cite{Gopakumar2012JACS,Strozecka2012,Liu2013,Heinrich2015Ani}. Most STM experiments assume that the tip is a non-invasive tool in the investigation. However, the STM tip can perturb the molecular properties when it is brought in sufficiently close proximity to the molecule. As such, the coupling of a molecule to the surface can be modified, which can lead to the appearance/disappearance of a Kondo resonance in case of magnetic molecules~\cite{Hiraoka2017,Abufager2017}. Similar changes had previously been only detected in two-terminal break junction experiments, where the stress upon opening/closing the junction is assumed to be very large~\cite{Parks2010}. Recently, it has been shown that far before contact formation, an STM tip may already affect intramolecular properties. In the far tunneling regime, the forces exerted by the STM tip are already sufficiently large to induce small conformational changes within the molecule, which affect the crystal field splitting of the central ion, and, hence, its magnetic anisotropy~\cite{Heinrich2015Ani}.
Here, we show that the tunneling regime bears even more degrees of freedom, which can be reversibly controlled by the STM tip.
For this purpose, we adsorb iron(III) octaethylporphyrin chloride (Fe-OEP-Cl) on a Pb(111) surface.
In the absence of the tip, the d$^5$ configuration of Fe$^\mathrm{III}$ gives rise to an $S=5/2$ spin state in the ligand field.
We make use of the possibility to controllably and reversibly manipulate the Fe--Cl bond by the force field of an STM tip. We show that the forces do not only lead to a change in the magnetic anisotropy as was reported previously\cite{Heinrich2013}, but that they can also be used to induce a reversible redox reaction, implying the reversible change from a half-integer to an integer spin state. Both states are not affected by sizable scattering from substrate electrons, i.e., their spin state is not quenched by the Kondo effect. In combination with the use of superconducting electrodes, which increase the excited state spin lifetime compared to metal electrodes\cite{Heinrich2013}, this renders this system suitable for potential spintronic applications.
The experiments are supported by theoretical simulations. By means of Density Functional Theory (DFT) and Wavefunction Theory (WFT) calculations, we determine the relative energies of different spin manifolds of Fe-OEP-Cl as a function of the Fe--Cl distance. Our simulations reproduce the experimentally observed changes in the magnetocrystalline anisotropy and indicate that at large Fe--Cl distances, the integer $S=2$ state is favored.
\section{Results and Discussion}
\begin{figure}[t]
\includegraphics [width=0.7\textwidth,clip=]{Fig1_dIdV_z_sweep}
\caption{Fe-OEP-Cl on Pb(111). (a) Topography of a monolayer island. Two types of molecules are observed: Fe-OEP (dim center) and Fe-OEP-Cl (bright center). (b) Conductance at 10\,mV vs. tip-sample distance $\Delta z$ measured with the tip place above the center of a Fe-OEP-Cl molecule. A sudden drop in the conductance is observed at $\Delta z \approx -130$\,pm. (c) $dI/dV$ spectra at different tip-sample distances $\Delta z$ show signatures of inelastic excitations at different energies.}
\label{FigIz}
\end{figure}
We deposited Fe-OEP-Cl molecules on an atomically clean Pb(111) surface held at $\approx 240\,$K. This leads to self-assembled mixed islands of iron-octaethylporphyrin-chloride (Fe-OEP-Cl) and iron-octaethylporphyrin (Fe-OEP) as can be seen in the STM images obtained at low temperature (Fig.~\ref{FigIz}a). The individual molecules can be identified by the eight ethyl legs. The chlorinated species appears with a bright protrusion in the center, whereas the dechlorinated species appears with an apparent depression in the center. The dechlorination process is probably catalytically activated by the surface and occurs upon adsorption.
The magnetic excitation spectra of both species have been characterized previously\cite{Heinrich2013,Heinrich2015Ani}.
As mentioned above, in Fe-OEP-Cl the Fe center lies in a $+3$ oxidation state with a spin of $S=5/2$. Dechlorination changes the oxidation state to $+2$ with the resulting spin state being $S=1$.
This change of oxidation and spin state is naturally expected when the Fe center is subject to such a drastic modification as removing one ligand. It was also observed that the STM tip can have an influence on both individual species. The Fe ions are subject to crystal fields, which split the $d$ levels and induce a sizable magnetocrystalline anisotropy and, hence, a zero field splitting (ZFS). By approaching the tip to the molecule, for both species the ZFS has been shown to be modified by the local potential of the tip\cite{Heinrich2015Ani}.
Here, we explore if the tip potential can be used to change the local energy landscape even further such that a reversible change not only of the ZFS, but also of the spin state takes place. For this, we target at the Fe-OEP-Cl molecule because it bears more flexibility with its additional ligand than the dechlorinated species. We first explore the change of junction conductance upon approaching the tip on top of the Fe-OEP-Cl molecule in Figure~\ref{FigIz}b. After a first exponential increase of the conductance with reduced distance ($\Delta z<0$), which is characteristic for the tunneling regime (regime I), there is a reversible drop in the conductance in the order of 30\% of the total signal. This can be linked to a sudden change of the junction geometry and/or the electronic structure of the complex. Further decreasing the distance increases the conductance again, yet with a smaller and decreasing slope (regime II). At $\Delta z\approx-330$\,pm, the conductance increases again with increasing slope.
The different slopes indicate different structural/ electronic configurations of the molecule in the junction. To characterize the magnetic properties of the molecule in these regimes, we record d$I/$d$V$\xspace\ spectra at different $\Delta z$ (Fig.~\ref{FigIz}c). In all spectra, we observe an excitation gap with a pair of quasiparticle excitation peaks at $\pm (\Delta_{tip}+\Delta_{sample})$, with $\Delta_{tip}$ and $\Delta_{sample}$ being the real part of the order parameter of the superconducting tip and sample, respectively.
In regime I ($\Delta z=-90$\,pm), there are two additional pairs of resonances outside the gap. These resonances are due to inelastic spin excitations on a superconductor \cite{Heinrich2013,Fransson2015}. Replicas of the quasiparticle resonances appear at the threshold energies for the opening of an inelastic tunneling channel. The first inelastic peak corresponds to a transition between the
$\mathrm{M}_{\mathrm{s}}={\pm{1}/{2}}$ to the $\mathrm{M}_{\mathrm{s}}={\pm{3}/{2}}$ state within the $S=5/2$ manifold. The second excitation from the $\mathrm{M}_{\mathrm{s}}={\pm{3}/{2}}$ state to the $\mathrm{M}_{\mathrm{s}}={\pm{5}/{2}}$ state becomes possible at large current densities due to spin pumping~\cite{Heinrich2013}.
In regime II, i.e., after the sudden change in the molecular conductance, both inelastic excitations vanish. Instead, an excitation of lower energy is observed close to the superconducting gap edge (see, e.g., the spectrum at $\Delta z=-160$\,pm). At even smaller tip--molecule distance (exemplary spectrum at $\Delta z=-320$\,pm), we observe an inelastic excitation of even lower energy, which merges with the coherence peak.
\begin{figure} [t]
\includegraphics [width=0.7\textwidth,clip=]{Fig2_2DdIdV}
\caption{(a) Pseudo-2D color plot of $dI/dV$ spectra of the molecule in Fig.\,\ref{FigIz}b and \ref{FigIz}c as a function of $\Delta z$. (b) Pseudo-2D color plot of $dI/dV$ spectra of a different molecule with a different tip. $\Delta z$ increment: 5\,pm. The spectrum at $\Delta z=-280$\,pm is shown on top and evidences a low-energy excitation close to the superconducting gap edge. For additional data on different molecules and with different tips see SI. }
\label{Fig2Dplot}
\end{figure}
To track the evolution of the inelastic excitations we record a series of spectra while approaching the STM tip in small intervals. A 2D color plot of these spectra is shown in Fig.~\ref{Fig2Dplot}a. Within regime I, we observe a shift of both excitations to larger energies upon tip approach as a result of an increased ZFS. This has been qualitatively explained as a response to a distortion of the ligand field~\cite{Heinrich2015Ani}. The sudden drop in conductance at the transition to regime II is also reflected in an abrupt change of the d$I/$d$V$\xspace spectra. The two excitations of the $S=5/2$ manifold disappear and a new excitation appears. It shifts to lower energies and merges with the quasiparticle excitation peaks when approaching the STM tip to $\Delta z\approx -200$\,pm. Eventually, a faint excitation emerges from the quasiparticle excitation resonances for $\Delta z< -220$\,pm and a second excitation departs from the gap edge at $\Delta z\approx -290$\,pm. A high-resolution 2D plot of regime II (for a different molecule) is shown in Fig.~\ref{Fig2Dplot}b. It resolves the latter two resonances more clearly (note that their appearance is shifted in $\Delta z$ with respect to the molecule shown in Fig.~\ref{Fig2Dplot}a).
\begin{figure}[t]
\includegraphics [width=0.7\textwidth,clip=]{Fig3_dIdV_Zeeman}
\caption{(a) $B$-field dependent $dI/dV$ spectra at $\Delta z=-150\,$pm. A shift of the excitation step to higher energies is observed with increasing $B$ field. (b) Step energies as a function of $B$-field (black squares).
The purple asterisk indicates the energy as extracted from the spectrum recorded with tip and sample in the superconducting state. The excitation shifts to higher energies with increasing $B$ field.}
\label{FigZeeman}
\end{figure}
Whereas the origin of the two excitations in regime I is already understood, regime II was not explored in earlier experiments. We note that regime II does not indicate the desorption of the Cl ligand, because Fe-OEP shows characteristic spin excitations at larger energies \cite{Heinrich2015Ani}, which are absent here (see Fig.\,S2 of the SI). Furthermore, the junctions are stable: when we retract the tip, we reversibly enter into regime I. Indeed, we can precisely and reversibly tune the junction conductance and, hence, the energy of the excitations. We first determine whether the new excitations are of magnetic origin. By applying an external magnetic field perpendicular to the sample surface, superconductivity in substrate and tip are quenched. Here, inelastic excitations give rise to steps in the d$I/$d$V$\xspace spectra at the excitation threshold as expected for a normal metal substrate (Fig.~\ref{FigZeeman}a). Importantly, we observe a shift of the steps to larger energies in response to an increasing external field. The extracted step energies agree with a Zeeman shift (Fig.~\ref{FigZeeman}b), which evidences their magnetic origin.
\begin{figure}[t]
\includegraphics [width=0.7\textwidth,clip=]{Fig4_theory_2}
\caption{(a) Potential energies of different spin manifolds as a function of Fe--Cl distance, determined
by a combination of DFT optimization and WFT single-point calculations (see text). Half-integer and integer $S$ values
refer to the neutral (Fe$^{\mathrm{III}}$) and
anionic (Fe$^{\mathrm{II}}$) complexes, respectively.
(b) ZFS of the $S=2$ manifold. The colors indicate contributions of the eigenstates of the spin projection operator. We calculated the average ($\bar{M}_s$) as the sum of the absolute eigenvalues multiplied by the weights of contributing eigenstates. Note that at 210\,pm the lowest two eigenstates are almost degenerate, while for Fe--Cl distances larger than 285\,pm, the two highest states are almost degenerate.
}
\label{Fig_theory}
\end{figure}
Summarizing the experimental results, there is evidence for a sharp transition from the $S=5/2$ state to a different spin state when approaching the STM tip on top of Fe-OEP-Cl. In principle, this could either be a spin crossover to another half-integer spin state or a change to an integer spin state, which necessarily needs to be accompanied by a change in oxidation state. In order to gain further insights, we performed DFT calculations using the B3LYP exchange-correlation functional\cite{Becke1993} as well as WFT based simulations in the form of the so-called CASSCF/NEVPT2 method\cite{Malmqvist1989,Angeli2001,Angeli2001a,Angeli2002,Angeli2006} according to Refs.\cite{Atanasov2011,Stavretis2015}. Details are described in the Methods section and in the SI.
We assume that the tip-molecule interaction,
in the sense of a Lennard-Jones potential, is attractive at not too short tip-molecule distances.
Upon approaching with the STM tip, the Cl ion will thus be attracted by the tip. We model the effect of such a potential in an isolated Fe-porphin molecule by increasing the Fe--Cl distance. In porphin, the eight ethyl groups are replaced by H atoms. We have previously checked that the ethyl legs do not affect the electronic and magnetic properties of the complex (see Fig. S6 in the SI).
We used DFT optimization to compute the relaxed structure while scanning and keeping the Fe--Cl distance fixed in three different spin states of the neutral Fe(III)-complex, i.e., with spin quantum numbers $S=5/2$; $S=3/2$; $S=1/2$. Based on these results, we calculated the potential energies by the CASSCF/NEVPT2 method at the DFT geometries (Fig.~\ref{Fig_theory}a; corresponding {\it pure} DFT potential curves are shown in Fig. S8 of the SI). In agreement with previous experiments and calculations, we find that the ground state exhibits a spin state of $S=5/2$. Here, the Fe--Cl distance amounts to \mbox{220\,pm}. This configuration represents the unperturbed molecule on the surface, i.e., when the tip is far away.
While the energy difference between the $S=5/2$ and $S=3/2$ states decreases, the latter remains at larger energies throughout the whole range of reasonable Fe--Cl distances. Hence, a spin crossover scenario as an explanation for the sudden change in experimental spin excitations from regime I to regime II can be excluded\footnote{We can also experimentally exclude such a spin crossover, because the appearance of two excitations in regime II cannot be captured in a $S=3/2$ system.}. The $S=1/2$ spin state lies even higher in energy and can therefore be excluded to play a role as well.
Hence, we also calculated the potential energies of the anionic Fe(II)-complex in its $S=2$ and $S=1$ spin states\footnote{The $S=0$ state can be disregarded as it would not show any spin excitations.}. The $S=2$ state is found at significantly lower energies than the $S=1$ state. Within increasing Fe--Cl bond length, the potential energy increases but not with the same slope as in the $S=5/2$ state. As a result, the
energy difference between $S=5/2$ and $S=2$, i.e., the electron affinity
of the neutral complex {\it increases} with increasing Fe--Cl distance.
Together with a resulting image charge stabilization of the anionic complex and a
relatively low work function of Pb of about 4\,eV, this is the main driving
force for an electron transfer from the tip or substrate to the molecule
at a certain critical, elongated Fe--Cl distance.
To check whether the experimentally observed spin excitations in regime II match with this scenario, we calculated the ZFS of the anionic system. The results are shown in Fig. \ref{Fig_theory}b. Overall it is observed that the state energies are very sensitive with respect to geometrical variation. The discussion of this behavior can be split in two parts: a first part (i) with Fe--Cl bond length shorter than \mbox{285\,pm} and a second part (ii) with Fe--Cl bond lengths longer that \mbox{285\,pm}.
In part (i) the ground state is composed of two states, which are almost degenerated. The eigenvalues of the spin projection operator on the $z$ axis have mainly $M_s=\pm 2$ character. The third and fourth state have eigenvalues of $M_s = \pm 1$, while the fifth state has the $M_s = 0$ eigenvalue. The splitting can be easily rationalized with the
help of an effective Hamiltonian method \cite{Ganyushin2006} and related to
so-called axial and rhombic ZFS parameters $D$ and $E$, respectively, as explained in the SI.
With increasing bond length the third, fourth and fifth state decrease in energy and mixing occurs between the states.
Inelastic electron tunneling requires $\Delta M_s=0,\pm 1$. Hence, one can observe excitations from the $M_s=\pm 2$ ground states to the $M_s = \pm 1$ states. The energy of this excitation is in the order of 1\,meV with decreasing energy until the bond length amounts to \mbox{285\,pm}. This behavior of one excitation with decreasing energy is found in experiment at the initial stages of regime II, i.e., right after the spin transition.
When the Fe--Cl bond length becomes longer than \mbox{285\,pm} [part (ii)], the lowest two states are again almost degenerate, but have changed their character. Both states are now mixtures of several eigenvectors of the spin projection operator. The first one is dominated by $M_s = 0$, the second by $M_s = \pm 1$. The third state is mainly of $M_s = \pm 1$ character, while the fourth and fifth are dominated by $M_s = \pm 2 $. The energy of the three highest states increase now monotonously with elongated Fe--Cl bond lengths. Since the nearly degenerate ground state has $M_s = \pm 1$ and $M_s = $0 character, inelastic excitations are allowed to both the $M_s = \pm 1$ and $M_s = \pm 2 $. These two excitations are indeed found in experiment in the later part of regime II. In agreement with theory, their energies increase with increased Fe--Cl bond length. The lower intensity of the second excitation in experiment may be ascribed to an underestimated energy splitting of the $M_s = \pm 1$ and $M_s = $0 state in the calculations. A finite splitting would require the second state to be thermally occupied in order to be able to observe an inelastic transition and thus be reduced in intensity.
In conclusion, we have established a fully reversible switching of the spin and oxidation state in a single molecule including the possibility to fine-tune the magnetocrystalline anisotropy. By approaching the STM tip on a Fe-OEP-Cl molecule, we have observed drastic changes in the spin excitation spectra. The force exerted by the STM tip leads to a deformation of the molecule, in particular to an elongation of the Fe--Cl bond length. Initially, this affects the magnetocrystalline anisotropy of the $S=5/2$ state. At a critical bond length, a sudden change of oxidation and spin state occurs, which leads to a distinctly different fingerprint in the excitation spectra. Our combined DFT and CASSCF/NEVPT2 simulations allow for an unambiguous identification of the experimentally observed excitations. Their behavior under the force field agrees well with the experiment. Our experiments provide an intriguing method for a reversible and controlled way of changing not only the magnetic anisotropy, but also and more drastically, the spin state. They open new pathways for controlling magnetic properties by external forces.
\section{Methods}
All experiments were performed in a SPECS JT-STM, a commercial low-temperature scanning tunneling microscope (STM) working under UHV conditions at a base temperature of 1.2\,K. We obtained clean and superconducting surfaces of the Pb(111) single crystal by cycles of Ne$^+$ ion sputtering and subsequent annealing to $\approx 430\,$K.
In order to circumvent the Fermi-Dirac limit in energy resolution, superconducting, Pb-covered W tips were used, which allow us to improve the energy resolution down to $\approx 60\,\mu$V when employing elaborated RF-filtering and grounding schemes~\cite{Ruby2015Pb}.
All spectra are recorded in constant-height mode with an offset $\Delta z$ (as indicated in the figures) in the tip--sample distance relative to the set point: $U=50\,$mV; $I=200$\,pA.
The differential conductance $dI/dV$ is measured
using standard lock-in technique with a modulation frequency $f=912$\,Hz and a voltage modulation $U_{mod}=35\,\mu$V$_{rms}$.
We used the ORCA (version 3.0.3) program package \cite{Neese2011} to optimize geometries and to calculate potential energy surface scans along the Fe--Cl distance for several
spin multiplicities. The B3LYP density functional \cite{Becke1993} in combination with the def2-TZVP basis set \cite{Weigend2005} was utilized.
For the geometries along the potential energy scans, the Zero-Field parameters have been computed by using a
correlated WFT methodology\cite{Atanasov2015,Ganyushin2006} based on the complete active space self-consistent field method (CASSCF)\cite{Malmqvist1989}. In order to take into account dynamical correlation, the NEVPT2 \cite{Angeli2001,Angeli2001a,Angeli2002,Angeli2006} method has been used on top of CASSCF.
Subsequently, the effects of spin-orbit and spin-spin interactions were included within the framework of quasi-degenerate perturbation theory. To accomplish this, the MRCI module of the ORCA program has been used, together with an effective one-electron mean field spin-orbit coupling operator \cite{SOMF}.
Further details are given in the SI.
\section{Acknowledgements}
We thank J.\ I.\ Pascual, who was involved in the first experiments on Fe-OEP-Cl. We acknowledge discussions with C.\ Herrmann. We gratefully acknowledge funding by the Deutsche Forschungsgemeinschaft through Sfb 658 and HE7368/2, as well as the European Research Council through the consolidator grant "NanoSpin".
|
\section{Introduction}\label{sec:Introduction}
The success of physics programs at CERN would be impossible without reliable
data management systems. During the Run 1 and Run 2 data taking periods
four experiments at the Large
Hadron Collider (LHC) in Geneva, Switzerland, demonstrated the ability to handle petabytes of data on a regular
basis. In CMS, we rely on a tiered computing infrastructure.
The data collected with the detector are streamed to the high level trigger (HLT)
farm and organized into trigger streams. Later they are archived at the Tier-0
center at CERN and distributed to CMS Analysis facilities at CERN
and Tier-1 centers around the globe. Many Tier-2 centers worldwide
share a portion of the data for further processing and Monte Carlo (MC) generation.
Finally, the Tier-3 centers (mostly at Universities) are used for various
analysis tasks. More details about CMS Computing Model can be found
elsewhere \cite{CMSComp}.
To accomplish various processing tasks, both real and MC data
are handled by a distributed set of agents at various computing centers.
The chain of high energy physics (HEP) workflows
is performed within an individual agent.
For instance, a job can run an MC
simulation where the CMS software (CMSSW) framework generates new data in a
distributed fashion, collects various pieces of data frame the workflow pipeline, and publishes
results in the central Data Bookkeeping System (DBS). At every step, the associated
metadata information are collected in ``framework job report'' (FWJR)
documents. This information can be further analyzed and used to
monitor the entire process both in terms of various metrics, e.g. successfull job throughput,
or capture various errors in different steps of the workflow pipeline. In addition, the information
can be used to plan future jobs and better manage the resources.
Initially, ad-hoc solutions were used by various data-operations teams, but
it soon became clear it was necessary to unify the various solutions under a single system.
Here, we present the Worklfow Management Archive system (WMArchive), which was
designed to provide a long term solution for FWJR document storage along with
flexible queries and visualization to help data ops in their
daily operations.
This paper is organized as follows. In Sec. \ref{sec:Architecture}
we present the overall architecture of the system starting by
describing the CMS workflow management system (Sec.~\ref{sec:WMAgents}),
followed by discussion of the requirements (Sec.~\ref{sec:Requirements}).
We then present the system components, storage layers, and
interfaces in Secs.~\ref{sec:Components}, \ref{sec:Storage}, and \ref{sec:Interfaces},
respectively. In Sec.~\ref{sec:Benchmarks} we show various benchmarks
and discuss query language used in the WMArchive in Sec.~\ref{sec:QL}. We finish our
discussion with current use of the system in Sec.~\ref{sec:Monitoring}
and provide an overview of the monitoring tools based on stored metadata.
\section{WMArchive architecture}\label{sec:Architecture}
The WMArchive system architecture is quite complex.
Here, we start our discussion with
overview of the CMS workflow management system followed by
set of requirements imposed by our data operations teams.
\subsection{CMS Workflow Management System}\label{sec:WMAgents}
The CMS Workflow Management System~\cite{WMAgent} contains several applications to process the
workflows defined by physics groups. The overview of the procedure is as
follows. When a workflow is fed into the system (Request Manager), it divides
the work into smaller units which are placed into a queue (WorkQueue). Individual
Workflow Manager Agents (WMAgents), distributed among computing centers,
pull down work from the queue depending on resource availability and other
various conditions, e.g. priority of the work, etc. Each WMAgent is responsible
for splitting the work into smaller chunks (jobs) and sending them to
the CMS Global Pool~\cite{GlobalPool}, an HTCondor~\cite{HTCondor} batch system overlayed,
using GlideinWMS~\cite{GlideinWMS}, on the CMS-accessible grid resources.
The batch system then distributes the jobs to
computing resources all over the world---the tiered system mentioned above. Each
WMAgent is also responsible for tracking and accounting for the jobs and
publishing the results in other applications in CMS (DBS, PhEDEx, WMArchive).
We have about a dozen WMAgents running concurrently.
When a job has finished
sucessfully or failed, WMAgent will generate an FWJR which
contains job specific information, such as the input/out datasets and files,
CPU and memory usage, log file and its location, and so on.
The current set of WMAgents can produce on average 500K FWJRs per day but could not keep
the documents in the system indefinitely. FWJR reports are published and stored
in WMArchive for further analysis and as a solution for long term storage
needs.
\subsection{Requirements}\label{sec:Requirements}
In CMS we rely on the distributed nature of WMAgents. At the time of initial design of the system, we assumed
the WMAgents would generate 200--300\,000 FWJR documents per day, with a document size of
$\mathcal{O}$(10~KB) which would yield 3~GB total/day or about 1--2~TB/year of
metadata information.
Each FWJR document is in a JSON data format whose structure is
specific to the job. For example, the FWJR of a successful job will be different
from a failed job. Here, is a list of requirements we gathered from CMS data
management and workflow management teams:
\begin{itemize}
\item store FWJR for the lifetime of the experiment in persistent storage for
data lookup anytime; losing data is not acceptable;
\item provide data-ops quick and flexible queries to retrieve the log files of
any particular job for debugging;
\item the statistics to be collected include, but are not limited to, CMSSW version,
CPU/wallclock and memory information, log files and their locations, input and
output datasets, number of files and events, job error codes, etc.
\item the system should be able to use a flexible query language over
unstructured data; there is no requirement to support live queries;
\item provide tools to query, aggregate, and analyze CMS jobs and to plan for
future processing and resource management;
\item provide a web interface to monitor job processing trends, gather statistics,
and aggregate information across multiple dimensions;
\item provide a web interface to hourly update job stats for debuggging and error
reporting of running jobs.
\item support a flexible schema to extend, add, or remove fields from the FWJR if necessary;
\item have a minimal impact on existing CMS infrastructure, in particular on production WMAgents;
\end{itemize}
\subsection{System architecture}\label{sec:Components}
WMArchive system is based on a RESTful server and series of tools to handle
various data management tasks. The overall architecture of the system is shown
in Fig. \ref{fig:Architecture}. The data are pushed to WMArchive by the
distributed set of WMAgents via REST APIs. All WMAgents are authenticated via
CMSWeb by providing proxy GRID certificates at front-end servers. The
pushed data are submitted as a bulk collection in certain time intervals defined
by WMAgent configuration.\footnote{Right now every WMAgent sends data every 5 minutes, 1000
docs per polling cycle with rate of 200 docs per single HTTP request.}
The received data are routed to internal
short-term storage (STS) based on the document-oriented MongoDB database~\cite{MongoDB}.
Then, a separate daemon reads data from STS, merges records
together, converts the JSON data to Avro~\cite{Avro} data format, and writes Avro files
to the local file system. Finally, we rely on a migration script to push data from the
local file system into long-term storage (LTS) based on HDFS. Once data are
written to LTS we use an aggregation script to collect hourly and daily
statistics and push them into STS and the CERN MONIT systems.
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{WMArchiveArchitecture.pdf}
\caption{WMArchive architecture. MongoDB represents short-term storage (STS)
and HDFS refers to long-term storage (LTS) discussed in a text.}
\label{fig:Architecture}
\end{figure}
\end{center}
\subsection{Storage layers}\label{sec:Storage}
Based on the system requirements (see discussion in Sec.~\ref{sec:Requirements}), we
decided to develop a set of abstract interfaces to provide access to available storage solutions.
Later, we provided a common implementation of BaseIO, FileIO, MongoIO, and
HdfsIO interfaces, which were used for rapid prototyping of APIs and
understanding interprocess communications. After several trials (see
Sec.~\ref{sec:Benchmarks}), we decided on the short and long term storage
solution described above. The former is responsible for keeping
up with the injection rate while the latter used for permanent storage of FWJR
documents. In other words, we use STS as a buffer to keep up with our agents'
load, while LTS serves as persistent storage.
Among the various candidates, we decided to use MongoDB with WiredTiger
document-level concurrency control as our STS solution. We estimated that our
injection rate should stay within 1 KHz and tested that MongoDB will sustain
such load. While we did not impose any constraint on the incoming document
structure, we complemented each document with auxiliary information such as
WMArchive unique id and storage type (e.g. \emph{fileio}, \emph{mongoio}, \emph{Avroio})
attributes. The latter was used by migration and clean-up scripts to
dynamically handle the size of STS within budget constraints.
For the LTS, we decided to use HDFS provided by the CERN IT infrastructure;
several issues we faced justifyied this choice. First, our FWJR documents were
quite small (about 10~KB) to be stored individually on HDFS.
Therefore, we accumulated our data into larger groups before storing them to HDFS.
However, we also needed to satisfy another requirement (see Sec.~\ref{sec:Requirements}),
and provide flexible queries over a large portion of stored data. As a result, we
decided to convert our data into the Avro~\cite{Avro} data format, which was a more
natural fit for both tasks. It guarantees data consistency within
a single file and optimization for the HDFS eco-system such as efficient use of
Spark jobs to process large amounts of data.
We complemented the WMArchive service with three cron jobs, one, to continuously
accumulate data from STS to Avro files on local filesystem, another to move
Avro files from staging area to HDFS, and a third to clean-up STS. The first
was designed to collect enough documents to fit into the HDFS block boundaries.
The CERN HDFS system is using 256~MB block boundaries and was adjusted
by CERN IT to the needs of experiments based on different use-cases. This constrain
was added as one of the configuration parameters of WMArchive system.
Within this boundary we collected roughly 50\,000 documents
into a single Avro file.
The actual number of documents in Avro file varied based on a file
size growth up-to a given threshold (in our case 256~MB) rather than fixed
number of stored documents.
These files were then moved into a staged area where they
were periodically purged once relocated to HDFS by the second cronjob. Afterwards,
the underlying script flipped the document's storage type attribute in STS from
\emph{mongoio} to \emph{Avroio}. These values of the attributes were used by the third cronjob which
performed a cleanup in STS of documents with the \emph{Avroio} storage type value based
on a predefined threshold. We determined this threshold based on the disk capacity
of the production system and the desire to keep documents in STS. In our case, we keep
documents in STS for 1 month after they are moved to LTS.
The data movement was optimized with respect to the accumulation rate of documents
into a single Avro file. Empirically, we found that a ten minute interval is
sufficient to accumulate 50\,000 (256~MB) docs into a single Avro file. Therefore, the
overall latency of data appearance on HDFS was approximately 10--20 minutes.
Minimization of this interval was important for the developement of monitoring
tools based on the aggregated information from LTS (see Sec.~\ref{sec:Monitoring}).
\subsection{Interfaces}\label{sec:Interfaces}
Users communicate with the WMArchive service via REST APIs. The POST APIs are
used both for data injection and placing queries into WMArchive. The latter
is done in real time in STS or via batch submission of Spark jobs on LTS.
To distinguish these use cases, we rely on a document representation rather than
individual APIs. For example, the data injection is done in a bulk using
the following document structure:
\begin{verbatim}
{"data":[list_of_documents]}
\end{verbatim}
while end-user queries are represented as:
\begin{verbatim}
{"spec":{conditions},
"fields":[list_of_attributes_to_retrieve]} .
\end{verbatim}
These requests are acknowledged by the system which provides a response with a
unique token id to look-up their data at a later time, e.g.
\begin{verbatim}
{"job": {"results":{results_of_job},
"wmaid":wma_unique_id}} .
\end{verbatim}
This approach allows us to process user-based queries on the HDFS/Spark cluster
independently from data injection tasks and other user communications with the system.
The results of the Spark jobs are stored back into STS and users are able to look-up
the results later via a GET request based on the provided token.
For example, to fetch the set of data from WMArchive end-users rely on the GET
API, e.g.
\begin{verbatim}/wmarchive/data/wma_unique_id .\end{verbatim}
To minimize the execution time and to avoid a situation when user queries can
span the entire data collection (millions of documents per year), we require each
query to provide a time range attribute which specifies the boundaries of scanned
documents on HDFS. If both ranges of the time range are below a certain threshold
(implied by STS capacity) the query is executed in real time on STS,
otherwise a job submission is executed on LTS. The LTS job, i.e. Spark job on
HDFS, is routed to the CERN analytics cluster. We performed various benchmarks to
determine latencies of queries executed on STS and LTS.
\section{WMArchive Benchmarks}\label{sec:Benchmarks}
To ensure that our python based code scales, we performed various benchmarks
using the STS and LTS back-ends. The benchmarks were designed to test data
injection, parsing, and access patterns on a local node and a Spark cluster
provided by CERN IT. For local tests we used a Linux node with
24~GB of RAM and 12 cores; the CERN Spark cluster consists of 39 nodes with
64~GB of RAM and 32 cores per node.
\subsection{Short-Term Storage benchmark}\label{sec:MongoDBIndex}
CouchDB and MongoDB have been used by the group in different projects.
Initially, CouchDB was chosen to implement the WMArchive DB short-term storage
backend. However, it was not possible to have flexible queries
nor the required volume of data storage. MongoDB was chosen for these reasons:
\begin{itemize}
\item it enables horizontal scaling of data;
\item it provides a flexible query language;
\item it is possible to exchange data beween the current CouchDB documents and MongoDB
(WMAgents use CouchDB as temporate storage for FWJRs).
\end{itemize}
One disadvantage of MongoDB is that the query speed depends on indices which should fit
into the RAM of the node, for details see \cite{MongoDBIndex}.
If not, the lookup time increases significantly due to reading indices from the disk.
We studied our data in MongoDB regarding the data size, index size, insertion speed, and query
response times.
We injected 1\,542\,513 real FWJR documents into MongoDB and created 13 indices on the
database. The total data size in the database was 15.3~GB and the total index size was 3.5~GB.
Based on this profile we estimated that we should be able to hold about 10 M docs with our hardware and still have all the indices be able to fit
into the RAM (24~GB). Found that the time to insert 1000 documents
in bulk is less than 0.5 seconds. In addition, the insertion speed was
not affected by the database size. Uploading data into MongoDB has a mininual
effect on WMAgent performance. Based on our studies, we concluded it was feasible to store one month of data
in MongoDB with the current hardware and still maintain acceptable performance.
\subsection{Data lookup on a single node}
To test data access on a single node we used custom code to translate FWJR documents
from the JSON format into the Avro format. We generated
an Avro file with 50\,000 FWJR documents of identical structure. The total file size
in Avro format was about 190~MB. The bzip'ed Avro file shrank to 26~MB. Even
though we achieved such a large compression level, we decided to proceed with plain
Avro files on HDFS while rolling our system to production. The usage of zipped
Avro files was postponed to a later time to become more familiar with all
aspects of the system in the production environment.
We used Spark scripts to run a simple map-reduce job over the set of Avro
files. For this benchmark we decided to use a rate of 200\,000 documents/day, so we put four
Avro files into a single directory. Then we cloned that directory to simulate
two months of CMS data. In total we collected 12 million records. We ran Spark jobs and
measured the time we spent to find individual records based on provided logical
file name (LFN) patterns. We found that one day of data can be processed in
about a minute, while 2 months of data (12 million records) will require about 1 hour
of processing time.
\subsection{Data lookup on a Spark cluster}
After benchmarking data access patterns on a single node, we moved our tests to
the CERN IT Spark cluster. In addition to the
search time we also measured the overhead between the Java and Python processes'
communication and found it satisfactory for our needs.
The tests were done in two modes:
\begin{itemize}
\item
{\bf yarn-cluster:} all Spark execution actors were included inside the cluster
(driver and executors). This is the default setup provided by CERN IT.
\item
{\bf yarn-client:} the Spark driver was spawned in the local node meanwhile executors
were inside the cluster. This setup was preferred for us since we were able to control
our production node utilization.
\end{itemize}
For both modes we used two months of data and placed a query to find a pattern
in task names across all documents stored on HDFS. The Spark jobs were configured to
use 4 executors, 4~GB of memory, and 4 cores each.
We obtained the following results: search results across one day of data was finished
in $\mathcal{O}(10)$ seconds, one month required $\mathcal{O}(100)$ seconds, and a search
pattern over two months of data fits into $\mathcal{O}(200)$ seconds.
\section{WMArchive Query Language}\label{sec:QL}
Our users, the CMS data-operations teams, require a flexible query language
(QL) to communicate with the WMArchive system. Since we didn't base our system on an RDBMS solution nor
impose live-query requirements, we opted in favor of a JSON based QL
similar to the MongoDB QL \cite{MongoQL}. Due to the nested structure of stored metadata
within FWJR documents we used dot notations, e.g. \emph{output.inputDataset},
to specify query conditions. We also allow users to
use flexible conditions such as $\$gt$, $\$lt$ operators as well as use an alternative set
of conditions via $${"\$or": [JSON1, JSON2]}$$ structure. These choices helped us
to initially place and test queries in STS and later adopt them to LTS via
customized python classes. To clarify QL syntax here is a simple example:
\begin{verbatim}
{"spec": {"task": "/Abc*",
"timerange":[20160801,20160820]},
"fields":["error.exitCode"]} ,
\end{verbatim}
which lists a set of conditions defined by a \emph{spec} dictionary, e.g. task
name pattern along with mandatory timerange constraint, and output
\emph{fields} which end-users requested for their result set, e.g. \emph{error.exitCode}
specify a list of error codes to retrive.
It is worth to mention that usage of non-SQL language (JSON) in queries
was easily adopted. Its rich syntax and expresiveness
covers many use cases and usage of JSON in other data-management
applications made this transition effortless.
\subsection{User queries}
Users place their queries via POST API and fetch the end results via GET API. Here, we provide a
complete example of an end-user interaction with the system via a python client.
A client posts a request which is immediately acknowledged by the system:
\begin{verbatim}
# post request to the server
wma_client.py --spec=query.json
# response from server
{"result": [
{"status": "ok",
"input": {"fields": ["wmaid"],
"spec": {"task": "/Abc*"}},
"storage": "mongodb",
"results": [{"wmaid": "6b0bac"}]}]} .
\end{verbatim}
The user query is supplied via the \emph{query.json} JSON file which has the following
content:
\begin{verbatim}
{"spec":{"task": "/Abc*",
"timerange":[20160801,20160820]},
"fields":["meta_data"]} .
\end{verbatim}
At a later time the results were accessible via GET API,
e.g. GET \emph{/wmarchive/data/6b0bac} yields the following results:
\begin{verbatim}
{"result": [
{"status": "ok",
"input": {"fields": ["meta_data"],
"spec": {"task": "/Abc*",
"timerange":[20160801,20160820]}},
"storage": "mongodb",
"results": [
{"meta_data":
{"agent_ver": "1.0.13.pre8",
"fwjr_id": "100-0",
"host": "vocms008.cern.ch",
"ts": 1454617125}}]}]} .
\end{verbatim}
We didn't provide any specific way to know up-front readiness of the query
and relied on HDFS native solution to look-up job status via its web UI.
\section{Usage of the system}\label{sec:Monitoring}
The WMArchive system was deployed to production in the middle of June 2016.
Since then it has proved to be stable and capable of sustaining the projected
load from the CMS production WMAgents. Fig.~\ref{fig:Rate} shows the data
injection rate for the period from July 2016 until mid March 2017.
\begin{center}
\begin{figure}
\includegraphics[width=0.5\textwidth]{wma_records.pdf}
\caption{Data rate from production CMS WMAgents for a period of July 2016 till March 2017.
More than 125M documents are collected in WMArchive over this period.}
\label{fig:Rate}
\end{figure}
\end{center}
We anticipated that load from production agents will eventually grow,
and, as can be seen from Fig \ref{fig:Rate} it was the case. Initially,
we operated at the level of $\mathcal{O}$(200-300K) docs/day and lately we started seeing
5 times higher rate, reaching 1.5M docs/day during busy time of
CMS production. These peaks correspond to a higher data production demand
of CMS collaboration, e.g. the time of the major conferences where end-users
requested more data from data-operations teams to prepare their results
for a conferences.
Since deploying the system into production we monitored its usage and didn't
see significant load on CPU and I/O of the backend node. The RAM usage is
correlated with the STS size (MongoDB) whose data-lookup load depends
on index size of the stored records, see discussion in Sec. \ref{sec:MongoDBIndex}.
Therefore we anticipate that scaling of WMArchive system is more related to Disk and
RAM components rather then CPU and I/O.
The overall cost of the system had minimal impact on CMS budget. We were able
to operate using two CERN VMs, and we relied on centraly supported CERN HDFS
cluster, see Sec \ref{sec:Benchmarks} for more details. The production and
development nodes were configured via puppet and maintained by CMS.
\subsection{Data lookup}\label{sec:DataLookup}
One of the use cases supported by WMArchive is flexible data lookup.
This task is divided into two categories, the data lookup in STS
and in LTS. The former was easy to achive since our QL, see Sec. \ref{sec:QL},
is natively translated into MongoDB queries. The harder part was to
provide ability to look up documents in LTS.
For that we developed custom Python based classes abstracted from a single
interface to follow the map-reduce paradigm. We provide code examples
for common use cases such as couting the number of documents for a given set of
conditions, finding records, finding failed records, and record aggregation.
These classes are fed into a Spark Python wrapper which applied their
logic to a set of documents found on HDFS. This approach helped us
to accommodate more sophisticated tasks required by CMS data-operations teams.
Here, we want to outline the most difficult use case we faced so far.
The task is to find LogCollect files for a given output LFN.
A workflow in an individual CMS WMAgent follows a series of steps where each step
creates its own set of FWJR documents. The normal processing chain contains the
following job sequence:
processing or production job $\rightarrow$ merge job which creates a final set of output files.
The merge jobs are omitted if the processing or production job creates a file bigger
than some threshold. In that case, the output of the production or processing job is the final output file.
For each job, log-archive files are created and the FWJR contains log-archive's
logical file name. Then a LogCollect job creates tar files when enough log-archive
files are collected. The LogCollect job is a separate and independent job which takes
log-archive files as input and produce a tar file as an output.
The common task is to find records from the output file created by merge job (or
processing, production job) and look-up a corresponding tar file created by the
LogCollect job. Also, to find the corresponding log-archive file from processing or
production job.
To accomplish this task we are forced to perform multi-step look-up of
documents stored on HDFS. Here, are the steps we followed:
\begin{itemize}
\item
look up FWJR documents with provided file name which is present in \emph{LFNArray}
FWJR list attribute;
\item
check the job type and, in the case of merge jobs, iterate and retrieve all input files read in the merge job;
\item
retrieve all unmerged input files from \emph{LFNArray} list;
\item
for each input file look-up FWJR documents associated with unmerged files in \emph{LFNArray};
\item
find the file which ends with \emph{logArchive.tar.gz} in \emph{LFNArray} list in the same
document;
\item
search again the documents with \emph{LogCollect} job type which contain
\emph{logArchive.tar.gz} file above as input file;
\item
return the list of intermediate files along with log-archive and \emph{LogCollect} job files.
\end{itemize}
This procedure quite often causes three iterations over the data on HDFS to find the
final results. But due to the excellent parallel processing pipeline on Spark platform
we were able to quickly ($\mathcal{O}(10)$ minutes/week worth of data) find the desired documents on HDFS.
\subsection{Data Aggregation and Monitoring}\label{sec:Aggregation}
Next to long-term storage and flexible access to individual FWJRs, it is one
objective of the WMArchive service to assist CMS data operators in monitoring
the CMS computing infrastructure through an interactive web interface. Of
course, to access the performance data in the long-term HDFS storage it is
necessary to schedule jobs that retrieve the data, and such tasks can take a
significant amount of time. So to provide a responsive user interface we
constructed an aggregation pipeline that regularly processes the distributed
database of \\
FWJRs to collect performance metrics and cache the aggregated data
in the MongoDB short-term storage, where it is quickly accessible by the
REST server and the user interface.
The primary aggregation procedure based on Apache Spark~\cite{Spark} reduces
the original long-term storage data to a cache of limited size
where only selected information is preserved that data operators may want to commonly
monitor. This cache is not data from each individual job report but is instead
aggregated data grouped only by a number of attributes, for example the job
success state, its host, or its processing site.
During implementation of the aggregation procedure we made several
choices to keep reasonable query times under control.
We took particular care in
selecting a suitable temporal granularity for the aggregation procedure. While
daily aggregated data turned out to be generally sufficient, we keep hourly
aggregated data for a limited time to allow data operators to find failures in
current operations. Stored back in the short-term
storage, a secondary MongoDB aggregation query over the cached
data are produced data for visualizations on timescales suitable for a
responsive user interface.
To interactively present the aggregated data, we originally developed the user
interface depicted in Fig. \ref{fig:WMArchiveWebUI} based on open web
technologies such as the D3.js Ja\-va\-Script data visualization
library~\cite{WMArchiveDeps}. We collected feedback from CMS data operators to
construct a user interface that allows them to comfortably perform commonly
needed queries. They may select any number of performance metrics, such as the
total job computation time, CPU consumption, and storage usage, as well as a
set of visualization axes from the attributes we chose to preserve in the
aggregation procedure. The service performs the secondary MongoDB aggregation
query for each combination of the selected metrics and axes in realtime and
visualizes the results in a way that is suitable for the query. The user may
refine the scope of his or her query through regular expression filters by any
of the available attributes, as well as by specifying a timeframe. Filters are
also applied by interactive user interface elements that allow data operators
to comfortably navigate through the data.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{wma_webui.png}
\caption{The WMArchive Performance Service web interface gives CMS data operators flexible access to a range of commonly needed data aggregation queries.}
\label{fig:WMArchiveWebUI}
\end{center}
\end{figure}
With the aggregated data available in the short-term storage we also
investigated visualization platforms such as Kibana~\cite{Kibana} and
Grafana~\cite{Grafana} for monitoring the performance data. This service is
provided by the CERN Monitoring system that we feed with the data from the
primary aggregation procedure. Complementing the WMAr\-chive Performance
Service web interface that is tailored to specific workflows of data operators,
these platforms allow for flexible inspection of the entire aggregated dataset
and particularly excel at presenting time series data, as shown in
Fig.~\ref{fig:Kibana}. Although these tools are well adopted and the majority
of CMS data-services use them for central monitoring dashboards we found
that there are certain limitations in their capacity above certain limit,
e.g. slowness of making plots from millions of documents, which are under investigation by CERN IT.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{kibana.png}
\caption{The Kibana data visualization platform exposes the entire aggregated WMArchive Performance Service dataset.}
\label{fig:Kibana}
\end{center}
\end{figure}
\subsection{Error handling}\label{sec:ErrorHandling}
The error handling is an important component of every system. Apart from usual
programming bugs catched via Exception mechanism in the Python code, we mostly
struggled to handle errors on the Spark side. The PySpark framework is a wrapper
around Java Spark libraries. As such, the nested, often very long and cumbersome,
errors were hard to debug and understand. Most of them came from JVM memory
issues due to incorrect data handling on a worker node. This was caused by a sequential data
model in our code rather than a distributed lazy processing offered by PySpark APIs.
We must admit that a there is learning curve to efficiently use PySpark and adapt it
to the Python code we were dealing with before implementing this system. Due to this nature,
the errors caused by our Spark jobs were separately examined and debugged in collaboration
with CERN IT, while other errors were captured in server logs and invetigated by developers.
\section{Summary}
We have provided a detailed description of the WMArchive system for
distributed workflow management agents in the CMS collaboration.
The system is designed according to a specific set of requirements
imposed by CMS data-operations teams and it has been deployed into
production since mid 2016. Since then we have collected more than
125 million FWJR documents and have not experienced any significant
problems with system maintenance. The introduction of short-term
and long-term storage systems has helped us to maintain the injection rate and isolate
it from user-based queries. We found that complex search queries
are desired by the CMS data-operations teams and we are able to accommodate them
via a flexible query language discussed in the paper. We plan to expand
our system and include a new data steam coming out from CMS
CRAB~\cite{CRAB} analysis facilities which process user-analysis jobs across
the globe. This will roughly double the demands on the deployed system where
we may face new challenges. But the flexible design of the system and almost a
year of running in production environment allow us to be optimistic about such
an expansion.
\begin{acknowledgements}
We would like to thank our colleagues Seangchan Ryu (FNAL) and Alan Malta
(Univ. of Nebraska) for numerous feedback and guidance across various
details of CMS Workflow Management System. Special thanks goes to Eric
Vaandering (FNAL) for initiating the idea of WMArchive system in CMS and
his constant support along development cycle. We also would like to thank
Luca Menichetti from CERN IT who provided support for development,
maintenance and deployment of our scripts on Spark platform.
\end{acknowledgements}
|
\section{Introduction}
Let $G$ be a compact Lie group with Lie algebra $\ggg$, and pick a maximal torus $T\subset G$. Spaces of the form $(\Hom \pi_1(\Sigma),G)/G$, where $\Sigma$ is a Riemann surface of genus $ g \geq 2$, arise in various branches of geometry. Such spaces have interpretations as the moduli space of flat connections up to gauge transformations (see \cite{ljjwtoric,ljjw, yoshida}); they also occur as a building block in the topological quantum field theoretical construction of invariants of 3-manifolds with boundary $\Sigma$. See also the related discussions of character varieties (\cite{sikora,simpson}), of parabolic Higgs bundles (\cite{hyperpolygon}) and of polygon spaces (\cite{polygon}).
A related space is obtained by marking a point $p \in \Sigma$ and prescribing the holonomy of the connections around that point. That is, fix a generator $c \in \pi_1(\Sigma \setminus p)$ that represents a small curve around $p$, and for $t \in T$, let $S_{g}(t):= \{\rho \in \Hom(\pi_1(\Sigma \setminus p), G \vert \rho(c) \sim t \} / G$, where $\sim$ denotes conjugacy in $G$. In the case where the holonomy $\zeta$ lies in the centre of $G$, the spaces $S_g(\zeta)$ are moduli spaces of stable holomorphic vector bundles over $\Sigma$ (see for example \cite{ab,tw,ek,ek2,kirwan,kirwan2}).
We will consider the space $S_g(t)$, where $t$ is a generic torus element with $\Stab t = T$. In this paper, we will take $G=SO(2n+1)$. Consider the torus bundle $V_g(t) \to S_g(t)$ given by $V_g(t):=\{\rho \in \Hom(\pi_1(\Sigma \setminus p),G) \vert \rho(c) =t\}$. If $\phi \in \Phi(\ggg)$ is a root of $\ggg$, let $L_\phi$ be the line bundle associated to $V_g(t)$ via the torus representation with weight $\phi$. We will construct geometric representatives for the first Chern classes of these line bundles. By considering these geometric representatives, we are able to identify several particular products of Chern classes which vanish in $H^*(S_g(t))$; we also give a combinatorial proof that any monomial in the $c_1(L_\phi)$ is equivalent in $H^*(S_g(t))$ to a combination of those particular monomials and hence also vanishes. Our geometric representatives are analogous to Schubert cycles for flag manifolds; however, a key difference is that there is no canonical complex structure on $S_g(t)$, and our geometric representatives for Chern classes will not generally be complex subvarieties of $S_g(t)$ with respect to an arbitrary choice of K\"ahler structure. Thus a feature of our topological approach is that it enables us to make use of these particular geometric representatives which would not show up under an algebraic geometric treatment of the subject.
This paper builds on our earlier work \cite{aj}, where we used a similar approach for the case $G=SU(n)$, making $S_g(t)$ the moduli space of parabolic holomorphic vector bundles over $\Sigma$. (For a different approach to finding generators and relations for the cohomology of this moduli space, see \cite{ek2}.) This itself was based on the earlier paper \cite{jw}, which used this geometric approach in the case $G=SU(2)$ to provide a geometric proof to a conjecture of Newstead (\cite{newstead}).
Let us now take $G=SO(2n+1)$, and fix generators $a_1,\ldots, a_g, b_1, \ldots, b_g, c$ for the fundamental group $\pi_1(\Sigma \setminus \{p\})$, such that $c$ represents the boundary of $\Sigma \setminus \{p\}$ and $\prod_{l=1}^g[a_l, b_l]=c$. Recall that the set of roots of $\ggg$ is $\Phi(\ggg)=\{\pm(\eta_i + \eta_j) \mid 1 \leq i \leq j \leq n\}\cup \{\pm(\eta_i - \eta_j) \mid 1 \leq i < j \leq n\}$. Choose the maximal torus $T \subset G$ consisting of elements
\[
\left(\begin{array}{cccccc}
\cos \theta_1 & -\sin \theta_1 & &&&\\
\sin \theta_1 & \cos \theta_1 &&&& \\
&& \ddots &&& \\
&&& \cos \theta_n & -\sin \theta_n & \\
&&& \sin \theta_n & \cos \theta_n & \\
&&&&&1\\
\end{array}\right);
\]
for ease of notation such elements will be denoted $(\theta_1, \ldots, \theta_n)$.
\begin{defn}\label{generic}
We say an element $t = (\theta_1, \ldots, \theta_n)$ is \emph{generic} if $\Stab t = T$, and if the only relation $\lambda_1 \theta_1 + \cdots + \lambda_n \theta_n \in 2\pi \mathbb{Z} $ with $\lambda_i \in \{\pm 1, 0\}$ is the trivial relation $\lambda_1 = \cdots = \lambda_n = 0$.
\end{defn}
Let $t \in T$ be generic, and set
\[S_{n,g}(t) := \{ \rho \in \Hom(\pi_1(\Sigma \setminus p), SO(2n+1))\mid \rho(c) \sim t\}/G.
\]
Let $\mathbb{C}_{(\pm(\eta_i \pm \eta_j))}$ denote the 1-dimensional torus representation
\begin{align*}
T \times \mathbb{C}_{(\pm(\eta_i \pm \eta_j))} & \to \mathbb{C}_{(\pm(\eta_i \pm \eta_j))} \\
(\theta_1, \ldots, \theta_n) \cdot z &\mapsto e^{\pm \sqrt{-1}(\eta_i \pm \eta_j)}z.
\end{align*}
For $\phi \in \Phi(\ggg)$ a root of $\ggg$, we consider line bundles $L_\phi:=(V_g(t)\times \mathbb{C}_{(\phi)})/T \to V_g(t)/T=S_{n,g}(t)$, where the quotient is by the diagonal $T$-action. We will also denote $L_{(\eta_i \pm \eta_j)}$ by $L_{ij}^\pm$; observe that $L_{-(\eta_i \pm \eta_j)}\cong (L_{\eta_i \pm \eta_j})^*$.
\begin{thm} \label{main} The product $\prod_{\phi \in \Phi(\ggg)}c_1(L_\phi)^{k_\phi} \in H^*(S_{n,g}(t);\mathbb{Q})$ vanishes whenever $\sum_{\phi \in \Phi(\ggg)}k_\phi \geq 2gn^2+ \frac{1}{2}(n-1)(n-2) $.\end{thm}
The dimension of $S_{n,g}(t)$ is $2gn(2n+1)-2n(n+1)$, so when $g \geq \frac{3n}{2}$, our theorem shows that the top cohomology classes of $S_{n,g}(t)$ must be generated by elements other than the $c_1(L_\phi)$.
We begin in the case $n=1$, the proof for which is analogous to Weitsman's proof (\cite{jw}) of the Newstead conjecture. When $n=1$, we have just the one pair of roots $\pm\phi$ of $\mathfrak{so(3)}$, and the corresponding pair of line bundles $L_\phi$ and $L_{-\phi}=L_\phi^*$; of course $c_1(L_\phi)=-c_1(L_\phi^*)=-c_1(L_{-\phi})$. In this case, the theorem is that $c_1(L_\phi)^{2g}=0$, as follows:
\begin{prop} Let $G=SO(3)$. Let $t \in T$ be generic, and let $L \to S_{1,g}(t)$ be the line bundle associated to $V_g(t)$ by the representation $\left( \begin{array}{ccc} \cos \theta & -\sin \theta & \\ \sin \theta & \cos \theta & \\ &&1 \end{array} \right)\cdot z = e^{2i\theta}z$. Then $(c_1(L))^{2g}=0$ in $H^{4g}(S_{1,g}(t); \mathbb{Q})$.
\end{prop}
\begin{proof}
For $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$, consider the section $s_x$ of $L$ induced by the $T$-equivariant map
\begin{align*} f_x: V_g(t) &\to \mathbb{C} \\ \rho &\mapsto (\rho(x))_{11}-(\rho(x))_{22} + i((\rho(x))_{21}+(\rho(x))_{12})
\end{align*} (where the subscript $ij$ denotes the $(i,j)^{th}$ matrix entry). If $s_x=0$, then
\[
\rho(x) = \left( \begin{array}{ccc} a & b & c \\ -b & a & d \\ e & f&g \end{array} \right) \in S0(3).
\]
But since $\rho(x) \in SO(3)$, we also have $\vert c \vert = \vert d \vert$ and $cd=0$, thus $c=d=0$. Similarly $e=f=0$, and so $a^2 + b^2 = 1$ and $g=1$; that is, $\rho(x) \in T$. Suppose $\rho(x)=0$ for all $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$. Then $\rho(a_1), \ldots, \rho(a_g), \rho(b_1), \ldots, \rho(b_g) \in T$. But then $\rho(c) = \prod_{i=1}^g [\rho(a_i), \rho(b_i)]=1\neq t$, so in fact there are no such $\rho$ in $V_g(t)$. Hence the section $(s_{a_1},\ldots, s_{a_g}, s_{b_1}, \ldots, s_{b_g})$ of $L^{\oplus 2g}$ is nowhere zero, so $c_{2g}(L^{\oplus 2g})=(c_1(L))^{2g}=0$.
\end{proof}
When $n > 1$, the combinatorics of the vanishing loci of the relevant sections becomes more complicated. The key idea in the proof of Theorem \ref{main} is similar to that in our earlier paper \cite{aj}, but the combinatorics required in this case is more intricate.
The outline of the rest of this paper is as follows.
In section \ref{five}, we prove the theorem for the case $G=SO(5)$. As in \cite{aj}, the proof begins by identifying specific products of Chern classes which must vanish because there are sections (constructed explicitly) of the relevant line bundles with no common zeros. The proof is completed by showing that any monomial of sufficiently high degree is equivalent to a combination of these products of Chern classes which have already been shown to vanish. In section \ref{purple} we allow $n$ to be an arbitrary positive integer, exhibit some collections of sections of the appropriate line bundles with no common zeros, and conclude that the corresponding products of Chern classes vanish. In section \ref{heart} we turn to the combinatorial heart of the argument, showing that every product of Chern classes of sufficient degree is equivalent in $H^*(S_{n,g}(t))$ to a combination of those shown to vanish in section \ref{purple}.
\section{$G=SO(5)$} \label{five}
We have chosen to devote a section to proving our result in the case $G=SO(5)$. This is intended to provide intuition for the general case in a more tractable context.
Let $G=SO(5)$ and let $T\subset G$ be the maximal torus
\[
T = \left\{\left(\begin{array}{cccccc}
\cos \theta_1 & -\sin \theta_1 & &&&\\
\sin \theta_1 & \cos \theta_1 &&&& \\
&&& \cos \theta_2 & -\sin \theta_2 & \\
&&& \sin \theta_2 & \cos \theta_2 & \\
&&&&&1\\
\end{array}\right)\right\} \subset G;
\]
to simplify notation we will also denote elements of $T$ by $(\theta_1,\theta_2)$. Choose an element $t \in T$ that is generic in the sense that $\theta_1 \pm \theta_2 \notin 2\pi\mathbb{Z}$. Recall that the roots of $\ggg$ are $\pm(2\eta_1, 2\eta_2, \eta_1 + \eta_2, \eta_1 - \eta_2)$. If $\phi = \pm(\eta_j \pm \eta_k)$ we denote by $\mathbb{C}_{(\phi)}$ the one dimensional torus representation $(\theta_1,\theta_2) \cdot z = e^{\pm i (\theta_j \pm \theta_k)}z$. Consider the line bundles $L_\phi = V_g(t) \times_T \mathbb{C}_{(\phi)} \to V_g(t)/T = S_{2,g}(t) $ for $\phi \in \Phi(G)$. Observe that $L_{-\phi} \cong L^*_{\phi}$, and $L_{\phi + \psi} \cong L_\phi \otimes L_\psi$. Hence $c_1(L^*_\phi)=-c_1(L_\phi)$, and $c_1(L_{\phi + \psi}) = c_1(L_\phi) + c_1(L_\psi)$. In this case, Theorem \ref{main} says
\begin{prop} \label{propfive} The product $\prod_{\phi \in \Phi(G)} c_1 (L_\phi)^{k_\phi}$ vanishes whenever $\sum_{\phi \in \Phi(G)}k_\phi \geq 8g$.
\end{prop}
As in the case of $G=SO(3)$, and following the strategy for $G=SU(n)$ in \cite{aj}, we prove this by identifying collections of sections of the $L_\phi$ with no common zeros, thus observing that the corresponding products of first Chern classes $c_1(L_\phi)$ vanish. We will identify three such vanishing products, and then prove that any other monomial of degree at least $8g$ is equivalent in $H^*(S_{2,g}(t))$ to a combination of those three.
\begin{defn} For $x \in \{a_l, b_l \mid 1 \leq l \leq g \} $, let $s^\pm_{ij}(x)$ be the section of $L_{\eta_i \pm \eta_j}$ induced by the $T$-equivariant map
\begin{align*}
V_g(t) \to & \mathbb{C}_{(\eta_i \pm \eta_j)}\\
\rho \mapsto & (\rho(x))_{2i-1,2j-1} \mp (\rho(x))_{2i,2j} + \sqrt{-1}((\rho(x))_{2i,2j-1}\pm(\rho(x))_{2i-1,2j}).
\end{align*}
(Note that $s^-_{ii}(x)$ is a section of the trivial bundle $L_0$.)
\end{defn}
Observe that if $s^{\pm}_{ij}(x)$ vanishes, then the $(i,j)^{th}$ 2-by-2 block in $\rho(x)$ takes the form $\left(\begin{array}{cc} a & b \\ \mp b & \pm a \end{array}\right)$; if for any given $i,j$, both $s^{+}_{ij}(x)$ and $s^{-}_{ij}(x)$ vanish, then the $(i,j)^{th}$ 2-by-2 block in $\rho(x)$ is zero.
To simplify notation, will will write $c^\pm_{ij} := c_1(L^\pm_{ij})$
\begin{lem} The monomial $z_1 := (c^+_{11}c^+_{12}c^-_{12})^{2g}$ vanishes in $H^{12g}(S_{2,g}(t))$.
\end{lem}
\begin{proof}Let $x \in \{a_l, b_l\}$, and consider the sections $s^+_{11}(x), s^+_{12}(x), s^-_{12}(x)$. If these all vanish, then
$
\rho(x) = \left(\begin{array}{ccccc}
a & b &0 &0&c\\
-b & a &0&0&d \\
*&*&*& *& *\\
*&*&*& * & * \\
*&*&*&*&*\\
\end{array}\right)$, for some $a,b,c,d \in \mathbb{R}$. Since $\rho(x) \in SO(5)$, $c=d=0$, so in fact
$\rho(x) = \left(\begin{array}{ccccc}
a & b &0 &0&0\\
-b & a &0&0&0 \\
0&0&*& *& *\\
0&0&*& * & * \\
0&0&*&*&*\\
\end{array}\right)$, for some $a, b \in \mathbb{R}$ with $a^2 + b^2 =1$. If all $6g$ sections $s^+_{11}(x), s^+_{12}(x), s^-_{12}(x)$, for $x = a_1, \ldots, a_g, b_1, \ldots, b_g$ vanish, then
\begin{align*}
\rho(c) &= \prod_{l=1}^g[\rho(a_l),\rho(b_l)]\\ &= \prod_{l=1}^g \left[\left(\begin{array}{ccccc}
\cos \alpha_l & -\sin \alpha_l &0 &0&0\\
\sin \alpha_l & \cos \alpha_l &0&0&0 \\
0&0&*& *& *\\
0&0&*& * & * \\
0&0&*&*&*\\
\end{array}\right),\left(\begin{array}{ccccc}
\cos \beta_l & -\sin \beta_l &0 &0&0\\
\sin \beta_l & \cos \beta_l &0&0&0 \\
0&0&*& *& *\\
0&0&*& * & * \\
0&0&*&*&*\\
\end{array}\right) \right] \\ &= \left(\begin{array}{ccccc}
1 & 0 &0 &0&0\\
0 & 1 &0&0&0 \\
0&0&*& *& *\\
0&0&*& * & * \\
0&0&*&*&*\\
\end{array}\right) \neq t
\end{align*}
(since $t$ was chosen to be generic). But $\rho(c) = t$ for $\rho \in V_g(t)$, and so these $6g$ sections have no common zeros; thus $(c^+_{11} c^+_{12}c^-_{12})^{2g}=0$ as claimed.
\end{proof}
Similarly, $z_2:=(c^+_{22} c^+_{12}c^-_{12})^{2g}=0$.
\begin{lem}\label{fiveplus}Suppose $A_l \in SO(5)$ for $1 \leq l \leq g$, and each $A_l$ has the form $\left(\begin{array}{ccccc}
&&&&0\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R_{\alpha_l}$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R_{\beta_l}$}}}&0\\
&&&&0\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R_{\gamma_l}$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R_{\delta_l}$}}}&0\\
0&0&0&0&1
\end{array}\right)$, where $R_\theta$ represents the 2-by-2 block $\left(\begin{array}{cc}\cos \theta & - \sin \theta \\ \sin \theta & \cos \theta \\\end{array}\right)$. Then $\prod_{l=1}^g [A_{2l-1},A_{2l}]$ cannot be a generic torus element.
\end{lem}
\begin{proof} Consider the map
\begin{align*}
\kappa: U(2) &\hookrightarrow SO(5) \\
\kappa: \left(\begin{array}{cc}a & b\\ c & d \\\end{array}\right) & \mapsto \left(\begin{array}{ccccc} \Re a & -\Im a & \Re b & - \Im b & 0 \\\Im a & \Re a & \Im b & \Re b & 0 \\ \Re c & -\Im c & \Re d & - \Im d & 0 \\\Im c & \Re c & \Im d & \Re d & 0 \\ 0&0&0&0&1\\ \end{array}\right).
\end{align*} This is an injective homomorphism, and $\kappa(M) \in T \iff M$ is diagonal in $U(2)$. Each $A_l$ is in the image of $\kappa$; say $A_l=\kappa(M_l)$, for $M_l \in U(2)$. Then
\begin{align*}
\prod_{l=1}^g [ A_{2l-1}, A_{2l}] &= \prod_{l=1}^g[\kappa(M_{2l-1}), \kappa(M_{2l})] \\
&= \kappa \left( \prod_{l=1}^g [M_{2l-1},M_{2l}]\right).
\end{align*}
Observe that $M=\prod_{l=1}^g [M_{2l-1}, M_{2l}]$ has determinant 1. If $M$ is not diagonal, then $\kappa(M) \neq t$. If $M$ is diagonal, then it must be $\left(\begin{array}{cc}e^{i\theta} & \\ & e^{-i\theta} \end{array} \right)$ for some $\theta$, so $\kappa(M) = \left(\begin{array}{ccc}R_\theta && \\ & R_{-\theta} & \\ &&1 \end{array} \right)$ is not generic. Thus in particular, $\prod_{l=1}^g [A_{2l-1}, A_{2l}] \neq t$.
\end{proof}
\begin{cor}The monomial $p=(c^+_{11}c^+_{12}c^+_{21}c^+_{22})^{2g}$ vanishes in $H^{16g}(S_{2,g}(t))$. \end{cor}
\begin{proof} For each $x \in \{a_l, b_l \mid 1 \leq l \leq g\}$, consider the sections $s^+_{11}(x), s^+_{12}(x), s^+_{21}(x)$, and $s^+_{22}(x)$. If these were all to vanish, then each $\rho(x)$ would have the form $\left( \begin{array}{ccccc} a&b&c&d&\\ -b&a&-d&c&\\e&f&g&h&\\-f&e&-h&g&\\&&&&1\end{array}\right)$, where $a,b,c,d,e,f \in \mathbb{R}$. By Lemma \ref{fiveplus}, if these $8g$ sections all vanish, then $\prod_{l=1}^g[\rho(a_l),\rho(b_l)]\neq t$. Thus these $8g$ sections have no common zeros, so $(c^+_{11}c^+_{12}c^+_{21}c^+_{22})^{2g}=0$.
\end{proof}
\begin{lem}\label{fivepm}Suppose $B_k \in SO(5)$ for $1 \leq k \leq 2g$, and each $B_k$ has the form $\left( \begin{array}{ccccc} a&b&c&d&\\ -b&a&d&-c&\\e&f&g&h&\\f&-e&-h&g&\\&&&&1\end{array}\right)$. Then $\prod_{k=1}^g[B_{2k-1},B_{2k}]\neq t$.\end{lem}
\begin{proof} Let $E = \left(\begin{array}{ccccc}0&1&&&\\ 1&0&&&\\ &&1&&\\&&&1&\\&&&&1\end{array}\right)$. Observe that $EB_kE^{-1}$ has the form described in Lemma \ref{fiveplus}. So by Lemma \ref{fiveplus}, $\prod_{k=1}^g[EB_{2k-1}E^{-1},EB_{2k}D^{-1}]$ cannot be a generic torus element. If $\prod_{k=1}^g[B_{2k-1},B_{2k}]= t$, then $\prod_{k=1}^g[EB_{2k-1}E^{-1},EB_{2k}D^{-1}]=EtE^{-1}$. Note that
\[E \left( \begin{array}{ccc}R_{\theta_1}&&\\&R_{\theta_2}&\\&&1\end{array}\right)E^{-1}= \left( \begin{array}{ccc}R_{-\theta_1}&&\\&R_{\theta_2}&\\&&1\end{array}\right), \]
so $EtE^{-1}$ is a generic torus element. This is impossible by Lemma \ref{fiveplus}, hence $\prod_{k=1}^g[B_{2k-1},B_{2k}]\neq t$. \end{proof}
\begin{cor}The monomial $q = (c^+_{11}c^+_{22}c^-_{12}c^-_{21})^{2g}$ vanishes in $H^{16g}(S_{2,g}(t))$. \end{cor}
\begin{proof} Consider the sections $s^+_{11}(x), s^+_{22}(x), s^-_{12}(x)$, and $s^-_{21}(x)$ for each $x \in \{ a_l, b_l \vert 1 \leq l \leq g\}$. If these all vanish, then each $\rho(x)$ has the form described in Lemma \ref{fivepm}, and so $\prod_{l=1}^g[\rho(a_l),\rho(b_l)]\neq t = \rho(c)$. So these $8g$ sections have no common zeros, so $(c^+_{11}c^+_{22}c^-_{12}c^-_{21})^{2g}=0$.\end{proof}
\begin{lem} Let $0 \leq m \leq 2g$. The monomials
\begin{equation} \label{typeone}y^+_m:=(c^+_{11})^{2g}(c^+_{22})^{2g}(c^+_{12})^{2g+m}(c^-_{12})^{2g-m}
\end{equation} and
\begin{equation} \label{typetwo} y^-_m:=(c^+_{11})^{2g}(c^+_{22})^{2g}(c^+_{12})^{2g-m}(c^-_{12})^{2g+m}
\end{equation}vanish in $H^{16g}(S_{2,g}(t))$.\end{lem}
\begin{proof}For fixed $x$, we have seen above that the sections $s^+_{11}(x), s^+_{12}(x), s^+_{21}(x)$, and $s^+_{22}(x)$ all vanish when $\rho(x)$ has the form
\begin{equation}\label{matformp}\left( \begin{array}{ccccc} a&b&c&d&\\ -b&a&-d&c&\\e&f&g&h&\\-f&e&-h&g&\\&&&&1\end{array}\right).\end{equation} Observe that if the sections $s^+_{11}(x), s^-_{12}(x), s^+_{12}(x)$, and $s^+_{22}(x)$ all vanish, then $\rho(x)$ has the form $\left( \begin{array}{ccccc} a&b&0&0&0\\ -b&a&0&0&0\\0&0&c&d&0\\0&0&-d&c&0\\0&0&0&0&1\end{array}\right)$, which is in particular also in the form \eqref{matformp}. Thus if $X_m$ is any $m$-element subset of $\{a_1, \ldots, a_g, b_1, \ldots, b_g\}$, then the collection of sections
\begin{align*}
s^+_{11}(x), s^+_{22}(x), s^+_{12}(x) & \text{ for all } x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\},\\
s^-_{12}(x) & \text{ for all } x \in X_m, \\
s^+_{21}(x) & \text{ for all } x \in X^c_m
\end{align*} has no common zeros, so the monomial \eqref{typeone} vanishes.
Similarly, the collection of sections $s^+_{11}(x), s^+_{22}(x), s^-_{12}(x)$, and $s^-_{21}(x)$ vanishes when $\rho(x)$ has the form $\left( \begin{array}{ccccc} a&b&c&d&\\ -b&a&d&-c&\\e&f&g&h&\\f&-e&-h&g&\\&&&&1\end{array}\right)$, and the collection $s^+_{11}(x), s^+_{22}(x), s^-_{12}(x)$, and $s^+_{12}(x)$ vanishes when $\rho(x)$ has the form $\left( \begin{array}{ccccc} a&b&0&0&0\\ -b&a&0&0&0\\0&0&c&d&0\\0&0&-d&c&0\\0&0&0&0&1\end{array}\right)$, so by Lemma \ref{fivepm} the collection of sections \begin{align*}
s^+_{11}(x), s^+_{22}(x), s^-_{12}(x) & \text{ for all } x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\},\\
s^+_{12}(x) & \text{ for all } x \in X_m, \\
s^-_{21}(x) & \text{ for all } x \in X^c_m
\end{align*} has no common zeros, so the monomial \eqref{typetwo} vanishes.
\end{proof}
\begin{prop}Suppose $\zeta$ is a monomial in the $c^\pm_{ij}$ with degree at least $8g$. Then $\zeta$ is a combination of monomials $z_1, z_2, y^+_m$, and $y^-_m$.
\end{prop}
\begin{proof}Use the relations
\begin{align*}\label{fiverelns}
c^+_{12}&=\frac{1}{2}(c^+_{11}+c^+_{22}) \\
c^-_{12} &= \frac{1}{2}(c^+_{11}-c^+_{22})\\
c^-_{ij} &= -c^-_{ji}
\end{align*} to write $\zeta$ as a sum of terms $(c^+_{11})^a(c^+_{22})^b$, where $a+b \geq 8g$, and consider each term separately. We must have $a \geq 2g$ or $b \geq 2g$; without loss of generality assume $a \geq 2g$. Rewrite this term as $(c^+_{11})^{2g}(c^+_{12}-c^+_{22})^{a-2g}(c^+_{22})^b$, and consider each term $\lambda(c^+_{11})^{2g}(c^+_{12})^{a-2g-m})(c^+_{22})^{b+m}$. Note that $a-2g+b \geq 6g$, so either $a-2g-m \geq 2g$ or $b+m \geq 2g$. \begin{itemize} \item If $a-2g-m \geq 2g$, rewrite this term as $\lambda'(c^+_{11})^{2g}(c^+_{12})^{2g}(c^-_{12}+c^+_{22})^{a-m-4g}(c^+_{22})^{b+m}$, and consider each term $\tilde{\lambda}(c^+_{11})^{2g}(c^+_{12})^{2g}(c^-_{12})^{a-m-4g-s}(c^+_{22})^{b+m+s}$ in the resulting expansion. Note that $a+b-4g \geq 4g$, so either $a-m-4g-s \geq 2g$ or $b+m+s \geq 2g$. \begin{itemize} \item If $a-m-4g-s \geq 2g$, this term is a multiple of $z_1$. \item If $b+m+s \geq 2g$, rewrite this term as $\bar{\lambda}(c^+_{11})^{2g}(c^+_{22})^{2g}(c^+_{12})^{2g}(c^-_{12})^{a-m-4g-s}(c^+_{12}-c^-_{12})^{b+m+s-2g}$. Each term in the expansion of this polynomial is a multiple of some $y^\pm_m$. \end{itemize} \item If $b+m \geq 2g$, rewrite the term as $\lambda(c^+_{11})^{2g}(c^+_{22})^{2g}(c^+_{12})^{a-2g-m}(c^+_{12}-c^-_{12})^{b+m-2g}$. Each term in the expansion of this polynomial is a multiple of some $y^\pm_m$.\end{itemize}Hence $\zeta$ is equal to a sum of multiples of the monomials $z_1, z_2$ and $y^\pm_m$ as claimed.
\end{proof}
\begin{cor} $\prod_{\phi \in \Phi(SO(5))}c_1(L_\phi)^{k_\phi}$ vanishes whenever $\sum_{\phi \in \Phi(SO(5))}k_\phi \geq 8g.$ \qed \end{cor}
\section{The general case} \label{purple}
We now begin our study of the general case $G=SO(2n+1)$. As we did for $SO(5)$ in the previous section, we will show that some particular products of the $c_1(L_\phi)$, for roots $\phi$ of $\mathfrak{so(2n+1)}$, vanish, by finding sections of the $L_\phi$ with no common zeros.
A $T$-equivariant map $V_g(t) \to \mathbb{C}_{(\phi)}$ induces a section of $L_\phi = (V_g(t) \times \mathbb{C}_{(\phi)})/T \to V_g(t)/T=S_{n,g}(t)$. Let $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$ be one of our chosen generators of $\pi_1(\Sigma\setminus \{p\})$ other than $c$, and consider the maps
\begin{align*}
f^\pm_{ij}(x): V_g(t)&\to \mathbb{C}_{(\eta_i \pm \eta_j)} \\
\rho & \mapsto (\rho(x))_{2i-1,2j-1}\mp (\rho(x))_{2i,2j}+ \sqrt{-1}((\rho(x))_{2i,2j-1} \pm (\rho(x))_{2i-1,2j}).
\end{align*} These maps are $T$-equivariant and induce sections $s^{\pm}_{ij}(x)$ of $L^\pm_{ij}$. The nature of the $T$ action on $SO(2n+1)$ is such that top left $2n$-by-$2n$ corner of elements of $SO(2n+1)$ can be divided into 2-by-2 matrices on which the behaviour of the torus action can be considered separately. Thus it will often be convenient to write elements of $SO(2n+1)$ in the form
$\left(\begin{array}{ccccccc}
&&&&&&b_1\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$A_{11}$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$A_{1n}$}}}&b_2\\
\multicolumn{2}{c}{\vdots}&\multicolumn{2}{c}{\ddots}&\multicolumn{2}{c}{\vdots}&\vdots\\
&&&&&&b_{2n-1}\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$A_{n1}$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$A_{nn}$}}}&b_{2n}\\
c_1&c_2&\multicolumn{2}{c}{{\cdots}}&c_{2n-1}&c_{2n}&d
\end{array}\right)$, adopting the notational convention that capital letters denote 2-by-2 arrays whereas lowercase letters represent real numbers.
The next lemma is the direct generalisation of Lemma \ref{fiveplus} to the general case.
\begin{lem}\label{genplus} Suppose $M_l \in SO(2n+1)$ for $1 \leq l \leq 2g$, and each $M_l$ has the form
\begin{equation}\label{plusmat}
\left(\begin{array}{ccccccc}
&&&&&&0\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R^{11}_l$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R^{1n}_l$}}}&0\\
\multicolumn{2}{c}{\vdots}&\multicolumn{2}{c}{\ddots}&\multicolumn{2}{c}{\vdots}&\vdots\\
&&&&&&0\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R^{n1}_l$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R_l^{nn}$}}}&0\\
0&0&\multicolumn{2}{c}{{\cdots}}&0&0&1
\end{array}\right),
\end{equation}where each $R^{ij}_l$ is a 2-by-2 block of the form $\left(\begin{array}{cc}x&y\\-y&x\end{array}\right)$, with $x, y \in \mathbb{R}$. Then $\prod_{l=1}^g[M_{2l-1},M_{2l}]$ cannot be a generic torus element. \end{lem}
\begin{proof}Consider the map
\begin{align*}
\kappa:U(n) &\hookrightarrow SO(2n+1)\\
\kappa: \left(\begin{array}{ccc}a_{11}&\cdots & a_{1n}\\ \vdots & \ddots & \vdots \\ a_{n1}& \cdots & a_{nn} \end{array}\right) & \mapsto \left(\begin{array}{cccccc} \Re a_{11} & -\Im a_{11} & \cdots &&&0\\ \Im a_{11}& \Re a_{11} &&&&0\\ &&\ddots &&&\vdots \\ &&&\Re a_{nn} & -\Im a_{nn} &0 \\ &&& \Im a_{nn} & \Re a_{nn}&0\\0&0&\cdots&0&0&1\end{array}\right).
\end{align*} This is an injective homomorphism, and $\kappa(M) \in T \iff M$ is diagonal in $U(n)$.Under the hypothesis of this lemma, each $M_l$ is in the image of $\kappa$; say $M_l = \kappa(N_l)$, for $N_l \in U(n)$. Then
\[\prod_{l=1}^g[M_{2l-1},M_{2l}] = \prod_{l=1}^g[\kappa(N_{2l-1}), \kappa(N_{2l})]=\kappa\left(\prod_{l=1}^g[N_{2l-1},N_{2l}]\right).
\] Observe that if $N=\prod_{l=1}^g[N_{2l-1},N_{2l}]$ has determinant 1. If $N$ is not diagonal, then $\kappa(N)$ is not a torus element. If $\kappa(N)$ is a torus element then $N$ is a diagonal matrix in $U(n)$ of determinant 1, so is $\left(\begin{array}{ccc}e^{i\theta_1}&&\\&\ddots&\\&&e^{i\theta_n}\end{array}\right)$, where $\theta_1 + \ldots + \theta_n \in 2\pi \mathbb{Z}$. Thus $\kappa(N)$ is $(\theta_1, \ldots, \theta_n) \in T$, which is not generic in the sense of Definition \ref{generic} because $\theta_1 + \ldots + \theta_n \in 2 \pi \mathbb{Z}$.
\end{proof}
\begin{cor}The sections $\{s^{+}_{ij}(x) \vert 1 \leq i, j \leq n, x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}\}$ of the line bundles $L^+_{ij}$ have no common zeros. \end{cor}
\begin{proof}Fix $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$ and consider the sections $\{s^{+}_{ij}(x) \mid 1 \leq i, j, \leq n \}$. These sections all vanish when
\[\rho(x)= \left( \begin{array}{cccccc} x_{11} & -y_{11} & \cdots & x_{1n} & -y_{1n} & z_1\\ y_{11} & x_{11} & \cdots & y_{1n} &x_{1n}& z_2\\ \vdots & \vdots & \ddots & \vdots & \vdots & \vdots \\x_{n1} & -y_{n1} & \cdots & x_{nn} & -y_{nn} & z_{2n-1} \\ y_{n1} & x_{n1} & \cdots & y_{nn} & x_{nn} & z_{2n} \\ w_1 & w_2 & \cdots & \cdots & w_{2n} & u \end{array}\right).
\] Since $\rho(x) \in SO(2n+1)$, we must have $z^2_{2l-1}=z^2_{2l}$, and $z_{2l-1}z_{2l}=0$, so $z_1 = \cdots = z_{2n} =0$. Similarly $w_1 = \cdots = w_{2n} = 0$, hence $u=1$, so $\rho(x)$ is in the form \eqref{plusmat} from Lemma \ref{genplus}. At points in $S_{n,g}(t)$ where all of our sections vanish, therefore, $\rho(x)$ is in this form for every $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$, so by Lemma \ref{genplus}, $\prod_{l=1}^g[\rho(a_l),\rho(b_l)]$ cannot be a generic torus element. However, $\prod_{l=1}^g[\rho(a_l),\rho(b_l)] = \rho(c) = t$ which was chosen to be a generic torus element. Thus there are no such points in $S_{n,g}(t)$, that is, the locus on which every one of these sections vanishes is empty. \end{proof}
In the coming discussion, we will make use of the following definition from our earlier paper \cite{aj}.
\begin{defn} \label{block}Let $X$ be a finite set. A \emph{block} $B$ in $X\times X$ is a subset of $X\times X$ of the form $V \times V^c$, where $\varnothing \subsetneq V \subsetneq X$ is a proper nonempty subset of $X$. We denote the set of all blocks in $X \times X$ by $\mathcal{B}[X]$. If $B=V \times V^c$, let $\bar{B}=V^c\times V$. We will also make use of the indicator functions
\[
\epsilon_B(i,j) := \begin{cases}
1 & (i,j) \in B\sqcup\bar{B}\\
-1 & \text{otherwise}
\end{cases}
\]
and \[
\epsilon_V(i) := \begin{cases}
1 & i \in V\\
-1 & \text{otherwise}
\end{cases}.
\]
\end{defn}
\begin{notn}For a positive integer $m$, we will denote by $[m]$ the set $\{1, \ldots, m\}$.
\end{notn}
For $l \in [n]$, let $E_l \in O(2n+1)$ be the matrix
\bordermatrix{&&&&&2l&&\cr
&1&&&&&\cr
&&\ddots&&&&\cr
&&&1&&\cr
&&&&0&1&&&\cr
2l&&&&1&0&\cr
&&&&&&1\cr
&&&&&&&\ddots\cr
&&&&&&&&1\cr}
(so conjugation by $E_l$ switches the $(2l-1)^{th}$ and $(2l)^{th}$ rows, and the $(2l-1)^{th}$ and $(2l)^{th}$ columns).
\begin{lem} \label{genminus} Let $B=V \times V^c \in \mathcal{B}[[n]]$. Suppose $M_l \in SO(2n+1)$ for each $1 \leq l \leq 2g$, and each $M_l$ has the form
\begin{equation} \label{minusmat}
\left(\begin{array}{ccccccc}
&&&&&&0\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$S^{11}_l$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$S^{1n}_l$}}}&0\\
\multicolumn{2}{c}{\vdots}&\multicolumn{2}{c}{\ddots}&\multicolumn{2}{c}{\vdots}&\vdots\\
&&&&&&0\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$S^{n1}_l$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$S_l^{nn}$}}}&0\\
0&0&\multicolumn{2}{c}{{\cdots}}&0&0&1
\end{array}\right),
\end{equation} where $S^{ij}_l$ takes the form
\[
\begin{cases} \left(\begin{array}{cc}x&-y \\ y & x \end{array}\right) & (i,j) \notin B \sqcup \bar{B}\\ \left( \begin{array}{cc} w& z \\ z & -w \end{array}\right) & \text{otherwise}.
\end{cases}
\]
Then $\prod_{l=1}^g[M_{2l-1},M_{2l}]$ cannot be a generic torus element.
\end{lem}
\begin{proof} Let $E= \prod_{k \in V} E_k$, and observe that if $M_l$ is in the form \eqref{minusmat} then $EME^{-1}$ is in the form \eqref{plusmat}. Thus $E\prod_{l=1}^g[M_{2l-1},M_{2l}]E^{-1}$ is not a generic torus element. Notice that if $h = ( \theta_1, \ldots, \theta_n)\in T$ is a generic torus element, then $EhE^{-1}= (\epsilon_V(1)\theta_1, \ldots, \epsilon_V(n) \theta_n)$ is also generic. Hence $\prod_{l=1}^g[M_{2l-1}, M_{2l}]$ is not a generic torus element.
\end{proof}
\begin{cor} Let $B \in \mathcal{B}[[n]]$. The sections
\begin{equation*}
\{s^{-\epsilon_B(i,j)}_{ij}(x) \vert 1 \leq i, j \leq n, x = a_1, \ldots, a_g, b_1, \ldots, b_g\}
\end{equation*}have no common zeros. \end{cor}
\begin{proof}Consider the sections $\{s^{-\epsilon_B(i,j)}_{ij}(x) \vert 1 \leq i, j \leq n, x = a_1, \ldots, a_g, b_1, \ldots, b_g\}$. These sections all vanish when each
\[
\rho(x) = \left(\begin{array}{ccccccc}
&&&&&&z_1\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R^{11}_x$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R^{1n}_x$}}}&z_2\\
\multicolumn{2}{c}{\vdots}&\multicolumn{2}{c}{\ddots}&\multicolumn{2}{c}{\vdots}&\vdots\\
&&&&&&z_{2n-1}\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R^{n1}_x$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$R_x^{nn}$}}}&z_{2n}\\
w_1&w_2&\multicolumn{2}{c}{{\cdots}}&w_{2n-1}&w_{2n}&u
\end{array}\right),
\] where $R^{ij}_x$ takes the form
\[\begin{cases} \left(\begin{array}{cc}x& y \\ -y & x \end{array} \right) & \epsilon_B(i,j)=-1 \\ \left( \begin{array}{cc}z&w \\ w & -z \end{array} \right) & \epsilon_B(i,j) = 1
\end{cases}.
\] Since $\rho(x) \in SO(2n+1)$, this forces $z_1 = \cdots = z_{2n} = w_1 = \cdots = w_{2n} = 0$, and thus $u = 1$. So the sections we are considering all vanish when every $\rho(x)$ is of the form \eqref{minusmat} described in Lemma \ref{genminus}, in which case $\prod_{l=1}^g[\rho(a_l), \rho(b_l)]$ cannot be a generic torus element. But again, $\prod_{l=1}^g[\rho(a_l), \rho(b_l)]= \rho(c) = t$ was chosen to be generic, so these sections have no common zeros.
\end{proof}
Now fix $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$, and let $C = U \times U^c \in \mathcal{B}[[n]]$ be another block (not necessarily distinct from $B$). Consider the sections
\begin{equation}\label{msections}\{s^{+}_{ij}(x) \vert \epsilon_B(i,j) = \epsilon_C(i,j) = -1 \} \cup \{s^{-}_{ij}(x) \mid \epsilon_B(i,j) = 1, \epsilon_c(i,j)=-1\} \cup \{ s^{+}_{ij}(x), s^{-}_{ij}(x) \mid (i,j) \in C\}.
\end{equation} If these sections all vanish, then
\[
\rho(x) = \left(\begin{array}{ccccccc}
&&&&&&z_1\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$T^{11}_x$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$T^{1n}_x$}}}&z_2\\
\multicolumn{2}{c}{\vdots}&\multicolumn{2}{c}{\ddots}&\multicolumn{2}{c}{\vdots}&\vdots\\
&&&&&&z_{2n-1}\\
\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$T^{n1}_x$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$\cdots$}}}&\multicolumn{2}{c}{\smash{\raisebox{.5\normalbaselineskip}{$T_x^{nn}$}}}&z_{2n}\\
w_1&w_2&\multicolumn{2}{c}{{\cdots}}&w_{2n-1}&w_{2n}&u
\end{array}\right),
\] where $T^{ij}_x$ takes the form
\[\begin{cases} \left( \begin{array}{cc}x&-y \\ y & x \end{array}\right) & \epsilon_B(i,j)=\epsilon_C(i,j)=-1 \\ \left( \begin{array}{cc}z&w \\ w & -z \end{array}\right) & \epsilon_B(i,j)=1 \text{ and } \epsilon_C(i,j) = 1 \\ \left( \begin{array}{cc}0&0 \\ 0 & 0\end{array}\right) & \epsilon_C(i,j)=1
\end{cases}.
\] Since $\rho(x) \in SO(2n+1)$, this forces $z_{2i-1}=z_{2i}=0$ for all $i \in U$; thus $\rho(x)$ is block diagonal up to reordering of basis elements, so also $T^{ij}_x= \left( \begin{array}{cc}0&0 \\ 0 & 0\end{array}\right)$ for $(i,j) \in \bar{C}$. Thus again, $z_1 = \cdots = z_{2n} = w_1 = \cdots = w_{2n} = 0$, and $u=1$. Observe that in particular, $\rho(x)$ is in the form \eqref{minusmat} described in Lemma \ref{genminus}, and further, that this form \eqref{minusmat} is independent of $C$. More generally, $\rho(x)$ takes the same form \eqref{minusmat} when the sections obtained from those in \eqref{msections} by replacing the block $C$ with a union of blocks all vanish. If we instead take $B$ to be the empty set, then $\rho(x)$ is in the form \eqref{plusmat} described in Lemma \ref{genplus}. This leads to the following lemma:
\begin{lem}\label{withunions}Let $C \in \mathcal{B}[[n]]\cup\{\varnothing \}$. For $ 1 \leq l \leq 2g$, let $D_l$ be a (possibly empty) union of blocks in $\mathcal{B}[[n]]$. Then the following collection of sections has no common zeros:
\begin{multline}\label{unionsections}
\bigcup_{l=1}^g \{s^+_{ij}(a_l), s^-_{ij}(a_l) \mid (i,j) \in D_l\} \cup \{s^{-\epsilon_C(i,j)}_{ij}(a_l) \mid (i,j) \notin D_l \cup \bar{D_l} \} \\ \cup \bigcup_{l=g+1}^{2g} \{s^+_{ij}(b_{l-g}),s^-_{ij}(b_{L_g}) \mid (i,j) \in D_l \} \cup \{s^{-\epsilon_C(i,j)}_{ij}(b_{l-g}) \mid (i,j) \notin D_l \cup \bar{D_l} \}.
\end{multline}
\end{lem}
\begin{proof} As discussed above, if these sections all vanish then $\rho(x)$ has the form \eqref{minusmat} for every $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$. Thus $\prod_{l=1}^g[\rho(a_l), \rho(b_l)]$ cannot be a generic torus element by Lemma \ref{genminus}. But $\prod_{l=1}^g[\rho(a_l), \rho(b_l)]=\rho(c) = t$ was chosen to be a generic torus element, so these sections have no common zeros.
\end{proof}
\begin{rmk} \label{blockrmk}
In Lemma \ref{genminus}, we proved that if each $\rho(x)$ is of the form \eqref{minusmat}, then $\prod_{l=1}^g[\rho(a_l),\rho(b_l)]$ cannot be a generic torus element. Observe that if we have $2g$ block diagonal matrices $A_1, \ldots, A_g, B_1, \ldots, B_g$, where the first block in each is a $k$-by-$k$ matrix of the form \eqref{minusmat}
$\left(\begin{array}{c|c}\eqref{minusmat}& 0 \\ \hline 0 & *\end{array}\right)$, then their product of commutators $\prod_{l=1}^g[A_l, B_l]$ also cannot be a generic torus element.
\end{rmk}
We will use this idea to find collections of sections with no common zeros by extending collections that worked for lower-rank cases. Note that in the example above, the matrices only need to be block diagonal up to reordering of the basis elements. Thus it will be useful to introduce the following notation:
\begin{notn} Let $G=SO(2n+1)$, let $X$ be a nonempty finite subset of $[n]$, let $B \in \mathcal{B}[X] \cup \{\varnothing\}$, and let $D$ be a $2g$-tuple of unions of blocks in $\mathcal{B}[X]$. Let $x_l= \begin{cases} a_l & 1 \leq l \leq g \\ b_{l-g} & g+1 \leq l \leq 2g \end{cases}$. Define the collection of sections
\[Q_n(X,B,D):=\bigcup_{l=1}^{2g} \{s^+_{ij}(x_l),s^-_{ij}(x_l) \mid (i,j) \in D_l \} \cup \{s^{-\epsilon_B(i,j)}_{ij}(x_l) \mid (i,j) \notin D_l \cup \bar{D_l} \}.
\] \end{notn}
When the sections in $Q_n(X,B,D)$ all vanish, the $\vert X \vert$-by-$\vert X \vert$ submatrix of each $\rho(x)$ induced by considering only the rows and columns $2i-1$ and $2i$ for elements $i$ in $X$ must take the form \eqref{minusmat}. Lemma \ref{withunions} says that the sections in $Q_n([n], B, D)$ have no common zeros.
\begin{defn} Consider sets $\mathcal{A}_n$, for $n \in \mathbb{N}$, defined recursively as follows: \begin{itemize} \item if $B\in \mathcal{B}[[n]]\cup \{\varnothing\}$ and $0 \leq r \leq 2g$, then $Q_n([n],B,D) \in \mathcal{A}_n$. \item if $X \sqcup Y = [n]$ is a partition, $\psi: [ \vert X \vert ] \to X$ is a bijection, and $P \in \mathcal{A}_{\vert X \vert}$, then $\psi_*P \cup \cup_{i \in X, j \in Y} \cup_{x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}} s^{+}_{ij}(x) s^{-}_{ij}(x) \in \mathcal{A}_n$. \end{itemize}
\end{defn}
This induced map $\psi_*$ sends a section $s^\pm_{ij}(x)$ of the line bundle $L^\pm_{ij}$ over $S_{\vert X \vert,g}(t)$ to the section $s^\pm_{\psi(i),\psi(j)}(x)$ of the line bundle $L^\pm_{\psi(i),\psi(j)}$ over $S_{n,g}(t)$.
\begin{prop} If $P \in \mathcal{A}_n$ is a collection of sections in the set just described, then the sections in $P$ have no common zeros.
\end{prop}
\begin{proof}An element $P \in \mathcal{A}_n$ takes the form
\[P = Q_n(X, B, D) \cup \bigcup_{k=1}^d \bigcup_{\substack{i \in X \cup Y_1 \cup \cdots \cup Y_{k-1}\\ j \in Y_k}}\bigcup_{x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}}s^{+}_{ij}(x) s^{-}_{ij}(x),
\] for some $0 \leq k \leq n$. Observe that this collection contains the collection of sections
\[Q_n(X,B,D) \cup \bigcup_{\substack{i \in X \\ j \in Y_1 \cup \cdots \cup y_k}}\bigcup_{x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}}s^{+}_{ij}(x) s^{-}_{ij}(x).
\]
When these sections all vanish, each $\rho(x)$ is (up to reordering) of the form
\[\left(\begin{array}{ccc|ccccccc|c}
&&&&&&&&&&z_1\\
&A_x&&&&&0&&&&\vdots\\
&&&&&&&&&&z_{2\vert X \vert}\\
\hline
&&&&&&&&&&\\
&&&&&&&&&&\\
&*&&&&&*&&&&*\\
&&&&&&&&&&\\
&&&&&&&&&&\\
\end{array}\right),
\] where each $A_x$ is of the form \eqref{minusmat}. But the form of $A_x$ forces $z_{2l-1}=z_{2l}=0$ for each $l \in X$, so each $\rho(x)$ is in fact block diagonal (up to reordering), of the form discussed in Remark \ref{blockrmk}. Hence these sections have no common zeros.
\end{proof}
\begin{cor} \label{goodthingsvanish}Suppose
\[ \bigcup_{\substack{1 \leq i,j \leq n \\ x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}}}(s^{+}_{ij}(x))^{k^+_{ij}(x)}(s^{-}_{ij}(x))^{k^-_{ij}(x)} \in \mathcal{A}_n,
\]where $k^+_{ij}(x)$ and $k^-_{ij}(x) \in \mathbb{N}$ for $1 \leq i, j \leq n$ and $x \in \{a_1, \ldots, a_g, b_1, \ldots, b_g\}$, and $k^-_{ij}(x)=0$ whenever $i=j$. Then the cohomology class
\[\prod_{1 \leq i, j \leq n}c_1(L^+_{ij})^{\sum_x k^+_{ij}(x)}c_1(L^-_{ij})^{\sum_xk^-_{ij}(x)}
\] vanishes in $H^*(S_{n,g}(t);\mathbb{Q})$. \qed \end{cor}
\section{The combinatorics} \label{heart}
So far we have found a class of ``good'' products of the $c_1(L^\pm_{ij})$ which vanish. The rest of this paper is devoted to proving that any product of at least $2gn^2 + \frac{1}{2}(n-1)(n-2)$ of the $c_1(L^\pm_{ij})$ is equivalent in $H^*(S_{n,g}(t);\mathbb{Q})$ to a combination of these ``good'' products, and hence also vanishes.
The proof is combinatorial, and makes extensive use of the relations
\begin{multline} \label{relns}
c^+_{ij} = c^+_{ji} \\
c^-_{ij} =-c^-_{ji} \\
c^+_{ij} -c^+_{jk}+c^-_{ki}=0.\\
\end{multline}
For the terms appearing in these relations to be defined, we must have fixed the rank $n$, but the same relations hold no matter which $n$ is chosen. In order to use inductive arguments we wish to be able to make statements that don't rely on having fixed the rank. We thus move away from considering the Chern classes themselves and instead study the combinatorial properties of another ring $R$ whose elements satisfy the same relations \eqref{relns} as our Chern classes.
\begin{defn} Let $X$ be a finite subset of $\mathbb{N}$. Define the associated auxiliary sets \begin{itemize}
\item $Y^+(X) := \{y^+_{ij} \mid i,j \in X\}$ \item $Y^-(X):= \{ y^-_{ij} \mid i, j \in X; i \neq j\}$ \item $Y(X):= Y^+(X) \sqcup Y^-(X)$,\end{itemize} as well as subsets $Y_z(X):=\{y^\pm_{ij}\in Y(X) \mid i = z \text{ or } j = z\}$ for each $x \in X$.
If $B \in \mathcal{B}[X]$ is a block as defined in Definition \ref{block}, then we also introduce
\begin{itemize} \item $Y^+_B(X):=\{y^{-\epsilon_B(i,j)}_{ij} \mid i, j\in X\}$ \item $Y^-_B(X):= \{y^{\epsilon_B(i,j)}_{ij} \mid i, j \in X ; i \neq j\}$
\end{itemize}
\end{defn}
We would like to form a quotient of $\mathbb{Q}[Y(X)]$ by relations corresponding to those satisfied by the $c_1(L^\pm_{ij})$ listed above \eqref{relns}.
\begin{defn}
Let $I \subset \mathbb{Q}[Y(X)]$ be the ideal generated by the elements
\begin{itemize}
\item $y^-_{ij} + y^-_{ji}$ \item $y^+_{ij}- y^+_{ji}$ \item $y^-_{ij} + y^-_{jk} + y^-_{ki}$ \item $y^+_{ij} -y^+_{jk} + y^-_{ki}$
\end{itemize} for all triples of elements $i,j,k \in X$. Let $R:= \mathbb{Q}[Y(X)]/I$ be the quotient of $\mathbb{Q}[Y(X)]$ by this ideal; if $p \in \mathbb{Q}[Y(X)]$ we will denote by $[p]$ its image in $R$. Note that this quotient preserves the grading by degree of $\mathbb{Q}[Y(X)]$.
\end{defn}
In the previous section we found sets $\mathcal{A}_n$ of collections of sections with no common zeros, and concluded that the corresponding products of Chern classes vanished in $H^*(S_{n,g}(t))$. We wish now to work in the ring $R$ and not in $H^*(S_{n,g}(t))$, so we introduce the map
\begin{align*}
\alpha: \cup_n \{\text{sections } s^{\pm}_{ij}(x) \text{ of }L^\pm_{ij} \to S_{n,g} \} & \to \mathbb{Q}[Y(X)] \\ s^{\pm}_{ij}(x) & \mapsto y^\pm_{ij}.
\end{align*}
Note that $\alpha$ is a map of sets, not a ring homomorphism; we will use the same notation for the map that takes a set of sections to the product of their images in $\mathbb{Q}[Y(X)]$. Observe that $\alpha$ forgets both $n$ and $x$.
The main goal of the remainder of this paper is to prove that for any monomial $q \in \mathbb{Q}[Y(X)]$ of degree at least $2gn^2+\frac{1}{2}(n-1)(n-2)$, there exist some $q_i \in \mathcal{A}_n$ and monomials $\theta_i \in \mathbb{Q}[Y(X)]$ such that $[q]=[\sum_i \theta_i \alpha(q_i)]$ in $R$. Recall that the sets $\mathcal{A}_n$ were defined recursively. We make this more explicit in the statement of the proposition:
\begin{prop} Let $X \subset \mathbb{N}$ with $\vert X \vert = n$. Let $p \in \mathbb{Q}[Y(X)]$ be a polynomial of degree at least $2gn^2+\frac{1}{2}(n-1)(n-2)$. Then for each $B = V \times V^c \in \mathcal{B}[X]$ and $C \in \mathcal{B}[X]\cup \{\varnothing\}$, we can find:
\begin{itemize} \item a finite set $\mathcal{D}$ whose elements are $2g$-tuples of unions of blocks in $\mathcal{B}[X]$ \item homogeneous polynomials $\theta_B, \phi_B, \psi_C \in \mathbb{Q}[Y(X)]$, as well as polynomials $\chi_D$ for each $D \in \mathcal{D}$ \item elements $p_V \in \mathcal{A}_{\vert V \vert}$ and $p_{V^c} \in \mathcal{A}_{\vert V^c \vert}$, and \item bijections $f_V:[\vert V \vert] \to V$ and $f_{V^c}:[\vert V^c \vert ] \to V^c$ such that\end{itemize}
\begin{multline*}[p] = [\sum_{B=V\times V^c \in \mathcal{B}[X]}\prod_{(i,j) \in B}(y^+_{ij} y^-_{ij})^{2g}(\theta_B\cdot\alpha((f_V)_*(p_V))+\phi_B\cdot \alpha((f_{V^c})_*(p_{V^c}))) \\ + \sum_{C \in \mathcal{B}[X] \cup \{\varnothing\}}\psi_C \sum_{D \in \mathcal{D}}\chi_D \prod_{l=1}^{2g}\left(\prod_{(i,j) \in D_l}y^+_{ij} y^-_{ij}\right)\left(\prod_{(i,j) \notin D_l \cup \bar{D_l}}y^{-\epsilon_C(i,j)}_{ij}\right)]
\end{multline*}
\end{prop}
There are several points during the course of the proof of this proposition when we apply the pigeonhole principle, generally in order to show that the restriction of a polynomial to some subring of $\mathbb{Q}[Y(X)]$ has high enough degree to apply an inductive hypothesis. In order to reduce clutter we collect the relevant calculations into the following five lemmas.
\begin{lem} \label{oddsomez} Let $X$ be a finite subset of $\mathbb{N}$ with $\vert X \vert = m > 3$, and let $p \in \mathbb{Q}[Y(X)]$ be a monomial of degree at least $2gm(m-1)-m+1$. Then there exists some $z \in X$ such that if we factorise $p$ as $p=q_zr_z$, where $q_z \in \mathbb{Q}[Y(X \setminus \{z\})$ and $r_z \in \mathbb{Q}[Y_z(X)]$ are monomials, then $q_z$ has degree at least $2g(m-1)(m-2)-m+2$.
\end{lem}
\begin{proof} Write
\[p=\lambda \prod_{i,j \in X }(y^+_{ij})^{d^+_{ij}}(y^-_{ij})^{d^-_{ij}}\]
(where $d^-_{ii}=0$ for all $i$). Given $z \in X$,
\[q_z = \prod_{i,j \in X \setminus \{z\} }(y^+_{ij})^{d^+_{ij}}(y^-_{ij})^{d^-_{ij}}.
\]
Note that each factor $y^+_{ij}$ of $p$ appears in $q_z$ precisely when $i, j \neq z$, and thus appears in at least $m-2$ of the $q_z$ as $z$ ranges over $X$ (exactly $m-2$ except for the $y^+_{ii}$ which appear in $m-1$). Hence $\prod_{z=1}^mq_z=p^{m-2}\prod_{i \in X \setminus \{z\}}(y^+_{ii})^{d^+_{ii}}$ has degree $\geq (m-2)(2gm(m-1)-m+1)$. Suppose by way of contradiction that each $q_z$ has degree at most $2g(m-1)(m-2)-m+1$. Then $\prod_{z=1}^mq_z$ has degree at most
\begin{align*} m(2g(m-1)(m-2)-m+1) & = 2gm(m-1)(m-2)-m+1\\ &< (m-2) 2gm(m-1)-(m-2)(m-1) \\&=(m-2)(2g(m-1)-m+1).
\end{align*} This is a contradiction, so the desired $z$ must exist.
\end{proof}
\begin{lem} \label{indind}
Let $X = \{e_1, \ldots, e_h, f_1, \ldots, f_w, z\}$ be a subset of $\mathbb{N}$ with $\vert X \vert = m$. Let $p \in \mathbb{Q}[Y_z(X)]$ be a monomial of degree at least $2gm(m-1)-m+1-4gwh$ that factorises as $p = p_h p_w$, where $p_h \in \mathbb{Q}[Y(\{e_1, \ldots, e_h, z\})]$ and $p_w \in \mathbb{Q}[Y(\{f_1, \ldots, f_w, z\})]$ are monomials. Then either $\deg p_h \geq 2gh(h+1)-h$, or $\deg p_w \geq 2gw(w+1)-w $.
\end{lem}
\begin{proof}Suppose $\deg p_h \leq 2gh(h+1)-h-1$. Then
\begin{align*} \deg p_w &\geq 2gm(m-1)-m+1-4gwh-2gh(h+1)+h+1 \\ &= 2gm(w+h)-(w+h)-4gwh -2gh(m-w)+(m-w) \\ &=2gw(m+h)-4gwh-w-h+m-w \\&=2gw(w+h+1+h)-4gwh-w+1 \\&= 2gw(w+1)-w+1 > 2gw(w+1)-w.\qedhere
\end{align*}
\end{proof}
\begin{lem} \label{bozsqind} Let $X = H \sqcup W$ be a finite subset of $\mathbb{N}$ with $\vert H \vert = h$, $\vert W \vert = w$, and $\vert X \vert = w + h = n$. Let $p \in \mathbb{Q}[Y(X)]$ be a monomial of degree at least $2gn^2+ \frac{1}{2}(n-1)(n-2)-4gwh$ that factorises as $p = p_w p_h$, where $p_w \in \mathbb{Q}[Y(W)]$ and $p_h \in \mathbb{Q}[Y(H)]$. Then either $\deg p_w \geq 2gw^2+ \frac{1}{2}(w-1)(w-2)$ or $\deg p_h \geq 2gh^2+ \frac{1}{2}(h-1)(h-2)$.
\end{lem}
\begin{proof} Suppose $\deg p_h \leq 2gh^2 + \frac{1}{2}(h-1)(h-2)-1$. Then
\begin{align*}\deg p_w &\geq 2gn^2+ \frac{1}{2}(n-1)(n-2)-4gwh-2gh^2-\frac{1}{2}(h-1)(h-2) +1 \\ &= 2g(h^2+2wh + w^2) + \frac{1}{2}(w+h-1)(w+h-2)-4gwh-2gh^2-\frac{1}{2}(h-1)(h-2)+1 \\ &=2gw^2+ \frac{1}{2}(h-1)(h-2)+ \frac{1}{2}w(h-2)+ \frac{1}{2}w(h-1)+\frac{w^2}{2}-\frac{1}{2}(h-1)(h-2)+1 \\
&= 2gw^2+ \frac{1}{2}(w^2-2w-w+2)+wh \\ &= 2gw^2 + \frac{1}{2} (w-1)(w-2)+wh > 2gw^2+ \frac{1}{2}(w-1)(w-2). \qedhere
\end{align*}
\end{proof}
\begin{lem} \label{2goroct} Let $p=qr$ be a monomial of degree at least $2gn^2+ \frac{1}{2}(n-1)(n-2)-n(n-1)g-2g(n-1)$. Then either $\deg q \geq 2g$ or $\deg r \geq n(n-1)g-n+2$.
\end{lem}
\begin{proof} Suppose $\deg r \leq n(n-1)g-n+1$. Then
\begin{align*} \deg q & \geq 2gn^2 + \frac{1}{2}(n-1)(n-2)-n(n-1)g-2g(n-1)-n(n-1)g+n-1 \\ &=2gn^2 -2gn(n-1)-2g(n-1)+ \frac{1}{2}(n-1)(n-2)+n-1 \\ &=2g + \frac{1}{2}n(n-1) \geq 2g. \qedhere
\end{align*}
\end{proof}
\begin{lem} \label{indoroct} Let $w, h \in \mathbb{N}$ with $w+h=m-1$, and let $p=qr$ be a monomial of degree at least $2gm(m-1)-m+1-4gwh-h(h+1)g$. Then either $\deg q \geq 2gw(w+1)-w$ or $\deg r \geq h(h+1)g-(h+1)+2$.
\end{lem}
\begin{proof} Suppose $\deg r \leq h(h+1)g-(h+1)+1$. Then \begin{align*} \deg q & \geq 2gm(m-1)-m+1-4gwh-h(h+1)g-h(h+1)g+(h+1)-1 \\ &= 2g(w+h+1)(w+h)-(w+h)-4gwh-2gh(h+1)+h \\ &= 2gw^2 + 2gh(h+1) + 2gw -w -2gh(h+1) \\ &= 2gw(w+1)-w . \qedhere
\end{align*}
\end{proof}
We will also need the following results about interactions between different blocks.
\begin{lem} \label{cd1sbextends} Suppose $X = \{e_1, \ldots, e_h, f_1, \ldots, f_w, z\}$. Let $B \in \mathcal{B}[X\setminus \{z\}]$ be the block $B = \{e_1, \ldots, e_h\} \times \{f_1, \ldots, f_w\}$, and let $C$ be a block in $\mathcal{B}[\{f_1, \ldots, f_w,z\}]$. Then the union $B \cup C \cup \bar{C}$ contains a block $D \in \mathcal{B}[X]$.
\end{lem}
\begin{proof} Suppose without loss of generality $C = \{f_1, \ldots, f_d\} \times \{f_{d+1}, \ldots, f_w, z\}$, for some $1 \leq d \leq w$. Then $D = \{ e_1, \ldots, e_h, f_1, \ldots, f_d\} \times \{f_{d+1}, \ldots, f_w, z\} \subset B \cup C \cup \bar{C}$.
\end{proof}
\begin{lem} \label{ebfbextendinb} Again, suppose $X = \{e_1, \ldots, e_h, f_1, \ldots, f_w, z\}$, and let $B=\{e_1, \ldots, e_h\} \times \{f_1, \ldots, f_w\} \in \mathcal{B}[X\setminus \{z\}]$. Let $C \in \mathcal{B}[\{e_1, \ldots, e_h, z\}]$ and $E \in \mathcal{B}[\{f_1, \ldots, f_w, z\}]$. Then there exists a block $A_{CE} \in \mathcal{B}[X]$ such that either $C\cup E \subset A_{CE} \subset C \cup E \cup B \cup \bar{B}$, or $C \cup \bar{E} \subset A_{CE} \subset C \cup \bar{E} \cup B \cup \bar{B}$.
\end{lem}
\begin{proof}Suppose without loss of generality $C = \{e_1, \ldots, e_k\} \times \{e_{k+1}, \ldots, e_h, z\}$, for some $1 \leq k \leq h$. Either $E$ or $\bar{E}$ has the form $\{f_1, \ldots, f_l \} \times \{f_{l+1}, \ldots, f_w, z\}$ (up to relabelling), so take $A_{CE}= \{ e_1, \ldots, e_k, f_1, \ldots, f_l\} \times \{e_{k+1}, \ldots, e_h, f_{l+1}, \ldots, f_w, z\}$.
\end{proof}
\begin{lem}[The symmetric difference of two blocks is a block] \label{symdiff} Let $B, C \in \mathcal{B}[X]$. Then there exists a block $D \in \mathcal{B}[X]$ such that $(B\cup \bar{B}) \triangle (C \cup \bar{C})=D \cup \bar{D}$.
\end{lem}
\begin{proof}
Suppose $B = \{e_1, \ldots, e_h\} \times \{f_1, \ldots, f_w\}$, and
\[
C = \{e_1, \ldots, e_s, f_1, \ldots, f_t\} \times \{e_{s+1}, \ldots, e_h, f_{t+1}, \ldots, f_w\}.
\] Then take $D = \{e_1, \ldots, e_s, f_{t+1}, \ldots, f_w\} \times \{e_{s+1}, \ldots, e_h, f_1, \ldots, f_t\}$.
\end{proof}
\begin{rmk} \label{restrictblock} Let $B \in \mathcal{B}[X]$ and $z \in X$. Then $B \vert_{X \setminus \{z\}}:=B \cap ((X \setminus \{z\}) \times (X \setminus \{z\})) \in \mathcal{B}[X\setminus \{z\}]$. \end{rmk}
We are now ready to study the quotient ring $R$. We begin by observing that the polynomial ring obtained by adjoining the elements in the subset $Y^-(X)\subset Y(X)$ to $\mathbb{Q}$ is isomorphic to the polynomial ring introduced in \cite{aj} to study the case $G=SU(n)$; we can therefore use the following result from our earlier work:
\begin{lem} \label{propfromaj} Let $X$ be a finite set. Let $p \in \mathbb{Q}[Y^-(X)]$ be a monomial of degree at least $\frac{1}{2}\vert X \vert(\vert X \vert -1)a-\vert X \vert +2$, for some $a \in \mathbb{N}$. Then for each $B \in \mathcal{B}[X]$ we can find a monomial $\phi_B \in \mathbb{Q}[Y^-(X)]$ such that
\[[p] = \left[\sum_{B \in \mathcal{B}[X]}\phi_B\prod_{(i,j) \in B}(y^-_{ij})^a\right].
\]
\end{lem}
\begin{proof} This follows from replacing $2g$ by $a$ in the proof of Proposition 3.6 in \cite{aj}. \end{proof}
\begin{rmk} Let $X$ be a finite set and let $B=V \times V^c$ be a block in $\mathcal{B}[X]$. Consider the isomorphism $f_B:\mathbb{Q}[Y(X)] \to \mathbb{Q}[Y(X)]$ given by
\[f_B:y^\pm_{ij} \mapsto \begin{cases} - y^\pm_{ij} & i, j \in V \\ y^\pm_{ij} & i, j \notin V \\ -y^\mp_{ij} & (i,j) \in B \\ y^\mp_{ij} & (i,j) \in \bar{B} \end{cases}.
\] This map restricts to an isomorphism $f_B:\mathbb{Q}[Y^-(X)] \to \mathbb{Q}[Y^-_B(X)]$; it is straightfoward to check that $f_B$ descends to an isomorphism on the quotient $R$.\end{rmk}
\begin{cor} \label{ajpropflipblock} Let $X$ be a finite set, let $B=V\times V^c \in \mathcal{B}[X]$ be a block, let $a \in \mathbb{N}$, and let $p \in \mathbb{Q}[Y^-_B(X)]$ be a monomial of degree at least $\frac{1}{2}\vert X \vert (\vert X \vert -1)a - \vert X \vert +2$. Then for each $C \in \mathcal{B}[X]$ we can find a monomial $\phi_C \in \mathbb{Q}[Y^-_B(X)]$ such that
\[[p]= \left[ \sum_{C \in \mathcal{B}[X]}\phi_C \prod_{(i,j) \in C}(y_{ij}^{\epsilon_B(i,j)})^a\right].
\]
\end{cor}
\begin{proof}
Consider the isomorphism $f_B:\mathbb{Q}[Y^-(X)] \to \mathbb{Q}[Y^-_B(X)]$ defined above. Applying Lemma \ref{propfromaj} to $f_B^{-1}(p)$, we can write
\[[f^{-1}_B(p)]=\left[\sum_{C \in \mathcal{B}[X]}\phi_C\prod_{(i,j) \in C}(y^-_{ij})^a\right].
\]
Thus
\[ [p]= \left[ \sum_{C \in \mathcal{B}[X]}f_B(\phi_C) \prod_{(i,j) \in C}(-1)^{\lambda_C}(y_{ij}^{\epsilon_B(i,j)})^a \right],
\] where $\lambda_C = \vert \{(i,j) \in C \mid i \in V\} \vert.$
\end{proof}
\begin{lem} \label{zroworoct} Let $X$ be a finite subset of $\mathbb{N}$ with $\vert X \vert \geq 2$, and let $B=V\times V^c \in \mathcal{B}[X]$. Let $z \in V^c$. Let $p \in \mathbb{Q}[Y(X)]$ be a homogeneous polynomial of degree at least $2gm(m-1)-m+1-(m-1)(m-2)g$. Then $[p]$ has a representative in $\mathbb{Q}[Y^-_B(X) \cup \{ y^{-\epsilon_V(i)}_{iz} \mid i \in X \setminus \{z\}\}]$; furthermore, this representative can be chosen to be a sum of monomials, each of which, when factorised as $qr$ with $q \in \mathbb{Q}[Y^-_B(X)]$ and $r \in \mathbb{Q}[\{y^{-\epsilon_V(i)}_{iz}\mid i \in X \setminus \{z\}]$, either satisfies $\deg q \geq m(m-1)g-m+2$, or $r=\theta \prod_{i \in X \setminus \{z\}}(y^{-\epsilon_V(i)}_{iz})^{2g}$ for some monomial $\theta$. \end{lem}
\begin{proof} We first claim that for any $ i \in X \setminus \{z\}$, the element $[y^{-\epsilon_V(i)}_{iz}]$ together with the images in $R$ of the elements of $Y^-_B(X)$ form a set of generators for $R$. To see this, fix $i \in X \setminus \{z\}$. Since $z \in V^c$, we know that $y^{\epsilon_V(i)}_{iz} \in Y^-_B(X)$, and so both $y^+_{iz}$ and $y^-_{iz}$ are in $Y^-_B(X) \cup \{ y^{-\epsilon_V(i)}_{iz}\}$. Observe that $[y^+_{ij} + y^-_{ij}]=[y^+_{iz}+y^-_{iz}]$, and for any $j \in X \setminus \{z, i\}$, either $y^\pm_{ij} \in Y^-_B(X)$ or $y^-_{ij} \in Y^-_B(X)$. Thus both $[y^+_{ij}]$ and $[y^-_{ij}\}$ are elements of $\langle [Y^-_B(X)], [y^{-\epsilon_V(i)}_{iz}] \rangle \subset R$. Similarly, given any $l\neq j \in X$, either $y^+_{lj} \in Y^-_B(X)$ or $y^-_{lj} \in Y^-_B(X)$. But $[y^+_{lj}+y^-_{lj}]=[y^+_{ij} + y^-_{ij}]$, so $[y^+_{lj}]$ and $[y^-_{lj}]$ are both in $\langle [Y^-_B(X) ], [y^{-\epsilon_V(i)}_{iz}]\rangle $. Finally, observe that $[y^+_{ll}]=\frac{1}{2}[y^+_{lj}+y^-_{lj}]$, so each $[y^+_{ll}]$ is also in $\langle [Y^-_B(X)],[y^{-\epsilon_V(i)}_{iz}]\rangle$.
Without loss of generality, suppose $X=[m]$ and $z=m$. Pick a representative for $[p]$ in $\mathbb{Q}[Y^-_B(X) \cup \{y^{-\epsilon_V(1)}_{1z}\}]$, and factorise each term as $[q_1(y^{-\epsilon_V(1)}_{1z})^{d_1}]$, with $q_1 \in \mathbb{Q}[Y^-_B(X)]$. If $\deg q_1$ is at least $m(m-1)g-m+2$, this term has the desired form; else
\begin{align*}
d_1 &\geq 2gm(m-1)-m+1-(m-1)(m-2)g-m(m-1)g+m-1 \\ & = 2g(m-1) \\ &\geq 2g.
\end{align*} In this case, pick a representative for $[q_1(y^{-\epsilon_V(1)}_{1z})^{d_1}]$ that is a sum of terms of the form $(y^{-\epsilon_V(1)}_{1z})^{2g}q_2(y^{-\epsilon_V(2)}_{2z})^{d_2}$, where $q_2 \in \mathbb{Q}[Y^-_B(X)]$, and consider each term separately. Any term where $\deg q_2 \geq m(m-1)g-m+2$ now has the desired form; else $d_2 \geq 2g(m-1)-2g=2g(m-2)\geq 2g$. Repeat this process: for each term, either $\deg q_i \geq m(m-1)g-m+2$, or $d_i \geq 2g(m-i)$. Thus any term with $\deg q_{m-1} < m(m-1)g-m+2$ has $d_i \geq 2g \forall 1 \leq i \leq m-1$, and so is a multiple of $\prod_{i \in X \setminus \{z\}}(y^{-\epsilon_V(i)}_{iz})^{2g}$.
\end{proof}
\begin{lem} \label{monster} Let $X$ be a finite subset of $\mathbb{N}$ with $\vert X \vert \geq 2$, and suppose $p \in \mathbb{Q}[Y(X)]$ is a monomial of degree at least $2g\vert X \vert (\vert X \vert -1)-\vert X \vert +1$. Then we can find a homogeneous polynomial $\chi_\varnothing \in \mathbb{Q}[Y(X)]$, together with polynomials $\psi_B, \chi_B \in \mathbb{Q}[Y(X)]$ for each block $B \in \mathcal{B}[X]$, such that
\[
[p]=\left[ \sum_{B \in \mathcal{B}[X]} \psi_B \prod_{(i,j) \in B}(y^+_{ij} y^-_{ij})^{2g} + \sum_{B \in \mathcal{B}[X] \cup \{\varnothing\}}\chi_B \prod_{(i,j) \in B}(y^-_{ij})^{2g}\prod_{\substack{(i,j) \notin B \cup \bar{B} \\ i < j}}(y^+_{ij})^{2g}\right].
\]
\end{lem}
\begin{ill}
Our goal is to use the relations between the $c_1(L^\pm_{ij})$ to express any monomial $[p] \in R$ as a sum of images of ``good'' collections of sections defined at the end of Section \ref{purple}.While we are working in $R$, we will illustrate monomials appearing in our argument with pictures of the form $\rho(x)$ takes in the vanishing loci of the corresponding sections $s^{\pm}_{ij}(x)$. Though $\rho(x) \in SO(2n+1)$, our illustrations will be $n\times n$ grids, in which we represent $2 \times 2$ blocks of the form $\left(\begin{array}{cc}x&-y\\y&x\end{array}\right)$ by \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture}, the $2 \times 2$ blocks of the form $\left(\begin{array}{cc}z&w\\w&-z\end{array}\right)$ by \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture}, and the $2 \times 2$ blocks of zeros (in the vanishing loci of $(s^+_{ij}(x), s^-_{ij}(x)$) by \begin{tikzpicture} \node[draw, circle, violet] at (0,0){$0$}; \end{tikzpicture}. We have chosen to illustrate our argument using blocks $B$ with the property that $i<j \forall (i,j) \in B$, since these produce the simplest visualisations; note that this is not an assumption we make in the proof. Lemma \ref{monster} says we can express any monomial of degree $\geq 2gn(n-1)-n+1$ as a sum of terms whose corresponding sections have vanishing loci taking one of the forms shown in Figure \ref{monsterstatement}.
\begin{figure}
\begin{subfigure}{.45\linewidth}
\centering
\begin{tikzpicture}[scale=1]
\draw [blue](4-0.3,3-0.3) rectangle (4+0.3,3+0.3);
\draw [blue](3-0.3,3-0.3) rectangle (3+0.3,3+0.3);
\draw [blue](2-0.3,3-0.3) rectangle (2+0.3,3+0.3);
\draw [red](1-0.3,4-0.3) rectangle (1+0.3,4+0.3);
\draw [red](3-0.3,2-0.3) rectangle (3+0.3,2+0.3);
\draw [red](4-0.3,2-0.3) rectangle (4+0.3,2+0.3);
\draw [red](4-0.3,1-0.3) rectangle (4+0.3,1+0.3);
\draw [fill=gray](0,4) circle [radius=0.1];
\draw [blue](3-0.3,4-0.3) rectangle (3+0.3,4+0.3);
\draw [blue](4-0.3,4-0.3) rectangle (4+0.3,4+0.3);
\draw [blue](2-0.3,4-0.3) rectangle (2+0.3,4+0.3);
\draw [fill=gray](2,2) circle [radius=0.1];
\draw [fill=gray](1,3) circle [radius=0.1];
\draw [fill=gray](3,1) circle [radius=0.1];
\draw [fill=gray](4,0) circle [radius=0.1];
\node [] at (2,3) {$-$};
\node [] at (4,3) {$-$};
\node [] at (3,3) {$-$};
\node [] at (2,4) {$-$};
\node [] at (4,4) {$-$};
\node [] at (3,4) {$-$};
\node [] at (1,4) {$+$};
\node [] at (3,2) {$+$};
\node [] at (4,2) {$+$};
\node [] at (4,1) {$+$};
\foreach \x in {0, ..., 3}
\foreach \y in {\x, ..., 3}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\end{tikzpicture}
\caption{A block of \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture}, otherwise \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture}, above the diagonal. (Or all \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture}, if $B = \varnothing$)} \label{fig:M1}
\end{subfigure}%
\hfill
\begin{subfigure}{.45\linewidth}
\centering
\begin{tikzpicture}[scale=1]
\foreach \x in {0, ..., 4}
{\draw [fill=gray] (\x, 4-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 3}
\foreach \y in {\x, ..., 3}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {2,...,4}
\foreach \y in {3,4}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\end{tikzpicture}
\caption{A block of zeros. Note that this alone does not make $\rho(x) \in SO(2n+1)$ block diagonal, since this is only a $2h$-by-$2(n-h)$ block of zeros.} \label{fig:M2}
\end{subfigure}
\caption{}
\label{monsterstatement}
\end{figure}
\end{ill}
\begin{proof} By induction on $\vert X \vert$.
For $\vert X \vert = 2$, without loss of generality take $X = \{1,2\}$. Then
\[
\mathbb{Q}[Y(X)]=\mathbb{Q}[y^+_{12},y^-_{12},y^+_{21} ,y^-_{21}].
\]
Since $[y^+_{12}]=[y^+_{21}]$ and $[y^-_{12}]=[-y^-_{21}]$ in $R$, any monomial $p \in \mathbb{Q}[Y(X)]$ of degree at least $4g-1$ is equivalent in $R$ to $\lambda(y^+_{12})^a(y^-_{12})^b$, where $\lambda \in \mathbb{Q}$ and $a+b \geq 4g-1$. So either $a \geq 2g$ or $b \geq 2g$. If $a \geq 2g$, take $\chi_\varnothing = (y^+_{12})^{a-2g}(y^-_{12})^b$, and $\psi_B=\chi_B=0$ for all $B \in \mathcal{B}[X]$. If $b \geq 2g$, take $\chi_{(1,2)}=(y^+_{12})^{a}(y^-_{12})^{b-2g}$, and $\psi_B=\chi_\varnothing =0$ for all $B \in \mathcal{B}[X]$. (\checkmark)
Now suppose $\vert X \vert = m \geq 3$. By Lemma \ref{oddsomez}, there is some $z \in X$ for which when we factorise $p$ as $q_z r_z$, with monomials $q_z \in \mathbb{Q}[Y(X \setminus \{z\})$ and $r_z \in \mathbb{Q}[Y_z(X)]$, the degree of $q_z$ is at least $2g(m-1)(m-2)-m+2$. By the inductive hypothesis, there exist homogeneous polynomials $\tilde{\psi_C}, \tilde{\chi_C}, \tilde{\chi_\varnothing} \in \mathbb{Q}[Y(X \setminus \{z\})$, for all $C \in \mathcal{B}[X\setminus \{z\}]$, such that
\[
[q_z]= \left[ \sum_{C \in \mathcal{B}[X\setminus \{z\}]}\tilde{\psi_C}\prod_{(i,j) \in C}(y^+_{ij} y^-_{ij} )^{2g} + \sum_{C \in \mathcal{B}[X\setminus \{z\}] \cup \{ \varnothing \}} \tilde{\chi_C} \prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{(i,j) \in X \times X \setminus C \cup \bar{C} \\ i < j }}\right].
\]
We will consider these terms separately, as it suffices to show each term has the desired form.
Fix a block $C \in \mathcal{B}[X\setminus \{z\}]$, say $C = \{e_1, \ldots, e_h\} \times \{f_1, \ldots, f_w\}$ (so $h+w=m-1$). Consider the term
\begin{equation} \label{codim1zeros}
\left[ r_z \tilde{\psi_C}\prod_{(i,j) \in C}(y^+_{ij} y^-_{ij})^{2g}\right]
\end{equation} in $[p]$ (see Figure \ref{codim1boz}).
\begin{figure}
\begin{center}
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \y in {4,5}
{\draw [fill=gray](2,\y) circle [radius=0.06];
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node [] at (2,5.6) {$z$};
\node [] at (5.6,3) {$z$};
\node [] at (3,5.6) {$f_1$};
\node [] at (5,5.6) {$f_w$};
\node [] at (5.6,5) {$e_1$};
\node [] at (5.6,4) {$e_h$};
\draw [green] (1.7,-0.3) rectangle (2.3,5.3);
\draw [green] (-0.3,2.7) rectangle (5.3,3.3);
\end{tikzpicture}
\end{center}
\caption{The inductive hypothesis gave us a block of zeros in $\mathcal{B}[X\setminus \{z\}]$, using a monomial of degree $4gwh$}
\label{codim1boz}
\end{figure}
Observe that $\deg r_z \tilde{\psi_C} \geq 2gm(m-1)-m+1-4gwh$. Since $[y^+_{ij}]=[y^+_{iz} + y^{-}_{zj}]$ and $[y^-_{ij}] = [y^-_{iz} + y^-_{zj}]$, each term of $r_z \tilde{\psi_C}$ can be expressed as a sum of terms of the form $[p_h p_w]$, where $p_h \in \mathbb{Q}[Y(\{ e_1, \ldots, e_h, z\})]$ and $p_y \in \mathbb{Q}[Y(\{f_1, \ldots, f_w, z\})]$ are monomials. By Lemma \ref{indind}, either $\deg p_h \geq 2gh(h+1)-h$, or $\deg p_w \geq 2gw(w+1)-w$ (see figure \ref{codim1bozwtriangles}).
\begin{figure}
\begin{center}
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \y in {4,5}
{\draw [fill=gray](2,\y) circle [radius=0.06];
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node [] at (2,5.6) {$z$};
\node [] at (5.6,3) {$z$};
\node [] at (3,5.6) {$f_1$};
\node [] at (5,5.6) {$f_w$};
\node [] at (5.6,5) {$e_1$};
\node [] at (5.6,4) {$e_h$};
\draw [green, line width=1mm](-0.6,5.3) -- (2.3,5.3) -- (2.3,2.4)-- (-0.6,5.3);
\node [green] at (1,5.6) {$T_h$};
\draw [cyan, line width=1mm](1.4,3.3) -- (5.4,3.3) -- (5.4,-0.6)-- (1.4,3.3);
\node [cyan] at (5.8,1.6) {$T_w$};
\draw [green] (1.7,-0.3) rectangle (2.3,5.3);
\draw [green] (-0.3,2.7) rectangle (5.3,3.3);
\end{tikzpicture}
\end{center}
\caption{We can apply the inductive hypothesis to one of these triangles $T_h$, $T_w$}
\label{codim1bozwtriangles}
\end{figure}
Consider one such monomial in the sum, and without loss of generality assume the former. By the inductive hypothesis, we can find homogeneous polynomials $\psi_D, \chi_D$ for each $D \in \mathcal{B}[\{e_1, \ldots, e_h, z\}]$, together with $\chi_\varnothing$, with $\psi_D, \chi_D, \chi_\varnothing \in \mathbb{Q}[Y(\{e_1, \ldots, e_h, z\})]$, such that
\[
[p_h] = \left[ \sum_{D \in \mathcal{B}[\{e_1, \ldots, e_h, z\}]} \psi_D \prod_{(i,j) \in D}(y^+_{ij} y^-_{ij})^{2g} + \sum_{D \in \mathcal{B}[\{e_1, \ldots, e_h, z\}] \cup \{\varnothing \}}\chi_D \prod_{(i,j) \in D} (y^-_{ij})^{2g} \prod_{\substack{i,j \in \{e_1, \ldots, e_h,z\} \\ (i,j) \notin D \cup \bar{D} \\ i < j}} (y^+_{ij})^{2g} \right]
\] (see figure \ref{indhypto1tri}).
\begin{figure}
\centering
\begin{subfigure}{.45\linewidth}
\begin{tikzpicture}[scale=1]
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\draw [violet, line width=0.7mm] (2.5,3.5) rectangle (5.5,5.5);
\node [violet] at (6.2,4.5) {$C$};
\foreach \y in {4}
{\draw [fill=gray](2,\y) circle [radius=0.06];
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\draw [green, line width=0.5mm](-0.6,5.3) -- (2.3,5.3) -- (2.3,2.4)-- (-0.6,5.3);
\node [green] at (-0.5,5.8) {$T_h$};
\foreach \x in {1,2}
{\node[draw, circle, line width=0.7mm, olive] at (\x, 5){$0$};
}
\draw [olive, line width=0.7mm] (0.6,4.6) rectangle (2.4,5.4);
\node [olive] at (1.5,5.8) {$D$};
\draw [blue, line width=0.7mm] (0.5,4.5) rectangle (5.5,5.5);
\node [blue] at (3.8,5.8) {$B_{CD}$};
\end{tikzpicture}
\caption{A block $D$ of zeros in $\mathcal{B}[\{e_1, \ldots, e_h, z\}]$. Observe that the union $C\cup D$ contains a full block $B_{CD}\in \mathcal{B}[X]$} \label{blockintri}
\end{subfigure}%
\hfill
\begin{subfigure}{.45\linewidth}
\begin{tikzpicture}[scale=1]
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node[draw, rectangle, red] at (2,4){$+$};
\node[draw, rectangle, blue] at (2,5){$-$};
\node[draw, rectangle, blue] at (1,5){$-$};
\node [] at (2,5.6) {$z$};
\node [] at (5.6,3) {$z$};
\node [] at (3,5.6) {$f_1$};
\node [] at (5,5.6) {$f_w$};
\node [] at (5.6,5) {$e_1$};
\node [] at (5.6,4) {$e_h$};
\draw [green, line width=0.5mm](-0.6,5.4) -- (2.4,5.4) -- (2.4,2.4)-- (-0.6,5.4);
\node [green] at (1,5.8) {$T_h$};
\end{tikzpicture}
\caption{A block of \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture}, the rest \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} in $T_h$}
\label{pmintri1}
\end{subfigure}
\caption[Applying the inductive hypothesis to $T_h$]{Applying the inductive hypothesis to $T_h$, we find either a block of zeros, or a block of \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture} with the rest \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture}}
\label{indhypto1tri}
\end{figure}
Fix a block $D = \{e_1, \ldots e_d\}\times \{e_{d+1}, \ldots, e_h, z\}$ (some $1 \leq d \leq h$), and consider the term
\[
\left[p_w \psi_D \prod_{(i,j) \in D }(y^+_{ij} y^-_{ij})^{2g} \prod_{(i,j) \in C}(y^+_{ij} y^-_{ij})^{2g}\right]
\] of $[p]$. By Lemma \ref{cd1sbextends}, there is a block $B_{CD} \in \mathcal{B}[X]$ with $B_{CD} \subset C\cup D \cup \bar{D}$, so our term is a multiple of $\prod_{(i,j) \in B_{CD}}(y^+_{ij} y^-_{ij})^{2g}$ and hence has the desired form (see Figure \ref{blockintri}).
Now consider the term
\[
\left[p_w \chi_D \prod_{(i,j) \in D}(y^-_{ij})^{2g}\prod_{\substack{i,j \in \{e_1, \ldots, e_h,z\} \\ (i,j) \notin D \cup \bar{D} \\ i < j}}(y^+_{ij})^{2g}\prod_{(i,j) \in C}(y^+_{ij} y^-_{ij}) ^{2g}\right]
\] of $[p]$ (see Figure \ref{pmintri1}).
Factorise $[p_w \chi_D]$ as a sum of terms of the form $[\alpha_w \alpha_h]$, where $\alpha_w \in \mathbb{Q}[Y(\{f_1, \ldots, f_w, z\})]$ and $\alpha_h \in \mathbb{Q}[Y^-_D(\{e_1, \ldots, e_h,z\})]$. By Lemma \ref{indoroct}, either $\deg \alpha_w \geq 2gw(w+1)-w$ or $\deg \alpha_h \geq h(h+1)g -(h+1) +2$ (see figure \ref{octorindhyp}).
\begin{figure}
\centering
\begin{tikzpicture}[scale=1]
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node[draw, rectangle, red] at (2,4){$+$};
\node[draw, rectangle, blue] at (2,5){$-$};
\node[draw, rectangle, blue] at (1,5){$-$};
\node [] at (2,5.6) {$z$};
\node [] at (5.6,3) {$z$};
\node [] at (3,5.6) {$f_1$};
\node [] at (5,5.6) {$f_w$};
\node [] at (5.6,5) {$e_1$};
\node [] at (5.6,4) {$e_h$};
\draw [green, line width=0.7mm](-0.6,5.4) -- (2.4,5.4) -- (2.4,2.4)-- (-0.6,5.4);
\node [green] at (1,5.8) {$T_h$};
\draw [cyan, line width=0.7mm](1.4,3.3) -- (5.4,3.3) -- (5.4,-0.6)-- (1.4,3.3);
\node [cyan] at (5.8,1.6) {$T_w$};
\draw [green, line width=0.7mm](-4.6,5.4) -- (-1.6,5.4) -- (-1.6,2.4)-- (-4.6,5.4);
\node [] at (-3,5.8) {$\bar{T_h}$};
\foreach \x in {0,...,2}
{\draw [fill=gray](\x-4,5-\x) circle [radius=0.1];
}
\node[draw, rectangle, blue] at (-2,4){$-$};
\node[draw, rectangle, red] at (-2,5){$+$};
\node[draw, rectangle, red] at (-3,5){$+$};
\end{tikzpicture}
\caption{We can apply either Corollary \ref{ajpropflipblock} to $\bar{T_h}$ (left), or the inductive hypothesis to $T_w$}
\label{octorindhyp}
\end{figure}
If $\deg \alpha_h \geq h(h+1)g -(h+1)+2$, then by Corollary \ref{ajpropflipblock}, for each block $E \in \mathcal{B}[\{e_1, \ldots, e_h, z\}]$ we can find a monomial $\beta_E$ such that
\[
[\alpha_h] = \left[ \sum_{E \in \mathcal{B}[\{e_1, \ldots, e_h,z\}]}\beta_E \prod_{(i,j) \in E}(y^{\epsilon_D(i,j)}_{ij})^{2g} \right].
\]
\begin{figure}
\begin{subfigure}{.45\linewidth}
\centering
\begin{tikzpicture}[scale=0.7]
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node[draw, rectangle, red] at (2,4){$+$};
\node[draw, rectangle, blue] at (2,5){$-$};
\node[draw, rectangle, blue] at (1,5){$-$};
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\draw [green, line width=0.7mm](-0.6,5.4) -- (2.4,5.4) -- (2.4,2.4)-- (-0.6,5.4);
\node [green] at (1,5.8) {$T_h$};
\draw [green, line width=0.7mm](-4.6,5.4) -- (-1.6,5.4) -- (-1.6,2.4)-- (-4.6,5.4);
\draw [magenta, line width =1mm] (-2.5,3.5) rectangle (-1.5,5.5);
\node [] at (-3,5.8) {$\bar{T_h}$};
\foreach \x in {0,...,2}
{\draw [fill=gray](\x-4,5-\x) circle [radius=0.1];
}
\node[draw, rectangle, blue] at (-2,4){$-$};
\node[draw, rectangle, red] at (-2,5){$+$};
\node[draw, rectangle, red] at (-3,5){$+$};
\node [] at (-1,2.5) {$\bigoplus$};
\node [] at (7,2.5) {\large $\leadsto$};
\end{tikzpicture}
\end{subfigure}
\hfill
\begin{subfigure}{.45\linewidth}
\centering
\begin{tikzpicture}[scale=0.9]
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node[draw, circle, purple] at (2,4){$0$};
\node[draw, circle, purple] at (2,5){$0$};
\node[draw, rectangle, blue] at (1,5){$-$};
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\draw [green, line width=0.7mm](-0.6,5.4) -- (2.4,5.4) -- (2.4,2.4)-- (-0.6,5.4);
\node [green] at (1,5.8) {$T_h$};
\draw [orange, line width=0.7mm] (1.5, 3.5) rectangle (5.5,5.5);
\end{tikzpicture}
\end{subfigure}
\caption[Applying Corollary \ref{ajpropflipblock}]{Applying Corollary \ref{ajpropflipblock} to $\overline{T_h}$ gives a block $E$ of \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} or \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture} as shown in $\overline{T_h}$, which combines with the \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} and \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture} in $T_h$ to get zeros everywhere in $E$. The union $C \cup E$ contains a full block $B_{CE} \in \mathcal{B}[X]$}
\label{blockint1makesblock}
\end{figure}
Then for fixed $E$, the term
\[\left[\alpha_w \beta_E \prod_{(i,j)\in E}(y^{\epsilon_D(i,j)}_{ij})^{2g}\prod_{(i,j) \in D}(y^-_{ij})^{2g}\prod_{\substack{i, j \in \{e_1, \ldots, e_h,z\} \\ (i,j) \notin D \cup \bar{D} \\ i<j}}(y^+_{ij})^{2g}\prod_{(i,j) \in C}(y^+_{ij} y^-_{ij} ) ^{2g}\right]
\] of $[p]$ contains the factor
\[\prod_{(i,j) \in E} ( y^+_{ij} y^-_{ij})^{2g} \prod_{(i,j) \in C} (y^+_{ij} y^-_{ij})^{2g}
\](up to replacing $y^+_{ji}$ with $y^+_{ij}$ whenever $(i,j) \in E$, $(i,j) \notin D \cup \bar{D}$, and $i<j$). By Lemma \ref{cd1sbextends}, $C \cup E \cup \bar{E}$ contains a block $B_{CE} \in \mathcal{B}[X]$, so this term has the desired form (see Figure \ref{blockint1makesblock}).
Suppose instead that $\deg \alpha_w\geq 2gw(w+1)-w$. Then by the inductive hypothesis,
\[[\alpha_w] = \left[ \sum_{F \in \mathcal{B}[\{f_1, \ldots, f_w, z\}]} \theta_F \prod_{(i,j) \in F} (y^+_{ij} y^-_{ij})^{2g} + \sum_{F \in \mathcal{B}[\{f_1, \ldots, f_w, z\}] \cup \{\varnothing\}} \phi_F \prod_{(i,j) \in F}(y^-_{ij})^{2g}\prod_{\substack{i,j \in \{f_1, \ldots, f_w,z\} \\ (i,j) \notin F \cup \bar{F} \\ i<j}}(y^+_{ij})^{2g} \right],
\] for some homogeneous polynomials $\theta_F$, $\phi_F \in \mathbb{Q}[Y(\{f_1, \ldots, f_w, z\})]$.
\begin{figure}
\begin{subfigure}{.48\linewidth}
\centering
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \x in {3,...,4}
{\draw [fill=gray](\x,3) circle [radius=0.06];
}
\node[draw, rectangle, red] at (2,4){$+$};
\node[draw, rectangle, blue] at (2,5){$-$};
\node[draw, rectangle, blue] at (1,5){$-$};
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\foreach \y in {1,...,3}
{\node[draw, circle, olive] at (5,\y){$0$};
}
\draw [cyan, line width=0.5mm](1.4,3.4) -- (5.4,3.4) -- (5.4,-0.6)-- (1.4,3.4);
\node [cyan] at (5.8,1.6) {$T_w$};
\draw [orange, line width=1mm] (4.5,0.5) rectangle (5.5,5.5);
\end{tikzpicture}
\caption{A block $F$ of zeros in $T_w$; the union $C \cup E$ contains a full block $B_{CF}\in \mathcal{B}[X]$ }
\label{zerosinwdone}
\end{subfigure}
\hfill
\begin{subfigure}{.48\linewidth}
\centering
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\node[draw, rectangle, red] at (2,4){$+$};
\node[draw, rectangle, blue] at (2,5){$-$};
\node[draw, rectangle, blue] at (1,5){$-$};
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\draw [cyan, line width=0.5mm](1.4,3.4) -- (5.4,3.4) -- (5.4,-0.6)-- (1.4,3.4);
\node [cyan] at (5.8,1.6) {$T_w$};
\node[draw, rectangle, red] at (3,3){$+$};
\node[draw, rectangle, red] at (5,1){$+$};
\node[draw, rectangle, blue] at (4,3){$-$};
\node[draw, rectangle, blue] at (5,3){$-$};
\node[draw, rectangle, blue] at (4,2){$-$};
\node[draw, rectangle, blue] at (5,2){$-$};
\draw [orange, line width=1mm] (0.5,4.5) rectangle (3.5,5.5);
\draw [orange, line width=1mm] (3.5,1.5) rectangle (5.5,4.5);
\end{tikzpicture}
\caption{A block of \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture}, the rest \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture}, in $T_w$. There is now a full block $B_{DF}\in \mathcal{B}[X]$ with \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture} (or zeros) in every position inside and \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} (or zero) in every position outside }
\label{bomrestpdone}
\end{subfigure}
\caption{Applying the inductive hypothesis in $T_w$ results in one of these two situations}
\label{caseonedone}
\end{figure}
Fix a block $F \in \mathcal{B}[\{f_1, \ldots, f_w, z\}]$ and consider each term separately. By Lemma \ref{cd1sbextends}, $C \cup F \cup \bar{F}$ contains a block $B_{CF} \in \mathcal{B}[X]$, and so the term
\[
\left[\alpha_h \theta_F \prod_{(i,j) \in F} (y^+_{iz} y^-_{ij})^{2g} \prod_{(i,j) \in D}(y^-_{ij})^{2g}\prod_{\substack{i,j \in \{e_1, \ldots, e_h,z\} \\ (i,j) \notin D \cup \bar{D} \\ i<j}} \prod_{(i,j) \in C}(y^+_{ij} y^-_{ij})^{2g} \right]
\] of $[p]$ contains the factor $\prod_{(i,j) \in B_{CF}}(y^+_{ij} y^-_{ij})^{2g}$ and hence is of the desired form (see Figure \ref{zerosinwdone}).
By Lemma \ref{ebfbextendinb}, there exists a block $B_{DF} \in \mathcal{B}[X]$ with either $D \cup F \subset B_{DF} \subset D \cup F \cup C \cup \bar{C}$, or $\bar{D} \cup F \subset B_{DF} \subset \bar{D} \cup F \cup C \cup \bar{C}$, and thus the term
\[
\left[ \alpha_h \phi_F \prod_{(i,j) \in F}(y^-_{ij})^{2g}\prod_{\substack{i,j \in \{f_1, \ldots, f_w,z\} \\ (i,j) \in F \cup \bar{F} \\ i < j}}(y^+_{ij})^{2g} \prod_{(i,j) \in D}(y^-_{ij})^{2g} \prod_{\substack{i,j \in \{e_1, \ldots, e_h,z\} \\ (i,j) \notin D \cup \bar{D} \\ i<j}}(y^+_{ij})^{2g} \prod_{(i,j) \in C} (y^+_{ij} y^-_{ij})^{2g}\right]
\] of $[p]$ contains the factor $\prod_{(i,j) \in B_{DF}}(y^-_{ij})^{2g}\prod_{\substack{i,j \in X \\ (i,j) \notin B_{DF} \cup \bar{B_{DF}} \\ i<j}}(y^+_{ij})^{2g}$, and hence is of the desired form (see Figure \ref{bomrestpdone}).
Hence each term $ \left[ r_z \tilde{\psi_C} \prod_{(i,j ) \in C}(y^+_{ij} y^-_{ij})^{2g} \right]$ has a representative in $\mathbb{Q}[Y(X)]$ of the desired form.
It remains to show that
\[ \left[ r_z \sum_{C \in \mathcal{B}[X\setminus \{z\}] \cup \{\varnothing\}} \tilde{\chi_C} \prod_{(i,j) \in C}(y^-_{ij})^{2g} \prod_{\substack{i,j \in X \setminus \{z\} \\ (i,j) \notin C \cup \bar{C} \\ i<j}}(y^+_{ij})^{2g} \right]
\] (Figure \ref{codim1bomrp}) has a representative in $\mathbb{Q}[Y(X)]$ of the desired form; again it suffices to show this for each term separately.
\begin{figure}
\centering
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, rectangle, blue] at (\x,\y){$-$};
}
\foreach \y in {4,5}
{\draw [fill=gray](2,\y) circle [radius=0.07];
}
\foreach \x in {3,...,5}
{\draw [fill=gray](\x,3) circle [radius=0.07];
}
\node [] at (2,5.6) {$z$};
\node [] at (5.6,3) {$z$};
\node [] at (3,5.6) {$f_1$};
\node [] at (5,5.6) {$f_w$};
\node [] at (5.6,5) {$e_1$};
\node [] at (5.6,4) {$e_h$};
\node [draw, rectangle, red] at (1,5) {$+$};
\node [draw, rectangle, red] at (4,2) {$+$};
\node [draw, rectangle, red] at (5,2) {$+$};
\node [draw, rectangle, red] at (5,1) {$+$};
\draw [green] (1.7,-0.3) rectangle (2.3,5.3);
\draw [green] (-0.3,2.7) rectangle (5.3,3.3);
\end{tikzpicture}
\caption[The result of the inductive hypothesis]{The inductive hypothesis gave us a block $C \in \mathcal{B}[X\setminus \{z\}]$of \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture}, and \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} everywhere else above the diagonal and away from $z$}
\label{codim1bomrp}
\end{figure}
Fix $C = V \times ((X \setminus \{z\}) \setminus V) \in \mathcal{B}[X\setminus \{z\}]$ and consider
\begin{equation}\label{rzchitildeetc}
r_z \tilde{\chi_C} \prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{i,j \in X \setminus \{z\} \\ (i,j) \notin C \cup \bar{C} \\ i<j}}(y^+_{ij})^{2g}.\end{equation}
Observe that $\deg r_z \tilde{\chi_C} \geq 2gm(m-1)-m+1 -g(m-1)(m-2)$. By Lemma \ref{zroworoct}, $[r_z \tilde{\chi_C}]$ has a representative in $\mathbb{Q}[Y(X)]$ that is a sum of terms of the form $q \prod_{i \in X \setminus \{z\}}(y^{-\epsilon_V(i)}_{iz})^{d_i}$, where either $\deg q \geq m(m-1)g-m+2$, or $d_i \geq 2g \forall i \in X \setminus \{z\}$ (Figure \ref{case2done}).
\begin{figure}
\begin{subfigure}{0.49\linewidth}
\centering
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, rectangle, blue] at (\x,\y){$-$};
}
\foreach \y in {4,5}
{\node[draw, rectangle, blue] at (2,\y) {$-$};
}
\foreach \x in {3,...,5}
{\node [draw, rectangle, red] at (\x,3) {$+$};
}
\draw [teal, line width=1mm] (1.5,3.5) rectangle (2.5,5.5);
\draw [teal, line width=1mm] (2.5,2.5) rectangle (5.5,3.5);
\draw [violet, line width=1mm] (2.6,3.6) rectangle (5.4,5.4);
\draw [magenta, line width=1mm] (1.4,3.4) rectangle (5.6,5.6);
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\node [magenta] at (1,4.4) {$\tilde{C}$};
\node [violet] at (4,4.5) {$C$};
\node [draw, rectangle, red] at (1,5) {$+$};
\node [draw, rectangle, red] at (4,2) {$+$};
\node [draw, rectangle, red] at (5,2) {$+$};
\node [draw, rectangle, red] at (5,1) {$+$};
\end{tikzpicture}
\caption{We can either get the appropriate pattern of \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} and \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture} along $z$ to extend $C$ to a full block of \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture} with everything else \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} \ldots}
\label{case2rowz}
\end{subfigure}
\hfill
\begin{subfigure}{0.49\linewidth}
\centering
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {3,...,5}
\foreach \y in {4,5}
{\node[draw, rectangle, pink] at (\x,\y){$+$};
}
\foreach \y in {4,5}
{\node[draw, rectangle, pink] at (2,\y) {$+$};
}
\foreach \x in {3,...,5}
{\node [draw, rectangle, cyan] at (\x,3) {$-$};
}
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\node [draw, rectangle, cyan] at (1,5) {$-$};
\node [draw, rectangle, cyan] at (4,2) {$-$};
\node [draw, rectangle, cyan] at (5,2) {$-$};
\node [draw, rectangle, cyan] at (5,1) {$-$};
\draw [magenta, line width=1mm] (1.4,3.4) rectangle (5.6,5.6);
\node [magenta] at (1,4.4) {$\tilde{C}$};
\draw[orange, line width=1mm] (3.5,1.5) rectangle (5.5,5.5);
\node [orange] at (6,2.4) {$G$};
\end{tikzpicture}
\caption{\ldots or we can apply Corollary \ref{ajpropflipblock} to find a full block $G$ in the subring $\mathbb{Q}[Y^-_{\tilde{C}}(X)]$}
\label{case2oct}
\end{subfigure}
\caption{Lemma \ref{zroworoct} puts us in one of these situations}
\label{case2done}
\end{figure}
In the latter case, the corresponding terms of $[p]$ have the factor
\[
\prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{i,j \in X \setminus \{z\} \\ (i,j) \notin C \cup \bar{C} \\ i<j}}(y^+_{ij})^{2g}\prod_{i \in X \setminus \{z\}}(y^{-\epsilon_V(i)}_{iz})^{2g} = \prod_{(i,j) \in \tilde{C}}(y^-_{ij})^{2g}\prod_{\substack{i,j \in X \\ (i,j) \notin C \cup \tilde{C}\\ i<j}}(y^+_{ij})^{2g},
\]where $\tilde{C} = V \times (X\setminus V)$ is the extension of the block $C$ obtained by adding $z$ to the second factor. Hence these terms have the desired form (see Figure \ref{case2rowz}).
If a term $q \prod_{i \in X \setminus \{z\}}(y^{-\epsilon_V(i)}_{iz})^{d_i}$ has $\deg q \geq m(m-1)g-m+2$, then by Corollary \ref{ajpropflipblock} there exist homogeneous polynomials $\gamma_G$, for each $G \in \mathcal{B}[X]$, such that
\[
[q] = \left[ \sum_{G \in \mathcal{B}[X]} \gamma_G \prod_{(i,j) \in G} (y^{\epsilon_{\tilde{C}}(i,j)}_{ij})^{2g}\right]
\]
(see Figure \ref{case2oct}). Then each such term \[\prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{i,j \in X \setminus \{z\} \\ (i,j) \notin C \cup \bar{C} \\ i<j}}(y^+_{ij})^{2g}\gamma_G\prod_{(i,j) \in G}(y^{\epsilon_{\tilde{C}}(i,j)}_{ij})^{2g}\] in \eqref{rzchitildeetc} contains the factor $\prod_{(i,j) \in G \vert_{X \setminus \{z\}}}(y^+_{ij} y^-_{ij})^{2g}$ (up to replacing $y^+_{ji}$ by $y^+_{ij}$ whenever $(i,j) \notin \tilde{C} \cup \bar{\tilde{C}}$, $(i,j) \in G$ and $i>j$). By Remark \ref{restrictblock}, $G \vert_{X \setminus \{z\}} \in \mathcal{B}[X\setminus \{z\}]$, and thus we are reduced to a monomial of the form \eqref{codim1zeros} (Figure \ref{case2now1}), which we dealt with above.
Thus we have found a representative for $[p]$ in $\mathbb{Q}[Y(X)]$ that has the desired form.
\begin{figure}
\centering
\begin{tikzpicture}
\foreach \x in {0, ..., 5}
{\draw [fill=gray] (\x, 5-\x) circle [radius=0.1];
}
\foreach \x in {0, ..., 4}
\foreach \y in {\x, ..., 4}
{\draw [fill=gray](\x,\y-\x) circle [radius=0.06];
}
\foreach \x in {4,...,5}
\foreach \y in {4,5}
{\node[draw, circle, violet] at (\x,\y){$0$};
}
\foreach \y in {4,5}
{\draw [fill=gray](2,\y) circle [radius=0.06];
}
\foreach \x in {4,...,5}
{\node [draw, rectangle, cyan] at (\x,3) {$-$};
}
\foreach \y in {4,5}
{\node[draw, rectangle, blue] at (3, \y) {$-$};
}
\draw [green] (1.7,-0.3) rectangle (2.3,5.3);
\draw [green] (-0.3,2.7) rectangle (5.3,3.3);
\node [] at (2,5.8) {$z$};
\node [] at (5.8,3) {$z$};
\node [] at (3,5.8) {$f_1$};
\node [] at (5,5.8) {$f_w$};
\node [] at (5.8,5) {$e_1$};
\node [] at (5.8,4) {$e_h$};
\node [draw, rectangle, red] at (1,5) {$+$};
\node [draw, circle, violet] at (4,2) {$0$};
\node [draw, circle, violet] at (5,2) {$0$};
\node [draw, rectangle, red] at (5,1) {$+$};
\draw[orange, line width=1mm] (3.5,1.5) rectangle (5.5,5.5);
\end{tikzpicture}
\caption[Combining the block $G$ with ]{Combining the block $G$ (Figure \ref{case2oct}) with \begin{tikzpicture} \node[draw, rectangle, blue] at (0,0){$-$}; \end{tikzpicture} everywhere in $C$ and \begin{tikzpicture} \node[draw, rectangle, red] at (0,0){$+$}; \end{tikzpicture} elsewhere away from $z$ (Figure \ref{codim1bomrp}) gives a block of zeros in $\mathcal{B}[X\setminus \{z\}]$, reducing us to the earlier case (Figure \ref{codim1boz})}
\label{case2now1}
\end{figure}
\end{proof}
\begin{lem} \label{octandonediag} Let $B \in \mathcal{B}[[n]]\cup \{ \varnothing \}$ be a block (or the empty set), let $p \in \mathbb{Q}[Y(X)]$ be a monomial, and fix $1 \leq i \leq n$. Then $[p]$ is equivalent in $R$ to a sum of terms of the form $[qr]$, where $q \in \mathbb{Q}[y^+_{ii}]$ and $r \in \mathbb{Q}[Y^-_B(X)]$.
\end{lem}
\begin{proof}We must check that the images in $R$ of the elements of $Y^-_B(X)$ together with $[y^+_{ii}]$ generate $R$. We showed in the proof of Lemma \ref{zroworoct} that the images of elements in $Y^-_B(X)$ together with $[y^{-\epsilon_B(i,z)}_{iz}]$ (for some $z \neq i$ generate $R$, so it suffices to show that $[y^{-\epsilon_B(i,z)}_{iz}] \in \langle [Y^-_B(X)], [y^+_{ii}]\rangle$. But this is clear, because $y^{\epsilon_B(i,z)}_{iz} \in Y^-_B(X)$, and $[y^{\epsilon_B(i,z)}_{iz}+y^{-\epsilon_B(i,z)}_{iz}]=[2y^+_{ii}]$.
\end{proof}
\begin{notn} If $p \in \mathbb{Q}[Y(X)]$ is a monomial, let $d^\pm_{ij}$ (for $i,j \in X$) be the integers such that $p=\lambda\prod_{i,j \in X}(y^\pm_{ij})^{d^\pm_{ij}}$ (where $\lambda \in \mathbb{Q}$), and define $d_{ij}(p):=d^+_{ij}+d^-_{ij}+d^+_{ji}+d^-_{ji}$.
\end{notn}
\begin{lem} \label{procedure} Let $X=[n]$ and let $p \in \mathbb{Q}[Y(X)]$ be a monomial of degree at least $2gn^2+\frac{1}{2}(n-1)(n-2)-2gn-n(n-1)g$. Let $C \in \mathcal{B}[X] \cup \{\varnothing\}$ be a fixed block (or the empty set). Then we can find a (finite) set $\mathcal{D}$ whose elements are $2g$-tuples of (possibly empty) unions of blocks in $\mathcal{B}[X]$, and homogeneous polynomials $\phi_D$ for each $D \in \mathcal{D}$ and $\theta_B$ for each $B \in \mathcal{B}[X]$, such that
\[[p]= \left[\sum_{B \in \mathcal{B}[X]} \theta_B \prod_{(i,j) \in B} (y^{\epsilon_C(i,j)}_{ij})^{2g} + \sum_{D \in \mathcal{D}}\left(\phi_D\prod_{m=1}^{2g}\left(\prod_{(i,j) \in D_m}y^{\epsilon_C(i,j)}_{ij}\right)\left(\prod_{\substack{(i,j) \notin D_m \cup \bar{D_m} \\ i<j}}y^{-\epsilon_C(i,j)}_{ij}\right)\right) \right].
\]
\end{lem}
\begin{proof} First factorise $p$ as $p_-p_+$, where $p_- \in \mathbb{Q}[Y^-_C(X)]$ and $p_+ \in \mathbb{Q}[Y^+_C(X)]$. If $\deg p_- \geq n(n-1)g-n+2$, we are done by Corollary \ref{ajpropflipblock}. Otherwise, let $a <2g$ be the largest integer such that $\deg p_- \geq \frac{1}{2}n(n-1)a-n+2$, and write
\[[p_-] = \left[\sum_{B \in \mathcal{B}[X]} \psi_B \prod_{(i,j) \in B} (y^{\epsilon_C(i,j)}_{ij})^a \right]
\]
(which is possible by Corollary \ref{ajpropflipblock}). Treat each monomial in the resulting polynomial separately.
Observe that for each $i < j$, we know $y^{\epsilon_C(i,j)}_{ij} \in \mathbb{Q}[Y^-_C(X)]$ and $y^{-\epsilon_C(i,j)}_{ij} \in \mathbb{Q}[Y^+_C(X)]$. Thus, if a monomial $q = q_+\prod_{(i,j) \in B}(y^{\epsilon_C(i,j)}_{ij})^a\prod_{l=1}^{2g}(\prod_{(i,j) \in D_m}y^{\epsilon_C(i,j)}_{ij})$ (where the $D_m$ are unions of blocks in $\mathcal{B}[X]$ and $q_+ \in \mathbb{Q}[Y^+_C(X)]$) satisfies $d_{ij}(q) \geq 2g \forall i \neq j$, then with a little rearranging (and swapping $y^+_{ij}$ with $y^+_{ji}$ or $y^-_{ij}$ with $-y^-_{ji}$ as necessary), we can write
\[[q] = \left[\lambda \left( \prod_{(i,j) \in B} (y^{\epsilon_C(i,j)}_{ij})^a \prod_{\substack{(i,j) \notin B \cup \bar{B} \\ i<j}}(y^{-\epsilon_C(i,j)}_{ij})^a\right) \prod_{l=1}^{2g}\left( \left(\prod_{(i,j) \in D_m} y^{\epsilon_C(i,j)}_{ij} \right)\prod_{\substack{(i,j) \notin D_m \cup \bar{D_m}\\i<j}}y^{-\epsilon_C(i,j)}_{ij} \right) \right].
\]
Thus $[q]$ has a representative of the desired form.
The following procedure takes monomials and finds them representatives of the desired form.
Start with a monomial $\alpha \beta \gamma$ of degree $\geq 2gn^2+\frac{1}{2}(n-1)(n-2)-2gn-n(n-1)g$, where $\alpha \in \mathbb{Q}[Y^-_C(X)]$, $\beta = \prod_{l=1}^s(\prod_{(i,j) \in D_m}y^{\epsilon_C(i,j)}_{ij})$ for some $0 \leq s \leq 2g$ and unions $D_m$ of blocks in $\mathcal{B}[X]$, and $\gamma \in \mathbb{Q}[Y^+_C(X)]$.
\begin{enumerate} \item \label{applycor} If $\deg(\alpha\beta )\geq n(n-1)g-n+2$, apply Corollary \ref{ajpropflipblock} and Stop. Otherwise, go to Step 2. \item \label{checkalpha} If $\deg(\alpha) \geq \frac{1}{2}n(n-1)-n+2$, go to Step \ref{hitalpha}. Otherwise go to Step 4. \item \label{hitalpha} Apply Corollary \ref{ajpropflipblock} to write $[\alpha]=\left[\sum_{B \in \mathcal{B}[X]}\phi_B\prod_{(i,j) \in B} y^{\epsilon_C(i,j)}_{ij}\right]$. Consider each term in the resulting polynomial separately. \begin{enumerate}\item If $s<2g$: Replace $\alpha$ with $\phi_B$. Set $D_{s+1}:=B$. Replace $\beta$ with $\beta\cdot \prod_{(i,j) \in B}y^{\epsilon_C(i,j)}_{ij}$. Go to Step \ref{checkalpha}. \item If $s=2g$: \begin{enumerate} \item If $B \subseteq D_m$ for all $1 \leq m \leq 2g$, then the monomial contains the factor $\prod_{(i,j) \in B}(y^{\epsilon_C(i,j)}_{ij})^{2g}$ so is in the desired form. Stop. \item Otherwise, pick $1 \leq m \leq 2g$ with $B \not\subseteq D_m$. Replace $D_m$ with $B \cup D_m$ in $\beta$. Go to Step \ref{checkalpha}. \end{enumerate} \end{enumerate}\item \begin{enumerate}\item \label{moveinaxles} If $d_{ij}(\beta \gamma)\geq 2g$ for all $i<j$, then $\beta \gamma$ has a representative of the desired form, as shown above, so Stop. \item Otherwise, pick $i<j$ with $d_{ij}(\beta \gamma) < 2g$.
We know $\deg(\alpha) \leq \frac{1}{2}n(n-1)-n+1$, whilst $\deg (\alpha \beta \gamma) \geq 2gn^2+\frac{1}{2}(n-1)(n-2)-2gn-n(n-1)g$, so $\deg(\beta \gamma) \geq n(n-1)g=(2g)\binom{n}{2}$. Thus there must be some $k,l \in X$ with $d_{kl} > \begin{cases} 2g & k \neq l,\\ 0 & k=l \end{cases}$ and $y^{\epsilon_C(k,l)}_{kl}$ must be a factor of $\gamma$. Replace one instance of $y^{\epsilon_C(k,l)}_{kl}$ in $\gamma$ with $y^{\epsilon_C(k,i)}_{ki}+y^{-\epsilon_C(i,j)}_{ij}+y^{\epsilon_C(l,j)}_{l,j}$ (the two are equivalent in $R$), and treat each term in the resulting polynomial separately. Go to Step \ref{applycor}.\end{enumerate}
\end{enumerate}
Observe that Step \ref{hitalpha} increases the number $b$ of blocks (counted with multiplicity) in $\cup_{r=1}^sD_s$, and Step \ref{moveinaxles} either increases $\deg(\alpha\beta)$ or decreases $d:=\sum_{\substack{i \neq j \\ d_{ij} < 2g}}(2g-d_{ij})$, for each term in the polynomials these steps create. None of the steps decrease $b$ or $\deg(\alpha\beta)$, or increase $d$. If $\deg(\alpha \beta) \geq n(n-1)g-n+2$, or $d=0$, or $b$ gets large enough to force every $D_m$ to contain some common block $B$ ($b \geq (2g-1)\vert \mathcal{B}[X] \vert +1$ will do), then the algorithm terminates. So even when a step results in a polynomial with more than one term, there are still only finitely many terms, each of which are also closer to a terminating condition than the previous monomial. So this procedure terminates in finite time, giving us a representative of the desired form for any monomial of sufficient degree.
\end{proof}
\begin{prop}\label{alltogethernow} Let $ X \subset \mathbb{N}$ with $\vert X \vert = n$. Let $p \in \mathbb{Q}[Y(X)]$ be a polynomial of degree at least $2gn^2 + \frac{1}{2}(n-1)(n-2)$. Then for each $B = V \times V^c \in \mathcal{B}[X]$ and each $C \in \mathcal{B}[X] \cup \{\varnothing\}$, we can find \begin{itemize} \item homogeneous polynomials $\theta_B, \phi_B, \psi_C \in \mathbb{Q}[Y(X)]$ \item elements $p_V \in \mathcal{A}_{\vert V \vert}$ and $p_{V^c} \in \mathcal{A}_{\vert V^c \vert}$ \item bijections $f_V:[\vert V \vert ] \to V$ and $f_{V^c}:[\vert V^c\vert] \to V^c$ \item a finite set $\mathcal{D_C}$ whose elements are $2g$-tuples of unions of blocks in $\mathcal{B}[X]$, and \item homogeneous polynomials $\chi_D \in \mathbb{Q}[Y(X)]$ for each $D \in \mathcal{D}_C$ \end{itemize} such that
\begin{multline}\label{desiredform}[p]=[\sum_{B = V \times V^c \in \mathcal{B}[X]}\prod_{(i,j) \in B}(y^+_{ij}y^-_{ij})^{2g}(\theta_B\cdot \alpha ((f_V)_*(p_V))+\phi_B \cdot \alpha((f_{V^c})_*(p_{V^c})))\\+\sum_{C \in \mathcal{B}[X] \cup \{\varnothing\}}\psi_C \sum_{D \in \mathcal{D}_C}\chi_D \prod_{m=1}^{2g}\prod_{(i,j) \in D_m}y^+_{ij}y^-_{ij} \prod_{(i,j) \notin D_m \cup \bar{D_m}}y^{-\epsilon_C(i,j)}_{ij}].
\end{multline}
\end{prop}
\begin{rmk} If this relation holds, then there exist $p_i \in \mathcal{A}_n$, monomials $\theta_i \in \mathbb{Q}[Y(X)]$, and a bijection $f_X:[n] \to X$ such that $[p] = [ \sum_i \theta_i \cdot \alpha (f_X)_*(p_i)]$.
\end{rmk}
\begin{proof} By induction on $n$.
For $n=1$, if $p \in \mathbb{Q}[y^+_{11}]$ is a monomial of degree $\geq 2g$, it is in the desired form. (\checkmark)
Suppose $n \geq 2$. By Lemma \ref{monster}, we can write
\[
[p]= \left[ \sum_{B \in \mathcal{B}[X]} \psi_B \prod_{(i,j) \in B} (y^+_{ij} y^-_{ij})^{2g} + \sum_{B \in \mathcal{B}[X] \cup \{\varnothing\}} \chi_B \prod_{(i,j) \in B}(y^-_{ij})^{2g}\prod_{\substack{(i,j) \notin B \cup \bar{B} \\ i<j}}(y^+_{ij})^{2g}\right].
\]
As usual, we treat each term separately. Fix a block $B \in \mathcal{B}[X]$ and consider a monomial $\psi_B \prod_{(i,j) \in B } ( y^+_{ij} y^-_{ij})^{2g}$. Suppose $B = \{e_1, \ldots, e_h\} \times \{f_1, \ldots, f_w\}$; then $\deg \psi_B \geq 2gn^2 + \frac{1}{2}(n-1)(n-2) -4gwh$. By using the relations $[y^\pm_{e_if_j}]=\frac{1}{2}[y^+_{e_ie_i}\pm y^+_{f_jf_j}]$, we can express $[\psi_B]$ as a sum of terms of the form $[\psi_h \psi_w]$, where $\psi_h \in \mathbb{Q}[Y(\{e_1, \ldots, e_h\})]$ and $\psi_w \in \mathbb{Q}[Y(\{f_1, \ldots, f_w\})]$ are monomials of degree $\geq 2gn^2+\frac{1}{2}(n-1)(n-2)-4gwh$. By Lemma \ref{bozsqind}, in each of these terms either $\deg \psi_h \geq 2gh^2+\frac{1}{2}(h-1)(h-2)$ or $\deg \psi_w \geq 2gw^2 + \frac{1}{2}(w-1)(w-2)$, and we can apply the inductive hypothesis to the appropriate monomial, thus getting an expression of the form
\[
\left[\psi_B \prod_{(i,j) \in B} (y^+_{ij} y^-_{ij})^{2g}\right] = \left[(\theta_B \cdot \alpha (f_V)_*(p_V))+\phi_B \cdot \alpha((f_{V^c})_*(p_{V^c})\right],
\] as desired.
Now fix $C \in \mathcal{B}[X] \cup \{\varnothing\}$ and consider the term $\chi_C \prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{(i,j) \notin C \cup \bar{C} \\ i<j}}(y^+_{ij})^{2g}$. Observe that $\deg \chi_C \geq 2gn^2 + \frac{1}{2} (n-1)(n-2) - n(n-1)g$. By repeated application of Lemmas \ref{2goroct} and \ref{octandonediag}, we can write $[\chi_C] = \left[ \prod_{i\in X}(y^+_{ii})^{2g} \theta + \phi r \right]$, where $r \in \mathbb{Q}[Y^-_C(X)]$ is a homogeneous polynomial with $\deg r \geq n(n-1)g-n+2$, and $\theta, \phi \in \mathbb{Q}[Y(X)]$. By Corollary \ref{ajpropflipblock}, $[r] = \left[\sum_{E \in \mathcal{B}[X]} \theta_E \prod_{(i,j) \in E} (y^{\epsilon_E(i,j)}_{ij})^{2g} \right]$, for homogeneous polynomials $\theta_E \in \mathbb{Q}[Y(X)]$. Thus
\[\left[ \phi r \prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{(i,j) \notin C \cup \bar{C} \\ i<j }}(y^+_{ij})^{2g}\right] = \left[ \sum_{E \in \mathcal{B}[X]} \psi_E \prod_{(i,j) \in E} (y^+_{ij} y^-_{ij})^{2g} \right]
\] (for some $\phi_E \in \mathbb{Q}[Y(X)]$), which is of the form considered above.
Finally, consider the monomial $\prod_{i \in X}(y^+_{ii})^{2g} \theta \prod_{(i,j) \in B}(y^-_{ij})^{2g}\prod_{(i,j) \notin B \cup \bar{B}}(y^+_{ij})^{2g}$. Note that $\deg \theta \geq 2gn^2 + \frac{1}{2}(n-1)(n-2) - \frac{1}{2}n(n-1)g-2gn$. Applying Lemma \ref{procedure},
\[
[\theta] = \left[ \sum_{F \in \mathcal{B}[X]} \phi_F \prod_{(i,j) \in F} (y^{-\epsilon_C(i,j)}_{ij})^{2g} + \sum_{D \in \mathcal{D}_C}\psi_D \prod_{m=1}^{2g}(\prod_{(i,j) \in D_m}y^{\epsilon_C(i,j)}_{ij})(\prod_{\substack{(i,j) \notin D_m \cup \bar{D_m} \\ i<j}}y^{-\epsilon_C(i,j)}_{ij}) \right].
\]
As usual, consider each monomial individually. Fix $F$, and consider
\[
\prod_{i \in X}(y^+_{ii})^{2g}\prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{(i,j) \notin C \cup \bar{C}\\i<j}}(y^+_{ij})^{2g}\phi_C \prod_{(i,j) \in F}(y^{-\epsilon_C(i,j)}_{ij})^{2g}.
\] This contains the factor $\prod_{(i,j) \in F}(y^+_{ij} y^-_{ij})^{2g}$, and so is of the form discussed above. The remaining term
\[
\prod_{i \in X} (y^+_{ii})^{2g} \prod_{(i,j) \in C}(y^-_{ij})^{2g}\prod_{\substack{(i,j) \notin C \cup \bar{C} \\ i<j}}(y^+_{ij})^{2g} \left( \sum_{D \in \mathcal{D}_C}\psi_D \prod_{m=1}^{2g}\prod_{(i,j) \in D_m}y^{\epsilon_C(i,j)}_{ij}\prod_{\substack{(i,j) \notin D_m \cup \bar{D_m} \\ i<j}}y^{-\epsilon_C(i,j)}_{ij}\right)
\] is already in the form \eqref{desiredform}.
\end{proof}
\subsection{Proof of the main theorem}
\begin{thm} For each $\phi \in \Phi(G)$, let $k_\phi$ be a nonnegative integer. Then the cohomology class
$
\prod_{\phi \in \Phi(G)}(c_1(L_\phi))^{k_\phi} \in H^{2 \sum_{\phi \in \Phi(G)}k_\phi}(S_{n,g}(t); \mathbb{Q})
$
vanishes whenever $\sum_{\phi \in \Phi(G)} k_\phi \geq 2gn^2 + \frac{1}{2}(n-1)(n-2)$.
\end{thm}
\begin{proof} Let $X = [n]$ and consider the rings $\mathbb{Q}[Y(X)]$ and $R = \mathbb{Q}[Y(X)] / I$. Let $J \subset H^*(S_{n,g}(t); \mathbb{Q})$ be the subring generated by the $c_1(L_\phi)$ for $\phi \in \Phi(G)$. Since the relations \eqref{relns} hold in $J$, the map
\begin{align*} \pi: R & \to J \\ [y^\pm_{ij}] & \mapsto c_1(L^\pm_{ij})
\end{align*}
defines a ring homomorphism. Consider the element $\prod_{\phi \in \Phi(G)} c_1(L_\phi)^{k_\phi} \in J$. It has a representative $[\prod(y^\pm_{ij})^{d^\pm_{ij}}]$ in $R$. Suppose $\sum_\phi k_\phi \geq 2gn^2 + \frac{1}{2}(n-1)(n-2)$. Then by Proposition \ref{alltogethernow}, $[p]$ is equivalent in $R$ to an expression of the form \eqref{desiredform}. So $\pi(p)$ vanishes in $H^*(S_{n,g}(t);\mathbb{Q})$ by Corollary \ref{goodthingsvanish}.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
A classical theorem by Ramsey from 1930~\cite{ramsey1930problem} proves the existence of (finite) Ramsey numbers. The multicolor Ramsey number $R(H_1,H_2,\ldots,H_k)$ is defined as the smallest positive integer $N$ such that for every coloring of the edges of the complete graph $K_N$ with $k$ colors there is a monochromatic subgraph $H_i$ in the $i$-th color. For convenience, we write $R_k(H)$ if $H_i = H$, for all $i\in [k]$.
The study of Ramsey numbers started with complete graphs but has been extended to general graphs and has found the interest of many researchers. Several surveys on the topic can be found in~\cite{ conlon2015recent,graham1990ramsey,radziszowski1994small}. Ramsey numbers for complete graphs are known to grow exponentially and are hard to analyze. Already the base of the exponential term is very much unknown, the most recent bound was by Conlon~\cite{conlon2009new}.
One of the most natural families of graphs to study are paths $P_n$ and cycles $C_n$. The two-color Ramsey numbers for paths are known since 1967~\cite{gerencser1967ramsey}. For more colors the problem is more difficult and it took until 2007 for Gy\'arf\'as, Ruszink\'o, S\'ark\"ozy and Szemer\'edi to prove that $R_3(P_n) = 2n + \mathcal{O}(1)$~\cite{Gyarfas2007}. This progress was initiated with the use of Szemer\'edi's regularity lemma~\cite{szemeredi1975regular}. The regularity lemma has proven to be a strong tool in many applications of extremal graph theory. Figaj and Łuczak~\cite{figaj2007ramsey} used it to show that in the case of Ramsey numbers asymptotically, avoiding \textit{connected matchings}, paths and cycles is the same (see Lemma~\ref{lem:CycMat} below). A matching $M$ is connected in $G$ if all edges
of $M$ are in the same component of $G$. The idea of using these connected matchings was suggested by Łuczak~\cite{luczak1999r}. This method has since been used in a series of papers (see e.g.~\cite{benevides20093,figaj2007ramsey,gyarfas2007three,luczak2012multi}).
These results show that finding large connected matchings is an essential step towards understanding the Ramsey numbers of paths and even cycles.
For $k$ colors, an easy application of the Erd\H{o}s-Gallai extremal theorem on each color class gives a bound of $R_k(P_n) \le k n$ for even $n$ .
This was best known until a recent breakthrough by S\'ark\"ozy~\cite{sarkozy2016multi}, who further improved this to $(k- \frac{k}{16 k^3 +1} )n$. Davies, Jenssen and Roberts~\cite{davies2017multicolour} refined his ideas to get $R_k(P_n) \le \left( k-\frac{1}{4} + \frac{1}{2k} \right) n$.
All of these recent results focus on the $k$-color Ramsey number for paths and obtain the bounds for connected matchings in similar ways. As graphs without connected matching seem to be conceptually easier to analyze than graphs without paths our main improvement comes from considering the former.
\paragraph{Contribution} In this paper, we analyze the structural properties of graphs not containing large connected matchings as a subgraph. We show the vertices can be very clearly partitioned into categories of `low', `intermediate' and `high' degree vertices. In particular, if a connected component is large compared to $n$, then only a small number of vertices have high degree. We introduce a tool for counting edges in such graphs.
While the arguments in the two most recent papers on $R_k(P_n)$ revolve around the overlap of dense connected components, we add the analysis on large connected components. This is important because the construction for the lower bound by Yongqi, Yuansheng, Feng and Bingxi~\cite{yongqi2006new}, who showed that for all $k\ge 3$ we have
$R_k(C_n) \ge \left( k-1 +o(1) \right) n ,$ also contains these large components. Note that the bound is given for cycles but easily extends to paths. We believe this is an essential step towards achieving exact bounds for the multicolor Ramsey numbers for paths and cycles.
For this, we consider the following variant of finding monochromatic connected matchings in $k$-colored dense graphs. This will allow us to generalize our results to paths and cycles using the regularity lemma.
\begin{theorem}
Let $k\geq 4$ be a positive integer and $\varepsilon,\delta$ constants, such that $0< \varepsilon \leq 1/2$ and $0\leq\delta<\frac{\varepsilon^3}{3k^2}$. Then for every even integer $n\geq 4$, every $k$-colored graph $G$ with $v(G)> (k-1/2+\varepsilon)n$ and $e(G)\geq (1-\delta)\binom{v(G)}{2}$ has a monochromatic connected matching of size $n/2$.
\label{thm:main}
\end{theorem}
We will then use the mentioned relationship between avoiding even cycles and connected matching to derive the following theorem.
\begin{theorem}
For every integer $k\geq 4$ and an even integer $n$
\[R_k(C_n)\leq \left(k-\frac{1}{2}+o(1)\right)n.\]
\label{thm:RCn}
\end{theorem}
As as immediate consequence of the above result we obtain an upper bound on the $k$-color Ramsey numbers for paths which follows from the fact that $P_n\subseteq C_n$.
\begin{cor}
For every integer $k\geq 4$ and an even integer $n$
\[R_k(P_n)\leq \left(k-\frac{1}{2}+o(1)\right)n.\]
\label{cor:RPn}
\end{cor}
\paragraph{Outline.} Our paper is organized as follows. In Section~\ref{sec:methods} we show the structural results for graphs avoiding connected matchings and two tools to apply these structures for showing Ramsey results.
In Section~\ref{sec:mainproof} we prove the main result in two steps. First, we use our structural results from
the preceding section to show that we can only have few high-degree vertices. Second, we remove
these vertices and show that the resulting connected components are then all small (of size at most n).
In Sections~\ref{sec:structure} and~\ref{sec:LemlossCorrloss} we prove the lemmas we used in Section~\ref{sec:methods}.
Finally, in Section~\ref{sec:conclusion} we summarize our results and give some ideas for future work.
\section{Methods}\label{sec:methods}
All graphs considered in this paper are simple, without loops or multiple edges. For a graph $G=(V,E)$ we denote by $V(G)$ and $E(G)$ the vertex set and the edge set of the graph $G$ and we set $v(G)$ and $e(G)$ to be the respective cardinalities.
As mentioned there is a relation between avoiding connected matchings and even cycles in form of the following lemma which is a variant of a lemma used in~\cite{figaj2007ramsey}.
\begin{lemma}[Lemma 8 in~\cite{luczak2012multi}]
Let a real number $c>0$ be given. If for every $\varepsilon >0$ there exists a $\delta >0$ and an $n_0$ such that for every even $n>n_0$ and any graph $G$ with $v(G)>(1+\varepsilon)cn$ and $e(G)\geq (1-\delta)\binom{v(G)}{2}$ and any $k$-edge-coloring of $G$ has a monochromatic component containing a matching of $n/2$ edges then
\[R_k(C_n)\leq \left(c+o(1)\right)n.\]
\label{lem:CycMat}
\end{lemma}
For $c\geq 1$ we surely have $(1+\varepsilon)cn \geq (c+\varepsilon)n$ which is more convenient for later calculations. Observe that Theorem~\ref{thm:RCn} follows from Theorem~\ref{thm:main} by using Lemma~\ref{lem:CycMat} with $c=k-1/2$.\\
\subsection{Structure}
The structure of graphs without large connected matchings play an important role in this paper. As a path on $n$ vertices contains a connected matching on $\lfloor n/2 \rfloor$ edges, extremal results for paths directly give an upper bound for connected matchings.
\begin{lemma}[Erd\H{o}s-Gallai~\cite{erdHos1959maximal}]
Let $H$ be a graph which does not contain an $n$-vertex path. Then \[e(H)\leq \frac{n-2}{2}v(H).\]
\label{lem:ErdösGallai}
\end{lemma}
\begin{cor}
Let $H$ be a graph which does not contain a connected matching of size $n/2$ for even $n$. Then
\[e(H)\leq \frac{n-2}{2}v(H).\]
\label{cor:ErdösGallai}
\end{cor}
The extremal graph in both cases consists of disjoint cliques of size $n-1$.
In~\cite{balister2008connected} the extremal configurations of connected graphs without a long path are discussed. We provide a structure for connected graphs without large connected matchings which is very similar. We capture this in the following lemma.
\begin{lemma}
For every connected graph $G=(V,E)$ without a matching of size $n/2$ there is a partition $S\cup Q\cup I$ of the vertex set such that \begin{enumerate}
\item $|Q|+ 2|S| = min\{v(G),n-1\} $, \label{lemstruc:Cond:1}
\item $I$ is an independent set; additionally, if $v(G)\leq n-1$, then $I=\emptyset$,\label{lemstruc:Cond:2}
\item every vertex in $Q$ has at most one neighbor in $I$,\label{lemstruc:Cond:3}
\item every vertex in $I$ has degree less than $n/2$. \label{lemstruc:Cond:4}
\end{enumerate}
\label{lem:structure}
\end{lemma}
We only want to give an intuition on the structure here and defer its proof to Section~\ref{sec:structure}. In Figure~\ref{fig:structure} an example of such a partition can be seen.
Our structure has a series of properties that we use later. From $|Q|+ 2|S| \leq n - 1$ we can easily deduce that $|S| < n/2$ and $|Q| < n$. From $|Q|+ 2|S| \leq v(G)$ we get $ |S|\le |I|$, with a strict inequality as soon as the number of vertices exceeds $n-1$. Note that the graph induced by $Q$ and $S$ can potentially be a clique, and we think of the vertices in $S$ as high degree vertices because they can potentially have edges to every other vertex. Vertices in $Q$ have at most $|S|+|Q| < n$ neighbors and if $v(G) < n$, then $I$ is empty so $|S| = 0$ and $|Q| = v(G)$ (using \eqref{lemstruc:Cond:1}).
\begin{figure}
\centering
\includegraphics[scale=0.8]{structure}
\caption{Visualization of the structure from Lemma~\ref{lem:structure} including extremal structures}
\label{fig:structure}
\end{figure}
From the above properties it also follows that the number of edges is at most $\binom{|Q|+|S|}{2} + |I| \cdot |S| + |Q|$. After some careful consideration one can see that this bound is stronger than the Erd\H{o}s-Gallai bound as soon as $v(G) > n+1$.
\subsection{Loss Function}
In the following we define a function which intuitively should capture the difference between the bound of Erd\H{o}s-Gallai and the bound implied by Lemma~\ref{lem:structure}. Remember Corollary~\ref{cor:ErdösGallai} gave us $e(G)\leq \frac{n-2}{2}v(G)$ for general graphs. To simplify computations, we use the bound
\[e(G)\leq \frac{n-1}{2}v(G)
\]
We denote by $\mathcal{G}$ the class of all graphs without a connected matching of size $n/2$. Then the function maps $\mathcal{G}$ to a positive rational number.
It is important to note that whenever we have connected components we can partition them as in Lemma~\ref{lem:structure}. This partition may not be unique but we can fix an arbitrary one to make the definitions consistent.
\begin{defi}
Let $\mathcal{G}$ be the class of all graphs without a connected matching of size $n/2$. We define $f\colon\mathcal{G}\to \mathbb{Q}_{\geq 0}$ as the difference between $\frac{n-1}{2}v(G)$ and the number of edges in $G$, i.e.
$$ f(G)=\frac{n-1}{2}v(G) -e(G).$$ \label{def:fG}
\end{defi}
Observe that Corollary~\ref{cor:ErdösGallai} guarantees that $f(G)\geq 0$ for any graph without a connected matching of size $n/2$. One way to think about $f(G)$ is as the number of edges $G$ loses in comparison to $\frac{n-1}{2}v(G)$. In the following we also refer to this value as the loss of $G$. Intuitively, this loss can happen for two reasons:
\begin{enumerate}
\item $G$ can lose edges because of large connected components. This part can be captured by the difference between $\frac{n-1}{2}v(G)$ and the bound of edges for connected graphs obtained from Lemma~\ref{lem:structure}.
\item Edges missing because $G$ is not saturated. These are the edges that are not present in $G$ but could be added without creating a connected matching of size $n/2$.
\end{enumerate}
Throughout this paper we only look at the number of lost edges, it will not be important which exact edges are lost or what the cause was.
We introduce a function distributing this loss among the vertices of $G$.
\begin{defi}
Let G be a graph without a connected matching of size $n/2$. We denote by $f(v)\colon V(G)\to \mathbb{Q}_{\geq0}$ the loss of edges in $G$ compared to $\frac{n-1}{2}$ caused by a single vertex $v\in V(G)$ defined as follows. Let $C$ be the connected component of $G$ including $v$ and let $S\cup Q \cup I$ be a partition of $V(C)$ as described in Lemma~\ref{lem:structure}. We set
\begin{align*}
f(v)=\begin{cases}
\frac{n-1}{4},&\text{if } v\in S,\\
\frac{n-1}{2}-\frac{\deg(v)}{2},&\text{if }v\in Q,\\
0,& \text{if }v\in I.
\end{cases}
\end{align*}\\
\label{def:fvW}
\end{defi}
From Lemma~\ref{lem:structure} we have that $\deg(v) \le n-1$ for vertices in $Q$ and so the function $f(v)$ is non-negative and well defined. Since we want to distribute the loss of the graph over the different vertices we need that the sum of the losses over all vertices is at most the loss of $G$. We claim that this holds for $f$ defined as above.
\begin{lemma}
Let $G$ be a graph without a connected matching of size $n/2$.
Then $$\sum_{v\in V(G)}f(v)\leq f(G).$$\label{lem:fconnloss}
\end{lemma}
We defer the proof of this to Section~\ref{sec:LemlossCorrloss}.\\
Let $G$ be a graph with a $k$ coloring of its edges. We introduce three different categories for the vertices of $G$. Let $v\in V(G)$ be some vertex of our graph and let $G_1,\ldots,G_k$ be the monochromatic subgraphs in each color. Then for every color $i$ the vertex $v$ is in exactly one component of $G_i$. We denote this component by $C_i^v$. Then we have $C_1^v,\ldots,C_k^v$ which are $k$ monochromatic components in different colors, each including the vertex $v$. We partition the vertices into three classes depending on their role in the $k$ components.
\begin{defi}
For a graph $G$ with a $k$-edge-coloring not containing a monochromatic connected matching of size $n/2$ and a vertex $v\in V(G)$, let $C_1^v,\ldots,C_k^v$ be the $k$ components containing $v$ in $G_1,\ldots ,G_k$. Consider the partition $C_i^v=S_i^v\cup Q_i^v \cup I_i^v$ given by Lemma~\ref{lem:structure} for every color $i$.
We call the vertex $v$
\begin{enumerate}
\item \textit{\textbf{strong}}\text{ if it is in $S_i^v$ for some color $i$,}\label{def:strong}
\item \textit{\textbf{Q-saturated}}\text{ if it is in $Q_i^v$ for every color $i$,}\label{def:qsat}
\item \textit{\textbf{small}} \text{ if it is in $I_i^v$ for some color $i$ and in $Q_j^v$ or $I_j^v$ for all other colors $j$.\label{def:small}}
\end{enumerate}
\label{defi:Vertexclass}
\end{defi}
Generalizing Definition~\ref{def:fG}, we define $F(G)$ to be the total loss of edges over all colors in a graph $G$ without a monochromatic connected matching of size $n/2$. We consider $G$ to be the union of $k$ monochromatic graphs, each of which avoids a connected matching of size $n/2$. Applying Corollary~\ref{cor:ErdösGallai}, again with a slight weakening of the bound, for each of the color classes, gives
\[e(G) \leq k\cdot \frac{n-1}{2}v(G).\]
This is the bound we use for comparison throughout the paper.
\begin{defi}
Let $\mathcal{G}_c$ be the class of all $k$-edge-colored graphs which do not contain a monochromatic connected matching of size $n/2$ and let $F\colon\mathcal{G}_c\to \mathbb{Q}_{\geq 0}$ be defined as the difference between $k\cdot\frac{n-1}{2}v(G)$ and the number of edges in $G$, i.e.
$$ F(G)=k\cdot\frac{n-1}{2}v(G)-e(G).$$
\end{defi}
Observe that $F(G) =k\cdot\frac{n-1}{2}v(G)-e(G) = \sum_{i=1}^{k} f(G_i).$
The above equality holds because every edge of $G$ is colored in exactly one color and thus counted in exactly one $G_i$. This also implies that $F(G)$ is non-negative by the non-negativity of $f(G_i)$.
We again distribute this loss over the vertices in $G$.
\begin{defi}
Let $G$ be a $k$-edge-colored graph without a monochromatic connected matching of size $n/2$ and let $F(v)\colon V(G)\to \mathbb{Q}_{\geq 0}$ denote the loss of edges in $G$ compared to $k\cdot \frac{n-1}{2}v(G)$ caused by a single vertex $v\in V(G)$. We set
\begin{align*}
F(v)=\begin{cases}
\frac{n-1}{4},& \text{if $v$ is strong},\\
k\cdot \frac{n-1}{2} -\frac{\deg(v)}{2},&\text{if $v$ is $Q$-saturated}, \\
0, & \text{if $v$ is small}.
\end{cases}
\end{align*}
\label{def:lossfctF}
\end{defi}
This is again well defined as a $Q$-saturated vertex has degree at most $n-1$ in every color by Lemma~\ref{lem:structure}. We conclude $\deg(v)\leq k\cdot(n-1)$ which directly implies that $F(v)$ is non-negative.
We deduce the following corollary from Lemma~\ref{lem:fconnloss} to again verify that we have a valid distribution of the loss. The proof is deferred to Section~\ref{sec:LemlossCorrloss}.
\begin{cor}
Let $G$ be a $k$-edge-colored graph without a monochromatic matching of size $n/2$. Then
\[\sum_{v\in V(G)}F(v)\leq F(G).\]
\label{cor:lossFv}
\end{cor}
\subsection{Small Components}
Finally, we introduce a lemma to find an upper bound on the number of edges of graphs whose color classes have the special property of consisting only of components which are not too large.
\begin{lemma}
For any two integers $k\geq 4$ and $n\geq 4$ let $G$ be a $k$-edge-colored graph on $(k-1/2)n$ vertices with color classes $G_1,\ldots,G_k$ such that no $G_i$ has a component of size larger than $n$.
Then we have
\[e(G)\leq \binom{v(G)}{2}-\frac{n^2}{32}.\]
\label{lem:eGsmallcomp}
\end{lemma}
\begin{proof}
For sake of contradiction we assume there exists a graph $G$ on $(k-1/2)n$ vertices with a $k$-coloring such that no color class contains a component of size larger than $n$ and
\[e(G) > \binom{v(G)}{2}-\frac{n^2}{32}.\numberthis\label{eq:eGsmallcontra}\]
Additionally, let $G$ be the graph such that the number of edges is maximal among all graphs satisfying the property above.
Choose any color from the graph and without loss of generality we assume it is blue.
Over all possible colorings satisfying that no color class has a component of size larger than $n$ we choose the coloring which maximizes the number of blue edges, i.e.~blue is the densest color. Let $G_B$ be the subgraph of $G$ induced by the blue edges. We take a closer look at the structure of the blue graph.
\begin{claim}
The number of components in $G_B$ is either $k$ or $k+1$.
\end{claim}
\begin{proof}
If we had only $k-1$ components, then we could cover at most $(k-1)n<(k-1/2)n$ vertices. Hence we can be sure we have at least $k$ components. By edge maximality of $G$ and the choice of the coloring we cannot have more than one component of size less than $n/2$ in the blue graph, as otherwise we could add or recolor any edge between the two components. But if we had $k+2$ components, then by convexity and the fact that there is at most one component of size less than $n/2$, the number of edges is maximized when we have $k-3$ cliques of size $n$ , one clique of size $n-1$, $3$ cliques of size $n/2$ and one isolated vertex.
Then
\begin{align*}
e(G_B)&\leq(k-2)\binom{n}{2}+3\binom{n/2}{2}\leq (k-2)\frac{n^2}{2} + 3\frac{n^2}{8} = \left(k-\frac{5}{4} \right) \frac{n^2}{2}.
\end{align*}
As blue was the densest color, the total number of edges is at most $k\cdot e(G_B)$. We conclude
$$e(G)\leq k\left(k-\frac{5}{4} \right) \frac{n^2}{2}.$$
A simple algebraic transformation gives
$$e(G)\leq\left(k^2-k+\frac{1}{4}-\frac{k-1/2}{n}\right)\frac{n^2}{2}-\left(\frac{k}{4}+\frac{1}{4}-\frac{k-1/2}{n}\right)\frac{n^2}{2}.$$
Using that the complete graph on $(k-1/2)n$ vertices has $e(K_{v(G)})=\binom{v(G)}{2}=\frac{n^2}{2}\left(k^2-k+\frac{1}{4}-\frac{k-1/2}{n}\right)$ edges and assuming $n\geq 4$ we get
$$e(G)\leq \binom{v(G)}{2}- \frac{n^2}{16}.$$
Thus, $G$ having at least $k+2$ components contradicts our assumption \eqref{eq:eGsmallcontra} so we conclude that $G_B$ has either $k$ or $k+1$ components.
\end{proof}
By convexity and our previous observation that the blue graph has either $k$ or $k+1$ components, we obtain the number of blue edges is maximized when there are $k$ components: $k-1$ of size $n$ and one of size $n/2$. In this case we have
\begin{align*}
e(G_B)&\leq (k-1)\frac{n^2}{2}+\frac{n^2}{8}=\left(k-\frac{3}{4}\right)\frac{n^2}{2}.
\end{align*}
Let us now see what this implies for the other components. Consider the color with the second most edges, say, red. Let $G_R$ be the subgraph of $G$ induced by the red edges. By the choice of $G$ and the coloring, the blue components must be cliques.
Thus for all components $C_B\subseteq G_B$ and $C_R\subseteq G_R$, the edges with both endpoints in $V(C_B)\cap V(C_R)$ are blue. Observe that as the blue color has at most $k+1$ components, this implies that we can view the red graph as a union of $(k+1)$-partite components, where the number of edges is maximized when all components are complete $(k+1)$-partite where every part has the same size. We observe that by the same reasons as seen for the blue graph, the red graph cannot have fewer than $k$ components.
We conclude that the number of edges in the red graph is maximized, when there are again $k$ components: $k-1$ of size $n$ and one of size $n/2$. As $G_R$ is $(k+1)$-partite we know that a component of size $n$ has at most $\left(\left(1-\frac{1}{k+1}\right)\frac{n^2}{2}\right)$ edges. We get
\begin{align*}
e(G_R)&\leq \left(1-\frac{1}{k+1}\right)\left((k-1)\frac{n^2}{2}+\frac{1}{4}\frac{n^2}{2}\right)\\
&\leq\left(1-\frac{1}{k+1}\right) \left( k-\frac{3}{4}\right) \frac{n^2}{2}.
\end{align*}
As red was the second densest color, the number of edges in every other color is at most the number of red edges. We conclude for the total number of edges
\begin{align*}
e(G)&\leq e(G_B)+(k-1)\cdot e(G_R)\\
&\leq \left(k-\frac{3}{4}\right)\frac{n^2}{2}+(k-1)\left(1-\frac{1}{k+1}\right) \left( k-\frac{3}{4}\right) \frac{n^2}{2}.
\end{align*}
Simplifying gives
\begin{align*}
e(G)\leq \left(k^2-\frac{7}{4}k+\frac{11}{4}-\frac{7}{2(k+1)}\right)\frac{n^2}{2}.
\end{align*}
Using the fact that the graph $G$ has $(k-\frac{1}{2})n$ vertices, $n\geq 4$ and $k\geq 4$, we conclude the proof of the lemma with
$$e(G)\leq \binom{v(G)}{2}-\frac{n^2}{32}.$$
\end{proof}
\section{Proof of the Main Theorem}\label{sec:mainproof}
The main part of this paper is the proof of Theorem~\ref{thm:main} about finding monochromatic connected
matchings in $k$-edge-colored dense graphs. We restate the theorem for convenience.\\
\\
\textbf{Theorem 1.} \textit{
Let $k\geq 4$ be a positive integer and $\varepsilon,\delta$ constants, such that $0< \varepsilon \leq 1/2$ and $0\leq\delta<\frac{\varepsilon^3}{3k^2}$. Then for every even integer $n\geq 4$, every $k$-colored graph $G$ with $v(G)> (k-1/2+\varepsilon)n$ and $e(G)\geq (1-\delta)\binom{v(G)}{2}$ has a monochromatic connected matching of size $n/2$. }\\
In the previous section we stated all the tools we need to prove Theorem~\ref{thm:main}, so all that remains is to put all the lemmas together.
\begin{proof}[Proof of Theorem~\ref{thm:main}]
For $0< \varepsilon \leq \frac{1}{2}$, let $\alpha=1/2-\varepsilon$ and $\delta<\frac{\varepsilon^3}{3k^2}$. We proceed to prove Theorem~\ref{thm:main} by contradiction. Let $G$ be an edge maximal graph on $(k-\alpha)n$ vertices with the property that there is a $k$-coloring of the edges of $G$ that avoids a monochromatic connected matching of size $n/2$.
Assume $G$ has many edges, \[e(G)\geq (1-\delta)\binom{v(G)}{2} > \binom{v(G)}{2}-\delta k^2\frac{n^2}{2}.\label{eq:maincontra}\numberthis\]
We then derive a contradiction by showing that $G$ cannot have enough edges.
Firstly, we show that the number of vertices of low degree in $G$ is small. Let $V_{\ell}$ be all vertices $v\in V(G)$ with $\deg(v)<(k-1/2)n$.
\begin{claim}
$|V_{\ell}|\leq \frac{\delta k^2 n}{\varepsilon}.$
\label{cl:main1}
\end{claim}
\begin{proof}
Observe that every vertex in $V_\ell$ misses at least $\varepsilon n$ incident edges. Thus we can find a lower bound on the number of edges missing in $G$ compared to the complete graph $K_{v(G)}$ by $\frac{|V_{\ell}|\varepsilon n}{2}$. As we cannot miss more than $\delta k^2\frac{n^2}{2}$ edges we conclude $|V_{\ell}|\leq \frac{\delta k^2 n}{\varepsilon}.$
\end{proof}
Observe next that every vertex in the set $V(G)\setminus V_{\ell}$ has degree at least $(k-1/2)n$. Recall that, by Lemma~\ref{lem:structure}, we can partition every color class $G_i$ into $S_i\cup Q_i\cup I_i$, such that all vertices with very large degree are in $S_i$ and all vertices in $I_i$ have degree less than $n/2$ and all vertices in $Q_i$ have degree less than $n$. Let $v\in V(G)\setminus V_{\ell}$ be a vertex which is in the $I_i$ for some color $i$ and thus has degree less than $n/2$ in this color. If this vertex was not in $S_j$ for any color $j$, then it has a maximum degree of $n-1$ in every other color, thus $\deg(v)< n/2 +(k-1)(n-1) < (k-1/2)n$, contradicting that $v\in V(G)\setminus V_{\ell}$. We conclude that no vertex in $V(G)\setminus V_{\ell}$ by Definition~\ref{defi:Vertexclass} is small and thus every vertex in $V(G)\setminus V_{\ell}$ is either strong or $Q$-saturated, meaning it is in $S_i$ for some color $i$ or it is in $Q_i$ for all colors $1\leq 1\leq k$.
Secondly, we show that the number of strong vertices in $G$ also has to be small. Let $V_s$ be the set of all strong vertices in $V(G)\setminus V_{\ell}$ and let $\beta=\frac{|V_s|}{n}$.
\begin{claim}
$\beta\leq\frac{2\delta k^2}{\varepsilon^2}. $
\label{cl:main2}
\end{claim}
\begin{proof}
Remember that we previously saw that we can use the function $F(v)$ from Definition~\ref{def:lossfctF} to capture the loss of edges in a graph caused by vertex $v\in V(G)$. By definition for $Q$-saturated vertices we have
\begin{align*}
F(v)&= k\cdot\frac{n-1}{2} -\frac{\deg(v)-1}{2}.\\
\end{align*}
As $\deg(v)\leq v(G)-1= (k-\alpha)n-1$ we conclude that for a $Q$-saturated vertex $v\in V(G)$
\begin{align*}
F(v)&\geq k\cdot\frac{n-1}{2} - \frac{(k-\alpha)n-1}{2}=\alpha\frac{n}{2}-\frac{k-1}{2}.\numberthis\label{eq:FQsatmain}
\end{align*}
By definition of $F$ we have $F(G)=k\cdot\frac{n-1}{2} v(G)-e(G)$ and Corollary~\ref{cor:lossFv} states $\sum_{v\in V(G)}F(v)\leq F(G)$, so we can bound the number of edges in $G$ by
\[e(G)\leq k\cdot\frac{n-1}{2} v(G)- \sum_{v\in V(G)}F(v).\]
Next, we use the loss function $F$ of the vertices to find an upper bound on the number of edges in $G$. All small vertices in $G$ have to be in $V_{\ell}$ and thus there cannot be too many of them, remember for strong vertices we have $F(v)=\frac{n-1}{4}$ by Definition~\ref{def:lossfctF}. Using this, Equation~\eqref{eq:FQsatmain} for $Q$-saturated vertices and the fact that there are only few small vertices allows us to find an upper bound on the number of strong vertices. We get
\begin{align*}
e(G) &\leq k\cdot\frac{n-1}{2}\left(k-\alpha\right)n -\sum_{v\in V(G)} F(v)\\
&\leq k\cdot\frac{n-1}{2}(k-\alpha)n -\oversetbrace{\substack{\text{vertices in } V(G)\setminus V_{\ell}\\\text{which are $Q$-saturated }}}{\left(k-\alpha -\beta - \frac{\delta k^2}{\varepsilon}\right)n\left(\alpha\frac{n}{2}-\frac{k-1}{2}\right)}-\oversetbrace{\substack{\text{vertices in } V(G)\setminus V_{\ell}\\\text{which are strong}}}{\beta n\left(\frac{n-1}{4}\right)}.
%
\end{align*}
Now it follows from algebraic transformations and simple estimations that
\begin{align*}
e(G) &\le \frac{n^2}{2}\left(k^2-2\alpha k +\alpha^2 -\frac{k-\alpha}{n}\right) -\frac{n^2}{2} \left(\frac{\beta}{2}-\alpha\beta -\frac{\delta k^2}{\varepsilon}\alpha\right).\\
\end{align*}
Observe that $\binom{v(G)}{2}=\frac{n^2}{2}\left(k^2-2k\alpha +\alpha^2-\frac{k-\alpha}{n}\right)$. Together with $\alpha=1/2-\varepsilon$ this gives
$$e(G)\leq\binom{v(G)}{2}-\frac{n^2}{2}\left(\varepsilon\beta-\frac{\delta k^2}{\varepsilon}\alpha\right).$$
By the assumption $e(G) \ge \binom{v(G)}{2}-\frac{ \delta k^2 n^2}{2}$ we conclude that $ \varepsilon\beta-\frac{\delta k^2}{\varepsilon}\alpha \leq \delta k^2.$
Using $\alpha \leq 1$ and $\varepsilon \leq 1$ we get a bound on the size of $\beta$.
$$ \beta \leq \frac{\delta k^2}{\varepsilon} + \frac{\delta k^2}{\varepsilon^2}\alpha\leq \frac{2\delta k^2}{\varepsilon^2} .$$
This concludes the proof of Claim 2.
\end{proof}
We look at the induced subgraph on the remaining vertices i.e.~$V(G)\setminus\{V_\ell\cup V_s\}$. Observe that we get using that $\delta<\frac{\varepsilon^3}{3k^2}$, $\alpha=1/2-\varepsilon$ and Claims~\ref{cl:main1} and~\ref{cl:main2}
\begin{align*}
\left|V(G)\setminus\left(V_{\ell}\cup V_s\right)\right|& \geq v(G) - \frac{\delta k^2n}{\varepsilon} -\frac{2\delta k^2n}{\varepsilon^2} \\\
&\geq v(G) - \frac{3\delta k^2n}{\varepsilon^2}\\
&> v(G)- \varepsilon n = (k-1/2)n.
\end{align*}
We now remove $\varepsilon n$ vertices from the graph, including the vertices from $V_s$ and $V_{\ell}$. We are left with a graph $G'$ on $(k-1/2)n$ vertices such that every vertex was $Q$-saturated in $G$. This means we remove from all components in all colors $i$ the vertices in $I_i$ and $S_i$ of the corresponding partition and we are thus left with only monochromatic components of size at most $n-1$. This allows us to apply Lemma~\ref{lem:eGsmallcomp} which gives us $e(G')\leq \binom{v(G')}{2}-\frac{n^2}{32}\leq \binom{v(G')}{2}- \delta k^2 \frac{n^2}{2}$. Which follows by the choice of $\varepsilon$ and $\delta$.
From here we conclude that there are at least $\delta k^2 \frac{n^2}{2}$ edges not present in $G'$. But then, even without considering the edges missing in the rest of $G$ we can conclude that $e(G)\leq \binom{v(G)}{2}-\delta k^2 \frac{n^2}{2}<(1-\delta)\binom{v(G)}{2}$ edges, yielding the desired contradiction.
\end{proof}
\section{Proof of Lemma~\ref{lem:structure}}
\label{sec:structure}
In the following we derive some properties of connected graphs without large matchings. For convenience we restate the lemma\\
\\
\textbf{Lemma~\ref{lem:structure}.} \textit{For every connected graph $G=(V,E)$ without a matching of size $n/2$ there is a partition $S\cup Q\cup I$ of the vertex set such that \begin{enumerate}
\item $|Q|+ 2|S| = min\{v(G),n-1\} $,
\item $I$ is an independent set; if $v(G)\leq n-1$, then $I=\emptyset$,
\item every vertex in $Q$ has at most one neighbor in $I$,
\item every vertex in $I$ has degree less than $n/2$.
\end{enumerate}}
\vspace{1em}
For a graph $G$ and $A \subseteq V(G)$ we define $G\setminus A$ as the subgraph of $G$ where all vertices from $A$ and their incident edges are removed.
For a matching $M\subseteq E$ we call all vertices which are not incident to any edge in $M$ the unmatched vertices in $M$. Furthermore, we denote by $q(G\setminus S)$ the number of odd components in $G\setminus S$.
We use a generalization of Tutte's Theorem by Berge~\cite{berge1958matchings} in our proof.
\begin{theorem}[\textbf{Berge~\cite{berge1958matchings}}]
Let $G=(V,E)$ be a graph. For any set $S\subseteq V$ and any matching $M$, the number of unmatched vertices in $M$ is at least $\q{G\setminus S}-|S|$. Moreover, there exists a set $S\subset V$ such that every maximum matching of $G$ misses exactly $\q{G\setminus S} - |S|$ vertices.
\label{thm:TutteBerge}
\end{theorem}
\begin{proof}[Proof of Lemma~\ref{lem:structure}]
We distinguish two cases in the proof.
\begin{case}[$v(G)\leq n-1$]
In this case we set $Q=V$ and $S=I=\emptyset$. It can be easily verified that all four conditions are satisfied in this case.
\end{case}
\begin{case}[$v(G)>n-1$]
Let $M$ be a maximum matching in $G$ and let $V_M$ be all vertices covered by the matching. As $G$ has no matching covering $n$ vertices and $n$ is even, $2|M| \leq n-2$.
Then we know by Theorem~\ref{thm:TutteBerge} that there exists a subset of $S\subset V$ such that $|V\setminus V_M|= \q{G\setminus S} -|S|$.
Let $\mathcal{Q}=\{Q_1,\ldots,Q_{q(G\setminus S)}\}$ be the set of all odd components in $G\setminus S$ where we assume without loss of generality that $|Q_1|\geq |Q_2|\geq \ldots\geq |Q_{q(G\setminus S)}|$. Let $I$ be be an arbitrary set of vertices from the odd components such that $|I\cap Q_1|=0$ and $|I\cap Q_i|=1$ for all $2\leq i\leq q(G\setminus S)$. Then clearly we have that $I$ is an independent set (proving \eqref{lemstruc:Cond:2}). Note that we have
\[|I|= q(G\setminus S) -1.\]
Let $Q = V\setminus (I\cup S)$ be the set of all remaining vertices. Then $S\cup Q \cup I$ is clearly a partition of $V$ and thus $|V|=|S|+|Q|+|I|$.
By Theorem~\ref{thm:TutteBerge} we know
\begin{align*}
n-2\geq 2|M|&=|V|-(q(G\setminus S)-|S|)\\
&= |S|+|Q|+|I| - q(G\setminus S) + |S|\\
&=2|S|+|Q|+1.
\end{align*}
We conclude $n-1\geq 2|S|+|Q|$. In case $n-1>2|S|+|Q|$ we move vertices from $I$ to $Q$ until the above holds with equality. As we are in the case $v(G) > n-1$ this proves \eqref{lemstruc:Cond:1}.
To see that \eqref{lemstruc:Cond:3} holds, recall that $I$ contains at most one vertex from each odd component. Every vertex in $Q$ can thus be adjacent to at most one of the vertices in $I$, the one which was in the same component in $G\setminus S$. As we by construction did not remove any vertex from the largest odd component, no vertex in $I$ can have more that $\frac{|Q|}{2}$ neighbors in $Q$. Together with the vertices from $S$ we conclude for $v\in I$ that \[\deg(v)\leq |S|+\frac{|Q|}{2}=\frac{n-1}{2},\]
proving \eqref{lemstruc:Cond:4}.
\end{case}
\end{proof}
\section{Proof of Lemma~\ref{lem:fconnloss} and Corollary~\ref{cor:lossFv}}
\label{sec:LemlossCorrloss}
\setcounter{case}{0
Next we proof Lemma~\ref{lem:fconnloss}, recall that
$$ f(G)=\frac{n-1}{2}v(G) -e(G)$$
and for every connected component $C$ and every vertex $v\in V(C)=S\cup Q\cup I$ we have
\begin{align*}
f(v)=\begin{cases}
\frac{n-1}{4},&\text{if } v\in S,\\
\frac{n-1}{2}-\frac{\deg(v)}{2},&\text{if }v\in Q,\\
0,& \text{if }v\in I.
\end{cases}
\end{align*}
\\
\textbf{Lemma~\ref{lem:fconnloss}.} \textit{
Let $G$ be a graph without a connected matching of size $n/2$.
Then $$\sum_{v\in V(G)}f(v)\leq f(G).$$}
\begin{proof}[Proof of Lemma~\ref{lem:fconnloss}]
Let $G$ be a graph, not necessarily connected, and let $C_1,\ldots,C_m$ be the connected components of $G$. Every vertex and every edge is in exactly one component thus we get
\begin{align*}
f(G)&=\frac{n-1}{2}v(G)-e(G)\\
&=\sum_{i=1}^{m} \frac{n-1}{2}v(C_i)-e(C_i)\\
&=\sum_{i=1}^{m}f(C_i).
\end{align*}
If we know that for every component we have $\sum_{v\in v(C_i)}f(v)\leq f(C_i)$, then
\begin{align*}
\sum_{v\in V(G)} f(v) = \sum_{i=1}^m\sum_{v\in v(C_i)}f(v)\leq \sum_{i=1}^{m}f(C_i) = f(G).
\end{align*}
We can thus without loss of generality assume $G$ is connected. This allows us to partition $G$ with Lemma~\ref{lem:structure}.
We distinguish two cases.
\begin{case}[$v(G)\leq n-1$]
In this case all vertices are in $Q$ as by Lemma~\ref{lem:structure} \eqref{lemstruc:Cond:2} we know $I=\emptyset$ and thus $|Q|+2|S|=v(G)$ implies $S=\emptyset$. We clearly have
\begin{align*}
f(G)&=\frac{n-1}{2}v(G) -e(G)= \sum_{v\in V(G)}\left(\frac{n-1}{2}-\frac{\deg(v)}{2}\right)=\sum_{v\in V(G)}f(v).
\end{align*}
\end{case}
\begin{case}[$v(G)>n-1$]
As in the previous case we rewrite the function $f$
$$f(G) = \frac{n-1}{2}v(G) -e(G)= \sum_{v\in V(G)}\left(\frac{n-1}{2}- \frac{\deg(v)}{2}\right).$$
Let $V(G)= S\cup Q \cup I$. We now distinguish between vertices which are in $Q$, $S$ or $I$.
By definition
\begin{equation}\label{eq:qvertices}
\sum_{v\in Q} \left(\frac{n-1}{2} - \frac{\deg(v)}{2}\right) = \sum_{v\in Q} f(v).
\end{equation}
Then we know that $\deg(v)$ for $v\in S$ is at most $v(G)-1$ and $\deg(v)$ for $v\in I$ is at most $|S|$ plus at most $|Q|$ edges in total from the set $I$ to $Q$. Therefore, for vertices in $S$ and $I$ we have the inequalities
\begin{align*}
\sum_{v\in S}\left( \frac{n-1}{2} - \frac{\deg(v)}{2} \right) &\ge \sum_{v\in S} \left( \frac{n-1}{2} - \frac{v(G)-1}{2} \right),\text{ and}\\
\sum_{v\in I}\left( \frac{n-1}{2} - \frac{\deg(v)}{2} \right) &\ge \sum_{v\in I} \left( \frac{n-1}{2} - \frac{|S|}{2} \right) - \frac{|Q|}{2}.
\end{align*}
Adding in the fact that $v(G) = |Q| + |I| + |S|$ and $n-1 = |Q|+2|S|$ we get
\begin{align*}
\frac{n-1}{2} - \frac{v(G)-1}{2} &= \frac{|S|-|I|+1}{2},\text{ and} \\
\frac{n-1}{2} - \frac{|S|}{2} &= \frac{|Q|+|S|}{2} .
\end{align*}
using the fact that $|I| \ge |S|+1$ we obtain
\begin{align*}
\sum_{v\in S\cup I}\left(\frac{n-1}{2}-\frac{\deg(v)}{2}\right)&\geq |S| \frac{|S|-|I|+1}{2} + |I| \frac{|Q|+|S|}{2} - \frac{|Q|}{2} \\&\ge |S| \frac{|Q|+|S|}{2} \ge |S|\frac{n-1}{4} \label{eq:svertices}.\numberthis
\end{align*}
Putting inequalities \eqref{eq:qvertices} and \eqref{eq:svertices} together finally gives
\begin{align*}
f(G) &= \sum_{v\in V(G)}\left( \frac{n-1}{2} - \frac{\deg(v)}{2}\right) \\&\ge \sum_{v\in Q} \left(\frac{n-1}{2} - \frac{\deg(v)}{2}\right) + |S|\frac{n-1}{4} = \sum_{v\in V(G)} f(v).
\end{align*}
This finishes the proof of Lemma~\ref{lem:fconnloss}.
\end{case}
\end{proof}
We deduce Corollary~\ref{cor:lossFv} from Lemma~\ref{lem:fconnloss}. For this consider $G$ to be a $k$-edge-colored graph. Recall that $G_i$ is the monochromatic induced subgraph in color $i$.
We denote the associated function from Definition~\ref{def:fG} and~\ref{def:fvW} for the graph $G_i$ in color $i$ with $f_i$.
\begin{proof}[Proof of Corollary~\ref{cor:lossFv}]
By definition of $F(G)$ and $f_i(G)$ we have that $F(G)=\sum_{i=1}^{k} f(G_i)$. We conclude that if $F(v)\leq \sum_{i=1}^{k} f_i(v)$, then this together with Lemma~\ref{lem:fconnloss} implies \begin{align*}
\sum_{v\in V(G)}F(v)\leq \sum_{v\in V(G)}\sum_{i=1}^{k} f_i(v)=\sum_{i=1}^{k} \sum_{v\in V(G)}f_i(v)\leq\sum_{i=1}^{k}f_i(G_i)=F(G).
\end{align*}
In the following we show that $F(v)\leq \sum_{i=1}^{k} f_i(v)$ indeed holds. For this we look at $F(v)$ depending on the class of the vertex $v$.
For every strong vertex $v$, we know that for at least one color $i$, $v\in S_i$, so we have $f_i(v)=\frac{n-1}{4}$. Hence then $F(v)=f_i(v)\leq \sum_{i=1}^{k} f_i(v)$ by the non-negativity of $f(v)$.\\
For every $Q$-saturated vertex by definition we have $f_i(v)=\frac{n-1}{2}-\frac{\deg_{G_i}(v)}{2}$. This means that for every $Q$-saturated vertex of $G$
\begin{align*}
\sum_{i=1}^{k}f_i(v)&=\sum_{i=1}^{k}\left(\frac{n-1}{2} -\frac{\deg_{G_i}(v)}{2}\right)= k\cdot\frac{n-1}{2} -\frac{\deg_{G}(v)}{2} = F(v)
\end{align*}
since the graphs $G_i$ are edge-disjoint and their union is $G$.\\
For every small vertex we have, by non-negativity of $f(v)$, that $F(v) = 0 \leq \sum_{i=1}^{k}f_i(v)$.
\end{proof}
\section{Conclusion}
\label{sec:conclusion}
In this paper, we provided some insight into the behavior of graphs avoiding connected matchings. We introduced strong properties for large connected components without a matching of size $n/2$. Also these directly imply a better bound for the multicolor Ramsey numbers of paths and cycles.
The analysis of these large connected components is important because of the different extremal constructions that exist for the lower bound of $R_k(P_n)$. The bound by Yongqi et al.~\cite{yongqi2006new} uses large connected components. However, a construction using finite affine planes (see~\cite{Bierbrauer1987123}) shows that similar bounds can be achieved when all colors have small connected components.
To prove the bound for $R_k(C_n)$, we first bound the Ramsey numbers for connected matchings (Theorem~\ref{thm:main}) and then conclude by applying Lemma~\ref{lem:CycMat}, which itself applies the regularity lemma. Because the regularity lemma is such a strong tool, it should be possible to deduce Ramsey bounds for other structures than the path or even cycle, such as bounded degree trees, by adjusting Lemma~\ref{lem:CycMat} (see~\cite{mota2015ramsey} for similar ideas with three colors).
Although we now have some consideration for large connected components and the overlap of small components, we do not look at the overlap of large components. Considering this might lead to better bounds. We would be interested to see these ideas used to prove an upper bound matching the lower bound.
\subsection*{Acknowledgements}
We would like to thank Rajko Nenadov for bringing this problem to our attention. We also thank Miloš Trujić for many helpful comments.
\bibliographystyle{abbrv}
|
\section{Introduction: the biological problem}
\label{intro}
The discovery of the biochemical basis of life is one of the great scientific success stories of the past century. Remarkably, the amazing diversity of life can be explained from a relatively small set of biochemical actors and their interactions. Heritable information is stored in chromosomes, very long polymers of double stranded DNA, which encode information as a sequence of symbols from a four letter alphabet, A,C,G,T, the nucleotides constituting the building blocks of DNA. Just as DNA is the universal information storage medium, information flow also follows a consistent biochemical pathway across all organisms. Stored information can be dynamically read through the process of {\it gene expression}, a two step process whereby DNA gets transcribed into RNA, an intermediate, single stranded polymer of nucleic acids (with the T nucleotide replaced by uracil, U), and RNA is subsequently translated into proteins, macromolecules formed of amino-acids which carry out most cellular functions. This process is of such fundamental importance in biology to have earned the moniker of {\it central dogma of molecular biology} \cite{crick1970central}; it constitutes the universal flow of information across all living creatures (the most notable exception being reverse transcription of viral RNA).
Not all DNA within a cell codes for proteins, and not all DNA is transcribed; indeed, genes, the stretches of DNA encoding some functionality (either protein or other classes of functional RNAs), constitute a small fraction of the overall genome. One of the surprising outcomes of the major genome-sequencing projects at the turn of the millennium was the realisation of just how little DNA codes for proteins (approximately 3\% of the human genome, with similar percentages in other higher eukaryotes). Moreover, the number of genes in different organisms is relatively constant across scales of organismal complexity: the humble baker's yeast {\it Saccharomyces cerevisiae} has approximately 6.000 genes, more than a quarter the number of genes in the human genome. Apart from raising overdue questions on our anthropocentric worldview, the natural corollary of this observation is that complexity in life does not arise from a disparity in the number of available components (genes), but from the nature and dynamics of the interactions between such components.
Measuring interactions is difficult within live cells. On the other hand, measuring components' abundances (e.g. mRNA levels) is considerably easier, and technological advances within the last two decades have enabled increasingly large-scale measurements of gene expression at steadily decreasing costs. This trend has provided a powerful motivation to attempt to reconstruct {\it computationally} the interaction structures underpinning patterns of gene expression: these interactions collectively are denoted as {\it Gene Regulatory Networks} (GRNs). Reconstructing such networks has been a central effort of the interdisciplinary field of {\it Systems Biology}.
In this chapter, we provide a tutorial overview of the field, aimed at a novice computational scientist or biologist wishing to approach the subject. We first provide a brief introduction to the core biological concepts, as well as the main sources of data currently available. We then introduce the core mathematical concepts, and briefly attempt a categorization of the main methodological approaches available. This chapter is intended to be a self-contained introduction which will provide some essential background to the book; later chapters will describe more advanced concepts, and associated tools for GRN reconstruction across the breadth of their biological application.
\subsection{Mechanisms of gene regulation}
The molecular bases of the transcription process have been intensely studied over the last 60 years. Many excellent monographs are available on the subject; we refer the reader in particular to the classic books by Ptashne and collaborators \cite{ptashne2002genes,ptashne2004genetic} (see also this recent review \cite{ptashne2014chemistry} for a historical perspective). Here we give a brief intuitive description of the process, taking, as an illustrative example, the transcriptional response of the bacterium {\it Escherichia coli} in response to changes in oxygen availability (see reference \cite{bettenbrock2014towards} for a modern review of this field). Transcription is carried out by the enzyme RNA polymerase (RNAP), that slides along the DNA, opening the double strand and producing a faithful RNA copy of the gene. The rate of recruitment of RNAP at a gene can be modulated by the presence or absence of specific {\it transcription factor} (TF) proteins, which contain a DNA-binding module that enables them to recognise specific DNA-sequence signals near the start of genes (promoter regions). The classical view of gene regulation holds that changes in cellular state are orchestrated by changes in binding by TFs.
For example, in {\it E. coli}, oxygen withdrawal leads to dimerisation of the master regulator protein Fumarate Nitrate Reductase (FNR); FNR dimers (but not monomers) can bind specifically to DNA, and change the rate of recruitment of RNAP at the FNR target genes, thereby changing their levels of expression to enable the cell to adapt to the changed conditions. However, FNR is not the only regulator responding to changes in oxygen availability: another master regulator, the two component system ArcAB, also senses oxygen changes, albeit through a different mechanism, and changes its binding to hundreds of genes as a result. FNR and ArcAB share many targets, and through their combined action they can give rise to highly complex dynamics \cite{partridge2007transition,rolfe2011transcript}.
Two important observations can be made from the previous discussion. Firstly, the regulation of gene expression levels is enacted through the action of gene products themselves: therefore, in principle, one may hope to be able to describe the dynamics of gene expression as an {\it autonomous} system. Secondly, even in the simple case of the bacterium {\it Escherichia coli}, regulation of gene expression is a complex process, likely to involve the interactions of several molecular players.
In higher organisms, the basic components of the transcriptional regulatory machinery are remarkably similar. However, many more levels of regulatory control are present: in particular, chemical modifications of the DNA itself (in particular methylation of C nucleotides) and of the structural histone proteins, around which DNA is wound, can affect the structural properties of the DNA, and hence the local accessibility to the transcriptional machinery. Such effects, collectively known as {\it epigenetic modifications}, have strong associations with transcription \cite{alberts1994molecular,bird2002dna,karlic2010histone,benveniste2014transcription}, and are generally thought to encode processes of cellular memory associated with long-term adaptation or cell-type differentiation.
Finally, while we have primarily focussed on transcription, subsequent steps of gene expression are also tightly regulated: RNA processing, translation and RNA and protein degradation all provide additional levels at which gene expression can be controlled. Mechanisms of post-transcriptional control of gene expression are less well explored, but it is widely believed that such processes, mostly effected through proteins or RNAs binding to RNA targets, may be as prevalent as transcriptional controls \cite{hogan2008diverse,tebaldi2012widespread}{(see also Chapter 15 for perspectives on incorporating post-transcriptional regulation in GRN inference)}. Therefore, while a gene may have no effect on the expression of another gene at the RNA level, it may well be extremely important for the protein expression.
\subsection{High throughput measurements techniques}
As we have seen in the previous subsection, the control of gene expression is effected through the action of gene products themselves. Naturally, in order to discover and quantify such controls, one must then be able to simultaneously measure the levels of expression of multiple genes. Measurements of gene expression have progressed dramatically in the last twenty years, with technological advances driving a seemingly unstoppable expansion in the scope of such experiments.
Proteins are the final product of the process of gene expression. Methods based on quantitative mass spectrometry have been highly effective in quantifying hundreds to thousands of proteins within samples. Despite that, intrinsic limits to their sensitivity and a relatively complex analysis pipeline mean that such methods do not yet reach the comprehensiveness of transcriptomic measurements \cite{bantscheff2012quantitative}.
Methods for assaying RNA levels have progressed immensely in the last two decades. Microarray technology first provided enormous impetus to the field in the late 90s \cite{brown1999exploring}. Microarrays consist of thousands of short fragments of DNA (probes) arranged on a substrate chip (usually glass or some other inert material); by designing probes to complement thousands of genomic regions from target organisms, one can obtain a readout of the (steady state) concentration of thousands of transcripts within a population of cells.
Microarrays represented a turning point in our ability to comprehensively measure genetic materials; however, the design of the probes implicitly defines what can be measured, biasing the assay and limiting the scope for discovery of unexpected biological facts, e.g. previously unobserved transcripts. Next generation sequencing (NGS) technologies proved revolutionary in this context. NGS provides a massively parallel implementation of DNA sequencing protocols, which enabled it to dramatically reduce costs and expand throughput. RNA-seq is the main NGS technology used to measure transcript abundances \cite{wang2009rna}: RNA from a population of cells is reverse transcribed (usually after a selective enrichment process to filter out highly abundant ribosomal RNAs), fragmented and the resulting complementary DNA is sequenced and mapped to a reference genome. The number of fragments mapped to a particular gene, suitably normalised \cite{evans2017selecting}, then gives a raw measurement of gene expression.
One of the major success stories of NGS technologies is the ability of combining them with a variety of biochemical assays, greatly expanding the scope of potential measurements. Of particular relevance for GRNs is the ability to select fragments of DNA bound to specific proteins by a process called immuno-precipitation. Genomic material is fragmented, and an antibody specific to a particular DNA binding protein is added, allowing separation by centrifugation of the protein. The bound DNA fragments are then released, sequenced and mapped to a reference genome to identify where the protein was bound. This technique, Chromatin Immuno-Precipitation followed by sequencing (ChIP-seq), has been instrumental in obtaining {\it in vivo} mappings of possible regulatory relationships \cite{park2009chip}.
\section{Introduction: the mathematical formulation}\label{graph-intro}
In the previous section, we have given a condensed tour of the fundamental biological problem addressed in this book. We have seen that interactions between gene products are the fundamental processes underpinning the cell's ability to modulate gene expression. High-throughput measuring techniques paved the way to the use of computational statistics techniques to reconstruct statistically such interactions, a process sometimes called {\it reverse engineering}. In this section we introduce some of the fundamental mathematical concepts common to all methods for reverse engineering GRNs, see e.g. \cite{west2001introduction} for a more comprehensive review of these concepts.
\begin{definition}[Network] A (directed) network or graph is a pair $(V,E)$ where $V$ is a finite set of {\it vertices} (or nodes) and $E$ is a set of {\it edges} (or arcs) connecting the vertices. If $\mathcal{I}$ is a set indexing the nodes, the set of edges is a subset of the Cartesian product $E\subset\mathcal{I}\times\mathcal{I}$, with element $(ij)$ indicating the presence of an edge between node $i$ and node $j$. An {\it undirected} network is a network where the edge set is symmetric under swapping the indices of the nodes, i.e. whenever edge $(ij)$ exists also edge $(ji)$ exists.
\end{definition}
Within the GRN context, network nodes universally represent the expression level of genes. Edges are intuitively linked to associations between genes, but the precise meaning of an edge depends on the mathematical model of the system.
Networks are abstract representations of systems, and {\it per se} do not have a semantic interpretation that could link the network to node behaviours, e.g. their collective dynamics. Nevertheless, the structure of a network (the {\it topology}) can provide an intuitively appealing visualisation of the system, and often be informative in itself. Informally, the aim of a network abstraction is to condense in a simple representation the complexity of interactions underpinning gene expression, see Figure \ref{abstraction} for a cartoon representation. One of the most important quantities in this regard is the {\it degree} of a node, i.e. the number of edges that are attached to the node, and the {\it degree distribution} of the network, i.e. the empirical distribution of degrees across all nodes in the network. Degree distributions often encode intuitively interpretable properties of networks such as the presence of hubs or the ability to reach rapidly any node from any starting node, and in many cases they can be related to distinct stochastic mechanisms by which the network may arise. In the case of directed networks, one may further distinguish between {\it in-degree} (also called {\it fan-in}), the number of edges terminating at a node, and {\it out-degree} (also called {\it fan-out}), the number of edges starting at a node.
\begin{figure}
\begin{center}\includegraphics[width=0.95\textwidth]{network_abstraction.pdf}\end{center}
\caption{A cartoon schematic of a gene regulatory network. A complex biophysical model describes the interaction between three genes, involving both direct regulation (gene 2 by gene 1) and combinatorial regulation via complex formation (gene 3 by genes 1 and 2). The abstracted structure of the system is given in the (directed) network on the right.}\label{abstraction}
\end{figure}
Finally, in many cases the bare topological description is insufficient in capturing aspects of interest, such as the different importances of different edges. To obviate this problem, one can consider {\it weighted} networks, where each edge is associated with a real number, its weight. We will see that in most cases reconstructed networks, the topic of this book, arise naturally as weighted networks, where the weight is intuitively associated with the support that the data offers for the existence of an edge. Weighted networks are often visualised as networks with edges of different thickness, retaining the visual immediacy of the network abstraction but effectively conveying more information.
A schematic example of a standard graphical representation for directed, undirected and weighted networks is given in Figure \ref{netTypes}.
\begin{figure}
\begin{center}\includegraphics[width=0.95\textwidth]{network_types.pdf}\end{center}
\caption{Examples of network types: directed (left), undirected (centre) and weighted (right), where the weights are represented by edge thickness. Note that a weighted network can be directed or undirected.}\label{netTypes}
\end{figure}
Network science is a rich interdisciplinary field of research, and this whistle-stop tour of the basic mathematical concepts cannot do justice to such a field. Nevertheless, we now have the essential tools to understand, at least at a high level, many of the common strategies for reconstructing GRNs.
\section{Data-driven methods}\label{score}
The first class of GRN reconstruction methods considers a fully connected network and associates a weight to each edge by estimating gene dependencies directly from the data. The output of such methods is therefore a weighted network, which can be suitably thresholded to yield the topology of the network. Such methods are generally simple to implement, computationally efficient (they scale with the number of possible edges, which is quadratic in the number of nodes) and have proved often remarkably accurate and effective. For these reasons, some of the most popular tools for GRN inference pertain to this category.
\subsection{Correlation networks}
The simplest score that one may associate to a pair of vector-valued measurements is their correlation. This is computed in the following way: given two zero-mean vectors $\mathbf{v}_i$ and $\mathbf{v}_j$, the (Pearson) correlation between the vectors is given by
\begin{equation}
\mathrm{corr}(\mathbf{v}_i,\mathbf{v}_j)=\rho_{ij}=\frac{\mathbf{v}_i\cdot\mathbf{v}_j}{\Vert\mathbf{v}_i\Vert\Vert\mathbf{v}_j\Vert}\label{Pearson}
\end{equation}
where $\cdot$ indicates the scalar product and $\Vert\mathbf{v}_i \Vert$ is the Euclidean norm of vector $\mathbf{v}_i$ (square root of the sum of the squares of the elements). Practically, given a set of $N$ expression measurements (e.g., different conditions) for $G$ genes, one arranges them into a data matrix $D\in\mathbb{R}^{N\times G}$. Computing correlations between the columns of $D$ yields a $G\times G$ matrix of pairwise gene correlations, which can be taken as the weights of an undirected network and suitably thresholded to obtain a network structure. Variations of this approach involve taking a different measure of correlation (e.g. Kendall's or Spearman's correlation), or raising each correlation to a power to effectively filter out spurious low correlations (weighted correlations).
Correlation networks are extremely simple to implement; their complexity scales linearly with the number of experiments and quadratically with the number of genes, so they can be easily deployed on genome-wide studies with very high numbers of experiments. The assumption that interacting genes should have correlated expression is biologically plausible, and methods such as WGCNA (weighted gene coexpression network analysis \cite{zhang2005general}) have proved consistently reliable and are widely adopted.
Correlation networks however also have some limitations. First, two genes might appear correlated not because they genuinely interact, but because of the effect of a third gene (or several other genes). For example, a high correlation might appear between two genes that share a common regulator. Correlation networks are also unable to distinguish between direct and indirect interactions: if gene $i$ regulates gene $j$ which in turn regulates gene $k$, it is likely that there will be a high correlation between gene $i$ and gene $k$. Correlation networks are therefore vulnerable to false positives. In this respect, {\it partial correlation networks} (see Section \ref{sec:GGM}) offer a conceptually appealing solution to the problem, at the cost of some additional assumptions. Another drawback of correlation networks is that limited sample sizes (which are common in small to medium scale studies) may produce apparent high correlations which are not statistically significant. Furthermore, Pearson correlation is a linear measure of correlation, therefore non-linear regulatory effects might easily be missed, creating a vulnerability to false negatives as well.
Since the correlation is a symmetric metric, correlation networks are intrinsically undirected. Also, correlation is purely a measure of statistical association; therefore, these models are not predictive, in the sense that knowledge of some node values would not allow us to make a quantitative prediction about the remaining nodes.
\subsection{Information theoretic scores}
As we have seen before, the linearity of Pearson correlation may limit its suitability to capture complex regulatory relations. To obviate this problem, several groups have considered alternative scores based on information theory. The main mathematical concept is the mutual information, defined as follows. Let $X$ and $Y$ be two discrete random variables, and let $P(X,Y)$ be their joint probability distribution. The mutual information between the two random variables is then defined as\begin{equation}
\mathrm{MI}[X,Y]=\sum_{x_i,y_j}P(x_i,y_j)\log\frac{P(x_i,y_j)}{P(x_i)P(y_j)}=\sum_{x_i,y_j}P(x_i,y_j)\log\frac{P(x_i\vert y_j)}{P(x_i)}\label{MutInf}
\end{equation}
where $x_j$ and $y_j$ are the values the two random variables can take, and $P(X)$ (resp. $P(Y)$) is the marginal distribution obtained by summing out the values of $Y$ (resp. $X$) in the joint distribution. Intuitively, the Mutual Information quantifies the degree of dependence of the two random variables: it is zero when the two random variables are independent (as is clear from the second formulation in equation \eqref{MutInf}), and, when the two variables are deterministically linked, it returns the entropy of the marginal distribution. The mutual information is still a symmetric score, so mutual information networks are naturally undirected. Nevertheless, it can accommodate more subtle dependencies than the linear correlation score in \eqref{Pearson}, therefore potentially catering for a broader class of regulatory interactions.
In the GRN context, the idea is to replace the probability distributions in \eqref{MutInf} with empirical distributions (estimated from the samples) of gene expression levels for each pair of genes. This gives a weight to each possible edge within a fully connected, weighted undirected network; thresholding at a user-defined parameter then returns a network topology {called {\it relevance network}} \cite{butte1999mutual}. A number of methods have been proposed to filter out indirect or spurious links in relevance networks, the most popular methods being ARACNE \cite{margolin2006aracne}, CLR \cite{faith2007large} and MRNET \cite{meyer2007information}.
Mutual information networks are among the most widely used GRN inference methods. They scale to genome-wide networks, even if they are slightly more computationally intensive than correlation-based methods, as their computational complexity is quadratic in the number of genes and samples. However, they also stop short of providing a predictive framework. Furthermore, estimation of the joint probabilities in equation \eqref{MutInf} might be highly sensitive to noise when the sample size is medium-small.
\subsection{Regression-based methods}
An alternative approach to quantify the dependence of two variables is to predict one from the other. In the simplest case, one may try a linear regression approach, where the slope of the regression line may be used to quantify the strength of the dependence. In a GRN context this would amount to regressing each gene in turn against all other genes in order to obtain network weights. Thus, for every gene $g$, denoting by $x_{gi}$ its expression level in sample $i$, we would solve the regression problem\begin{equation}
x_{gi}=\sum_{j\neq g}w_jx_{ji}+\epsilon_i,\label{LinReg}
\end{equation}
with $\epsilon_i$ a noise term, and use the resulting weight $w_j$ as the weight associated with the network edge between gene $j$ and gene $g$. Notice that in this case the regression formulation naturally gives a direction to the network (even though bidirectional edges are clearly possible).
This idea is at the core of several successful methods for GRN reconstruction. TIGRESS \cite{haury2012tigress} adopts directly the framework of equation \eqref{LinReg}, introducing a L1 regularisation penalty{, which forces some of the weights $w_j$ to be strictly zero,} to ensure the identifiability of the system (in general, unless the number of samples is higher than the number of genes, these are overparametrised systems). An alternative idea is to replace the linear regression model of \eqref{LinReg} with a more flexible, non-parametric regression model. GENIE3 \cite{huynhthu2010inferring}, another widely used method, and subsequent developments \cite{van2014gene,huynh2015combining} also follow this strategy, replacing linear regression with an ensemble of {regression} trees. The score for the edge $(jg)$ is then the importance of gene $j$ in the tree model predicting gene $g$, which can be interpreted as the fraction of variance of the expression of gene $g$ that can be explained by gene $j$. Finally, the regression approach is also extremely popular to handle time series data, with the simple modification that the expression of gene $g$ at time $t$ is regressed against the expression of the other genes {\it at the previous time point} $t-1$ ({\it autoregressive model}) \cite{michailidis2013autoregressive}.{In this book, regression-based methods are discussed in Chapters 8 and 9.}
Methods based on a regression approach are amongst the most popular and scalable approaches for reconstructing directed networks. Compared to other data-driven methods, they are generally computationally more intensive, but they have predictive capability, in the sense that, given the expression of a subset of genes, one may in principle predict the expression levels of the remaining genes. {Moreover, regression-based methods are potentially able to capture high-order conditional dependencies between gene expression patterns, while correlation- and mutual information-based methods only focus on pairwise dependencies.} Practically, the identifiability of {regression} models from limited data may be problematic: different genes often have strongly correlated expression patterns, and (regularised) regression with correlated covariates is notoriously prone to spurious results.
\section{Probabilistic models}
The data-driven based methods described before all start from some statistical or information theoretic measure of dependence, but do not explicitly formulate a model of the data in terms of probabilities. In this section, we briefly introduce two classes of methods that start explicitly from a probabilistic model of the data, using global measures of fit (joint likelihood) or Bayesian approaches to identify the network structure.
\subsection{Gaussian Graphical Models}\label{sec:GGM}
The simplest probabilistic model one may wish to consider is a multivariate normal distribution. The probability density for a multivariate normal vector $\mathbf{x}\in\mathbb{R}^G$ is given by\begin{equation}
p(\mathbf{x}\vert\mathbf{m},\Sigma)=\frac{1}{\sqrt{2\pi\vert\Sigma\vert}}\exp\left[-\frac{1}{2}\left(\mathbf{x}-\mathbf{m}\right)^T\Sigma^{-1}\left(\mathbf{x}-\mathbf{m}\right)\right]\label{MVNdensity}
\end{equation}
where the mean vector $\mathbf{m}$ and variance-covariance matrix $\Sigma$ represent the parameters of the distribution. The off-diagonal entries of the symmetric matrix $\Sigma$ give the covariance between different entries of the random vector $\mathbf{x}$, which is related to the correlation via multiplication by the marginal standard deviations.
An important result is that the inverse of the variance-covariance matrix, the {\it precision matrix} $C=\Sigma^{-1}$, contains the {\it partial correlations} between entries in the random vector $\mathbf{x}$. The partial correlation represents the residual correlation between two variables once the effect of {all the other} variables has been removed. As such, it provides a better measure of association than simple correlation, as it is less vulnerable to spurious associations.
This insight has been effectively used in the context of GRNs by a class of models known as {\it Gaussian Graphical Models} \cite{schafer2004empirical}. The idea is to treat gene expression measurements as a multivariate normal random vector (each entry of the vector representing the expression of one gene), and then estimate the precision matrix from multiple conditions using maximum likelihood estimation. Since this requires estimating a number of parameters which is proportional to the square of the number of genes, regularisation techniques are needed; sparse regularisation techniques such as L1 regularisation (also known as {\it graphical lasso} \cite{graphlasso}) have the added advantage of returning a more interpretable result, with the non-zero entries of the precision matrix representing the edges of the (undirected) regulatory network. Several algorithmic approaches have been proposed to carry out this estimation efficiently, and Gaussian Graphical Models represent a popular network inference approach.{Within this book, Chapter 6 discusses the most recent developments in Gaussian Graphical Models usage.}
While Gaussian Graphical Models are certainly a success story, as usual they come with limitations. Estimating a high-dimensional precision matrix from limited data is difficult, and, while using a consistent estimator such as penalised maximum likelihood brings guarantees in the infinite sample limit, the accuracy of the reconstruction for finite samples is more difficult to quantify a priori. More problematically, Gaussian Graphical Models assume normality of the data, which implies linearity in the relationship between the various genes. While this can be a reasonable approximation, and surprisingly effective inferentially, it certainly is a strong modelling limitation.
\subsection{Bayesian Networks}
All methods described so far address the problem of network reconstruction from a top-down approach: start with a fully connected network, compute pairwise scores (or estimate jointly a precision matrix in the case of Gaussian Graphical Models), and then threshold/ regularise to obtain a sparse network structure. In this subsection we will briefly introduce a very popular class of methods that takes the opposite approach, constructing a joint probabilistic model out of local conditional terms, {\it Bayesian networks}.
The starting point is the product rule of probability, which holds that for any two random variables $X$ and $Y$, $P(X,Y)=P(X\vert Y)P(Y)$. Applying this rule recursively, one has that for $G$ variables\begin{equation}
P(X_1,\ldots,X_G) = P(X_1) \prod_{i=2}^G P(X_i \vert X_1, \ldots, X_{i-1})
\label{prodRule}
\end{equation}
This factorisation is general and clearly not unique, since the ordering of the random variables is arbitrary. Bayesian networks start from this general factorisation, but create structure by imposing that only a subset of all possible variables are relevant in the conditioning set \cite{pearl2014probabilistic}. More formally, for each variable $X_i$, we define the set of {\it parents} of $X_i$, $\pi_i\subset\{X_1,\ldots,X_{i-1},X_{i+1},\ldots,X_G\}$. We then construct a directed network by connecting parents and children (the direction of the arrow goes from parents to children); the network structure corresponds to a special factorisation of the joint probability as \begin{equation}
P(X_1,\ldots,X_G\vert\mathcal{G})=\prod_{i=1}^G P(X_i\vert X_{\pi_i})
\end{equation}
where we introduce the variable $\mathcal{G}$ to denote the graph structure of the Bayesian network. When the parent set $\pi_i$ is empty, the conditional distribution $P(X_i\vert X_{\pi_i})$ is equal to the marginal distribution $P(X_i)$. See Figure \ref{BayesNet} for an example. Two remarks are important: not all parents-children assignments will lead to a valid factorisation of the joint probability distribution. A fundamental result is that only networks without directed loops ({\it directed acyclic graphs, DAGs}) specify valid probability distributions (i.e. you cannot return to the same place walking on the network along the direction of the arrows). This global constraint poses considerable difficulties to reconstruction algorithms. Furthermore, even with the DAG constraint, the correspondence between networks and probability distributions is not one-to-one. As already highlighted in the case of the factorisation \eqref{prodRule}, there can be multiple valid factorisations of a joint probability distribution, leading to different networks encoding exactly the same probability distribution. This issue is known as {\it Markov equivalence} in probability theory; see e.g. \cite{barber2012bayesian} Ch. 3 for more details about the mathematical aspects of graphical statistics.
\begin{figure}[t]
\sidecaption[t]
\includegraphics[width=3cm]{bayesian_network.pdf}
\caption{Example of a valid Bayesian Network with four nodes and four edges. Given this structure $\mathcal{G}$, the joint distribution $P(A,B,C,D|\mathcal{G})$ factorises as $P(A)P(B\vert A)P(C \vert A)P(D \vert B,C)$.}
\label{BayesNet}
\end{figure}
Within a GRN context, Bayesian networks have been hugely popular due to the simplicity with which prior information (e.g. in the form of known interactions) can be incorporated{(see for example Chapter 7 for applications of this paradigm to modern problems)}. As usual, gene expression levels are taken to represent the nodes of the network. For computational convenience, all conditional distributions are generally assumed to be Gaussian or discrete (multinomial), which enables the distributional parameters to be efficiently marginalised. In this way, one can easily compute the marginal likelihood function by evaluating the probability of the data under the model. The outstanding problem then remains the identification of the network structure. This is a very difficult combinatorial optimisation problem. Greedily searching the space of networks structures for an optimum of the likelihood was an early solution \cite{friedman2000using}: although this can be surprisingly effective,
in practice the cardinality of the space of network structures increases super-exponentially with the number of nodes, creating a formidable computational problem. This problem is compounded by the existence of multiple optima (due to Markov equivalence) and by the fact that the search must be constrained by the global DAG condition. As an alternative, Bayesian statistical methods have been extensively studied. This approach usually proceeds by constructing a biased random walk in the space of allowable network structures such that structures with a higher posterior probability are visited more often (a procedure called Markov Chain Monte Carlo) \cite{friedman2003being}. The Bayesian approach has considerable advantages in the ease with which prior information can be encoded, and in the way the intrinsic uncertainty in the system is represented: typically, such methods return an ensemble of plausible network structures, weighted by their posterior probability. Nevertheless, Bayesian methods suffer from considerable computational overheads and, despite recent advances \cite{hill2012bayesian}, the scalability of Bayesian network methods to genome-wide data sets remains a challenge.
\section{Dynamical models}
One of the central questions in biology is how organisms adapt to changing conditions. Therefore, a substantial fraction of high-throughput experiments have a time series design, e.g. they assay the same system at different time points to follow the evolution of the system in time. GRNs play a fundamental role in the mathematical modelling of such processes; unsurprisingly, several GRN reconstruction techniques are tailored towards the analysis of time series data. In this section, we introduce two broadly used classes of methods to infer network structures from dynamic data.
\subsection{Dynamic Bayesian Networks}
As we have seen in the previous chapter, a fundamental requirement on the structure of a Bayesian network is the absence of loops (DAG condition). Within the GRN context, this has long been seen as one of their main limitations: biological systems often exhibit feedback loops as a mechanism to engender robustness and stability. An elegant solution is provided by {\it Dynamic Bayesian Networks} (DBNs), a special class of Bayesian networks adapted for time series data.
DBNs work around the DAG condition by expanding the set of random variables under consideration, so that the nodes of the network now represent expression of genes {\it at a specific time point}. Network edges may now only connect nodes pertaining to different time points, so that a gene can only influence the expression of another gene (or, indeed, itself) at a later time point ({see Figure \ref{DBN} for an example}). In this way, the DAG condition is automatically satisfied, while at the same time biologically plausible features such as feedback mechanisms can be easily incorporated. In most cases, the dynamic structure of a DBN is chosen such that edges are only present between nodes at consecutive time points, with time-independent transition probabilities: this assumption of a homogeneous, first order Markov process is a plausible approximation in many cases, and, particularly when the conditional distributions are chosen to be Gaussian, it allows the modeller to leverage a rich literature on signal processing in autoregressive models.
\begin{figure}
\begin{center}\includegraphics[width=0.75\textwidth]{dynamic_bayesian_network.pdf}\end{center}
\caption{Example of a Dynamic Bayesian Network with four nodes: static representation (with cycles) on the left, and unrolled dynamic representation on the right.}\label{DBN}
\end{figure}
DBNs are extremely popular in the GRN context, and are implemented in several software tools (see \cite{oates2012network} for a recent review,{and also Chapters 2 and 3 in the present volume}). Structure learning within DBNs is easier than in standard Bayesian Networks, since the DAG condition is automatically satisfied, however it still remains computationally demanding, particularly in a Bayesian setting. From the modelling point of view, most implementations assume a linear dynamic model, which is clearly a limitation. Extensions exist which include nonlinear mappings between time points \cite{morrissey2010reverse,aijo2009learning} or that relax the time-homogeneity assumption \cite{grzegorczyk2011non}, however these incur generally higher computational costs and/or place strong restrictions on the class of nonlinear functions allowed. Most often, DBNs are implemented so that each time point in the model corresponds to an observation time. While this is somewhat natural, it constrains all biological processes to have essentially the same time-scale, which can be a serious limit; this is addressed by using a continuous-time semantic within the model, as in the case of continuous time Bayesian Networks \cite{nodelman2002continuous} or, more generally, of differential equation models.
\subsection{Differential equation methods}
Differential equations represent perhaps the best studied and most widely used class of dynamical models in science and engineering. They provide an infinitesimal description of the system dynamics by relating the rate of change (time derivative) of a variable to its value, \begin{equation}
\frac{d\mathbf{x}}{dt}=\mathbf{f}\left(\mathbf{x},\Theta,\mathbf{u}(t),t\right).\label{ODE}
\end{equation}
Here $\mathbf{f}$ is a general, time dependent, vector valued function of the variable of interest $\mathbf{x}$ itself, taking as additional inputs a set of parameters $\Theta$ and possibly also a set of external signals $\mathbf{u}(t)$. When the function $\mathbf{f}$ does not depend explicitly on time, the system is said to be {\it time homogeneous}, and when it does not depend on external inputs it is said to be {\it autonomous}.
Within a GRN context, the variables $\mathbf{x}$ are the expression levels of the set of genes we are interested in modelling, and the interactions between genes are encoded in the parameters $\Theta$. By far the most widely used class of models are linear, autonomous and time homogeneous models, where equation \eqref{ODE} simplifies to\begin{equation}
\frac{d\mathbf{x}}{dt}=A\mathbf{x}\label{linearODE}
\end{equation}
where the parameters $\Theta$ form the interaction matrix $A$. A non-zero entry $A_{ij}$ signifies an influence of gene $j$ on the time evolution of gene $i$, and hence a directed edge between $j$ and $i$ in the GRN.
Equation \eqref{linearODE} or variants thereof are at the core of several methods for inferring GRNs. The Inferelator \cite{bonneau2006inferelator} is one such popular approach, where the derivative on the left hand side of \eqref{linearODE} is approximated with the difference of observed values at consecutive time points, and the network structure is recovered via L1 regularised regression. Other approaches solve directly the differential equation \eqref{linearODE}, positing the solution to be a linear combination of basis functions \cite{trejo2015bayesian} or a draw from a Gaussian process \cite{dondelinger2013ode}, and then take a Bayesian approach to infer the parameters of the differential equation under a suitable, sparsity inducing prior distribution. Finally, the restriction to linear dynamics is not central to methods based on differential equations, and indeed methods using non-linear dynamics (such as Hill kinetics \cite{mcgoff2016local}) have been proposed.{See Chapter 16 for a comprehensive description of state-of-the-art methods for inferring GRNs using differential equations}.
Differential equation models offer several potential advantages: their continuous-time semantics is closer to the class of models used in biophysical approaches to systems biology \cite{alon2006introduction}, so that in principle such approaches can benefit from a more mechanistic interpretation. Employing a continuous-time semantics also has the added advantage of limiting the influence of experimental design decisions (e.g. choice of time points/sampling frequencies) on the final result. In other respects, differential equation models are subject to the same computational hurdles as other methods, and they suffer from similar identifiability issues.
\section{Multi-network models}
All of our previous discussion has assumed that all the data can be explained by a single network structure. While this may be reasonable when all data comes from similar conditions, it is a very strong assumption when one is trying to jointly model data from heterogenous scenarios, as different biological conditions may lead to different pathways being activated, so that effectively different network structures may be more appropriate.
This idea has been fruitfully exploited in two main directions. Several papers have considered the scenario where data (e.g. time series) is available from different, but related conditions. Therefore, one may reasonably assume some commonalities between the underlying network structures, so that methods that can transfer information across conditions are needed. This transfer can be achieved via introducing a shared diversity penalty within different optimisation problems \cite{niculescu2007inductive,Chiquet2011}. Equivalently but more flexibly, {the joint reconstruction of the different networks can be achieved} by adopting a hierarchical Bayesian approach \cite{werhli2007reconstructing,penfold2012nonparametric}.
Another direction that has seen considerable interest is the idea of {\it time-varying networks}. Here, the assumption is that the network structure itself can rewire across time, for example to account for checkpoints during development or cancer evolution. The solution is generally composed of two steps: the identification of the change-points, and a joint learning of related networks across the homogeneous stretches of the time series. This idea has been explored both in the context of optimisation approaches \cite{ahmed2009recovering,robinson2010learning}, and more extensively in a Bayesian scenario \cite{lebre2010statistical,thorne2012inference,dondelinger2013non}.
Some of these ideas are explored in Chapters 2, 10 11 and 13 of this volume.
\section{Evaluation}
During our discussion of various methods for GRN inference, we have often referred to several methods as successful or effective, without specifying how the performance of a particular method may be evaluated. This is a difficult issue: GRN inference methods are motivated precisely by the difficulty of directly measuring regulatory relationships between genes, therefore almost by definition gold standard scenarios where such interactions are known with high confidence are rare. One possibility is the recourse to simulated data. One may employ a biochemically plausible interaction model to generate some simulated gene expression measurements, and then evaluate the accuracy of the method against this gold standard. This strategy has been advocated by major international initiatives such as the Dialogue for Reverse Engineering of Models (DREAM) \cite{marbach2012wisdom}, which has organised a long-running challenge on GRN inference, providing both a stimulus and a benchmark for methodological development. Another direction has been the use of synthetic gene circuits as a benchmark for network reconstruction algorithms. The most well known example of this is probably the IRMA network \cite{cantone2009yeast}, a synthetic network of 5 genes engineered within living yeast cells. While this synthetic biology approach is appealingly close to biological reality, so far technological limitations mean that such an approach has been limited to small networks containing a handful of genes.
Having decided on a benchmark data generation procedure, the next step in evaluating a GRN inference algorithm is the choice of a suitable metric. Naively, one may consider thresholding the algorithm's outputs and reporting an average accuracy in detecting presence or absence of edges. This strategy is however flawed since GRNs are typically very sparse, so that an algorithm constantly predicting the absence of edges would potentially achieve high accuracy. A better strategy is to consider the fraction of true positive calls relative to all positives ({\it sensitivity} or {\it recall}) and the fraction of true positive calls out of all positive calls ({\it precision} or {\it positive predictive value}).
Naturally, precision and recall depend on the threshold chosen: with a very lax cutoff, we will likely recall many true positives (high recall), at the cost of many false positives (low precision). To elucidate the effectiveness of an algorithm in handling the precision/recall trade-off, a visually appealing strategy is the use of {\it precision-recall curves}. These are constructed as follows: given the output of a GRN inference algorithm as a weighted network, one starts by thresholding at a very strict (high) threshold, where precision {is expected to be} high and recall will be low. Decreasing the threshold, one will progressively lower precision by introducing some false positives, but also increase recall, until at zero threshold (fully connected network) recall is 1 and precision is the fraction of actual edges over possible edges (positives fraction). This procedure results in a curve in precision-recall space (see Figure \ref{ROC_PR}, right panel, for an illustration) indicative of the overall performance of the inference algorithm: a random predictor will always have an expected precision equal to the positives fraction, while an ideal algorithm will have precision 1 for any {recall between 0 and 1}. These observations justify the use of the {\it area under the curve} as a global metric of performance for an algorithm, a choice almost universally adopted in evaluating GRN inference methods.
\begin{figure}
\begin{center}\includegraphics[width=0.95\textwidth]{ROC_PR_curves.pdf}\end{center}
\caption{Evaluation of inferred networks: an algorithm typically outputs a ranked list of edges, with the weight of each edge being given by either a score or a posterior probability (left panel, where true and false edges are coloured in yellow and red, respectively). By progressively lowering the threshold for acceptance, one can construct either a ROC curve (central panel) or a precision-recall curve (right panel).}\label{ROC_PR}
\end{figure}
{Alternatively, a receiver operating characteristic (ROC) curve may be used to evaluate a weighted network against a gold standard. A ROC curve plots the recall versus the {\it false positive rate} (the fraction of false positive calls relative to all negatives) for different thresholds on the weights, again progressively lowering the threshold. Precision-recall curves are however more suited than ROC curves for problems where the number of negatives is much higher than the number of positives, which is typically the case of GRNs} \cite{davis2006relationship}.
\section{Software tools}
Most of the methods described above have been implemented in software tools which have been made freely available to the community. As it is perhaps to be expected of such a diverse and dynamic field, no single method has yet emerged as an industry standard, and tools differ widely in their usability and implementation. We provide here a summary list of some of the main software tools, as a reference list for the practitioner. All information is up-to-date at the time of writing (November 2017), and may clearly change. Naturally, this list is incomplete, and we would like to stress that any omissions do not reflect a judgement on the methods, but rather a restriction in space.
\begin{itemize}
\item{WGCNA, weighted correlation network analysis, an R package available from the comprehensive R archive CRAN.\\\small\url{https://labs.genetics.ucla.edu/horvath/CoexpressionNetwork/Rpackages/WGCNA/index.html}}
\item{ARACNe, mutual information based network inference approach. Source code in C++ available, as well as several OS-compatible versions and plugins.\\
\small\url{http://califano.c2b2.columbia.edu/aracne/}}
\item{CLR, context likelihood of relatedness, mutual information based network inference approach, originally implemented in MATLAB.\\
\small\url{http://m3d.mssm.edu/network_inference.html}}
\item{MRNET, mutual information based network inference approach. R implementation available in the Bioconductor package minet (also contains R implementations of ARACNe and CLR).\\
\small\url{https://www.bioconductor.org/packages/release/bioc/html/minet.html}}
\item{GENIE3 and other tree based methods, available as MATLAB, Python and R packages.\\
\small\url{http://www.montefiore.ulg.ac.be/~huynh-thu/software.html}}
\item{GeneNet, R package implementing Gaussian Graphical Models network inference, available from CRAN.\\\small\url{https://cran.r-project.org/web/packages/GeneNet/index.html}}
\item{CatNet, R package for (discrete) Bayesian Network structure learning, available from CRAN.\\
\small\url{https://cran.r-project.org/web/packages/catnet/index.html}}
\item{Banjo, Java package for Bayesian Networks structure learning.\\
\small\url{https://users.cs.duke.edu/~amink/software/banjo/}}
\item{G1DBN, R package for Dynamic Bayesian Network inference, available from CRAN.\\
\small\url{https://cran.r-project.org/web/packages/G1DBN/index.html}}
\item{GRENITS, Bioconductor package for Dynamic Bayesian Network inference.\\
\small\url{https://bioconductor.org/packages/release/bioc/html/GRENITS.html}}
\item{TSNI, differential equations based method, available as MATLAB package.\\
\small\url{http://dibernardo.tigem.it/softwares/time-series-network-identification-tsni}}
\item{Inferelator, differential equations based method.\\
\small\url{http://bonneaulab.bio.nyu.edu/networks.html}}
\item{netbenchmark, R package for benchmarking GRN inference methods (also contains R implementations of several methods such as ARACNe, C3NET, CLR, GeneNet and GENIE3).\\
\small\url{https://www.bioconductor.org/packages/release/bioc/html/netbenchmark.html}}
\end{itemize}
\section{Discussion and outlook}
GRN inference is a mature field of methodological research, with widespread and increasing applications in biomedical research. In this chapter, we have attempted a broad brush introduction to the field, highlighting the biological motivation and the technological advances in data collection that have underpinned its recent flourishing. We then proceeded to give a bird's eye view of the statistical principles underpinning some of the most popular methodologies for GRN inference. Our focus has been on the foundations, attempting a coarse categorisation of different methods based on their assumptions and semantics. Of course, many interesting contributions fall at the intersection of different categories, and are not well accommodated by our simplifying approach.
Naturally, it is impossible to do justice to a rich and wide research area within a short introductory review. Our aim here is to prepare the reader for more advanced concepts to be described in subsequent chapters of this book; nevertheless, we hope that this chapter will also form a worthwhile introduction for the novice to the field, and have attempted to make it as self contained as possible.
\begin{acknowledgement}
GS acknowledges support from the European Research Council under grant MLCS 306999. VAHT is a Post-doctoral Fellow of the F.R.S.-FNRS.
\end{acknowledgement}
\bibliographystyle{spbasic_unsort}
|
\section*{Prologue}
The set of all projective structures on a surface turns out to be too large to yield interesting general results, and one therefore restricts to what are called \emph{properly convex projective structures} or \emph{strictly convex projective structures}. The foundations of this study were laid by the work of Kuiper (1953, \cite{Kuiper-surfaces-1953, Kuiper-convex-1954}), Vinberg and Kac (1967, \cite{Vinberg-quasi-1967}) and Benzecri (1960, \cite{Benzecri-varietes-1960}). Parameterisations, structure and dimension of the moduli spaces were given by Goldman (1990, \cite{Goldman-convex-1990}), Choi--Goldman (1993, \cite{Choi-convex-1993}) and Marquis (2010, \cite{Marquis-espace-2010}). Finding good parameterisations is the key to applications and further advances of the theory. In these notes, a particular parameterisation due to Fock and Goncharov~(2007, \cite{Fock-moduli-2007}) is used to give a self-contained treatment of the key facts about the moduli spaces of these structures on surfaces. We have elaborated many subtle details that are not made explicit in \cite{Fock-moduli-2007}, and hope that this expanded treatment of their work gives a nice introduction to the study of moduli spaces of real projective structures on surfaces.
For foundational and further material on real projective manifolds we refer to Goldman~\cite{Goldman-convex-1990}, Marquis~\cite{Marquis-espace-2010}, Benoist~\cite{Benoist-survey-2008}, Cooper-Long-Tillmann~\cite{Cooper-convex-2015} and the references therein.
Much of the material presented here is based on a seminar series at the University of Sydney, and we heartily thank the following people who lectured or participated in the talks: Grace Garden, Montek Gill, Robert Haraway, James Parkinson, and Robert Tang. The last author learned most of what he knows about projective geometry from Daryl Cooper, who introduced him to this rich, puzzling and at times amusing field of research. He also thanks Bill Goldman for first bringing the work of Fock and Goncharov to his attention. The authors acknowledge support by the Commonwealth of Australia and the Australian Research Council (DP140100158).
\newpage
\begin{spacing}{0.5}
\tableofcontents
\end{spacing}
\newpage
\input{01_intro}
\newpage
\input{03_flags}
\newpage
\input{04_modulispace}
\input{05_poisson}
\newpage
\input{06_examples}
\newpage
\addcontentsline{toc}{section}{References}
\bibliographystyle{plain}
\section{Introduction}
Let $\RR\PP^2$ denote the real projective plane and $\PGL(3, \mathbb{R})$ denote the group of projective transformations $\RR\PP^2\to\RR\PP^2$. A \emph{convex projective surface} is a quotient $\Omega / \Gamma,$ where $\Omega \subset \RR\PP^2$ is an open convex subset and $\Gamma$ is a discrete (torsion-free) subgroup of $\PGL(3, \mathbb{R})$ that leaves $\Omega$ invariant. Examples of convex projective surface include Euclidean tori (via the embedding of the Euclidean plane in $\RR\PP^2$ as an affine patch), and hyperbolic surfaces (via the Klein model of the hyperbolic plane). Some convex projective structures on the once-punctured torus are indicated in Figure~\ref{fig:clover_pictures}.
\begin{figure}[h]
\centering
\includegraphics[width=4cm]{sage3.pdf} \qquad
\includegraphics[width=4cm]{sage5.pdf} \qquad
\includegraphics[width=4cm]{sage6.pdf}
\caption{Developing maps for convex projective structures on the once-punctured torus \cite{Haraway-tessellating-2018}}
\label{fig:clover_pictures}
\end{figure}
This paper is concerned with \emph{moduli spaces} of convex projective structures on surfaces satisfying additional hypotheses, which will now be described.
Let $S$ be a smooth surface. A \emph{marked properly convex projective structure} on $S$ is a triple $( \Omega, \Gamma, f),$ where
\begin{itemize}
\item $\Omega \subset \RR\PP^2$ is an open subset of the real projective plane whose closure is contained in an affine patch and which is convex in that patch,
\item $\Gamma$ is a discrete (torsion-free) subgroup of the group $\PGL(3, \mathbb{R})$ of projective transformations that leaves $\Omega$ invariant,
\item $f \co S \to \Omega / \Gamma$ is a homeomorphism.
\end{itemize}
Denote $\widetilde{S}$ the universal cover of $S.$ The map $f \co S \to \Omega / \Gamma$ lifts to the \emph{developing map} $\dev \co \widetilde{S} \to \RR\PP^2$ with image equal to $\Omega$, and induces the \emph{holonomy representation} $\hol \co \pi_1(S) \to \PGL(3, \mathbb{R})$ with image equal to $\Gamma.$
Two marked properly convex projective structures $( \Omega_0, \Gamma_0, f_0) $ and $( \Omega_1, \Gamma_1, f_1)$ are equivalent if there is $A \in \PGL(3, \mathbb{R})$ such that $A \cdot \Omega_0 = \Omega_1,$ $A \cdot \Gamma_0 \cdot A^{-1} = \Gamma_1$ and $f_1$ is isotopic to $A^* \cdot f_0$, where $A^*$ is the homeomorphism $\Omega_0 / \Gamma_0 \rightarrow \Omega_1 / \Gamma_1$ induced by $A$. The set of all such equivalence classes, which we will refer to as the \emph{moduli space of marked properly convex projective structures}, is denoted $\mathcal{T}_3(S)$. As such, we view a moduli space as a space of objects, and seek suitable parameterisations of this moduli space.
The subscript alludes to the place this moduli space holds in the more general study of discrete and faithful representations of $\pi_1(S)$ into $\PGL(n, \mathbb{R})$; see \cite{Fock-moduli-2006, Labourie-cross-2007, Guichard-convex-2008, Bonahon-parameterizing-2014, Gueritaud-compactification-2017, Kapovich-dynamics-2018}. In the case of properly convex projective structures on a closed surface, no further qualification is necessary:
\begin{thm}[Goldman 1990, \cite{Goldman-convex-1990}]\label{thm:Goldman-main}
If $S$ is a closed, orientable surface of negative Euler characteristic, then $\mathcal{T}_3(S)$ is an open cell of dimension $- 8 \chi(S).$
\end{thm}
Suppose that $S_g$ is a closed, orientable surface of finite genus $g.$ Let $D_1, \ldots, D_n$ be open discs on $S_g$ with pairwise disjoint closures. Choose $p_k \in D_k.$ Then $S_{g,n} = S_g \setminus \{p_1, \ldots, p_n\}$ is a \emph{punctured surface} and we call $E_k = \overline{D_k} \setminus \{p_k\}$ an \emph{end} of $S_{g,n}.$ A \emph{compact core} of $S_{g,n}$ is $S_{g,n}^c = S_g \setminus \cup D_k.$ We also write $S_g = S_{g,0}.$
Goldman's theorem states that $\mathcal{T}_3(S_g)$ is an open cell of dimension $16g-16.$ The most immediate generalisation of Goldman's theorem to punctured surfaces makes the additional hypothesis that the structures on $S$ have \emph{finite volume}. The volume arises in this context by taking the Hausdorff measure computed from the \emph{Hilbert metric} on the domain $\Omega,$ though we remark that there are other ways to define volume that can be used in our context. Denote $\mathcal{T}_3^{f}(S) \subseteq \mathcal{T}_3(S)$ the subset of all finite volume structures.
\begin{thm}[Marquis 2010, \cite{Marquis-espace-2010}]\label{thm:Marquis-main-intro}
If $S = S_{g,n}$ has negative Euler characteristic and at least one puncture, then $\mathcal{T}_3^{f}(S)$ is an open cell of dimension $16g-16+6n.$
\end{thm}
Note that $S = S_{g,n}$ has negative Euler characteristic if and only if $2g+n>2.$ All ends in a structure of finite volume are cusps in the same manner that non-compact complete hyperbolic surfaces of finite volume have cusps at their ends.
When considering structures of infinite area in $\mathcal{T}_3(S)$, the first consideration is to enumerate the various possibilities for the geometry at each end of $S$. Consider the interior of a hyperbolic surface $F$ with totally geodesic boundary. One can, in hyperbolic space, equivariantly add a convex collar to the boundary. This collar could be smooth or piecewise linear or a combination thereof. Hence there are infinitely many inequivalent points in $\mathcal{T}_3(F)$ whose structures agree on a compact core of $F.$ In particular, all of these points have the same holonomy. Returning to the general case of a surface $S = S_{g,n}$, denote by $\Hol(S)$ the set of all holonomies of strictly convex projective structures on $S.$ The equivalence relation between structures that defines $\mathcal{T}_3(F)$
descends to the action of $\PGL(3,\mathbb{R})$ by conjugation on $\Hol(S),$ and we denote the quotient by
\[ \mathfrak{X}(S) = \Hol(S)/\PGL(3,\mathbb{R}).\]
Following Fock and Goncharov~\cite{Fock-moduli-2007}, we will now \emph{frame} the holonomies. For each end $E_k$ of $S,$ there is a \emph{peripheral subgroup} $\Gamma_k\le \Gamma$ corresponding to $\im (\pi_1(E_k) \to \pi_1(S)).$ Since $\Gamma_k \cong \ZZ,$ it fixes at least one point in $\RR\PP^2$ and preserves at least one line in $\RR\PP^2$ through that point.
A \emph{framing} of the holonomy $\hol \co \pi_1(S) \to \PGL(3, \mathbb{R})$ consists of the choice, for each end $E_k$, of such an invariant flag consisting of a point and a line. The set of framed holonomies is denoted
\[\Hol^\times(S_{g,n}) \subset \Hom\big(\pi_1(S_{g,n}), \PGL(3, \mathbb{R})\big) \times \big( \RR\PP^2 \times (\RR\PP^2)^*\big)^n\]
There is a natural action of $\PGL(3, \mathbb{R})$ on $\Hol^\times(S),$ where the action of an element $A$ of $\PGL(3, \mathbb{R})$ on the dual plane $(\RR\PP^2)^*$ is given by taking the inverse of the transpose of $A.$
We denote the quotient by
\[ \mathfrak{X}^\times(S) = \Hol^\times(S)/ \PGL(3, \mathbb{R}). \]
\begin{thm}[Fock-Goncharov 2007, \cite{Fock-moduli-2007}]\label{thm:FG-main-X}
If $S = S_{g,n}$ has negative Euler characteristic and at least one puncture, then $\mathfrak{X}^\times(S)$ is an open cell of dimension $- 8 \chi(S) = 16g-16+8n.$
\end{thm}
The above result is not stated in this form in \cite{Fock-moduli-2007}. Indeed, what is hidden in the statement is the ingenious observation in \cite{Fock-moduli-2007} that there is a section of $\Hol^\times(S)\to \mathfrak{X}^\times(S)$ which consists of the so-called \emph{positive representations}. To describe them, we will introduce two constructions of \emph{framed} marked properly convex projective structures on $S$ using developing maps of ideal triangulations. The distinction between these constructions is not made explicit in \cite{Fock-moduli-2007}, so we will briefly explain it.
In the above example of the hyperbolic surface $F$, we concluded that there are infinitely many inequivalent points in $\mathcal{T}_3(F)$ whose structures agree on a compact core of $F.$ This situation was remedied by focussing on the associated holonomies, which were subsequently decorated with flags. Similarly, one can frame a properly convex projective structure as follows. Given $( \Omega, \Gamma, f),$ a framing is the $\Gamma$--invariant choice for each peripheral subgroup of a fixed point $p$ in the frontier of $\Omega$ and an invariant supporting line to $\Omega$ at $p.$ Denote the resulting space $\mathcal{T}_3^\times(S).$ Then there is a natural map $\mathcal{T}_3^\times(S)\to \mathfrak{X}^\times(S).$
The flags can then be used to construct two different sections for this map.
The first section chooses, for each element in $\mathfrak{X}^\times(S)$ the interior of the convex hull of the $\Gamma$--orbits of the fixed points in all flags, framed with the corresponding lines. This convex hull will be represented as an infinite \emph{union} of ideal triangles, and the set of these framed structures is denoted $\mathcal{T}_3^+(S)\subset \mathcal{T}_3^\times(S).$ The second section chooses, for each element in $\mathfrak{X}^\times(S)$ the intersection of the $\Gamma$--orbits of half-spaces associated to the lines in all flags. This is represented as an infinite \emph{intersection} of ideal triangles and the set of these with the appropriate framing is denoted $\mathcal{T}_3^-(S)\subset \mathcal{T}_3^\times(S).$
In particular, Theorem~\ref{thm:FG-main-X} follows from the following result.
\begin{thm}[Fock-Goncharov 2007, \cite{Fock-moduli-2007}]\label{thm:FG-main}
If $S = S_{g,n}$ has negative Euler characteristic and at least one puncture, then $\mathcal{T}_3^+(S)$ and $\mathcal{T}_3^-(S)$ are open cells of dimension $- 8 \chi(S) = 16g-16+8n.$
\end{thm}
The key lies in a parametrisation of these spaces using an ideal triangulation of $S.$ Fock and Goncharov associate to each triangle in the triangulation a \emph{triple ratio of flags} and to each oriented edge a \emph{quadruple ratio of flags}, and show how to turn this into bijections $\psi^+ \co \mathbb{R}_{>0}^{- 8 \chi(S)}\to \mathcal{T}_3^+(S)$ and $\psi^- \co \mathbb{R}_{>0}^{- 8 \chi(S)}\to \mathcal{T}_3^-(S)$ with the property that the following diagram commutes:
\begin{center}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em] {
& \mathcal{T}_3^+(S) & \\
\mathbb{R}_{>0}^{- 8 \chi(S)} & & \mathfrak{X}^\times(S) \\
& \mathcal{T}_3^-(S) & \\};
\path[-stealth]
(m-2-1) edge node [above] {$\psi^+$} (m-1-2)
(m-2-1) edge node [below] {$\psi^-$} (m-3-2)
(m-1-2) edge node [above]{$\mu^+$} (m-2-3)
(m-3-2) edge node [below]{$\mu^-$} (m-2-3);
\end{tikzpicture}
\end{center}
The map $\mu^+\circ\psi^+ = \mu^-\circ\psi^-$ has a lift $\mathbb{R}_{>0}^{- 8 \chi(S)}\to \Hol^\times(S)$ that is explicitly described in terms of rational functions where each numerator and denominator is a polynomial with only positive coefficients. These are the so-called \emph{positive representations}. For each $\star \in \{ +, -\}$, the corresponding \emph{monodromy map} $\mu^\star\co \mathcal{T}_3^\star(S)\to \mathfrak{X}^\times(S)\to \mathfrak{X}(S)$ is surjective and generically $6^n$--to--$1$, and corresponds to an action of (the direct product of $n$ copies of) the symmetric group in three letters.
We remark that the set $\mathcal{T}_3^\pm(S) = \mathcal{T}_3^+(S) \cap \mathcal{T}_3^-(S)$ where the two constructions agree has the property that the composition $\mathcal{T}_3^\pm(S)\to \mathfrak{X}^\times(S)\to \mathfrak{X}(S)$ is surjective and generically $2^n$--to--$1.$ Many people we talked to in the initial phase of this project had the impression that Fock and Goncharov's coordinates parameterise the set $\mathcal{T}_3^\pm(S).$ This may arise from the comment after Definition 2.1 in \cite{Fock-moduli-2007} that the set $\mathcal{T}_3^+(S)$ is a $2^n$--to--$1$ cover of the space of non-framed convex real projective structures on $S$ \emph{with geodesic boundary}.
In this interpretation, one compactifies those ends of $S= \Omega/\Gamma$ by adding a circle, where a corresponding supporting line in the framing meets the closure of $\Omega$ in more than one point.
The duality principle of projective geometry gives rise to natural \emph{duality maps} $\sigma^+ \co \mathcal{T}_3^+(S) \to \mathcal{T}_3^-(S)$ and $\sigma^- \co \mathcal{T}_3^-(S) \to \mathcal{T}_3^+(S)$ defined by taking a framed structure to a dually framed dual structure. These are inverse to each other. There also is a natural dual map $\sigma \co \mathfrak{X}^\times(S) \to \mathfrak{X}^\times(S)$ taking the character of the representation $\rho$ to the character of the inverse transpose of the representation, and each flag to the dual flag. We again obtain a commutative diagram:
\begin{center}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em] {
\mathcal{T}_3^+(S) & \mathfrak{X}^\times(S) \\
\mathcal{T}_3^-(S) & \mathfrak{X}^\times(S)\\};
\path[-stealth]
(m-1-1) edge node [above] {$\mu^+$} (m-1-2)
(m-1-1) edge [bend right=10] node [left] {$\sigma^+$} (m-2-1)
(m-2-1) edge [bend right=10] node [right] {$\sigma^-$} (m-1-1)
(m-2-1) edge node [above]{$\mu^-$} (m-2-2)
(m-1-2) edge node [left]{$\sigma$} (m-2-2)
(m-2-2) edge (m-1-2);
\end{tikzpicture}
\end{center}
In particular, the action of duality on the coordinate space $\mathbb{R}_{>0}^{- 8 \chi(S)}$ is given by $(\psi^-)^{-1} \circ \sigma^+ \circ \psi^+ = (\psi^+)^{-1} \circ \sigma^- \circ \psi^-.$
We will give a proof of Theorem~\ref{thm:FG-main} using the methods of \cite{Fock-moduli-2007}. We also highlight various properties of the moduli space, including an explicit description of the duality map and the action of the mapping class group, in \S\ref{sec:Properties of the parameterisation}.
As consequences of Theorem~\ref{thm:FG-main} we then deduce Goldman's Theorem~\ref{thm:Goldman-main}, Marquis' Theorem \ref{thm:Marquis-main-intro}, as well as a classical result due to Fricke and Klein about the classical Teichm\"uller space of finite area hyperbolic structures on $S$ in \S\ref{sec:Applications of the parameterisation}.
As a last application, we focus on the Poisson structure on the moduli space $\mathcal{T}_3^+(S)$ in \S \ref{sec:poisson-structure}.
Natural symplectic structures and Poisson structures may often be used to distinguish between homeomorphic spaces and open the door to the world of integrable systems; see Audin~\cite{Audin-lectures-1997} for a discussion in the context of this paper.
In the case of closed surfaces, Goldman~\cite{Goldman-symplectic-1984} gives a completely general result via the intersection pairing. Let $G$ be a Lie group preserving a non-degenerate bilinear form on its Lie algebra., and let $\mathfrak{X}_G(S) := \Hom(\pi_1(S), G)\mathbin{/\mkern-6mu/} G$ be the $G$--character variety of $S.$ Then Goldman \cite{Goldman-symplectic-1984} showed that $\mathfrak{X}_G(S_{g,0})$ has a natural symplectic structure.
In the case of a punctured surface, where a natural symplectic structure is not available on $\mathfrak{X}_G(S_{g,n})$, it can still be foliated by symplectic leaves making it into a Poisson variety \cite{Guruprasad-group-1997}. For $G = \SL(m,\mathbb{R})$, character varieties of surfaces with the same Euler characteristic are isomorphic as varieties, while they may be distinguished using their Poisson structure. A natural question to ask is whether this also applies to the moduli space $\mathcal{T}_3^+(S).$ Following Fock and Goncharov~\cite{Fock-moduli-2007}, we explicitly construct a Poisson bracket on $\mathcal{T}_3^+(S).$
The monodromy map $\mu^+$ induces an injective homomorphism $\mu^*$ between the spaces of smooth functions on $\mathcal{T}_3^+(S)$ and on $\mathfrak{X}(S)$. When we restrict to the trace algebra $\Tr(\mathfrak{X}(S))$, we have the following result (see Theorem \ref{thm:compatibility-poisson-bracket} for details):
\begin{thm}
$\mu^*$ is a Poisson homomorphism between $\mathcal{T}_3^+(S)$ and $\mathfrak{X}(S)$.
\end{thm}
We provide a detailed geometric proof of this result, which appears to be known to experts (see, for example, Audin \cite{Audin-lectures-1997} or Xie \cite{Xie-structures-2013}), but the homomorphism does not appear to have been described explicitly before.
\section{Ratios and Configurations of Flags} \label{sec:ratios-and-config-flags}
This section is devoted to the development of the main tools used in the proof of Theorem \ref{thm:global-coord}, in \S\ref{sec:Moduli-Spaces}. Most of the definitions and conventions are taken from \cite{Fock-moduli-2007}. We introduce flags of $\RR\PP^2$ and study their properties. The rest of the chapter is completely devoted to developing the necessary tools to prove the main results in \S\ref{sec:Moduli-Spaces}: triple ratios, cross ratios, quadruple ratios and configurations of flags.
\subsection{Flags}
We introduce some notation and recall well-known results from projective geometry. The homomorphism $\GL(3,\mathbb{R}) \to \SL(3,\mathbb{R})$ defined by
\[ A \mapsto (\det(A))^{-1/3} A\]
descends to an isomorphism $\PGL(3,\mathbb{R}) \to \SL(3,\mathbb{R}).$ We will therefore work with $\SL(3,\mathbb{R}),$ which allows us to talk about eigenvalues of maps, and it will be clear from context whether an element of $\SL(3,\mathbb{R})$ acts on $\RR\PP^2$ or on $\mathbb{R}^3.$ We also remark that the action on $\mathbb{R}^3$ is by orientation preserving maps.
Points and lines of $\RR\PP^2$ are denoted by column and row vectors, respectively. In particular, a line $\left[ a : b : c \right]$ corresponds to the set of points $\left[ x : y : z \right]^t$ of $\RR\PP^2$, satisfying $ax + by + cz = 0$. Thus a point $P$ belongs to a line $l$ if and only if $l(P) = 0$. Four points (resp. four lines) of $\RR\PP^2$ are in \emph{general position} if no three are collinear (resp. no three are incident). They are often referred as a \emph{projective basis}, because:
\begin{itemize}
\item $\SL(3,\mathbb{R})$ is simply transitive on ordered $4$--tuples of points in general position;
\item $\SL(3,\mathbb{R})$ is simply transitive on ordered $4$--tuples of lines in general position.
\end{itemize}
Up to projective transformation, there is a canonical dual map between points and lines of $\RR\PP^2$. That is the following one:
\begin{align*}
\perp : \RR\PP^2 &\rightarrow (\RR\PP^2)^*,\\
P = \left[
\begin{array}{c}
x\\ y\\ z
\end{array} \right] &\mapsto P^\perp = \left[ x : y : z\right].
\end{align*}
A \emph{flag} $\mathcal{F}_i := (V_i,\eta_i)$ of $\RR\PP^2$ is a pair consisting of a point $V \in \RR\PP^2$ and a line $\eta \subset \RR\PP^2$ passing through $V$. An $m$--tuple of flags $\{\mathcal{F}_1,\dots,\mathcal{F}_m \}$ is in \emph{general position} if
\begin{itemize}
\item no three points are collinear;
\item no three lines are coincident;
\item $\eta_i(V_j) = \delta_{ij}$, the Kronecker delta.
\end{itemize}
Henceforth, we will denote by $((\mathcal{F}_1,\dots,\mathcal{F}_m))$ a cyclically ordered $m$--tuple of flags, as opposed to an ordered $n$--tuple $(\mathcal{F}_1,\dots,\mathcal{F}_m)$. The group of projective transformations acts on the space of flags via
$$
T \cdot F_i := ( T \cdot V_i , T \cdot \eta_i ), \qquad T \in \SL(3,\mathbb{R}).
$$
The action naturally extends to $m$--tuples, ordered $m$--tuples and cyclically ordered $m$--tuples of flags.
\subsection{Triple ratio}
\label{subsec_triple_ratio}
Suppose $\mathfrak{F} = ((\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2))$ is a cyclically ordered triple of flags in general position. We define the \emph{triple ratio} of $\mathfrak{F}$ as
$$
\cancel{3}(\mathfrak{F}) := \frac{\overline{\eta_0}(\overline{V_1}) \cdot \overline{\eta_1}(\overline{V_2}) \cdot \overline{\eta_2}(\overline{V_0})}{\overline{\eta_0}(\overline{V_2}) \cdot \overline{\eta_1}(\overline{V_0}) \cdot \overline{\eta_2}(\overline{V_1})},
$$
where $\overline{V_i}$ and $\overline{\eta_i}$ are some class representatives. The above definition is well-defined, as it is independent of the choice of representatives, and manifestly invariant under a cyclic permutation of the flags. The following property is easy to verify.
\begin{lem} \label{lem:triple_ratio_proj_invariant}
$\cancel{3}$ is invariant under projective transformations, $i.e.$ for all $T \in\SL(3,\mathbb{R})$
$$
\cancel{3}(T \cdot \mathfrak{F}) = \cancel{3}(\mathfrak{F}).
$$
\end{lem}
\subsection{Cross ratio}
\label{subsubsec:positivity_of_cross}
Let $l \subset \RR\PP^2$ be a line. Given $P_0,P_1,P_2,P_3 \in l$ with $P_0,P_1,P_2$ pairwise distinct, let $T: l \rightarrow \mathbb{R} \PP^1$ be the unique projective map such that $T(P_0) = \infty$, $T(P_1) = -1$ and $T(P_2) = 0$. Then the \emph{cross ratio} of the ordered quadruple $(P_0,P_1,P_2,P_3)$ is
$$
\cross(P_0,P_1,P_2,P_3) = T(P_3).
$$
Just as for the triple ratio, it is easy to check that the cross ratio is invariant under projective transformation.
Moreover, if $x_0,x_1,x_2,x_3$ are local coordinates for $P_0,P_1,P_2,P_3 \in l$, then
$$
\cross(P_0,P_1,P_2,P_3) = \frac{(x_0-x_1)(x_2-x_3)}{(x_0-x_3)(x_1-x_2)}.
$$
It follows that, if $\sigma$ is a permutation on four symbols and $\lambda = \cross(P_0,P_1,P_2,P_3)$, then
$$
\cross(P_{\sigma(0)},P_{\sigma(1)},P_{\sigma(2)},P_{\sigma(3)}) \in \left\{ \lambda,\frac{1}{\lambda},1-\lambda,\frac{1}{1-\lambda},1-\frac{1}{\lambda},\frac{\lambda}{\lambda-1}\right\}.
$$
In particular
$\cross(P_3,P_2,P_1,P_0) = \cross(P_0,P_1,P_2,P_3)$.
Similarly, we can define the cross ratio of four incident lines via duality. Let $P \in \RR\PP^2$ be a point and $l_0,l_1,l_2,l_3$ lines through $P$ with $l_0,l_1,l_2$ pairwise distinct. Then the \emph{cross ratio} of the ordered quadruple $(l_0,l_1,l_2,l_3)$ is
$$
\cross(l_0,l_1,l_2,l_3) := \cross(l_0^\perp,l_1^\perp,l_2^\perp,l_3^\perp).
$$
A straightforward argument shows:
\begin{lem} \label{lem:cr_lines_points}
Let $m$ be a line intersecting the lines $l_0,l_1,l_2,l_3$ transversely in $Q_0,Q_1,Q_2,Q_3$ respectively, then
$$
\cross(l_0,l_1,l_2,l_3) = \cross(Q_0,Q_1,Q_2,Q_3).
$$
\end{lem}
The convention for the cross ratio used here was suggested by Fock and Goncharov~\cite[pg. 253]{Fock-moduli-2007}. It is motivated by the fact that, if $P_0,P_1,P_2$ are pairwise distinct points on a line $l \subset \RR\PP^2$, we would like $\cross(P_0,P_1,P_2,P_3) > 0$ when $P_3$ and $P_1$ lie in different components of $l\setminus \{P_0,P_2\}$. Lemma \ref{lem:cr_lines_points} implies that if $l_0,l_1,l_2$ are pairwise distinct lines passing through a point $P \in \RR\PP^2$, then $\cross(l_0,l_1,l_2,l_3) > 0$ when $l_3$ and $l_1$ lie in different regions of $\RR\PP^2 \setminus \{l_0,l_2\}$. See Figure~\ref{cross_ratios}.
\begin{figure}[ht]
\centering
\includegraphics[width=12.5cm]{cross_ratios.png}
\caption{$\cross(P_0,P_1,P_2,P_3) > 0$ when $P_3$ and $P_1$ lie in different components of $l\setminus \{P_0,P_2\}$. Likewise, $\cross(l_0,l_1,l_2,l_3) > 0$ when $l_3$ and $l_1$ lie in different regions of $\RR\PP^2 \setminus \{l_0,l_2\}$.}
\label{cross_ratios}
\end{figure}
Henceforth, if $P,Q \in \RR\PP^2$ are two points and $m,l \subset \RR\PP^2$ are two lines, we will denote by $PQ$ the line passing through $P$ and $Q$, and by $lm$ the point of intersection between $l$ and $m$.
\begin{thm} \label{thm:cross_triple}
Let $\mathfrak{F} = ((\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2))$ be a cyclically ordered triple of flags in general position where $\mathcal{F}_i = (V_0,\eta_0)$. Then
$$
\cross(\eta_0,V_0V_1, V_0(\eta_1\eta_2),V_0 V_2) =\cancel{3}(\mathfrak{F}).
$$
\end{thm}
\begin{proof}
Both cross ratio and triple ratio are projectively invariant, and the points $V_0,V_1,V_2, \eta_1 \eta_2$ are in general position so we may assume, without loss of generality, that
$$
V_0 = \left[
\begin{matrix}
0 \\ 0 \\ 1
\end{matrix} \right], \quad
V_1 = \left[
\begin{matrix}
1 \\ 0 \\ 1
\end{matrix} \right], \quad
V_2 = \left[
\begin{matrix}
0 \\ 1 \\ 1
\end{matrix} \right], \quad
\eta_1 \eta_2 = \left[
\begin{matrix}
1 \\ 1 \\ 1
\end{matrix} \right].
$$
It follows that
$$
\eta_1 = \left[
\begin{matrix}
1 : 0 : -1
\end{matrix} \right], \quad
\eta_2 = \left[
\begin{matrix}
0 : 1 : -1
\end{matrix} \right].
$$
The projective line $\eta_0$ passes through $V_0$ but does not pass through $V_2$ or $V_1$, so $\eta_0 = [A:1:0]$ for some $A \in \mathbb{R} \setminus \{0\}$. By Lemma \ref{lem:cr_lines_points},
$$
\cross(\eta_0,V_0V_1, V_0(\eta_1\eta_2), V_0 V_2 ) = \cross( \eta_0 \eta_2 , (V_0V_1)\eta_2, (V_0(\eta_1\eta_2))\eta_2 , (V_0 V_2)\eta_2 ).
$$
A direct calculation shows that
\[
\cancel{3}(\mathfrak{F}) = A = \cross( \eta_0 \eta_2 , (V_0V_1)\eta_2 , (V_0(\eta_1\eta_2))\eta_2 , (V_0 V_2)\eta_2 ).
\]
This completes the proof.
\end{proof}
\begin{figure}[t]
\centering
\includegraphics[width=5cm]{cross_and_triple.png}
\caption{The cross ratio $\cross( V_0 V_2 , V_0(\eta_1\eta_2) , V_0 V_1 , \eta_0 )$ .}
\label{cross_and_triple}
\end{figure}
\subsection{Quadruple ratio}
\label{subsubsec:Quadruple ratio}
Let $\mathfrak{F} = (\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3)$ be an ordered quadruple of flags in general position, $\mathcal{F}_i = ( V_i , \eta_i )$. We define the \emph{quadruple ratio} of $\mathfrak{F}$ as
$$
\cancel{4}(\mathfrak{F}):= \cross( \eta_0 , V_0V_3 , V_0V_2 , V_0V_1 ).
$$
$\cancel{4}(\mathfrak{F})$ is sometimes referred as \emph{edge ratio} with respect to $V_0 V_2$. We give an intuitive description of this definition in $\S~\ref{subsubsec:param_of_flags}$. It follows from Theorem~\ref{thm:cross_triple} that
\[
\cancel{4}(\mathfrak{F}) = \cancel{3}\big( \big( ( V_0 , \eta_0 ) , ( V_3 , V_2 V_3 ) , ( V_1 , V_1V_2 ) \big) \big).
\]
The requirement that $(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3)$ are flags in general position may be relaxed, but this is not needed for this paper.
\begin{figure}[t]
\centering
\includegraphics[width=12.5cm]{quadruple_ratio.png}
\caption{The quadruple ratio $\cancel{4}(\mathfrak{F})$ can be read off both in terms of $\cross( \eta_0 , V_0V_3 , V_0V_2 , V_0V_1 )$ and $\cancel{3}\left( \left( ( V_0 , \eta_0 ) , ( V_3 , V_2V_3 ) , ( V_1 , V_1V_2 ) \right) \right) $.}
\label{quadruple_ratio}
\end{figure}
\subsection{Pairs of convex polygons and configuration of flags } \label{subsec_def_toymodel}
When $m$ flags are configured properly, they form two convex polygons, one inscribed into the other. In the remainder of this section we parametrise such configurations for $m =3,4$ using the triple ratio and quadruple ratio. This will be a key ingredient in the proof of Theorem \ref{thm:global-coord}.
We say that a non-degenerate $m$--gon $\sigma_1$ is \emph{strictly inscribed} into another non-degenerate $m$--gon $\sigma_2$ if each edge of $\sigma_2$ contains one vertex of $\sigma_1$ in its interior. A pair of oriented $m$--gons is \emph{oriented} if the two $m$--gons have the same orientation. Furthermore, an oriented pair of strictly inscribed $m$--gons is \emph{marked} by a choice of a preferred vertex of the inscribed polygon (equiv. a preferred edge of the circumscribed polygon) and subsequent ordering of the vertices of the inscribed polygon (equiv. the edges of the circumscribed polygon) starting from the preferred vertex and following the orientation of the pair.
Let $\mathcal{P}_m$ be the space of oriented pairs of strictly inscribed convex $m$--gons in $\RR\PP^2$, and $\mathcal{P}^*_m$ be the space of marked oriented pairs of strictly inscribed convex $m$--gons in $\RR\PP^2$, both modulo the action of $\SL(3,\mathbb{R})$.
Similarly, $\Conf_m$ is the space of cyclically ordered $m$--tuples of flags in general position and $\Conf^*_m$ is the space of ordered $m$--tuples of flags in general position, both modulo the action of $\SL(3,\mathbb{R})$. There are natural maps
$$
\xi_m : \mathcal{P}_m \rightarrow \Conf_m \qquad \mbox{ and } \qquad \xi^*_m : \mathcal{P}^*_m \rightarrow \Conf^*_m,
$$
defined by deleting the interior of the edges of the inscribed polygon and extending the edges of the circumscribed polygon to lines. The fact that the polygons are assumed to be strictly inscribed one into the other ensures that the corresponding flags are in general position. Moreover, two pairs of polygons are projectively equivalent if and only if the associated flags are projectively equivalent. Hence $\xi_m$ and $\xi^*_m$ are injective and hence $\mathcal{P}_m$ and $\mathcal{P}^*_m$ can be identified with their images in $\Conf_m$ and $\Conf^*_m$ respectively. On the other hand, these maps are not surjective (see \S\ref{subsubsec:not injective} below). We give a characterisation of $\mathcal{P}_3$ and $\mathcal{P}^*_4$ in $\S\ref{subsubsec:param_of_flags}$.
\subsection{Inscribed triangles and triples of flags}
\label{subsubsec:not injective}\label{rem:example}
We present an example where we relate $\mathcal{P}_3$, $\Conf_3$ and convex domains in $\RR\PP^2$. Let
\[
v_A := \begin{bmatrix} 1 \\ 0 \\ 0 \end{bmatrix}, \qquad
v_B := \begin{bmatrix} 0 \\ 0 \\ 1 \end{bmatrix}, \qquad
v_C := \begin{bmatrix} 1 \\ 1 \\ 1 \end{bmatrix}.
\]
Now consider the cyclically ordered triple of flags $\mathfrak{F} = ((\mathcal{F}_A, F_B, \mathcal{F}_C))$, where
\[
\mathcal{F}_A:= ( v_A , v_B^{\perp} ), \qquad
\mathcal{F}_B:=( v_B , v_A^{\perp} ), \qquad
\mathcal{F}_C:= ( v_C , \eta_C ), \qquad \eta_C := [ -t : 1 + t : -1 ], \quad t \in \mathbb{R}.
\]
This triple is in general position if and only if $t \neq 0, -1$, as these are the cases where $ \eta_C $ passes through the points $v_A$ and $v_A^\perp v_B^\perp$ respectively. The case $t = \infty$ would also violate generality as this would ensure that $\eta_C$ passes through $ v_B $ .
There are four triangular regions in $\RR\PP^2$ with vertices $v_A, \; v_B$ and $v_C$. The triple $\mathfrak{F}$ is in the image of $\xi_3$ if and only if one of these triangles is disjoint from each of the lines $ v_B^\perp $, $ v_A^\perp $ and $ \eta_C $. The lines $ v_B^\perp $ and $ v_A^\perp $ pass through three of the four potential triangles. The only remaining option is the triangle which is strictly contained in the affine patch $ \mathcal{A} := \{ [x: y: 1 ] \ | \ x,y \in \mathbb{R} \} \subset \RR\PP^2$. Therefore, $\mathfrak{F} \in \xi_3(\mathcal{P}_3)$ if and only if $ \eta_C $ intersects the line $ v_A v_B $ at a point of the form $[x : 0 : 1]$, where $x<0$. This is the case if and only if $t > 0$.
A direct calculation shows that $\cancel{3}(\mathfrak{F}) = t$. So the triple ratio may be used to determine whether $\mathfrak{F} \in \xi_3(\mathcal{P}_3)$. Moreover, there is a unique conic $\mathcal{C}$ passing through $v_A, v_B$ and $v_C$ with supporting projective lines $v_A^{\perp} $ and $ v_B^{\perp} $, namely the set of points
\[
\begin{bmatrix}
x^2 \\ xz \\ z^2
\end{bmatrix}, \qquad (x, z) \in \mathbb{R}^2 \backslash (0,0).
\]
The line defined by $ \eta_C $ is a supporting line to $\mathcal{C}$ at $ v_C $ if and only if $t=1$. As projective transformations preserve the set of conics, this triple ratio test may be used to determine whether an element of $\mathcal{P}_3$ inscribes a conic. We will expand further upon this in \S\ref{subsec:classical-teich}.
\begin{figure}[ht]
\centering
\includegraphics[width=12.5cm]{example.png}
\caption{When $t=1$, $((\mathcal{F}_A, \mathcal{F}_B, \mathcal{F}_C))$ corresponds to an element of $\mathcal{P}_3$ and the line $\eta_{C} = [-1:2:-1]$ is a supporting projective line to $\mathcal{C}$ at the point $v_C$.}
\label{example}
\end{figure}
\subsection{Parametrisation of the spaces $\mathcal{P}_3$ and $\mathcal{P}^*_4$} \label{subsubsec:param_of_flags}
We give a parametrisation of $\mathcal{P}_3$ and $\mathcal{P}^*_4$ using triple ratios and quadruple ratios. This will used in the proof of Theorem \ref{thm:global-coord}. Define the map $ \cancel{3} \colon \mathcal{P}_3 \rightarrow \mathbb{R} \PP^1 $ as follows. For $\chi \in \mathcal{P}_3$, choose a representative $\mathfrak{F}$ of $\xi_3(\chi)$. Then $\cancel{3}(\chi) := \cancel{3}(\xi_3(\chi)) = \cancel{3}(\mathfrak{F})$ is well-defined by Lemma \ref{lem:triple_ratio_proj_invariant}.
\begin{thm} \label{thm:param_P_3}
$\cancel{3}$ establishes a bijection between $\mathcal{P}_3$ and $\mathbb{R}_{>0}$.
\end{thm}
\begin{proof}
Let $\chi \in \mathcal{P}_3$ and fix a representative $\mathfrak{F} = ((\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2))$ of $\xi_3(\chi)$, where $\mathcal{F}_i = ( V_i , \eta_i )$ for $ i= 0, 1, 2$. As in Theorem \ref{thm:cross_triple}, after a projective transformation, we can assume
$$
V_0 = \left[
\begin{matrix}
0 \\ 0 \\ 1
\end{matrix} \right], \quad
V_1 = \left[
\begin{matrix}
1 \\ 0 \\ 1
\end{matrix} \right], \quad
V_2 = \left[
\begin{matrix}
0 \\ 1 \\ 1
\end{matrix} \right], \quad
\eta_1 \eta_2 = \left[
\begin{matrix}
1 \\ 1 \\ 1
\end{matrix} \right],
$$
$$
\eta_1 = \left[
\begin{matrix}
1 : 0 : -1
\end{matrix} \right], \quad
\eta_2 = \left[
\begin{matrix}
0 : 1 : -1
\end{matrix} \right], \qquad
\eta_0 = \left[
\begin{matrix}
t : 1 : 0
\end{matrix} \right], \quad t \in \mathbb{R}.
$$
Recall that $\cancel{3}(\mathfrak{F}) \in \mathbb{R} \setminus \{0\}$ because $ \eta_0 $ is a line through $ V_0 $ which is disjoint from $ V_2 $ or $ V_1 $. In this setting, $\chi$ is uniquely determined by $\cancel{3}(\mathfrak{F})$ so injectivity is immediate. Theorem \ref{thm:cross_triple} and the discussion in \S\ref{subsubsec:positivity_of_cross} imply that $\cancel{3}(\mathfrak{F})$ is positive if and only if $ \eta_0 $ and $ V_0 (\eta_1 \eta_2 )$ lie in different regions of $\RR\PP^2 \setminus \{ V_0V_1 , V_0V_2 \}$. It follows that $\cancel{3}(\mathfrak{F})>0$ if and only if $\mathfrak{F}$ represents an element of $\mathcal{P}_3$, proving that $\cancel{3} (\mathcal{P}_3) = \mathbb{R}_{>0}$.
\end{proof}
Now we are going to show something similar for $\mathcal{P}^*_4$ using both triple ratios and quadruple ratios. We define $g \colon P^*_4 \rightarrow (\mathbb{R} \PP^1)^4 $ as follows. For $\chi \in \mathcal{P}^*_4$, choose a representative $\mathfrak{F} = (\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3)$ of $\xi^*_4(\chi)$, with $\mathcal{F}_i = ( V_i , \eta_i )$ for $ i = 0,1,2,3$. Then $g(\chi) := (t_{012},t_{023},e_{02},e_{20})$, where
\begin{align*}
t_{012} :=& \cancel{3}((\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2)),\\
t_{023} :=& \cancel{3}((\mathcal{F}_0,\mathcal{F}_2,\mathcal{F}_3)),\\
e_{02} :=& \cancel{4}(\mathcal{F}_0, \mathcal{F}_1, \mathcal{F}_2, \mathcal{F}_3) = \cancel{3}(( ( V_0 , \eta_0 ),( V_3 , V_2V_3 ) , ( V_1 , V_1V_2 ) )), \\
e_{20} :=& \cancel{4}(\mathcal{F}_2, \mathcal{F}_3, \mathcal{F}_0, \mathcal{F}_1) = \cancel{3}(( ( V_2 , \eta_2 ),( V_1 , V_0V_1 ) , ( V_3 , V_3V_0 ) )).
\end{align*}
Once again, $g$ is well-defined due to the projective invariance of triple ratio. One may visualise $\chi$ as in Figure \ref{fig:edge_coordinate}, with an additional edge crossing from $V_0$ to $V_2$. This motivates why $\cancel{4}(\mathfrak{F})$ is also called the \emph{edge ratio} of the oriented edge $V_0 V_2$.
\begin{figure}[ht]
\centering
\includegraphics[width=5cm]{edge_coordinate.png}
\caption{$e_{02}$ is the edge oriented from $V_0$ to $V_2$, adjacent to the two triangles $t_{012},t_{023}$.}
\label{fig:edge_coordinate}
\end{figure}
\begin{thm} \label{thm:param_P^*_4}
$g$ establishes a bijection between $\mathcal{P}^*_4$ and $\mathbb{R}^4_{>0}$.
\end{thm}
\begin{proof}
Let $\chi \in \mathcal{P}^*_4$ and choose a representative $\mathfrak{F} = (\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3)$ of $\xi^*_4(\chi)$, with $\mathcal{F}_i = ( V_i , \eta_i )$. As in Theorem \ref{thm:param_P_3}, we assume without loss of generality that $V_0$, $V_2$, $V_3$ and $\eta_2 \eta_0$ are fixed at an arbitrary generic quadruple of points in some affine patch $\mathcal{A}$ of $\RR\PP^2$.
\begin{figure}[ht]
\centering
\includegraphics[width=.99\linewidth]{triangulation_induction.png}
\caption{$V_1V_0$ and $V_1V_2$ are uniquely determined by $\cancel{4}(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3)$ and $\cancel{4}(\mathcal{F}_2,\mathcal{F}_3,\mathcal{F}_0,\mathcal{F}_1)$, while $\eta_1$ is uniquely determined by $V_1$ and $\cancel{3}((\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2))$.}
\label{fig:triangulation_induction}
\end{figure}
Having fixed these points, the lines $\eta_2$ and $\eta_0$ are determined. By Theorem \ref{thm:param_P_3}, the line $\eta_3$ is uniquely determined by $t_{023}$, and any assignment of positive real number to $t_{023}$ gives rise to a unique, well-defined element of $\mathcal{P}_3$. This already ensures the injectivity of $g$.
If necessary, we rearrange the affine patch $\mathcal{A}$ to contain this pair of triangles. In $\mathcal{A}$, a polygon may be referred to as the convex hull of its vertices without ambiguity.
Recall from \S\ref{subsubsec:Quadruple ratio} that the quadruple ratios $e_{02}$ and $e_{20}$ can be expressed as triple ratios
$$
e_{02} = \cancel{3}( ( V_0 , \eta_0 ) , ( V_3 , V_2V_3 ) , ( V_1 , V_1V_2 ) ) \mbox{ and } e_{20} = \cancel{3}( ( V_2 , \eta_2 ) , ( V_1 , V_1V_0 ) , ( V_3 , V_3V_0 ) ).
$$
As in Theorem \ref{thm:param_P_3}, the lines $V_0V_1$ and $V_2V_1$ are uniquely determined by $e_{02}$ and $e_{20}$, so $V_1$ is uniquely determined. Furthermore,
$V_1$ belongs to the triangle $\langle V_0,V_2,\eta_0 \eta_2 \rangle$ if and only if $e_{02} >0$ and $e_{20}>0$. Equivalently, the quadrilateral $\langle V_0,V_1,V_2,V_3 \rangle$ in $\mathcal{A}$ is convex if and only if $e_{02}, e_{20} \in \mathbb{R}_{>0}$. This construction is shown in Figure \ref{fig:triangulation_induction}.
Given that $V_1$, $\mathcal{F}_0$ and $\mathcal{F}_2$ are now fixed, the line $\eta_1$ is uniquely determined by $t_{012}$, once again appealing to Theorem \ref{thm:param_P_3}. Since $\chi$ belongs to $\mathcal{P}^*_4$, $\eta_1$ does not intersect the triangle $\langle V_0,V_1, V_2 \rangle$. That happens if and only if $t_{012} >0$. Together with the previous discussion on $t_{012},t_{023}$ and $e_{02}$, this concludes the proof that $g$ surjects onto $\mathbb{R}^4_{>0}$.
\end{proof}
The \emph{cyclic group of order four} $C_4 := \langle \alpha \ | \ \alpha^4 = 1 \rangle$, acts on $\mathcal{P}^*_4$ by removing the marking, namely $\mathcal{P}^*_4 / C_4 \cong \mathcal{P}_4$. The corresponding action of $C_4$ on $\mathbb{R}^4_{>0}$ can be thought of as the change of coordinates:
$$
\alpha \cdot \left( t_{012},t_{023},e_{02},e_{20} \right) := \left(t_{123},t_{013},e_{13},e_{31} \right),
$$
where
\begin{align*}
t_{123}= \frac{t_{012} ( e_{02} t_{023} e_{20}+ e_{02} t_{023}+ e_{02}+1)}{ e_{02} t_{012} e_{20}+t_{012} e_{20}+ e_{20}+1},\quad & \quad
t_{013}= \frac{t_{023} ( e_{02} t_{012} e_{20}+t_{012} e_{20}+ e_{20}+1)}{ e_{02} t_{023} e_{20}+ e_{02} t_{023}+ e_{02}+1},\\
e_{13} = \frac{ e_{20}+1}{( e_{02}+1) t_{012} e_{20}},\quad & \quad
e_{31} = \frac{ e_{02}+1}{ e_{02} t_{023} ( e_{20}+1)}.
\end{align*}
This change of coordinates will be analysed again in $\S\ref{subsec:change-of-coord}$.
\begin{figure}[h]
\centering
\includegraphics[width=.9\linewidth]{change_of_coord_C4.png}
\caption{The change of coordinates for the action of $C_4$. The quadrilateral formed by $V_i$, $i=0, \dots, 3$ is depicted in an affine patch such that the triangles are oriented anticlockwise.}
\label{fig:change_of_coord_C4}
\end{figure}
\section{The parameterisation of the moduli space} \label{sec:Moduli-Spaces}
Ratios of flags can be viewed as the algebraic underpinning of Fock and Goncharov's parameterisation of the moduli spaces of interest in this paper. In addition, the parameterisation is based in geometry and topology through framings of the ends and ideal triangulations.
We will now describe these ingredients and then give a self-contained construction of the parameterisation of Fock-Goncharov. Our proofs in this section are different from those found in \cite{Fock-moduli-2007}.
\subsection{Framing the ends}
\label{subsec:classification_of_ends}
In the introduction, we described the spaces $\mathcal{T}_3^+(S)$ and $\mathcal{T}_3^-(S)$ corresponding to framed structures where each end is either \emph{minimal} or \emph{maximal}, and the distinction between the spaces arises from the framings that are allowed for a hyperbolic end. In this section, we give a different definition of these spaces and show in the proofs of Theorems~\ref{thm:global-coord} and \ref{thm:global-coord-dual} that these definitions are in fact equivalent.
We recall the notation from the introduction: $S_g$ denotes a closed, orientable surface of finite genus $g$; $D_1, \ldots, D_n$ are open discs on $S_g$ with pairwise disjoint closures; and $p_k \in D_k.$ Then $S_{g,n} = S_g \setminus \{p_1, \ldots, p_n\}$ is a punctured surface and $E_k = \overline{D_k} \setminus \{p_k\}$ is an end of $S_{g,n}.$ A compact core of $S_{g,n}$ is $S_{g,n}^c = S_g \setminus \cup D_k.$ We also write $S_g = S_{g,0}.$ Furthermore, we fix a marked convex projective structure $(\Omega,\Gamma,f)$ on $S= S_g$, with $\Omega = \dev(\widetilde{S})$ and $\Gamma = \hol(\pi_1(S))$. For each end $E_k$, we identify $\pi_1(E_k)$ with its image in $\pi_1(S)$ and call it a \emph{peripheral subgroup} of $\pi_1(S).$ Similarly, $\Gamma_k := \hol(\pi_1(E_k))$ is called a \emph{peripheral subgroup} of $\Gamma$.
Let $\gamma$ be a generator of the peripheral subgroup $\pi_1(E)$ of an end $E$. Since $\hol(\gamma)$ preserves the convex domain $\Omega = \dev(\widetilde{S})$, $\hol(\gamma) \in \SL(3,\mathbb{R})$ has three positive real eigenvalues, counted with multiplicity. By a standard argument of convex projective geometry (see \cite[section $2$, pg. 193]{Cooper-convex-2015} ), $\hol(\gamma)$ is conjugate to one of the following three Jordan forms:
$$
\stackrel{
J_1 = \left(
\begin{array}{ccc}
\lambda_1 & 0 & 0 \\
0 & \lambda_2 & 0 \\
0 & 0 & \frac{1}{ \lambda_1 \lambda_2} \\
\end{array}
\right)
\qquad
J_2 = \left(
\begin{array}{ccc}
\lambda_1^2 & 0 & 0 \\
0 & \frac{1}{\lambda_1 } & 1 \\
0 & 0 & \frac{1}{\lambda_1} \\
\end{array}
\right)
\qquad
J_3 = \left(
\begin{array}{ccc}
1 & 1 & 0 \\
0 & 1 & 1 \\
0 & 0 & 1 \\
\end{array}
\right)
}{\lambda_1 > \lambda_2 > \frac{1}{ \lambda_1 \lambda_2} \hspace{3cm} \lambda_1 \neq 1 \hspace{4cm} }
$$
In the first case, $\hol(\gamma) \sim J_1$ fixes three pairwise distinct points of $\RR\PP^2$ and preserves the lines through them. We say that it is \emph{totally hyperbolic}. We refer to $E$ as a \emph{hyperbolic end}. In the second case, $\hol(\gamma) \sim J_2$ fixes only two distinct points of $\RR\PP^2$, preserves the line through them and acts as a unipotent transformation on a second line. In this case we will say that $\hol(\gamma)$ is \emph{quasi-hyperbolic} and $E$ is a \emph{special end}. In the last case, $\hol(\gamma) \sim J_3$ is \emph{parabolic}, it fixes a unique point and preserves a unique line through it. Hence $E$ is a \emph{cusp}.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\textwidth]{peripheral_action.png}
\caption{Actions of a totally hyperbolic, quasi hyperbolic and parabolic peripheral element, respectively.}
\label{peripheral_action}
\end{figure}
The action of $\hol(\pi_1(E))$ preserves $\Omega$, so $\partial\cOmega$ always contains the attracting and repelling fixed points of the peripheral element $\hol(\gamma)$ (whether or not they are distinct). In particular, when $E$ is a hyperbolic or special end, the interior of a segment between the the attracting and repelling fixed points is either contained in $\Omega$ or $\partial\cOmega$. Fix an affine patch $\mathcal{A}$ containing $\overline{\Omega}$. A hyperbolic end is \emph{maximal} if $\Omega$ contains the interior (with respect to $\mathcal{A}$) of a triangle spanned by the three fixed points of $\hol(\gamma)$, and it is \emph{minimal} if $\Omega$ is contained in the complement of this triangle. The triangle in question is shaded in the left image of Figure \ref{peripheral_action}. Note that $\Omega$ cannot be equal to the interior of such a triangle since otherwise $S$ would itself be an end, hence an annulus, contradicting our hypothesis on the topology of $S$. We define $\mathcal{T}_3^{m}(S) \subset \mathcal{T}_3(S)$ to be the subset of structures for which \emph{every hyperbolic end is either maximal or minimal}.
Each end may also be endowed with a \emph{framing}. There are two definitions which are related by projective duality.
A \emph{positive framing} of $(\Omega, \Gamma, f) \in \mathcal{T}_3^{m}(S)$ is a choice, for each end $E_k$, of a pair consisting of:
\begin{itemize}
\item if $E_k$ is a maximal hyperbolic end: the $\Gamma$--orbit of the saddle point $V$ of $\Gamma_k$ and the $\Gamma$--orbit of a supporting line $\eta$ through $V$ that is invariant under $\Gamma_k$;
\item otherwise: the $\Gamma$--orbit of a fixed point $V$ of $\Gamma_k$ in the frontier of $\Omega$, and the $\Gamma$--orbit of a supporting line $\eta$ through $V$ that is invariant under $\Gamma_k$.
\end{itemize}
Under this definition, there are two possible positive framings for each maximal end, for each special end there are three, and for each minimal end there are four. Each cusp has a unique positive framing. Two positively framed structures are equivalent if and only if they represent the same element of $\mathcal{T}_3^\times(S)$. We denote by $\mathcal{T}_3^+(S)$ the set of equivalence classes of \emph{positively framed marked properly convex projective structures with maximal or minimal hyperbolic ends} on $S$. We call $\pi^+$ the natural projection of $\mathcal{T}_3^+(S)$ onto $\mathcal{T}_3^{m}(S)$ defined by forgetting the framing.
In terms of notation, we denote the structure $(\Omega, \Gamma, f)$ with a fixed positive framing by $(\Omega, \Gamma, f)^+.$
A \emph{negative framing} differs from a positive framing only at hyperbolic ends. In the former, the choice of the pair consisting of a non-saddle point and the line through it and the saddle point was only allowed for minimal ends. In a \emph{negative framing}, this choice is only allowed for maximal ends instead, while keeping everything else the same. This means that for each maximal end there are four possible choices, while for each minimal end there are now two. We denote the structure $(\Omega, \Gamma, f)$ with a fixed negative framing by $(\Omega, \Gamma, f)^-.$
The set $\mathcal{T}_3^-(S)$ is the set of equivalence classes of \emph{negatively framed marked properly convex projective structures with maximal or minimal hyperbolic ends} on $S$, and $\pi^- : \mathcal{T}_3^-(S) \rightarrow \mathcal{T}_3^{m}(S)$ is the associated projection map.
Projective duality induces natural \emph{duality maps} $\sigma^+ \co \mathcal{T}_3^+(S) \to \mathcal{T}_3^-(S)$ and $\sigma^- \co \mathcal{T}_3^-(S) \to \mathcal{T}_3^+(S)$ which are mutually inverse. This will be discussed in detail in \S\ref{subsec:dual-proj-st}.
It will be shown in \S\ref{subsec:fin_vol} that finite-volume structures on $S$ are those in which every end is a cusp. In particular, finite-volume structures have a unique framing so there are injective maps $\mathcal{T}_3^{f}(S) \hookrightarrow{} \mathcal{T}_3^+(S)$ and $\mathcal{T}_3^{f}(S) \hookrightarrow{} \mathcal{T}_3^-(S)$.
\subsection{On cusps and boundary components}
It was already mentioned in the introduction that Fock and Goncharov offer an interpretation of cusps and geodesic boundary components depending on the flags used to decorate the ends; the different cases can be glimpsed from Figure~\ref{peripheral_action}.
In this section, we offer a different interpretation, using \emph{algebraic horospheres}, which results in a different classification for special ends.
Let us first recall some definitions (see \cite[Section $3$ pg. 16]{Cooper-convex-2015} for details). Let $\Omega \subset \RR\PP^2$ be a properly convex domain, $p \in \partial \overline{\Omega}$ and $H \subset \RR\PP^2$ a supporting hyperplane to $\Omega$ at $p$. Define $\SL(H,p)$ to be the subgroup of $\SL(3,\mathbb{R})$ which preserves both $H$ and $p$, and $\mathcal{G} = \mathcal{G}(H,p)$ to be the subgroup of $\SL(H,p)$ of those elements $A \in \SL(H,p)$ which satisfy:
\begin{itemize}
\item $A$ acts as the identity on $H$,
\item $A(\eta) = \eta$ for every line $\eta$ in $\RR\PP^2$ passing through $p$.
\end{itemize}
If $\eta$ is a line containing $p$ that is not contained in $H$, then $\mathcal{G}$ acts on $\eta$ as the group $\Par(\eta,p)$ of parabolic transformations of $\eta$ fixing $p$, thus there is an isomorphism $(\mathbb{R},+) \cong \Par(\eta,p) \cong \mathcal{G}$.
Let $\mathcal{S}_0 \subset \partial \overline{ \Omega}$ be the subset of $\partial \overline{ \Omega}$ obtained by deleting $p$ and all line segments in $\partial \overline{ \Omega}$ with one endpoint at $p$. A \emph{generalized horosphere centred on $(H,p)$} is the image $\mathcal{S}_A := A(\mathcal{S}_0)$ of $\mathcal{S}_0$ under $A \in \mathcal{G}(H,p)$. An \emph{algebraic horosphere}, or simply \emph{horosphere}, is a generalized horosphere contained in $\Omega$.
If $B \in \SL(H,p)$ preserves $\Omega$, it is well-known that $B$ acts on horospheres by
$$
B(\mathcal{S}_A) = B A (\mathcal{S}_0) = B A B^{-1}(B \mathcal{S}_0) = B A B^{-1} (\mathcal{S}_0) = \mathcal{S}_{BAB^{-1}}.
$$
Such action can be expressed in terms of the \emph{exponential horosphere displacement function} $\tau$, defined as follows. Given $B \in \SL(H,p)$, let $\lambda_+(B)$ be the eigenvalue for the eigenvector $p$. If $v \in \RR\PP^2 \setminus H$, then $Bv + H = \lambda_-v + H$ and $\lambda_- = \lambda_-(B)$ is another eigenvalue of $B$. This does not depend on the choice of $v$. The \emph{exponential horosphere displacement function} with respect to $(H,p)$ is the homomorphism $\tau : \SL(H,p) \rightarrow (\mathbb{R}^*, \times)$ given by
$$
\tau(B) = \frac{\lambda_+(B)}{\lambda_-(B)}.
$$
Choose an isomorphism from $(\mathbb{R},+)$ to $\Par(\eta,p)$ given by $ t \mapsto A_t$ and denote
$$
\mathcal{S}_t = A_t (\mathcal{S}).
$$
\begin{thm}[\cite{Cooper-convex-2015}] \label{thm_CLT_horospheres}
If $B \in \SL(H,p)$ preserves $\Omega$, then $B(\mathcal{S}_t) = \mathcal{S}_{\tau(B)t}$.
\end{thm}
Let $x \in \mathcal{T}_3^+(S)$, and let $\gamma$ be a generator of the peripheral subgroup $\pi_1(E)$. Let $\Omega := \dev (\widetilde{S})$. If $\hol (\gamma)$ is totally hyperbolic, it fixes three pairwise distinct points $\{V_{\lambda_1},V_{\lambda_2},V_{\lambda_3} \}$ of $\RR\PP^2$ and preserves the lines $\{H_{12},H_{13},H_{23}\}$ through them (Figure \ref{peripheral_action} on the left). Moreover, $\Omega$ may or may not contain the shaded triangle. Horospheres centred at $(H_{ij},V_{\lambda_i})$ look differently depending on the triple $(\Omega,V_{\lambda_i}, H_{ij})$, (the different horospheres are depicted in Figure \ref{horospheres_totally_hyperbolic}). In all cases $\tau(B) \not= 1$, therefore $\hol (\gamma)$ does not preserve the horospheres.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\textwidth]{horospheres_totally_hyperbolic.png}
\caption{Horospheres are not preserved by a totally hyperbolic peripheral element}
\label{horospheres_totally_hyperbolic}
\end{figure}
When $\hol(\gamma)$ is parabolic, it fixes a unique point $V_{\lambda_1} \in \partial \overline{\Omega}$ and preserves a unique supporting hyperplane $H$ through $V_{\lambda_1}$. Since $\tau(B) = 1$, Theorem \ref{thm_CLT_horospheres} implies that $\hol(\gamma)$ preserves horospheres at $(H,p)$, as depicted in Figure \ref{horospheres_parabolic}. Hence horospheres are $\hol(\gamma)$--orbits in $\Omega$.
\begin{figure}[h!]
\centering
\includegraphics[width=4cm]{horospheres_parabolic.png}
\caption{Horospheres are preserved at a cusp}
\label{horospheres_parabolic}
\end{figure}
When $\hol(\gamma)$ is quasi-hyperbolic, we a have more subtle situation, in between the two previous ones. In this case $\hol(\gamma)$ fixes two points $\{V_{\lambda_1},V_{\lambda_2} \}$ on $\partial \overline{\Omega}$, preserves the line $H_{12}$ through them and acts as a unipotent transformation on a second line $H_1$ containing only $V_{\lambda_1}$. On one hand, the exponential horosphere displacements with respect to $(H_{12},V_{\lambda_2})$ and $(H_1,V_{\lambda_1})$ are non-trivial, behaving like a hyperbolic end. Whereas for $(H_{12},V_{\lambda_1})$, $\tau(\hol(\gamma)) = 1$ and algebraic horospheres are preserved, analogously to the parabolic case. We demonstrate this situation in Figure \ref{horospheres_quasi_hyperbolic}.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\textwidth]{horospheres_quasi_hyperbolic.png}
\caption{Special boundary components preserve some horospheres and permute others}
\label{horospheres_quasi_hyperbolic}
\end{figure}
\subsection{Ideal triangulations of surfaces} \label{sec:Ideal_triangulations _of_surfaces}
Here we present some well-known results which we will use extensively. By construction, $S_{g,n} = S_g \setminus \mathcal{P},$ where $\mathcal{P} = \{p_1, \ldots, p_n\}.$ An \emph{essential arc} in $S_{g,n}$ is the intersection with $S_{g,n}$ of a simple arc embedded in $S_{g}$ that has endpoints in $\mathcal{P},$ interior disjoint from $\mathcal{P}$ and is not homotopic (relative to $\mathcal{P}$) to a point in $S_{g}.$ An \emph{ideal triangulation} $\mathrm{T}$ of $S_{g,n}$ is a union of pairwise disjoint and non-homotopic essential arcs. The components of $S_{g,n}\setminus \mathrm{T}$ are \emph{ideal triangles}, and we regard two ideal triangulations of $S_{g,n}$ as equivalent if they are isotopic via an isotopy of $S_{g}$ that fixes $\mathcal{P}$.
\begin{lem} \label{ideal_triangulation}
The surface $S_{g,n}$ admits an ideal triangulation. Moreover, every ideal triangulation has $-2 \chi(S_{g,n})$ ideal triangles.
\end{lem}
An \emph{edge flip} on an ideal triangulation consists of picking two distinct ideal triangles sharing an edge, removing that edge and replacing it with the other diagonal of the square thus formed.
For instance, any ideal triangulation of $S_{1,1}$ comprises three essential arcs and divides the surface into two ideal triangles. All of these ideal triangulations are combinatorially equivalent. However, performing an edge flip results in a non-isotopic ideal triangulation. The space of isotopy classes of ideal triangulations of the once-punctured torus naturally inherits the structure of the infinite trivalent tree, where vertices correspond to isotopy classes of ideal triangulations, and there is an edge between two such classes if and only if they are related by an edge flip. A well-known geometric realisation of this was described by Floyd and Hatcher \cite{Floyd-incompressible-1982}. In general, we have:
\begin{lem}[Hatcher~\cite{Hatcher-triangulations-1991}, Mosher~\cite{Mosher-tiling-1988}]\label{lem:finite_sequence_of_elementary_moves}
Any two ideal triangulations of $S$ are related by a finite sequence of edge flips.
\end{lem}
Now suppose $S$ has a strictly convex real projective structure of finite volume $(\Omega, \Gamma, f)$. An ideal triangulation of $S$ is \emph{straight} if each ideal edge is the image of the intersection of $\Omega$ with a projective line. We remark that the word \emph{straight} has been chosen instead of \emph{geodesic} so that the terminology can be transferred to the more general case of properly convex projective structures, where geodesics (with respect to the Hilbert metric) generally do not develop into straight lines in the projective plane.
\begin{lem} \label{lem:geodesic_ideal_triangulation}
Every ideal triangulation of $S$ is isotopic to a straight ideal triangulation.
\end{lem}
\begin{proof}
The ideal triangulation $\mathrm{T}$ of $S$ can be lifted to a $\Gamma$--equivariant topological ideal triangulation $\widetilde{\mathrm{T}}$ of $\Omega.$ There is a homeomorphism of $\Omega$ that fixes the frontier $\partial \overline{\Omega}$ and takes each topological ideal edge to a segment of a projective line. Since $\overline{\Omega}$ is a closed disc, this homeomorphism is isotopic to the identity. Whence the topological ideal triangulation $\widetilde{\mathrm{T}}$ of $\Omega$ is isotopic to a straight ideal triangulation. Since the set of ideal endpoints of edges is $\Gamma$--equivariant and any two such endpoints determine a unique segment of a projective line in $\Omega,$ the straight ideal triangulation is $\Gamma$--equivariant. We may therefore choose a $\Gamma$--equivariant isotopy between the ideal triangulations of $\Omega$ and push this down to an isotopy of $S.$
\end{proof}
\subsection{The canonical bijection for positively framed structures}
For a fixed ideal triangulation $\mathrm{T}$ of $S$, let $\triangle$ be the set of all ideal triangles and $\underline{E}$ be the set of all oriented edges. Given a point in $\mathcal{T}_3^+(S),$ to each ideal vertex of the lift of $\mathrm{T}$ to the universal cover of $S$ there is both an associated point in the projective plane and a line though that point, a flag. Whence to each ideal triangle there is an associated triple of flags, and hence (after choosing a cyclic order) a triple ratio. Similarly, to each oriented edge, there is an associated quadruple ratio of flags. This gives rise to a map $\mathcal{T}_3^+(S) \rightarrow \mathbb{R}^{\triangle \cup \underline{E}}_{>0}.$ The proof of the below theorem will make the association precise and turn it into a well-defined bijective map. The \emph{Fock--Goncharov moduli space} is then the set of all functions $\mathbb{R}^{\triangle \cup \underline{E}}_{>0} := \{\triangle \cup \underline{E} \rightarrow \mathbb{R}_{>0} \}.$ We note that $|\triangle \cup \underline{E}| = - 8 \chi(S).$
\begin{thm}[Fock-Goncharov 2007, \cite{Fock-moduli-2007}] \label{thm:global-coord}
Let $S = S_{g,n}$ be a surface of negative Euler characteristic with at least one puncture. For each ideal triangulation $\mathrm{T}$ of $S$ and each orientation $\nu$ on $S$ there is a canonical bijection
\[
\phi^+_{\mathrm{T}, \nu}: \mathcal{T}_3^+(S) \rightarrow \mathbb{R}^{\triangle \cup \underline{E}}_{>0}.
\]
\end{thm}
By fixing an orientation of $S$, we may suppress $\nu$ from the notation and simply write $\phi^+_{\mathrm{T}, \nu} = \phi^+_{\mathrm{T}}.$ In the introduction we suppressed both the triangulation and the orientation from the notation and used the inverse $\psi^+ = (\phi^+_{\mathrm{T}, \nu})^{-1}.$
We will call $\phi^+_{\mathrm{T}}(x)$ the \emph{Fock--Goncharov coordinate} of $x \in \mathcal{T}_3^+(S)$.
\begin{proof}
Assume $S$ is oriented using $\nu$. We first define $\phi^+_{\mathrm{T}, \nu} = \phi^+_{\mathrm{T}}$.
Let $\widetilde{\mathrm{T}}$ be the lift of $\mathrm{T}$ to the universal cover $(\widetilde{S}, \pi)$. Let $x = (\dev_x, \hol_x) \in \mathcal{T}_3^+(S)$. The idea is to choose a developing map in its isotopy class that is adapted to the framing and has straight ideal triangles, in order to define triple and quadruple ratios.
Fix an affine patch $\mathcal{A}$ in which $\Omega$ is strictly contained. We determine a developing map in its isotopy class by mapping ideal vertices of $\widetilde{\mathrm{T}}$ to the points of the framing. This choice is allowed, as the ideal vertices are fixed points of subgroups which are conjugate to the peripheral subgroups, contained in the frontier of $\Omega$. This also determines the images of the edges in $\widetilde{\mathrm{T}}$ as straight line segments between the vertices.
Recall that each point in the framing is paired up with a supporting line. Hence to each ideal triangle we have an associated triple of flags. More specifically, let $t_i \in \triangle$ be an ideal triangle. Any lift $\widetilde{t_i}$ of $t_i$ to $\widetilde{S}$ is assigned a triple of flags $\mathfrak{F}_i := ((\mathcal{F}_{i,0}, \mathcal{F}_{i,1}, \mathcal{F}_{i,2}))$, cyclically ordered according to the orientation induced on $t_i$ by $\nu$. Since $\Omega$ is properly convex, $\mathfrak{F}_i$ represents an element of $\mathcal{P}_3$. Hence define
\[
\phi^+_{\mathrm{T}}(x)(t_i) := \cancel{3}(\mathfrak{F}_i) \in \mathbb{R}_{>0}.
\]
There is a choice of $\widetilde{t_i}$ in this definition. Any two such choices differ by a deck transformation. Since $\dev_x$ is $\hol_x$--equivariant, any two resulting triples of flags $\mathfrak{F}_i$ and $\mathfrak{F}_i'$ differ by a projective transformation and therefore represent the same element of $\mathcal{P}_3$. Lemma \ref{lem:triple_ratio_proj_invariant} ensures that $\cancel{3}(\mathfrak{F}_i) = \cancel{3}(\mathfrak{F}_i')$.
Let $e_i \in \underline{E}$ be an oriented edge. Choose a lift $\widetilde{e_i}$ of $e_i$. $\widetilde{e_i}$ is shared by two triangles in $\widetilde{\mathrm{T}}$, each of which is assigned a triple of flags. Thereby we can consider $\widetilde{e_i}$ to be uniquely associated to an ordered quadruple of flags $\mathfrak{F}_i := (\mathcal{F}_{i,0}, \mathcal{F}_{i,1}, \mathcal{F}_{i,2}, \mathcal{F}_{i,3})$, ordered according to the orientation of $\Omega$, where $\mathcal{F}_{i,0}$ and $\mathcal{F}_{i,2}$ represent the vertices at the tail and head of $\dev_x(\widetilde{e_i})$ respectively. As $\Omega$ is properly convex, $\mathfrak{F}_i$ is identified with an element of $\mathcal{P}^*_4$. Thereby we define
\[
\phi^+_{\mathrm{T}}(x)(e_i) := \cancel{4}(\mathfrak{F}_i) \in \mathbb{R}_{>0}.
\]
As above, the choice of $\widetilde{e_i}$ only changes $\mathfrak{F}_i$ by a projective transformation and does not change $\cancel{4}(\mathfrak{F}_i)$. The positivity of $\phi^+_{\mathrm{T}}(x)(e_i)$ follows from Theorem \ref{thm:param_P^*_4}.
The above construction gives rise to a well-defined map
\[
\phi^+_{\mathrm{T}} : \mathcal{T}_3^+(S) \rightarrow \mathbb{R}^{\triangle \cup \underline{E} }_{>0}.
\]
It remains to show that this map is a bijection.
Let $x,y \in \mathcal{T}_3^+(S)$ and $\phi^+_{\mathrm{T}}(x) = \phi^+_{\mathrm{T}}(y)$. For any triangle $t \in \triangle$,
\[
\phi^+_{\mathrm{T}}(x)(t_0) = \phi^+_{\mathrm{T}}(y)(t_0).
\]
Fix an arbitrary lift $\widetilde{t}$ of $t$ to $\widetilde{S}$. Then $\widetilde{t}$ is assigned triples of flags $\mathfrak{F}^{\widetilde{t}}_x$ and $\mathfrak{F}^{\widetilde{t}}_y$ by the framings of $x$ and $y$ respectively. By Theorem \ref{thm:param_P_3}, there exists $A \in \SL(3, \mathbb{R})$ such that
\[
A \cdot \mathfrak{F}^{\widetilde{t}}_x = \mathfrak{F}^{\widetilde{t}}_y.
\]
Following Lemma \ref{lem:geodesic_ideal_triangulation}, we assume that edges of ideal triangles are segments of projective lines, which are thus uniquely determined by $\mathfrak{F}^{\widetilde{t}}_x$ and $ \mathfrak{F}^{\widetilde{t}}_y$. Therefore $A \cdot \dev_x(\widetilde{t}) = \dev_y(\widetilde{t})$. In particular, the restrictions of $A \cdot \dev_x$ and $\dev_y$ to $\widetilde{t}$ are the same up to an isotopy of their image that fixes the vertices..
We claim that $A \cdot \Omega_x = \Omega_y$ and furthermore that $A \cdot \dev_x$ is isotopic to $\dev_y$ over the whole domain. Let $t'$ be a triangle adjacent to $t$ and $\widetilde{t}'$ the lift of $t'$ which is adjacent to $\widetilde{t}$. Let $\widetilde{t}$ and $\widetilde{t}'$ share the edge $e$ and let $v'$ be the ideal vertex of $t'$ not contained in $e$. By proper convexity of $\Omega_x$ and $\Omega_y$, the flags at the vertices of $\widetilde{t}$ and $\widetilde{t}'$ form two pairs of strictly inscribed convex quadrilaterals $\chi_x$ and $\chi_y$. Cyclically order them according to $\nu$, and endow them with the same marking by choosing an endpoint of $e$. Thus they are elements of $\mathcal{P}^*_4$ having same triple ratios and quadruple ratios. By Theorem \ref{thm:param_P^*_4}, $\chi_x$ and $\chi_y$ are projectively equivalent. Since $A \cdot \dev_x(\widetilde{t}) = \dev_y(\widetilde{t})$, it is also the case that $A \cdot \chi_x = \chi_y$. In particular the projective lines $\eta'_x$ and $\eta'_y$ at $\dev_x(v')$ and $\dev_y(v')$ respectively, are projectively equivalent via $A$. Lemma \ref{lem:geodesic_ideal_triangulation} ensures once again that $ A \cdot \dev_x(\widetilde{t}') = \dev_y(\widetilde{t}')$, and $A \cdot \dev_x$ and $\dev_y$ are isotopic on $\widetilde{t} \cup \widetilde{t}'$.
Continuing in this manner we see that the image of $\widetilde{\mathrm{T}}$ is uniquely determined by $\phi^+_{\mathrm{T}}(x) = \phi^+_{\mathrm{T}}(y)$. Moreover, since both developing maps are equivariant with respect to the action of their corresponding holonomy groups, we have
\[
\hol_x (\gamma) = A^{-1} \cdot \hol_y (\gamma) \cdot A, \quad \forall \; \gamma \in \pi_1(S).
\]
Therefore $A \cdot \dev_x$ and $\dev_y$ differ by a holonomy-equivariant isotopy, which may be pushed down to $S$. Since the framing is determined by the flags at vertices of the ideal triangles, we can conclude that $x$ and $y$ define equivalent framed convex projective structures on $S.$ This completes the proof that $\phi^+_{\mathrm{T}}$ is injective.
It remains to show that $\phi^+_{\mathrm{T}}$ is surjective. Suppose each triangle and each oriented edge in $\mathrm{T}$ were assigned a positive real number. Lift these assignments to $\widetilde{\mathrm{T}}$. We will explicitly construct a domain $\Omega^+ \subset \RR\PP^2$ and holonomy representation $\hol : \pi_1(S) \rightarrow \SL(3,\mathbb{R})$ such that $ \Omega^+ / \hol( \pi_1(S) ) $ is a convex projective surface homeomorphic to $S$. Finally we will show that the original assignment of numbers to $\mathrm{T}$ determines a canonical choice of framing, so that the convex projective structure on $S$ thus defined is extended to an element of $\mathcal{T}_3^+(S)$.
Let $t \in \triangle$. Fix a lift $\widetilde{t}$ of $t$ to $\widetilde{S}$ with orientation inherited from $S$. By Theorem \ref{thm:param_P_3}, the positive real number assigned to $\widetilde{t}$ uniquely determines an element of $\mathcal{P}_3$. Hence fix a cyclically ordered representative $\mathfrak{F}^{\widetilde{t}}$. This triple of flags defines three points of $ \partial \overline{ \Omega^+}$ and a triple of supporting lines to $\Omega^+$ at those points. The rest of the construction of $\Omega^+$ proceeds as in the proof of Theorem \ref{thm:param_P^*_4} where, given a triple of flags, a fourth flag of an adjacent triangle is uniquely determined by two edge ratios and a triple ratio. Iterating this procedure, $\Omega^+$ is defined to be the union of the inscribed triangles determined by all these elements of $\mathcal{P}_3$, and the flags will be the framing for the projective structure.
We describe $\Omega^+$ in more detail. Since the set of triangles of $\widetilde{\mathrm{T}}$ is countable, we can index them as $\{ \widetilde{T}_i \}_{i \in \NN}$ so that $\cup_{i=0}^m \widetilde{T}_i$ is connected for all $m \in \NN$, and $\cup_{i=0}^\infty \widetilde{T}_i = \widetilde{\mathrm{T}}$. In the above construction we assigned to each triangle $\widetilde{T}_i \in \widetilde{\mathrm{T}}$ an element $\chi_i$ of $\mathcal{P}_3$. Let $\mathcal{T}_i$ be the inscribed triangle of $\chi_i$ with the vertices removed, thus define
\[
\Omega^+_m := \bigcup_{i = 0}^m \mathcal{T}_i\qquad \mbox{ and } \qquad \Omega^+ := \bigcup_{i = 0}^\infty \mathcal{T}_i.
\]
We are going to show that $\Omega^+$ is properly convex. Since $\Omega^+_0$ is clearly properly convex, we will proceed by induction on $m$. Suppose $\Omega^+_{m-1}$ is properly convex and fix an affine patch $\mathcal{A}$ strictly containing $\Omega^+_{m-1}$. Let $e$ be the common edge between $\Omega^+_{m-1}$ and $\mathcal{T}_m$ and let $V$ be the vertex of $\mathcal{T}_m$ disjoint from $e$. The edge $e$ is equipped with two edge ratios and a flag $(V_i, \eta_i)$ for each endpoint. We remark that $\eta_i$ is a supporting line for the domain $\Omega^+_{m-1}$. If necessary, we rearrange the patch $\mathcal{A}$ to contain the triangle $\Delta$ defined by $\eta_0,\eta_1$ and $e$, disjoint from $\Omega^+_m$. Since the edge ratios are strictly positive, Theorem \ref{thm:param_P^*_4} implies that $V$ lies strictly within the triangle $\Delta$. It follows that $\Omega^+_m$ is also convex and properly contained in the affine patch $\mathcal{A}$. To conclude the induction argument it is enough to observe that the positivity of the triple ratio associated to $\mathcal{T}_m$ implies that $\eta_V$, the line attached to $V$, is also a supporting line for $\Omega^+_m$. Hence $\Omega^+$ is properly convex.
The set $\Omega^+$ is defined as a union of ideal triangles, thus it has a natural combinatorial structure. By construction, we have a combinatorial isomorphism from the universal cover $\widetilde{S}$ to $\Omega^+$, and we now define an action of $\pi_1(S)$ by projective maps on $\Omega^+$ that has this combinatorial isomorphism as a developing map.
We now define the holonomy representation $\hol: \pi_1(S) \rightarrow \Gamma < \SL(3, \mathbb{R})$. Fix an arbitrary lift $\widetilde{t}$ of some $t \in \triangle$. Let $\mathfrak{F}^{\widetilde{t}}$ be the cyclically ordered triple of flags associated to $\widetilde{t}$. For each $\gamma \in \pi_1(S)$, $\widetilde{t}' := \gamma (\widetilde{t})$ is another lift of $t$, which is assigned another triple of flags $\mathfrak{F}^{\widetilde{t}'}$ with the same triple ratio as $\mathfrak{F}^{\widetilde{t}}$. Once again, Theorem \ref{thm:param_P_3} applies to provide a unique projective transformation $T_{\gamma} \in \SL(3, \mathbb{R})$ such that
\[
T_{\gamma} \cdot \mathfrak{F}^{\widetilde{t}} = \mathfrak{F}^{\widetilde{t}'}.
\]
Now $T_{\gamma}$ is an element of $\SL(\Omega^+)$, namely it preserves $\Omega^+$. Indeed, $\Omega^+$ is defined exclusively in terms of triangle parameters and edge parameters, which are invariant under projective transformations. More precisely, given two triangles $\widetilde{t},\widetilde{t}' \subset \widetilde{\mathrm{T}}$,
\[
\widetilde{t} = \gamma (\widetilde{t}') \iff \dev(\widetilde{t}) = T_{\gamma} \cdot \dev(\widetilde{t}').
\]
Hence we have
\[
\hol(\gamma) := T_{\gamma}, \quad \forall \; \gamma \in \pi_1(S).
\]
The developing map is equivariant with respect to $\hol \colon \gamma \mapsto T_\gamma$, which is an isomorphism onto its image $\Gamma := \hol(\pi_1(S))$. In particular, $\dev$ induces an isomorphism $f$ of CW--complexes from $S$ to $\Omega^+ / \Gamma$. The triple $(\Omega^+, \Gamma, f)$ thereby defines an element of $\mathcal{T}_3(S)$.
The next step is to show that $(\Omega^+, \Gamma, f) \in \mathcal{T}_3^{m}(S)$ and that it has a natural framing given by the flags at each ideal vertex. Before we analyse maximal and minimal ends, we characterise flags of $\Omega^+$ as fixed elements of conjugates of peripheral subgroups of $\Gamma$.
Recall that every vertex $V$ of $\Omega^+$ has a line $\eta$ assigned to it, forming the flag $(V,\eta)$. As $\dev$ is a combinatorial equivalence, $V$ corresponds to an ideal vertex $\widetilde{E}$ of $\widetilde{\mathrm{T}}$, which is a lift of an end $E$ of $S$. Let $\gamma_0$ be a generator of $\pi_1(E)$, and $\gamma$ the element conjugate to $\gamma_0$ fixing $\widetilde{E}$. Then $V$ is fixed by $\hol(\gamma)$. Furthermore, if $\widetilde{t}$ is a triangle in $\Omega^+$ with vertex $V$, then $\hol(\gamma)(\widetilde{t})$ is, by construction, another triangle of $\Omega^+$ sharing the same flag $(V,\eta)$. Therefore $\eta$ is also preserved by $\hol(\gamma)$. Summarising, every flag of $\Omega^+$ is a pair consisting of a fixed point and an invariant line of some (conjugate of a) peripheral subgroup of $\Gamma$. Thus when $E$ is a special end or a cusp, the flag $(V,\eta)$ is an admissible choice of framing at $E$.
Now suppose that $E$ is hyperbolic, namely $\hol(\gamma)$ fixes three distinct points of $\RR\PP^2$, one attracting $V_+$, one repelling $V_-$, and a saddle $V_0$. By the previous discussion there are six possibilities for $(V,\eta)$, given that $V \in \{V_+,V_-,V_0\}$, $\eta \in \{ V_+V_-, V_+V_0, V_-V_0\}$ and $V \in \eta$. For convenience, we fix an affine patch $\mathcal{A}$ strictly containing $\Omega^+$. Recall from \S \ref{subsec:classification_of_ends} that the interior of the segment between $V_+$ and $V_-$ in $\mathcal{A}$, say $s$, is either contained in $\Omega^+$ or in the frontier of $\Omega^+$, and $\Omega^+$ can not be equal to the interior of a triangle bounded by $s$ and $V_0$.
If $V = V_0$, then both $V_0$ and $s$ are subsets of $\overline{ \Omega^+}$. By proper convexity, $\Omega^+$ must contain the interior of the triangle spanned by $V_+,V_-,V_0$, and therefore $E$ must be maximal.
Suppose $V \not= V_0$. Let $t$ be an ideal triangle of $\Omega^+$ with vertex $V$. If $s$ was contained in $\Omega^+$, then by the construction $s$ would be covered by some triangles of $\Omega^+$. However, the orbit of $t$ under $\langle \hol(\gamma) \rangle$ accumulates at $s$, so $s$ can not intersect the interior of any triangle. At the same time, $s$ can not be the edge of a triangle, as $\hol(\gamma)$ acts on $s$ by translation. It follows that $s$ must be in the frontier of $\Omega^+$, and $E$ is minimal.
In conclusion, every hyperbolic end $E$ is either maximal or minimal, and $(V,\eta)$ is an admissible choice of positive framing also at $E$. Framing each end in this way gives $(\Omega^+, \Gamma, f)^+ \in \mathcal{T}_3^+(S)$.
\end{proof}
\subsection{A natural action of the symmetric group} \label{subsubsec:sym_gp}
The \emph{symmetric group} on three letters, $\Sym(3)$, acts by permuting the set of framings of a given end. This induces a $\Sym(3)^n$--action on $\mathcal{T}_3^+(S)$.
\begin{thm} \label{sym_group_action}
Let $S = S_{g,n}$ be a surface of negative Euler characteristic with $n>0$ punctures. There is a faithful but not free action $\Sym(3)^n \curvearrowright \mathcal{T}_3^+(S)$, whose orbits are framed convex projective structures having conjugate holonomy representations. In particular, the map $\mu^+\co\mathcal{T}_3^+(S) \to \mathfrak{X}^\times(S)$ is an isomorphism and the quotient space $\mathcal{T}_3^+(S) / \Sym(3)^n$ is naturally identified with $\mathfrak{X}(S)$.
\end{thm}
\begin{proof}
First we construct a $\Sym(3)$--action on the moduli space of a once punctured surface $S=S_{g,1}$. Denote the end of $S$ by $E$ and let $\gamma \in \pi_1(S)$ be a generator of the peripheral subgroup. Every point $x = (\Omega^x,\Gamma^x,f^x)^+ \in \mathcal{T}_3^+(S)$ comes with a holonomy representation $\hol_x$ and a $\Gamma^x$--orbit of flags constituting the framing of $E$. Let $(V, \eta)$ be the unique element in that orbit such that $V$ is fixed by $\hol_x(\gamma)$. As a projective transformation, $\hol_x(\gamma)$ fixes three points (counted with multiplicity) $V_1,V_2$ and $V_3$ in $\RR\PP^2$. We order them so that:
\begin{enumerate}
\item $V_1 = V$;
\item $V_2 \in \eta$,
\item if $E$ is special, $V_1 = V_2$ if and only if $V_3 \notin \eta$.
\end{enumerate}
Thereby, $x$ is assigned the $\Gamma^x$--orbit of an ordered triple $(V_1, V_2, V_3)$. For $\alpha \in \Sym(3)$, we define $y := \alpha \cdot x$ to be that structure whose assigned $\Gamma^x$--orbit of ordered triple is $(V_{\alpha(1)}, V_{\alpha(2)}, V_{\alpha(3)} )$. If $E$ is not hyperbolic, $y$ has equivalent Hilbert geometry to $x= (\Omega^x,\Gamma^x,f^x)$, but it is framed according to the conditions $1) - 3)$ above, applied to the triple $(V_{\alpha(1)}, V_{\alpha(2)}, V_{\alpha(3)} )$. If $E$ is hyperbolic, one may need to retract or expand $\Omega^x$ to match the correct framing. In this case, the flag $(V_{\alpha(1)}, V_{\alpha(1)}V_{\alpha(2)} )$ uniquely determines whether $E$ is maximal or minimal, hence define $\Omega^y$ accordingly. Then $\Omega^x / \Gamma^x$ differs from $\Omega^y / \Gamma^x$ by at most an annulus, so there is a natural homeomorphism $g : \Omega^x / \Gamma^x \rightarrow \Omega^y / \Gamma^x$ which is the identity outside a sufficiently small regular neighbourhood of $E$. We define $f^y := f^x \circ g$ and $y$ is the structure $(\Omega^y,\Gamma^x,f^y)$, framed as per $(V_{\alpha(1)}, V_{\alpha(2)}, V_{\alpha(3)} )$. We remark that $x$ and $y$ have the same holonomy representations.
The above construction extends to the case $n \geq 1$, by acting on a different boundary component with each distinct copy of $\Sym(3)$. The stabilizer of a structure with all ends hyperbolic is trivial, therefore the action is faithful.
Structures with only cuspidal ends are global fixed points, hence the action is not free. The existence of both types of structures follows by a standard construction in hyperbolic geometry.
It remains to show that the $\Sym(3)^n$--orbit of a point $x$ is precisely the set of those structures in $\mathcal{T}_3^+(S)$ with holonomy conjugate to $x$. One inclusion is trivial as the $\Sym(3)$--action preserves the holonomy. For the other, suppose $x = (\Omega^x,\Gamma^x,f^x)$ and $y = (\Omega^y,\Gamma^y,f^y)$ have conjugate holonomies. By acting with the appropriate element of $\SL(3,\mathbb{R})$, we can fix a representative of $y$ which has the same holonomy as $x$. It follows that $\Gamma^x = \Gamma^y$ and they share the same peripheral subgroups, hence have the same fixed points in $\RR\PP^2$. In particular, for each end $E_k$ of $S$ there is an element $\alpha_k \in \Sym(3)$ mapping the ordered triple associated to the framing of $x$ at $E_k$ to the ordered triple associated to the framing of $y$ at $E_k$. This is enough to prove that $\Omega^x$ and $\Omega^y$ may only differ at the hyperbolic ends, which can be maximal or minimal.
In particular, analogous to the construction of the developing map in the proof of Theorem~\ref{thm:global-coord}, the straightening procedure of ideal triangulation can now be applied to convex cores of the structures $x$ and $y$ respectively, to give a homeomorphism $g : \Omega^x / \Gamma^x \rightarrow \Omega^y / \Gamma^y$ which is the identity outside a small neighbourhood of $\cup_{k} E_k$, such that $f^y := f^x \circ g$. This concludes the proof that $y$ is indeed $\alpha \cdot x$, for some $\alpha = (\alpha_1,\dots,\alpha_n) \in \Sym(3)^n$.
Associating flags to the holonomies gives the claim regarding the map to $\mathfrak{X}^\times(S).$
\end{proof}
Global fixed points of the action are those structures whose ends are all cusps. In \S \ref{subsec:fin_vol} we show that this set is in fact $\mathcal{T}_3^{f}(S)$. In \S \ref{subsubsec:max-min-coordinates} we give a parametrisation of a fundamental domain for the action of $\Sym(3)^n$.
\subsection{Duality} \label{subsec:dual-proj-st}
Every convex projective structure has a dual structure, constructed using the self-duality of the projective plane. This induces an isomorphism $\sigma^+ \co \mathcal{T}_3^+(S) \rightarrow \mathcal{T}_3^-(S)$. In this section we recall the main points of this construction, and refer the reader to \cite{Goldman-notes-1988, Vinberg-theory-1963, Cooper-marked-2010} for details and proofs.
Let $\Omega$ be a properly convex domain of $\RR\PP^2$. Its \emph{dual domain} is defined to be the set
$$
\Omega^* := \{ P \in \RR\PP^2 \ | \ P^\perp \cap \overline{\Omega} = \emptyset \}.
$$
Then $\Omega^*$ is also a properly convex domain and $\Omega = (\Omega^*)^*$. Supporting lines to $\Omega$ correspond to points on the frontier of $\Omega^*$, and vice-versa. Furthermore, there is a diffeomorphism $\Phi: \Omega \rightarrow \Omega^*$, called the \emph{dual map}. The map $\Phi$ has an elementary interpretation as follows. For $P \in \Omega$, $P^\perp$ is a line disjoint from $\overline{ \Omega^*}$, thus $\mathcal{A}_P = \RR\PP^2 \setminus P^\perp$ is an affine patch strictly containing $\Omega^*$. Then $\Phi(P)$ is the centre of mass of $\Omega^*$ in $\mathcal{A}_P$.
In general, $\Phi$ is not a projective transformation, unless $\Omega$ is a conic. However, for $A \in \PGL(\Omega)$ and $P \in \Omega$, $\Phi(A(P)) = (A^t)^{-1} (\Phi(P)).$ Hence we define the \emph{dual group} of a subgroup $\Gamma \leq \PGL(3,\mathbb{R})$ to be the group
$$
\Gamma^* := \{ (A^t)^{-1} \ | \ A \in \Gamma \}.
$$
Clearly $\Gamma$ is isomorphic to $\Gamma^*$. The upshot of this summary of duality is the following result.
\begin{thm} \label{thm:involution}
Let $(\Omega,\Gamma,f) \in \mathcal{T}_3(S)$. Then $\Phi$ induces a diffeomorphism $\overline{\Phi} : \Omega / \Gamma \rightarrow \Omega^* / \Gamma^*$. In particular, $(\Omega^*,\Gamma^*,f^*)$ is an element of $\mathcal{T}_3(S)$, where $f^* := \overline{ \Phi} \circ f$.
\end{thm}
Theorem \ref{thm:involution} defines an involution $\sigma \co \mathcal{T}_3(S) \rightarrow \mathcal{T}_3(S)$, taking a structure to its dual structure. Duality inverts the eigenvalues of the holonomy, therefore $\overline{ \Phi}$ preserves cusps and maps hyperbolic (resp. quasi-hyperbolic) ends to hyperbolic (resp. quasi-hyperbolic) ends. However it reverses inclusions,
$$
\Omega_1 \subset \Omega_2 \iff \Omega_1^* \supset \Omega_2^*.
$$
Hence a hyperbolic end $E$ is maximal for a structure $x$ if and only if it is minimal for $\sigma(x)$. It follows that $\sigma$ restricts to $\mathcal{T}_3^{m}(S)$. When $x \in \mathcal{T}_3^{m}(S)$ is enriched with a positive framing, its dual structure $\sigma(x)$ has a natural negative framing. Specifically, suppose $(V,\eta)$ is a flag in the $\Gamma$--orbit of the framing of an end $E$, fixed by an element $\hol(\gamma)$. Then $(\eta^\perp,V^\perp)$ is fixed by $(\hol(\gamma)^t)^{-1}$ and its $\Gamma^*$--orbit is an admissible dual framing of the same end $E$. In particular, if $V_0,V_+,V_-$ and $W_0,W_+,W_-$ are the fixed points of $\hol(\gamma)$ and $(\hol(\gamma)^t)^{-1}$, the dual relation between them is
\begin{align*}
V_+^\perp = W_-W_0, \qquad &V_-^\perp = W_+W_0, \qquad V_0^\perp = W_+W_-,\\
(V_+V_-)^\perp = W_0, \qquad &(V_-V_0)^\perp = W_+, \qquad (V_+V_0)^\perp = W_-.
\end{align*}
Therefore $\sigma$ lifts to an isomorphism $\sigma^+\co \mathcal{T}_3^+(S) \rightarrow \mathcal{T}_3^-(S)$, that we call the \emph{duality map}. Moreover, the above argument gives a natural isomorphism $\sigma^-\co \mathcal{T}_3^-(S) \rightarrow \mathcal{T}_3^+(S)$ which is inverse to $\sigma^+.$
It follows from this discussion, that the composition
$\phi^+_{\mathrm{T}, \nu}\circ\sigma^- \co \mathcal{T}_3^-(S) \to \mathbb{R}^{\triangle \cup \underline{E}}_{>0}$ gives a natural bijection. However, the associated conjugacy classes of framed holonomies for the same point in
$\mathbb{R}^{\triangle \cup \underline{E}}_{>0}$ will be projectively dual to each other. This can be remedied by \emph{defining} the map $\phi^-_{\mathrm{T}, \nu}$ as
\[\phi^+_{\mathrm{T}, \nu} \circ (\mu^+)^{-1} \circ \sigma \circ \mu^+ \circ \sigma^{-1},\]
where $\sigma \co \mathfrak{X}^\times(S) \to \mathfrak{X}^\times(S)$ takes the character of the representation $\rho$ to the character of the inverse transpose of the representation, and each flag to the dual flag.
Instead, we \emph{construct} the map $\phi^-_{\mathrm{T}, \nu}$ in the next section in such a way that the identity
\[ \phi^-_{\mathrm{T}, \nu} = \phi^+_{\mathrm{T}, \nu} \circ(\mu^+)^{-1} \circ \sigma \circ\mu^+ \circ \sigma^{-1} \]
holds.
\subsection{The canonical bijection for negatively framed structures}
In \S\ref{subsec:classification_of_ends} we defined the negative framing of a convex projective structure and the moduli space $\mathcal{T}_3^-(S)$. The proof of the following result shows how to parametrise $\mathcal{T}_3^-(S)$ in a manner analogous to that of $\mathcal{T}_3^+(S).$
\begin{thm} \label{thm:global-coord-dual}
Let $S = S_{g,n}$ be a surface of negative Euler characteristic with at least one puncture. For each ideal triangulation $\mathrm{T}$ of $S$ and each orientation $\nu$ on $S$ there is a canonical isomorphism
\[
\phi_{\mathrm{T}, \nu}^*: \mathcal{T}_3^-(S) \rightarrow \mathbb{R}^{\triangle \cup \underline{E}}_{>0}.
\]
\end{thm}
Once again, we fix an orientation of $S$, and suppress $\nu$ from the notation to simply write $\phi_{\mathrm{T}, \nu}^* = \phi_{\mathrm{T}}^*.$ The map in the introduction is $\psi^- =
(\phi_{\mathrm{T}, \nu}^*)^{-1}.$
\begin{proof}
One can follow the proof of Theorem \ref{thm:global-coord} verbatim until the end of the proof of injectivity. We recall that $\phi^+_{\mathrm{T}}$ was defined by mapping the lift to $\widetilde{S}$ of the ideal triangulation of $S$ into $\Omega$ with vertices at the peripheral fixed points in the frontier of $\Omega,$ and then taking triple ratios and quadruple ratios of the appropriate flags associated to edges and vertices in the triangulation. In the proof of the injectivity, it turned out that $\Omega = \Omega^+,$ the domain obtained as the image of the ideal triangulation of $\widetilde{S}.$
The same construction is now applied to define $\phi^-_{\mathrm{T}}$ and hence produce a point in $\mathbb{R}^{\triangle \cup \underline{E}}_{>0}$ associated to an element of $\mathcal{T}_3^-(S).$ However, in certain cases, the negatively framed domain $\Omega$ strictly contains the image $\Omega^+$ of $\widetilde{S}.$ This happens precisely when we frame a maximal hyperbolic end with one of the two non-saddle points, and the line through the saddle point. Then $\Omega\setminus \Omega^+$ consists of the orbits of maximal cusps framed in this way. One can now produce an element of $\mathcal{T}_3^-(S)$ by thickening the structure at these ends. Alternatively, one may appeal to duality as follows, which in particular proves the claim that the new domain is the intersection of the circumscribing triangles.
Referring to the setting and the notation in the proof of Theorem~\ref{thm:global-coord}, we define $\mathcal{T}_i$ to be the circumscribed triangle in $\chi_i$ (containing the inscribed triangle), with the vertices removed. Then
\[
\Omega^- := \bigcap_{i = 0}^\infty \mathcal{T}_i.
\]
In this case, $\Omega^-$ is clearly properly convex as it is the countable intersection of properly convex sets. We have the action of $\Gamma$ on $\Omega^-,$ and are required to describe a homeomorphism $f\co S \to \Omega^-/\Gamma.$ In the proof of Theorem~\ref{thm:global-coord} this followed from the combinatorial map between the universal cover of $S$ with the induced triangulation to the triangulated domain. However, here we do not have such a triangulation of $\Omega^-.$
By duality, $(\Omega^-)^*$ is a properly convex domain, union of the triangles $\mathcal{T}_i^*$. Following the discussion in the proof of Theorem \ref{thm:global-coord}, $(\Omega^-)^*$ is combinatorially isomorphic to $\widetilde{\mathrm{T}}$, and there is the dual group $\Gamma^*$ of $\PGL((\Omega^-)^*)$ acting as a simplicial isomorphism on $(\Omega^-)^*$. In particular, $(\Omega^-)^* / \Gamma^* \cong S$ and Theorem~\ref{thm:involution} therefore gives a homeomorphism $f\co S \to (\Omega^-)^* / \Gamma^* \to \Omega / \Gamma.$ Whence $(\Omega,\Gamma,f)$ with the given framing is an element of $\mathcal{T}_3^-(S)$.
\end{proof}
The same Fock-Goncharov coordinate (for a fixed triangulation and orientation of $S$) gives structures in both $\mathcal{T}_3^+(S)$ and $\mathcal{T}_3^-(S).$ The respective domains $\Omega^+$ and $\Omega^-$ for these structures only differ (up to projective equivalence) at those hyperbolic ends whose flags comprise one of the two non-saddle points and the line through the saddle point. In fact, in these cases, $\Omega^-$ contains the triangle spanned by the three peripheral fixed points, hence the end is maximal, while $\Omega^+$ does not.\footnote{We note that this distinction does not arise in \cite{Fock-moduli-2007}. Indeed, the proof of Theorem 2.5 in \cite{Fock-moduli-2007} claims that
$\Omega^+$ and $\Omega^-$ are always equal domains.}
However, the proofs of Theorems~\ref{thm:global-coord} and \ref{thm:global-coord-dual} show that the structures have equivalent framed holonomies, as the developing maps and framed domains are constructed from equivalent sets of inscribed and circumscribed triangles. We record this in the next result.
\begin{cor}
The following diagram commutes:
\begin{center}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em] {
& \mathcal{T}_3^+(S) & \\
\mathbb{R}_{>0}^{\triangle \cup \underline{E}} & & \mathfrak{X}^\times(S) \\
& \mathcal{T}_3^-(S) & \\};
\path[-stealth]
(m-2-1) edge node [above] {$(\phi^+_{\mathrm{T}, \nu})^{-1} \hspace{1cm} $} (m-1-2)
(m-2-1) edge node [below] {$(\phi_{\mathrm{T}, \nu}^-)^{-1} \hspace{1cm} $} (m-3-2)
(m-1-2) edge node [above]{$\mu^+$} (m-2-3)
(m-3-2) edge node [below]{$\mu^-$} (m-2-3);
\end{tikzpicture}
\end{center}
\end{cor}
\section{Properties of the parameterisation}
\label{sec:Properties of the parameterisation}
\subsection{Change of coordinates} \label{subsec:change-of-coord}
Let $S$ be a surface as in Theorem \ref{thm:global-coord}. Having fixed a triangulation $\mathrm{T}$ and orientation $\nu$, $\mathcal{T}_3^+(S)$ can be canonically parametrised by positive real numbers. A different choice of $\nu$ or $\mathrm{T}$ may be interpreted as a change of coordinates. Similarly, the duality map $\sigma : \mathcal{T}_3^+(S) \rightarrow \mathcal{T}_3^-(S)$, defined by taking a structure to its dual, induces an involution on Fock-Goncharov moduli space. We explicitly construct these transition maps.
We remark that also the action of $\Sym(3)^n$ on $\mathcal{T}_3^+(S)$ descends to a change of coordinates. However, even in the simplest cases, the transition functions are quite convoluted and we do not have a local description of them.
\subsubsection{Transition maps for a different orientation} \label{subsubsec:change-of-orientation}
The transition map associated to a switch in the orientation of $S$ is simple to describe. Denote by $-\nu$ the opposite orientation of $\nu$. Then for all $q \in \triangle \cup \underline{E}$,
$$
\phi_{\mathrm{T}, -\nu}(x)(q) = \frac{1}{\phi_{\mathrm{T}, \nu}(x)(q)}.
$$
Indeed triple ratios are computed with respect to flags with the opposite cyclical order, and edge ratios are computed after permuting the second and final arguments.
\subsubsection{Duality map in Fock-Goncharov coordinates}\label{subsubsec:change-of-coord-involution}
Fix a triangulation $\mathrm{T}$ and an orientation of $S$. Two dual structures $x \in \mathcal{T}_3^+(S)$ and $y \in \mathcal{T}_3^-(S)$, are generally represented by different points of $\mathbb{R}^{\triangle \cup \underline{E}}_{>0}$. That is, the composition map
$$
\phi^-_{\mathrm{T}} \circ \sigma^+ \circ (\phi^+_{\mathrm{T}})^{-1} = \phi^+_{\mathrm{T}} \circ \sigma^- \circ (\phi^-_{\mathrm{T}})^{-1}
\colon \mathbb{R}^{\triangle \cup \underline{E}}_{>0} \rightarrow \mathbb{R}^{\triangle \cup \underline{E}}_{>0}
$$
is not trivial. Recall from \S\ref{subsec:dual-proj-st} that duality maps a flag $(V,\eta)$ to its dual $(\eta^\perp,V^\perp)$. Referring to the left hand side of Figure~\ref{fig:ARS_flip}, a straightforward calculation shows that the change of coordinates is locally:
$$
e_{20} \mapsto \frac{ e_{02} t_{012} ( t_{023}+1 ) } { t_{012}+1 },
\qquad
e_{02} \mapsto \frac{ e_{20} t_{023} ( t_{012}+1 ) } { t_{023}+1 }
\qquad
\text{and}
\qquad
t_{012} \mapsto \frac{1}{t_{012}}.
$$
This transformation manifestly has order two. Its set of fixed points is the algebraic variety $\mathcal{V}$ defined by the polynomials
\[\{e_{ij} - e_{ji}, \ t_{ijk} - 1 \ | \ e_{ij} \in \underline{E}, \ t_{ijk} \in \triangle \}.\]
The duality of framings (see also the classification in \S\ref{subsubsec:max-min-coordinates})
implies that structures in $(\phi^+_{\mathrm{T}})^{-1}(\mathcal{V}) \subset \mathcal{T}_3^+(S)$ may only have cusps or minimal hyperbolic ends, where the line in the framing of each minimal hyperbolic end must pass through the saddle point. In particular, $(\phi^+_{\mathrm{T}})^{-1}(\mathcal{V})$ strictly contains the classical Teichm\"uller space, namely the set of finite-area hyperbolic structures (see Lemma~ \ref{lem:coordinates-of-conic} in \S\ref{subsec:classical-teich}). On the other hand, one can think of the rest of the structures in $(\phi^+_{\mathrm{T}})^{-1}(\mathcal{V})$ as infinite-area hyperbolic structures with a preferred framing (and dual framing).
There is a rational re-parametrisation of the edge ratios due to Parreau~\cite{Parreau-invariant-2015}, which reduces the change of coordinates $\phi^-_{\mathrm{T}} \circ \sigma^+ \circ (\phi^+_{\mathrm{T}})^{-1}$ into an even simpler form. If $e_{ik}$ is the coordinate of the edge oriented from $V_i$ to $V_k$, and $V_j$ is the third vertex of the triangle with ordered triple of vertices $(V_i,V_j,V_k)$, then the new edge parameter is
$$
s_{ik} := \frac{e_{ik}t_{ijk}}{1 + t_{ijk}}.
$$
The duality map on the edge ratios is then reduced to
$$
s_{ik} \mapsto s_{ki}.
$$
We will not make use of this reparametrisation of the edge ratios since this will results in a more complicated monodromy map, which is discusses in \S\ref{subsect_monodromy_operator}.
\subsubsection{Transition maps for a different triangulation} \label{subsubsec:change-of-coord}
The transition map induced by a change of triangulation is slightly more complicated. Henceforth we fix an orientation $\nu$ on $S$ and simplify the notation to $\phi_{\mathrm{T}} = \phi_{\mathrm{T}, \nu}$. Recall that any two ideal triangulations $\mathrm{T}$ and $\mathrm{T}'$ of $S$ differ by a finite sequence of edge flips (cf. Lemma \ref{lem:finite_sequence_of_elementary_moves}), where a flip along an edge $e$ is the removal of $e$ and insertion of the other diagonal into the arising quadrilateral (see Figure \ref{fig:ARS_flip}). In particular, $\mathrm{T}$ and $\mathrm{T}'$ have the same number of vertices, edges and triangles. Let $\triangle \cup \underline{E}$ and $\triangle' \cup \underline{E}'$ denote triangles and oriented edges of the triangulations $\mathrm{T}$ and $\mathrm{T}'$ respectively. Let
$$
\Phi_e \colon \mathbb{R}^{\triangle \cup \underline{E}}_{>0} \rightarrow \mathbb{R}^{\triangle' \cup \underline{E}'}_{>0}
$$
be the coordinate change induced by a flip along $e$. If $(e_0,\dots,e_k)$ is the sequence of edges along which we flip in order to get from $\mathrm{T}$ to $\mathrm{T}'$, then the coordinate change between $\mathrm{T}$ and $\mathrm{T}'$ is the composition map $\Phi_{e_k} \circ \dots \circ \Phi_{e_0}$. We explicitly give $\Phi_e$ below according to Figure \ref{fig:ARS_flip}. To simplify the notation, we denote $\phi_\mathrm{T}(x)(q),\phi_{\mathrm{T}'}(x)(q)$ by $q,q'$, for all $q \in \triangle \cup \underline{E}$ and $q' \in \triangle' \cup \underline{E}'$. Hence:
\begin{align*}
e_{01}' = \frac{e_{01} e_{02}}{ e_{02}+1}, \quad & \quad
e_{10}' = \frac{e_{10} ( e_{02}+1) t_{012} e_{20}}{ e_{02} t_{012} e_{20}+t_{012} e_{20}+ e_{20}+1},\\
e_{12}' = \frac{e_{12} ( e_{02} t_{012} e_{20}+t_{012} e_{20}+ e_{20}+1)}{ e_{20}+1},\quad & \quad
e_{21}' = e_{21} ( e_{20}+1),\\
e_{23}' = \frac{e_{23} e_{20}}{ e_{20}+1},\quad & \quad
e_{32}' = \frac{e_{32} e_{02} t_{023} ( e_{20}+1)}{ e_{02} t_{023} e_{20}+ e_{02} t_{023}+ e_{02}+1},\\
e_{30}' = \frac{e_{30} ( e_{02} t_{023} e_{20}+ e_{02} t_{023}+ e_{02}+1)}{ e_{02}+1},\quad & \quad
e_{03}' = e_{03} ( e_{02}+1),\\
e_{13}' = \frac{ e_{20}+1}{( e_{02}+1) t_{012} e_{20}},\quad & \quad
e_{31}' = \frac{ e_{02}+1}{ e_{02} t_{023} ( e_{20}+1)},\\
t_{130}'= \frac{t_{023} ( e_{02} t_{012} e_{20}+t_{012} e_{20}+ e_{20}+1)}{ e_{02} t_{023} e_{20}+ e_{02} t_{023}+ e_{02}+1},\quad & \quad
t_{123}'= \frac{t_{012} ( e_{02} t_{023} e_{20}+ e_{02} t_{023}+ e_{02}+1)}{ e_{02} t_{012} e_{20}+t_{012} e_{20}+ e_{20}+1}.
\end{align*}
\begin{figure}[ht]
\centering
\includegraphics[width=14cm]{ARS_flip.png}
\caption{The change of coordinates for a flip. The quadrilateral formed by $V_i$, $i=0, \dots, 3$ is assumed to be in an affine patch such that the triangles are oriented anticlockwise.}
\label{fig:ARS_flip}
\end{figure}
\subsection{Monodromy map} \label{subsect_monodromy_operator}
In this paragraph we develop an efficient way to compute the monodromy map $\mu: \mathcal{T}_3^+(S) \rightarrow \mathfrak{X}_3(S)$. This has an immediate application in \S\ref{subsubsec:max-min-coordinates} and plays a key role in the proofs of \S\ref{subsec:fin_vol} and \S\ref{subsec:compatibility_brackets}.
Henceforth, $S$ is assumed to be as in Theorem~\ref{thm:global-coord}, endowed with a framed convex projective structure $x \in \mathcal{T}_3^+(S)$. We fix an orientation $\nu$ and an ideal triangulation $\mathrm{T}$, with triangles and oriented edges $\triangle$ and $\underline{E}$. The canonical isomorphism of Theorem \ref{thm:global-coord} is given by $\phi = \phi_{\mathrm{T}, \nu}$. As earlier, we simplify the notation by reducing $\phi(x)(q), \phi(x)(q)$ to $q, q'$, for all $q \in \triangle \cup \underline{E}$ and $q' \in \triangle' \cup \underline{E}'$. It will always be clear from context whether $q$ stands for a triangle, edge or coordinate.
\subsubsection{The monodromy graph} \label{subsubsect:monodromy-graph}
We construct a sub-triangulation $\mathrm{T}'$ of $\mathrm{T}$, by connecting the vertices of every triangle with its barycentre. The \emph{monodromy graph} $\mathcal{G}$ is the dual spine of $\mathrm{T}'$. Let $\widetilde{\mathcal{G}}$ and $\widetilde{\mathrm{T}}$ be lifts of $\mathcal{G}$ and $\mathrm{T}$ to $\Omega$, via the developing map of $x$ (Figure \ref{monodromy_graph}).
Let $ \underline{E}(\widetilde{\mathcal{G}})$ be the set of oriented edges of $\widetilde{\mathcal{G}}$. We now define the map
\[
\Mon \co \underline{E}(\widetilde{\mathcal{G}}) \rightarrow \SL(3,\mathbb{R}).
\]
Using the construction of Theorem \ref{thm:global-coord}, each triangle $t$ of $\widetilde{\mathrm{T}}$ corresponds to a cyclically ordered triple of flags $\mathcal{F}^t := (\mathcal{F}^t_i, \mathcal{F}^t_j, \mathcal{F}^t_k)$ in $\RR\PP^2$.
\begin{itemize}
\item[$(i)$] Suppose $e \in \underline{E}(\widetilde{\mathcal{G}})$ is strictly contained in $t$. Then $e$ forms part of the boundary of a triangle in $\widetilde{\mathcal{G}}$, which itself is contained in $t$. It makes sense to compare the orientation of $e$ with that of $t$.
$\Mon(e)$ is defined as follows.
\begin{itemize}
\item
If the orientation of $e$ agrees with that of $t$ then
\[
\Mon(e)(\mathcal{F}_i,\mathcal{F}_j,\mathcal{F}_k) := (\mathcal{F}_j,\mathcal{F}_k,\mathcal{F}_i),
\]
\item
If the orientation of $e$ does not agree with that of $t$ then
\[
\Mon(e)(\mathcal{F}_i,\mathcal{F}_j,\mathcal{F}_k) := (\mathcal{F}_k,\mathcal{F}_i,\mathcal{F}_j).
\]
\end{itemize}
Such an edge is referred to as a \emph{$\triangle$--edge} and is depicted as $e_0$ in Figure \ref{example_monodromy}.
\item[$(ii)$] Otherwise, $e \in \underline{E}(\widetilde{\mathcal{G}})$ intersects an edge of $\widetilde{\mathcal{G}}$, say between $V_j$ and $V_k$. Let $t'$ be the triangle associated with the triple $\mathcal{F}^{t'} := (\mathcal{F}^{t'}_k, \mathcal{F}^{t'}_j, \mathcal{F}^{t'}_l)$, and suppose $e$ is oriented from $t$ to $t'$. Then $\Mon(e)$ is the unique projective transformation defined by
\[
\Mon(e)(V_i,\mathcal{F}_j,\mathcal{F}_k) = (V_l,\mathcal{F}_k,\mathcal{F}_j).
\]
In this case $e$ is called an \emph{$\underline{E}$--edge}. Such an edge is depicted as $e_1$ in Figure \ref{example_monodromy}.
\end{itemize}
\begin{minipage}[t]{0.45\textwidth}
\centering
\includegraphics[height=5cm]{monodromy_graph.png}
\captionof{figure}{The development of the monodromy graph is the dual spine of the development of the triangulation $\mathrm{T}'$.}
\label{monodromy_graph}
\end{minipage}
\begin{minipage}[t]{0.55\textwidth}
\centering
\includegraphics[height=5.5cm]{example_monodromy.png}
\captionof{figure}{Assuming the upper triangle is oriented anticlockwise with respect to the chosen affine patch, then $\Mon(e_0)(\mathcal{F}_i,\mathcal{F}_j,\mathcal{F}_k)=(\mathcal{F}_j,\mathcal{F}_k,\mathcal{F}_i)$ and $\Mon(e_1)(V_i,V_j,V_k,\eta_j\eta_k)=(V_l,V_k,V_j,\eta_j\eta_k)$.}
\label{example_monodromy}
\end{minipage}
We claim that $\Mon$ is well-defined. In case $(i)$, such a projective transformation always exists by Theorem~ \ref{thm:param_P_3} and the fact that $\cancel{3}(\mathcal{F}_i,\mathcal{F}_j,\mathcal{F}_k) = \cancel{3}(\mathcal{F}_j,\mathcal{F}_k,\mathcal{F}_i)$. Meanwhile, the operator defined in case $(ii)$ can be restated as the unique projective transformation such that
\[
\Mon(e)(V_i,V_j,V_k,\eta_j\eta_k) = (V_l,V_k,V_j,\eta_j\eta_k),
\]
from which existence and uniqueness are clear by the fact that both quadruples form a projective basis.
\begin{lem} \label{lem_compute_monodromy}
Let $\gamma \in \pi_1(S)$. There exists a finite sequence $(e_0 , \dots , e_m)$ of oriented edges $e_i \in \underline{E}(\widetilde{\mathcal{G}})$ such that the path $e_0 \cdot e_{1} \cdots e_m \subset \widetilde{\mathcal{G}}$ is a lift of a loop $\widehat{\gamma} \subset \mathcal{G}$, freely homotopic to $\gamma$. Furthermore,
$$
\mu(x)(\gamma) = \left[ \prod_{i=0}^m \Mon(e_{m-i}) \right].
$$
\end{lem}
\begin{proof}
The graph $\mathcal{G}$ contracts in $S$ to the dual spine of $\mathrm{T}$ by collapsing all trivial loops. This implies that the natural homomorphism $\pi_1(\mathcal{G}) \rightarrow \pi_1(S)$ is surjective.
Let $\gamma \in \pi_1(S)$ and choose a loop in $\mathcal{G}$ which is freely homotopic to $\gamma$. Lift this loop to $\widetilde{\mathcal{G}}$ and let $w := e_0 \cdot e_{1} \cdots e_m$ be the sequence of edges defined by the chosen lift. Denote by $(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2)$ the flags at the vertices of the triangle $t_{012}$ of $\widetilde{\mathrm{T}}$ from which $w$ starts, and $(\mathcal{F}_3,\mathcal{F}_4,\mathcal{F}_5)$ the flags at the vertices of the triangle $t_{345}$ where $w$ ends. Note that $t_{012}$ and $t_{345}$ are lifts of the same triangle in $T$ as $w$ is a lift of a closed loop.
It is clear from the definition of $\Mon$ that $\prod_{i=0}^m \Mon(e_{m-i})$ is a projective transformation which maps the vertices and two flags of $t_{012}$ to the vertices and two flags of $t_{345}$. However $\cancel{3}(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2) = \cancel{3}(\mathcal{F}_3,\mathcal{F}_4,\mathcal{F}_5)$ so invoking Theorem~\ref{thm:param_P_3},
\[
\prod_{i=0}^m \Mon(e_{m-i}) \cdot (\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2) = (\mathcal{F}_3,\mathcal{F}_4,\mathcal{F}_5).
\]
It follows that $\prod_{i=0}^m \Mon(e_{m-i})$ agrees on $(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2)$ with a representative in $\mu(x)(\gamma)$. The claim follows from the fact that $\mu(x)(\gamma)$ is (the conjugacy class of) a projective transformation mapping $(\mathcal{F}_0, \mathcal{F}_1, \mathcal{F}_2)$ to $(\mathcal{F}_3, \mathcal{F}_4, \mathcal{F}_5)$ and that such a projective transformation, hence its conjugacy class, is unique.
\end{proof}
We remark that there was a choice of lift in the above construction. If another lift $w'$ is chosen, giving rise to a product $\prod_{i=0}^m \Mon(e'_{m-i})$, and $A$ is the projective transformation taking the starting triangle of $w'$ to the starting triangle of $w$, then
\[
\prod_{i=0}^m \Mon(e'_{m-i}) = A^{-1} \cdot \prod_{i=0}^m \Mon(e_{m-i}) \cdot A.
\]
Hence the product $\prod_{i=0}^m \Mon(e_{m-i})$ is defined up to conjugation in $\SL(3, \mathbb{R})$. However, this is the best we can hope for as the monodromy is defined only up to conjugation. Moreover, the triangulation $\mathrm{T}$ is only defined up to a projective transformation.
\begin{rem} \label{rem_sequence_edges}
The sequence in Lemma \ref{lem_compute_monodromy} can be chosen to alternate between $\triangle$--edges and $\underline{E}$--edges. More precisely, we can choose the $e_i$'s so that $(e_0 \cdot \dots \cdot e_m)$ is a sequence of pairs of the form ($\underline{E}$--edge, $\triangle$--edge).
\end{rem}
\subsubsection{How to compute the monodromy} \label{subsubsec:computing-monodromy}
Consider the following special case.\\
Let $\{ \mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3 \}$ be the following generic quadruple of flags:
\begin{align*}
\mathcal{F}_0 = (V_0,\eta_0) = \left( \left[ \begin{array}{c} 0 \\ 0 \\ 1 \end{array} \right] , \left[ 1:0:0 \right] \right), \quad & \quad \mathcal{F}_1 = (V_1,\eta_1) = \left( \left[ \begin{array}{c} 1 \\ -1 \\ 1 \end{array} \right] , \left[t_{012}:t_{012}+1:1 \right] \right),\\
\mathcal{F}_2 = (V_2,\eta_2) = \left( \left[ \begin{array}{c} 1 \\ 0 \\ 0 \end{array} \right] , \left[ 0:0:1 \right] \right), \quad & \quad \mathcal{F}_3 = (V_3,\eta_3) = \left( \left[ \begin{array}{c} e_{02} e_{20} \\ e_{20} \\ 1 \end{array} \right] , \eta_3 \right).
\end{align*}
Then $\{V_0,V_1,V_2,\eta_0 \eta_2\}$ is a projective basis of $\RR\PP^2$. Moreover, $\eta_1$ and $V_3$ are uniquely determined by the parameters:
$$
\cancel{3}((\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2)) = t_{012},
$$
$$
\cancel{4}(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3) = e_{02} \qquad \mbox{ and } \qquad \cancel{4}(\mathcal{F}_2,\mathcal{F}_3,\mathcal{F}_0,\mathcal{F}_1) = e_{20}.
$$
Suppose $\{\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3\}$ are flags assigned to the vertices of some triangulation $\mathrm{T}$, with monodromy graph $\mathcal{G}$. Let $e_0$ be a $\triangle$--edges in $t_{012}$ and $e_1$ the $\underline{E}$--edge crossing the interval from $V_0$ to $V_2$. Assume $e_0$ is oriented according to the orientation of $t_{012}$, and $e_1$ is oriented towards $t_{023}$. See Figure \ref{example_monodromy} by setting $(k,i,j,l)=(0,1,2,3)$. Then one easily computes:
$$
\Mon(e_0) = \frac{1}{\sqrt[3]{t_{012}}}
\left(
\begin{array}{ccc}
0 & 0 & 1 \\
0 & -1 & -1\\
t_{012} & t_{012}+1 & 1
\end{array}
\right), \qquad \qquad
\Mon(e_1) = \sqrt[3]{\frac{e_{20}}{e_{02}}}
\left(
\begin{array}{ccc}
0 & 0 & e_{02} \\
0 & -1 & 0\\
\frac{1}{e_{20}} & 0 & 0
\end{array}
\right).
$$
If we denote by $e_0^{-1},e_1^{-1}$ the same edges with opposite orientation, we have that $\Mon(e_0^{-1}) = \Mon(e_0)^{-1}$ and $\Mon(e_1^{-1}) = \Mon(e_1)^{-1}$.
We remark that the appearance of cubic roots can be avoided by replacing triangle invariants and edge invariants by their cubes; this was done in \cite{Haraway-tessellating-2018}.
\begin{lem} \label{lem_example_monodromy}
Let $e$ be an oriented edge of $\widetilde{\mathcal{G}}$ with tail in some triangle $t_{ijk}$ of $\widetilde{\mathrm{T}}$, having flags $\mathcal{F}_i,\mathcal{F}_j$ and $\mathcal{F}_k$ assigned to its vertices. Let $P \in \SL(3,\mathbb{R})$ be the projective transformation mapping $(V_i,V_j,V_k,\eta_k\eta_j)$ to the projective basis $(V_1,V_2,V_0,\eta_0\eta_2)$ as above, and let
$$
T(z) := \frac{1}{\sqrt[3]{z}}
\left(
\begin{array}{ccc}
0 & 0 & 1 \\
0 & -1 & -1\\
z & z+1 & 1
\end{array}
\right), \qquad \qquad
E(x,y) := \sqrt[3]{\frac{x}{y}}
\left(
\begin{array}{ccc}
0 & 0 & y \\
0 & -1 & 0\\
\frac{1}{x} & 0 & 0
\end{array}
\right).
$$
\begin{itemize}
\item[$i)$] If $e$ is a $\triangle$--edge oriented as $t_{ijk}$, and $\cancel{3}((\mathcal{F}_i,\mathcal{F}_j,\mathcal{F}_k)) = t_{ijk}$, then
$$
\Mon(e) = P^{-1} \cdot T(t_{ijk}) \cdot P.
$$
\item[$ii)$] If $e$ is an $\underline{E}$--edge crossing the segment of endpoints $(\mathcal{F}_i,\mathcal{F}_k)$ oriented away from $t_{ijk}$, and $e_{ki},e_{ik}$ are the corresponding edge ratios, then
$$
\Mon(e) = P^{-1} \cdot E(e_{ki},e_{ik}) \cdot P.
$$
\end{itemize}
\end{lem}
\begin{proof}
In both cases, right hand side and left hand side agree on the four points $\{V_i,V_j,V_k,\eta_k\eta_j\}$, so the statement follows.
\end{proof}
We may now compute $\mu(x)$ using Theorem~\ref{thm:global-coord}.\\
Let $\gamma$ be an oriented closed curve on $S$, in general position with respect to $\mathrm{T}$ (and itself). Let $(t_0,\dots,t_m)$ be the sequence of triangles of $\triangle$ crossed by $\gamma$, cyclically ordered with respect to the orientation of $\gamma$ (possibly with repetitions). For each $i \in \{1,\dots, m\}$, suppose $t_i$ has the orientation induced by $S$ and that $\gamma$ enters $t_i$ through the edge $e_i' \in \underline{E}$ and leaves $t_i$ through $e_i \in \underline{E}$, where $e_i, e'_i$ are oriented according to $t_i$. A diagram of this situation is shown in Figure~\ref{epsilon_rightleft_turn}. We define the following quantities:
\[
E_i := E\left( \phi(x)(e_i) , \phi(x)(e_i^{-1} )\right) , \quad T_i := T\left( \phi(x)(t_i) \right),
\]
\[
\epsilon_i :=
\begin{cases}
+1 & \text{ if } e_i \cap e_i' \text{ is the head of } e_i', \\
-1 & \text{ if } e_i \cap e_i' \text{ is the tail of } e_i'.
\end{cases}
\]
\begin{figure}[h]
\centering
\includegraphics[height=4cm]{epsilon_rightleft_turn.png}
\captionof{figure}{ If we fix an affine patch in which $t_i$ is oriented anticlockwise, $\epsilon_i = +1$ or $-1$ respectively when $\gamma$ ``turns right" or ``turns left".}
\label{epsilon_rightleft_turn}
\end{figure}
\begin{thm}\label{thm_monodromy}\label{thm:monodromy}
$\mu(x)(\gamma)$ is the conjugacy class $[T_0^{\epsilon_0}\cdot E_0 \cdots T_m^{\epsilon_m}\cdot E_m]$. Furthermore, if $\gamma$ represents a peripheral element of $\pi_1(S)$ and $\epsilon_0 = 1$ (resp. $-1$), then $T_0^{\epsilon_0}\cdot E_0 \cdots T_m^{\epsilon_m}\cdot E_m$ is lower triangular (resp. upper triangular).
\end{thm}
\begin{proof}
It follows from Lemma~\ref{lem_compute_monodromy}, Remark~\ref{rem_sequence_edges} and Lemma~ \ref{lem_example_monodromy} that $\mu(x)(\gamma)$ is the conjugacy class of
\[
Q := \prod_{i=0}^m P_{m-i}^{-1} \cdot E_{m-i} \cdot T_{m-i}^{\epsilon_{m-i}} \cdot P_{m-i},
\]
where $P_i \in \SL(3,\mathbb{R})$ is the projective transformation corresponding to the triangle $t_i$, as defined in Lemma~ \ref{lem_example_monodromy}.\\
Observe that $P_j \cdot P_{j-1}^{-1} = T_j^{-\epsilon_j} \cdot E_{j-1}^{-1}$ for all $j \in \{1,\dots,m\}$, as both transformations agree on the four points $\{V_0,V_2,E_{j-1}(V_1),\eta_0 \eta_2\}$. Hence
\begin{align*}
Q &= \prod_{i=0}^m P_{m-i}^{-1} \cdot E_{m-i} \cdot T_{m-i}^{\epsilon_{m-i}} \cdot P_{m-i}\\
&= \left( \prod_{i=0}^{m-1} P_{m-i}^{-1} \cdot E_{m-i} \cdot T_{m-i}^{\epsilon_{m-i}} \cdot T_{m-i}^{-\epsilon_{m-i}}\cdot E_{m-(i+1)}^{-1} \cdot P_{m-(i+1)} \right) \cdot P_{0}^{-1} \cdot E_{0} \cdot T_{0}^{\epsilon_{0}} \cdot P_{0}\\
&= P_{m}^{-1} \cdot E_{m} \cdot T_{0}^{\epsilon_{0}} \cdot P_{0}.
\end{align*}
A simple inductive argument shows that
$P_{m}^{-1} = P_{0}^{-1} \cdot \left( \prod_{i=0}^{m-1} E_{i} \cdot T_{i+1}^{\epsilon_{i+1}} \right)$. It follows that
$$
Q = P_{0}^{-1} \cdot E_{0} \cdot \left( \prod_{i=1}^{m} T_{i}^{\epsilon_{i}} \cdot E_{i} \right) \cdot T_{0}^{\epsilon_{0}} \cdot P_{0}.
$$
With regard to the second part of the statement, we remark that $\gamma$ represents a peripheral element of $\pi_1(S)$ if and only if it circuits some ideal vertex of $\mathrm{T}$. In this case $\epsilon_i = \epsilon_j$ for all $i,j \in \{0,\dots,m\}$. Since $T_i \cdot E_i$ and $T_i^{-1} \cdot E_i$ are lower- and upper-triangular respectively, the product $T_0^{\epsilon_0}\cdot E_0 \dots T_m^{\epsilon_m}\cdot E_m$ is triangular as claimed.
\end{proof}
\subsection{Maximal and minimal ends in coordinates}
\label{subsubsec:max-min-coordinates}
In this paragraph, we show a straightforward application of Theorem~\ref{thm_monodromy}. We provide a criterion to determine the geometry of an end using only the Fock-Goncharov coordinates of the structure. More specifically, not only may one distinguish among hyperbolic ends, special ends and cusps, but also between maximality and minimality as well as the different framings.
Let $E_i$ be an end of $S$. Let $V_i$ be a vertex of $\widetilde{\mathrm{T}}$ developing $E_i$, and $\gamma_i \in \pi_1(S)$ a peripheral element whose holonomy fixes $V_i$. Let $\mathcal{F}_i = (V_i,\eta_i)$ be the framing at $V_i$. Denote by $e_{ik}$ for $k=1, \dots, m$ the edges of $\mathrm{T}$ oriented away from $E_i$, and $e_{ki}$ the edges oriented toward $E_i$. Let $t_k^i$ be the triangle in $\mathrm{T}$ appearing after the edge $e_{ik}$ according to the orientation of $\gamma_i$.
In Theorem~\ref{thm_monodromy} we show that the conjugacy class of $\hol(\gamma_i)$ has a representative of the form
\[
\prod_{k=1}^m T^{-1}(t_k^i) E(e_{ik}, e_{ki}) = \prod_{k=1}^m \sqrt[3]{ \frac{t_k^i e_{ik}}{e_{ki}}}
\begin{pmatrix} \frac{1}{t_k^i e_{ik}} & \frac{t_k^i+1}{t_k^i} & e_{ki} \\ 0 & 1 & -e_{ki} \\ 0 & 0 & e_{ki} \end{pmatrix}.
\]
The eigenvalues are
$$
\lambda_1 = \prod_{k=1}^m(e_{ki} (t_k^i)^2 e_{ik}^2)^{-\frac{1}{3}}, \qquad \lambda_2 = \prod_{k=1}^m(e_{ki}^{-1} t_k^i e_{ik})^{\frac{1}{3}}, \qquad \lambda_3 = \prod_{k=1}^m(e_{ki}^{2} t_k^i e_{ik})^{\frac{1}{3}}.
$$
As an artefact of the construction in Lemma~\ref{lem_example_monodromy}, $\lambda_1$ is always the eigenvalue corresponding to the eigenvector $V_i$, and $\lambda_2$ corresponds to the other eigenvector contained in $\eta$. One easily deduces the following classification:
\begin{enumerate}
\item $E_i$ is a hyperbolic end:
\begin{enumerate}
\item $E_i$ is maximal, i.e. $V_i$ is the saddle eigenvector:
\begin{itemize}
\item $\eta_i$ passes through the attracting eigenvector:
\[\prod_{k=1}^m t_k^i e_{ik} > \; 1 \ \text{and} \ \prod_{k=1}^m t_k^i e_{ik} e_{ki} < \; 1 \]
\item $\eta_i$ passes through the repelling eigenvector:
\[\prod_{k=1}^m t_k^i e_{ik} < \; 1 \ \text{and} \ \prod_{k=1}^m t_k^i e_{ik} e_{ki} > \; 1\]
\end{itemize}
\item $E_i$ is minimal, i.e. $V_i$ is not the saddle eigenvector:
\begin{itemize}
\item $V_i$ is the attracting eigenvector and $\eta_i$ passes through the saddle eigenvector:
\[\prod_{k=1}^m t_k^i e_{ik} < \; 1 \ \text{and}\ \prod_{k=1}^m e_{ki} < \; 1\]
\item $V_i$ is the attracting eigenvector and $\eta_i$ passes through the repelling eigenvector:
\[\prod_{k=1}^m t_k^i e_{ik} e_{ki} < \; 1 \ \text{and} \ \prod_{k=1}^m e_{ki} > \; 1 \]
\item $V_i$ is the repelling eigenvector and $\eta_i$ passes through the saddle eigenvector:
\[\prod_{k=1}^m t_k^i e_{ik} > \; 1 \ \text{and}\ \prod_{k=1}^m e_{ki} > \; 1\]
\item $V_i$ is the repelling eigenvector and $\eta_i$ passes through the attracting eigenvector:
\[\prod_{k=1}^m t_k^i e_{ik} e_{ki} > \; 1 \ \text{and} \ \prod_{k=1}^m e_{ki} < \; 1 \]
\end{itemize}
\end{enumerate}
\item $E_i$ is a special end:
\begin{enumerate}
\item $V_i$ is the attracting eigenvector of multiplicity two:
\begin{itemize}
\item $\eta_i$ does not pass through the repelling eigenvector:
$$\prod_{k=1}^m t_k^i e_{ik} = \; 1 \ \text{and} \ \prod_{k=1}^m e_{ki} < \; 1 $$
\item $\eta_i$ passes through the repelling eigenvector:
$$\prod_{k=1}^m t_k^i e_{ik} e_{ki} = \; 1 \ \text{and} \ \prod_{k=1}^m e_{ki} > \; 1$$
\end{itemize}
\item $V_i$ is the repelling eigenvector of multiplicity two:
\begin{itemize}
\item $\eta_i$ does not pass through the attracting eigenvector:
$$\prod_{k=1}^m t_k^i e_{ik} = \; 1 \ \text{and} \ \prod_{k=1}^m e_{ki} > \; 1 $$
\item $\eta_i$ passes through the attracting eigenvector:
$$\prod_{k=1}^m t_k^i e_{ik} e_{ki} = \; 1 \ \text{and} \ \prod_{k=1}^m e_{ki} < \; 1$$
\end{itemize}
\item $V_i$ is the attracting eigenvector of multiplicity one:
$$\prod_{k=1}^m t_k^i e_{ik} < \; 1 \ \text{and}\ \prod_{k=1}^m e_{ki} = \; 1$$
\item $V_i$ is the repelling eigenvector of multiplicity one:
$$\prod_{k=1}^m t_k^i e_{ik} > \; 1 \ \text{and}\ \prod_{k=1}^m e_{ki} = \; 1$$
\end{enumerate}
\item $E_i$ is a cusp:
$$\prod_{k=1}^m t_k^i e_{ik} = \; 1 \ \text{and}\ \prod_{k=1}^m e_{ki} = \; 1$$
\end{enumerate}
\section{Applications of the parameterisation}
\label{sec:Applications of the parameterisation}
\subsection{Structures of finite area}\label{subsec:fin_vol}
We seek now to determine necessary and sufficient conditions for a framed convex projective structure to have finite area, given its representative in $\mathbb{R}^{\triangle \cup \underline{E}}_{>0}$. This leads to a new proof of a result from Marquis~ \cite{Marquis-surface-2012}.
Throughout this section, $S = S_{g,n}$ will be a surface of negative Euler characteristic, with at least one puncture and orientation $\nu$. Fix a framed convex projective structure on $S$ with holonomy $\hol$ and developing map $\dev$. Furthermore, fix an ideal triangulation $\mathrm{T}$ of $S$, let $\Omega := \dev(\widetilde{S}) \subset \RR\PP^2$ and denote by $\widetilde{\mathrm{T}}$ the lift of $\mathrm{T}$ to $\Omega$. The map $\phi = \phi_{\mathrm{T}, \nu}$ is the canonical isomorphism defined in Theorem~\ref{thm:global-coord}. We identify $e_{ij}, t_{ijk} \in \mathrm{T}$ with $\phi(x)(e_{ij})$ and $\phi(x)(t_{ijk})$ respectively, so as to simplify notation.
A proof of the following lemma was given by Marquis~\cite{Marquis-surface-2012}. In order to keep this part of the notes self-contained, we give an adaptation of that proof here.
\begin{lem}\label{lem:para-fin}
The area of $S$ is finite if and only if every end of $S$ is a cusp.
\end{lem}
\begin{proof}
The surface $S$ has finite area if and only if each ideal triangle in $\widetilde{\mathrm{T}}$ has finite area. Vertices of those ideal triangles correspond to ends of $S$, therefore it is enough to show that an ideal triangle has finite area if and only if its vertices correspond to cusps.
Fix an ideal triangle $t$ of $\widetilde{\mathrm{T}}$ and denote its vertices by $V_0, V_1$ and $V_2$. Let $E$ be the end corresponding to $V_0$ and let $\gamma \in \pi_1(S)$ be a generator of the subgroup conjugate to $\pi_1(E)$ such that $\hol(\gamma)$ fixes $V_0$. Firstly, we show that if $E$ is not a cusp, the area of $t$ is unbounded. There are two cases, depending on the regularity of $\partial \overline{ \Omega}$.
Suppose $E$ is either hyperbolic, or special where the eigenvalue of $V_0$ has multiplicity two. These are the cases where $\partial \overline{ \Omega}$ is $C^0$ but not $C^1$ at $V_0$. Let $\eta_1$ and $\eta_2$ be distinct supporting lines to $\Omega$ at $V_0$. One may construct a triangle $\Delta$ with vertex $V_0$, strictly contained in a common affine patch with $\Omega$, having two edges on $\eta_1$ and $\eta_2$, and strictly containing $\Omega$. Observe that $\Delta$ and $t$ share the vertex $V_0$. We are going to show that $\mu_{\Delta}(t) = \infty$, from which it follows that $\mu_{\Omega}(t) = \infty$. There is a projective transformation of $\RR\PP^2$ such that
$$
V_0 = \left[
\begin{array}{c}
0 \\ 0 \\ 1
\end{array}
\right]
\qquad
\text{ and }
\qquad
\Delta = \left\{ \left[
\begin{array}{c}
x \\ y \\ 1
\end{array}
\right] \ | \ x, y > 0 \right\}.
$$
In the affine patch $\mathcal{A} : \{ z = 1 \}$, $V$ is the origin and $\Delta$ is the positive quadrant. Let $\epsilon>0$ and $0 < a < b$ such that
$$
t_0 = \{ (x,y) \in \mathcal{A} \ | \ 0<x<\epsilon \text{ and } ax < y < bx \} \subset t.
$$
Let $l_a : \{y=ax\}$ and $l_b : \{y=bx\}$. A direct calculation shows that
$$
d_\Delta(l_a,P) = \text{constant} \qquad \text{and} \qquad d_\Delta(V_0,P) = \infty, \quad \forall \; P \in l_b \cap \{0<x< \epsilon\},
$$
therefore $\mu_{\Delta}(t_0) = \infty$ and $\mu_{\Delta}(t) = \infty$.
Now suppose $E$ is special, but $V_0$ has eigenvalue of multiplicity one. In this case $\partial \overline{\Omega}$ is $C^1$ at $V_0$, but not $C^2$, therefore there is a unique supporting line $\eta$ to $\partial \overline{\Omega}$ at $V_0$. We are going to apply the same principle as in the previous case, but with respect to a different set $\Delta$. Let $\lambda$ be the eigenvalue of $\hol(\gamma)$ with multiplicity two. Then there is a projective transformation of $\RR\PP^2$ such that
$$
\hol(\gamma) =
\begin{pmatrix}
\lambda & 0 & 0 \\
\lambda & \lambda & 0 \\
0 & 0 & 1/\lambda^2
\end{pmatrix}.
$$
In this situation, $V_0 = \left[0 : 0 : 1\right]^t$. For all points $P = \left[ x_0 : y_0 : 1 \right]^t \in \RR\PP^2$ with $x_0>0$, there is a curve $\mathcal{C}_P \subset \RR\PP^2$ starting at $V_0$ passing through $P$ and preserved by $\hol(\gamma)$. Namely
$$
\mathcal{C}_P(t) = \left\{ \left.
\left[
\begin{array}{c}
t \\ t ( \frac{y_0}{x_0}+ \frac{\log(t) - \log(x_0) }{3 \log(\lambda) } ) \\ 1
\end{array}
\right]
\ \right| \ t > 0 \right\}.
$$
If $P \notin \overline{\Omega}$, then $\mathcal{C}_P$ is disjoint $\Omega$. Hence choose such a $P$ and fix $\mathcal{C}_P$. After the affine transformation
$$
\left(x,y\right) \mapsto \left( \frac{ e^{\frac{y_0}{x_0}} \lambda^3 }{x_0} x , \frac{ \ln(\lambda^3) e^{\frac{y_0}{x_0}} \lambda^3 }{x_0}y \right),
$$
$\mathcal{C}_P$ is mapped to $ \mathcal{C}'_P(t) := \left\{ \left. \left[ t : t \log(t) : 1 \right]^t \ \right| \ t >0 \right\}$. Thus $\Omega$ is strictly contained in the domain
$$
\Delta := \left\{ \left[
\begin{array}{c}
x \\ y \\ 1
\end{array}
\right] \ | \ 0 < x, \ x \log(x) < y \right\}.
$$
Once again, $\Delta$ and $t$ share the vertex $V_0$. By working in the affine patch $\mathcal{A} : \{ z = 1 \}$, one can apply a similar argument as in the previous case to show that $\mu_{\Delta}(t) = \infty$. The salient difference is that the ``width" of $t$ in $\Delta$ is not constant, but decreases sufficiently slowly.
It remains to show that if each ideal vertex of $t$ corresponds to a cusp, then $t$ has finite area. This is the case where $\partial \Omega$ is smooth at the vertices. Making use of an elegant proof from Colbois, Vernicos and Verovic \cite{Colbois-aire-2004}, the fact that $\partial \Omega$ is $C^2$ at $V_i$, ensures the existence of an open subset $B_i$, strictly contained in $\Omega$, whose frontier is a conic sharing the supporting line at $V_i$ with $\Omega$. The conic $\partial \overline{B_i}$ intersects $\partial \overline{T}$ in three points, one of which is $V_i$, and the others will be denoted $t_{i,1}$ and $t_{i,2}$. Let $t_i$ be the triangle in the interior of $t$ with vertices $V_i, \; t_{i,1}$ and $t_{i,2}$. Then
\[
4 \pi = \mu_{B_i} (t_i) \; > \; \mu_{\Omega}(t_i).
\]
The set $C := \overline{t \; \backslash \; ( t_0 \cup t_1 \cup t_2)} \subset \Omega $ is compact so $\mu_{\Omega}(C) < \infty$. It follows that $\mu_{\Omega}(C \cup t_0 \cup t_1 \cup t_2) < \infty$ and therefore $\mu_{\Omega }(t) < \infty$.
\end{proof}
\begin{thm}[Marquis 2010, \cite{Marquis-espace-2010}]\label{thm:Marquis-main}
The space of finite-area structures $\mathcal{T}_3^{f}(S)$ is an open cell of dimension $16g-16+6n = -8 \chi(S) - 2n$.
\end{thm}
\begin{proof}
We have seen in Theorem~\ref{thm:global-coord} that $\mathcal{T}_3^+(S)$ is an open cell of dimension $- 8 \chi(S) = 16g-16+8n.$ Moreover, it is shown in Lemma~\ref{lem:para-fin} that the finite-area condition is equivalent to the requirement that every end of $S$ is cuspidal.
Fix an ideal vertex $V_i$ of $\mathrm{T}$ and let $\gamma_i \in \pi_1(S)$ be a peripheral element around $V_i$ representing a simple loop. Denote by $e_{ik}$ for $k \in \{1, \dots, m_i\}$ the edges of $\mathrm{T}$ oriented away from $V_i$ and $e_{ki}$ the edges oriented toward $V_i$. Let $t_k^i$ be the triangle in $\mathrm{T}$ appearing after the edge $e_{ik}$ according to the orientation of $\gamma_i$. From the discussion in \S\ref{subsubsec:max-min-coordinates}, $\hol(\gamma)$ is parabolic if and only if
\[
\prod_{k=1}^{m_i} e_{ki} =: \mathcal{X}_i = 1 \quad \text{ and} \quad \prod_{k=1}^{m_i} e_{ik} t_k^i =: \mathcal{Y}_i = 1.
\]
As conjugation preserves eigenvalues, the parabolicity conditions reduce to two such equations for each of the $n$ vertices in $\mathrm{T}$. It remains to show that the $2n$ equations thus obtained are independent. Suppose the converse, then there is some equation of the form
\begin{equation}
\mathcal{X}_1^{c_1} \mathcal{Y}_1^{d_1} \dots \mathcal{X}_n^{c_n} \mathcal{Y}_n^{d_n } =1 \label{eq:Vi}
\end{equation}
which is non-trivial in the sense that it holds when $c_k$ and $d_j$ are non-zero for some values of $k$ and $j$.
Each oriented edge $e_{ij}$ has endpoints in only two vertices. Therefore this edge appears in $\mathcal{X}_i$ for exactly one value of $i$ and in $\mathcal{Y}_j$ for exactly one value of $j$. Suppose the edge $e_{ij}$ is oriented from a vertex $V_i$ to a vertex $V_j$. It follows that for (\ref{eq:Vi}) to hold, we must have $c_i = - d_j $.
Now consider the triangle in $\mathrm{T}$ with vertices $V_1$, $V_2$ and $V_3$. Using the above consideration, it follows that
\[
c_1 = - d_2 = c_3 = - d_1 = c_2 = - d_3.
\]
Completing this for all triangles in $\mathrm{T}$, it follows that $c_i = \pm d_j \quad \forall \; i ,j \in \{ 1, \ldots n\}$. Now consider the degree of each triangle parameter term in (\ref{eq:Vi}). Each such term appears exactly $3$ times in (\ref{eq:Vi}). The degree of the term $t_i$ in the left hand side of (\ref{eq:Vi}) must be in the set $\{\pm c_1, \pm 3 c_1 \}$. Hence $c_i = d_j =0 \quad \forall \; i,j$ and equation (\ref{eq:Vi}) may only hold trivially.
It follows that the $2n$ equations generated by the parabolicity conditions are independent and
\[
\text{dim}_{\mathbb{R}}(\mathcal{T}_3^{f}(S)) = \text{dim}_{\mathbb{R}}(\mathcal{T}_3^+(S)) -2n = 16g-16+6n.
\]
To show that $\mathcal{T}_3^{f}(S)$ is a topological cell, it suffices to note that $\mathcal{T}_3^+(S)$ is a cell and that imposing the parabolicity conditions amounts to retracting $\mathcal{T}_3^+(S)$ onto an algebraic subvariety.
\end{proof}
\subsection{The classical Teichm\"uller space} \label{subsec:classical-teich}
Having obtained a subvariety of $\mathcal{T}_3^+(S)$ comprising structures of finite area there is another obvious subvariety that we wish to interrogate -- the set of points which are invariant under the projective duality $\phi^-_{\mathrm{T}} \circ \sigma^+ \circ (\phi^+_{\mathrm{T}})^{-1}$ defined in \S\ref{subsubsec:change-of-coord-involution}. This is of interest because, for instance, results dating back to Kay~\cite{Kay-ptolemaic-1967} indicate that these structures should be identified with hyperbolic structures on $S$. To that end we use this section to characterise the Fock-Goncharov coordinates of these structures and in so doing reproduce a classical result of Fricke and Klein~\cite{Fricke-vorlesungen-1965} regarding the toplogy of the Teichm\"uller space $\Teich(S)$ of $S$.
As usual $S = S_{g,n}$ is a surface of negative Euler characteristic, with at least one puncture and orientation $\nu$. We fix an ideal triangulation $\mathrm{T}$ of $S$, and let $\phi = \phi_{\mathrm{T}, \nu}$ be the canonical isomorphism constructed in Theorem~\ref{thm:global-coord}.
A quick calculation shows that the Fock-Goncharov coordinates which are invariant under projective duality are
\begin{align}
\begin{split}
\phi(x)(e) &= \phi(x)(e^{-1}), \quad \forall \; e \in \underline{E},\label{equ:FG duality} \\
\phi(x)(t) &= 1, \hspace{2cm} \forall \; t \in \triangle,
\end{split}
\end{align}
One can easily deduce from the classification in \S\ref{subsubsec:max-min-coordinates} that this can only happen when each end is either a cusp or a minimal hyperbolic end, framed by the line through the saddle point. If we restrict our attention to the former, the following result elucidates how we identify the convex projective structure in question with a hyperbolic structure.
\begin{lem} \label{lem:coordinates-of-conic}
Let $x = (\Omega, \Gamma, f) \in \mathcal{T}_3^+(S)$. For every vertex $V_i$ of $\mathrm{T}$, let $e_{ik} \in \underline{E}, \ k = 1,\dots,m_i$, be the oriented edges starting from $V_i$. Then $\partial \overline{ \Omega}$ is a conic of $\RR\PP^2$ if and only if
\begin{align}
\begin{split}
\phi(x)(e) &= \phi(x)(e^{-1}), \quad \forall \; e \in \underline{E}, \label{eq:finite-area-hyperbolic-equations}\\
\phi(x)(t) &= 1, \hspace{2cm} \forall \; t \in \triangle, \\
\prod_{k=1}^{m_i} \phi(e_{ik}) &= 1, \hspace{2cm} \forall \; i =1,\dots,n.
\end{split}
\end{align}
\end{lem}
\begin{proof}
We begin with a preliminary observation. Suppose $\{\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3\}$ is a generic quadruple of flags of $\RR\PP^2$, where $\mathcal{F}_i = (V_i,\eta_i)$. Let $\mathcal{C}$ be the unique conic passing through $V_0,V_1,V_2$ and tangent to $\eta_0$ and $\eta_1$. A straightforward computation shows that $\mathcal{C}$ is tangent to $\eta_2$ if and only if $\cancel{3}((\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2)) = 1$ (see Remark~\ref{rem:example}). Furthermore, $V_4\in \mathcal{C}$ if and only if $\cancel{4}(\mathcal{F}_0,\mathcal{F}_1,\mathcal{F}_2,\mathcal{F}_3) = \cancel{4}(\mathcal{F}_2,\mathcal{F}_3,\mathcal{F}_0,\mathcal{F}_1)$.
Let $\widetilde{\mathrm{T}}$ be the lift of $\mathrm{T}$ to $\Omega$. By Lemma~\ref{lem:para-fin}, the projective structure $x$ has finite volume if and only if all ends are cusps. In that case, the set of vertices of $\widetilde{\mathrm{T}}$ is dense in $\partial \overline{\Omega}$, therefore $\partial \overline{\Omega}$ is a conic if and only if there is a conic passing through all the vertices of $\widetilde{\mathrm{T}}$. We can conclude that $\partial \overline{ \Omega}$ is a conic if and only if
\begin{itemize}
\item[$a)$] for each vertex $V_i$,
\begin{equation}
\prod_{k=1}^{m_i} t_k^i e_{ik} = \; 1 \qquad \text{and} \qquad \prod_{k=1}^{m_i} e_{ki}^{-1} = \; 1, \label{eq:first-equation}
\end{equation}
where $\{t_k^i\}_{k = 1,\dots,m_i}$ is the set of triangles in $\mathrm{T}$ with a vertex at $V_i$. These conditions force the ends to be cusps, as underlined in the classification in \S\ref{subsubsec:max-min-coordinates}.
\item[$b)$] for all edges $e \in \underline{E}$ and triangles $t \in \triangle$,
\begin{equation}
\phi(x)(e) = \phi(x)(e^{-1}) \qquad \text{and} \qquad \phi(x)(t) = 1. \label{eq:second-equation}
\end{equation}
This follows from the preliminary discussion. In fact, assuming all ends are cusps, one can construct a conic through all the vertices of $\mathrm{T}$, tangent to the supporting lines defined by the flags, if and only if every triple ratio is equal to $1$ and edge ratios of opposite oriented edges are equal.
\end{itemize}
Combining equations (\ref{eq:first-equation}) and (\ref{eq:second-equation}) gives precisely the conditions of the lemma.
\end{proof}
The Klein disc embeds conformally in $\RR\PP^2$ as the interior of a conic and the group of orientation-preserving hyperbolic isometries, $\PSL(2, \mathbb{R})$, is isomorphic to $PO^+(1,2) \leq \SL(3, \mathbb{R})$. Hence every hyperbolic structure on $S$ is naturally a convex projective structure on $S$. Conversely if $\Omega$ is a conic then, up to conjugation, the corresponding holonomy group is a subgroup of $ \text{Aut}(\Omega) \cong PO^+(1,2)$. Moreover the Hilbert metric on a conic is twice the natural hyperbolic metric in the Klein model. Two hyperbolic structures are equivalent as hyperbolic structures if and only if they are equivalent as projective structures.
We have seen that finite-area structures admit a unique natural framing (resp. dual framing) so one can canonically identify $\Teich(S)$ with a subvariety of $\mathcal{T}_3^+(S)$ despite the fact that $\Teich(S)$ contains no information regarding a framing. Thus we deduce the following result.
\begin{thm}[Fricke-Klein 1926, \cite{Fricke-vorlesungen-1965}]
The Teichm\"uller space of $S=S_{g,n}$ is homeomorphic to $\mathbb{R}^{6g-6 + 2n}$.
\end{thm}
\begin{proof}
The parameter space $\Teich(S)$ is the set of those structures $(\Omega, \Gamma, f) \in \mathcal{T}_3^+(S)$ such that $\partial \overline{ \Omega}$ is a conic of $\RR\PP^2$. Recall from Theorem \ref{thm:global-coord} that $\mathcal{T}_3^+(S)$ is an open cell of dimension $- 8 \chi(S) = 16g-16+8n$. By Lemma~\ref{lem:coordinates-of-conic}, $\Teich(S)$ is a subvariety of $\mathcal{T}_3^+(S)$ defined by the $(-5 \chi(S) + n)$ equations in (\ref{eq:finite-area-hyperbolic-equations}). This set of equations is independent, thus
$$
\dim_{\mathbb{R}}(\Teich(S)) = -3 \chi(S) - n = 6g-6+3n - n = 6g-6+2n.
$$
Furthermore, there is a natural retraction of $\mathcal{T}_3^+(S)$ onto $\Teich(S)$, therefore $\Teich(S)$ is also an open cell, and the claim follows.
\end{proof}
\subsection{Closed surfaces}\label{subsec:closed_surfaces}
A result of Goldman~\cite{Goldman-convex-1990} determines that if $S$ is an orientable, closed surface then $\mathcal{T}_3(S)$ is an open cell of dimension $- 8 \chi(S)$. We are going to give a different proof of the same result, using Fock-Goncharov coordinates. This requires a notion of how to glue two ends of a convex projective surface so as to induce a convex projective structure on the resulting glued surface. The more technical details of this gluing are omitted, for which we refer the reader to the original result.
Bonahon and Dreyer~\cite{Bonahon-parameterizing-2014} also use Fock and Goncharov's coordinates to prove this result employing geodesic laminations in place of ideal triangulations. We present this version here to motivate the usefulness of having a simple expression for the monodromy map.
\begin{thm}[Goldman 1990, \cite{Goldman-convex-1990}]
If $S$ is a closed, orientable surface of negative Euler characteristic, then $\mathcal{T}_3(S)$ is an open cell of dimension $- 8 \chi(S)$.
\end{thm}
\begin{proof}
Suppose $S$ is of genus $g$, endowed with a convex projective structure $(\Omega, \Gamma, f)$. That is, $S$ is the quotient of a properly convex domain $\Omega \subset \RR\PP^2$ by the action of a discrete subgroup $\Gamma < \SL(3, \mathbb{R})$ marked by $f$. Denote the holonomy representation of $S$ by $\hol$ and developing map by $\dev$.
Choose an essential, simple, closed, non-separating geodesic on $S$ and denote it by $\gamma$. Let $F$ be the interior of the surface obtained from $S$ by cutting along $\gamma$, and let $f_0$ be the restriction of $f$ to $F$. Then $F$ is a topologically finite surface, of the same Euler characteristic as $S$, with two ends, say $E_+$ and $E_-$. Let $\gamma_\pm$ be the generator of the peripheral subgroup $\pi_1(E_\pm)$, homotopic to $\gamma$ in $S$. The restriction of $\hol$ to curves in $F$ endows $F$ with a holonomy representation which we denote by $\hol_0$. Moreover, $F$ inherits a developing map from $S$, as the universal cover $\widetilde{S}$ of $S$ is tiled by copies of the universal cover of $F$. Choose one such copy of the universal cover of $F$ and name it $\widetilde{F}$. Let $\dev_0$ be the developing map obtained by restricting $\dev$ to $\widetilde{F}$. It is clear that $\Omega_0 := \dev_0(\widetilde{F}) \subset \RR\PP^2$ is a convex domain preserved by the action of $\Gamma_0 := \hol_0(\pi_1(F))$, thus $(\Omega_0,\Gamma_0,f_0)$ is a convex projective structure on $F$.
The surface $F$ has minimal hyperbolic ends at both $E_+$ and $E_-$ as the peripheral elements at those ends are inherited from the geometry of $S$. We frame $E_+$ and $E_-$ by choosing the $\Gamma_0$--orbit of the attracting fixed point and the line through the saddle point. This choice is arbitrary, but it gives a well-defined map
\[
\Psi : \mathcal{T}_3(S) \rightarrow \mathcal{T}_3^+(F).
\]
By construction, $\Psi(\mathcal{T}_3(S))$ is a subset of the set of structures $\mathcal{V} \subset \mathcal{T}_3^+(F)$, wherein:
\begin{itemize}
\item[$(a)$] All ends are minimal hyperbolic ends, framed by the attracting fixed points and the lines through the saddle points,
\item[$(b)$] The transformation $\hol_0(\gamma_+)$ is conjugate to $\hol_0(\gamma_-)$.
\end{itemize}
From the classification in \S\ref{subsubsec:max-min-coordinates}, we know that the set of structures satisfying $(a)$ is an open cell of $\mathcal{T}_3^+(F)$, hence of full dimension. On the other hand, condition $(b)$ amounts to requiring that the eigenvalues of $\hol_0(\gamma_+)$ and $\hol_0(\gamma_-)$ are equal. This amounts to two equations easily deducible from the standard form of a peripheral element, as shown in \S\ref{subsubsec:max-min-coordinates}. Therefore $\mathcal{V}$ is an open cell of dimension $-8 \chi(F) - 2$.
We are going to show that $\Psi(\mathcal{T}_3(S)) = \mathcal{V}$. Let $(\Omega_0, \Gamma_0, f_0) \in \mathcal{V}$, with holonomy $\hol_0$ and developing map $\dev_0$. Let $\gamma_\pm$ be the generator of $\pi(E_\pm)$ such that the attracting fixed point of $\hol_0(\gamma_\pm)$ is one of the vertices of the framing of $E_\pm$. By definition, $\hol_0(\gamma_+)$ and $\hol_0(\gamma_-)$ are conjugate, so let $A \in \SL(3,\mathbb{R})$ be a conjugating matrix
$$
\hol_0(\gamma_+) = A^{-1} \cdot \hol_0(\gamma_-) \cdot A
$$
with the property that $A$ maps the attracting (resp. repelling, saddle) fixed point of $\hol_0(\gamma_+)$ to the attracting (resp. repelling, saddle) fixed point of $\hol_0(\gamma_-).$ This is done with respect to a lift of $\Omega$ to $\mathbb{S}^2.$
Define the group $\Gamma := \langle \Gamma_0, A \rangle < \SL(3,\mathbb{R})$. Let $s$ be the open segment between the attracting and repelling fixed points of $\hol_0(\gamma_-)$. Hence define $\Omega$ to be the orbit of $\Omega_0 \cup s$ under $\Gamma$.
We claim that $\Omega$ is properly convex, tesselated by copies of $\Omega_0,$ and that $\Omega / \Gamma \cong S$. By definition, $\Omega_0$ and $A \cdot \Omega_0$ share supporting projective lines at the endpoints of $s$. This ensures that
$A \cdot \Omega_0$ is contained in the interior of the closed triangle $\delta$ defined by the fixed points of $\hol_0(\gamma_-)$ which contains the interval $s$ and which does not contain $\Omega_0$. Both $\Omega_0$ and $A \cdot \Omega_0$ are convex, and so
$\Omega_0 \cup A \cdot \Omega_0$ is convex. Iterating this argument at all lifts of $E_+$ and $E_-$ shows that $\Omega$ is convex and tesselated by copies of $\Omega_0.$ This also implies that $\Omega / \Gamma \cong S.$
Let $f$ be the unique homeomorphism extending $f_0$ to $S$. Then $(\Omega, \Gamma, f)$ is a convex projective structure of $S$, which maps onto $(\Omega_0, \Gamma_0, f_0)$ via $\Psi$. It follows that $\Psi$ surjects onto $\mathcal{V}$.
To conclude the proof, it is sufficient to show that the fibers of $\Psi$ are $2$--dimensional discs. This follows from the fact that the centraliser of $ \hol_0(\gamma_\pm)$ in $\SL(3, \mathbb{R})$ is a $2$--dimensional. Since $A$ is conjugate to a diagonal matrix with positive eigenvalues, this implies there is a $2$--disc worth of choices for $A$.
\end{proof}
In fact, more is known about the moduli space for a closed surface.
Denote $\mathfrak{X}_3(S) = \Hom(\pi_1(S), \SL(3, \mathbb{R}))\mathbin{/\mkern-6mu/} \SL(3, \mathbb{R})$ the \emph{character variety} of $S.$ We have a well-defined map $\mathcal{T}_3(S) \to \mathfrak{X}_3(S)$ and denote its image $\mathfrak{X}(S).$ In the case of a closed surface, this is known to give a Zariski component:
\begin{thm}[Choi-Goldman 1997, \cite{Choi-classification-1997}]\label{thm:Goldman-Choi-main}
If $S$ is a closed, orientable surface of negative Euler characteristic, then $\mathcal{T}_3(S)$ is homeomorphic to a connected component of $\mathfrak{X}_3(S).$
\end{thm}
\subsection{A Poisson Structure on the Moduli Space} \label{sec:poisson-structure}
This section is devoted to the construction and analysis of a Poisson structure on the moduli space $\mathcal{T}_3^+(S_{g,n})$. We start with a brief summary of basic notions on symplectic structures and Poisson structures. In \S~\ref{subsec:goldman-symplectic-structure} we remind the reader about Goldman's results on character varieties, with particular emphasis on Goldman's formula (Theorem~ \ref{thm:Gol-formula}). Next, we shift our attention to the moduli space $\mathcal{T}_3^+(S_{g,n})$, and explicitly construct a Poisson bracket via Theorem~\ref{thm:global-coord} and Fock-Goncharov moduli space. Finally, we concentrate on the relation between the Poisson structures on $\mathcal{T}_3^+(S_{g,n})$ and on $\mathfrak{X}_3(S_{g,n})$. The material in \S~\ref{subsec:compatibility_brackets} is devoted to developing the necessary tools to show that the structures are homomorphic.
\subsubsection{Symplectic structures and Poisson structures} \label{subsec:symplectic_bascis}
Over Riemannian manifolds, a smooth scalar function $f$ differentiates to a vector field $\nabla f$ via the metric. Symplectic manifolds are similarly rich in structure, they relate scalar functions to Hamiltonian vector fields.\\
A symplectic form $\omega$ on a smooth manifold $M$ is a closed differential $2$--form which is not degenerate in that the kernel of the operation of contracting vector fields is trivial. The pair $(M,\omega)$ is a \emph{symplectic manifold}. Because $\omega$ is non-degenerate, each smooth scalar function $f$ corresponds to a unique \emph{Hamiltonian vector field} $X_f$, the vector field satisfying
$$
\omega(X_f(x),v_x) = -D_xf(v_x) \qquad \mbox{for all } x \in M \mbox{ and } v_x \in T_xM.
$$
A \emph{Poisson manifold} is a smooth manifold $M$ endowed with a \emph{Poisson bracket}, $i.e.$ a Lie bracket
$$
\{-,-\} : C^{\infty}(M) \times C^{\infty}(M) \rightarrow C^{\infty}(M)
$$
on the algebra $C^{\infty}(M)$ of smooth functions, satisfying the Leibniz Rule.
\begin{lem}
Let $(M,\omega)$ be a symplectic manifold. For all $f,g \in C^{\infty}(M)$, the bracket $\{f,g\}_\omega := \omega( X_f, X_g)$ is a Poisson bracket.
\end{lem}
Symplectic manifolds are Poisson manifolds, but the converse is not true. A classical example of a Poisson manifold which may not give rise to a compatible symplectic structure is the linear Poisson structure on the dual of a Lie algebra. The fact that a symplectic form must be non-degenerate ensures that a symplectic manifold must have even dimension which clearly does not have to be the case for a Lie algebra.
If the Poisson bracket of a Poisson manifold $M$ restricts to a submanifold $N \subset M$, then $N$ is said to be a \emph{Poisson sub-manifold}. If moreover the restricted Poisson bracket on $N$ can be defined by a symplectic form, then $N$ is a \emph{symplectic submanifold}. Any Poisson manifold admits a \emph{symplectic foliation}, a possibly singular foliation whose leaves are the maximal symplectic submanifolds.
\subsubsection{Goldman's Poisson structure and formula} \label{subsec:goldman-symplectic-structure}
Let $S_{g,n}$ be a compact oriented surface with fundamental group $\pi_1$. Let $G$ be a Lie group preserving a non-degenerate bilinear form on its Lie algebra and $\mathfrak{X}_G$ the $G$--character variety $ \Hom(\pi_1,G) \mathbin{/\mkern-6mu/} G$. We begin by recalling the following:
\begin{thm}[Goldman 1984, \cite{Goldman-symplectic-1984}] \label{thm:Goldman-symplectic}
If $S_{g,n}$ is closed $(n=0)$, then $\mathfrak{X}_G$ has a natural symplectic structure.
\end{thm}
Theorem \ref{thm:Goldman-symplectic} is valid for any reductive group $G$, in particular $\SL(m,\mathbb{R})$. Goldman's result was generalised to compact oriented surfaces with non-empty boundary in the following sense.
\begin{thm}[Guruprasad-Huebschmann-Jeffrey-Weinstein 1997, \cite{Guruprasad-group-1997}] \label{thm:GHJW}
If $S_{g,n}$ has $n>0$ boundary components, then $\mathfrak{X}_G$ has a natural Poisson structure.
\end{thm}
On each leaf of its symplectic foliation, the symplectic form is that defined by Goldman. For the purpose of describing this foliation, we fix a presentation of $\pi_1(S_{g,n})$ based at $p \in S_{g,n}$:
$$
\pi_1 := \pi_1(S_{g,n},p) = \langle x_1,y_1,\dots,x_g,y_g,b_1,\dots,b_n \ | \ \prod_{i=1}^g [x_i,y_i] \prod_{j=1}^n b_j =1 \rangle.
$$
For $n>0$, $\pi_1$ is a free group of rank $r = 2g + n - 1 = 1 - \chi(S_{g,n})$. The \emph{$i^{th}$ boundary map} $\mathbf{b_i} : [\rho] \mapsto [\rho_{| b_i}]$ sends the extended class of a representation to the class corresponding to the restriction of $\rho$ to the boundary loop $b_i$. The \emph{boundary map}
$$
\mathbf{b} = (\mathbf{b_1},\dots,\mathbf{b_s}) : \mathfrak{X}_G \rightarrow (G \mathbin{/\mkern-6mu/} G)^{\times n}
$$
is a submersion on a Zariski-open subset $\mathfrak{X}_G^{\mathbf{b}} \subset \mathfrak{X}_G$ (cf. Hartshorne \cite{Hartshorne-algebraic-1977}, page 271), and $\mathfrak{X}_G^{\mathbf{b}}$ is foliated by non-singular submanifolds $\mathbf{F}_y : = \mathfrak{X}^{\mathbf{b}} \cap \mathbf{b}^{-1}(y)$. The leaves $\mathbf{F}_y$ are symplectic submanifolds (cf. Lawton \cite{Lawton-poisson-2009}), which extend to make the complement of the singular locus in $\mathfrak{X}_G$ a Poisson manifold. Furthermore, the Poisson structure extends continuously over singularities in $\mathfrak{X}_G$ (cf. Goldman \cite{Goldman-symplectic-1984}), making it into a \emph{Poisson variety}, an affine variety whose coordinate ring is a Poisson algebra.
As further motivation for investigating structures on the $G$--character variety, we recall an application of Theorem~\ref{thm:GHJW} to the special case $G = \SL(m,\mathbb{R})$, where the Poisson structure of $\mathfrak{X}_{\SL(m,\mathbb{R})}$ distinguishes the topology of the surface, whereas the algebraic structure alone does not.
\begin{thm}[Lawton 2009, \cite{Lawton-poisson-2009}] \label{thm:poisson_distinguish_varieties}
Let $S$ and $S'$ be compact connected orientable surfaces with boundary, and let $\mathfrak{X}_n(S)$ and $\mathfrak{X}_n(S')$ be their respective $\SL(m,\mathbb{R})$--character varieties. Then:
\begin{itemize}
\item[$(1)$] $\mathfrak{X}(S)$ and $\mathfrak{X}(S')$ are isomorphic as varieties if and only if $\chi(S) = \chi(S')$;
\item[$(2)$] $\mathfrak{X}(S)$ and $\mathfrak{X}(S')$ are isomorphic as Poisson varieties if and only if $ S \cong S'$.
\end{itemize}
\end{thm}
When $G$ is a classic matrix group (for example $\GL$, $\SL$ or $\SU$), the Poisson structure of $\mathfrak{X}_G$ has a deep geometric characterisation (see Goldman \cite{Goldman-symplectic-1984} and Lawton \cite{Lawton-poisson-2009}). Henceforth we restrict our attention to the case $G = \SL(m,\mathbb{R})$ which is of main interest here ($m = 3$), and abbreviate the notation of the $\SL(m,\mathbb{R})$--character variety to $\mathfrak{X}_m$.
Let $\tr : \SL(m,\mathbb{R}) \rightarrow \mathbb{R}$ be the trace function. For all $\alpha \in \pi_1$, define
\begin{align*}
\tr_\alpha \colon \mathfrak{X}_m \rightarrow \mathbb{R} \\
[\rho] \mapsto \tr(\rho(\alpha))
\end{align*}
The subalgebra of $C^\infty(\mathfrak{X}_m)$ generated by functions of this form is the \emph{trace algebra} $\Tr(\mathfrak{X}_m)$.
Let $\alpha$ and $\beta$ be immersed curves in $S$ representing elements in $\pi_1$ (in a general position). Denote the set of (transverse) double point intersections of $\alpha$ and $\beta$ by $\alpha \cap \beta$, and the oriented intersection number of $\alpha$ and $\beta$ at $p \in \alpha \cap \beta$ by $\epsilon(p,\alpha,\beta) \in \{-1,1\}$. Additionally, let $\alpha_p \in (\pi_1,p)$ represent the curve $\alpha$ based at $p$.
\begin{thm}[Goldman 1986 ~\cite{Goldman-invariant-1986}, Lawton 2009 ~\cite{Lawton-poisson-2009}] \label{thm:Gol-formula}
Let $\{-,-\}_{\Gol}$ be the Poisson structure on $\mathfrak{X}_m$. Then
\begin{equation} \label{eq:Gol-formula}
\{\tr_\alpha, \tr_\beta\}_{Gol} = \sum_{p \in \alpha \cap \beta} \epsilon(p,\alpha,\beta) (\tr_{\alpha_p \beta_p} - \frac{1}{n} \tr_\alpha \tr_\beta).
\end{equation}
defines a Poisson bracket on $\Tr(\mathfrak{X}_m)$.
\end{thm}
It is worth noticing that one may rewrite (\ref{eq:Gol-formula}) as
\begin{equation}
\{\tr_\alpha, \tr_\beta\}_{Gol} = \left(\sum_{p \in \alpha \cap \beta} \epsilon(p,\alpha,\beta) \tr_{\alpha_p \beta_p} \right) - \frac{\iota(\alpha,\beta)}{n} \tr_\alpha \tr_\beta,
\end{equation}
for the algebraic intersection number $\iota(\alpha,\beta)$. However we can not simplify it any further as $\tr_{\alpha_p \beta_p}$ heavily depends on $p$, as shown in Figure \ref{fig:curves_example}.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{curves_example.png}
\captionof{figure}{$\alpha_{p_1} \beta_{p_1}$ and $\alpha_{p_2} \beta_{p_2}$ are very different curves, in particular they are not homotopic.}
\label{fig:curves_example}
\end{figure}
It is known for $\SL(m,\mathbb{C})$ that the coordinate ring $\mathbb{C}[\mathfrak{X}_{\SL(m,\mathbb{C})}]$ of the $\SL(m,\mathbb{C})$--character variety is generated by trace functions, hence Theorem \ref{thm:Gol-formula} is true for $\Tr(\mathfrak{X}_m) = \mathbb{C}[\mathfrak{X}_{\SL(m,\mathbb{C})}]$ (cf. Procesi~\cite{Procesi-invariant-1976}). However, it is still an open problem for $\SL(m,\mathbb{R})$.
\subsubsection{Fock and Goncharov's Poisson structure} \label{subsec:FG-poisson-structure}
We recall the setting and notation of \S~\ref{sec:Moduli-Spaces}. As usual, $\mathrm{T}$ denotes a fixed ideal triangulation of a surface $S = S_{g,n}$ ($2g + n > 2$) and $\triangle,\underline{E}$ the sets of its triangles and oriented edges respectively. To frame our subsequent definition, we will make the following assumption:
\begin{itemize}
\item[(I)] \emph{every edge in $\mathrm{T}$ is contained in two distinct triangles, or equivalently every triangle has three distinct edges}
\end{itemize}
Such a triangulation always exists, and can be constructed algorithmically from any given triangulation. When $t$ is a triangle with two edges identified, we perform an edge flip on the third edge of $t$. We underline that this assumption is not necessary, but its inclusion avoids a lot of technicalities.
The Fock--Goncharov moduli space is the set of functions $\mathbb{R}^{\triangle \cup \underline{E}}_{>0} := \{ \triangle \cup \underline{E} \rightarrow \mathbb{R}_{>0}\}$, and it is isomorphic to $\mathcal{T}_3^+(S)$ by Theorem \ref{thm:global-coord}.
For $q \in \triangle \cup \underline{E}$, the \emph{coordinate function of $q$} is the map
\begin{align*}
q^* : \mathbb{R}^{ \triangle \cup \underline{E}}_{>0} &\rightarrow \mathbb{R}_{>0}\\
v &\mapsto v(q)
\end{align*}
and the \emph{$q$--coordinate} is the element $\overline{q} \in \mathbb{R}^{ \triangle \cup \underline{E}}_{>0}$, such that
$$
\overline{q}(q') =
\begin{cases}
1 &\mbox{ if } q = q',\\
0 & \mbox{ otherwise }.
\end{cases}
$$
It is clear from the definitions that $\frac{\partial q_i^*}{\partial \overline{q_j} } = \delta_{ij}$. For all $f,g \in C^{\infty}(\mathbb{R}^{\triangle \cup \underline{E}}_{>0})$, we define
\begin{equation} \label{eq:triangulation-bracket}
\{ f,g \}_{\mathrm{T}} := \sum_{q_1,q_2 \in \triangle \cup \underline{E} } 2 \epsilon(q_1,q_2) q_1^* q_2^* \frac{\partial f}{\partial \overline{q}_1} \frac{\partial g}{\partial \overline{q}_2},
\end{equation}
where $\epsilon : \triangle \cup \underline{E} \times \triangle \cup \underline{E} \rightarrow \{-1,0,1\}$ is the skew-symmetric integral valued function defined as follows. Given $e \in \underline{E}$, we will say that
\begin{itemize}
\item $e$ and $t \in \triangle$ are \emph{in special position} if $e \subset t$ and $e,t$ are oriented coherently (where $t$ has orientation induced from $S$).
\item $e$ and $e' \in \underline{E}$ are \emph{in special position with respect to $e$} if
\begin{itemize}
\item There is a unique $t \in \triangle$ such that $e,e' \subset t$ and $e,e'$ share the same starting endpoint;
\item $e$ and $t$ are in special position.
\end{itemize}
\end{itemize}
Hence we define:
\begin{align*}
\epsilon(e_1,e_2) &=
\begin{cases}
1 & \mbox{ if } e_1,e_2 \mbox{ are in special position with respect to } e_1;\\
-1 & \mbox{ if } e_1,e_2 \mbox{ are in special position with respect to } e_2;\\
0 & \mbox{ otherwise }.
\end{cases}\\
\epsilon(e,t) &=
\begin{cases}
1 & \mbox{ if } e^{-1},t \mbox{ are in special position};\\
-1 & \mbox{ if } e,t \mbox{ are in special position};\\
0 & \mbox{ otherwise }.
\end{cases}\\
\epsilon(t,e) &= - \epsilon(e,t) \qquad \mbox{ and } \qquad \epsilon(t_1,t_2) = 0,
\end{align*}
for all $e,e_1,e_2 \in \underline{E}$ and $t,t_1,t_2 \in \triangle$. We remark that this is the only place where we make use of assumption $(I)$. As mentioned before, one does not need such condition, but disposing of it makes the above definitions much more involved.
\begin{rem}
The function $\epsilon$ can also be understood graphically by considering the diagram shown in Figure~\ref{epsilon_diagram}. Here we symbolically represent a triangle by a point in its interior and an oriented edge by a point on the edge, closer to the starting endpoint. Then $\epsilon(q_1,q_2)$ is the difference between the number of arrows from $q_1$ to $q_2$ and the number of arrows from $q_2$ to $q_1$.
\end{rem}
\begin{figure}[ht]
\centering
\includegraphics[height=5cm]{epsilon_diagram.png}
\caption{$\epsilon$ can be read off of a graph. Here we have shown edges connecting marked points belonging to one triangle only. Points of other triangles are connected by arrows in the same ways. Points connected by edges without arrows are not taken into account.}
\label{epsilon_diagram}
\end{figure}
\begin{pro}
$\{ -,- \}_{\mathrm{T}}$ is a Poisson bracket.
\end{pro}
\begin{proof}
$\mathbb{R}$--bilinearity, skew-symmetry and Leibniz's product rule are simple to check. We only show that $\{ -,- \}_{\mathrm{T}}$ satisfies the Jacobi identity, that is
$$
\{f,\{g,h\}_{\mathrm{T}} \}_{\mathrm{T}} + \{g,\{h,f\}_{\mathrm{T}} \}_{\mathrm{T}} + \{h,\{f,g\}_{\mathrm{T}} \}_{\mathrm{T}} = 0,
$$
for all $f,g,h \in C^{\infty}(\mathbb{R}^{\triangle \cup \underline{E}}_{>0})$. To make the computations more readable, henceforth we write $\boldsymbol{f},\boldsymbol{g},\boldsymbol{h}$ and simplify the notation as follows
\begin{align*}
\{ \boldsymbol{f},\boldsymbol{g} \}_{\mathrm{T}} :=& \sum_{q_1,q_2 \in \triangle \cup \underline{E} } 2 \epsilon(q_1,q_2) q_1^* q_2^* \frac{\partial \boldsymbol{f}}{\partial \overline{q}_1} \frac{\partial \boldsymbol{g}}{\partial \overline{q}_2} \\
=& \sum_{q_i,q_j \in \triangle \cup \underline{E} } \left[2 \epsilon_{ij} q_i^* q_j^*\right] \frac{\partial \boldsymbol{f}}{\partial \overline{q}_i} \frac{\partial \boldsymbol{g}}{\partial \overline{q}_j} \\
=& \;\sum_{i,j} \omega_{ij} \partial_i \boldsymbol{f} \partial_j \boldsymbol{g}.
\end{align*}
Observe that $\omega_{ij} = - \omega_{ji}$. It follows that
\begin{align*}
\{\boldsymbol{f},\{\boldsymbol{g},\boldsymbol{h}\}_{\mathrm{T}} \}_{\mathrm{T}} &= \sum_{i,j,k,l} \omega_{kl} \partial_k \boldsymbol{f} \left( \partial_l \omega_{ij} \partial_i \boldsymbol{g} \partial_j \boldsymbol{h} + \omega_{ij} \partial_l \partial_i \boldsymbol{g} \partial_j \boldsymbol{h} + \omega_{ij} \partial_i \boldsymbol{g} \partial_l \partial_j \boldsymbol{h} \right),\\
\{\boldsymbol{g},\{\boldsymbol{h},\boldsymbol{f}\}_{\mathrm{T}} \}_{\mathrm{T}} &= \sum_{i,j,k,l} \omega_{kl} \partial_k \boldsymbol{g} \left( \partial_l \omega_{ij} \partial_i \boldsymbol{h} \partial_j \boldsymbol{f} + \omega_{ij} \partial_l \partial_i \boldsymbol{h} \partial_j \boldsymbol{f} + \omega_{ij} \partial_i \boldsymbol{h} \partial_l \partial_j \boldsymbol{f} \right),\\
\{\boldsymbol{h},\{\boldsymbol{f},\boldsymbol{g}\}_{\mathrm{T}} \}_{\mathrm{T}} &= \sum_{i,j,k,l} \omega_{kl} \partial_k \boldsymbol{h} \left( \partial_l \omega_{ij} \partial_i \boldsymbol{f} \partial_j \boldsymbol{g} + \omega_{ij} \partial_l \partial_i \boldsymbol{f} \partial_j \boldsymbol{g} + \omega_{ij} \partial_i \boldsymbol{f} \partial_l \partial_j \boldsymbol{g} \right).
\end{align*}
We factorise in terms of the partial derivatives of $\boldsymbol{f},\boldsymbol{g},\boldsymbol{h}$ to get the coefficients
\begin{align}
\partial_k \boldsymbol{f} \partial_i \boldsymbol{g} \partial_j \boldsymbol{h} &: \qquad \sum_l \left(\omega_{kl} \partial_l \omega_{ij} + \omega_{il} \partial_l \omega_{jk} + \omega_{jl} \partial_l \omega_{ki} \right), \hspace{2.5cm} \forall i,j,k, \label{eq:Jacobi-id-1} \\
\partial_k \boldsymbol{f} \partial_l \partial_i \boldsymbol{g} \partial_j \boldsymbol{h} &: \qquad \omega_{kl} \omega_{ij} + \omega_{jl} \omega_{ki}, \hspace{6cm} \forall i,j,k,l,\label{eq:Jacobi-id-2} \\
\partial_k \boldsymbol{f} \partial_i \boldsymbol{g} \partial_l \partial_j \boldsymbol{h} &: \qquad \omega_{kl} \omega_{ij} + \omega_{il} \omega_{jk}, \hspace{6cm} \forall i,j,k,l,\label{eq:Jacobi-id-3} \\
\partial_k \boldsymbol{g} \partial_i \boldsymbol{h} \partial_l \partial_j \boldsymbol{f} &: \qquad \omega_{kl} \omega_{ij} + \omega_{il} \omega_{jk}, \hspace{6cm} \forall i,j,k,l \label{eq:Jacobi-id-4}.
\end{align}
When $i = l$, the coefficient in $(\ref{eq:Jacobi-id-2})$ is $\omega_{ki} \omega_{ij} + \omega_{ji} \omega_{ki} = 0$.\\
When $i \not= l$, $\partial_k \boldsymbol{f} \partial_i \partial_l \boldsymbol{g} \partial_j \boldsymbol{h} = \partial_k \boldsymbol{f} \partial_l \partial_i \boldsymbol{g} \partial_j \boldsymbol{h}$ and we can sum their coefficients to get
$$
\omega_{kl} \omega_{ij} + \omega_{jl} \omega_{ki} + \omega_{ki} \omega_{lj} + \omega_{ji} \omega_{kl} = \omega_{kl}(\omega_{ij} + \omega_{ji}) + \omega_{ki}(\omega_{jl} + \omega_{lj}) = 0.
$$
A similar argument applies to $(\ref{eq:Jacobi-id-3})$ and $(\ref{eq:Jacobi-id-4})$. Our claim follows from showing that the coefficients in $(\ref{eq:Jacobi-id-1})$ are also zero. Since $\partial_l \omega_{ij} = 2 \epsilon(q_i,q_j)\partial_l(q_i^* q_j^*) = 2 \epsilon_{ij} \left( \delta_{il} q_j^* + \delta_{jl} q_i^* \right)$, by substituting in $(\ref{eq:Jacobi-id-1})$, we conclude that
\begin{align*}
&\sum_l \left(\omega_{kl} \partial_l \omega_{ij} + \omega_{il} \partial_l \omega_{jk} + \omega_{jl} \partial_l \omega_{ki} \right)\\
=&\sum_l 2 q_l^* \left[ \delta_{jl} q_i^*q_k^* \left( \epsilon_{kl}\epsilon_{ij} + \epsilon_{il} \epsilon_{jk} \right) + \delta_{il} q_j^*q_k^* \left( \epsilon_{kl}\epsilon_{ij} + \epsilon_{jl} \epsilon_{ki} \right) + \delta_{kl} q_i^*q_j^* \left( \epsilon_{il}\epsilon_{jk} + \epsilon_{jl} \epsilon_{ki} \right) \right]\\
=&\sum_l 2 q_l^* \left[ q_i^*q_k^* \left( \epsilon_{kj}\epsilon_{ij} + \epsilon_{ij} \epsilon_{jk} \right) + q_j^*q_k^* \left( \epsilon_{ki}\epsilon_{ij} + \epsilon_{ji} \epsilon_{ki} \right) + q_i^*q_j^* \left( \epsilon_{ik}\epsilon_{jk} + \epsilon_{jk} \epsilon_{ki} \right) \right]= 0.
\end{align*}
\end{proof}
The next result shows that $\{ -,- \}_{\mathrm{T}}$ does not depend on $\mathrm{T}$.
\begin{thm}\label{thm:bracket-tri-independent}
If $\mathrm{T}$ and $\mathrm{T}'$ are ideal triangulations of $S$, then $(\mathbb{R}^{\triangle \cup \underline{E}}_{>0},\{-,-\}_\mathrm{T})$ and $(\mathbb{R}^{\triangle' \cup \underline{E}'}_{>0},\{-,-\}_{\mathrm{T}'})$ are naturally isomorphic Poisson manifolds.
\end{thm}
\begin{proof}
Let $\phi_e : \mathbb{R}^{\triangle \cup \underline{E}}_{>0} \rightarrow \mathbb{R}^{\triangle' \cup \underline{E}'}_{>0}$ be the change of coordinates induced by flipping along an edge $e$ of $\mathrm{T}$ (see \S~\ref{subsec:change-of-coord} for details). Since $\phi_e$ is a bijection and any two ideal triangulations of $S$ differ by a finite sequence of edge flips, it is enough to prove that for all $f,g \in C^{\infty}(\mathbb{R}^{\triangle' \cup \underline{E}'}_{>0})$,
$$
\{f,g\}_{\mathrm{T}'} \circ \phi_e = \{ f \circ \phi_e, g \circ \phi_e \}_{\mathrm{T}},
$$
or equivalently
\begin{equation} \label{eq:bracket_tri_indep}
\left( \sum_{q_1',q_2' \in \triangle' \cup \underline{E}' } \epsilon(q_1',q_2') q_1'^* q_2'^* \frac{\partial f}{\partial \overline{q}_1'} \frac{\partial g}{\partial \overline{q}_2'} \right) \circ \phi_e = \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \frac{\partial f \circ \phi_e}{\partial \overline{q}_1} \frac{\partial g \circ \phi_e}{\partial \overline{q}_2}.
\end{equation}
Denote by $(\phi_e)_i := q_i'^* \circ \phi_e$ the $q_i'^*$ component of $\phi_e$. Then
$$
\text{LHS} = \sum_{q_1',q_2' \in \triangle' \cup \underline{E}' } \epsilon(q_1',q_2') (\phi_e)_1 (\phi_e)_2 \left(\frac{\partial f}{\partial \overline{q}_1'}\circ \phi_e\right) \left( \frac{\partial g}{\partial \overline{q}_2'} \circ \phi_e\right),
$$
\begin{align*}
\text{RHS} &= \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \left[ \sum_{q_1',q_2' \in \triangle' \cup \underline{E}' } \left(\frac{\partial f}{\partial \overline{q}_1'} \circ \phi_e \frac{ \partial (\phi_e)_1}{\partial \overline{q}_1}\right)\left(\frac{\partial g}{\partial \overline{q}_2'} \circ \phi_e \frac{ \partial(\phi_e)_2}{ \partial \overline{q}_2}\right) \right] \\
&= \sum_{q_1',q_2' \in \triangle' \cup \underline{E}' } \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \left(\frac{\partial f}{\partial \overline{q}_1'} \circ \phi_e \frac{ \partial (\phi_e)_1}{\partial \overline{q}_1}\right)\left(\frac{\partial g}{\partial \overline{q}_2'} \circ \phi_e \frac{ \partial (\phi_e)_2}{\partial \overline{q}_2}\right).
\end{align*}
Hence we rewrite equation (\ref{eq:bracket_tri_indep}) as
$$
\sum_{q_1',q_2' \in \triangle' \cup \underline{E}' } \left(\frac{\partial f}{\partial \overline{q}_1'} \circ \phi_e \right)\left(\frac{\partial g}{\partial \overline{q}_2'} \circ \phi_e\right) \left[ \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \frac{ \partial (\phi_e)_1}{ \partial \overline{q}_1} \frac{ \partial (\phi_e)_2}{ \partial \overline{q}_2} - \epsilon(q_1',q_2') (\phi_e)_1 (\phi_e)_2 \right] = 0
$$
By differentiating the functions in section \ref{subsec:change-of-coord}, one can easily check that for all $i,j$
$$
\sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \frac{ \partial (\phi_e)_1}{ \partial \overline{q}_1} \frac{ \partial (\phi_e)_2}{ \partial \overline{q}_2} - \epsilon(q_i',q_j') (\phi_e)_i (\phi_e)_j = 0.
$$
\end{proof}
Let $\eta$ be the change of coordinates induced by the sequence of edge flips between $\mathrm{T}$ and $\mathrm{T}'$. The following diagram commutes
\begin{center}
\begin{tikzpicture}
\matrix (m) [matrix of math nodes,row sep=3em,column sep=4em,minimum width=2em] {
\mathcal{T}_3^+(S) & \mathbb{R}^{\triangle \cup \underline{E}}_{>0} \\
& \mathbb{R}^{\triangle' \cup \underline{E}'}_{>0} \\ };
\path[-stealth]
(m-1-1) edge node [above] {$\phi_\mathrm{T}$} (m-1-2)
(m-1-2) edge node [right] {$\eta$} (m-2-2)
(m-1-1) edge node [below]{$\phi_{\mathrm{T}'}$} (m-2-2);
\end{tikzpicture}
\end{center}
and the pull-back of $\{-,-\}_\mathrm{T}$ via $\phi_\mathrm{T}$ does not depend on the coordinate system. Therefore for all $f,g \in C^{\infty}(\mathcal{T}_3^+(S) )$ and any ideal triangulation $\mathrm{T}$ of $S$, we define
$$
\{f,g\}_{FG} : = \{f \circ \phi^{-1}_\mathrm{T}, g \circ \phi^{-1}_\mathrm{T} \}_{\mathrm{T}},
$$
which makes $(\mathcal{T}_3^+(S),\{-,-\}_{FG})$ a Poisson manifold.
\subsubsection{Compatibility of Poisson brackets} \label{subsec:compatibility_brackets}
Let $\Tr_3 := \Tr(\mathfrak{X}_3(S)) \subset C^{\infty}(\mathfrak{X}_3(S))$ be the trace algebra with respect to $\SL(3,\mathbb{R})$ and let
\begin{align*}
\mu^* \colon C^{\infty}(\mathfrak{X}_3(S)) &\rightarrow C^{\infty}(\mathcal{T}_3^+(S))\\
f &\mapsto f \circ \mu
\end{align*}
be the injective homomorphism induced by the monodromy operator $\mu \colon \mathcal{T}_3^+(S) \rightarrow \mathfrak{X}_3(S)$. This section is devoted to the proof of the following result.
\begin{thm} \label{thm:compatibility-poisson-bracket}
The restriction of $\mu^*$ to $\Tr_3$ is a Poisson homomorphism, $i.e.$ for all $f,g \in \Tr_3$,
\begin{equation} \label{eq:compatibility-poisson-bracket}
2 \ \mu^* \circ \{f,g \}_{Gol} = \{\mu^* \circ f,\mu^* \circ g \}_{FG}.
\end{equation}
\end{thm}
The strategy we adopt consists of the following $5$ steps. First we apply Theorem~\ref{thm:global-coord} to reword the problem in terms of $\{-,-\}_{\mathrm{T}}$ for a fixed triangulation $\mathrm{T}$ of $S$. That allows us to compare the definition of $\{-,-\}_{\mathrm{T}}$ in (\ref{eq:triangulation-bracket}) with the formulation of $\{-,-\}_{Gol}$ in (\ref{eq:Gol-formula}). Next, we break up $\alpha$ and $\beta$ into pieces and focus our attention on \emph{tunnels and galleries}, where curves travel next to each other (see \S~\ref{subsubsec:proof-part2} for details). We then show in \S~\ref{subsubsec:proof-part3} how the global quantity $\{ \tr_\alpha \circ \mu, \tr_\beta \circ \mu\}_{\mathrm{T}}$ translates into a local behaviour $\{ \tr_\alpha, \tr_\beta \}_{Gol}$ at the intersection points of $\alpha$ and $\beta$. A special case when $\alpha$ and $\beta$ are \emph{parallel} is treated in \S~\ref{subsubsec:proof-part4}. All of the pieces are put together to conclude the proof in \S~\ref{subsubsec:proof-part5}.
\subsubsection{Proof part 1: set up} \label{subsubsec:proof-part1}
Fix an ideal triangulation $\mathrm{T}$ of $S$ and the corresponding isomorphism $\phi_{\mathrm{T}} : \mathbb{R}^{\triangle \cup \underline{E}}_{>0} \rightarrow \mathcal{T}_3^+(S)$. Let $(\mu \circ \phi_{\mathrm{T}})^* \colon C^{\infty}(\mathfrak{X}_3(S)) \rightarrow C^{\infty}(\mathbb{R}^{\triangle \cup \underline{E}}_{>0} )$ be the map $(\mu \circ \phi_{\mathrm{T}})^*(f) = f \circ \mu \circ \phi_{\mathrm{T}}$, then $\mu^*$ is a Poisson homomorphism on $\Tr$ if and only if $(\mu \circ \phi_{\mathrm{T}})^*$ is a Poisson homomorphism on $\Tr$. Henceforth we identify $\mathcal{T}_3^+(S)$ with $\mathbb{R}^{\triangle \cup \underline{E}}_{>0}$ and denote by $\mu,\mu^*$ the maps $\mu \circ \phi_{\mathrm{T}}, (\mu \circ \phi_{\mathrm{T}})^*$.
Recall that $\Tr_3$ is the sub-algebra of $C^\infty(\mathfrak{X}_3(S))$ generated by elements of the form $\tr_\alpha$ for $\alpha \in \pi_1(S)$. By Theorem~\ref{thm:Gol-formula} and the defining properties of Poisson brackets, it is enough to show that for all representatives $\alpha,\beta \subset S$ of immersed oriented closed curves in general position,
\begin{equation} \label{eq:poisson-equality}
\sum_{p \in \alpha \cap \beta} \epsilon(p,\alpha,\beta) \left( \tr_{\alpha_p \beta_p}\circ \mu - \frac{1}{3} (\tr_\alpha \circ \mu )(\tr_\beta \circ \mu ) \right) = \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \frac{\partial (\tr_\alpha \circ \mu)}{\partial \overline{q}_1} \frac{\partial (\tr_\beta \circ \mu)}{\partial \overline{q}_2}.
\end{equation}
Here $\alpha$ and $\beta$ are assumed to be in general position both with respect to each other and with respect to $\mathrm{T}$.
We start by recalling some notation from \S~\ref{subsect_monodromy_operator}. For $\gamma \in \pi_1(S)$, we showed in Lemma~\ref{lem_compute_monodromy} that there exists a finite sequence $(e^\gamma_0 \cdot \dots \cdot e^\gamma_n)$ of $n+1$ oriented edges $e^\gamma_i \in \underline{E}^\mathcal{G}$ such that $e^\gamma_0 \cdot e^\gamma_{1} \cdot \dots \cdot e^\gamma_n \subset \mathcal{G}$ is the development of a loop in $\mathcal{G}_0$ homotopic equivalent to $\gamma$. Let $(s^\gamma_0 \cdot \dots \cdot s^\gamma_n)$ be such a loop, where $s^\gamma_i \in \underline{E}^{\mathcal{G}_0}$ corresponds to $e^\gamma_i \in \underline{E}^\mathcal{G}$. Then we showed in Theorem~\ref{thm:monodromy} that $\mu(x)(\gamma) $ is the conjugacy class of a product $S^\gamma_0\cdot S^\gamma_1 \cdots S^\gamma_n$, where $S^\gamma_i$'s are matrices of the type $T,T^{-1}$ or $E$. More precisely, $S^\gamma_i$ is of type $T,T^{-1}$ (respectively type $E$) if and only if $s^\gamma_i$ is a $\triangle$--edge (resp. $\underline{E}$--edge), as described at the beginning of \S~\ref{subsect_monodromy_operator}. Furthermore, we can assume that $s^\gamma_{i}$ and $s^\gamma_{i+1}$ are of different type for all $i \in \ZZ_{n+1}$.
For the remainder of this chapter we will work with fixed sequences $e_i^\alpha,e_i^\beta$, loops $s_i^\alpha,s_i^\beta$ and matrices $S_i^\alpha,S_i^\beta \in \SL(3,\mathbb{R})$ for $\alpha$ and $\beta$.
For convenience, we provide here a simple result which will be used later on.
\begin{lem} \label{lemma_basic_derivatives}
Let $T(z),E(x,y) \in \SL(3,\mathbb{R})$ be the matrices defined in Lemma~\ref{lem_example_monodromy}, and let
$$
M_1 := \left(
\begin{array}{ccc}
-2 & 0 & 0\\
0 & 1 & 0\\
0 & 0 & 1\\
\end{array} \right),
M_2 := \left(
\begin{array}{ccc}
-1 & 0 & 0\\
0 & -1 & 0\\
0 & 0 & 2\\
\end{array} \right),
M_3 := \left(
\begin{array}{ccc}
2 & 3 & 0\\
0 & -1 & 0\\
0 & 0 & -1\\
\end{array} \right),
M_4 := \left(
\begin{array}{ccc}
1 & 0 & 0\\
0 & 1 & 0\\
0 & -3 & -2\\
\end{array} \right).
$$
Then
$$
\frac{\partial E(x,y)}{\partial x} = \frac{1}{3x}E(x,y) \cdot M_1, \qquad \qquad
\frac{\partial E(x,y)}{\partial y} = \frac{1}{3y}E(x,y) \cdot M_2,
$$
$$
\frac{\partial T(z)}{\partial z} = \frac{1}{3z}T(z) \cdot M_3, \qquad \qquad
\frac{\partial T(z)^{-1}}{\partial z} = \frac{1}{3z}T(z)^{-1} \cdot M_4.
$$
\end{lem}
\subsubsection{Proof part 2: tunnels and galleries} \label{subsubsec:proof-part2}
The function $\{f,g\}_{FG}$ does not depend on the chosen triangulation $\mathrm{T}$ so we may assume, without loss of generality, that:
\begin{itemize}
\item[(I)] \emph{every edge in $\mathrm{T}$ is contained in two distinct triangles, or equivalently, every triangle has three distinct edges.}
\end{itemize}
As before, we underline that this assumption is not necessary and we will specify where we use it.
Denote by $\partial_i f := \frac{\partial f }{\partial \overline{q}_i}$, then by the product rule
$$
\partial_i (\tr_\gamma \circ \mu) = \partial_i (\tr \mu(\gamma)) = \tr (\partial_i \mu(\gamma)) = \sum_j \tr \left(S^\gamma_0 \cdots \partial_i S^\gamma_j \cdots S^\gamma_n\right).
$$
Denote $\partial_i s^\gamma_j := S^\gamma_0 \cdots \partial_i S^\gamma_j \cdots S^\gamma_n$ for all $i,j$. We remind the reader that $s^\gamma_j$ is a segment of a loop homotopic to $\gamma$, so the notation $\partial_i s^\gamma_j$ is chosen to emphasize that the partial derivative is ``taken at $s^\gamma_j$".
It follows that
$$
\partial_1 (\tr_\alpha \circ \mu) \cdot \partial_2 (\tr_\beta \circ \mu) = \sum_{i,j} \tr (\partial_1 s^\alpha_i) \cdot \tr (\partial_2 s^\beta_j),
$$
and the right hand side of (\ref{eq:poisson-equality}) can be reformulated as
$$
\sum_{i,j} \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \left(\tr (\partial_1 s^\alpha_i) \cdot \tr (\partial_2 s^\beta_j)\right).
$$
\begin{rem} \label{rem_segment_derivative}
Lemma \ref{lemma_basic_derivatives} implies that $\partial_i s^\gamma_j = \frac{1}{3 q_i^* } \left(S^\gamma_0 \cdots S^\gamma_j \cdot X \cdots S^\gamma_n \right)$, where $X$ is:
\begin{itemize}
\item[$i)$] $M_1$ or $M_2$, if $s^\gamma_j$ is an $\underline{E}$--edge crossing an oriented edge of $\mathrm{T}$ with edge ratio $q_i$;
\item[$ii)$] $M_3$ or $M_4$, if $s^\gamma_j$ is a $\triangle$--edge contained in a triangle of $\mathrm{T}$ with triple ratio $q_i$;
\item[$iii)$] the zero matrix otherwise.
\end{itemize}
\end{rem}
We are now going to investigate the set of pairs $(s^\alpha_i,s^\beta_j)$ and their contribution
$$
C(s_i^{\alpha},s_j^{\beta}) := \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \left(\tr (\partial_1 s^\alpha_i) \cdot \tr (\partial_2 s^\beta_j)\right)
$$
in the above sum.\\
Clearly $C(s_i^{\alpha},s_j^{\beta}) = 0$ when $s^\alpha_i $ and $ s^\beta_j$ do not intersect a common triangle. If $\alpha$ and $\beta$ never cross a common triangle, then they must be disjoint, hence both right hand side and left hand side in (\ref{eq:poisson-equality}) are $0$.\\
For that reason, we assume from now on that
\begin{itemize}
\item[(II)] \emph{there exist segments $s^\alpha_i $ and $ s^\beta_j$ intersecting a common triangle $\hat{T}$.}
\end{itemize}
When $t$ is a common triangle to two segments $s^\alpha_i $ and $ s^\beta_j$ it follows from assumption (I) that there is a unique (oriented) connected component $t(s^\alpha_i)$ (resp. $t(s^\beta_j)$) of $t \cap \alpha$ (resp. $t \cap \beta$) intersecting $s^\alpha_i$ (resp. $s^\beta_j$). We will say that $t$ has the \emph{configuration type} of:
\begin{itemize}
\item a \emph{starting triangle} (resp. \emph{ending} triangle) with respect to $s^\alpha_i $ and $ s^\beta_j$, if $t(s^\alpha_i)$ and $t(s^\beta_j)$ cross a unique common edge of $t$ and $t(s^\alpha_i)$ leaves (resp. enters) $t$ from such edge;
\item a \emph{transition triangle} with respect to $s^\alpha_i $ and $ s^\beta_j$, if $t(s^\alpha_i)$ and $t(s^\beta_j)$ cross the same two edges of $t$.
\end{itemize}
Up to interchanging $\alpha$ and $\beta$, there are only $7$ possible configurations. They are listed in Figure~\ref{triangle_config_types} and call $jS$ the configuration of type $j$ after interchanging.
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{triangle_config_types.png}
\caption{Triangle configuration types.}
\label{triangle_config_types}
\end{figure}
Furthermore, we will say that $w_\gamma = (w^\gamma_0 \cdots w^\gamma_k)$ is a \emph{walk along $(s^\gamma_0 \cdots s^\gamma_n)$} of length $k+1 \in \NN$ based at $s^\gamma_{j}$ if either
$$
w^\gamma_i = s^\gamma_{[j+i]_{n+1}} \qquad \mbox{ or } \qquad w^\gamma_i = s^\gamma_{[j-i]_{n+1}},
$$
where $[m]_{n+1} := m \mod (n+1)$. When $\gamma$ is represented by a loop $(s^\gamma_0 \cdots s^\gamma_n)$, we will abbreviate the notation by saying that $w_\gamma$ is a \emph{walk along $\gamma$}.
\begin{rem}
Broadly speaking $w_\gamma$ is a walk along $(s^\gamma_0 \cdots s^\gamma_n)$ starting at $s^\gamma_{j}$ which might wrap around the curve multiple times, possibly backwards with respect to its orientation. In particular, $(s^\gamma_0 \cdot s^\gamma_1 \cdots s^\gamma_n)$ and $(s^\gamma_0 \cdot s^\gamma_n \cdots s^\gamma_1)$ are walks of length $n+1$ based at $s^\gamma_{0}$ in different directions.
\end{rem}
For the moment, we will make the extra assumption that
\begin{itemize}
\item[$(\star)$] \emph{$\alpha$ and $\beta$ are \emph{non-parallel}}
\end{itemize}
that is $[\alpha] \not= [\beta^m]$ for all $m \in \ZZ$. We come back to this particular case later in \S~\ref{subsubsec:proof-part4}.\\
Since $\alpha$ and $\beta$ are closed non-parallel curves, if $t$ is a common triangle to two segments $s^\alpha_i $ and $ s^\beta_j$ then there exist:
\begin{itemize}
\item a sequence of adjacent (not necessarily distinct) triangles $(t_0,\dots,t_q) \subset \mathrm{T}$ containing $t$;
\item a walk $w_{\alpha} := (w^\alpha_0 \cdot \dots \cdot w^\alpha_{2q+2})$ along $\alpha$ containing $s^\alpha_i $;
\item a walk $w_{\beta} := (w^\beta_0 \cdot \dots \cdot w^\beta_{2q+2})$ along $\beta$ containing $s^\beta_j $;
\end{itemize}
such that $t_0$ is a starting triangle with respect to the pairs $(w^\alpha_i,w^\beta_i)_{0\leq i \leq 2}$, $t_q$ is an ending triangle with respect to the pairs $(w^\alpha_i,w^\beta_i)_{2q \leq i \leq 2q+2}$, and $t_j$ is a transition triangle with respect to the pairs $(w^\alpha_i,w^\beta_i)_{2j \leq i \leq 2j+2}$ for all $1\leq j \leq q-1$.\\
In this case we will say that $(w_{\alpha}, w_{\beta})$ is a \emph{tunnel through} $(t_0,\dots,t_q)$ \emph{containing} $(s^\alpha_i ,s^\beta_j)$. Furthermore, for all $k \in \{0 \dots q \}$ there is a preferred triple of pairs $(w^\alpha_i,w^\beta_i)_{2k \leq i \leq 2k+2}$ and a triangle $t_k$ associated to it, which will be called the \emph{$k^{th}$ gallery} of the tunnel. By definition every $j^{th}$ and $(j+1)^{th}$ gallery share a common pair $(w^\alpha_{2j+2},w^\beta_{2j+2})$.
\begin{rem}
A good way to visualise a tunnel is to pass to a universal cover of the surface. If $(w_{\alpha}, w_{\beta})$ is a tunnel through $(t_0,\dots,t_q)$ and we lift $\alpha$ and $\beta$ to the universal cover $\widetilde{S}$ based at a some point $p$ in any of the triangles $t_i$, then the two curves in $\widetilde{S}$ travel through common triangles only along the tunnel while apart from each other everywhere else.
\end{rem}
Henceforth, when we talk about a tunnel containing the pair $(s^\alpha_i, s^\beta_j)$, we will further assume that $s^\alpha_i$ and $ s^\beta_j$ are contained in the same gallery. In this way, for every pair intersecting a common triangle there is a unique tunnel containing them. We stress again that the assumption $(\star)$ is necessary for the existence of at least one tunnel, as the reader can be easily convinced by taking simple closed curves $\alpha = \beta$.
\subsubsection{Proof part 3: encoding the contribution along a tunnel} \label{subsubsec:proof-part3}
Let $(w_{\alpha}, w_{\beta})$ be a tunnel $\tau$ through $(t_0,\dots,t_q)$. We now develop a handy way to encode the \emph{contribution $C_k(\tau)$ of the $k^{th}$ gallery of $\tau$}, that is the quantity
\begin{eqnarray} \label{eq_gallery_contrib}
C_k(\tau) := \sum_{2k \leq i,j \leq 2k+2} \left( \sum_{q_1,q_2 \subset t_k } \epsilon(q_1,q_2) q_1^* q_2^* \left(\tr (\partial_1 w^\alpha_i) \cdot \tr (\partial_2 w^\beta_j)\right)\right).
\end{eqnarray}
Following the notation of the previous paragraph, let $W^\gamma_{i}$ be the matrix in $S^\gamma_0\cdot S^\gamma_1 \cdots S^\gamma_n$ corresponding to the segment $w_{i}^\gamma$ in $(s^\gamma_0 \cdot \dots \cdot s^\gamma_n)$. Then for all $k \in \{0,\dots,q\}$, $m \in \{0,1,2\}$ and $\gamma \in \{\alpha,\beta \}$,
\begin{align*}
q_i^* \tr( \partial_i w_{2k+m}^\gamma ) & = q_i^* \tr( S^\gamma_0 \cdots \partial_i W^\gamma_{2k+m} \cdots S^\gamma_n )\\
& = \frac{1}{3} \tr( S^\gamma_0 \cdots W^\gamma_{2k+m} \cdot \widetilde{X} \cdots S^\gamma_n ) \qquad \mbox{ see Remark~\ref{rem_segment_derivative}, }\\
& = \frac{1}{3} \tr( W^\gamma_{2k+m} \cdot \widetilde{X} \cdot \cdots S^\gamma_n \cdot S^\gamma_0 \cdots ) \qquad \mbox{ by conjugation invariance of } \tr,\\
& = \frac{1}{3} \tr( X \cdot W^\gamma_{2k+1} \cdot \cdots S^\gamma_n \cdot S^\gamma_0 \cdots ),\\
\end{align*}
where $\widetilde{X} = \widetilde{X}(\gamma,q_i,k,m)$ is a matrix dependent on $(\gamma,q_i,k,m)$ and
\begin{itemize}
\item for $\gamma = \alpha$,
$$
X = \left(\prod_{i = 1}^{m} W^\alpha_{2k+i} \right)\cdot \widetilde{X} \cdot \left(\prod_{i = 1}^{m} W^\alpha_{2k+i} \right)^{-1};
$$
\item for $\gamma = \beta$, if $\beta$ is oriented as $\alpha$ along the tunnel,
$$X = \left(\prod_{i = 1}^{m} W^\beta_{2k+i} \right)\cdot \widetilde{X} \cdot \left(\prod_{i = 1}^{m} W^\beta_{2k+i} \right)^{-1};
$$
otherwise
$$
X = \left(\prod_{i = 0}^{1-m} W^\beta_{2k+1-i} \right)\cdot \widetilde{X} \cdot \left(\prod_{i = 0}^{1-m} W^\beta_{2k+1-i} \right)^{-1}.
$$
\end{itemize}
In short we can write
$$
q_i^* \tr( \partial_i w_{2k+m}^\gamma ) = \frac{1}{3} \tr( X \cdot \mathbf{S}(w^\gamma_{2k+1}) ),
$$
where $X = X(\gamma,q_i,k,m)$ is a matrix dependent on $(\gamma,q_i,k,m)$ and
$$
\mathbf{S}(w^\gamma_{2k+1}) := W^\gamma_{2k+1} \cdot \cdots S^\gamma_n \cdot S^\gamma_0 \cdots,
$$
is the matrix conjugate to $S^\gamma_0 \cdots W^\gamma_{2k+1} \cdots S^\gamma_n$ obtained by moving all the matrices on the left side of $W^\gamma_{2k+1}$ to the right side of $S^\gamma_n$.
\begin{lem} \label{lem_trace_factorisation}
Let $X,Y \in M_{3 \times 3}(\mathbb{R})$ be real matrices and let $\mathbf{E}_{ij} \in M_{3 \times 3}(\mathbb{R})$ be the matrix with $(i,j)^{th}$ entry $1$ and zeros everywhere else. Then
$$
\tr(X \cdot Y) = \sum_{i,j} x_{ij} \tr(\mathbf{E}_{ij} \cdot Y),
$$
where $x_{ij}$ is the $(i,j)^{th}$ entry of $X$.
\end{lem}
By Lemma~\ref{lem_trace_factorisation} and the above discussion we can conclude that
\begin{align*}
C_k(\tau) = & \sum_{0 \leq m_1,m_2 \leq 2} \left( \sum_{q_1,q_2 \subset t_k } \epsilon(q_1,q_2) q_1^* q_2^* \left(\tr (\partial_1 w^\alpha_{2k+m_1}) \cdot \tr (\partial_2 w^\beta_{2k+m_2})\right)\right) \\
=& \frac{1}{9} \sum_{0 \leq m_1,m_2 \leq 2} \left( \sum_{q_1,q_2 \subset t_k } \epsilon(q_1,q_2) \bigg( \tr \Big( X(\alpha,q_1,k,m_1) \cdot \mathbf{S}(w^\alpha_{2k+1}) \Big) \cdot \tr \Big( X(\beta,q_2,k,m_2) \cdot \mathbf{S}(w^\beta_{2k+1}) \Big) \bigg) \right) \\
=& \frac{1}{9} \sum_{l,m} \sum_{0 \leq m_1,m_2 \leq 2} \left( \sum_{q_1,q_2 \subset t_k } \epsilon(q_1,q_2) \bigg( x^{(\alpha,q_1,k,m_1)}_{lm} \tr \Big( \mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1}) \Big) \cdot \tr \Big( X(\beta,q_2,k,m_2) \cdot \mathbf{S}(w^\beta_{2k+1}) \Big) \bigg) \right) \\
=& \frac{1}{9} \sum_{l,m} \tr \Big( \mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1}) \Big) \sum_{0 \leq m_1,m_2 \leq 2} \left( \sum_{q_1,q_2 \subset t_k } \epsilon(q_1,q_2) \bigg( x^{(\alpha,q_1,k,m_1)}_{lm} \cdot \tr \Big( X(\beta,q_2,k,m_2) \cdot \mathbf{S}(w^\beta_{2k+1}) \Big) \bigg) \right) \\
=& \frac{1}{9} \sum_{l,m} \tr \Big( \mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1}) \Big) \cdot \tr \Big( \left( \sum_{0 \leq m_1,m_2 \leq 2} \sum_{q_1,q_2 \subset t_k } \epsilon(q_1,q_2) x^{(\alpha,q_1,k,m_1)}_{lm} X(\beta,q_2,k,m_2) \right) \cdot \mathbf{S}(w^\beta_{2k+1}) \Big) \\
=&\frac{1}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( U^{(l,m)} \cdot \mathbf{S}(w^\beta_{2k+1}) ).
\end{align*}
where
\begin{equation}\label{eq:contribution-matrix-entries}
U^{(l,m)} := \sum_{0 \leq m_1,m_2 \leq 2} \sum_{q_1,q_2 \subset t_k } \epsilon(q_1,q_2) x^{(\alpha,q_1,k,m_1)}_{lm} X(\beta,q_2,k,m_2).
\end{equation}
with $x^{(\alpha,q_1,k,m_1)}_{lm}$ being the $(l,m)^{th}$ entry of $X(\alpha,q_1,k,m_1)$.\\
Let $\mathbf{M}_k(\tau) \in M_{3\times 3} (M_{3\times 3}(\mathbb{R}))$ be the $3 \times 3$ matrix whose $(l,m)^{th}$ entry is the $3 \times 3$ matrix $U^{(l,m)}$. Then $\mathbf{M}_k(\tau)$ encodes $C_k(\tau)$ in the sense that
$$
C_k(\tau) = \frac{1}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( (\mathbf{M}_k(\tau))_{lm} \cdot \mathbf{S}(w^\beta_{2k+1}) ).
$$
The matrix $\mathbf{M}_k(\tau)$ is unique and uniquely determines $C_k(\tau)$, thus we call $\mathbf{M}_k(\tau)$ the \emph{contribution matrix} of the $k^{th}$ gallery of $\tau$.
\begin{exa}
We show how to compute the contribution matrix for a gallery with an explicit example. Consider the $0{th}$ gallery of a tunnel $\tau$ where $t_0$ is a starting triangle of type $1$ with respect to the pairs $(w^\alpha_i,w^\beta_i)_{0\leq i \leq 2}$ as in Figure~\ref{example_gallery_contrib}.
\begin{figure}[ht]
\centering
\includegraphics[height=5cm]{example_gallery_contribution.png}
\caption{Example of contribution along a gallery with triangle of type $1$.}
\label{example_gallery_contrib}
\end{figure}
For some matrices $A,B$,
\begin{align*}
\mu(x)(\alpha) &= [E(a,b) \cdot T(z)^{-1} \cdot E(c,d) \cdot A],\\
\mu(x)(\beta) &= [E(e,f) \cdot T(z) \cdot E(c,d) \cdot B].
\end{align*}
We apply Lemma~\ref{lemma_basic_derivatives} and the conjugacy invariance of the trace operator to obtain:
\begin{align*}
q_j^* \tr (\partial_j w^\alpha_i) &= \frac{1}{3} \tr( X^{(\alpha,q_j,i)} \cdot T(z)^{-1} \cdot E(c,d) \cdot A \cdot E(a,b) ),\\
q_j^* \tr (\partial_j w^\beta_i) &= \frac{1}{3} \tr( X^{(\beta,q_j,i)} \cdot T(z) \cdot E(c,d) \cdot B \cdot E(e,f) ),
\end{align*}
where explicitly (see Lemma~\ref{lemma_basic_derivatives} for the notation):
\begin{align*}
X^{(\alpha,a,0)} = M_1 = X^{(\beta,e,0)} &\qquad X^{(\alpha,b,0)} = M_2 = X^{(\beta,f,0)}\\
X^{(\alpha,z,1)} = T(z)^{-1} \cdot M_4 \cdot T(z) &\qquad X^{(\beta,z,1)} = T(z) \cdot M_3 \cdot T(z)^{-1}\\
X^{(\alpha,c,2)} = T(z)^{-1} \cdot E(c,d) \cdot M_1 \cdot (T(z)^{-1} \cdot E(c,d))^{-1} &\qquad X^{(\beta,c,2)} = T(z) \cdot E(c,d) \cdot M_1 \cdot (T(z) \cdot E(c,d))^{-1}\\
X^{(\alpha,d,2)} = T(z)^{-1} \cdot E(c,d) \cdot M_2 \cdot (T(z)^{-1} \cdot E(c,d))^{-1} &\qquad X^{(\beta,d,2)} = T(z) \cdot E(c,d) \cdot M_2 \cdot (T(z) \cdot E(c,d))^{-1}\\
X^{(\alpha,q_j,i)} = X^{(\beta,q_j,i)} = &\mbox{ the zero matrix otherwise}.
\end{align*}
We can now compute the entries of $\mathbf{M}_0(\tau)$ from equation (\ref{eq:contribution-matrix-entries}), for example:
\begin{align*}
U^{(1,1)} = &x^{(\alpha,a,0)}_{11}\left( \epsilon(a,f)X^{(\beta,f,0)} + \epsilon(a,z)X^{(\beta,z,1)} \right) + x^{(\alpha,b,0)}_{11}\left( \epsilon(b,d)X^{(\beta,d,2)} + \epsilon(b,z)X^{(\beta,z,1)} \right) +\\ &x^{(\alpha,z,1)}_{11}\left( \epsilon(z,f)X^{(\beta,f,0)} + \epsilon(z,e)X^{(\beta,e,0)} + \epsilon(z,c)X^{(\beta,c,2)} + \epsilon(z,d)X^{(\beta,d,2)} \right) +\\ &x^{(\alpha,c,2)}_{11}\left( \epsilon(c,e)X^{(\beta,e,0)} + \epsilon(c,z)X^{(\beta,z,1)} \right) + x^{(\alpha,d,2)}_{11}\left( \epsilon(d,z)X^{(\beta,z,1)} \right).
\end{align*}
It follows that the contribution matrix of the $0^{th}$ gallery is
$$
\mathbf{M}_0(\tau) =
\left(\begin{array}{ccc}
\left(\begin{array}{ccc}
3 & 0 & 0\\
-6 & -3 & 0\\
3(z+2) & 3(z+1) & 0
\end{array}\right) &
\left(\begin{array}{ccc}
\frac{3(z+1)}{z} & 0 & 0\\
-9 & -\frac{6(z+1)}{z} & 0\\
9(z+1) & \frac{9(z+1)^2}{z} & \frac{3(z+1)}{z}
\end{array}\right) &
\left(\begin{array}{ccc}
\frac{3(z-1)}{z} & 0 & 0\\
0 & -\frac{3(z+2)}{z} & 0\\
0 & \frac{9(z+1)}{z} & \frac{3(2z+1)}{z}
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
3 & 3 & 0\\
-3(2z+1) & -6(z+1) & -3
\end{array}\right) &
\left(\begin{array}{ccc}
3 & 0 & 0\\
0 & 3 & 0\\
0 & -9 & -6
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
-3 & 0 & 0\\
3 & 0 & 0\\
3(z-1) & 3(z+1) & 3
\end{array}\right)\\
\end{array}\right)
$$
and
$$
C_0(\tau) = \frac{1}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot T(z)^{-1} \cdot E(c,d) \cdot A \cdot E(a,b)) \cdot \tr( (\mathbf{M}_0(\tau))_{lm} \cdot T(z) \cdot E(c,d) \cdot B \cdot E(e,f) ).
$$
\end{exa}
\begin{lem} \label{lem:sum-of-contributions}
Let $C_{k-1}(\tau),C_{k}(\tau)$ and $\mathbf{M}_{k-1}(\tau), \mathbf{M}_{k}(\tau)$ be the contributions and contribution matrices of two consecutive galleries in a tunnel $\tau$, and let
\begin{itemize}
\item $a^{(l,m)}_{ij}$ be the $(i,j)$--th entry of the matrix $A^{(l,m)} = (W^\alpha_{2k-1} \cdot W^\alpha_{2k})^{-1} \cdot \mathbf{E}_{lm} \cdot W^\alpha_{2k-1} \cdot W^\alpha_{2k}$,
\item if $\alpha,\beta$ are oriented in the same direction along the galleries,
$$
B^{(l,m)} = (W^\beta_{2k-1} \cdot W^\beta_{2k})^{-1} \cdot (\mathbf{M}_{k-1}(\tau))_{lm} \cdot W^\beta_{2k-1} \cdot W^\beta_{2k},
$$
otherwise
$$
B^{(l,m)} = W^\beta_{2k+1} \cdot W^\beta_{2k} \cdot (\mathbf{M}_{k-1}(\tau))_{lm} \cdot ( W^\beta_{2k+1} \cdot W^\beta_{2k} )^{-1}.
$$
\end{itemize}
Then
$$
C_{k-1}(\tau) + C_{k}(\tau) = \frac{1}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( ({\mathbf{M}_{(k-1,k)}}(\tau))_{lm} \cdot \mathbf{S}(w^\beta_{2k+1}) ),
$$
where $\mathbf{M}_{(k-1,k)}(\tau) \in M_{3\times 3} (M_{3\times 3}(\mathbb{R})) $ is the matrix with $(l,m)^{th}$ entry
$$
(\mathbf{M}_{(k-1,k)}(\tau))_{lm} = \sum_{i,j} a^{(i,j)}_{lm} B^{(i,j)} + (\mathbf{M}_k(\tau))_{lm}.
$$
\end{lem}
\begin{proof}
\begin{align*}
C_{k-1}(\tau) & = \frac{1}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k-1})) \cdot \tr( (\mathbf{M}_{k-1}(\tau))_{lm} \cdot \mathbf{S}(w^\beta_{2k-1}) )\\
& = \frac{1}{9}\sum_{l,m} \tr(A^{(l,m)} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( B^{(l,m)} \cdot \mathbf{S}(w^\beta_{2k+1}) ) \qquad \mbox{ by conjugation,}\\
& = \frac{1}{9}\sum_{l,m} \sum_{i,j} a^{(l,m)}_{ij} \tr( \mathbf{E}_{ij} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( B^{(l,m)} \cdot \mathbf{S}(w^\beta_{2k+1}) ) \qquad \mbox{ from Lemma~\ref{lem_trace_factorisation},}\\
& = \frac{1}{9}\sum_{i,j} \tr(\mathbf{E}_{ij} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( \sum_{l,m} a^{(l,m)}_{ij} \cdot B^{(l,m)} \cdot \mathbf{S}(w^\beta_{2k+1}) )\\
& = \frac{1}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( \sum_{i,j} a^{(i,j)}_{lm} \cdot B^{(i,j)} \cdot \mathbf{S}(w^\beta_{2k+1}) ) \qquad \mbox{ after renaming indices.}
\end{align*}
But
$$
C_k(\tau) = \frac{1}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( (\mathbf{M}_k(\tau))_{lm} \cdot \mathbf{S}(w^\beta_{2k+1}) ),
$$
hence the result follows.
\end{proof}
\begin{rem}
From a computational point of view, it is interesting to observe that the matrix $\sum_{i,j} a^{(i,j)}_{lm} \cdot B^{(i,j)}$ is the $(l,m)^{th}$ entry of $(W^\alpha_{2k-1} \cdot W^\alpha_{2k})^{-1} \cdot \mathbf{B} \cdot W^\alpha_{2k-1} \cdot W^\alpha_{2k}$, where $\mathbf{B} \in M_{3\times 3} (M_{3\times 3}(\mathbb{R})) $ is the matrix with matrix entries $B^{(i,j)}$.
\end{rem}
For each starting triangle configuration, there are only two compatible ending triangles, which makes a total of $8$ possible tunnels without transition triangles (cf. Figure~\ref{fig:length-two-tunnel}).
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{length_two_tunnels.png}
\caption{Length two tunnels.}
\label{fig:length-two-tunnel}
\end{figure}
\begin{lem} \label{lem:two-galleries-contribution}
The contribution along tunnels of length two equals
$$
\epsilon(p,\alpha,\beta) \left( \tr_{\alpha_p \beta_p}\circ \mu - \frac{1}{3} (\tr_\alpha \circ \mu )(\tr_\beta \circ \mu ) \right),
$$
where $p$ is the point of intersection (if any) of $\alpha$ and $\beta$ in the tunnel.
\end{lem}
\begin{proof}
The same technique applies to each case, hence we here our attention to just one of them. We will briefly explain at the end how to treat the others.
Let $(w^{\alpha}_i, w^{\beta}_i)_{0\leq i \leq 4}$ be a tunnel $\tau$ through $(t_0,t_1)$ of type C (as in Figure~\ref{fig:length-two-tunnel}, namely $t_0$ is a starting triangle of type $4S$ (which is type $4$ with $\alpha$ and $\beta$ switched) with respect to the pairs $(w^{\alpha}_i, w^{\beta}_i)_{0\leq i \leq 2}$, and $t_1$ is an ending triangle of type $4$ for $(w^{\alpha}_i, w^{\beta}_i)_{2\leq i \leq 4}$. According to the labels shown in Figure~\ref{example_type_c}, we have
\begin{align*}
\mu(x)(\alpha) &= [E(f,e) \cdot T(z) \cdot E(d,c) \cdot T(w) \cdot E(g,h) \cdot A],\\
\mu(x)(\beta) &= [E(j,i) \cdot T(w) \cdot E(c,d) \cdot T(z) \cdot E(a,b) \cdot B],
\end{align*}
for some matrices $A,B$.
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{example_type_c.png}
\caption{Example of type C.}
\label{example_type_c}
\end{figure}
The contribution matrices $\mathbf{M}_0(\tau), \mathbf{M}_1(\tau)$ at $t_0,t_1$ are:
$$
\mathbf{M}_0(\tau) =
\left(\begin{array}{ccc}
\left(\begin{array}{ccc}
0 & \frac{3(z+1)}{z} & \frac{3(2z+1)}{z}\\
0 & -3 & -6\\
0 & 0 & 3
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & \frac{9(z+1)}{z}\\
0 & 0 & -9\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & \frac{9}{z}\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
-3 & -\frac{6(z+1)}{z} & -\frac{3(z+2)}{z}\\
0 & 3 & 3\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & -\frac{9}{z} & -\frac{9}{z}\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
3 & \frac{3(z+1)}{z} & -\frac{3(z-1)}{z}\\
0 & 0 & 3\\
0 & 0 & -3
\end{array}\right)\\
\end{array}\right),
$$
$$
\mathbf{M}_1(\tau) =
\left(\begin{array}{ccc}
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 3 & 0\\
0 & 0 & -3
\end{array}\right) &
\left(\begin{array}{ccc}
-\frac{3(w+1)}{w} & 0 & 0\\
0 & \frac{6(w+1)}{w} & \frac{9}{w}\\
0 & 0 & -\frac{3(w+1)}{w}
\end{array}\right) &
\left(\begin{array}{ccc}
-\frac{3(2w+1)}{w} & -\frac{9(w+1)}{w} & -\frac{9}{w}\\
0 & \frac{3(w+2)}{w} & \frac{9}{w}\\
0 & 0 & \frac{3(w-1)}{w}
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
3 & 0 & 0\\
0 & -3 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
6 & 9 & 0\\
0 & -3 & 0\\
0 & 0 & -3
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
-3 & 0 & 0\\
0 & 0 & 0\\
0 & 0 & 3
\end{array}\right)\\
\end{array}\right).
$$
Hence we apply Lemma~\ref{lem:sum-of-contributions} to check that $\mathbf{M}_{(0,1)}(\tau)$ is the zero matrix, which implies that the contribution along the tunnel of type C is zero. Since there is no point of intersection $p$ in the tunnel $\epsilon(p,\alpha,\beta) = 0$ and the statement is proved for type C.
Cases A, B and D are treated verbatim, while for E, F, G and H one needs only to check that
$$
\epsilon(p,\alpha,\beta) \frac{\mathbf{M}_{(0,1)}(\tau) }{9} =
\left(\begin{array}{ccc}
\left(\begin{array}{ccc}
\frac{2}{3} & 0 & 0\\
0 & -\frac{1}{3} & 0\\
0 & 0 & -\frac{1}{3}
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
1 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
1 & 0 & 0
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 1 & 0\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
-\frac{1}{3} & 0 & 0\\
0 & \frac{2}{3} & 0\\
0 & 0 & -\frac{1}{3}
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 0\\
0 & 1 & 0
\end{array}\right)\\
\left(\begin{array}{ccc}
0 & 0 & 1\\
0 & 0 & 0\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
0 & 0 & 0\\
0 & 0 & 1\\
0 & 0 & 0
\end{array}\right) &
\left(\begin{array}{ccc}
-\frac{1}{3} & 0 & 0\\
0 & -\frac{1}{3} & 0\\
0 & 0 & \frac{2}{3}
\end{array}\right)\\
\end{array}\right).
$$
Indeed, it is easy to check that such matrix encodes precisely the contribution
$$
\left( \tr_{\alpha_p \beta_p}\circ \mu - \frac{1}{3} (\tr_\alpha \circ \mu )(\tr_\beta \circ \mu ) \right),
$$
and the proof is complete.
\end{proof}
\begin{lem} \label{lem:tunnel-contribution}
The contribution along any tunnel equals
$$
\epsilon(p,\alpha,\beta) \left( \tr_{\alpha_p \beta_p}\circ \mu - \frac{1}{3} (\tr_\alpha \circ \mu )(\tr_\beta \circ \mu ) \right),
$$
where $p$ is the point of intersection (if any) of $\alpha$ and $\beta$ in the tunnel.
\end{lem}
\begin{proof}
Let $(w_{\alpha}, w_{\beta})$ be a tunnel $\tau$ through $(t_0,\dots,t_q)$. Lemma~\ref{lem:two-galleries-contribution} applies for $q=1$. Hence assume $q>1$. For all $0<k<q$, $t_k$ is a transition triangle whose type depends on the relative orientation of $\alpha$ and $\beta$ along the tunnel. In particular, if $t_0$ is a starting triangle of type $1, 1S$ (resp. $2, 4S$), then only transition triangles of type $5, 7$ (resp. $6, 6S$) are allowed.
Let $C_k(\tau)$ and $\mathbf{M}_k(\tau)$ be the contribution and the contribution matrix of the $k^{th}$ gallery. Once can apply Lemma~\ref{lem:sum-of-contributions} to check that $\mathbf{M}_{(0,1)}(\tau) = \mathbf{M}_0(\tau)$ for all allowed type combinations of $t_0$ and $t_1$. It follows by induction that $\mathbf{M}_{(0,\dots,k)}(\tau) = \mathbf{M}_0(\tau)$ for all $k<q$, where $\mathbf{M}_{(0,\dots,k)}(\tau) \in M_{3\times 3} (M_{3\times 3}(\mathbb{R}))$ is the matrix contribution of the first $k$ galleries, namely the matrix for which
$$
\sum_{i = 0}^{k}C_i(\tau) = \frac{2}{9}\sum_{l,m} \tr(\mathbf{E}_{lm} \cdot \mathbf{S}(w^\alpha_{2k+1})) \cdot \tr( ({\mathbf{M}_{(0,\dots,k)}(\tau)})_{lm} \cdot \mathbf{S}(w^\beta_{2k+1}) ),
$$
In conclusion $\mathbf{M}_{(0,\dots,q-1)}(\tau) = \mathbf{M}_0(\tau)$ and we reduced to the case of length two tunnels, for which Lemma~\ref{lem:two-galleries-contribution} applies.
\end{proof}
\subsubsection{Proof part 4: parallel curves and closed tunnels} \label{subsubsec:proof-part4}
When $\alpha$ and $\beta$ are parallel, that is $[\alpha] = [\beta^n]$ for some $n \in \ZZ \setminus \{0\}$, a slightly different kind of tunnel appears.\\
We will say that $(w_{\alpha},w_{\beta})$ is a \emph{closed tunnel through} $(t_0,\dots,t_q)$ if $t_j$ is a transition triangle with respect to the pairs $(w^\alpha_i,w^\beta_i)_{2j \leq i \leq 2j+2}$ for all $j \in \{ 0,\dots, q\}$, and $w^\gamma_0 = w^\gamma_{2k+2}$ for $\gamma \in \{\alpha,\beta\}$.
\begin{lem} \label{lem_closed_tunnel_contribution}
The contribution along a closed tunnel is $0$.
\end{lem}
\begin{proof}
Let $(w_{\alpha},w_{\beta})$ be a closed tunnel $\tau$ through $(t_0,\dots,t_q)$. We will show that $(w_{\alpha},w_{\beta})$ can be modified to an ``open" tunnel with contribution $0$. More precisely, we claim that the contribution along $\tau$ equals the contribution along the ``open" tunnel $\widetilde{\tau} = (\widetilde{w_{\alpha}},\widetilde{w_{\beta}})$ through $(t_0,\dots,t_q,t_{q+1})$ defined by:
\begin{itemize}
\item $t_{q+1} := t_0$;
\item $\widetilde{w_{\alpha}} := ( w_0^{\alpha} w_1^{\alpha} w_2^{\alpha} w_3^{\alpha} \dots w_{2q+2}^{\alpha} w_1^{\alpha} w_2^{\alpha} )$;
\item $\widetilde{w_{\beta}} := ( \widetilde{w_0} \widetilde{w_1} w_2^{\beta} w_3^{\beta} \dots w_{2q+2}^{\beta} \widetilde{w_2} \widetilde{w_3} )$ where $\widetilde{w_0}, \widetilde{w_1}, \widetilde{w_2}, \widetilde{w_3}$ are as in Figure~\ref{fig:closed-tunnel}.
\end{itemize}
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{closed_tunnel.png}
\caption{Local modification of $\beta$ in a closed tunnel.}
\label{fig:closed-tunnel}
\end{figure}
It is clear that $C_k(\tau) = C_k(\widetilde{\tau})$ for all $1 \leq k \leq q$, as we have only modified $\beta$ locally at $t_0$. Nevertheless, the contribution of the $0^{th}$ gallery in $(w_{\alpha},w_{\beta})$ is
$$
C_0(\tau) = \sum_{0 \leq i,j \leq 2} \left( \sum_{q_1,q_2 \subset t_0 } \epsilon(q_1,q_2) q_1^* q_2^* \left(\tr (\partial_1 w^\alpha_i) \cdot \tr (\partial_2 w^\beta_j)\right)\right).
$$
In such a gallery, we replace
\begin{align*}
\tr (\partial_i w^\beta_1) & = \tr( S^\beta_0 \cdots \partial_i W^\beta_{1} \cdots S^\beta_n )\\
& = \tr( S^\beta_0 \cdots \partial_i \left((W^\beta_{1})^{-1} \widetilde{W} \ \widetilde{W}^{-1} (W^\beta_{1})^{-1}\right) \cdots S^\beta_n ) \\
& = \tr (\partial_i \widetilde{w_0}) + \tr (\partial_i \widetilde{w_1} ) + \tr (\partial_i \widetilde{w_2} ) + \tr (\partial_i \widetilde{w_3} ),
\end{align*}
where $\widetilde{W}$ is the matrix associated to the $\underline{E}$--edge corresponding to $\widetilde{w_0}$ (and $\widetilde{w_3}$). It follows that $C_0(\tau) = C_0(\widetilde{\tau}) + C_{q+1}(\widetilde{\tau})$ and the claim is proved.
Finally, since $(\widetilde{w_{\alpha}},\widetilde{w_{\beta}})$ is a tunnel with no intersection points, the conclusion is implied by Lemma~\ref{lem:tunnel-contribution}.
\end{proof}
\subsubsection{Proof part 5: conclusion} \label{subsubsec:proof-part5}
Now it is time to put all the pieces together and prove Theorem~\ref{thm:compatibility-poisson-bracket}. Recall that our goal is to show that
$$
\sum_{p \in \alpha \cap \beta} \epsilon(p,\alpha,\beta) \left( \tr_{\alpha_p \beta_p}\circ \mu - \frac{1}{3} (\tr_\alpha \circ \mu )(\tr_\beta \circ \mu ) \right) = \sum_{i,j} \sum_{q_1,q_2 \in \triangle \cup \underline{E} } \epsilon(q_1,q_2) q_1^* q_2^* \left(\tr (\partial_1 s^\alpha_i) \cdot \tr (\partial_2 s^\beta_j)\right).
$$
Firstly, assume $\alpha$ and $\beta$ are not parallel. If $(s^\alpha_i,s^\beta_j)$ do not cross a common triangle, then we can discard such pair from the right hand sum as it contributes nothing. Otherwise, $(s^\alpha_i,s^\beta_j)$ is in the gallery of a unique tunnel. By Lemma~\ref{lem:tunnel-contribution}, if $\alpha$ and $\beta$ are not intersecting along such tunnel, then the sum of the contributions of all pairs in the tunnel is zero and we can discard them too. Hence we are left only with pairs which are contained in a tunnel where $\alpha$ and $\beta$ do intersect. But since each intersection point is contained in a unique tunnel, we can apply Lemma~\ref{lem:tunnel-contribution} again to conclude the proof.
When $\alpha$ and $\beta$ are parallel, the only difference from the previous argument is the possible presence of closed tunnels. However Lemma~\ref{lem_closed_tunnel_contribution} ensures that there is no contribution from those tunnels so we may discard them and follow the above construction verbatim. This concludes the proof of Theorem~\ref{thm:compatibility-poisson-bracket}.
\section{Examples} \label{sec:Examples}
In this section we provide examples of the properties of $\mathcal{T}_3^+(S_{g,n})$ discussed throughout this paper. As usual $S_{g,n}$ is the oriented surface of genus $g$ with $n$ punctures and we fix an ideal triangulation of $S_{g,n}$ whose set of oriented edges is $\underline{E}$ and whose set of triangles is $\triangle$.
\subsection{The once-punctured torus}
Let $S=S_{1,1}$ as depicted by the polygonal map as in Figure~\ref{fig:fund_dom_s11}. The edge pairings are indicated by common edge invariants.
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{Fund_Dom_S11.png}
\caption{A polygonal map for $S_{1,1}$.}
\label{fig:fund_dom_s11}
\end{figure}
Fock and Goncharov's Theorem~\ref{thm:global-coord} shows that the assignment of triple ratios and edge ratios indicated in Figure~\ref{fig:fund_dom_s11} uniquely determines an element of $\mathcal{T}_3^+(S)$. Moreover we can read off information about the corresponding holonomy group $\Gamma$ with knowledge of the edge ratios and triple ratios alone using the method of \S\ref{subsubsec:computing-monodromy}. We denote by $\alpha, \beta \in \Gamma$ the transformations corresponding to the face pairings between the vertical and horizontal edges respectively. It follows that $\Gamma \cong \langle \alpha, \beta \rangle$. The method of \S\ref{subsubsec:computing-monodromy} determines that, up to a common conjugation
\begin{align*}
& \alpha = T(t_{143}) E(e_{13}, e_{31}) T(t_{132}) E(e_{14}, e_{41}) T(t_{143}) \\
& \beta = E(e_{21}, e_{12}) T(t_{132}) E(e_{31}, e_{13}) T(t_{143})^{-1} \\
\text{where} \hspace{0.5cm} & T(z) := \frac{1}{\sqrt[3]{z}}
\left(
\begin{array}{ccc}
0 & 0 & 1 \\
0 & -1 & -1\\
z & z+1 & 1
\end{array}
\right), \hspace{2cm}
E(x,y) := \sqrt[3]{\frac{x}{y}}
\left(
\begin{array}{ccc}
0 & 0 & y \\
0 & -1 & 0\\
\frac{1}{x} & 0 & 0
\end{array}
\right)
\end{align*}
A direct calculation of the generator $\alpha \beta \alpha^{-1} \beta^{-1}$ of the peripheral subgroup of $\hol(S_{1,1})$ yields the requisite information for further inquiry into the geometry of $\mathcal{T}_3^+(S_{1,1})$.
\[
\alpha \beta \alpha^{-1} \beta^{-1} = \frac{1}{ t_{132} t_{143} }
\begin{bmatrix} (e_{12} e_{13} e_{21} e_{31} e_{41} e_{14})^{-1} & 0 & 0 \\
* & 1 & 0 \\
* & * & e_{12} e_{13} e_{21} e_{31} e_{41} e_{14} (t_{132} t_{143})^3
\end{bmatrix}
\]
This is the normal form we expect to arise following the construction in \S\ref{subsubsec:computing-monodromy}. Recall from \S\ref{subsec:fin_vol}, or by inspection of the above matrix, that the following monomials determine whether the end of $S_{1,1}$ is hyperbolic, special or cuspidal.
\begin{align*}
\mathcal{X}_1 :=& e_{12}e_{21}e_{14}e_{41}e_{13}e_{31} \\
\mathcal{Y}_1 :=& (t_{132}t_{143})^3 e_{12}e_{21}e_{14}e_{41}e_{13}e_{31}
\end{align*}
These monomials are our primary tools in identifying interesting subvarieties of $\mathcal{T}_3^+(S_{1,1})$. Let us begin by identifying which coordinates give rise to structures which have minimal and maximal ends. This is a simple calculation following the classification in \S\ref{subsec:fin_vol}. The region in which coordinates give rise to structures on $S_{1,1}$ with a maximal end are:
\begin{align*}
&\mathcal{X}_1 < \mathcal{Y}_1\mathcal{X}_1 < 1 \quad \text{or} \\
&\mathcal{X}_1 > \mathcal{Y}_1\mathcal{X}_1 > 1
\end{align*}
Similarly, those structures with minimal ends are given by:
\begin{align*}
\mathcal{Y}_1 &< 1 \quad \text{and} \quad \mathcal{X}_1 < 1 \quad \text{or} \\
\mathcal{Y}_1\mathcal{X}_1 &< 1 \quad \text{and} \quad \mathcal{X}_1 > 1 \quad \text{or} \\
\mathcal{Y}_1 &< 1 \quad \text{and} \quad \mathcal{X}_1 > 1 \quad \text{or} \\
\mathcal{Y}_1\mathcal{X}_1 &> 1 \quad \text{and} \quad \mathcal{X}_1 < 1
\end{align*}
We see that generically (on structures with hyperbolic ends) there is a natural $2:1$ correspondence between structures with minimal ends and those with maximal ends, as noted in the discussion of the $\Sym(3)$-action on $\mathcal{T}_3^+(S_{1,1})$ in \S\ref{subsubsec:sym_gp}.
We will now identify the symplectic leaves of $\mathcal{T}_3^+(S_{1,1})$. Following Goldman~\cite{Goldman-symplectic-1984}, or by a brief calculation, we find that $\mathcal{X}_1$ and $\mathcal{Y}_1$ are Casimirs for the Poisson structure on $\mathcal{T}_3^+(S_{1,1})$ featured in \S\ref{sec:poisson-structure}. Note that Casimirs are constant on symplectic leaves. The condition of $\mathcal{X}_1$ and $\mathcal{Y}_1$ being constant simplifies to
\[
t_{132} t_{143} = c_0 \quad \text{ and } \quad e_{12} e_{21} e_{14} e_{41} e_{13} e_{31}= c_1 \label{level_set}
\]
for any positive real constants $c_0$ and $c_1$. In particular the symplectic leaves of $\mathcal{T}_3^+(S_{1,1})$ are, at largest, open cells of $6$ real dimensions defined by the conditions~\ref{level_set}. Denote a generic level set of equations~\ref{level_set} by $\mathcal{A}$. Suppose there exists function $f \in C^{\infty}(A)$ such that
\[
\{ f, \cdot \} : C^{\infty}(\mathcal{A}) \rightarrow C^{\infty}(\mathcal{A}) \label{Poisson_contraction}
\]
is the zero function. A coordinate system on $\mathcal{A}$ is determined by the functions
\[
\mathcal{P} = \{ t^*_{143}, e^*_{12}, e^*_{21}, e^*_{14}, e^*_{41}, e^*_{13} \} \subset \mathcal{C}^{\infty}(\mathcal{A})
\]
where $q^*$ for $q \in \triangle \cup \underline{E}$ are \emph{coordinate functions} on $\mathcal{T}_3^+(S_{1,1})$ defined in \S\ref{subsec:FG-poisson-structure}. Computing $\{ f, q^* \}$ for $ q^* \in \mathcal{P}$ shows that $\partial f / \partial q^* = 0$ for all such $q*$. In particular $f$ is constant on $\mathcal{A}$. Therefore, $\{ \cdot, \cdot \}$ is non-degenerate on $\mathcal{A}$ in the sense that the function \ref{Poisson_contraction} is uniformly the zero function if and only if $f$ is constant. It follows that $\mathcal{A}$ is a symplectic leaf.
Increasing once more in specificity, recall from \S\ref{subsec:fin_vol} that the finite-area structures on $S_{1,1}$ arise when $\mathcal{X}_1= \mathcal{Y}_1 =1$. Therefore the finite-area conditions for $S_{1,1}$ reduce to the particularly simple condition that $c_0=c_1=1$ in equations~\ref{level_set}.
To determine $\Teich(S_{1,1}) \subset \mathcal{T}_3^+(S_{1,1})$ recall that the conditions for FG coordinates to represent a finite-area hyperbolic structure described in \S\ref{subsec:classical-teich} are:
\begin{align*}
&\phi(x)(e) = \phi(x)(e^{-1}), \quad \forall \; e \in \underline{E} \quad \text{and } \quad \phi(x)(t) = 1, \quad \forall \; t \in \triangle; \\
&\prod_{k=1}^{m_i} \phi(e_{ik}) = 1, \qquad \forall \; i =1,\dots,n.
\end{align*}
In the present example this simplifies to
\[
t_{143} = t_{132} = 1, \qquad e_{ij} = e_{ji} \quad \forall i,j \quad \text{and} \qquad e_{21}e_{41}e_{31}=1
\]
So $\Teich(S_{1,1}) \cong \mathbb{R}^{2}$ as we should expect from the classical theory.
\subsection{The thrice-punctured sphere}
We use the same setup as above with $S=S_{0,3}$. Just as in the previous example Figure~\ref{fig:fund_dom_s03} denotes a topological $S_{0,3}$ with edge pairings indicated by the common edge ratios. Moreover the edge ratios and triple ratios uniquely determine an element of $\mathcal{T}_3^+(S_{0,3})$.
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{Fund_Dom_S03.png}
\caption{A polygonal map for $S_{0,3}$.}
\label{fig:fund_dom_s03}
\end{figure}
The surface $S_{0,3}$ has holonomy group isomorphic to the free group on two generators. Once again the holonomy group is generated by transformations corresponding to the edge pairings $\gamma$ and $\delta$ which glue the two edges on the left and the two edges on the right respecitvely. Their $SL(3, \mathbb{R})$-representatives are given below, up to a common conjugation, using the method of \S\ref{subsubsec:computing-monodromy}:
\[
\gamma = E(e_{21}, e_{12}) T(t_{142}) = \begin{bmatrix}
e_{12}t_{142} & e_{12}(t_{132}+1) & e_{12} \\
0 & 1 & 1\\
0 & 0 & e_{21}^{-1}
\end{bmatrix}
\]
\[
\delta = T(t_{142}) E(e_{24}, e_{42}) T(t_{243}) E(e_{32}, e_{23}) T(t_{243}) E( e_{42}, e_{24} ) T(t_{142})^{-1}
\]
We have chosen a conjugacy class of the holonomy group so that $\gamma$ is in a convenient form. However as there are three distinct peripheral subgroups of $\pi_1(S_{0,3})$, the other generators of these subgroups, $\delta$ and $\gamma \cdot \delta$ are not written as succinctly. Their eigenvalues are still easily calculable however for more complicated examples it is more convenient to make use of Theorem~\ref{thm:monodromy} to infer that the following monomials govern the boundary geometry of $S_{0,3}$:
\begin{align*}
\mathcal{X}_1 :=& e_{21} \qquad \text{and} \qquad \mathcal{Y}_1 := e_{12} t_{142};\\
\mathcal{X}_2 :=& e_{32}e_{42}e_{12}e_{24} \qquad \text{and} \qquad \mathcal{Y}_2 := e_{23}t_{243}^2 e_{24}t_{142}^2 e_{21}e_{42};\\
\mathcal{X}_3 :=& e_{23} \qquad \text{and} \qquad \mathcal{Y}_3 := e_{32}t_{243}
\end{align*}
Therefore there are at least six independent Casimirs described by the equations $\mathcal{X}_i = c_i$ and $\mathcal{Y}_i = d_i$ for positive real constants $c_i$ and $d_i$, $i= 1, 2, 3$. The equations defining the level sets of these Casimirs simplify somewhat to:
\begin{align*}
e_{21} =& c_1 \qquad \text{and} \qquad e_{12}t_{142} = d_1; \\
\frac{e_{42}e_{24}}{e_{32}e_{12} } =& c_2 \qquad \text{and} \qquad e_{32}e_{12} = \left( \frac{c_3 c_2 c_1 (d_1d_3)^2}{d_2} \right)^{\frac{1}{3}}; \\
e_{23}=& c_3 \qquad \text{and} \qquad e_{32}t_{243} = d_3
\end{align*}
This ensures that the maximal symplectic leaves are at most $2$-dimensional. Symplectic manifolds have even dimension so if $\{ \cdot, \cdot \}$ is not uniformly zero then the subvarieties defined by level sets of the six Casimirs above must determine symplectic leaves of the Poisson structure on $\mathbb{R}^{\triangle \cup \underline{E}}$.
A direct calculation shows that $\{ t^*_{142}, e^*_{24} \} \neq 0$ where the arguments of the Poisson bracket are two \emph{coordinate functions} defined in \S\ref{subsec:FG-poisson-structure}. In particular, the maximal symplectic leaves of $\mathcal{T}_3^+(S_{0,3})$ are $2$-dimensional whereas the symplectic leaves of $\mathcal{T}_3^+(S_{1,1})$ are $6$-dimensional. This allows us to distinguish between $\mathcal{T}_3^+(S_{1,1})$ and $\mathcal{T}_3^+(S_{0,3})$ using their intrinsic Poisson geometry despite the fact that they are homeomorphic.
The structures on $S_{0,3}$ with finite area are described by $c_i = d_i = 1$ for $i=1,2,3$ so $\mathcal{T}_3^{f}(S_{0,3})$ is defined by the conditions
\begin{align*}
e_{21}=e_{23}=1; \\
e_{42}e_{24}=e_{32}e_{12}=1; \\
e_{12}t_{142} = e_{32}t_{243} = 1
\end{align*}
Therefore $\mathcal{T}_3^{f}(S_{0,3}) \cong \mathbb{R}^3$. Imposing the finite-area conditions and the criteria presented in \S\ref{subsec:classical-teich} shows that the classical Teichm\"uller space is identified with:
\[
e_{ij} = 1 \quad \forall \ i, j \qquad \text{and} \qquad t_{142} = t_{243} = 1
\]
so $\Teich(S_{0,3})$ is a single point as we should expect from the classical theory.
\subsection{The twice-punctured torus}
The surface $S = S_{1,2}$ is denoted in Figure~\ref{fig:fund_dom_s12}.
\begin{figure}[h]
\centering
\includegraphics[height=5cm]{Fund_Dom_S12.png}
\caption{A polygonal map for $S_{1,2}$.}
\label{fig:fund_dom_s12}
\end{figure}
The four Casimirs arising from the peripheral elements of $\pi_1(S_{1,2})$ are determined by the monomials $\mathcal{X}_i$ and $\mathcal{Y}_i$ for $i=1,2$ as defined in \S\ref{subsec:fin_vol}. They are given explicitly as follows:
\begin{align*}
\mathcal{X}_1 :=& e_{21} e_{61}e_{53} e_{23} e_{25} e_{35} \qquad \text{and} \qquad \mathcal{Y}_1 := e_{12} t_{162} e_{16} t_{354}^2 e_{35} t_{253}^2 e_{32} t_{265} e_{52}e_{53} \\
\mathcal{X}_2 :=& e_{12} e_{62} e_{52} e_{32} e_{26} e_{16} \qquad \text{and} \qquad \mathcal{Y}_2 := e_{21} t_{162}^2 e_{26} t_{265}^2 e_{25} t_{253} e_{23} e_{62} e_{61} t_{354}
\end{align*}
The boundary geometry of $S_{1,2}$ is determined by the monomials $\mathcal{X}_i = c_i$ and $\mathcal{Y}_i = d_i$ for positive real constants $c_i$ and $d_i$ for $i=1,2$.
Setting $c_i = d_i = 1$ for $i=1,2$ we find that $\mathcal{T}_3^{f}(S_{1,2})$ is defined by
\begin{align*}
e_{21} e_{61}e_{53} e_{23} e_{25} e_{35}=1 \qquad \text{and} \qquad t_{162}t_{354}^2 e_{35} t_{253}^2 t_{265} e_{53} = e_{62}e_{26} \\
e_{12} e_{62} e_{52} e_{32} e_{26} e_{16}=1 \qquad \text{and} \qquad t_{162}^2e_{26}t_{265}^2t_{253}e_{62}t_{354}=e_{53}e_{35}
\end{align*}
It follows that $\mathcal{T}_3^{f}(S_{1,2}) \cong \mathbb{R}^{12}.$
Imposing both the conditions for finite-area and hyperbolicity as seen in \S\ref{subsec:classical-teich} we identify $\Teich(S_{1,2})$ with the subvariety defined by:
\begin{align*}
& e_{ij} = e_{ji} \quad \forall \ i \text{ and } j \qquad \text{and} \qquad t_{ijk} = 1 \quad \forall \ i,j \text{ and } k\\
& e_{53}^2 e_{21} e_{61} e_{23} e_{25} = 1 \qquad \text{and} \qquad e_{62} = e_{53}
\end{align*}
Therefore $\Teich(S_{1,2}) \cong \mathbb{R}^4$ in accordance with the classical theory.
\subsection{The closed genus two surface}
Now we will consider $S=S_{2,0}$ and use the method of \S\ref{subsec:closed_surfaces} to construct $\mathcal{T}_3(S_{2,0})$ from $\mathcal{T}_3^+(S_{1,2})$ by gluing boundary components.
In order to glue the boundary components of $S_{1,2}$ in the manner of Goldman~\cite{Goldman-convex-1990} we must ensure that both ends of $S_{1,2}$ are minimal. Using the notation from the previous example those conditions are, for $i=1$ and $i=2$:
\begin{align}
&\mathcal{Y}_i >1 \quad \text{and} \quad \mathcal{X}_i >1 \quad \text{or} \label{minimality1_s12} \\
&\mathcal{X}_i <1 \quad \text{and} \quad \mathcal{Y}_i <1 \quad \text{or} \\
&\mathcal{X}_i\mathcal{Y}_i <1 \quad \text{and} \quad \mathcal{X}_i > 1 \quad \text{or} \\
&\mathcal{X}_i\mathcal{Y}_i >1 \quad \text{and} \quad \mathcal{X}_i < 1 \label{minimality4_s12}
\end{align}
Note that $\mathcal{T}_3(S_{2,0})$ does not contain information regarding framing so we will `forget' the framing information contained in $\mathcal{T}_3^+(S_{1,2})$. This is tantamount to restricting our attention to exactly one of the regions defined by \ref{minimality1_s12}--\ref{minimality4_s12} for each of $i=1$ and $i=2$. We assume without loss of generality that equation \ref{minimality1_s12} holds for $i=1$ and $i=2$.
The second requirement in performing convex gluing is that the peripheral transformations at the ends of $S_{1,2}$ be conjugate. Therefore, following the classification of peripheral elements and their eigenvalues in \S\ref{subsubsec:max-min-coordinates} we assume that
\[
\mathcal{X}_1\mathcal{Y}_1^2 = \mathcal{X}_2\mathcal{Y}_2^2 \qquad \text{and}\qquad \mathcal{X}_1^2\mathcal{Y}_1 =\mathcal{X}_2^2 \mathcal{Y}_2
\]
Given that all variables must be positive and real we have one pair of solutions of the form $\mathcal{X}_1 = \mathcal{X}_2$ and $\mathcal{Y}_1 = \mathcal{Y}_2$. The solutions $\mathcal{X}_1=\mathcal{Y}_1$ and $\mathcal{X}_2 = \mathcal{Y}_2$ are disregarded in order to ensure that the convex gluing induces a coherent orientation on $S_{2,0}$. We are left with the following conditions which parameterise convex projective structures on $S_{1,2}$ whose ends are `gluable'.
\begin{align}
e_{21} e_{61} e_{53} e_{23} e_{25} e_{35} = e_{12} e_{62} e_{52} e_{32} e_{26} e_{16} > 1 \label{gluable1_s12} \\
(e_{53} e_{35})^2 t_{354} t_{253}= (e_{26} e_{62})^2 t_{162} t_{265} > 1 \label{gluable2_s12}
\end{align}
It remains to show that these equations define a topological cell of real dimension $16$ thus verifying Goldman's~\cite{Goldman-convex-1990} result. Following the method used in Theorem~\ref{thm:Marquis-main}, it is sufficient to note that the equations~\ref{gluable1_s12}-\ref{gluable2_s12} are independent. Recall that as long as the peripheral elements of $S_{1,2}$ are hyperbolic, there is a $2$-real parameter space of potential gluings for each pair of `gluable' ends. For an explanation of this see \S\ref{subsec:closed_surfaces} or the original proof in Goldman~\cite{Goldman-convex-1990}. Therefore $\mathcal{T}_3(S_{2,0})$ is a trivial $2$-dimensional vector bundle over the subvariety of $\mathcal{T}_3^+(S_{1,2})$ defined by the conditions~\ref{gluable1_s12}-\ref{gluable2_s12}. Ultimately this shows that $\mathcal{T}_3(S_{2,0}) \cong \mathbb{R}^{16}$, reproducing Goldman's result.
Once again we can use this to calculate a subvariety of $\mathcal{T}_3(S_{2,0})$ which is homeomorphic to $\Teich(S_{2,0})$. First note that any hyperbolic structure on $S_{2,1}$ must restrict to a hyperbolic structure on $S_{1,2}$. Therefore we impose the hyperbolicity and finite-area conditions on $\mathcal{T}_3^+(S_{1,2})$ defined in \S\ref{subsec:classical-teich} to obtain:
\[
t_{ijk} =1 \quad \forall \ i,j \ \text{and } k \qquad \text{and} \qquad e_{ij}=e_{ji} \quad \forall \ i \ \text{and } j
\]
Further imposing these conditions on the `gluable' equations we are left with:
\begin{align}
& e_{35} = e_{62} \label{fin_hyp1_S20} \\
& t_{ijk} =1 \quad \forall \ i,j \ \text{and } k; \label{fin_hyp2_S20} \\
& e_{ij}=e_{ji} \quad \forall \ i \ \text{and } j \label{fin_hyp3_S20}
\end{align}
To recover $\Teich(S_{2,0})$ note that instead of a $2$ parameter family of convex gluings we may only glue the ends of $S_{1,2}$ by hyperbolic isometries if we wish to give rise to a hyperbolic structure on $S_{2,0}$. Therefore, $\Teich(S_{2,0})$ may be realised as a trivial real $1$-dimensional vector bundle over the subvariety defined by the conditions~\ref{fin_hyp1_S20}-\ref{fin_hyp3_S20}. In particular $\Teich(S_{2,0}) \cong \mathbb{R}^{6}$ which accords with the classical theory.
|
\section{Introduction}
Silica aerogels perform a vital role in nuclear and particle physics
experiments as radiators for \u{C}erenkov detectors. These aerogels cover a
range of refractive index from about 1.01 to 1.08, and can be both highly
transparent and hydrophobic. The often-large volume of aerogel required for
experiments can be very costly, and commercial alternatives to products such as
those produced by Matsushita Electric Works could be beneficial. The work
presented in this paper is part of an effort to test the performance of new
hydrophobic aerogels produced by Aspen Aerogels in prototype \u{C}erenkov
detectors, and answers the call for proposals found in the U.S. DOE Small
Business Innovation Research (SBIR) Funding Opportunity Annoucements (FOAs)
(see for example page 103 of Phase I Release 1 found in \cite{SBIRFOAs}). The
SBIR FOAs call for cost-effective \u{C}erenkov radiators with indices between
those of gases and liquids, e.g., aerogels. The ultimate goal is the
qualification of commercially produced, cost-efficient, and high-quality
aerogel for use in the physics community by benchmarking the performance of the
aerogel in both threshold and ring-imaging \u{C}erenkov detectors. In this
paper, we report on aerogel production and testing performed using a
threshold-type counter.
A number of aerogel tiles have been produced by Aspen Aerogels, and a small
selection have been used for testing in the research presented here. A summary
of the techniques used in the production of the aerogels is given in section~\ref{SectionAerogel} along with a description of the technique used for
measuring the refractive index of the tiles. In this section, we also present
the distribution of refractive index measurements taken.
The performance testing was achieved by measuring the performance of a
prototype \u{C}erenkov threshold counter using the new aerogels.
Quantification of the performance is given in terms of both photoelectron yield
from the photomultiplier tubes (PMTs) that view the counter, as well as
calculated figures of merit. This approach is an alternative to measuring
individual optical properties of the aerogel, and instead aims to provide a
benchmark in the context of \u{C}erenkov particle detection. The prototype
counter was evaluated using an electron test beam at DESY, Hamburg. Details of
the experimental setup and test beam can be found in section~\ref{SectionSetup}.
Analysis of the experimental data is discussed in section~\ref{SectionAnalysis}, where a variation on a common and robust method of
photoelectron quantification is presented. The analysis focuses on
determination of the relative inefficiency of detecting electrons in order to
extrapolate to average photoelectron yield using the application of Poisson
statistics. This avoids difficulties often encountered when complex PMT
responses are fit. Another desireable feature of this approach is that it
provides a means of measuring systematic error in the determination of
photoelectron yields.
In section~\ref{SectionFoM} we describe a figure of merit (FoM) intended to
quantify performance in a way that is comparable to different geometries. This
FoM is derived from, and nearly identical to, that described by D.W.
Higinbotham \cite{Higinbotham1998332}, and as such provides a direct comparison
to previous experimental results analyzed in \cite{Higinbotham1998332}. The
aerogel configurations tested, with tile sizes up to 20~cm by 20~cm, are
described in section~\ref{SectionResults} along with photoelectron yields and
figures of merit. In addition, dependence of the photoelectron yields on
aerogel depth and incident beam position are presented and discussed.
\section{Silica aerogels}
\label{SectionAerogel}
\subsection{Sol-gel chemistry}
Fabrication of aerogels generally involves two major steps: formation of a wet
gel, and drying of the wet gel to form an aerogel. The vast majority of silica
aerogels prepared today utilize silicon alkoxide precursors; this route avoids
the formation of undesirable salt by-products and allows more control over the
final product. The most common alkoxide precursors are tetramethylorthosilicate
(TMOS) and tetraethylorthosilicate (TEOS), however, many others can be used to
impart different properties to the gel. The reaction is typically performed in
ethanol, with the aerogel density dependent on the concentration of alkoxide.
Hexamethyldisilazane (HMDZ) is often used during aging of TEOS or TMOS gels to
make the silica surface hydrophobic. However, this treatment takes time, and the
transparency of the aerogel is affected by the amount of HMDZ. In this program,
a methylsiliconate precursor was used as the co-precursor with TEOS, which
showed great results in terms of aerogel hydrophobicity and transparency.
\subsection{Aging process}
It is often assumed that the hydrolysis and condensation reactions of the
silicon alkoxides are complete when a sol reaches the gel point, however the
silica backbone of the gel still contains a significant number of unreacted
alkoxide groups. In fact, hydrolysis and condensation reactions continue after
several days beyond the gelation point. Failure to realize and accommodate this
is a common mistake preparing silica aerogel monoliths, as weak aerogels will
result.
Strengthening the gel is a result of the aging process, and is enhanced by
controlling the pH and water content of the aging bath. Common aging procedures
for base catalyzed gels involve soaking the gel in an alcohol/ammonia mixture,
at a pH of 8-9 for up to 48 hours at 60 \textdegree C. The time required for
this process depends on the gel thickness. Any water left in the gel will not
be removed by supercritical drying, and will lead to an opaque, white, and very
dense aerogel.
\subsection{Supercritical drying}
The final process step for silica aerogels is supercritical drying. This is
where the liquid within the gel is removed, at supercritical conditions, leaving
only the linked silica network \cite{doi:10.1021/ja01250a034},
\cite{supercritfluidextract}. The process can be performed by venting the
ethanol above its critical point or by prior solvent exchange with CO$_2$
followed by supercritical CO2 venting \cite{tewari1986hydrolysis},
\cite{Tewari1985363}. The gels are put in a high pressure vessel with liquid
CO$_2$ and solvent exchange occurs between ethanol and CO$_2$. Once the ethanol
is extracted, the gels are brought to the critical point of CO$_2$ (31.06
\textdegree C and 1050 psi) and CO$_2$ is slowly vented.
Once the pressure reaches ambient, the vessel is opened and crack-free aerogel
monoliths are obtained. In this program, a 60 liter vessel was used to fabricate
the aerogel tiles for optical characterization. Batches of different refractive
index were fabricated for \u{C}erenkov detection performance evaluation at
Arizona State University.
\subsection{Refractive index}
\begin{figure}[h] \begin{center}
\includegraphics[width=4in]{figures/IoRMeasurement.pdf}
\end{center}
\caption{A diagram representing the apparatus used to
measure refractive indices of the aerogel tiles - The method of measuring the
index of refraction for a tile of aerogel makes use of the simple prism effect.
This method is very precise for low indices such as those of aerogel.}
\label{fig:IoRDiagram}
\end{figure}
Figure~\ref{fig:IoRDiagram} shows the setup for measuring the index of
refraction for a tile of aerogel. The method is common (see
\cite{aThresholdCerenkov} and \cite{iormeasurement} for other accounts), and
easily yields precise measurements for low index materials like aerogel.
Refractive indices are calculated by measuring the minimum deflection of a laser
beam through the corner of an aerogel tile. The relevant relationship is given
by the common formula for a prism,
\begin{equation}
\label{eq:ior}
n=\frac{\sin{\frac{\theta+\delta}{2}}}{\sin{\frac{\theta}{2}}},
\end{equation}
where $n$ is the refractive index at the wavelength of the laser,
$\delta$ is the minimum angle of deflection, and $\theta$ is the prism angle
(in this case, about $\pi/2$). The minimum angle is found by measuring the
minimum deflection of the laser beam on a ruled screen.
The aerogel is pivoted about the intersection of the undeflected laser line, and
a line bisecting the corner of the tile. As the aerogel is pivoted back and
forth, a minimum deflection is easily found.
The laser used is at a wavelength of 532 nm, and lies comfortably in the
spectral range of bialkali photocathodes, such as those used by the
photomultiplier tubes in our \u{C}erenkov test counter.
\begin{figure}[h] \begin{center}
\includegraphics[width=4.5in]{figures/IoRHisto.pdf}
\end{center}
\caption{Histogram of the refractive index
measurements - For each tile, the refractive index was measured using the above
method on more than one corner of the tile. The histogrammed data represent
the range and distribution of indices produced for evaluation.}
\label{fig:IoRHisto}
\end{figure}
In Figure~\ref{fig:IoRHisto}, a histogram of all of the refractive index
measurements represents the range of indices produced for evaluation. In this
paper, measurements of only a portion of the tiles are presented.
\section{Experimental setup}
\label{SectionSetup}
\subsection{\u{C}erenkov Counter}
\label{SectionCounter}
\begin{figure}[h]
\begin{center}
\includegraphics[width=4in]{figures/DESYTestBeamCounterPic.png}
\end{center}
\caption{Counter Geometry - The solid filled regions in this image
represent aerogel (cyan), air surrounded by diffuse reflector (white),
aluminum light guides (grey), and PMT mock-ups (black).
Represented in wireframe, the aluminum housing (grey) and magnetic
shielding (light blue) are also shown. The red, green,
and blue axis lines
represent the $\hat{x}$, $\hat{y}$, and $\hat{z}$ axes respectively, where the
$\hat{z}$-axis is parallel to the beam.}
\label{FigurePrettySimulation}
\end{figure}
A diffusively reflective threshold counter was developed for testing the new
aerogels. It consists of an aluminum light box with dimensions 36.32 x 20.32 x
20 cm$^3$ in the $\hat{x}$, $\hat{y}$, and $\hat{z}$ directions respectively
(see Figure~\ref{FigurePrettySimulation} for axes), a 20.32 x 20.32 x 6.4 cm$^3$
box extruded upstream from the light box to house aerogel, and two 5'' Photonis
XP4500B photomultiplier tubes (PMTs) coupled to the sides of the light box. A
0.6 mm-thick aluminum window is seen by the beam before passing through the
aerogel. The light box and aerogel box are lined on the inside with a white,
diffusively reflective ``paper'', Millipore GVHP00010.
The PMTs are coupled to the light box with polished circular-cylindrical
aluminum light guides with an inside diameter of 13.6 cm. The PMTs use a UV
glass and a bi-alkali photocathode, and manufacturer's specifications for the
Photonis XP4500B state a peak quantum efficiency of 24\%, with a spectral range
of about 200 - 650 nm where the efficiency falls off below 10\% of peak.
The purpose of the aluminum light guides is to couple the PMTs at a distance so
that shielding for static magnetic fields can be effectively used. The test
counter is not tested in a significant magnetic field here, but the performance
of the counter itself in this configuration is of interest for future
applications.
In this setup, magnetic shields produced by MuShield are in place which reduce
ambient magnetic field at the photocathodes by a factor of 100.
\subsection{Data acquisition}
\label{SectionDAQ}
For readout, we use a 12-bit 250 MS/s digitizer (CAEN V1720). The PMTs coupled
with the Photonis VD105k bases produce a signal with a FWHM $\approx4$ ns. This
is too sharp of a pulse for the digitizer sample rate, so the PMT signals are
fed through passive 32 MHz low-pass LC filters.
When an external trigger is received, a number of digitized samples recorded in
the vicinity of the trigger are stored for offline analysis. These samples
include data prior to the trigger, so that a pedestal can be calculated for each
event.
\subsection{Test beam}
\label{SectionTestBeam}
The T22 e-/e+ test beam at the DESY facility in Hamburg was used to perform
tests of the aerogel/counter combination. The beam is generated by converting
bremsstrahlung radiation from carbon fiber targets in the DESY II lepton storage
ring. The leptons in this storage ring are bunched, and revolve once every
microsecond \cite{desytestbeam}.
For this setup, converted electrons at an energy of 2 GeV are selected, with an
energy spread $\approx5\%$.
The beam is shaped by a 20 x 20 mm$^2$ primary collimator, and then a 6 x 6
mm$^2$ secondary collimator. Located immediately downstream of the secondary
collimator is a broad scintillator (S1). Approximately 5 m further downstream
is the \u{C}erenkov test counter, and another 5 m downstream is a 2 x 2 cm$^2$
scintillator (S2). The test counter was mounted to an X-Y translation table
for relative beam positioning, and to measure position dependence. A
trigger signal is produced with coincidence between S1 and S2.
The average trigger rate is $\approx4$ kHz.
For simplicity, it's beneficial to argue that on each trigger, only one electron
passes through the counter. Given the trigger rate and the bunched nature of
the storage ring, we estimate that there are an average of 1.002 electrons
passing through the counter on triggered events. This estimate relies on the
assumption that all electrons passing through the counter produce a trigger.
Based on the observed narrowness of the beam, this is a reasonable assumption.
\begin{comment}
The remainder of subsection~\ref{SectionTestBeam} will be
an argument that this is essentially true.
Because the storage ring is bunched,
it can be concluded that the average number of particles that pass through the
counter \emph{on triggered events} is
\begin{equation}
\label{eq:prob}
\left<N_{e}\right>\approx\frac{\lambda}{1-P_0},
\end{equation}
where $\lambda$ is the average
number of triggerable particles \emph{per bunch revolution}, and $P_0$ is the
probability of the beam failing to yield a triggerable particle on a bunch revolution
(such is the case when no appropriately energetic electrons are produced via
bremsstrahlung conversion).
The above statement can be viewed as a consequence of the probabilistic
axiom \cite{probability}
\begin{equation}
\label{eq:axiom}
P(A|B)=\frac{P(A\cap B)}{P(B)}.
\end{equation}
If we say that the event A is the case of having $i$ triggerable electrons in a
bunch revolution, and event B is the case of having a minimum
of one, then Equation~\ref{eq:axiom} reads, ``The probability of
having $i$ electrons given that there is at least one is equal to the probability
of having both $i$ electrons \emph{and} at least one, divided by the probability
of having at least one.''
Given the applicability of Poisson statistics, $\left<N_{e}\right>$ can be
further approximated by using
\begin{equation}
\lambda=-\ln{P_0}.
\end{equation}
We then have $\left<N_{e}\right>$ purely in terms of known quantities, and we can
substitute with
\begin{equation}
P_0 = \frac{N_{rev}-N_{trig}}{N_{rev}}\approx0.996
\end{equation}
where $N_{rev}$ is the number or rate of bunch revolutions, and $N_{trig}$ is the
number or rate of triggers. This gives an $\left<N_{e}\right>$ value of 1.002,
which is a small enough deviation from unity to be ignored.
Equation~\ref{eq:prob} is an approximation mostly because it is based on the
assumption that all particles passing through the counter also produce a trigger
signal.
It is considered a good approximation based on the observed narrowness of the
collimated beam.
In hindsight, a scintillator paddle covering the entire effective area of the
counter could have been used to
eliminate this assumption.
\end{comment}
\section{Signal analysis}
\label{SectionAnalysis}
\subsection{Pulse binning}
\label{SectionPulseBinning}
\begin{figure}[h] \begin{center}
\includegraphics[width=4in]{figures/averagePulse.pdf}
\end{center}
\caption{PMT Response - An average digitized PMT
pulse is shown with the height inverted, so that the pulses discussed here are
positive. The region in blue is used to calculate the electronic pedestal for
each event. This allows long-term variations in pedestal to be ignored. The
region in green is integrated relative to the pedestal for a value proportional
to the PMT anode charge.}
\label{FigureWaveform}
\end{figure}
As described in section~\ref{SectionDAQ}, the PMT waveforms are digitized to be
analyzed offline. The timing of the signal is tuned so that 70 samples of
electronic pedestal are obtained before the desired PMT signal
(Figure~\ref{FigureWaveform}). The pedestal for each event and for each PMT is
therefore amply measured in order to produce a sharp zero-photoelectron peak,
on which our analysis ultimately relies. For each event, the latter portion of
the waveform in Figure~\ref{FigureWaveform} is integrated relative to the
calculated pedestal and binned into a pulse-height spectrum
(Figure~\ref{FigureSpectrum}).
\begin{figure}[h]
\begin{center}
\includegraphics[width=4in]{figures/SESAndGelSpec.pdf}
\end{center}
\caption{Pulse-Height Spectrum - The data shown are generated
by binning the measured PMT charge for each triggerable event. The
black line represents a measurement with aerogel in the counter, while
the blue line represents a measurement with only air in order to
identify a single photoelectron peak. The key features are the
zero-photoelectron (pedestal) peak on the left, and the single- and
multiple-photoelectron bumps to the right. The pedestal peak falls off
neatly to zero on the left in a Gaussian fashion, as can be seen in
Figure~\protect\ref{FigureDeconvFit}.}
\label{FigureSpectrum}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=4in]{figures/pedDevHisto.pdf}
\end{center}
\caption{Histogram of pedestal RMS values - The desirable
Gaussian-like peak of the distribution continues into a region where the tails
of thermal dark current pulses begin to dominate the RMS. This produces artifacts in the
pedestal calculation which assumes high-frequency stochastic fluctuations.
Therefore, a cut is placed on the pedestal RMS (the rejected events shown by the
grey region) where the distribution fails to be dominated by such fluctuations.
The calculated photoelectron yield has negligible sensitivity to the placement
of this cut in the range of 1.0 - 1.2.}
\label{FigurePedCut}
\end{figure}
In order to minimize the effects of dark current, a cut is placed on the
standard deviation of the pedestal samples to enforce the assumption of a
constant signal with a small random error. This cut, shown in
Figure~\ref{FigurePedCut}, is necessary for minimizing artifacts in the
zero-photoelectron peak due to dark current.
\subsection{Determination of average photoelectron counts}
The pulse-height spectrum shown in Figure~\ref{FigureSpectrum} has a sharp peak
at the left corresponding to the absence of photomultiplication, and its
integrated value divided by the total number of events is therefore considered
to be the probability that zero photoelectrons are produced in the PMT for a
triggerable event in a given counter configuration. The probability
distribution for the number of photoelectrons emitted from the PMT cathodes in a
triggerable event is well approximated by a Poisson, given the low
probability of a radiated photon producing a photoelectron. We therefore claim
that the average number of photoelectrons produced for such an event is equal to
\begin{equation}
\label{EqPoissonSolution}
\left<N_{PE}\right> = -\ln{\frac{N_{ped}}{N_{tot}}}
\end{equation}
where $N_{ped}$ and $N_{tot}$ are the pedestal event and total event counts
respectively. This general approach has been compared to other techniques in
\cite{dossi_methods_2000}, where R.~Dossi et al. also incorporate a model for
single- and multiple-photoelectron peaks for a $\chi^2$ fit to data. We have
chosen this approach since it has the advantage of few assumptions, and we have
found it to be robust in our application.
Systematic error in the determination of $N_{ped}$ comes from overlap
of the pedestal peak with contributions from non-zero photoelectron signals. This
is potentially largely due to incomplete photomultiplication in the large PMTs.
In practice, a sharp pedestal peak is necessary and a cutoff is placed on
the integration of the pedestal peak and varied to examine systematic effects.
As a guide for this cutoff, a method of determining the width of the pedestal
peak will be incorporated in order to place the cutoff at an estimated
probability of pedestal noise excursions.
\subsubsection{Deconvolution}
\begin{figure}[h]
\begin{center}
\includegraphics[width=4in]{figures/pedOverlay.pdf}
\end{center}
\caption{Convolution Fit - The data shown are from the
zero-photoelectron (pedestal) peak on the left side of Figure~\protect\ref{FigureSpectrum}.
The red line represents a reconvolution of the deconvolved spectrum
which has been fit to the data to extract the width of the PSF (point-spread
function).
The reduced $\chi^2$ of the fit is $102/97$. The green line
shows the zero pulse-height contribution from the deconvolved spectrum.
The three grey regions illustrate the pedestal integration to 90\%,
99\%, and 99.9\% of the total probability (see
Equation~\protect\ref{EqPedIntegral}).}
\label{FigureDeconvFit}
\end{figure}
To determine the width of the pedestal peak, a deconvolution method is applied to the
pulse-height spectrum with an estimated point-spread function (PSF). The PSF is
approximated by a Gaussian with an unknown width, and the reconvolved spectrum
is fit to the measured spectrum using a least-squares search method over the PSF
width parameter.
Figure~\ref{FigureDeconvFit} gives an example of the result of such a fit. The
purpose of such a process is to gain a precise form of the PSF.
By constraining the deconvolved spectrum to exist in a positive semi-definite
range, we can demand that the true charges deposited on the PMT anode are of only one
sign (negative, since the sign is inverted for a positive spectrum). This allows
the pedestal region of the spectrum to provide a stable constraint on the PSF.
In this way, the fitted PSF can be said to represent the random error in
the measurement of charge collected by the PMT anode.
An iterative deconvolution approach known as the Lucy-Richardson algorithm is
used \cite{citeulike:1719872}. This approach provides a straightforward means
of obtaining a maximum-likelihood solution while enforcing the condition $\psi_i
\geq 0$, where $\psi_i$ represent the discrete values of the deconvolved
spectrum.
\subsubsection{Pedestal integration}
Figure~\ref{FigureDeconvFit} shows the pedestal peak of a pulse-height spectrum
along with the predicted pedestal contribution from the deconvolved spectrum.
To determine the pedestal contribution from the data, the pulse-height spectrum
is integrated from left to right to specific probability intervals based on the
fitted PSF. The pedestal integral is taken to be
\begin{equation}
\label{EqPedIntegral}
N_{ped,i} = \displaystyle \int_{-\infty}^{\text{erfc}^{-1}\left[2(1-P_i)\right]\sqrt{2}
\sigma}
{\tilde\phi\left(x\right)dx}
\end{equation}
where $\text{erfc}^{-1}$ is the inverse complementary error function -- $P_i$ are 0.9,
0.99, and 0.999 -- $\sigma$ is the width of the Gaussian PSF --
and $\tilde\phi\left(x\right)$ is the measured spectrum. By combining Equation~\ref{EqPoissonSolution} and~\ref{EqPedIntegral}, $P_i$ provide average
photoelectron values ($\left<N_{PE,i}\right>$) for a mid-range result at
99\%, and a reasonable range from the upper bound at 90\% and the lower bound
at 99.9\%.
\section{Figure of Merit}
\label{SectionFoM}
D.W. Higinbotham \cite{Higinbotham1998332} presents a useful Figure of Merit
for threshold
\u{C}erenkov counters in the form of
\begin{equation}
\label{EquationHiginFoM}
H_{D.W.H.} \equiv \frac{\left<N_{PE}\right>}
{L\left(1-\frac{1}{\beta^2n^2}\right)}*\frac{1-\eta\left(1-\epsilon\right)}{\epsilon},
\end{equation}
where $L$ is the depth of the aerogel, $\eta$ is the average light box reflectivity
(using
reasonable values for reflector and aerogel reflectivity given by
\cite{Higinbotham1998332} for uniformity), and $\epsilon$ is the fraction of the
light box covered by photosensitive detector.
The factor $L\left(1-\frac{1}{\beta^2n^2}\right)$ is
approximately proportional to the number of photons radiated from the aerogel, and ${\epsilon}/\left[{1-\eta
\left(1-\epsilon\right)}\right]$ is the limit of
\begin{equation}
\displaystyle
\sum^{\infty}_{n = 0}{\epsilon\left[\eta\left(1-\epsilon\right)\right]},
\end{equation}
which estimates the probability of a photon
reaching a photocathode under the assumption that there is a flat distribution
of reflections across the surface of the light box.
$H$ then contains information on light losses in the aerogel, quantum efficiency of
photon detectors, and higher-order geometrical effects.
In the results of this paper, we will utilize this FoM in a slightly modified form.
Due to the depth of our counter and the rapidity of the beam used for testing,
\u{C}erenkov radiation in the air contributes non-trivially to our measured
photoelectron yield, making a FoM of the form of Equation~\ref{EquationHiginFoM}
artificially large. We will account for this by taking
\begin{equation}
L\left(1-\frac{1}{\beta^2n^2}\right) \rightarrow L_g\left(1-\frac{1}{\beta^2n_g^2} \right) +
L_a\left(1-\frac{1}{\beta^2n_a^2}\right)
\end{equation}
so that
\begin{equation}
\label{EquationFoM}
H \equiv \frac{\left<N_{PE}\right>}{L_g\left(1-\frac{1}{\beta^2n_g^2}\right)
+ L_a\left(1-\frac{1}{\beta^2n_a^2}\right)}*\frac{1-\eta\left(1-\epsilon\right)}
{\epsilon},
\end{equation}
where subscripts $g$ and $a$ represent aerogel and air respectively. This
provides an $O\left(10\%\right)$ correction that reduces the FoM to compare more
directly with counters that have less contribution from air.
\section{Aerogel configurations and results}
\label{SectionResults}
\begin{table*}[h]
\caption{Aerogel configuration results - The table below lists for each
aerogel configuration the batch number, total aerogel depth, average photoelectron yield,
mean refractive index, standard deviation of index measurements, and a Figure of Merit
(Equation~\protect\ref{EquationFoM})
in the style of D.W. Higinbotham \cite{Higinbotham1998332}.}
\label{TableResults}
\begin{center}
\begin{tabular}{c|c|c|c|c|c}
Batch & Depth (mm) & $\left<N_{PE}\right>$ & $\left<n\right>$ & $\sigma_n$ & $H$
(cm$^{-1}$) \\
\hline
Empty (air) & 0 & $0.58\substack{+0.19 \\ -0.02}$ & $\approx1.00028$ & N/A &
N/A\\
\hline
\multirow{3}{*}{1} & 28 & $3.81\substack{+0.29 \\ -0.08}$ & \multirow{3}{*}{$1.026$}
& \multirow{3}{*}{$0.0016$} & 40\\
& 41 & $4.91\substack{+0.32 \\ -0.10}$ & & & 36\\
& 56 & $6.04\substack{+0.34 \\ -0.12}$ & & & 32\\
\hline
2 & 53 & $6.36\substack{+0.34 \\ -0.12}$ & $1.028$ & 0.0007 & 34\\
\hline
3 & 62 & $5.60\substack{+0.28 \\ -0.09}$ & $1.028$ & 0.0023 & 28
\end{tabular}
\end{center}
\end{table*}
The performance of the \u{C}erenkov threshold test counter has been analyzed with three
different batches of aerogel. The batches provide a
small spread in refractive index and two different aerogel tile form factors.
The tiles from each batch are 1 - 2 cm deep squares.
\begin{description}
\item[Batch 1] has a mean refractive index of
1.026 and has a square area $\approx\text{10 x 10 }\text{cm}^2$.
\item[Batch 2] has a mean refractive index of 1.028, and also has a square area
$\approx\text{10 x 10 }\text{cm}^2$.
\item[Batch 3] also has a mean refractive index of 1.028, but with a square area
$\approx\text{20 x 20 }\text{cm}^2$. This is considered a large tile.
\end{description}
Table~\ref{TableResults} lists measurements for a series of aerogel configurations.
For each configuration, the total aerogel depth, average number of photoelectrons
$\left<N_{PE}\right>$, the mean refractive index, the spread in index measurements,
and a Figure of Merit have been calculated.
Figure~\ref{FigureDepthDep} shows a plot of the first four configurations listed in
Table~\ref{TableResults}, along with a fit to the values showing the non-linear depth
dependence of Batch 1.
This fit is of the form used by I. Adachi et al.
\cite{Adachi1995390}, and is given by
\begin{equation}
\label{eqFit}
\left<N_{PE}\right>_d=\left<N_{PE}\right>_\infty\times\left(1-\exp{\frac{-d}{L_{eff}}}\right)+C,
\end{equation}
where $d$ is the depth of the aerogel stack, and the effective absorption
length $L_{eff}$ contains information about absorption and scattering
in the aerogel convolved with the geometry and reflectivity of the test counter.
The spatial dependence of the test counter response with the Batch 3 large tiles
has been measured by scanning the counter transversely with respect to the beam,
and is shown in Figure~\ref{FigureSpatialScan}.
\begin{figure}[h] \begin{center}
\includegraphics[width=4in]{figures/npeVsDepth.pdf}
\end{center}
\caption{Batch 1 depth dependence - The red curve
below represents a fit to the data meant to extract the effective aerogel
absorption length $L_{eff}$ and a maximum photoelectron yield for an arbitrarily
deep stack of aerogel. The $\chi^2$ value for the fit is 0.01, which is not
unreasonable for a single degree of freedom, but tends to suggest that the
systematic errors stated in Table~\protect\ref{TableResults} are at high confidence
intervals as intended.}
\label{FigureDepthDep}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[width=4.5in]{figures/spatialScan2-3.pdf}
\end{center}
\caption{Spatial scan - The spatial dependence of the average photoelectron
yield
$\left<N_{PE}\right>$ of the
test counter with the Batch 3 large tile configuration is shown below. A large dip
down the
center is clearly illustrated, which is due to a metal support in the test counter
blocking
the \u{C}erenkov light cone.
A slight preference in the +x
direction exists due to differences in quantum efficiency between the PMTs on
either side of the aerogel.}
\label{FigureSpatialScan}
\end{figure*}
\subsection{Large tile performance}
Examining the features in Figure~\ref{FigureSpatialScan}, we find that a 1
cm-wide strip of aluminum that was used to hold the aerogel tiles in place
during testing produced a significant effect in the spatial scan as indicated by
the large dip in the center of the scan. There is also a small asymmetry in the
response of the counter on either side of the aluminum support, indicating a
difference in quantum efficiencies between the two PMT photocathodes. Finally,
a gentle slope exists toward the edges of the scan as a result of the geometry
of the light box. A Monte-Carlo simulation was performed which qualitatively
replicated each of these features.
The spatial scan shows that the uniformity of the counter in the regions not
affected by the aluminum aerogel support is quite good. Each of the
measurements listed in Table~\ref{TableResults} were obtained with the beam
center varying by no more than a few mm relative to the counter, and are each
around 50,-50 mm. We conclude that the small variations in relative beam center
produced variations between batches in the average number of measured
photoelectrons on the order of a percent.
\section{Summary}
\label{SectionSummary}
Hydrophobic aerogels have been produced for use as \u{C}erenkov radiators in a
threshold counter. Measurements of the performance of the aerogel/counter
combination with aerogel refractive indices from 1.026 to 1.028 and various
depths have been performed, yielding photoelectron counts between 3.8 and 6.4
(Table~\ref{TableResults}). Using a figure of merit in the style of D.W.
Higinbotham \cite{Higinbotham1998332}, low-order effects from the specific
refractive indices used -- as well as from the geometry (aerogel depth, PMT
area, etc) -- have been removed to produce values from 28 - 40 cm$^{-1}$ that
can be used to compare to previous experimental results.
\acknowledgments
This material is based upon work supported by the Department of Energy under
Award Number DE-SC0004290 and by the National Science Foundation Awards 0969654
and 1306547. The authors would also like to thank the OLYMPUS collaboration for
the beam time at the DESY detector testing facility.
\paragraph{Disclaimer}
This report was prepared as an account of work sponsored by an agency of the
United States Government. Neither the United States Government nor any agency
thereof, nor any of their employees, makes any warranty, express or implied, or
assumes any legal liability or responsibility for the accuracy, completeness, or
usefulness of any information, apparatus, product, or process disclosed, or
represents that its use would not infringe privately owned rights. Reference
herein to any specific commercial product, process, or service by trade name,
trademark, manufacturer, or otherwise does not necessarily constitute or imply
its endorsement, recommendation, or favoring by the United States Government or
any agency thereof. The views and opinions of authors expressed herein do not
necessarily state or reflect those of the United States Government or any agency
thereof.
\bibliographystyle{JHEP}
\section{First section}
Tasdfah asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi\footnote{Footnotes in the text are numbered.} soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd.
\subsection{Subsection}
Tasdfah asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd.
\section{Second section}
\subsection{Subsection}
Tasdfah asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
\subsubsection{Subsubsection}
Tasdfah asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd.
\section{Section}
Tasdfah asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
\acknowledgments
Tasdfah asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu
hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh
siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi
soiu hskjd asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd
asoiuh siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh
siuhihosi soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi
soiu hskjd. Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
Asoiuh siuhihosi soiu hskjd asoiuh siuhihosi soiu hskjd.
|
\section{Introduction}
The interest in the field of clusters is growing due to the increasing demand for nano-technology.
Besides the relevance for technology, clusters are also fundamentally very interesting. They behave
in general very different from their bulk counterpart. Also their electronic and magnetic properties
can drastically change by just adding or removing one atom.\cite{jena,chris,lars1,lars2} For example,
in a recent work on Nb$_{x}$Co clusters it is demonstrated that Nb$_{5}$Co and Nb$_{7}$Co are non-magnetic,
while Nb$_{4}$Co and Nb$_{6}$Co are strongly magnetic.\cite{larsrem} The physical origin of this behavior
can be traced back to the drastic change of the electronic structure as a function of cluster size.
The study of bimetallic clusters offers a broader playground than for pure clusters. This has resulted
in a number of intriguing observations.\cite{larsrem,lars3,lars4,pram,pram2,tay} For example, from an
anion photo-electron spectroscopy study on bimetallic Nb$_{x}$Co clusters, it was observed that for $x=6$
due to the addition of one Co atom to the Nb$_{x}$ host, the electronic structure resembles that of a typical bulk semi-conductor.\cite{pram} Therefore, this cluster was then proposed as a candidate for semiconductor materials.
From the theoretical side the bond properties and electronic structure of (NbCo)$_{x}$ clusters has
been investigated by means of relativistic density functional theory (DFT).\cite{wang} Further, the
geometry, stability and electronic properties of neutral and anionic Nb$_{x}$Co clusters is compared
with pure Nb$_{x}$ clusters within a DFT study.\cite{hong} Recently a combined theoretical
and experimental investigation has been performed on Nb$_{x}$Co clusters.\cite{larsrem} In this work
the geometry is obtained from a comparison of experimentally and theoretically obtained vibrational
spectra. With the geometry established the electronic structure is investigated in order to explain
the magnetic properties obtained from magnetic deflection experiments.
To our knowledge a systematic assessment of screening and correlation effects in Nb$_{x}$Co clusters does
not exist. This information is crucial in order to obtain a proper fundamental understanding of the system.
Namely correlation effects among the electrons inhibit in general an exact solution. Therefore, approximate
methods are required in practice. The choice of a suitable approximate method requires knowledge of the
effective Coulomb interaction in the system. More precisely, the gradient of the effective Coulomb
interaction is of importance.\cite{mot1,mot2} A very small gradient means that the effective Coulomb
interaction is merely constant, while a very large gradient indicates a purely local effective Coulomb
interaction. In the former case a mean-field treatment, i.e. single-particle approach, is probably a good
choice, while for the latter it might be the (generalized) Hubbard model.
The aim of the present work is the \textit{ab-initio} determination of the Coulomb interaction for NbCo to Nb$_{7}$Co
and Nb$_{9}$Co clusters. Besides being fundamentally interesting, such information is crucial to select an
adequate theoretical method for a further investigation of the system. The geometries of Nb$_{3}$Co to
Nb$_{7}$Co and Nb$_{9}$Co are well established from a comparison of theoretically and experimentally
obtained vibrational spectra.\cite{larsrem} In addition NbCo and Nb$_{2}$Co are considered, since the number
of isomers is very small. All these clusters are known to be magnetic except Nb$_{5}$Co and Nb$_{7}$Co,
which are non-magnetic. By employing the full-potential linearized augmented plane wave (FLAPW) method
using Wannier functions in conjunction with the random-phase approximation (RPA)\cite{ferd1,miy1,ers1},
it is found that in these clusters the onsite Coulomb interaction in RPA is well screened, while the
inter-site Coulomb interactions are barely screened. Interestingly for NbCo the inter-site interaction is
unscreened, while for Nb$_{4}$Co even anti-screening occurs. The important consequence being that the screened
Coulomb interaction is almost constant throughout the clusters. For completeness these results are compared
with pure Nb bulk for which only the onsite Coulomb interaction is appreciable, while the inter-site Coulomb
interactions are almost completely screened. Moreover, our constrained RPA calculations reveal that the
Co 3\emph{d} $\rightarrow$ 3\emph{d} channel only plays a minor role in the screening of the onsite
Coulomb interaction of the Co 3\emph{d} electrons. Finally, we find that both the onsite and inter-site
Coulomb interaction parameters decrease in a reasonable approximation linearly with the cluster size and
for clusters having more than 20 Nb atoms a transition from 0D to 3D screening is expected to take place. The
rest of the paper is organized as follows. The method and computational details are presented in
Section\,\ref{section-2}. Section\,\ref{section-4} deals with the results and discussion and finally in
Section\,\ref{section-5} we give the conclusions.
\section{Method and computational details}
\label{section-2}
In this work the screening of the Coulomb interaction in Nb$_x$Co clusters is calculated by means of the
ab-initio random phase approximation (RPA) method. The non-interacting reference system required for this method
comes from a preceding DFT calculation. In the following we shortly explain the RPA method and for details we
refer to Ref.~\onlinecite{lars3}. The screened Coulomb interaction is defined as
\begin{equation}
W(\boldsymbol{r},\boldsymbol{r}',\omega)=\int d\boldsymbol{r}'' \epsilon^{-1}(\boldsymbol{r},\boldsymbol{r}'',\omega) v(\boldsymbol{r}'',\boldsymbol{r}'),
\label{fullysw}
\end{equation}
where $\epsilon(\boldsymbol{r},\boldsymbol{r}'',\omega)$ is the dielectric function and $v(\boldsymbol{r}'',\boldsymbol{r}')$
is the bare Coulomb interaction potential. Since an exact expression for the dielectric function is not accessible,
an approximation is required. In the RPA the dielectric function is approximated by
\begin{equation}
\epsilon(\boldsymbol{r},\boldsymbol{r}',\omega)=\delta(\boldsymbol{r}-\boldsymbol{r}')-\int d\boldsymbol{r}'' v(\boldsymbol{r},\boldsymbol{r}'')P(\boldsymbol{r}'',\boldsymbol{r}',\omega),
\label{rpadiel1}
\end{equation}
where the polarization function $P(\boldsymbol{r}'',\boldsymbol{r}',\omega)$ is given by
\begin{equation}
\begin{gathered}
P(\boldsymbol{r},\boldsymbol{r}',\omega)=\\
\sum_{\sigma} \sum_{\boldsymbol{k},m}^{occ} \sum_{\boldsymbol{k}',m'}^{unocc} \varphi_{\boldsymbol{k}m}^{\sigma}(\boldsymbol{r}) \varphi_{\boldsymbol{k}'m'}^{\sigma*}(\boldsymbol{r}) \varphi_{\boldsymbol{k}m}^{\sigma*}(\boldsymbol{r}') \varphi_{\boldsymbol{k}'m'}^{\sigma}(\boldsymbol{r}') \\
\times\Bigg[ \frac{1}{\omega-\Delta_{\boldsymbol{k}m,\boldsymbol{k}'m'}^{\sigma}} - \frac{1}{\omega+\Delta_{\boldsymbol{k}m,\boldsymbol{k}'m'}^{\sigma}} \Bigg].
\end{gathered}
\label{rpapol1}
\end{equation}
\newline
Here $\Delta_{\boldsymbol{k}m,\boldsymbol{k}'m'}^{\sigma}=\epsilon_{\boldsymbol{k}'m'}^{\sigma}-\epsilon_{\boldsymbol{k}m}^{\sigma}-i\eta$
with $\epsilon_{\boldsymbol{k}m}^{\sigma}$ the single particle Kohn-Sham eigenvalues obtained from DFT and $\eta$ a positive infinitesimal.
Further, the $\varphi_{\boldsymbol{k}m}^{\sigma}(\boldsymbol{r})$ are the single particle Kohn-Sham eigenstates with
spin $\sigma$, wavenumber $\boldsymbol{k}$ and band index $m$. The tags $occ$ and $unocc$ above the summation symbol indicate that the
summation is respectively over occupied and unoccupied states only.
The Eqs.~(\ref{fullysw}), (\ref{rpadiel1}), and (\ref{rpapol1}) constitute what is called the RPA of the dynamically
screened Coulomb interaction. It is also possible to exclude certain screening contributions from Eq.~(\ref{rpapol1}),
which is referred to as \textit{constrained} RPA (cRPA). In this work the screening of the Coulomb interaction for the Co
3\textit{d} electrons and Nb 4\textit{d} electrons are investigated. One could for example exclude the screening contribution
coming from the Co 3\textit{d} states to obtain insight in their contribution to the total screening. More details on
the method used in this work to exclude certain screening contributions can be found in Ref.~\onlinecite{ers1}. Note that recently cRPA has become a very popular method to calculate Coulomb interaction parameters for different classes of materials.~\cite{Kotani,Schnell,Solovyev,Cococcioni,cRPA1,Zhang,Hunter}
The DFT calculations, providing the input of Eq.~\ref{rpapol1}, are performed with the FLEUR code. This code is
based on a FLAPW implementation.\cite{fleur} All calculations are performed with an exchange-correlation functional
in the generalized gradient approximation (GGA) as formulated by Perdew, Burke and Ernzerhof (PBE).\cite{gga} Further, all calculations are without spin orbit coupling. As will be demonstrated the effect of screening is on the $eV$ energy scale, while for the Co 3\textit{d} and Nb 4\textit{d} electrons the spin orbit coupling strength is at least an order of magnitude smaller. In addition it will be shown (Table~\ref{tabuco} and Fig.~\ref{fig3}) that the contribution of these electrons to the screening is small with respect to the other electrons, the Co 4\textit{sp} and Nb 5\textit{sp} electrons. For such extended states the spin orbit coupling strength is even smaller than for the Co 3\textit{d} and Nb 4\textit{d} electrons. Therefore, spin orbit coupling effects are expected to be small for the consideration of effective interactions. Since FLEUR is a {\bf k}-space code, a supercell approach was employed for the cluster calculations, with a large vacuum between clusters that were repeated in a periodic lattice. In order to prevent the interaction between clusters of different unit cells we performed tests for different unit cell sizes. We found that for a large unit cell of 12~\AA ~dimensions the results are converged to within a few percent. Therefore, this unit cell size is used for our calculations. Further, for the cluster calculations
the cutoff for the plane waves is 4.0~Bohr$^{-1}$, $l_{cut}=8$ and the $\Gamma$ point is the only {\bf k}-point
considered. For the calculations of bulk bcc Nb we use the same parameters with a $20 \times 20 \times 20$
\textbf{k}-point mesh and experimental lattice parameter of 3.3~\AA~of the bcc lattice. The ground state
geometric and magnetic structure of the Nb$_{3}$Co to Nb$_{7}$Co and Nb$_{9}$Co clusters is obtained from
Ref.~\onlinecite{larsrem} (see also Fig.~\ref{figclus}). More precisely, the geometries and magnetic structure
are obtained from a comparison of calculated and measured vibrational spectra. Structures of NbCo, Nb$_{2}$Co
and Nb$_{8}$Co were not addressed in Ref.~\onlinecite{larsrem}. Since the structure for Nb$_{8}$Co is unclear
due to the many possible isomers, we will only address NbCo and Nb$_{2}$Co in addition. In order to obtain the
ground state geometry of NbCo and Nb$_{2}$Co we performed the ATK-DFT calculations \cite{ATK-DFT} using the
GGA-PBE exchange-correlation functional \cite{ATK-GGA} and the SG15-Medium combination of norm-conserving
pseudopotentials and LCAO basis sets.\cite{Stradi,SG15} The total energy and forces have been converged at least
to ∼ $10^{-4}$ eV and 0.01 eV/\AA, respectively.
The DFT calculations are used as an input for the SPEX code to perform RPA and cRPA calculations for the screened
and partially screened (Hubbard $U$) Coulomb interaction.\cite{spex} The SPEX code uses the Wannier90 library to
construct the maximally localized Wannier functions.\cite{wan90,wan902} For this construction we used per spin channel six
states per Co atom, i.e. five 3$d$ states and one 4$s$ state, and nine states per Nb atom, five 4$d$ states,
one 5$s$ state and three 5$p$ states. More precisely, the maximally localized Wannier functions are used to
project the screened (bare) Coulomb interaction of Eq.~(\ref{fullysw}) on,
\begin{equation}
\begin{gathered}
U_{in_{1},jn_{3},in_{2},jn_{4}}^{\sigma_{1},\sigma_{2}}(\omega)= \\
\int \int d\boldsymbol{r}d\boldsymbol{r}' w_{in_{1}}^{\sigma_{1}*}(\boldsymbol{r}) w_{jn_{3}}^{\sigma_{2}*}(\boldsymbol{r}') W(\boldsymbol{r},\boldsymbol{r}',\omega) w_{jn_{4}}^{\sigma_{2}}(\boldsymbol{r}') w_{in_{2}}^{\sigma_{1}}(\boldsymbol{r}).
\end{gathered}
\label{hubudef31}
\end{equation}
\newline
Here $w_{in}^{\sigma}(\boldsymbol{r})$ is a maximally localized Wannier function located at site $i$ and spin $\sigma$.
In this work we only consider the static limit ($\omega=0$). Furthermore, we use Slater parametrization,
\begin{equation}
\begin{gathered}
U_{i}=\frac{1}{(2l+1)^{2}}\sum_{m,m'} U_{im,im',im,im'}^{\sigma_{1},\sigma_{2}}(\omega=0) \quad \text{and} \\
V_{ij}=\frac{1}{(2l+1)^{2}}\sum_{m,m'} U_{im,jm',im,jm'}^{\sigma_{1},\sigma_{2}}(\omega=0).
\end{gathered}
\label{fhuburpa}
\end{equation}
\newline
Here $U_{i}$ is the screened (bare) onsite Coulomb interaction at site $i$ and $V_{ij}$ the screened (bare) inter-site Coulomb
interaction between sites $i$ and $j$. Note that although the matrix elements of the Coulomb potential are formally
spin-dependent due to the spin dependence of the Wannier functions, we find that this dependence is negligible in
practice.
\section{Results and Discussion}
\label{section-4}
In Fig.\,\ref{figclus} the geometry of the NbCo to Nb$_{7}$Co and Nb$_{9}$Co clusters is depicted. The blue spheres
correspond to the Nb atoms and the red spheres to the Co atoms. Between brackets the point symmetry group of the
clusters is indicated. In the following we first address the fully screened (RPA) and partially screened without
the Co 3\emph{d} $\rightarrow$ 3\emph{d} channel (cRPA) onsite Coulomb interaction matrix elements of the Co
3\emph{d} electrons for the Nb$_{x}$Co clusters. This provides insight
into the contribution of the Co 3\emph{d} $\rightarrow$ 3\emph{d} channel to the total screening process. Second, the
fully screened onsite and inter-site Coulomb interaction matrix elements of the Nb 4\emph{d}
and Co 3\emph{d} orbitals are investigated. Finally, we make a comparison with pure bcc Nb bulk and investigate the influence of the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel on the screening of the onsite and inter-site Coulomb interaction of the Nb 4\emph{d} electrons. Note that the partially screened onsite Coulomb interaction is usually referred to as Hubbard $U$ and is what enters effective models, e.g. the Hubbard model.
\begin{figure}[t]
\begin{center}
\includegraphics[scale=0.45]{fig1.eps}
\end{center}
\vspace{-0.3 cm}
\caption{The geometry of the NbCo to Nb$_{7}$Co and Nb$_{9}$Co clusters.
Between brackets the point symmetry group of the cluster is indicated.
The blue spheres correspond to the Nb atoms and the red spheres to the Co atoms. }
\label{figclus}
\end{figure}
In Table~\ref{tabuco}, the bare, partially screened without the Co 3\emph{d} $\rightarrow$ 3\emph{d} channel (cRPA) and fully screened (RPA) average onsite Coulomb interaction
matrix elements of the Co 3\emph{d} electrons are presented. As it is seen the bare interaction is constant as
function of cluster size, while the partially and fully screened interactions decrease with size. This can be attributed
to the increase of screening channels with increasing cluster size rather than the delocalization of the Wannier functions.
Note that very similar matrix elements for the onsite bare Coulomb interaction for all clusters reflect the fact that
the localization of the Wannier functions does not change with increasing the cluster size. Furthermore, the obtained cRPA
and RPA Coulomb matrix elements for the Co 3\textit{d} orbitals are very close to each other, which means that the
contribution of the Co 3\emph{d} $\rightarrow$ 3\emph{d} channel to the total screening is very small compared to the
other screening channels (see Table~\ref{tabuco}).
\begin{table}[t]
\begin{center}
\caption{The bare, partially screened without the Co 3\emph{d} $\rightarrow$ 3\emph{d} channel (cRPA) and fully screened (RPA) average onsite Coulomb interaction
parameters for the Co 3\emph{d} orbitals of the NbCo-Nb$_{7}$Co and Nb$_{9}$Co clusters obtained from \emph{ab-initio}
calculations.}
\begin{ruledtabular}
\begin{tabular}{lccc}
Cluster & Bare (eV) & cRPA (eV) & RPA (eV) \\ \hline
NbCo & 22.2 & 7.9 & 7.7 \\
Nb$_{2}$Co & 22.2 & 5.9 & 5.8 \\
Nb$_{3}$Co & 22.3 & 5.6 & 5.5 \\
Nb$_{4}$Co & 22.6 & 5.2 & 5.0 \\
Nb$_{5}$Co & 22.7 & 4.7 & 4.6 \\
Nb$_{6}$Co & 22.7 & 4.4 & 4.3 \\
Nb$_{7}$Co & 22.7 & 4.1 & 4.1 \\
Nb$_{9}$Co & 22.9 & 3.9 & 3.8 \\
\end{tabular}
\end{ruledtabular}
\label{tabuco}
\end{center}
\end{table}
In Table~\ref{tabx} and~\ref{tabxx}, the bare (fourth column) and fully screened (fifth column)
onsite and inter-site average Coulomb interaction parameters for Nb 4\emph{d} and Co
3\emph{d} orbitals are presented for the Nb$_{x}$Co clusters. Due to the symmetry of
some clusters (see Fig.~\ref{figclus}), some Nb atoms are equivalent. In Fig.~\ref{figclus}
for Nb$_{2}$Co atoms 1 and 2 are equivalent, for Nb$_{3}$Co 1, 2 and 3 are equivalent,
for Nb$_{4}$Co 1, 2 and 3 are equivalent, for Nb$_{5}$Co there are no equivalent atoms,
for Nb$_{6}$Co 3, 5 and 6, and 1, 2 and 4 are equivalent, for Nb$_{7}$Co 5 and 7, and 4
and 6 are equivalent and for Nb$_{9}$Co 1, 4, 6 and 8, and 2, 5, 7 and 9 are equivalent.
In Table~\ref{tabx} and~\ref{tabxx} only symmetry unequivalent interactions are shown. Further, in
the first column $U_{1}$ corresponds to the onsite Coulomb interaction of atom 1 and $V_{1,2}$
to the inter-site Coulomb interaction between atoms 1 and 2 (see Fig.~\ref{figclus}). The
second column indicates between what type of atoms this refers and the third column correponds
to the distance in \AA ~between them. From this table it can be seen that besides for
the Co 3\emph{d} electrons also for the Nb 4\emph{d} electrons the onsite Coulomb interaction
is well screened and decreases with cluster size. On the other hand, the inter-site Coulomb
interaction is much less screened and is more or less constant as a function of interatomic
distance. This appears to be due to a decrease in the screening as a function of increasing
interatomic distance. Interestingly for NbCo the inter-site interaction is unscreened, while
for Nb$_{4}$Co there is even anti-screening present between Nb and Co at an interatomic distance
of 3.89~\AA. Anti-screening means that the the screened interaction is larger than the bare
interaction. By considering the effective interaction between two point charges in a medium,
screening is understood to be due to the response (polarization) of the medium to these charges.
Similarly anti-screening occurs when the medium is polarized in such a way to increase the bare
interaction between the two point charges. This situtation is known to occur only for low dimensional
systems such as carbon nanotubes, nanoribbons, wires, molecules and clusters.\cite{mot1,mot2}
From a simplistic point of view the two induced point charges can be considered as giving rise
to point-dipoles at the positions of the polarizable atoms that constitute the medium. Each point-dipole
produces an electric field and depending on its orientation it either increases or decreases
the bare electric field coming from the two point charges. Roughly the point-dipoles in between
the two point charges are oriented to increase, anti-screen, the bare interaction, whereas the
other surrounding point-dipoles lead to a reduction, screening, of the bare interaction.\cite{mot2}
Therefore, the occurence of anti-screening crucially depends on the dimensionality of the system and
distance between the induced point charges. More precisely, for low dimensional systems the ratio
of the region between the point charges and the rest of the medium is larger.
Anti-screening was also recently found in Fe$_{x}$O$_{y}$ clusters by means of ab-initio
calculations.\cite{lars3} However, the anti-screening appears to be more pronounced in Fe$_{x}$O$_{y}$
clusters than in Nb$_{x}$Co clusters. In order to qualitatively understand this, the microscopic
point-dipole model can be used. Within this model the atoms of the system are considered as
classical polarizable point-dipoles. These point-dipoles are then allowed to respond to an
external perturbation, e.g. induced point charges. From investigations on low-dimensional
systems by means of this microscopic point-dipole model, it is well established that anti-screening
delicately depends on the geometry and polarizability of the atoms constituting the system.\cite{mot1,mot2}
However, in general it is demonstrated that the inter-atomic distance at which anti-screening
occurs increases with increasing polarizability (see for example Fig.~1.10 of Ref.~\onlinecite{mot2}).
Further, from for example ab-initio calculations on isolated atoms it is known that the polarizability
of Nb is larger than that of Fe, Co and O.\cite{atompol} Fe and Co have a similar polarizability,
which is again larger than that of O. Based on these observations anti-screening in
Nb$_{x}$Co clusters is expected to occur at larger inter-atomic distances compared to Fe$_{x}$O$_{y}$
clusters, which explains why anti-screening is more pronounced in the latter.
The discussion above on the difference in anti-screening between Nb$_{x}$Co and Fe$_{x}$O$_{y}$ clusters
is based on the microscopic point-dipole model. It is however not clear if these clusters can be modeled
by a collection of point-dipoles. Therefore, it is instructive to also discuss anti-screening differences
based on Eq.~(\ref{rpapol1}). It is known that anti-screening only occurs in low-dimensional semiconductors
and insulators.\cite{mot1,mot2,Louie,Nomura} As mentioned above, the critical distance for the appearance of
anti-screening increases with increasing polarizability, which can be traced back to the distribution of the
occupied and unoccupied electronic states around the Fermi energy (strictly speaking chemical potential
for the clusters). In Fig.\,\ref{fig2} we present the density of states for Fe$_2$O$_3$ and Nb$_3$Co
clusters, which is calculated using the Gaussian method with a broadening parameter of 0.1 eV.
The polarizability (see Eq.~(\ref{rpapol1})) is inversely proportional to the energy difference between
occupied and unoccupied states, i.e., the smaller the energy difference the larger the polarizability. Indeed, as
seen in Fig.\,\ref{fig2} the Nb$_3$Co cluster has more states around the chemical potential with respect to
the Fe$_2$O$_3$ cluster despite similar HOMO-LUMO energy gaps of both clusters. As a consequence, the contribution
of the term between square brackets in Eq.~(\ref{rpapol1}) is larger for the Nb$_3$Co cluster giving rise to
smaller Coulomb matrix elements and absence of anti-screening for inter-site Coulomb interactions. Thus, similar as
for the microscopic point-dipole model, a small polarization or equivalently polarizability of the system
is required to observe anti-screening at short distances.
\begin{table}[H]
\begin{center}
\caption{The bare and fully screened (RPA) average Coulomb interaction
parameters for the Nb 4\emph{d} and Co 3\emph{d} orbitals for the NbCo to Nb$_{5}$Co
obtained from \emph{ab-initio} calculations. Here $U_{1}$
corresponds to the onsite Coulomb interaction of atom 1 and $V_{1,2}$ to the inter-site
Coulomb interaction between atoms 1 and 2 (see Fig.~\ref{figclus}). The second column
indicates between what type of atoms this refers and the third column correponds to
the distance in \AA{ }between them. Note that due the symmetry of some clusters, some
Nb atoms are equivalent.}
\begin{ruledtabular}
\begin{tabular}{lcccc}
& & NbCo && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 11.2 & 7.2 \\
$U_{2}$ & Co & 0 & 22.2 & 7.7 \\
$V_{1,2}$ & Nb-Co & 1.99 & 7.0 & 7.0 \\ \hline
&&& \\
& & Nb$_{2}$Co && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 10.0 & 5.2 \\
$U_{3}$ & Co & 0 & 22.2 & 5.8 \\
$V_{1,2}$ & Nb-Nb & 2.16 & 6.0 & 4.9 \\
$V_{1,3}$ & Nb-Co & 2.33 & 6.1 & 5.0 \\ \hline
&&& \\
& & Nb$_{3}$Co && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 10.7 & 5.0 \\
$U_{4}$ & Co & 0 & 22.3 & 5.5 \\
$V_{1,2}$ & Nb-Nb & 2.40 & 5.6 & 4.5 \\
$V_{1,4}$ & Nb-Co & 2.47 & 5.7 & 4.6 \\ \hline
&&& \\
&& Nb$_{4}$Co && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 11.0 & 4.6 \\
$U_{4}$ & Nb & 0 & 10.8 & 4.5 \\
$U_{5}$ & Co & 0 & 22.6 & 5.0 \\
$V_{1,5}$ & Nb-Co & 2.40 & 6.0 & 4.2 \\
$V_{1,4}$ & Nb-Nb & 2.52 & 5.5 & 4.1 \\
$V_{1,2}$ & Nb-Nb & 2.61 & 5.4 & 4.1 \\
$V_{4,5}$ & Nb-Co & 3.89 & 4.0 & \textbf{4.1} \\ \hline
&&& \\
&& Nb$_{5}$Co && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 11.2 & 4.3 \\
$U_{2}$ & Nb & 0 & 10.8 & 4.3 \\
$U_{3}$ & Nb & 0 & 11.0 & 4.3 \\
$U_{4}$ & Nb & 0 & 11.2 & 4.3 \\
$U_{5}$ & Nb & 0 & 11.2 & 4.3 \\
$U_{6}$ & Co & 0 & 22.7 & 4.6 \\
$V_{1,6}$ & Nb-Co & 2.27 & 6.3 & 3.9 \\
$V_{4,6}$ & Nb-Co & 2.28 & 6.2 & 3.8 \\
$V_{5,6}$ & Nb-Co & 2.32 & 6.1 & 3.8 \\
$V_{3,5}$ & Nb-Nb & 2.40 & 5.7 & 3.8 \\
$V_{1,2}$ & Nb-Nb & 2.44 & 5.6 & 3.8 \\
$V_{2,3}$ & Nb-Nb & 2.49 & 5.5 & 3.8 \\
$V_{2,4}$ & Nb-Nb & 2.50 & 5.5 & 3.8 \\
$V_{3,4}$ & Nb-Nb & 2.63 & 5.3 & 3.7 \\
$V_{1,5}$ & Nb-Nb & 2.69 & 5.3 & 3.7 \\
$V_{3,6}$ & Nb-Co & 2.81 & 5.1 & 3.8 \\
$V_{1,4}$ & Nb-Nb & 2.92 & 4.9 & 3.7 \\
$V_{2,5}$ & Nb-Nb & 2.99 & 4.8 & 3.7 \\
$V_{2,6}$ & Nb-Co & 3.27 & 4.5 & 3.8 \\
$V_{4,5}$ & Nb-Nb & 3.68 & 4.1 & 3.7 \\
$V_{1,3}$ & Nb-Nb & 3.72 & 4.1 & 3.7 \\
\end{tabular}
\end{ruledtabular}
\label{tabx}
\end{center}
\end{table}
\begin{table}[H]
\begin{center}
\caption{The same as Table~\ref{tabx} for Nb$_{6}$Co, Nb$_{7}$Co and Nb$_{9}$Co clusters. Here $U_{1}$
corresponds to the onsite Coulomb interaction of atom 1 and $V_{1,2}$ to the inter-site
Coulomb interaction between atoms 1 and 2 (see Fig.~\ref{figclus} for the geometry of the corresponding clusters).}
\begin{ruledtabular}
\begin{tabular}{lcccc}
&& Nb$_{6}$Co && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 11.3 & 4.0 \\
$U_{3}$ & Nb & 0 & 11.3 & 4.0 \\
$U_{7}$ & Co & 0 & 22.7 & 4.3 \\
$V_{3,7}$ & Nb-Co & 2.33 & 6.2 & 3.5 \\
$V_{3,4}$ & Nb-Nb & 2.53 & 5.5 & 3.5 \\
$V_{1,2}$ & Nb-Nb & 2.73 & 5.2 & 3.4 \\
$V_{3,5}$ & Nb-Nb & 2.88 & 5.0 & 3.4 \\
$V_{2,3}$ & Nb-Nb & 3.78 & 4.0 & 3.3 \\
$V_{2,7}$ & Nb-Co & 3.91 & 3.9 & 3.4 \\ \hline
&&& \\
&& Nb$_{7}$Co && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 11.3 & 3.9 \\
$U_{2}$ & Nb & 0 & 11.2 & 3.9 \\
$U_{3}$ & Nb & 0 & 11.2 & 3.9 \\
$U_{4}$ & Nb & 0 & 11.3 & 3.9 \\
$U_{5}$ & Nb & 0 & 11.4 & 3.9 \\
$U_{8}$ & Co & 0 & 22.7 & 4.1 \\
$V_{5,8}$ & Nb-Co & 2.30 & 6.2 & 3.4 \\
$V_{1,8}$ & Nb-Co & 2.43 & 5.9 & 3.3 \\
$V_{2,4}$ & Nb-Nb & 2.46 & 5.6 & 3.4 \\
$V_{1,4}$ & Nb-Nb & 2.53 & 5.5 & 3.4 \\
$V_{3,5}$ & Nb-Nb & 2.54 & 5.5 & 3.3 \\
$V_{2,3}$ & Nb-Nb & 2.56 & 5.4 & 3.3 \\
$V_{3,4}$ & Nb-Nb & 2.83 & 5.0 & 3.3 \\
$V_{1,5}$ & Nb-Nb & 2.85 & 5.0 & 3.3 \\
$V_{1,2}$ & Nb-Nb & 2.86 & 5.0 & 3.3 \\
$V_{5,7}$ & Nb-Nb & 2.90 & 5.0 & 3.3 \\
$V_{1,3}$ & Nb-Nb & 3.15 & 4.6 & 3.3 \\
$V_{3,8}$ & Nb-Co & 3.78 & 4.1 & 3.3 \\
$V_{4,8}$ & Nb-Co & 3.97 & 3.9 & 3.3 \\
$V_{4,6}$ & Nb-Nb & 3.98 & 3.9 & 3.2 \\
$V_{2,5}$ & Nb-Nb & 4.08 & 3.8 & 3.2 \\
$V_{4,5}$ & Nb-Nb & 4.26 & 3.7 & 3.2 \\
$V_{2,8}$ & Nb-Co & 4.88 & 3.4 & 3.3 \\ \hline
&&& \\
&& Nb$_{9}$Co && \\
U/V & Atom & Distance (\AA) & Bare (eV) & RPA (eV) \\ \hline
$U_{1}$ & Nb & 0 & 11.5 & 3.4 \\
$U_{2}$ & Nb & 0 & 11.4 & 3.5 \\
$U_{3}$ & Nb & 0 & 11.4 & 3.4 \\
$U_{10}$ & Co & 0 & 22.9 & 3.8 \\
$V_{1,10}$ & Nb-Co & 2.42 & 6.0 & 3.0 \\
$V_{1,7}$ & Nb-Nb & 2.53 & 5.5 & 2.9 \\
$V_{2,3}$ & Nb-Nb & 2.57 & 5.5 & 2.9 \\
$V_{1,4}$ & Nb-Nb & 2.82 & 5.1 & 2.9 \\
$V_{2,5}$ & Nb-Nb & 2.87 & 5.0 & 2.9 \\
$V_{2,10}$ & Nb-Co & 3.94 & 3.9 & 2.9 \\
$V_{1,8}$ & Nb-Nb & 3.99 & 3.9 & 2.9 \\
$V_{2,7}$ & Nb-Nb & 4.05 & 3.8 & 2.8 \\
$V_{1,3}$ & Nb-Nb & 4.11 & 3.8 & 2.8 \\
$V_{1,2}$ & Nb-Nb & 4.22 & 3.7 & 2.8 \\
$V_{3,10}$ & Nb-Co & 4.96 & 3.3 & 2.9 \\
\end{tabular}
\end{ruledtabular}
\label{tabxx}
\end{center}
\end{table}
\noindent A similar discussion holds for all other clusters, for
instance the NbCo cluster has a similar molecular energy level distribution around the chemical potential as Fe$_{3}$O$_{4}$
(not shown). Then, anti-screening is expected to occur at similar inter-site distances in these clusters. For
Fe$_{3}$O$_{4}$ this is expected to occur a bit below $3.4$~\AA~(see Table~I of Ref.~\onlinecite{lars3}), while
for NbCo indeed just above $3.0$~\AA. Further, although Nb$_{2}$Co and Nb$_{3}$Co show a similar molecular energy
spectrum around the chemical potential as the Fe$_{4}$O$_{6}$ cluster, anti-screening is not observed, because
the inter-site distances are too small compared to Fe$_{4}$O$_{6}$. For Nb$_{4}$Co and larger clusters the density of
molecular energy levels around the chemical potential increases and is quite a bit denser than for the Fe$_{x}$O$_{y}$
clusters. Therefore, anti-screening in these clusters is only expected for large inter-site distances. For example,
in Nb$_{4}$Co it occurs at $3.89$~\AA, while for Nb$_{7}$Co at an inter-site Nb-Co distance of $4.88$~\AA~the situation
is very close to anti-screening.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.67]{fig2.eps}
\end{center}
\vspace{-2.4 cm}
\caption{Calculated spin-resolved total density of states for Fe$_{2}$O$_{3}$ and Nb$_{3}$Co clusters.
The Fermi energy (chemical potential) is set to zero.}
\label{fig2}
\end{figure}
It is instructive to compare the Nb$_{x}$Co results with pure Nb bulk. In Table~\ref{tabbulk} the bare,
partially screened (without the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel), and fully screened onsite
and inter-site Coulomb interaction matrix elements of the Nb 4\emph{d} electrons are presented. From this
table it is clear that the inter-site Coulomb interaction in RPA is almost completely screened. This is in
contrast with what is observed for the clusters (Table~\ref{tabx} and~\ref{tabxx}). Further, the fully
screened onsite Coulomb interaction is more screened than for the clusters. The important observation
for pure Nb bulk is that the effective Coulomb interaction is not constant throughout the system. Instead,
it is localized, i.e. short ranged with a large gradient. Further, the contribution of the Nb 4\emph{d}
$\rightarrow$ 4\emph{d} channel to the screening can be investigated from Table~\ref{tabbulk}. Both for
the onsite and inter-site effective interaction this contribution is small, about 1.8~eV and 0.07~eV
(for the nearest-neighbor interaction), compared to the contribution of about 11~eV and 4.9~eV of
the other channels. The main screening contribution comes from the 5\emph{s} states, which are present around the Fermi level.
\begin{table}[!b]
\begin{center}
\caption{The bare, partially screened without the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel (cRPA), and fully screened (RPA) average onsite and inter-site Coulomb interaction parameters for the Nb 4\emph{d} orbitals of pure Nb bulk. Here the
first column refers to the distance in~\AA{ } between two Nb atoms, i.e. zero corresponds to the onsite
interaction.}
\begin{ruledtabular}
\begin{tabular}{lccc}
Distance (\AA) & Bare (eV) & cRPA (eV) & RPA (eV) \\ \hline
0 & 13.81 & 2.62 & 0.83 \\
2.86 & 5.01 & 0.08 & 0.01 \\
3.30 & 4.35 & 0.04 & 0.00 \\
4.67 & 3.11 & 0.01 & 0.00 \\
5.72 & 2.57 & 0.00 & 0.00 \\
\end{tabular}
\end{ruledtabular}
\label{tabbulk}
\end{center}
\end{table}
For the Nb$_{x}$Co clusters the influence of the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel
can be obtained from an inspection of Fig.~\ref{fig3}. Here an average of the partially screened
(without the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel) and fully screened onsite and
nearest-neighbor inter-site Coulomb interaction parameters for the Nb 4\emph{d} orbitals are
presented as function of cluster size. The cluster size is indicated by $x$, which represents
the number of Nb atoms in the Nb$_{x}$Co clusters. It appears that the contribution of the Nb
4\emph{d} $\rightarrow$ 4\emph{d} channel to the screening of the onsite and inter-site effective
interaction increases with cluster size. For instance for the onsite interaction this contribution
is about 0.3~eV for Nb$_{2}$Co and becomes about 0.9~eV for the Nb$_{9}$Co cluster. In case of
the nearest-neighbour inter-site interaction the contribution for Nb$_{2}$Co is almost 0~eV and
becomes about 0.3~eV for Nb$_{9}$Co. Compared to the contributions of the other channels (see
Table~\ref{tabx} and~\ref{tabxx} for the unscreened bare values), it can be concluded that the
contribution of the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel to the screening is small.
Namely for the onsite interaction the contribution of the other channels is about 5~eV for
Nb$_{2}$Co and becomes about 8~eV for Nb$_{9}$Co. In case of the nearest-neighbor inter-site
interaction this is about 1~eV for Nb$_{2}$Co and 2.5~eV for Nb$_{9}$Co.
It is interesting to obtain insight at what cluster size the behavior of the screened Coulomb interaction
becomes bulk like. For this purpose Fig.~\ref{fig3} is used again, where the averaged onsite and nearest-neighbor
inter-site screened and partially screened Coulomb interaction parameters between Nb 4\emph{d} electrons
are presented as function of cluster size and compared with the pure Nb bulk values of Table~\ref{tabbulk}
(green dashed and solid lines for the onsite cRPA and RPA interactions, respectively). The blue and red solid
(dashed) lines correspond to a linear extrapolation of the Nb$_{x}$Co cluster data points for which the smallest
cluster ($x=1$) is ignored. From these extrapolations it appears that both the averaged onsite and inter-site
screened and partially screened Coulomb interaction depend in a reasonable approximation linearly on the cluster
size. At a cluster size of $x=20$ both the onsite and nearest-neighbor inter-site screened (partially screened)
interaction have reached their corresponding bulk values, i.e. 0.83 eV (2.62 eV) and 0.01~eV (0.08 eV), respectively.
Therefore, we expect Nb$_{x}$Co clusters with $x$ larger than 20 to have a bulk like behavior.
\begin{figure}[!t]
\begin{center}
\includegraphics[scale=0.7]{fig3.eps}
\end{center}
\vspace{-0.5 cm}
\caption{The averaged partially screened without the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel and fully
screened onsite $U_{avg}$ (onsite) and nearest-neighbour inter-site $V_{avg}$ (nn) matrix elements
between Nb 4\emph{d} electrons as function of cluster size ($x$) for the Nb$_{x}$Co clusters. Here $x$
corresponds to the number of Nb atoms in the Nb$_{x}$Co cluster. Ignoring the smallest
cluster ($x=1$), the blue and red solid and dashed lines represent an extrapolation of the data points.
The green solid and dashed lines represent respectively the onsite fully screened and partially screened
Coulomb interaction for pure Nb bulk.}
\label{fig3}
\end{figure}
Finally, we would like to comment on the strength of the electronic correlations in the Nb$_{x}$Co clusters
and Nb bulk. As shown in Table~\ref{tabx} and~\ref{tabxx} the effective Coulomb interaction is more or less
constant throughout the clusters. In contrast for Nb bulk it has a strong gradient and is local in nature.
Although the effective Coulomb interaction in Nb bulk is mainly local, it should not be considered as a
(strongly) correlated material. For this purpose the band width should also be taken into account. The band
width is about 7.5~eV~\cite{nb1}, which is much larger than the effective onsite Coulomb interaction of 0.83~eV (see Table~\ref{tabbulk}). Therefore, it should be interpreted as a weakly correlated material and standard DFT is
expected to provide a good description of the essential physics. This is confirmed by DFT studies on the elastic
properties, band structure and electron-phonon coupling of Nb bulk, which are in good agreement with experiments.\cite{nb1,nb2,nb3} Due to the almost constant effective Coulomb interaction in the Nb$_{x}$Co clusters, it is also expected that DFT
should be able to capture the essential physics. This is confirmed by a comparison of the vibrational spectra
obtained within DFT and experiments.\cite{larsrem} Furthermore, in Ref.~\onlinecite{hong} it is correctly predicted
within DFT that Nb$_{7}$Co should be non-magnetic. The wrong prediction of Nb$_{5}$Co to be magnetic is probably
due to the consideration of the wrong geometry (see Ref.~\onlinecite{larsrem}). Besides providing an explanation
for the success of DFT in these clusters, our results are crucial to select an adequate method for future
investigations on many-body effects, e.g. quasi-particle life times. For example, intuitively one might expect DFT
in combination with the dynamical mean-field theory (DMFT)~\cite{dmft} to be suitable for this purpose, because the
Co atom can be interpreted as an impurity in a Nb$_{x}$ host. Since DMFT only properly treats local correlations,
while we have demonstrated non-local correlations to be also important, this is not a justified choice. Therefore,
an extended Hubbard-like model or the consideration of the cluster within multiplet ligand-field theory~\cite{mlft}
are probably more suitable choices.
In addition we expect, that due to the almost constant effective interaction in the Nb$_{x}$Co clusters, the observed trends are robust with respect to the choice of the exchange-correlation functional. For example, the local density approximation (LDA)~\cite{lda1} and GGA are expected to perform similar due to the constant interaction, because both methods are derived in the limit of a (nearly) uniform electron gas. As a test we made for all clusters a comparison between the density of states in GGA and LDA. Since they were found to be very similar around the Fermi level, it is indeed expected based on Eq.~\ref{rpapol1} that our results are robust with respect to the choice of the exchange-correlation functional.
\section{Conclusion}
\label{section-5}
We have performed RPA and cRPA calculations to reveal the screening of the Coulomb interaction in Nb$_{x}$Co ($1 \leq x \leq 9$)
clusters and pure Nb bulk. We have found that in both the clusters and the bulk the onsite Coulomb interaction in RPA
is well screened. On the other hand the inter-site Coulomb interaction is much less screened in the clusters resulting in an almost
constant interaction throughout the clusters. This is in contrast with pure Nb bulk, where the inter-site Coulomb interaction
in RPA is almost completely screened. Our cRPA calculations have shown that the contribution of the Co 3\emph{d} $\rightarrow$
3\emph{d} channel to the total screening process of the onsite Coulomb parameters of the Co 3\emph{d} electrons is negligible. Further, for the clusters investigated the contribution of the Nb 4\emph{d} $\rightarrow$ 4\emph{d} channel to effective onsite
and inter-site Coulomb parameters of the Nb 4\emph{d} electrons appears to be small compared to that of the total screening
contribution. Based on our findings we expect both for the Nb$_{x}$Co clusters and Nb bulk that correlation effects play a
minor role and that standard DFT is able to capture the essential physics. For the clusters this is due to the almost constant
effective Coulomb interaction and for the bulk due to the band width being much larger than the essentially local effective
Coulomb interaction. Finally, it has been found that both the onsite and inter-site Coulomb interaction parameters decrease
in a reasonable approximation linearly with cluster size and for Nb$_{x}$Co clusters having more than 20 Nb atoms a transition
from 0D to 3D screening is expected to take place.
\subsection*{Acknowledgements}
The Nederlandse Organisatie voor Wetenschappelijk Onderzoek (NWO) and SURFsara
are acknowledged for the usage of the LISA supercomputer and their support. L.P. and
M.I.K. acknowledges a support by European ResearchCouncil (ERC) Grant No. 338957. E. S. and I. M. greatly acknowledge the funding provided by the European Union (EFRE).
|
\section*{Methods}
The analyses described here were based on the PRESTO\cite{Ran01}, PSRCHIVE\cite{Str12}, and DSPSR\cite{Str11} pulsar software suites, as well as custom-written Python scripts for linking utilities into reduction pipelines, fitting the data, and plotting.
\subsubsection*{Observations and burst search}
\noindent {\it Arecibo} \\
We observed using the Arecibo `C-band' receiver (dual linear receptors), in the frequency range $4.1-4.9$\,GHz, and the Puerto-Rican Ultimate Pulsar Processing Instrument (PUPPI) backend recorder. The full list of observations is reported in Extended Data Table~\ref{tab:observations}. We operated PUPPI in its `coherent search' mode, which produced $10.24$\,$\mu$s samples and $512 \times 1.56$\,MHz frequency channels, each coherently dedispersed to $\text{DM} = 557.0$\,pc\,cm$^{-3}$. Coherent dedispersion within each 1.56-MHz channel means that the intra-channel dispersive smearing is $< 2$\,$\mu$s even if the burst DM is $10$\,pc\,cm$^{-3}$ higher/lower than the fiducial value of $557.0$\,pc\,cm$^{-3}$ used in the PUPPI recording. The raw PUPPI data also provide auto- and cross-correlations of the two linear polarizations, which can be converted to Stokes I, Q, U, and V parameters in post-processing. Before each observation, both a test scan on a known pulsar (PSR~B0525+21) and a noise-diode calibration scan (for polarimetric calibration) were performed.
Dedispersed time series with DM = $461-661$\,pc\,cm$^{-3}$, in trial steps of $1$\,pc\,cm$^{-3}$, were searched using PRESTO's {\tt single\_pulse\_search.py}, which applies a matched-filter technique to look for
bursts with durations between $81.92$\,$\mu$s to $24576$\,$\mu$s (for any putative burst that only has a single peak with width $< 81.92$\,$\mu$s, the sensitivity will be degraded by a factor of a few, at most).
The resulting DM-time-S/N events were grouped into plausible astrophysical burst candidates using a custom sifting algorithm and then a dynamic spectrum of each candidate was plotted for human inspection and grading. We found 16 bursts of astrophysical origin, and used the DSPSR package to form full-resolution, full-polarization PSRCHIVE `archive' format files for each burst.
\noindent {\it Green Bank Telescope}
On August 26, 2017, we observed FRB\,121102\ using the GBT `C-band' receiver ($4-8$\,GHz, with dual linear receptors) as part of a program of monitoring known FRB positions. Observations were conducted with the Breakthrough Listen Digital Backend\cite{Mac17}, which allowed recording of baseband voltage data across the entire nominal 4-GHz bandwidth of the selected receiver.
Scans of a noise-diode calibration, of the flux calibrator 3C161 and of the bright pulsar PSR~B0329+54 supplemented the observations.
In post-processing, a total intensity, low-resolution filterbank data product was searched
for bursts with DM$ = 500 - 600$\,pc\,cm$^{-3}$, using trial DMs in steps of 0.1\,pc\,cm$^{-3}$ and a GPU-accelerated search package to perform the incoherent dedispersion\cite{Bar12}. We detected\cite{Gaj17} 15 bursts with S/N $> 10$. Here we present the properties of just the two brightest GBT bursts in order to confirm the large RM observed by Arecibo and to quantify its variation in time. A detailed analysis of all GBT detections is presented in Gajjar et al. (in prep.).
A section of raw voltage data (1.5\,s) around each detected burst was extracted and coherently
dedispersed to a nominal DM of 557.91\,pc\,cm$^{-3}$ using the DSPSR package.
Final PSRFITS format data products have time and spectral resolutions of $10.24$\,$\mu$s and $183$\,kHz, respectively.
\subsubsection*{Data analysis}
\noindent {\it Calculation of burst RMs} \\
We calibrated the burst `archives' using the PSRCHIVE utility {\tt pac} in `SingleAxis' mode. This calibration strategy uses observations of a locally generated calibration signal (pulsed noise diode) to correct the relative gain and phase difference between the two polarization channels, under the assumption that the noise source emits equal power and has zero intrinsic phase difference in the two hands. This calibration scheme does not correct for cross-coupling or leakage between the polarizations. While leakage must be present at some level, the high polarization fraction, complete lack of circular polarization, and consistency of the test pulsar observations with previous work all give us confidence that calibration issues are not a significant source of error for the RM determination. In addition, the flux density of GBT observations was calibrated using the flux calibrator.
We initially performed a brute force search for peaks in the linear polarization fraction (Extended Data Fig.~\ref{fig:LI_RM}), and discovered $\text{RM}_{\rm obs} \sim +10^5$\,rad\,m$^{-2}$ in the Arecibo data.
Each burst was corrected for Faraday rotation using the best-fit RM value for that burst.
Residual variations in the resulting PA($\lambda$) were used to refine the initial values by fitting
\begin{eqnarray}
\text{PA}(\lambda) = \text{RM}\lambda^2 + \text{PA}_{\infty}.
\end{eqnarray}
Subsequently, the equation
\begin{eqnarray}
\hat{L}=\exp\left\lbrace i\cdot 2(\text{RM}\lambda^2+\text{PA}_{\infty})\right\rbrace,
\end{eqnarray}
where $\hat{L}$ is the unit vector of the linear polarization, was used to fit the whole sample of bursts together, imposing a different RM per day and a different PA$_{\infty}$ per telescope.
The results of these fits are reported in Table~\ref{tab:bursts} and an example is shown in Fig.~\ref{fig:UQ}.
Applying the optimal RM value to each burst, we produced polarimetric profiles showing that each burst is consistent with being $\sim$100\% linearly polarized after accounting for the finite widths of the PUPPI frequency channels (Fig.~\ref{fig:bursts_zoom}; Extended Data Fig.~\ref{fig:LI_f}).
In fact, the measured Arecibo bursts are depolarized to $\sim$98\%, consistent with an uncorrected intra-channel Faraday rotation of
\begin{eqnarray}\label{eq:faraday_smearing}
\Delta\theta = \frac{{\rm RM}_{\rm obs}c^2\Delta\nu}{\nu_c^{3}},
\end{eqnarray}
where $c$ is the speed of light, $\Delta{\nu}$ is the channel width, and $\nu_c$ is the central channel observing frequency. At 4.5\,GHz this corresponds to $\sim$9$^{\circ}$, and the depolarization fraction is
\begin{eqnarray}\label{eq:depol}
f_{\rm depol} = 1 - \frac{{\rm sin}(2\Delta{\theta})}{2\Delta{\theta}} = 1.6\%.
\end{eqnarray}
We supplemented our above analysis with a combination of RM Synthesis and {\tt RMCLEAN} (e.g. Extended Data Fig.~\ref{fig:rmclean}).
Ensuring the presence of minimal Faraday complexity is possible by integrating across the full bandwidth and taking advantage of a Fourier transform relation between the observed $\vec{L}(\lambda^2)$ values and the Faraday spectrum (the polarized brightness as a function of RM). This approach is commonly known as RM Synthesis\cite{Bre05}, and can be coupled with a deconvolution procedure ({\tt RMCLEAN}) to estimate the intrinsic Faraday spectrum\cite{Hea09}. While RM Synthesis and {\tt RMCLEAN} can have difficulty in properly reconstructing the intrinsic Faraday spectrum under certain circumstances, the spread of clean components is a reliable indicator of spectra that contain more than a single Faraday-unresolved source\cite{And15}.
At each observed frequency, we integrated Stokes Q and U values across the pulse width and normalized using Stokes I. Due to the normalization we only used frequency bins that had a Stokes I signal-to-noise ratio of at least 5. We computed a deconvolved Faraday spectrum for each burst separately, on a highly oversampled RM axis ($\delta\mathrm{RM}\approx\,10^{-4}$ of the nominal FWHM of the RM resolution element). We used a relatively small gain parameter (0.02) and terminated the deconvolution when the peak of the residual decreased to $2\sigma$ above the mean. The algorithm typically required $50-80$ iterations to converge. This combination of settings permits us to carefully consider the cumulative distribution of {\tt RMCLEAN} components along the RM axis, and thus constrain the intrinsic width of the polarized emission to $\lesssim0.1\%$ of the typical RM uncertainty. We found that this value scales with $\delta\mathrm{RM}$ because the peak of the Faraday spectrum rarely lands precisely on an individual pixel. To a high degree of confidence, there is no evidence for emission at more than one RM value, nor for a broadened (``Faraday thick'') emitting region; we therefore forego more complicated QU-fitting\cite{O'Sul12}. Results of this analysis are shown in Extended Data Table~\ref{tab:rm_synth}, and are consistent with the simplified QU-fitting results described above.
\noindent {\it Calculation of burst properties}
As in previous studies\cite{Spi16,Sch16}, a search for periodicity in the burst arrival times remains inconclusive.
Determining the exact DMs of the bursts is complicated by their changing morphology with radio frequency\cite{Spi16,Sch16}. Measuring DM based on maximizing the peak S/N of the burst often leads to the blurring of burst structure and, in the case of FRB\,121102, an overestimation of DM. We have thus chosen to display all bursts dedispersed to the same nominal DM from Burst \#6 (Fig.~\ref{fig:bursts_zoom} and Extended Data Fig.~\ref{fig:bursts}).
Taking advantage of the narrowness of Burst \#6, we estimated its optimal DM by minimizing its width at different DM trials.
We measured burst widths at half the maximum by fitting von Mises functions using the PSRCHIVE routine {\tt paas} (Table~\ref{tab:bursts}).
These widths correspond to the burst envelope in the case of multi-component bursts.
Flux densities of the Arecibo bursts were estimated using the radiometer equation to calculate the equivalent RMS flux density of the noise:
\begin{eqnarray}
\sigma_{\rm noise} = \frac{T_{\rm sys}}{G\sqrt{2Bt_{\rm int}}},
\end{eqnarray}
where $T_{\rm sys} \sim 30$\,K and $G \sim 7$\,K\,Jy$^{-1}$ are the system temperature and gain of the receiver, respectively, $B=800$\,MHz is the observing bandwidth and $t_{\rm int} = 10.24$\,$\mu$s is the sampling time.
GBT observations were instead calibrated using a flux calibrator as discussed above.
Due to the complicated spectra of the bursts, we quote average values across the frequency band (Table~\ref{tab:bursts}).
The burst dynamic spectra in Extended Data Fig.~\ref{fig:bursts} show narrow-band striations that are consistent with diffractive interstellar scintillations caused by turbulent plasma in the Milky Way. Autocorrelation functions (ACFs) of burst spectra show three features: a very narrow feature from radiometer noise, a narrow but resolved feature corresponding to the striations, and a broad feature related to the extent of the burst across the frequency band. The striation feature has a half width that varies from $\sim$2 to 5\,MHz from burst to burst and is comparable to the scintillation bandwidth expected from the Milky Way in the direction of FRB\,121102. The NE2001 electron density model\cite{Cor02} provides an estimate $\tau \sim 16~\mu$s for the pulse broadening at 1 GHz. This predicts a scintillation bandwidth $\sim \nu^{4.4} / 2\pi\tau$ that ranges from 5 to 11~MHz across the 4.1 to 4.9~GHz band.
We conclude that the measured ACFs and the NE2001 model prediction are consistent to within their uncertainties and that the narrow striations are due to Galactic scintillations.
\subsubsection*{A model for FRB\,121102's rotation measure (RM) and scattering measure (SM)}
\noindent {\it RM constraints}
The measured RM$_{\rm obs} \sim +1\times10^5\ \rm rad \ m^{-2}$ implies a source frame value
\begin{eqnarray}
{\rm RM}_{\rm src} = (1+z)^2 {\rm RM}_{\rm obs} \sim +1.4\times 10^5 \ \rm rad \ m^{-2}.
\end{eqnarray}
We can use the previously estimated\cite{Ten17} $\text{DM}_{\rm Host} \sim 70$--270\,pc\,cm$^{-3}$ (in the source frame) and RM$_{\rm src}$ to constrain the properties of the region in which the Faraday rotation
occurs. In the absence of other information, we can set a constraint on the average magnetic field
along the line of sight in the Faraday region with the ratio
\begin{eqnarray}
B_{\parallel} = \frac{{\rm RM}_{\rm src}}{0.81 {\rm DM}_{\rm Host}} = [0.6\ {\rm mG}, 2.4 \ {\rm mG}].
\end{eqnarray}
If only a small portion of FRB\,121102's total DM is from the highly magnetized region, the field could be much higher.
\noindent {\it SM constraints}
The best constraint on pulse broadening comes from the measurement of the scintillation (diffraction) bandwidth
of $\Delta\nu_{\rm d} \sim 5$~MHz at 4.5\,GHz (see above). This implies a pulse broadening time at 1\,GHz:
\begin{eqnarray}
\tau(1\ {\rm GHz}) \approx (2\pi \Delta\nu_{\rm d})^{-1} \times (4.5~{\rm GHz}/1~{\rm GHz})^{4.4} = 24\ \mu s.
\end{eqnarray}
This scattering time is consistent with that expected from the Milky Way using the NE2001 model\cite{Cor02} and therefore is an upper bound on any contribution from the host galaxy. Compared to scattering in the Milky Way, this upper bound is below the mean trend for any of the plausible values of DM$_{\rm Host}$, especially when the correction from spherical to plane waves is taken into account\cite{Cor16b}.
The ratio host-galaxy $\tau / {\rm DM}$ is a factor $(1+z)^2 = 1.42 $ larger in the source frame but that is still far from sufficient to account for the apparent scattering deficit compared to the Galactic $\tau$-DM relation. Given the apparent extreme conditions of the plasma in the host galaxy, it would not be surprising if its turbulence properties cause a scattering deficit. For example, scattering is reduced if the inner scale is comparable or larger than the Fresnel scale, either due to a large magnetic field or a high temperature.
\subsubsection*{Constraints on the properties of the Faraday region}
Comparison of the magnetic field and thermal energy densities enables us to constrain the density ($n_e$), electron temperature ($T_e$), and length scale ($L_{\rm RM}$) of the region responsible for the observed Faraday rotation.
We parametrize this relation with
\begin{equation}
\beta \frac{B^2}{8\pi} = 2 n_e k_B T_e,
\end{equation}
where $\beta$ is a scaling factor, $B$ is the magnetic field strength, and $k_B$ is the Boltzmann constant. This assumes a 100\% ionized gas of pure hydrogen with temperature equilibration between protons and electrons. Under equipartition, $\beta=1$. In more densely magnetized regions, $\beta \ll 1$. Field reversals will reduce the total RM, requiring a lower value of $\beta$ in order to match constraints. The absence of free-free absorption at a frequency of $\sim$1\,GHz sets an additional constraint on the permitted parameter space.
In Extended Data Fig.~\ref{fig:dmrm}, we explore a range of physical environments. We consider a smaller lower limit, i.e. ${\rm DM} = 1$\,pc\,cm$^{-3}$, on the dispersion measure than the previously estimated\cite{Ten17} $\text{DM}_{\rm Host} \sim 70$--270\,pc\,cm$^{-3}$, because not all of the DM may originate from the Faraday region. Galactic HII regions typically show $| {\rm RM} | \lesssim 3 \times 10^2$\,rad\,m$^{-2}$ and weak magnetic fields\cite{Har11} with $\beta \gtrsim 1$, although calculations suggest it is possible for HII regions to achieve high RMs under some circumstances\cite{Sic17}. Parameter space for typical HII region plasma at $T_e=10^4$\,K is almost entirely excluded, and considering a range of possible HII regions sizes and densities\cite{Hunt09} shows that these are incompatible with the $\text{DM}_{\rm Host}$ constraints. At higher $T_e$, wide ranges of parameter space are permitted. In the case of equipartition, we have explicit unique solutions. For $T_e=10^6$\,K, we find a density of $n_e\sim 10^2\,{\rm cm^{-3}}$ on a length scale $L_{\rm RM}\sim 1$\,pc, i.e., comparable to the upper limit on the size of the persistent source. Higher temperature gas ($T_e=10^8$\,K) can be extended to $L_{\rm RM}\sim 100$\,pc. For both of these solutions, the characteristic magnetic field strength is $\sim$1\,mG.
The large RM of FRB\,121102\ is similar to those seen toward massive black holes; notably, $\text{RM} \sim -5 \times 10^5$\,rad\,m$^{-2}$ is measured toward Sgr~A*, the Milky Way's central black hole, and probes scales of $< 10^4$ Schwarzschild radii ($\sim$0.001\,pc)\cite{Bow03,Mar07}. The constraints on $n_e$, $T_e$, and $L_{\rm RM}$ are also consistent with the environment around Sgr~A* (Extended Data Fig.~\ref{fig:dmrm}). The high RM toward the Galactic Centre magnetar
PSR~J1745$-$2900 (Fig.~\ref{fig:DM_RM_all}), $\text{RM} = -7 \times 10^4$\,rad\,m$^{-2}$,
at a projected distance of $\sim$0.1\,pc from Sgr~A* \cite{Eat13,Sha13}, is evidence
for a dynamically organized magnetic field around Sgr~A* that extends out to the magnetar's distance\cite{Eat13}. Notably, $\sim$4.5 years of radio monitoring of PSR~J1745$-$2900 has shown a $\sim$5\% decrease in the magnitude of the observed RM, while the DM remained constant at the $\sim$1\% level (Desvignes et al., in prep.). This suggests large fluctuations in magnetic field strength in the Galactic Centre, on scales of roughly $10^{-5}$ parsec.
The high RM and the rich variety of other phenomena\cite{Spi14,Spi16,Sch16,Cha17,Ten17,Mar17,Sch17,Bas17} displayed by the FRB\,121102\ system suggest that the persistent radio counterpart to FRB\,121102\ could represent emission from an accreting massive black hole, with the surrounding star formation representing a circum-black-hole starburst. Given the mass of the host galaxy, and typical scaling relationships\cite{Rei15}, the mass of the black hole would be $\sim$10$^{4-6}$\,M$_\odot$. The observed radio brightness, compactness, and the current optical and X-ray
non-detections\cite{Ten17,Mar17,Sch17} are compatible with such a black hole and an inefficient accretion state ($\sim$10$^{-6}$--$10^{-4}$\,$\mathrm{L_{Edd}}$, where $L_{\rm Edd}$ is the Eddington luminosity).
While models considering the presence of only a massive black hole have been proposed\cite{Vie17}, there is no observational precedent for microsecond bursts created in such environments.
Rather, the FRB\,121102\ bursts themselves could arise from a neutron star, perhaps highly magnetized and rapidly spinning, near an accreting massive black hole. The proximity of PSR J1745$-$2900 to Sgr A* demonstrates that such a combination is
possible. In this model, the black hole is responsible for the observed persistent source, whereas the bursts are created in the magnetosphere of the nearby neutron star\cite{Pen15}.
Alternatively, the association of FRB\,121102\ with a persistent radio source has been used to argue that the radio bursts are produced by a young magnetar powering a luminous wind nebula\cite{Met17,Kas17}. This model is not well motivated by Galactic examples, since the most luminous (non-magnetar powered) Galactic pulsar wind nebula is only $2\times 10^{-6}$ times as luminous as the persistent source coincident with FRB\,121102, and Galactic magnetars have no detectable persistent radio wind nebulae\cite{Hes08,Rey17}. Also, while giant flares from magnetars can produce relativistic outflows\cite{Fra99}, an upper limit on the RM from one such outburst\cite{Gae05} is 4 orders of magnitude below that observed for FRB\,121102.
Nonetheless, under the millisecond magnetar model, the properties of the persistent source constrain the putative magnetar's age to be between several years and several decades with a spin-down luminosity of $10^8$~to $10^{12}$ times higher than any local analog\cite{Met17}. Furthermore, the millisecond magnetar model predicts that the nebula magnetic field strength scales with the integrated spin-down luminosity of the magnetar\cite{Met17,Kas17}.
Extended Data Fig.~\ref{fig:dmrm} describes a range of sizes, densities, and temperatures for the Faraday-rotating medium that are consistent with Crab-like pulsar wind nebulae, known supernova remnants, and a simple model for swept-up supernova ejecta.
\subsubsection*{Data availability}
The calibrated burst data are available, upon request, from the Corresponding Author.
\subsubsection*{Code availability}
The code used to analyse the data is available at the following sites:\\
PRESTO (\url{https://github.com/scottransom/presto}),\\
PSRCHIVE (\url{http://psrchive.sourceforge.net}),\\
DSPSR (\url{http://dspsr.sourceforge.net}).
\clearpage
|
\section{Introduction}\label{Section1}
Internet of Things (IoT) and mobile devices, such as sensors and smart-phones, are typically powered by batteries that have limited operation time. Sensors for IoT, especially sensors that being deployed in special environments such as volcanoes, are difficult to be charged. Meanwhile, carrying a power cord and looking for a power supplier to charge mobile devices incur great inconvenience. An alternative is thus to transfer power wirelessly, which virtually provides perpetual energy supplies. Hence, wireless power transfer (WPT) or wireless charging attracts great attention recently.
\begin{figure}
\centering
\includegraphics[scale=0.4]{adhocnetwork.pdf}
\caption{Distributed Laser Charging Applications}
\label{adhocnetwork}
\end{figure}
Three major wireless charging technologies are surveyed in \cite{wirelesstechniques,electromagnetic}. Inductive coupling is safe and simple for implementation. However, it is limited by a short charging distance from a few millimeters to centimeters, which is only suitable for contact-charging devices like toothbrush. Magnetic resonance coupling has high charging efficiency. However, it is restricted by short charging distances and big coil sizes, which fits home appliances like TV. Electromagnetic (EM) radiation has long effective charging distances. However, it suffers from low charging efficiency and is unsafe when the EM power density exposure is high, hence is only favorable for low-power devices like sensors. In a nutshell, these traditional WPT technologies provide great wireless charging abilities for different application scenarios, whereas it is still challenging to offer sufficient power over long distance for safely charging IoT and mobile devices, e.g., smart-sensor, smart-phone, laptop, drone, etc., which usually need Watt-level power over meter-level distances.
To support the power and distance requirements for IoT and mobile devices, a distributed laser charging (DLC) system is presented in \cite{liu2016dlc}, which could transfer 2-Watt power over a 5-meter distance \cite{wi-charge}. By using inductive coupling or magnetic resonance coupling, IoT and mobile devices, say sensors and smart-phones, should typically be placed in a special charging cradle with a particular position. However, the DLC's self-aligning feature provides a more convenient way of charging IoT and mobile devices without specific positioning or tracking, as long as the transmitter and the receiver are in the line of sight (LOS) of each other. Different from EM radiation, DLC's wireless power transfer can be stopped immediately when this LOS is blocked by any object, which ensures the safety of DLC system. The size of the DLC receiver is sufficiently small to be embedded in a sensor or a smart-phone. The DLC transmitter can be installed on the ceiling like a lightbulb. In addition, multiple devices can be charged simultaneously by a single DLC transmitter \cite{niu2013optimal,anna2003opportunistic,zhou2004performance}. Therefore, DLC can provide IoT and mobile devices with safe WPT capability, which enables people to charge their devices with the similar experience as WiFi communications.
Fig.~\ref{adhocnetwork} illustrates the DLC potential applications.
In Fig.~\ref{adhocnetwork}, in the room, DLC Transmitter-1 is combined with a light-emitting diode (LED) array and become a DLC-equipped lightbulb. Thus Transmitter-1 can be conveniently installed on the ceiling, and then provide wireless power to IoT and mobile devices within its coverage. In the outdoor scenario, Drone-1 is equipped with a DLC transmitter, which can charge IoT and mobile devices on demand. At the same time, a DLC receiver is also embedded in Drone-1. Thus, it can be remotely charged by DLC Transmitter-2, which acts as the power-supply base station on the ground. In addition, Drone-2 equipped with both DLC transmitter and receiver can play the role of a relay to receive power from DLC Transmitter-2 and transmit power to Drone-1 simultaneously.
Similar to the maximization of the \emph{information transmission capacity} of wireless channels in wireless information transfer (WIT), an important research topic in WPT is to maximize the \emph{power or energy transmission efficiency} \cite{zhangruimimo}. The wireless charging efficiency of a DLC system is affected by many factors, including laser wavelength, electricity-to-laser conversion efficiency, laser transmission attenuation, and laser-to-electricity conversion efficiency \cite{810nmtransmitter,1550nmtransmitter,laserenergy2009,green2015solar}. In this paper, we focus our study on the modeling of DLC system and its performance evaluation. In order to understand the fundamental mechanism of DLC system, we separate the DLC system into multiple conceptually independent modules. Thus, the corresponding power conversion or transmission for each module can be investigated individually, considering the impacts of laser wavelength, transmission attenuation, and photovoltaic-cell (PV-cell) temperature. Finally, the maximum power transmission efficiency in closed-form can be obtained from this modular analysis.
In this paper, a multi-module system model is proposed to describe the DLC system. The physical mechanism and mathematical formula are presented to describe the relationship between the stimulating electrical power and the output power, as well as the efficiency. The relationship between the supply power and the laser power, the relationship between the received laser power and the output power, and thus the relationship between the output power and the supply power are all depicted by both analytical results and illustrative graphs. The relationship between the electricity-to-laser conversion efficiency and the supply power, the relationship between the laser-to-electricity conversion efficiency and the received laser power, and thus the relationship between the maximum power transmission efficiency and the supply power are captured by closed-form expressions as well as being illustrated by figures. As a result, this work not only provides the insight of DLC in theory, but also offers the design guideline for DLC system implementation in practice.
In the rest of this paper, we will first review the DLC system and present the multi-module system model. Then, we will illustrate the analytical modeling of each module to investigate the corresponding working principles. After that, we will evaluate the performance of each module and derive the maximum DLC power transmission efficiency in closed-form. Finally, we will give summarizing remarks and discuss open issues for future research.
\section{DLC system}\label{Section2}
DLC is a WPT technology based on the distributed resonating laser presented in \cite{liu2016dlc}. Traditional laser systems belong to the scope of \emph{integrated resonating laser}, since all optical components are integrated in one single device. However, in DLC systems, the optical components are divided into two separate parts, the transmitter and the receiver, respectively. Therefore, the laser in DLC systems falls within the scope of \emph{distributed resonating laser}.
\begin{figure}
\centering
\includegraphics[scale=0.43]{DLCmodule.pdf}
\caption{Distributed Laser Charging System Diagram}
\label{adaptivepower}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.43]{Overallefficiency.pdf}
\caption{Distributed Laser Charging System Model}
\label{effieiencyprocedure}
\end{figure}
Fig.~\ref{adaptivepower} shows the DLC system diagram described in \cite{liu2016dlc}. A retro-reflector mirror R1 with 100\% reflectivity and a gain medium are implemented at the transmitter. While in the receiver, a retro-reflector mirror R2 with exemplary 95\% partial reflectivity is contained. R1, R2 and the gain medium consist the resonant cavity, within which photons are amplified and form intra-cavity resonating laser. Photons that pass through R2 generates the external-cavity laser. The external-cavity laser power can be converted to electrical power by a photovoltaic-panel (PV-panel) installed behind mirror R2, which is similar to a solar panel. Fig.~\ref{adaptivepower} includes the power supplier at the transmitter and the power output at the receiver for the comprehensive DLC system design.
As specified in \cite{liu2016dlc}, in the DLC system, photons is amplified without concerning about the incident angle, as long as they travel along LOS of R1 and R2. Hence, the intra-cavity laser generated by the resonator can be self-aligned without specific positioning or tracking. This feature enables users to charge their devices without placing them in a specific position cautiously. Besides self-alignment, the DLC system is intrinsically-safe, since objects blocking the line-of-sight of intra-cavity laser can stop the laser immediately. These features offer DLC the capability of safely charging devices over long distance.
Fig. \ref{effieiencyprocedure} presents the system model to elaborate the wireless power transfer in the DLC system. This model illustrates a theoretical framework of power transfer by electricity-to-laser conversion, laser transmission, and laser-to-electricity conversion. The physical fundamentals and mathematical formulations of this modular model will be specified in the following section.
\section{Analytical Modeling}\label{Section3}
In this section, we will discuss each module of the DLC model in Fig.~\ref{effieiencyprocedure} and describe its wireless power transfer mechanism analytically. At the DLC transmitter, the power supplier provides electrical power to generate the intra-cavity laser. We will first introduce the electricity-to-laser conversion. Then, the intra-cavity laser will travel through the air and arrive at the DLC receiver. We will discuss the intra-cavity laser power attenuation along its transmission. At the DLC receiver, the intra-cavity laser will partially go through the mirror R2 and form the external-cavity laser, then the external-cavity laser will be converted into electricity by a PV-panel. We will analyze this laser-to-electricity conversion based on the PV engineering. Finally, the PV-panel output electrical power can be used to charge electronics. Based on the above analytical modeling, we will obtain the power conversion and transmission efficiency of each module and the overall power transmission efficiency.
\subsection{Electricity-to-Laser Conversion}\label{}
At the DLC transmitter, the electrical power $P_s$ is provided by the power supplier, which depends on the stimulating current $I_{t}$ and voltage $V_{t}$ as:
\begin{equation}\label{ps}
P_s = I_{t} V_{t}.
\end{equation}
The supply power $P_s$ can stimulate the gain medium to generate laser. Thus, the electrical power can be converted to the laser power.
We denote $P_l$ as the external-cavity laser power when the intra-cavity laser transmission efficiency is 100\%. It is well-known that laser can be generated, only when $I_{t}$ provided by the power supplier is over a certain threshold \cite{laserdiodes}. In the laser diode physics, the laser power $P_l$ relies on $I_{t}$
. Their relationship can be depicted as \cite{laserdiodes}:
\begin{equation}\label{powervscurrent}
P_{l}=\zeta \frac{h\upsilon}{q}(I_{t}-I_{th}),
\end{equation}
where $\zeta$ is the modified coefficient, $h$ is the Plunk constant, $\upsilon$ is the laser frequency, $q$ is the electronic charge constant, and $I_{th}$ is the current threshold.
Thus, the electricity-to-laser conversion efficiency $\eta_{el}$ can be figured out as:
\begin{equation}\label{etael}
\eta_{el} = \frac{P_{l}}{P_s}.
\end{equation}
\subsection{Laser Transmission}\label{}
Laser power transmission attenuation means that laser power decreases along with its transmission through the air, which is similar to EM wave propagation power loss \cite{attenuation}. The laser power attenuation level depends on the transmission distance and air quality \cite{JMLiuphotonic,foghaze}. Relying on the above laser-generation mechanism, the intra-cavity laser can transmit from the transmitter to the receiver. During the transmission, laser may experience power attenuation. For simplicity, we assume that the laser diameter is a constant. This assumption could be validated by controlling aperture diameters of the DLC transmitter and receiver \cite{JMLiuphotonic}.
The laser transmission efficiency $\eta_{lt}$ can be modeled as \cite{JMLiuphotonic}:
\begin{equation}\label{etalt}
\eta_{lt}=\frac{P_{r}}{P_l}= e^{-\alpha d},
\end{equation}
where $P_{r}$ is the external-cavity laser power received at the DLC receiver, $\alpha$ is the laser attenuation coefficient, and $d$ is the distance. When $d$ is close to zero, the laser transmission efficiency approaches 100\%. In this situation, $P_{r}$ is approximate to $P_{l}$.
$\alpha$ can be depicted as:
\begin{equation}\label{alphac}
\alpha = \frac{\sigma}{\kappa} \Big(\frac{\lambda}{\chi}\Big)^{-\rho},
\end{equation}
where $\sigma$ and $\chi$ are two constants, $\kappa$ is the visibility, $\lambda$ is the wavelength, and $\rho$ is the size distribution of the scattering particles. $\rho$ depends on visibility, which will be discussed later.
\subsection{Laser-to-Electricity Conversion}\label{}
\begin{figure}
\centering
\includegraphics[scale=0.8]{single-diodeonly.pdf}
\caption{PV-panel Power Conversion Circuit Model}
\label{singlediodeonly}
\end{figure}
\begin{table}[b]
\centering
\caption{Transmission or Conversion Efficiency}
\begin{tabular}{C{6.0cm} C{1.5cm}}
\hline
\textbf{Parameter} & \textbf{Symbol} \\
\hline
\bfseries{Electricity-to-laser conversion efficiency} & {${\eta}_{el}$} \\
\bfseries{Laser transmission efficiency} & {${\eta}_{lt}$} \\
\bfseries{Laser-to-electricity conversion efficiency} & {${\eta}_{le}$} \\
\hline
\bfseries{The overall DLC power transmission efficiency} & {${\eta}_{o}$} \\
\hline
\label{convertefficiency}
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[scale=0.6]{810pi.eps}
\caption{Electricity-to-Laser Conversion Power, Voltage and Current (810nm)}
\label{810ptI}
\end{figure}
At the DLC receiver, the external-cavity laser power can be converted to electrical power. To illustrate the laser-to-electricity conversion mechanism, the single-diode equivalent circuit model of a PV-panel is depicted in Fig.~\ref{singlediodeonly} \cite{secondauthorPV}. The PV-panel output voltage $V_{o}$, and current $I_{o}$ can be characterized as \cite{secondauthorPV}:
\begin{equation}\label{Io}
I_{o}=I_{sc} - I_{s}(e^{V_{o}/V_m}-1),
\end{equation}
where $I_{sc}$ is the PV-panel short-circuit current, $I_s$ is the saturation current, i.e., the diode leakage current density in the absence of light, and $V_m$ is the ``thermal voltage'', which can be defined as:
\begin{equation}\label{Vm}
V_m=\frac{nkT}{q},
\end{equation}
where $n$ is the PV-panel ideality factor, $k$ is the Boltzmann constant, and $T$ is the absolute PV-cell temperature. Then, the PV-panel output power $P_{o}$, which relies on $I_{o}$ and $V_{o}$, can be obtained as:
\begin{equation}\label{po}
P_{o}=I_{o}V_{o}.
\end{equation}
Therefore, the laser-to-electricity conversion efficiency, i.e. the PV-panel conversion efficiency, $\eta_{le}$ depends on $P_{o}$ and $P_{r}$, which can be depicted as:
\begin{equation}\label{etapv}
{\eta}_{le}=\frac{P_{o}}{P_{r}}=\frac{I_{o}V_{o}}{P_{r}}.
\end{equation}
In summary, the PV-panel converts the received laser power $P_{r}$ to the output power $P_{o}$ with the efficiency $\eta_{le}$.
\subsection{DLC Power Transmission Efficiency}\label{}
Based on the above analysis for each individual module of the DLC system model, the DLC power transmission efficiency from the power supplier at the transmitter to the power output at the receiver can be depicted as:
\begin{equation}\label{etao}
{\eta}_{o}={\eta}_{el} {\eta}_{lt} {\eta}_{le}.
\end{equation}
The conversion or transmission efficiency of each module and the DLC power transmission efficiency are listed in Table \ref{convertefficiency}.
The numerical evaluation of the DLC system model will be presented in the next section.
\section{Numerical Evaluation}\label{Section4}
Based on the analytical modeling in the previous section, we can find that the DLC system efficiency varies with laser wavelength, transmission attenuation and PV-cell temperature. Their impacts on the performance of each module as well as the overall DLC system will be discussed in this section. The numerical evaluation is implemented in MATLAB and Simulink.
\subsection{Electricity-to-Laser Conversion}\label{}
\begin{table}[b]
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\centering
\caption{Electricity-to-Laser Conversion Parameters}
\begin{tabular}{C{3.4cm} C{0.1cm} C{4.0cm}}
\hline
\textbf{Parameter} & \textbf{Symbol} & \tabincell{c}{\textbf{Value} \\ \textbf{810nm} \qquad \textbf{1550nm}} \\
\hline
\bfseries{Boltzmann constant} & {k} & $1.38064852\times10^{-23} J/K$ \\
\bfseries{Planck constant} & {h} & $6.62606957\times10^{-34} J\cdot s$ \\
\bfseries{Electronic charge constant} & {q} & $1.6\times10^{-19} C$ \\
\bfseries{Laser wavelength } & {$\lambda$} & \tabincell{c}{$810nm$ \qquad $1550nm$} \\
\bfseries{Laser frequency} & {$\upsilon$} & \tabincell{c}{$3.7\times10^{14} Hz$ \qquad $1.9\times10^{14} Hz$} \\
\bfseries{Stimulation current threshold} & {I$_{th}$} & \tabincell{c}{$0.5 A$ \qquad $0.6 A$}\\
\bfseries{Modified coefficient} & {$\zeta$} & \tabincell{c}{$1.5$ \qquad $3.52$} \\
\bfseries{$\textbf{P}_{\textbf{l}}$-$\textbf{P}_{\textbf{s}}$ curve fitting parameter} & {$a_1$} & \tabincell{c}{$0.445$ \qquad $0.34$} \\
\bfseries{$\textbf{P}_{\textbf{l}}$-$\textbf{P}_{\textbf{s}}$ curve fitting parameter} & {$b_1$} & \tabincell{c}{$-0.75$ \qquad $-1.1$} \\
\hline
\label{paramaters}
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[scale=0.6]{1550pi.eps}
\caption{Electricity-to-Laser Conversion Power, Voltage and Current (1550nm)}
\label{1550ptI}
\end{figure}
Electrical supply power $P_{s}$ provided by the power supplier at the transmitter depending on the stimulating current $I_{t}$ and voltage $V_{t}$, as in \eqref{ps}. Based on the measurement of $I_{t}$, $V_{t}$, and thus $P_{s}$, for the laser systems ($\lambda$ is 800-820nm and 1540-1560nm, respectively) in \cite{810nmtransmitter,1550nmtransmitter}, the measured supply power $P_{s}$, the measured laser power $P_{l}$, the stimulating current $I_{t}$, and the stimulating voltage $V_{t}$ are shown for 810nm and 1550nm in Fig.~\ref{810ptI} and Fig.~\ref{1550ptI}, respectively. From the dashed-lines for the measured laser power in Fig.~\ref{810ptI} and Fig.~\ref{1550ptI}, the modified coefficient $\zeta$ in \eqref{powervscurrent} can be determined and listed in Table \ref{paramaters}. Thus, from \eqref{powervscurrent}, the formulated laser power curves are given as the solid-lines in Fig.~\ref{810ptI} and Fig.~\ref{1550ptI}, respectively.
\begin{figure}
\centering
\includegraphics[scale=0.6]{8101550plps.eps}
\caption{Laser Power vs. Supply Power \protect\\ \qquad}
\label{8101550plps}
\end{figure}
In Fig.~\ref{810ptI} and Fig.~\ref{1550ptI}, the relationship between $P_{l}$ and $P_{s}$ is illustrated in Fig.~\ref{8101550plps}. We adopt the linear formula to approximate this power conversion as:
\begin{equation}\label{plps}
P_{l} \approx a_{1} P_{s}+b_{1}.
\end{equation}
The measured and formulated curves in Fig.~\ref{8101550plps} depict the linear approximation between $P_{l}$ and $P_{s}$ based on \eqref{plps}, when the wavelength $\lambda$ is about 810nm and 1550nm, respectively. We can find that the fitting curves match the measurement very well in the given supply power and laser power range in Fig.~\ref{8101550plps}.
From \eqref{etael} and \eqref{plps}, we can obtain the electricity-to-laser conversion efficiency $\eta_{el}$ as:
\begin{equation}\label{etaelall}
\eta_{el}= \frac{P_{l}}{P_{s}} = a_{1}+\frac{b_{1}}{P_{s}}.
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.6]{8101550etael.eps}
\caption{Electricity-to-Laser Conversion Efficiency vs. Supply Power}
\label{8101550etael}
\end{figure}
The solid-line and dashed-line in Fig.~\ref{8101550etael} illustrate $\eta_{el}$ for 810nm and 1550nm, respectively. The initial $P_{s}$ supply power threshold in Fig.~\ref{8101550etael} is corresponding to the current threshold $I_{th}$ for $P_{l}$ in Fig.~\ref{810ptI} and Fig.~\ref{1550ptI}. In Fig.~\ref{8101550etael}, $\eta_{el}$ starts to increase dramatically from the initial supply power $P_{s}$ threshold and will reach the plateau as $P_{s}$ increases. The plateau of $\eta_{el}$ for 810nm laser is around 43\%, which is higher than 31\% for 1550nm laser.
\subsection{Laser Transmission}\label{}
From \eqref{etalt} and \eqref{alphac}, the laser power attenuation coefficient in transmission can be determined under three typical scenarios, i.e., clear air, haze, and fog. For the three scenarios, the size distribution of the scattering particles $\rho$ in \eqref{alphac} can be specified as \cite{foghaze}:
\begin{equation}\label{rhoc}
\rho = \left\{
\begin{array}{lr}
1.3 \qquad\qquad\qquad\ \mathrm{for \ clear \ air} \quad\ (6km\leq \kappa \leq 50km), & \\
0.16 \kappa+0.34 \qquad\ \mathrm{for \ haze} \qquad\quad (1km\leq \kappa \leq 6km),\\
0 \qquad\qquad\qquad\quad\ \mathrm{for \ fog} \qquad\quad\ \ (\kappa \leq 0.5km), &
\end{array}
\right.
\end{equation}
where $\kappa$ is the visibility.
\begin{figure}
\centering
\includegraphics[scale=0.6]{transmittance.eps}
\caption{Laser Transmission Efficiency vs. Distance}
\label{transmittance}
\end{figure}
\begin{table}[b]
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\centering
\caption{Laser Transmission Parameters}
\begin{tabular}{C{1.0cm} C{7.0cm}}
\hline
\textbf{Parameter} &\tabincell{c}{\textbf{Value} \\ \quad\quad \textbf{Clear Air} \qquad\qquad\quad \textbf{Haze} \qquad\qquad\qquad \textbf{Fog}} \\
\hline
\bfseries{$\sigma$} & {\quad\qquad $3.92$} \\
\bfseries{$\chi$} & \quad\qquad $550nm$ \\
\bfseries{$\kappa$} & {\quad\qquad \ $10km$ \qquad\qquad\quad $3km$ \qquad\qquad\quad\ $0.4km$} \\
\bfseries{$\rho$} & {\quad\qquad $1.3$ \qquad\qquad $0.16\kappa+0.34$ \qquad\qquad $0$} \\
\hline
\label{pathloss}
\end{tabular}
\end{table}
Along with $\rho$, the other attenuation parameters are listed in Table \ref{pathloss}. Thus, the relationship between $\eta_{lt}$ and the transmission distance $d$ can be obtained from \eqref{etalt} and \eqref{alphac}, which is illustrated in Fig.~\ref{transmittance}. It is clear that $\eta_{lt}$ decays exponentially to zero as $d$ increases. Meanwhile, for the same laser wavelength, laser power attenuation depends on the visibility $\kappa$. Laser power attenuation increases when $\kappa$ decreases. As can be seen in Fig.~\ref{transmittance}, for clear air, haze and fog, given the same $d$, the laser power attenuation for short-wavelength is more than that of long-wavelength. For clear air and haze, laser attenuation for 810nm is much more than that of 1550nm. However, for fog, since $\rho$ takes 0 for both 810nm and 1550nm, the coefficient $\alpha$ has the same value. Therefore, the laser attenuation in fog does not dependent on $\lambda$.
\begin{figure}
\centering
\includegraphics[scale=0.6]{810Iirradiance.eps}
\caption{PV-panel Output Current vs. Voltage ($\lambda$ = 810nm)}
\label{810Iirradiance}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.6]{1550Iirradiance.eps}
\caption{PV-panel Output Current vs. Voltage ($\lambda$ = 1550nm)}
\label{1550Iirradiance}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.6]{810Pirradiance.eps}
\caption{PV-panel Output Power vs. Voltage ($\lambda$ = 810nm)}
\label{810Pirradiance}
\end{figure}
\begin{table
\newcommand{\tabincell}[2]{\begin{tabular}{@{}#1@{}}#2\end{tabular}}
\centering
\caption{Laser-to-Electricity Conversion Parameters}
\begin{tabular}{C{1.7cm} C{0.3cm} C{4.5cm}}
\hline
\textbf{Parameter} & \textbf{Symbol} & \tabincell{c}{\textbf{Value} \\ \textbf{810nm} \qquad \textbf{1550nm}} \\
\hline
\bfseries{Short-circuit current} & {$I_{sc}$} & \tabincell{c}{$0.16732 A$ \qquad $0.305 A$\qquad\ }\\
\bfseries{Open-circuit voltage} & {$V_{oc}$} & \tabincell{c}{$1.2 V$ \qquad $0.464 V$} \\
\bfseries{Irradiance used for measurement} & {$I_{r0}$} & \tabincell{c}{$36.5 W/cm^2$ \qquad\qquad $2.7187 W/cm^2$} \\
\bfseries{Laser frequency} & {$\upsilon$} & \tabincell{c}{$3.7037\times10^{14} Hz$ \quad $1.9355\times10^{14} Hz$} \\
\bfseries{Quality factor} & {$n$} & \tabincell{c}{$1.5$ \qquad\quad $1.1\qquad$} \\
\bfseries{Number of series cells} & {$N$} & $72$ \\
\bfseries{PV-panel material } & { } & \tabincell{c}{GaAs-based \qquad GaSb-based} \\
\bfseries{Measurement temperature} & {$T$} & \tabincell{c}{$25^{\circ}C$ \qquad\ $120^{\circ}C$} \\
\bfseries{Simulation temperature} & {} & \tabincell{c}{$0^{\circ}C$ / $25^{\circ}C$ / $50^{\circ}C$} \\
\bfseries{$\textbf{P}_{\textbf{m}}$-$\textbf{P}_{\textbf{r}}$ curve fitting parameter} & {$a_2$} & \tabincell{c}{0.546/0.541/0.537 \ 0.543/0.498/0.453} \\
\bfseries{$\textbf{P}_{\textbf{m}}$-$\textbf{P}_{\textbf{r}}$ curve fitting parameter} & {$b_2$} & \tabincell{c}{-0.213/-0.231/-0.249 \ -0.276/-0.299/-0.321} \\
\hline
\label{pvparamaters}
\end{tabular}
\end{table}
\subsection{Laser-to-Electricity Conversion}\label{}
At the DLC receiver, PV-panel takes the role of converting laser power to electrical power. PV-panel conversion efficiency relies on laser power, wavelength, and cell temperature. With reference to \eqref{Io}-\eqref{Vm}, we can obtain the PV-panel output current, voltage, and thus power, given the parameters listed in Table \ref{pvparamaters}. Fig.~\ref{810Iirradiance}-\ref{1550Ptemperature} demonstrate their relationships for different laser wavelength using the standard \emph{solar cell} Simulink model \cite{solarcell}.
Fig.~\ref{810Iirradiance} shows the relationship between PV-panel output current $I_{o}$ and voltage $V_{o}$ with different input laser power, i.e., the external-cavity laser power $P_{r}$ at the receiver, for the GaAs-based PV-panel with 810nm laser at 25$^{\circ}$C \cite{810nmpv}. Similarly, Fig.~\ref{1550Iirradiance} is for the GaSb-based PV-panel with 1550nm laser at 25$^{\circ}$C \cite{1550nmpv}. The PV-panel output power $P_{o}$ can be derived from the corresponding $I_{o}$ and $V_{o}$ based on Fig.~\ref{810Iirradiance} and Fig.~\ref{1550Iirradiance}. Thus, Fig.~ \ref{810Pirradiance} and Fig.~\ref{1550Pirradiance} depict the relationship between $P_{o}$ and $V_{o}$ for 810nm and 1550nm, respectively.
\begin{figure}
\centering
\includegraphics[scale=0.6]{1550Pirradiance.eps}
\caption{PV-panel Output Power vs. Voltage ($\lambda$ = 1550nm)}
\label{1550Pirradiance}
\centering
\includegraphics[scale=0.6]{810Itemperature.eps}
\caption{PV-panel Output Current vs. Voltage ($\lambda$ = 810nm)}
\label{810Itemperature}
\centering
\includegraphics[scale=0.6]{1550Itemperature.eps}
\caption{PV-panel Output Current vs. Voltage ($\lambda$ = 1550nm)}
\label{1550Itemperature}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.6]{810Ptemperature.eps}
\caption{PV-panel Output Power vs. Voltage ($\lambda$ = 810nm)}
\label{810Ptemperature}
\centering
\includegraphics[scale=0.6]{1550Ptemperature.eps}
\caption{PV-panel Output Power vs. Voltage ($\lambda$ = 1550nm)}
\label{1550Ptemperature}
\centering
\includegraphics[scale=0.6]{810prpm.eps}
\caption{Maximum Output Power vs. Received Laser Power ($\lambda$ = 810nm)}
\label{810prpm}
\end{figure}
From Fig.~\ref{810Pirradiance} and Fig.~\ref{1550Pirradiance}, given $P_{r}$, we can figure out the maximum output power, which is defined as the maximum power point (MPP) and marked by the dots on the corresponding output power curves. We denote $P_m$ as the MPP of $P_o$. From \cite{onlyMPP2}, $P_m$ is proved as the unique output power, i.e., the corresponding current and voltage are unique, given the received laser power $P_r$. For example, given $P_{r} = 10$W, the MPP is unique as 4.64W for 1550nm, which is depicted by the dots in Fig.~\ref{1550Iirradiance} and \ref{1550Pirradiance}. The corresponding unique $I_{o}$ and $V_{o}$ are 121.3mA and 38.3V, respectively.
In Fig.~\ref{810Iirradiance} and Fig.~\ref{1550Iirradiance}, given $P_{r}$, $I_{o}$ keeps almost a constant when $V_{o}$ is below the MPP. However, $I_{o}$ drops rapidly when $V_{o}$ is over the MPP. For the same $V_{o}$, $I_{o}$ increases when $P_{r}$ increases. When $I_{o}$ is close to zero, $V_{o}$ is the open-circuit voltage, which increases when $P_{r}$ increases. From Fig.~\ref{810Pirradiance} and Fig.~\ref{1550Pirradiance}, given $P_{r}$, $P_{o}$ increases when $V_{o}$ increases until it reaches the MPP. However, $P_{o}$ drops dramatically when $V_{o}$ is above the corresponding voltage for MPP. For a given voltage $V_{o}$, the output power $P_{o}$ increases when the input laser power $P_{r}$ increases.
Besides input laser power, PV-cell temperature also impacts the PV-panel output current, voltage, and power. Given the three cell temperatures (0$^{\circ}$C, 25$^{\circ}$C, 50$^{\circ}$C), for $\lambda$ = 810nm and $P_r$ = 10W power, Fig.~\ref{810Itemperature} and Fig.~\ref{810Ptemperature} depict the variation of $I_{o}$ and $P_{o}$ on different $V_{o}$, respectively. Similarly, for $\lambda$ = 1550nm and $P_r$ = 10W power, Fig.~\ref{1550Itemperature} and Fig.~\ref{1550Ptemperature} show the PV-panel output $I_{o}$, $V_{o}$ and $P_{o}$ for these cell temperatures.
\begin{figure}
\centering
\includegraphics[scale=0.6]{1550prpm.eps}
\caption{Maximum Output Power vs. Received Laser Power ($\lambda$ = 1550nm)}
\label{1550prpm}
\end{figure}
From Fig.~\ref{810Itemperature} and Fig.~\ref{1550Itemperature}, $I_{o}$ keeps almost as a constant when $V_{o}$ is below a certain value. Given different cell temperatures, $I_{o}$ curves start dropping at different $V_{o}$. The turning voltage is low when the temperature is high. From Fig.~\ref{810Ptemperature} and Fig.~\ref{1550Ptemperature}, $P_{o}$ is low when the temperature is high. Additionally, the MPP increases as the cell temperature declines.
\begin{figure}
\centering
\includegraphics[scale=0.6]{810etale.eps}
\caption{Maximum Laser-to-Electricity Conversion Efficiency vs. Received Laser Power ($\lambda$ = 810nm)}
\label{810etale}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.6]{1550etale.eps}
\caption{Maximum Laser-to-Electricity Conversion Efficiency vs. Received Laser Power ($\lambda$ = 1550nm)}
\label{1550etale}
\end{figure}
Based on the MPP dots in Fig.~\ref{810Iirradiance} and Fig.~\ref{810Pirradiance} for different $P_r$ and Fig.~\ref{810Itemperature} and Fig.~\ref{810Ptemperature} for different cell temperatures, we can obtain the MPP dots in Fig.~\ref{810prpm}, which illustrates $P_m$ versus $P_r$ for 810nm. Similarly, Fig.~\ref{1550prpm} demonstrates $P_m$ versus $P_r$ for 1550nm. In order to evaluate the relationship between $P_m$ and $P_r$, we adopt the approximation formula by using the curve fitting method as:
\begin{equation}\label{prpmf}
P_{m} \approx a_2 P_{r} + b_2,
\end{equation}
where $a_2$ and $b_2$ are the linear curve fitting coefficients for different wavelengths and cell temperatures, which are listed in Table \ref{pvparamaters}. From Fig.~\ref{810prpm} and Fig.~\ref{1550prpm}, we can find that the approximate lines based on \eqref{prpmf} matches the MPP dots very well.
We denote $\eta_{lem}$ as the maximum PV-panel conversion efficiency when $P_o$ is $P_m$. Based on \eqref{etapv} and \eqref{prpmf}, $\eta_{lem}$ can be depicted as:
\begin{equation}\label{etalem}
\eta_{lem} = \frac{P_{m}}{P_{r}}= a_2+\frac{b_2}{P_{r}}.
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.6]{810pspm.eps}
\caption{Maximum Output Power vs. Supply Power ($\lambda$ = 810nm)}
\label{810pspm}
\centering
\includegraphics[scale=0.6]{1550pspm.eps}
\caption{Maximum Output Power vs. Supply Power ($\lambda$ = 1550nm)}
\label{1550pspm}
\centering
\includegraphics[scale=0.6]{810etao.eps}
\caption{Maximum Power Transmission Efficiency vs. Supply Power ($\lambda$ = 810nm)}
\label{810etao}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale=0.6]{1550etao.eps}
\caption{Maximum Power Transmission Efficiency vs. Supply Power ($\lambda$ = 1550nm)}
\label{1550etao}
\centering
\includegraphics[scale=0.6]{etaodc.eps}
\caption{Maximum Power Transmission Efficiency vs. Distance (Clear Air)}
\label{etaodc}
\centering
\includegraphics[scale=0.6]{etaodh.eps}
\caption{Maximum Power Transmission Efficiency vs. Distance (Haze)}
\label{etaodh}
\end{figure}
Fig.~\ref{810etale} and Fig.~\ref{1550etale} show how $\eta_{lem}$ varies with the received laser power $P_{r}$ for 810nm and 1550nm, respectively. From Fig.~\ref{810etale} and Fig.~\ref{1550etale}, the changing trend of $\eta_{lem}$ is similar with that of $\eta_{el}$ in Fig.~\ref{8101550etael}.
$\eta_{lem}$ is low when cell temperature is high. The impact of cell temperature on $\eta_{lem}$ is bigger for 1550nm than that of 810nm, comparing Fig.~\ref{810etale} and Fig.~\ref{1550etale}.
\subsection{DLC Power Transmission Efficiency}\label{}
The relationship between the laser power $P_l$ and the supply power $P_s$ is demonstrated by \eqref{plps}, when the transmission distance $d$ is close to zero in the electricity-to-laser conversion. The relationship between $P_l$ and the received laser power $P_r$ due to laser transmission is illustrated in \eqref{etalt}. The relationship between $P_r$ and the maximum PV-panel output power $P_m$ in the laser-to-electricity conversion is shown in \eqref{prpmf}. Thus, from \eqref{plps}, \eqref{etalt} and \eqref{prpmf}, we can obtain the relationship between $P_s$ at the transmitter and $P_m$ at the receiver as:
\begin{equation}\label{gspspm}
\begin{aligned}
P_{m} = a_2 \eta_{lt} P_l+ b_2 \qquad\qquad\qquad\ \\
= a_1 a_2 \eta_{lt} P_{s} + (a_2 b_{1} \eta_{lt} + b_2).
\end{aligned}
\end{equation}
Fig.~\ref{810pspm} depicts the linear relationship between $P_m$ and $P_s$ for $\eta_{lt}$ = 100\% and $\eta_{lt}$ = 50\%, respectively, when PV-cell temperature is 0$^{\circ}$C, 25$^{\circ}$C, 50$^{\circ}$C, and $\lambda$ = 810nm. Meanwhile, Fig.~\ref{1550pspm} illustrates the similar circumstances for $\lambda$ = 1550nm.
From \eqref{etao}, we denote $\eta_{om}$ as the maximum power transmission efficiency, when $P_{o}$ is $P_{m}$, i.e. $\eta_{le}$ approaches $\eta_{lem}$. From \eqref{etael}, \eqref{etalt}, \eqref{etapv}, \eqref{etao}, \eqref{etaelall} and \eqref{etalem}, the maximum power transmission efficiency $\eta_{lem}$ can be obtained as:
\begin{equation}\label{etaoall}
\begin{aligned}
\eta_{om}
&=\eta_{el}\eta_{lt}\eta_{lem} \qquad\qquad\quad \\
&=\eta_{el}\eta_{lt}(a_2+\frac{b_2}{\eta_{el}\eta_{lt}P_{s}}) \\
&=a_1 a_2 \eta_{lt} + \frac{a_2 b_{1} \eta_{lt} + b_2}{P_{s}} \\
&=a_1 a_2 e^{-\alpha d} + \frac{a_2 b_{1} e^{-\alpha d} + b_2}{P_{s}}.
\end{aligned}
\end{equation}
\begin{figure}
\centering
\includegraphics[scale=0.6]{etaodf.eps}
\caption{Maximum Power Transmission Efficiency vs. Distance (Fog)}
\label{etaodf}
\end{figure}
Fig.~\ref{810etao} shows the relationship between $\eta_{om}$ and $P_{s}$ when $\eta_{lt}$ are 100\% and 50\% and cell temperatures are 0$^{\circ}$C, 25$^{\circ}$C, and 50$^{\circ}$C for 810nm. Fig.~\ref{1550etao} shows the same circumstances for 1550nm. $\eta_{om}$ raises up with $P_{s}$ increasing at first, then it reaches the plateau. The growth pattern of $\eta_{om}$ in Fig.~\ref{810etao} and Fig.~\ref{1550etao} is similar as $\eta_{el}$ in Fig.~\ref{8101550etael} and $\eta_{lem}$ in Fig.~\ref{810etale} and Fig.~\ref{1550etale}.
$\eta_{om}$ depends not only on the supply power $P_{s}$ but also on the distance $d$. Fig.~\ref{etaodc} depicts the relationship between $\eta_{om}$ and $d$ for different laser wavelength and PV-cell temperature, when $P_{s}$ = 40W and air quality is clear. Fig.~\ref{etaodh} and Fig.~\ref{etaodf} illustrate $\eta_{om}$ for the similar situation when air condition is haze and fog, respectively. Fig.~\ref{etaopl} describes how $\eta_{om}$ changes over $\eta_{lt}$ under clear air when $P_{s}$ is 40W.
From Fig.~\ref{etaodc} and Fig.~\ref{etaodh}, $\eta_{om}$ decreases when $d$ increases. $\eta_{om}$ of 810nm laser is higher than that of 1550nm laser when $d$ is short. However, $\eta_{om}$ for 810nm is lower than that of 1550nm when $d$ is long. From Fig.~\ref{etaodf}, $\eta_{om}$ of 810nm laser always keep higher than that of 1550nm laser until $\eta_{om}$ decrease to 0. At the same time, as described above, the cell temperature has bigger impact on $\eta_{om}$ for 1550nm than that of 810nm.
From Fig.~\ref{etaopl}, $\eta_{om}$ increases linearly as $\eta_{lt}$ enhances based on \eqref{etaoall}. Fig.~\ref{etaopl} provides a guideline of designing the DLC systems. For example, if 20\% of DLC maximum transmission efficiency is expected, the 1550nm DLC system can not meet the requirement, however, the 810nm DLC system is preferred. Meanwhile, when deploying the DLC system, the transmission efficiency at a certain distance provides the theoretical reference to determine the radius, i.e., the coverage, which is similar to the base station coverage analysis in mobile communications \cite{zhou2008energy,Erchin2007novel}. Therefore, the maximum economic benefits can be obtained by minimizing the number of DLC transmitters to cover a given area \cite{gtzhou2010}. This analysis provides a guideline for the efficient deployment of the DLC systems.
In summary, the numerical evaluation in this section validates the analytical model presented in Section \ref{Section3}. At first, for the three modules: electricity-to-laser conversion, laser transmission, laser-to-electricity conversion, the conversion or transmission efficiency of each module is quantitatively analyzed. Secondly, through numerical analysis, we obtain the approximate linear relationship between the supply power $P_s$ at the transmitter and the maximum PV-panel output power $P_m$ at the receiver. Next, the maximum DLC power transmission efficiency $\eta_{om}$ in closed-form is derived. Finally, based on the maximum power transmission efficiency, DLC system design and development guidelines are provided, for example, how to select the laser wavelength and determine the coverage of the DLC systems.
\begin{figure}
\centering
\includegraphics[scale=0.6]{etaopl.eps}
\caption{Maximum Power Transmission Efficiency vs. Laser Transmission Efficiency}
\label{etaopl}
\end{figure}
\section{Conclusions}\label{Section5}
This paper presents the distributed laser charging technology for wireless power transfer. The multi-module analytical modeling of distributed laser charging provides the in-depth view of its physical mechanism and mathematical formulation. The numerical evaluation illustrates the power conversion or transmission in each module under the impacts of laser wavelength, transmission attenuation, and PV-cell temperature. The linear approximation is adopted and validated by measurement and simulation for electricity-to-laser and laser-to-electricity power conversion. Thus the maximum power transmission efficiency in closed-form is derived and its performance depending on the supply power, laser wavelength, transmission distance, and PV-cell temperature is illustrated by figures. Therefore, this paper not only provides the theoretical insight, but also offers the practical guideline in system design and deployment of distributed laser charging.
Due to the space limitation, there are serval important issues unaddressed in this paper and left for our future work, some of which are briefly discussed here:
\begin{itemize}
\item The PV-panel efficiency used in the DLC system is about 50\%, which is not much efficient. More studies on the PV-panel types, the efficiency analyzation, and the total efficiency of the DLC system could be improved in the future.
\item Only 810nm and 1550nm laser wavelengths are considered in this paper. Wider range of wavelengths can be studied to make the DLC system more universal in the future work.
\item To convert the PV-panel output current and voltage to different preferred charging current and voltage for different applications, the circuit or device that can convert a source of direct current from one voltage level to another is worth to be discussed in the future.
\item The point-to-point charging procedure is well illustrated in this paper. On this basis, the accessing protocols, the scheduling algorithms, the influencing factors of power conversion and transmission, and the system optimization for charging batteries adaptively in the the point-to-multiple-point wireless charging scenario should be another interesting topic to discuss.
\item Since point-to-multiple-point wireless power transfer is naturally supported by distributed laser charging systems, the network architecture of WPT becomes an interesting research topic worthy of further investigation. The related protocols and algorithms to effectively operate WPT networks could be developed, e.g., WPT network access protocols, WPT scheduling algorithms and so on \cite{wirelesstechniques}.
\item It is interesting to investigate the potential simultaneous wireless information and power transfer (SWIPT) in distributed laser charing systems. Due to the huge available bandwidth conveyed by laser, such high-power and high-rate SWIPT systems have the potential to support the demanding applications, e.g., IoT, mobile virtual reality/augmented reality, ultra-high-definition video streaming and so on \cite{zhangruimimo}.
\end{itemize}
\bibliographystyle{IEEEtran}
\bibliographystyle{unsrt}
|
\section{Introduction}
\noindent The Kepler problem
\begin{equation}
\label{eq:Kepler}
\ddot{q}+\frac{q}{\|q\|^3}=0, \quad q \in \mathbb{R}^2\setminus\{0\}.
\end{equation}
is a classical reduction of the two-body
problem~\cite{arnol2013mathematical}. Here, we think of $q$ as the position of a
spacecraft, and of the attraction as the action of the Earth. We are
interested in controlling the transfer of the spacecraft from one
Keplerian orbit towards another, in the plane. Denoting
$v = \dot q$ the velocity, and the adjoint variables of $q$ and $v$ by $p_q$ and
$p_v$, the minimum time dynamics is a Hamiltonian system with
\begin{equation}
\label{eq:0}
H(q,v,p_q,p_v)=p_q.v-\frac{p_v.q}{\|q\|^3}+\|p_v\|,
\end{equation}
as is explained in section~\ref{sec:reminder}.
Prior studies of this problem can be found in \cite{caillau2012minimum, caillau2001}.
The controlled Kepler problem can be embedded in the two parameter
family obtained when considering the control of the circular restricted three-body
problem:
\begin{equation} \label{eq51}
\ddot{q}+\nabla_q\Omega_\mu(t,q) = \varepsilon u,
\end{equation}
where
\begin{eqnarray*}
\Omega_\mu(t,q) &=&
-\frac{1-\mu}{\sqrt{(q_1+\mu\cos t)^2+(q_2+\mu\sin t)^2}}\\
&& -\frac{\mu}{\sqrt{(q_1-(1-\mu)\cos t)^2+(q_2-(1-\mu)\sin t)^2}}
\end{eqnarray*}
is the potential parameterized by the ratio of masses, $\mu \in [0,1/2]$,
and where $u \in \mathbb{R}^2$ is the control, whose amplitude is modulated by the second
parameter, $\varepsilon \geq 0$. Alternatively to time minimization, minimization of the
$\L^2$ norm of the control can be considered,
\[ \int_0^{t_f} u^2(t)\,\d t \to \min. \]
This is the so-called energy cost. In the uncontrolled model ($\varepsilon=0$), it is
well known that the Kepler case ($\mu=0$) is integrable and geodesic (there exists a Riemannian metric such that Keplerian curves are geodesics of this metric \cite{moser-1970a,osipov-1977a})
while there are obstructions to integrability for positive $\mu$.
In the controlled case ($\varepsilon>0$), the Kepler
problem for the energy cost has been shown to be integrable (and geodesic)
when suitably averaged (see \cite{caillau2009} for a survey). The aim of this paper is
to study the integrability properties of the Kepler problem for time minimization.
The pioneering work of Ziglin in the 80's \cite{ziglin1982branching},
followed by the modern formulation of differential Galois theory in
the late 90's by Moral{\`e}s, Ramis and Sim{\'o}
\cite{morales2001galoisian,morales2007integrability}, have led to a
very diverse literature on the integrability of Hamiltonian systems.
According to Pontrjagin's Maximum principle, one can turn general
optimization problems with dynamical constraints into Hamiltonian
systems, which are generally not everywhere differentiable. Optimal
control theory thus provides an abundant class of dynamical systems
for which integrability is a central question. Yet,
differential Galois theory has not so often been applied in this context
(see, \emph{e.g.}, \cite{combot-2015a}),
in part because of the difficulty brought by the singularities.
Notwithstanding theses singularities (vanishing of the adjoint variable $p_v$, here),
we show how to apply these ideas to the system~\eqref{eq:0}.
\section{Setting}
\subsection{The minimum time controlled Kepler problem}
\label{sec:reminder}
We first recall some classical facts on optimal control. We refer for
example to the book of Agrachev and Sachkov \cite{agrachev13} for more details.
Let $M$ be an $n$-dimensional smooth manifold and $U$ an arbitrary
subset of $\mathbb{R}^m$ (typically a submanifold with boundary).
A controlled dynamical system is a smooth family of vector fields
\[f : M\times U\rightarrow TM\]
parameterized by the control values.
Admissible controls are measurable functions valued in the subset $U$.
A preliminary question is the following:
Is some final state $x_f$ accessible from some initial state $x_0$,
\emph{i.e.} does the system
\[ \dot{x}(t)=f(x(t),u(t)),\quad u(t) \in U, \]
\[ x(0)\ =x_0,\quad x(t_f)=x_f, \]
have a solution for some admissible control? The system
is said to be controllable if the answer is positive for all
possible initial and final states $x_0,x_f\in M$.
The controlled Kepler problem, associated with~\eqref{eq:Kepler}, is
\[ \ddot{q}+\frac{q}{\|q\|^3}=u, \quad q \in \mathbb{R}^2\setminus\{0\}, \quad
u_1^2+u_2^2\leq1, \]
\[ (q(0),\dot q(0))\ \ =(q_0,v_0),\quad
(q(t_f),\dot q(t_f))=(q_f,v_f), \]
where $q$ is the position vector of a spacecraft and where
the control $u$ is the thrust of the engine. The
thrust is obviously bounded; here we assume that it is valued in the
Euclidean unit ball. (Note that, with respect to (\ref{eq51}),
we have chosen $\varepsilon=1$; as will be clear from Section~\ref{sproof},
this does not restrict the generality of the analysis.)
\begin{proposition}[\cite{caillau2001}]
The Kepler problem is controllable.
\end{proposition}
\noindent This is a consequence of two facts: The Lie algebra generated by the
drift and the vector field supporting the control generate the whole
tangent space at each point (which entails some local
controllability), and the uncontrolled flow (or \emph{drift}) of the
Kepler problem is recurrent. Under some additional convexity and compactness assumptions, one is
then able to retrieve existence of optimal controls.
We now deal with such optimal controls.
We restrict ourselves to integral cost functions, that is to problems of the form
\begin{equation}
\label{eq:opti}
\begin{cases}
\dot{x}(t)=f(x(t),u(t)),\\
x(0)\ =x_0,\quad x(t_f)=x_f,\\
\int_{0}^{t_f} L(x(t),u(t))\, \text{dt}\rightarrow \min
\end{cases}
\end{equation}
where the final time $t_f$ can be fixed or not, and $L:M\times U\rightarrow\mathbb{R}$
is a smooth function.
In the early 60's, Pontrjagin and his coauthors realized that necessary conditions
for optimality could be stated in Hamiltonian terms. By $T^*M$ we denote the cotangent
bundle of the manifold $M$.
\begin{definition} \label{def1}
The associated pseudo-Hamiltonian is
\[H: T^*M\times\mathbb{R}\times U \to \mathbb{R}, \quad (x,p,p^0,u) \mapsto \langle
p,f(x,u)\rangle+p^0 L(x,u).\]
\end{definition}
\noindent The following fundamental result is Pontrjagin Maximum Principle
\cite{pontryagin1987mathematical} (see
\cite{agrachev13} for a modern presentation).
\begin{theorem} \label{th1}
If $(x,u)$ solves \eqref{eq:opti}, there exists a Lipschitzian function
$p(t)\in T^*_{x(t)}M$, $t\in[0,t_f]$, a constant $p^0\leq 0$,
$(p(t),p^0)\neq 0$, such that, almost everywhere,
\begin{itemize}
\item[(i)] $(x,p)$ is a solution of the Hamiltonian system associated with
$H(\cdot, \cdot, u(t))$:
\[ \dot{x}=\frac{\partial H}{\partial p}(x,p,u),\quad
\dot{p}=-\frac{\partial H}{\partial x}(x,p,u), \]
\item[(ii)] $H(x(t),p(t),u(t))=\max_{v \in U}H(x(t),p(t),v)$.
\end{itemize}
Such curves $(x,p)$ are called extremals. As a consequence of the
maximization condition, the pseudo-Hamiltonian evaluated along an extremal is
constant. Moreover, if the final time is free then this constant is zero.
\end{theorem}
\noindent This powerful result has some downsides. The Hamiltonian is
defined on the cotangent bundle of the original phase space, and thus
the dimension is doubled. Besides, the maximization condition, which "eliminates the
control" and allows to obtain a truly Hamiltonian system in $(x,p)$ only, might generate
singularities (that is non-differentiability points of the maximized Hamiltonian which is in general only Lipschitzian as a function of time when evaluated along an extremal).
The above theorem applies to time minimization with $L\equiv 1$ (and free final
time). In this case, the non-positive constant $p^0$ is only related to the level of the
Hamiltonian, and we will not mention it in the sequel as we will not discuss the
implications of having \emph{normal} ($p^0 \neq 0$) or \emph{abnormal} ($p^0=0$) extremals.
\subsection{Main result}
The minimum time Kepler problem can be stated according to
\begin{equation}\label{eq:mt}
\begin{cases}
\ddot{q}+\frac{q}{\|q\|^3} =u,\quad \|u\|\leq 1,\\
(q(0),\dot q(0)) = (q_0,v_0),\quad (q(t_f),\dot q(t_f)) = (q_f,v_f),\\
t_f\rightarrow \min,
\end{cases}
\end{equation}
where, as before, $q \in \mathbb{R}^2$ is the position vector and $u \in \mathbb{R}^2$
the control. It will be convenient to use the same notations as in the
general problem~\eqref{eq:opti} and let
\[q=(x_1,x_2), \quad \dot q=(x_3,x_4),\]
be the coordinates on the initial phase space
$M = (\mathbb{R}^2 \setminus \{0\}) \times \mathbb{R}^2 $. According to Definition~\ref{def1},
the pseudo-Hamiltonian is then
\begin{equation} \label{eq20}
H(x,p,u) = p_1x_3+p_2x_4 - \frac{p_3x_1+p_4x_2}{(x_1^2+x_2^2)^{3/2}}
+ p_3u_1+p_4u_2.
\end{equation}
According to Theorem~\ref{th1}, minimizing trajectories must be projections on $M$ of
integral curves of the Hamiltonian that has to be maximized over the unit disk.
The maximized Hamiltonian is readily equal to
\[H(x,p) = p_1x_3+p_2x_4 - \frac{p_3x_1+p_4x_2}{(x_1^2+x_2^2)^{3/2}} +
\sqrt{p_3^2+p_4^2}\]
on $T^*M$, while the control is given by
\[ u = \frac{1}{\sqrt{p_3^2+p_4^2}}(p_3,p_4) \]
whenever $p_3$ and $p_4$ do not vanish simultaneously.
Now, let
\[\mathcal{M}=\{(x,p,r)\in\mathbb{C}^8\times \mathbb{C}_*^2,\;
r_1^2=x_1^2+x_2^2,\;\; r_2^2=p_3^2+p_4^2\}\]
be the Riemann surface of $H$. It is a complex symplectic manifold
(with local Darboux coordinates $(x,p)$ outside the singular
hypersurface $r_1r_2=0$), over which $H$ extends meromorphically, and
even rationally, since
\begin{equation}
\label{eq:H}
H(x,p,r)=p_1x_3+p_2x_4-\frac{p_3x_1+p_4x_2}{r_1^3}+r_2.
\end{equation}
The Hamiltonian $H$ has four degrees of freedom, hence (see
\cite{arnol2013mathematical}) the meromorphic Liouville integrability
of $H$ over $\mathcal{M}$ would mean that there would exist three independent
first integrals, in addition to $H$ itself, almost everywhere in
$\mathcal{M}$. The aim of this paper is to prove that it is not the case.
\begin{theorem}
The minimum time Kepler problem is not meromorphically Liouville
integrable on $\mathcal{M}$.
\label{thm:ni}
\end{theorem}
\noindent It is well known that the classical Kepler problem is integrable, and
even super integrable (since there are more first integrals than
degrees of freedom, as a result of Kepler's first law and of the
dynamical degeneracy of the Newtonian potential---see
for instance \cite{Fejoz:2004a}).
On the opposite, the three-body problem is
not as is known after the seminal work of Poincar{\'e} (for recent accounts on this topic see, \emph{e.g.}, \cite{Combot:2012:nbp, Julliard:2000:Bruns,
Poincare:1892, Tsygvintsev:2003}). Similarly, the above theorem
asserts that lifting the Kepler problem to the cotangent bundle and
introducing the singular control term $r_2$ breaks integrability.
This result prevents the existence of enough complex analytic (and even
meromorphic) first integrals to ensure integrability over
$\mathcal{M}$. Or course, it does not prevent the existence of an
additional real first integral
which would have a natural frontier asymptotic to the
real domain and thus, would not extend to the complex plane.
Future work might be dedicated to investigate either or not
Theorem \ref{thm:ni} holds for real first integrals.
\section{Proof of Theorem~\ref{thm:ni}} \label{sproof}
\noindent The rest of the article is devoted to proving the theorem. Our proof
consists in studying the variational equation along some integral
curve of~\eqref{eq:H}. In order to carry out this computation, we
choose a collision orbit, with the drawback that it requires some
regularization. We also note that there exist effective tools to perform this kind of
computations (see, \emph{e.g.}, \cite{weil-2012a}).
The algebraic obstruction to Liouville integrability
comes from the theorem below of Moral{\`e}s and Ramis, which we now recall. We follow the presentation of Singer in \cite{singer2009}.
\subsection{Some facts of Galois differential theory}
Consider a linear differential equation $(L) :Y'=AY$, $A\in M_n(k)$,
$k$ being a differential field whose field of constants $k_0$ is algebraically closed, and of characteristic zero. We want the Galois group to be the group of symmetries preserving all algebraic and differential relations of a basis of solutions. We consider the polynomial ring
$$S=k[Y_{1,1},\dots,Y_{n,n},1/\hbox{det}(Y)]$$
where $Y$ is an $n\times n$ matrix. This ring has a derivation provided by the differential system $Y'=AY$. We now consider a maximal differential ideal $M$ of $S$, and the quotient $R=S/M$. This quotient satisfies the following
\begin{definition}[Picard-Vessiot field]
A Picard-Vessiot ring for $Y' = AY$ is a differential ring $R$ over $k$ such that
\begin{itemize}
\item[(i)] The only differential ideals of $R$ are $(0)$ and $R$.
\item[(ii)] There exists a fundamental matrix $Z\in GL_n(R)$ for the equation
$Y' = AY$.
\item[(iii)] $R$ is generated as a ring by $k$, the entries of $Z$ and $1/\hbox{det}(Z)$ .
\end{itemize}
\end{definition}
It turns out that the choice of the maximal differential ideal $M$ always gives the same Picard-Vessiot ring up to isomorphism. This ring is also a domain, thus allowing to consider the quotient field, the Picard-Vessiot field.
\begin{definition}[Galois group]
The differential Galois group of $R$ over $k$ is the group of differential automorphism of $R$ preserving $k$, noted $Gal(R/k)$.
\end{definition}
\noindent For a differential system $Y'=AY$, if there is no ambiguity on the base field $k$. (For the case treated in this paper, the base field $k$ is $\mathbb{C}(z)$.)
Given a fundamental matrix of solution $Z$ and a Galois group element $\sigma$, we have $Z'=AZ$, and thus applying $\sigma$, we also have $\sigma(Z)'=A\sigma(Z)$. Thus $\sigma(Z)$ is also a matrix of solutions; there exists a constant matrix $C$ such that $\sigma(Z)=ZC$, and as $\sigma$ is an automorphism, $C$ has to be invertible. So $Gal(R/k)$ can be represented as a group of $n\times n$ matrices.
\begin{proposition}
The Galois group $Gal(R/k)\subset GL_n(k_0)$ is a linear algebraic
group, i.e. the zero set in $GL_n(k_0)$ of a system of polynomials
over $k_0$ in $n^2$ variables.
\end{proposition}
\noindent\textbf{Proof.}
This can be obtained by letting a Galois group element $\sigma$ act (right multiplication by a matrix) on the differential ideal $I=(f_1,\dots,f_p)$. We can moreover assume that $f_i\in k[Y]$. As this does not change the degrees in the $Y_{i,j}$ and since $I$ must be stabilized, $\sigma(f_i)$ must belong to $I\cap k_{\max(\deg f_1,\dots, \deg f_p)}[Y]$. This condition is a condition of membership to a vector space, which provides algebraic conditions on the entries of the matrix $\sigma$.
\begin{proposition}[Fundamental Theorem of Differential Galois
Theory]
Let $K$ be a Picard-Vessiot field with differential Galois group $G$ over $k$.
\begin{itemize}
\item[(i)] There is a one-to-one correspondence between Zariski-closed subgroups
$H\subset G$ and differential subfields $F$, $k\subset F\subset K$, given by
$$H \subset G \rightarrow K^H = \{a \in K, \sigma (a) = a\; \forall \sigma \in H\}$$
$$F \rightarrow Gal(K/F)= \{\sigma \in G, \sigma(a)=a \; \forall a\in F \}$$
\item[(ii)] A differential subfield $F$, $k\subset F \subset K$, is a Picard-Vessiot extension of $k$ if and only if $Gal(K/F)$ is a normal subgroup of $G$, in which case $Gal(F/k) \simeq G/Gal(K/F)$.
\end{itemize}
\end{proposition}
We are interested in non integrability for nonlinear Hamiltonian systems. The link with the Hamiltonian world is given by the celebrated theorem of
Moral\'es-Ramis below. We recall that an algebraic group $G$ is said to be virtually Abelian if
its connected component containing the identity is an Abelian subgroup of $G$.
\begin{theorem}[Moral{\`e}s-Ramis \cite{morales2001galoisian}]
\label{thm:mr}
Let $H$ be an analytic Hamiltonian on a complex analytic
symplectic manifold and $\Gamma$ be a non constant solution. If $H$ is
integrable in the Liouville sense with meromorphic first integrals,
then the first order variational equation along $\Gamma$ has a
virtually Abelian Galois group over the base field of meromorphic
functions on $\Gamma$.
\end{theorem}
\noindent The main idea behind this theorem is that if $H$ is Liouville integrable, then so are the linearized equations near a non constant solution $\Gamma$. More precisely, thanks to Ziglin's Lemma below, the first integrals of $H$ can be transformed in such a way that their first non trivial term in their series expansion near $\Gamma$ are functionally independent.
\begin{lemma}[Ziglin's Lemma]
Let $\Phi_1,\dots,\Phi_r\in k(x_1,\dots,x_n)$ be functionally independent functions. We consider $\Phi_1^0,\dots,\Phi_r^0$ the lowest degree homogeneous term for some fixed positive weight homogeneity in $x_1,\dots,x_n$. Assume $\Phi_1^0,\dots,\Phi_{r-1}^0$ are functionally independent. Then there exists a polynomial $\Psi$ such that the lowest degree homogeneous term $\Psi^0$ of $\Psi(\Phi_1,\dots,\Phi_r)$ is such that $\Phi_1^0,\dots,\Phi_{r-1}^0,\Psi^0$ are functionally independent.
\end{lemma}
\noindent Applying this Lemma recursively, we prove that if a Hamiltonian system admits a set of commuting, functionally independent meromorphic first integrals on a neighbourhood of a curve, then their first order terms, after possibly polynomial combinations of them, are also commuting, functionally independent meromorphic first integrals of the linearized system along the curve.
Moral{\`e}s-Ramis \cite{morales2001galoisian} precisely proved that symplectic linear differential systems having such first integrals have a Galois group whose identity component is Abelian. This result can be expected knowing that the Galois group leaves invariant every first integral, so the more first integrals, the smaller the Galois group.
We will need the definition of the monodromy group of a
linear differential equation. Let us consider a differential system $Y'=AY,\; A\in M_n(\mathbb{C}(x))$. We note $S=\mathbb{P}^1 \setminus \{\hbox{singularities of } A\} $.
Let us consider a point $z_0\in S$ and a closed oriented curve $\gamma\subset S$, with $x_0 \in\gamma$. There exists a basis of solutions $Z$ on a neigbourhood of $x_0$, holomorphic in $z$. We now use analytic continuation along the loop $\gamma$ to extend this basis of solutions. However, it cannot \emph{a priori} be extended to a whole neighbourhood of $\gamma$, because after one loop, the basis of solutions $Z_\gamma$ at $x_0$ could be different. This defines a matrix $D_\gamma \in GL_n(\mathbb{C})$ such that $Z_\gamma =Z D_\gamma$ and thus a homomorphism
$$\hbox{Mon} :\pi_1(S,x_0) \rightarrow GL_n(\mathbb{C}), \quad \hbox{Mon}(\gamma)= D_\gamma.$$
This homomorphism carries the group structure of $\pi_1(S,x_0)$, and thus its image is also a group.
\begin{definition}
The image of the application $\hbox{Mon}$ is called the monodromy group.
\end{definition}
\noindent Note that the monodromy group depends on the choice of $Z$, so it is only
determined up to conjugation. Since analytic continuation preserves
analytic relations, the monodromy group is a subset of
the differential Galois group over the base field of meromorphic functions on $S$; in particular, it is included in the differential Galois group over the base field of rational functions. For Fuchsian systems (all singularities are regular singularities, i.e. the growth at singularities of solutions is at most polynomials), we have moreover the following.
\begin{theorem}[Schlesinger density theorem \cite{sauloy16}]\label{thmschles}
Let $(E):\ Y'=AY$ be a Fuchsian differential linear equation with coefficients in $\mathbb{C}(x)$ and let
$\Pi$ be its monodromy group. Then $\Pi$ is dense for the Zariski
topology in the Galois group of the Picard-Vessiot extension of
$(E)$ over the base field of rational functions: $\overline{\Pi} = \textrm{Gal}(A)$.
\end{theorem}
\subsection{A collision orbit}
In order to find an explicit solution of \ref{eq20}, let us define the
$4$-dimensional symplectic submanifold
\[ S = \{(x,p,r)\in \mathcal{M}\ |\ x_2=x_4=p_2=p_4=0,r_1=x_1,r_2=-p_3\}. \]
As $S$ is the phase space of the controlled Kepler problem on the line
(collision orbit)
parameterized by $q_1$, it is invariant. On the interior of $S$,
$(x_1,x_3,p_1,p_3)$ is a set of (Darboux) coordinates and, in
restriction to $S$, the Hamiltonian reduces to
\[H(x,p)=p_1x_3-\frac{p_3}{x_1^2}-p_3,\]
so the Hamiltonian vector field on $S$ is
\[
\begin{cases}
\dot x_1 = x_3\\
\dot x_3 =- 1 -\frac{1}{x_1^2}\\
\dot p_1 = - \frac{2p_3}{x_1^3}\\
\dot p_3 = - p_1.
\end{cases}
\]
In particular,
\begin{equation}
\label{eq:x1p3}%
\begin{cases}
\ddot{x}_1=-1-\frac{1}{{x_1}^2} \\
\ddot{p}_3-\frac{2p_3}{x_1^3}=0.
\end{cases}
\end{equation}
As is known since the work of Charlier and Saint Germain on the Kepler problem with a
constant force (see \cite{beletsky01}), the function
\[C=\frac{1}{2}x_3^2+x_1-\frac{1}{x_1}\]
is a first integral on $S$ and $H_{|S}$ is integrable. Let us change time to $s=x_1(t)$ and denote by $'=\frac{d}{ds}$ the derivation with respect to this new time. It suffices to find an obstruction in this modified time, as
explained at the end of the proof.
Using~\eqref{eq:x1p3}, we see that the variable $p_3$ satisfies the linear differential equation
\[2\left(C+\frac{1}{x_1}-x_1\right)p_3''(x_1) -
\left(1+\frac{1}{{x}_1^2}\right)p_3'(x_1)-\frac{2p_3(x_1)}{x_1^3}=0,\]
which yields
\[p_3(x_1) = \frac{\sqrt{-Cx_1+x_1^2-1}}{\sqrt{x_1}}
\left(c_1\int\frac{x_1^{3/2}}{(-Cx_1+x_1^2-1)^{3/2}}dx_1+c_2\right)\]
for some constants of integration $c_1$ and $c_2$. Here the symbol
$\int f(x_1)dx_1$ denotes some primitive of $f$ with respect to the
variable $x_1$. It suffices to find one particular integral curve along which the
variational equation has a non virtually Abelian Galois group. To this end, we
consider the simple---but rich enough---case $c_1=0$, $c_2=1$.
\[p_3(x_1)=\frac{\sqrt{-Cx_1+x_1^2-1}}{\sqrt{x_1}}\cdot \]
Using the expression of the first integral $C$ and of the vector
field, we deduce
\[x_3(x_1)=\sqrt{2}\frac{\sqrt{-Cx_1+x_1^2-1}}{\sqrt{x_1}}, \;\;
p_1(x_1)=-\frac{1}{\sqrt{2}}\frac{x_1^2+1}{x_1^2}\cdot\]
Choosing $C=2i$ and some determination of the squares yields
a particularly simple solution $\Gamma$ drawn on
$S\subset \mathcal{M}$,
\begin{equation}\label{eqGamma}
\begin{cases}
x_1 = x_1,\\
x_2 = 0,\\
x_3 = \sqrt{2}\frac{x_1-i}{\sqrt{x_1}},\\
x_4 = 0,\\
\end{cases}
\quad
\begin{cases}
p_1 = -\frac{x_1^2+1}{\sqrt{2} x_1^2},\\
p_2 = 0,\\
p_3 = \frac{x_1-i}{\sqrt{x_1}},\\
p_4 = 0.
\end{cases}
\end{equation}
\subsection{Normal variational equation}
In the initial time, the linearized equation along $\Gamma$ is the
Hamiltonian vector field associated with the Hamiltonian $DH$ along
$\Gamma$:
\[\dot{Z}(t)=A(t)Z(t), \quad A(t)=J \, D^2 H(\Gamma(t)),\]
where $J$ is the Poisson structure. In the coordinates
$(x_1,...,x_4,p_1,...,p_4)$,
\[
J=\begin{pmatrix}
0_4 & I_4\\
-I_4 & 0_4
\end{pmatrix}.\]
We will keep on using time $x_1$, instead of the initial time $t$,
writing
\[Z'(x_1(t))=\frac{1}{x_3(t)}A(x_1(t))Z(x_1(t)).\]
Let us now reorder coordinates according to $(x_1,x_3,p_1,p_3,x_2,x_4,p_2,p_4)$.
Since $S$ is an invariant submanifold, the $8\times8$ matrix $A$ has
an upper triangular bloc structure
\[A=\begin{pmatrix}
A_1 & A_2\\
0 & A_3
\end{pmatrix}\]
with
\[A_3=\begin{pmatrix}
0 & 0 & 0 & \frac{1}{\sqrt{2}p_3}\\
-\frac{1}{\sqrt{2}p_3^2} & 0 & -\frac{1}{\sqrt{2}x_1^3p_3} & 0\\
0 & \frac{1}{\sqrt{2}p_3} & 0 & 0\\
-\frac{1}{\sqrt{2}x_1^3p_3} & 0 & \frac{3}{\sqrt{2}x_1^4} & 0
\end{pmatrix}.\]
Moral\`es-Ramis Theorem gives necessary conditions for Liouville
integrability in terms of the Galois group of this linear differential
system over the base field of meromorphic functions \emph{on}
$\Gamma$. Looking at the expression \eqref{eqGamma} of $\Gamma$, we
see that meromorphic functions on $\Gamma$ are just meromorphic
functions in $\sqrt{x_1}\in \mathbb{C}\setminus \{0,\pm \sqrt{i}\}$.
The block $A_3$ corresponds to infinitesimal variations in the normal
direction to $S$, which is the part where interesting phenomena might
occur. As the Picard-Vessiot field is generated by all the components
of the solutions, the Picard-Vessiot field $K$ generated by the normal
variational equation
\[(L): \quad X'=A_3X, \quad X=(X_1,X_2,X_3,X_4)\]
is a subfield of the Picard-Vessiot field of the whole variational
equation, and thus $\hbox{Gal}(A) \supset \hbox{Gal}(A_3)$. That
$\hbox{Gal}(A_3)$ is not virtually Abelian will thus imply that
$\hbox{Gal}(A)$ itself is not virtually Abelian.
In order to reduce the system to a one dimensional linear equation, we
use the cyclic vector method on $A_3$: From $(L)$ we get
$X_1'=L_1(X_1,X_2,X_3,X_4)$, where $L_1$ is a linear form on $\mathbb{R}^4$,
thus by derivation,
\begin{align*}
X_1''=&L_1(X_1',X_2,X_3,X_4) + L_1(X_1,X_2',X_3,X_4)\\
+&L_1(X_1,X_2,X_3',X_4) + L_1(X_1,X_2,X_3,X_4')\\
=&L_2(X_1,X_2,X_3,X_4).
\end{align*}
Iterating, we obtain
\begin{empheq}[left = \empheqlbrace]{align*}
X_1&=X_1,\\
X_1'&=L_1(X_1,X_2,X_3,X_4),\\
X_1''&=L_2(X_1,X_2,X_3,X_4),\\
X_1^{(3)}&=L_3(X_1,X_2,X_3,X_4),\\
X_1^{(4)}&=L_4(X_1,X_2,X_3,X_4).
\end{empheq}
The $L_i$'s are five linear forms on $\mathbb{R}^4$, so $X_1$ must satisfy some linear differential equation of order $4$ that we compute to be
\begin{multline}
X_1^{(4)}+
\frac{2(3i-5x_1)}{x_1(i-x_1)} X_1^{(3)} +
\frac{(-3x_1+i)(-29x_1+23i)}{4(x_1-i)^2x_1^2}X_1''\\
-\frac{(i-3x_1)(7x_1+i)}{4(x_1-i)^2x_1^3}X_1'
+\frac{3x_1+i}{4(x_1-1)^3x_1^4}X_1=0.
\label{hyp}
\end{multline}
We find a solution of this equation of the form \small
\[X_1(x_1)=\frac{i-x_1}{\sqrt{x_1}}\left( c_1+c_2\int
\sqrt{x_1}(1+ix_1)^{-\frac{3}{2}-i\frac{\sqrt{3}}{2}}.{} _2{F}_1
(\gamma(x_1))dx_1 \right),\]
where ${}_2F_1$ is the Gauss hypergeometric function and
\[\gamma(x_1) =
\left(\frac{5}{2} - i\frac{\sqrt{3}}{2}, \frac{1}{2} +
i\frac{\sqrt{3}}{2}, 1+i\sqrt{3},1+i x_1\right).\]
The Picard-Vessiot field $K$ contains this solution and, as it is a
differential field, it also contains
\[\sqrt{x_1}(1+ix_1)^{-\frac{3}{2}-i\frac{\sqrt{3}}{2}} {}_2{F}_1(\gamma(x_1)).\]
Noting $\tilde{K}$ the differential field generated by this function, we have $\tilde{K} \subset K$. Now the Galois group of ${}_2{F}_1(\gamma(x_1))$ over
$\mathbb{C}(x_1)$ is $SL_2(\mathbb{C})$ (see Kimura's table, \cite{kimura1969riemann}). By Galois correspondence, the Galois group of (\ref{hyp}) over the rational functions in $x_1$ admits $SL_2(\mathbb{C})$ as a subgroup.
The hypergeometric equation \eqref{hyp} is Fuchsian
(all its singular points are regular), so thanks to Theorem~\ref{thmschles},
we know that its Galois group over the field of rational functions is the closure of its monodromy group.
Besides, the Galois group over meromorphic functions contains the
monodromy group, and of course, is included in the Galois group over
rational functions. Eventually, the Galois group of (\ref{hyp}) over meromorphic functions in
$x_1$ also contains $SL_2(\mathbb{C})$. Thus,
adding the algebraic extension $\sqrt{x_1}$, the Galois group can be
reduced to at most one subgroup of index $2$: The only possibility is that the identity component contains
$SL_2(\mathbb{C})$ again.
So the Galois group of $K$ over the base field of meromorphic functions
in $\sqrt{x_1}\in \mathbb{C}\setminus \{0,\pm \sqrt{i}\}$ contains
$SL_2(\mathbb{C})$ and is not virtually Abelian. According to Morales-Ramis,
this concludes the proof.
\section*{References}
|
\section{Introduction}
Coronal Mass ejections (CMEs) are violent phenomena of solar activity with repercussions throughout the entire heliosphere.
Their manifestations into interplanetary space are responsible for major geomagnetic storms, hence the prediction of the arrival of interplanetary coronal mass ejections (ICMEs) at 1AU is one of the primary subjects of the space-weather forecasting \cite[e.g.:][]{Daglis2001, Schrijver+2010}.\\
Several forecasting methods have been proposed over the last two decades.
On one hand, there are the approaches relying on statistical-empirical relations established through observations between coronagraphically measured parameters and quantities related to their heliospheric propagation \citep[e.g. ][]{Brueckner+1998}.
Another approach is represented by numerical MHD-based models of the heliospheric propagation of ICME, generally requiring a detailed knowledge of the state of the heliosphere and large computational facilities \citep[as is the case for the WSA-ENLIL model][]{Enlil99, Enlil04, Enlil1, Enlil2}.
The numerical models are fairly accurate \citep{Vrsnak+2014}, and highly sensitive to the quality of the input parameters \citep{Falk+2010}, as one may expect.
Recently, the WSA-ENLIL model started to be employed also in a probabilistic approach \citep{Mays+2015, Cash+2015, Pizzo+2015} to quantify the prediction uncertainties and to determine the forecast confidence.
However this approach, interesting as it is, is not widely enough used for real-time space-weather forecasting due to the demanding computational needs.\\
The last category, somewhat lying in between the previous ones, is that employing an MHD- or HD-based simplified description of the interactions an ICME may be subjected to during its interplanetary travel \cite{Gopalswamy+2000, Vrsnak+2002, Michalek+2002, Schwenn+2005}.
Such an approach leads to analytical models or empirical models which require modest computational power.
Those models assume a morphology (either simple as in \citet{Mostl+2011} or more complex as in \citet{Isavnin2016}), a fixed direction and a velocity evolution for the CME and can predict an arrival time and speed from relatively limited initial information on the CME onset conditions.
Such initial conditions can be obtained from several sources, such as LASCO C2 and C3 coronal imagers on-board the \textit{SOlar and Heliospheric Observatory} \citep[SOHO][]{Domingo+1995} and, more recently, from COR1 and COR2 and the \textit{Heliospheric Imager} (HI) on-board the \textit{Solar Terrestrial Relation Observatory} \citep[STEREO][]{Kaiser+2008} spacecraft either separately or using appropriate tools to merge the measures taken from the different instruments and the different points of view \citep[][]{Lugaz+2009, Davies+2012, Mostl+2013, Mostl+2014}.
In this paper, we focus our attention on a model belonging to the last category, the \emph{drag-based model} \cite[DBM ][]{Vrsnak+2013}.
The model hypothesizes a simple interaction between the ICME plasma and the solar wind that works to equalize the ICME velocity to that of the solar wind itself.
This is consistent with the measures of the ICME speeds in the near Earth environment which are typically confined in the 400-700km/s range and the estimates of the initial velocity of the plasma ejecta near the Sun, which range between 100km/s and 2000km/s.
This process has been modeled analogously as an aerodynamic or viscous drag by several authors \citep{Cargill+1995, Vrsnak+2002,Shi+2015}.
It makes use of the initial CME velocity, its distance from the Sun at the moment of the measure, and the solar wind speed to compute the travel time at 1AU.\\
The DBM has already generated a whole family of approaches, which may differ for the way to evaluate the initial parameters, or how the CME is propagated in the heliosphere.\\
The difference may arise from the peculiarity of the data used (type and source) and from the interpretation of such data to estimate the CME onset parameters, which ultimately depends on the shape assumed for the CME itself (the Fixed-$\phi$ method in \citet{Sheeley+1999, Rouillard+2008,Mostl+2014}, the Harmonic mean method in \citet{Howard+2009,Lugaz+2009}, the Self Similar Expansion method in \citet{Davies+2012,Mostl+2013}, the Graduated Cylindrical Shell method in \citet{Thernisien+2006,Thernisien+2009}, the Elliptic Conversion method in \citet{Rollett+2016}, to cite the most used).
Or, the difference can be in the way the drag effect is approximated and the velocity of the CME evolves as in \citet{Hess+2015}, \citet{Zic+2015} and \citet{Rollett+2016}.\\
In the cited literature, much attention is paid to get the best estimate of the DBM parameters and an evaluation of the associated errors, but none of the mentioned DBM approaches takes into consideration this last information in the implementation of the forecast.\\
In this work, we apply a statistical approach on the DBM for the computation of ICME travel times, by introducing the probability distributions rather than exact values for the input parameters.
This approach has the non-trivial advantage to provide also an evaluation of the uncertainty on the arrival time.
In Section \ref{DBM} we rapidly revisit the equations of the DBM model and introduce its probabilistic version.
We also present and discuss the Probability Distribution Functions (PDFs) that we assume to compute the most probable ICME travel times.
Section \ref{data} presents the dataset of CME speeds, onset times and travel times that we use to compute and then compare (Section \ref{test}) the forecast travel times and associated errors.
In Section \ref{conclusions} we comment on our results, comparing with those already present in the literature, and discuss possible applications and further evolutions of this model.\\
For the sake of clarity, we specify that in this work the terms CME and ICME are referred to the plasma and magnetic field structure expelled from the Sun, without the shock that precedes it.\\
\section{The drag-based model}
\label{DBM}
\subsection{General description}
The drag-based model relies on the hypothesis that all the interactions responsible for the launch of the CME cease in the upper corona, and that, beyond a certain distance, the dynamics of ICME propagation are governed mainly by its interaction with the ambient solar wind.
The DBM considers such an interaction by means of a drag force analogous to that experienced by a body immersed in a fluid.
The idea of an MHD analogous "hydrodynamical" drag is supported by the observation that ICMEs which are faster than the solar wind are decelerated, whereas those slower than the solar wind are accelerated by the ambient flow \citep{Gopalswamy+2000,Manoharan2006}.\\
Following \citet{Cargill2004}, we consider the relative speed dependence of the drag force in the radial direction:
\begin{equation}
F_d = -C_d A \rho(v-w)|v-w|
\end{equation}
where $v$ is the ICME radial speed and ${w}$ that of the solar wind, $A$ is the ICME cross-section, $\rho$ is the solar wind density and $C_d$ is an dimensionless coefficient for the drag force.
In a classical Newton's law framework, this leads to a radial drag acceleration in the form:
\begin{equation}
\label{dragaccel}
a = -\gamma({v}-{w})|{v}-{w}|
\end{equation}
where $\gamma$ is the so-called \emph{drag parameter} which contains the information about the ICME shape, mass, and in general about the effectiveness of the drag effect.\\
Considering the solar wind speed and the drag parameter as constants \citep[which is a good approximation beyond $20-40\,R_{\sun}$ ][]{Cargill2004, Vrsnak+2013}, equation (\ref{dragaccel}) can be solved explicitly, obtaining as functions of time the ICME speed:
\begin{equation}
v(t)=\frac{v_0-w}{1 \pm \gamma(v_0-w)t} + w
\label{vDBM}
\end{equation}
and the heliospheric distance:
\begin{equation}
r(t)=\pm \frac{1}{\gamma} \ln{\Big[ 1 \pm \gamma(v_0-w)t\Big]} + wt + r_0
\label{rDBM}
\end{equation}
where the $\pm$ signs apply to the cases $v_0>w$ and $v_0<w$, respectively, and $r_0$ and $v_0$ are the CME distance from the Sun and velocity at the onset time $t_0$. %
In this framework, the model needs four quantities, $[r_0, v_0, \textit{w}, \gamma]$, to compute the heliospheric distance and velocity of the ICME at any $t$.\\
The shape of the ICME we are modeling corresponds to type A) in Fig.9 of \cite{Schwenn+2005}, i.e. the front of the CME is a section of a sphere concentric with the Sun.
\subsection{The Probabilistic drag-based model}
As just stated, the DBM needs four quantities to be computed, namely $[r_0, v_0, \textit{w}, \gamma]$.
The first two quantities suffer from measure errors, while the last two are, in general, unknown.\\
If we consider the measure errors to be described by Gaussian PDFs, and assume a priori PDFs for both \textit{w} and $\gamma$, we can extend the DBM into a probabilistic approach.\\
The Probabilistic drag-based model (P-DBM henceforth), is a Monte-Carlo evaluation of the time of arrival and the velocity of the ICME at a chosen distance from the Sun, transforming the PDFs associated to the inputs into PDFs for the outputs, thus generating best estimates and errors for both the time of arrival and the velocity.
For each ICME whose $r_0$ and $v_0$ are measured, we can generate $N$ different $[r_0, v_0, \textit{w}, \gamma]$ initial conditions sets, randomly chosen from the relative PDFs, to compute via eqs. (\ref{vDBM}) and (\ref{rDBM}) the transit time and the velocity at 1 AU, for example.
This process generates the PDFs associated to $t_{1AU}$ and $v_{1AU}$, which can be used to estimate the ICME most probable time of arrival and velocity and their associated uncertainties at 1 AU.\\
Of course, the robustness of the results strongly depends on the validity of the assumptions, the realism of the PDFs, and on a thorough exploration of the parameter space, i.e. how large is $N$.
Given the simplicity of eq. (\ref{vDBM}) and (\ref{rDBM}) and the present computing capabilities, $N$ of the order of $10^4 - 10^6$ can be used to explore the parameter space and obtain nicely sampled output PDFs in a matter of seconds.
\subsection{PDFs for the input quantities}
In this section we introduce the PDFs which will be used for the four input quantities.
\begin{figure*}
\includegraphics[width=8.8cm]{hist_sw.jpg}
\includegraphics[width=8.8cm]{hist_gamma.jpg
\caption{Histograms of $w$ (a) and $\gamma$ (b) obtained by the inversion of \citet{Schwenn+2005} and \citet{Manoharan2006} catalogs.}
\label{hist_sw}
\end{figure*}
\begin{figure*}
\includegraphics[width=8.8cm]{pdf_sw.jpg}
\includegraphics[width=8.8cm]{pdf_gamma.jpg}
\caption{(a) PDF adopted for the random generation of $w$ in the P-DBM, with the slow \textit{w} represented by a Gaussian PDF centered at 400km/s with $\sigma=33~\textrm{km/s}$, and the fast \textit{w} represented by a Gaussian PDF centered at 600km/s with $\sigma=66~\textrm{km/s}$. (b) PDF adopted for the random generation of $\gamma$ in the P-DBM, modeled by a Log-Normal function with $\mu=-0.70$ and $\sigma=1.01$.}
\label{pdf_sw}
\end{figure*}
As \citet{Vrsnak+2013} have shown, the two equations (\ref{vDBM}) and (\ref{rDBM}) can be inverted to obtain the drag parameter $\gamma$ and the solar wind speed \textit{w}, if the initial position $r_0$ and speed $v_0$ of an ICME and its time of arrival $t_{1AU}$ and velocity $v_{1AU}$ at 1 AU are known.
\begin{equation}
\gamma=\frac{ (v_0-v_{1AU}) } {(v_0 - w) (v_{1AU}-w)t_{1AU}}
\label{gamma_inv}
\end{equation}
This equation can be used to compute directly $\gamma$ once one has numerically solved:
\begin{equation}
\frac{(v_0 - w) (v_{1AU}-w)t_{1AU}}{(v_0-v_{1AU})} \ln{\left[\frac{(v_0-v_{1AU})} {(v_{1AU}-w)} +1\right]} + w\,t_{1AU} + r_0 - r_{1AU} = 0
\label{rDBM_inv}
\end{equation}
to obtain $w$.\\
As in \citet{Vrsnak+2013}, we use the catalogs of ICMEs by \citet{Schwenn+2005} and \citet{Manoharan2006} to compute this inversion.
The first list consists of 91 CMEs between 1991 and 2001 for which the authors were able to uniquely associate ICME signatures in front of the Earth after careful inspections of the SOHO/LASCO CME catalog
and the complete LASCO/EIT data set.
The second list by Manoharan consists of 30 CME events between 1998 and 2004 whose heliospheric evolution has been investigated between the Sun and the Earth using LASCO coronagraphic images and interplanetary scintillation images of the inner heliosphere.
Therefore, these lists include CME events for which a safe association between a remote coronagraphical observation and an in-situ signature has been established, allowing the knowledge of quantities such as transit time and initial and final speed, required for the inversion.
From the results, we obtain the histograms reported in Fig.\ref{hist_sw} for \textit{w} and $\gamma$.
We choose larger bins than in the original work to make the obtained distributions more robust at the expense of sampling.
Apart from those differences, these distributions are of course consistent with \citet{Vrsnak+2013} results.\\
For \textit{w}, we can complement the distribution obtained by using the values of the solar wind recorded by SOHO, ACE, ULYSSES, HELIOS \citep{SOHO1, SOHO2, ACE1, ULYSSES1,ULYSSES2, HELIOS1} and many other missions.\\
The common understanding \citep[see][for a review]{Schwenn2006} is that there exist two different PDFs for the so-called slow (below $500~\textrm{km/s}$) and fast solar wind, the latter originated from the coronal holes, which are regions on the Sun with depressed UV emission and low magnetic activity.
The \textit{w} probability densities that we assume are plotted in Fig.\ref{pdf_sw}a, with the slow \textit{w} represented by a Gaussian PDF centered at $400~\textrm{km/s}$ with $\sigma=33~\textrm{km/s}$, and the fast \textit{w} represented by a Gaussian PDF centered at $600~\textrm{km/s}$ with $\sigma=66~\textrm{km/s}$.
Of course, such PDFs are limited to positive values of \textit{w}.
Following the works of \citet{Robbins+2006, Vrsnak+2007}, we adopt the fast \textit{w} PDF in those cases where there is a prominent coronal hole in the center on the disk, the slow \textit{w} PDF in all the other cases.\\
\hspace{1cm} For $\gamma$, we note that the the distribution retrieved by the inversion has a peak in the first bin ($0.2-0.4 \times 10^{-7}$ km$^{-1}$) and then decays, with an extended tail up to $\simeq 4 \times 10^{-7}$km$^{-1}$.
The skewed shape of the distribution suggested to fit ($\tilde{\chi}^2 = 1.13$) such a distribution with the 2-parameter Log-Normal function:
\begin{equation}
\label{Log-Normal}
f(x) = \frac{1}{\sigma\sqrt{2\pi}} e^{- \frac{(\ln x - \mu)^2}{2\sigma^2} }
\end{equation}
retrieving $\mu=-0.70$ and $\sigma=1.01$, to obtain an analytic form for the PDF, which is shown in Fig.\ref{pdf_sw}b.\\
Despite the fact that we are not putting forward any physical model for the CME kinematics, we must note that the Log-Normal distribution has been found to describe several aspect of solar wind plasma \cite[see][and references therein]{Burlaga+2000} and even the CME speed distribution \citep{Yurchyshyn+2005}.
In our case, the Log-Normal distribution just provides a good fit to the observed distribution, capturing its properties in just two parameters.\\
\hspace{1cm} For $r_0$, we consider that CME detection algorithms have inherent uncertainties for the CME location and the moment and duration of the CME liftoff.
From that, we assume that the PDF of $r_0$ can be modeled by a Gaussian PDF whose average is the last height derived by the CME tracking algorithm at the onset time and whose sigma is estimated from the associated error
\citep[$3\sigma\simeq1 R_{\odot}$ in the case of][]{Shi+2015}.\\
\hspace{1cm} Also for $v_0$, we assume a Gaussian PDF whose average value is the velocity measured by the CME tracking algorithm and whose sigma is the uncertainty associated to the measurement.\\
\subsection{P-DBM step-by-step}
To resume, here is a step-by-step description of how the P-DBM performs a prediction on the arrival of an ICME:
\begin{enumerate}
\item the position PDF is generated using the last measured CME height within coronagraph images and its associated error;
\item the velocity PDF is generated using the measured velocity and its associated error;
\item the Log-Normal PDF described by Eq.\ref{Log-Normal} ($\mu=-0.70$ and $\sigma=1.01$) is considered for the drag parameter;
\item a Gaussian PDF is chosen for the solar wind velocity, selecting either fast solar wind conditions (600 $\pm$ 66 $km/s$) in the case of a coronal hole in a relevant position of the solar disk, or slow solar wind otherwise (400 $\pm$ 33 $km/s$);
\item N initial condition sets $[r_0,v_0,\gamma,w]$ are randomly generated from those PDFs;
\item N different time of arrivals at 1AU $t_{1AU}$ are computed from Eq. \ref{rDBM}, by setting $t=t_{1AU}$ and $r(t_{1AU})=1AU$, and computing $t=t_{1AU}$ as the root of the equation via an iterative algorithm;
\item the time of arrival PDF is evaluated from the N $t_{1AU}$ values;
\item the best estimate for $t_C$ and its associated error are evaluated as the mean and the root mean square of the time of arrival PDF;
\item Steps 6, 7, 8 are also applied to Eq. \ref{vDBM} to evaluate the best estimate for $v_C$ and its associated error.
\end{enumerate}
\section{The Dataset}
\label{data}
In order to test the P-DBM described in the previous section, we use a sample of events from \citet{Shi+2015}.
For such events, a reconstruction of the ICME shape and speed has been obtained with the Graduated Cylindrical Shell model \cite[GCS - ][]{Thernisien+2006, Thernisien+2009} by means of triangulation of coronagraphic images taken from both STEREO and LASCO.
Following \citet{Shi+2015}, we excluded from the original sample those ICME which probably had interactions with the background magnetic field or other CMEs.
We also excluded entry 11 from the original sample which was most probably not correctly associated with the ICME arrival time \cite[cf. ][]{Mostl+2014}.
The details about how the CGS model has been used to fit the CME shapes and to determine the CME initial speeds and heights are reported in the original paper of \citet{Shi+2015}.
Here, we only recall that the authors estimated the errors of their detection and tracking procedure and that the CME speeds were evaluated through linear fits of the height versus time curves, and the associated error is the uncertainty of the linear fitting.
This will be used to estimate the width of the Gaussian PDFs associated to the CME position and velocity uncertainties and to update the original onset times of the events in \citet{Shi+2015} which are reported in the second column of Table \ref{TongSTEREO}.\\
These onset times are associated to the first detection in the instrument FOV, that is at 2.5 $R_{\odot}$.
In order to employ the DBM in the proper range of heliospheric distances, we choose to move the onset positions at the last useful detection in the instrument FOV at 15 $R_{\odot}$.
Consequently, we re-evaluated the onset time for each CME by adding a delay of $12.5R_{\odot}/v_0$, using the velocity $v_0$ (third column in Table \ref{TongSTEREO}) obtained by \citet{Shi+2015} through the linear fits of the CME positions exactly between $2.5R_{\odot}$ and $15R_{\odot}$.
The new onset times are reported in the fourth column of Table \ref{TongSTEREO}.
Consequently, for the purpose of this work, we can assume for each event a normal distribution of the height $r_0$ at the new onset time, with mean value $<r_0>=15\,R_{\odot}$ and standard deviation $\sigma_r=0.33\, R_{\sun}$.\\
Furthermore, it must be observed that for the events from the paper by \citet{Shi+2015} the ICME arrival time is referred to the time of first occurrence of an ICME signature in the near Earth environment, which in most cases is the time of arrival of the fore-shock.
To perform a correct validation of the CME transit time forecast, we want to consider the arrival of the ICME leading edge, instead of that of the shock \citep[see also the discussion in][]{Schwenn+2005, Vrsnak+2014}.
To this purpose, for each event, we checked for the time of arrival of a plasma driven effect (Magnetic clouds or Ejecta), as reported in the GMU CME/ICME list compiled by Phillip Hess and Jie Zhang (\url{http://solar.gmu.edu/heliophysics/index.php/GMU_CME/ICME_List}).
In 10 out of 14 cases, we could correct the arrival times.
Column five of Table \ref{TongSTEREO} reports the arrival date and time of the CME, taking into account this update.\\
The last column of Table \ref{TongSTEREO} reports the condition of the solar wind associated with the CME, obtained by the inspection of suitable coronal images and verified by using data recorded by ACE \citep{ACE1}.
\begin{table*}[ht!]
\centering
\footnotesize
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{c c c c c c}
\# &Onset time @ 2.5 $R_{\odot}$&{v}$_{0}\pm 3\sigma$ [km/s] &Onset time @ 15 $R_{\odot}$ &Arrival $@$ 1AU &SW\\
1 &12 Dec 2008, 08:37 &$363 \pm 23$ &12 Dec 2008, 15:16 &17 Dec 2008, 02:00 &S\\% &113.4
2 &03 Apr 2010, 09:54 &$864 \pm 7$ &03 Apr 2010, 12:42 &05 Apr 2010, 08:00 &F\\%&46.1 \\
3 &08 Apr 2010, 03:39 &$512 \pm 34$&08 Apr 2010, 08:22 &12 Apr 2010, 02:00 &S\\
4 &16 Jun 2010, 14:39 &$222 \pm 2$ &17 Jun 2010, 01:32 &21 Jun 2010, 08:00 &S\\
5 &15 Feb 2011, 02:24 &$769 \pm 12$&15 Feb 2011, 05:33 &18 Feb 2011, 03:00 &S\\%&72.6 \\
6 &04 Aug 2011, 04:39 &$1512 \pm 90$&04 Aug 2011, 06:15 &06 Aug 2011, 12:00 &F\\%&38.4 \\
7 &14 Sep 2011, 00:39 &$505 \pm 5$ &14 Sep 2011, 05:26 &17 Sep 2011, 10:00 &S\\%&73.3 \\
8 &19 Jan 2012, 14:54 &$1299 \pm 16$&19 Jan 2012, 16:46 &23 Jan 2012, 00:00 &S\\%&62.1 \\
9 &13 Mar 2012, 17:39 &$1616 \pm 17$&13 Mar 2012, 19:09 &15 Mar 2012, 19:00 &F\\%&43.4 \\
10 &19 Apr 2012, 15:39 &$607 \pm 15$ &19 Apr 2012, 19:38 &23 Apr 2012, 12:00 &S\\%&82.8 \\
11 &12 Jul 2012, 16:39 &$1224\pm14$ &12 Jul 2012, 18:37 &15 Jul 2012, 06:00 &F\\%&48.4 \\
12 &28 Sep 2012, 00:39 &$1104 \pm 112$&28 Sep 2012, 02:50 &01 Oct 2012, 12:00 &S\\%&70.3 \\
13 &05 Oct 2012, 03:39 &$558 \pm 21$&05 Oct 2012, 07:59 &08 Oct 2012, 18:00 &S\\%&73.3 \\
14 &27 Oct 2012, 16:54 &$340 \pm 28$&28 Oct 2012, 00:00 &01 Nov 2012, 00:00 &S\\
\end{tabular}
\caption{Sample of events from \citet{Shi+2015} employed to test the P-DBM.
Columns are in order: CME index number, CME onset date and time (UT) at 2.5 $R_{\odot}$, CME initial speed with associated uncertainty, CME onset date and time (UT) at $15 R_{\odot}$, arrival date and time (UT) of the ICME at 1A, solar wind (Slow/Fast) during the CME propagation.}
\label{TongSTEREO}
}
\end{table*}
\section{Validation of the P-DBM}
\label{test}
We apply the probabilistic approach in order to generate the transit time distribution for each event in Table \ref{TongSTEREO}.
For this run, the number of forecast realizations has been set to $N=50000$ and it took less than a minute to obtain the results on a desktop PC.\\
As example, we show in Fig. \ref{Final}a the distribution of the transit times $t_i$ computed by the P-DBM for the first CME of the sample.
As a result from the input distributions, the travel times range from 80h to 120h, with a median value of 103.8h.
The distribution is not symmetric, slightly skewed towards the shorter times.
However, it is viable to describe this distribution by its mean value $t_C=103.1\textrm{h}$ and its root mean square $\sigma=4.4 \textrm{h}$.\\
The results for the whole sample, with $t_C$ and $\sigma$ of the arrival time distributions taken as the measure of the predicted arrival time, are reported in the third column of Table \ref{Table2}.
In the second column, instead, we report the observed transit time $t_O$ computed as the difference between the onset time and the arrival time at 1AU of Table \ref{TongSTEREO}.\\
\begin{table}[ht!]
\centering
\footnotesize
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{c c c c c }
\# &$t_O$ [h] &$<t_C>\pm \sigma$ [h] & $t_O-<t_C>$ [h] \\
1 & 106.7 & $103.1\pm 4.4$ & 3.6 \\
2 & 43.3 & $53.4\pm 5.8$ & -10.1 \\
3 & 89.6 & $83.9\pm 6.0$ & 5.8 \\
4 & 102.5 & $120.9\pm 14.3$ & -18.4 \\
5 & 69.4 & $72.5\pm 10.7$ & -2.9 \\
6 & 53.7 & $44.9\pm 9.3$ & 8.8 \\
7 & 76.6 & $84.3\pm 5.8$ & -7.7 \\
8 & 79.2 & $62.7\pm 15.2$ & 16.5 \\
9 & 47.8 & $44.2\pm 9.7$ & 3.7 \\
10 & 88.4 & $78.3\pm 8.1$ & 10.1 \\
11 & 59.4 & $47.7\pm 8.1$ & 11.7 \\
12 & 81.2 & $65.5\pm 14.0$ & 15.7 \\
13 & 82.0 & $80.8\pm 7.0$ & 1.2 \\
14 & 96.0 & $106.8\pm 5.8$ & -10.8 \\
\end{tabular}
\caption{Results from the P-DBM statistical simulation for the events in Table~\ref{TongSTEREO}. In the first column the CME index as in \ref{TongSTEREO}, in the second column the ICME transit time $t_O$ from 15$ R_{\odot}$ to 1AU, in the third column the computed CME transit time $t_C$ with the associated error $\sigma$. In the fourth column the difference $t_O-t_C$.}
\label{Table2}
}
\end{table}
\noindent
Fig. \ref{Final}b shows a plot of $t_C$ with $1 \sigma$ error bars versus $t_O$ and a least squares linear fit to these data.
The two datasets are evidently highly correlated, with a correlation coefficient $R=$0.87.
The linear fit performed on the data ($\tilde{\chi}^2=$1.66) retrieved a slope of $1.00\pm0.1$ and a constant value of $3\textrm{h}\pm8\textrm{h}$.
Given those values, the P-DBM results are compatible with the $t_C=t_O$ hypothesis.\\
Similarly to \citet{Colaninno+2013}, we plot the residuals $t_O-t_C$ and the error associated to $t_C$ for the 14 CMEs in Fig. \ref{hist_residui}a to allow an easy comparison of the forecast results.
In particular, for 7 CMEs out of 14 the forecast residuals are within the error.
Also, we report in Fig. \ref{hist_residui}b the histogram of the residuals.
It can be noted that $80\%$ of the forecasts are within 15h of the actual $t_O$, and just one is beyond 20h.
The distribution is compatible with a Gaussian function, centered in zero and with a $\sigma \simeq 10.6\textrm{h}$, with a marginal partiality towards forecasts behind of the observed times.
To conclude, we computed the average of the absolute value of the residuals $<|\Delta t|>$ = 9.1h, which is often used in the literature to assess the forecast accuracy.\\
\begin{figure*}
\includegraphics[width=8.8cm]{event1histogram_freq_in_percent1.jpg}
\includegraphics[width=8.8cm]{all_events_fit.png}
\caption{(a) Distribution of the transit times $t_i$ calculated for event \#1 in Table~\ref{TongSTEREO}. N=50000 initial conditions are generated in the P-DBM. (b) Dots with error bars are the forecast transit times $t_C$ versus observed transit times $t_O$. The solid line shows a linear fit to the data.}
\label{Final}
\end{figure*}
\begin{figure*}
\includegraphics[width=8.8cm]{residui.png}
\includegraphics[width=8.8cm]{hist_residui.png}
\caption{(a) The residuals $t_O-t_C$ and the error associated to $t_C$ for the 14 CMEs. (b) Distribution of the residuals $t_O-t_C$.}
\label{hist_residui}
\end{figure*}
\section{Conclusions and future work}
\label{conclusions}
In this work, we predicted the transit time between the Sun and the Earth for a sample of 14 CME events.
These events were selected among the database of \cite{Shi+2015}, for which the onset time, initial velocity and transit time are known.
By using the DBM \citep{Vrsnak+2013} and a probabilistic approach, we were able to associate an error to the transit time we computed, assuming that all the input parameters could be described by suitable PDFs.\\
For the shape we adopted to model the ICME and since all these ICMEs hit Earth, we did not use the CME principal direction nor the angular width to compute the transit time from the initial parameters.
Nevertheless, it is straightforward to modify the P-DBM to include a different CME shape and to consider the PDFs also for those two input parameters.
Given the very short time needed to compute a CME transit time distribution with this approach, adding two dimensions to the parameter space to be explored should be still feasible with undemanding computational resources.\\
Even with a model as simple as this, the results of the probabilistic approach are extremely promising:\\
$\cdot$ the scatter plot of $t_C$ vs $t_O$ has a slope which is unity within the errors:\\
$\cdot$ the histogram of the residuals $\Delta~t= t_o -t_C$ has a Gaussian shape, centered in zero and with a $\sigma \simeq 10.6\text{h}$;\\
$\cdot$ the average of the absolute value of the residuals is $<|\Delta~t|>$ = 9.1h.\\
However, less than half of the residuals is within the $1\sigma$ error associated to $t_C$, which is under-performing for a Gaussian distribution of the associated error.
This can either be due to a statistical fluctuation (given the small dimension of the test set) or to an under-estimate of the input PDF widths.
Of course, this disagreement may also arise from the model assumptions. In its simplicity, this DBM implementation models the ICME front as a portion of a sphere concentric with the Sun, therefore neglecting the difference between the ICME apex position and velocity and the ICME position and velocity on the ecliptic plane. On the other hand, this assumption reduces the number of PDFs needed by the model. While it is unclear whether increasing the model complexity will significantly reduce the discrepancy or not, especially considering the intrinsic difficulty in measuring the actual travel times (errors and bias can arise both from the onset and the arrival time estimates), it is instead possible that the PDF we used for $\gamma$, evaluated from actual data, may have incorporated most of such complexity, thus including these effects in the model in a statistical way.
At present, we can conclude that the chosen PDFs led to good estimations of the average times on transit time forecasts, but we need a larger sample to properly evaluate the robustness of the associated errors.
This is probably the main task for future work.
\\
There is a vast literature to compare our results with.
We limit ourselves to cases where the authors employed data with projection effects eliminated (measures in quadrature or multi-spacecraft plus CGS model), as in our test.\\
\cite{Gopalswamy+2001} found an empirical relation between the initial CME velocity and its acceleration and applied this relation to a model to compute the Time of Arrival (ToA) at Earth.
They were able to forecast the ICME ToA at 1AU with a mean error $<|\Delta~t|> = 10.7\textrm{h}$ and $72\%$ of the events had ToA within $\pm 15\textrm{h}$ from the predicted values\\
\cite{Owens+2004} tested on a 35 CME sample three different models: a model with a constant acceleration \cite{Gopalswamy+2000}, a model with an acceleration which ceased before 1AU \cite{Gopalswamy+2001}, and the original aerodynamic drag model \cite{Vrsnak+2002}.
These three model were best fitted on the sample and their $<|\Delta~t|>$ varied from 12h to 9h.\\
\cite{Schwenn+2005} derived an empirical correlation between halo CME expansion speeds and travel times to 1AU, fitting a straight forward deceleration model assuming viscous drag on the data from 75 halo CME events.
For $95\%$ of those events, the shock associated to the CME arrived within $\pm 24\textrm{h}$ of the predicted time.\\
\cite{Colaninno+2013} found that a first-order polynomial to the height-time measurements beyond $50 R_{\sun}$ (0.23 AU) was the best parameter for predicting the CME ToA at 1AU.
For a sample of 9 CME, they were are able to predict their ToA to within $\pm ̇13\textrm{h}$. It is worth to stress that they supplemented their data with STEREO/HI observations, thus increasing the accuracy of their CME initial parameter estimation.\\
\citet{Taktakishvili+2009} instead evaluated the performances of the ENLIL MHD simulation fed with a cone model of CME for a sample of 14 events. They reported an average absolute error of 6 hours, which is also very similar to the error reported by \citet{Millward+2013} of 7.5 hours, obtained again with ENLIL simulation initialized with CME parameters obtained via the CME Analysis Tool (CAT), but on a larger (25 events) set.
\cite{Vrsnak+2014} compared the CME arrival time prediction based on the DBM against ENLIL.
They reported estimation errors of about 14h with standard deviation ranges from 14h to 19h, depending on the sample and method.\\
\cite{Shi+2015} used a multi-parametric best fit on the transit time versus the initial speed for different drag based regimes.
Depending on the regime, they were able to reach a mean error $<|\Delta~t|>$ down to 6.7h.
Since we employed exactly their sample to test the P-DBM, we can note that their model performed better than ours on this sample.\\
It is worth to stress that \cite{Shi+2015} (and all the authors previously cited) fitted their distribution to the data, therefore optimizing the model to that dataset.
Our approach, in contrast, used two datasets to build the PDFs and was tested against an independent dataset, thus providing a true a-priori forecast test.\\
As already stated, we are aware that our results are based on the analysis and comparison of a very limited dataset.
Among the next steps in the further validation of this approach, is the test with a larger database of ICME.
Since databases which provide information sufficient to fully characterize the ICME are difficult to retrieve, we are already taking into consideration the possibility of having much less information on the ICME onset and morphology.\\
Therefore, we are working to include both the uncertainty on the angular extension, the uncertainty on the main direction and on the de-projected velocity of the CME in the P-DBM, again, modeled by PDFs.
At present, we are also working on a real-time implementation of the P-DBM which ingests the parameters of ICMEs tracked by the CACTUS software \citep{Robbrecht+2004} and forecast the time of arrival at 1AU of the ICMEs and their velocity, of course with the associated errors.
As a result, we will build up a database of the results and we plan to verify and possibly re-consider the PDFs we have chosen for the input parameters.\\
Also, we are pondering an evolution to consider a different morphology of the ICME, passing from the cone model (and its intersection with the ecliptic plane) to a 3-D light-bulb model or a similar model \citep[e.g.:][]{Kleimann2012}.\\
All these effects significantly alter the travel time and it is worth to explore how the P-DBM can include them in its probabilistic approach.\\
Since a complete and real-time stereoscopic determination of the CME morphology and propagation will not be available in the near future, all the 3D effects which are not taken into account in the present P-DBM should be considered as partially unknown variables
and should be modeled by suitable PDFs, constrained by as much information as available.
As example, the real width, direction and velocity of the CME have to be evaluated from images which suffer from projection effects. There are several ways to de-project the data, which imply different assumptions. One of the simplest \citep{Zhao+2002} assumes that the cone-shaped CME has its vertex in the Sun's center and its axis normal to the solar surface at the position of a relevant solar magnetic feature (erupting filament or flaring AR). In such a case, this information, error propagation theory and previous CME parameters statistics could be used to generate the PDFs needed to propagate the ICME with the P-DBM.
It is likely that adding other uncertainties from other input parameters will enlarge the error associated with the forecast, but it is important to stress again that the P-DBM light computation needs make it interesting to evaluate the propagation of any ICME in any portion of the inner solar system.\\
To conclude, the accurate prediction of the time of arrival of an ICME to Earth or other interesting part of the Heliosphere \citep[e.g.:][]{Falkenberg+2010MARS} is of critical importance for our high-technology society and for any future manned exploration of the solar system.
We think that as critical as the prediction accuracy is the knowledge of precision, i.e. the error associated to the forecast.
The method we presented here, building on the DBM model of \cite{Vrsnak+2013}, is capable to predict the arrival time of ICMEs to the Earth and its uncertainty with minor computation necessities, providing a forecast of the space weather in the near Earth environment with a 2-day horizon.\\
\acknowledgement{This research work has been partly supported by the Italian MIUR-PRIN grant 2012P2HRCR on "The active Sun and its effects on Space and Earth climate" and by Space Weather Italian COmmunity (SWICO) Research Program, from the Regione Lazio FILAS-RU-2014-1028 grant on "Banca Dati di Space Weather da Strumenti nello Spazio ed a Terra", and from the EC Tender No.434/PP/GRO/RCH/15/8381 for the "Ionosphere Prediction Service".\\
GN wishes to take this opportunity to express his sincere appreciation for the PhD grant from the Universit\`a degli Studi dell'Aquila, for the supplies and facilities placed at his disposal, and for providing the opportunity to work on this project.\\
The authors thank R. Schwenn for sharing the CME database used in \cite{Schwenn+2005}.\\
The authors thank the anonymous referees for their insightful and helpful comments on earlier versions of the paper.
The editor thanks two anonymous referees for their assistance in evaluating this paper.\\
}
|
\section{Introduction}
In the context of General Relativity, photos emitted from
distant galaxies are continuously deflected by the intervening
mass field of the large-scale structures \citep{1992grle.book.....S,
2006glsw.conf.....M, 2016arXiv161206535B}. This
gravitational lensing effect, referred to as ``cosmic shear", produces
some coherent distortions of the observed galaxy images,
which can be measured to probe the matter distribution in the universe
\citep{1999ARA&A..37..127M, 2001A&A...374..757V, 2003astro.ph..9482K,
2008A&A...479....9F, 2015RPPh...78h6901K, 2016MNRAS.463.3326F}.
Great progress has been made in using cosmic shears to constrain
cosmological models (see \citealt[][]{2015RPPh...78h6901K} for a review), to estimate the dark
energy parameter $w$ \citep{2007NJPh....9..444B, 2009JCAP...06..026L,
2015JCAP...04..048B}, and to test theories of modified
gravity \citep{2015PhRvD..92f4024L, 2016MNRAS.459.2762H}.
Observational results from recent weak-lensing
surveys, such as Canada-France-Hawaii Telescope Lensing Survey
\citep[CFHTLenS, ][]{2012MNRAS.427..146H, 2013MNRAS.432.2433H}
and Deep Lens Survey
\citep[DLS, ][]{2013ApJ...765...74J, 2016ApJ...824...77J, 2016AAS...22730707J},
demonstrate that cosmic shears can be combined with other observations,
such as the cosmic microwave background (CMB), baryon acoustic
oscillations (BAO), and galaxy cluster abundance, to break the degeneracy
among different cosmological parameters (e.g., $\Omega_{\rm m}$-$\sigma_{8}$).
Thus, accurate measurement of weak lensing effects has been one of the main
goals of many ongoing and upcoming galaxy surveys, such as the Kilo-Degree Survey
\citep[KiDS,][]{2015A&A...582A..62D}, Dark Energy Survey
\citep[DES,][]{2016MNRAS.460.1270D}, Hyper Suprime-Cam Survey
\citep[HSC, ][]{2012SPIE.8446E..0ZM}, Euclid
\citep{2011arXiv1110.3193L} and the Large Synoptic Survey Telescope
\citep[LSST, ][]{2009arXiv0912.0201L}.
These surveys will provide high-quality data with very wide
sky coverages, and the statistical uncertainties in weak lensing
measurements are expected to be small. However,
high accuracy analyses of cosmic shear also requires
understanding the systematics in the data, such as those in
measurements of galaxy ellipticity and the point spread function.
In addition, accurate theoretical modeling is also necessary in order to
interpret the observed data.
One of the most serious astrophysical systematic
effects in the era of accurate weak-lensing analyses is the
intrinsic alignment (IA) of galaxies (\citealt{2000MNRAS.319..649H,
2001ApJ...559..552C, 2002MNRAS.335L..89J}; see
\citealt{2015SSRv..193..139K, 2015PhR...558....1T} for a review),
which can mimic the gravitational lensing signal, thereby contaminating
the measurements of cosmic shears.
A significant signal of intrinsic alignments has been detected by
\citet{2006MNRAS.367..611M} in the luminous red galaxies from
the Sloan Digital Sky Survey \citep[SDSS,][]{2000AJ....120.1579Y}, and
the authors concluded that neglecting such alignments can lead to an underestimate
of the linear amplitude of density fluctuations by $20\%$ for cosmic
shear surveys at $z \sim 1$. Clearly, an accurate model for galaxy IA,
which is capable of describing its dependence on redshift and galaxy
properties, is crucial for maximizing the science returns of
ongoing and future weak lensing surveys \citep{2016MNRAS.456..207K}.
There have been numerous investigations on galaxy IA.
Based on the tidal field theory \citep{2001MNRAS.320L...7C},
\citet{2004PhRvD..70f3526H} developed a linear model for galaxy IA,
which was later improved to include some non-linear effects
\citep{2007NJPh....9..444B, 2012JCAP...05..041B}. A useful formula
with a single parameter was introduced by \citet{2011A&A...527A..26J},
which can easily be included in the analyses of cosmic shears
from observational data
\citep{2010MNRAS.408.1502K, 2013MNRAS.432.2433H, 2016ApJ...824...77J,
2017MNRAS.465.1454H, 2017MNRAS.465.2033J}.
As a more accurate description of galaxy IA, a halo model
is developed \citep{2010MNRAS.402.2127S}, which can predict the
IA signal as a function of galaxy properties. However, as pointed in
\citet{2013MNRAS.431..477J}, most of these simple IA models
are expected to work only at low $z$, and it is still unclear how
galaxy IA varies as a function of galaxy properties at high $z$.
$N$-body and Hydro-dynamical simulations are also extensively used to
study galaxy IA. When $N$-body simulations are used for the purpose,
assumptions about the connection between galaxy shape and dark matter halo shape
have to be made. \citet{2007MNRAS.378.1531K} used $N$-body simulations to
explain the observed small-scale alignment of satellites galaxies
around central galaxies in the SDSS data \citep{2006MNRAS.369.1293Y}.
They found that the orientations of elliptical galaxies follow that of the host halos,
albeit with some mis-alignment, and that the spins of spiral galaxies
follow that of their host halos. This assumption is later confirmed
\citep[e.g.,][]{2009RAA.....9...41F,
2009ApJ...694..214O, 2010ApJ...709.1321A}. With similar assumptions
about how galaxies are aligned with dark matter halos,
\citet{2013MNRAS.436..819J} measured galaxy IA on large scales from
the Millennium Simulations \citep{2005Natur.435..629S} and found that
early-type galaxies are strongly aligned with each other, but
spiral galaxies do not show significant correlation signals between their intrinsic
ellipticities. This dependence on galaxy-type agrees with observational results.
\citep[e.g.,][]{ 2011A&A...527A..26J,
2013MNRAS.432.2433H, 2011MNRAS.410..844M}. More recently, cosmological
hydrodynamical simulations have been used to predict the galaxy
IA \citep[e.g.,][]{2014ApJ...791L..33D, 2014MNRAS.441..470T,
2015MNRAS.454.2736C, 2016MNRAS.461.2702C, 2017MNRAS.472.1163C,
2015MNRAS.454.3328V, 2017MNRAS.468..790H, 2016MNRAS.462.2668T}.
The main merit of using a hydro-dynamical simulation is that galaxy shapes
are directly predicted by the simulation. In agreement with previous
analytical models and $N$-body simulations, these hydro-dynamical
simulations also indicate that elliptical galaxies have a stronger tendency
to align with each other on large scales than do spiral galaxies.
However, due to limited volumes of these simulations (often around
100Mpc/$h$) and different treatment of baryonic physics, the predicted
galaxy IA signal and its dependence on galaxy properties and redshift
still varies from simulation to simulation.
Although an accurate model for galaxy IA is still not available at
the present, the main assumption, adopted in $N$-body
simulations, that elliptical galaxies follow the shapes,
while spirals follow the spins, of host halos
\citep[e.g.,][]{2013MNRAS.436..819J}, can be checked by
comparing real and mocked observational data of galaxy shear
correlations. This can be achieved by using ray tracing in an $N$-body
simulation combined with a model of galaxy formation which can predict
galaxy shapes, luminosities and positions. With such an approach, we can
produce observable images of galaxies and obtain the auto- and
cross-correlation functions between gravitational shear and galaxy
intrinsic ellipticity at different redshifts. We can then compare
model predictions with results obtained from two recent surveys, KiDS and
DLS, and examine the importance of galaxy IA. The results of these two
surveys show a $\sim 2\sigma$ tension in
$S_{8} \equiv \sigma_{8} \sqrt{\Omega_{\rm m}/0.3}$,
with KiDS giving $S_{8} = 0.745 \pm 0.039$ and DLS giving
$0.818_{-0.026}^{+0.034}$. The main goal of this paper is to use
such approach to constrain galaxy IA models and to examine
the contamination from IA in the 2-point correlation functions
of the cosmic shear.
As a `standard' algorithm, the multiple-plane ray-tracing simulation
with the flat-sky approximation
\citep[e.g.,][]{2000ApJ...530..547J, 2004APh....22...19W, 2009A&A...499...31H}
has been widely used to simulate lensing
maps for small-field survey. It also roughly works for hundreds of square degree
surveys, such as KiDS with 450 square degrees \citep[hereafter KiDS-450,][]{2017MNRAS.465.1454H},
but will not suitable for even large-filed surveys such as Euclid and LSST \citep{2016arXiv161104954K, 2017arXiv170205301K, 2017JCAP...05..014L}.
To quantify the effect of cosmic variance in the small-field surveys,
one needs construct a lot of light cones to simulated different realizations.
In this paper, we adopt ray-tracing code on a curved sky to simplify
this procedure and to prepare for these large-field surveys.
Full-sky weak-lensing maps have already been constructed in
a number of papers
\citep{2009A&A...497..335T, 2013MNRAS.435..115B,
2008MNRAS.391..435F, 2015MNRAS.447.1319F,
2015MNRAS.453.3043S}. These simulations usually cover a sufficiently large
volume to compute a full-sky convergence (and shear) maps, and explore
the lensing power at both the linear and nonlinear regimes.
In this paper, we follow the ray-tracing method of
\citet{2008ApJ...682....1D, 2009A&A...497..335T, 2013MNRAS.435..115B}.
We perform high resolution (both in space and in mass) lensing simulations,
using an iterative scheme of spherical harmonic analysis, to model
lensed properties of 'semi-analytic' galaxies in the simulation.
These simulated galaxies allow us to study the statistical properties
of galaxy alignments, and to compare our mock observations with the
observational results from both DLS \citep{
2016ApJ...824...77J, 2016AAS...22730707J} and KiDS-450
\citep{2017MNRAS.465.1454H} using tomographic correlation functions.
The paper is organized as follows. In Section 2 we first summarize the
basic theoretical background of weak lensing, focusing on the power
spectrum and shear correlation analyses. In Section 3 we introduce
the simulations and the spherical ray-tracing technique. Section 4 describes
how we model galaxy properties, such as luminosity, morphology and shape, from the
semi-analytic model, and we also present results of
intrinsic alignments of galaxies and their dependence on galaxy type
and halo mass. In Section 5, we describe the tomographic analyses of cosmic
shears in our lensing simulation, compare model predictions with
observational data, and quantify the contributions of the
intrinsic-intrinsic (II) shear correlation and the gravitational shear-intrinsic
(GI) shear correlation by spiral and elliptical galaxies. Conclusions and
discussions are given in Section 6.
\section{COSMOLOGICAL WEAK LENSING}
In this section, we briefly summarize the theoretical background for the
analyses of weak gravitational lensing and describe some basics about
intrinsic alignment and shear correlations.
\subsection{Basics}
In general, for a source galaxy with the observed angular position
$\bm{\theta}$ and its real position $\bm{\beta}$, one can
characterize the deformation effect of cosmic shear through
the distortion matrix \citep{1992grle.book.....S,
2000ApJ...530..547J},
\begin{equation} \label{equ:matrixA}
\mathcal{A}(\bm{\theta})=\frac{\partial\bm{\beta}}{\partial\bm{\theta}}
\equiv\begin{pmatrix}1-\kappa-\gamma_1 & -\gamma_2-\omega \\ -\gamma_2+\omega & 1-\kappa+\gamma_1 \end{pmatrix},
\end{equation}
where $\kappa$ is the convergence,
$\bm{\gamma}=\gamma_1+{\rm i}\gamma_2$ defines the complex shear in
lensing, and the additional antisymmetric quantity, $\omega$, describes
an overall rotation in the lensed images. In the weak lensing regime
(i.e., $\kappa,\bm{\gamma} \ll 1$) and to the linear order,
the components of the matrix are related to the second derivatives of the
gravitational potential as \citep{2001PhR...340..291B,
2009A&A...499...31H, 2015RPPh...78h6901K}
\begin{equation}
\mathcal{A}_{ij}(\bm{\theta}, \chi)=\delta_{ij}-\frac{2}{c^2}\int_{0}^{\chi} {\rm d}\chi'\frac{r(\chi-\chi')r(\chi')}{r(\chi)}\Phi_{,ij}(r(\chi')\bm{\theta},\chi'),
\end{equation}
where $\delta_{ij}$ is the Kronecker delta, $c$ is the speed of
light, $\chi$ is the comoving distance and $r(\chi)$ the
comoving angular diameter distance. According to the Poisson
equation, gravitational potential $\Phi$ can be related to the density
contrast $\delta$. Hence, convergence $\kappa$ can be
expressed as a weighted integral of the over-density $\delta$ along the line
of sight,
\begin{equation}
\kappa(\bm{\theta}, \chi)=\frac{3H_{0}^{2}\Omega_{{\rm m}}}{2c^2}\int_{0}^{\chi}{\rm d}\chi'\frac{r(\chi-\chi')r(\chi')}{r(\chi)}\frac{\delta(r(\chi')\bm{\theta},\chi')}{a(\chi')},
\label{equ:kappa}
\end{equation}
where $H_{0}$ is the Hubble constant, $\Omega_{{\rm m}}$ is
the matter density in units of the critical density,
and $a(\chi')$ is the scale factor at $\chi'$.
\subsection{Power Spectrum of Weak Lensing Field}\label{sec:introPS}
In the flat-sky limit, the power spectrum of the convergence
$C^{\kappa\kappa}(\ell)$ on the modulus $\ell$ is known as the two-point
correlation in Fourier space,
\begin{equation}
\langle\tilde{\kappa}(\bm{\ell}) \tilde{\kappa}^{*}(\bm{\ell}')\rangle=(2\pi)^2\delta_{\rm D}(\bm{\ell}-\bm{\ell}')C^{\kappa\kappa}(\ell),
\end{equation}
where $\delta_{\rm D}(\bm{\ell})$ is the Dirac delta function.
Using equation~(\ref{equ:kappa}), one can derive the angular power
spectrum of the convergence field in the Limber approximation,
\begin{equation}
C^{\kappa\kappa}(\ell) = \int^{\chi_{\rm H}}_{0} {\rm d}\chi \frac{W(\chi)^{2}}{r(\chi)^{2}} P_{\delta}\left(k=\frac{\ell}{r(\chi)}, \chi\right),
\label{eq:Ckk}
\end{equation}
where $P_{\delta}(k,\chi)$ is the 3-D power spectrum of the matter
distribution at the given comoving distance $\chi$, and the integral
is calculated along the line of sight to the comoving horizon distance
$\chi_{\rm H}$. Here the weight function $W(\chi)$ is defined as,
\begin{equation}
W(\chi) = \frac{3H_{0}^{2}\Omega_{{\rm m}}}{2c^2}\frac{r(\chi_{\rm H}-\chi)r(\chi)}{r(\chi_{\rm H})} \frac{1}{a(\chi)}.
\end{equation}
From the non-linear theoretical models, such as the Halofit model
\citep{2003MNRAS.341.1311S, 2012ApJ...761..152T}, one can predict the
convergence power spectrum $C^{\kappa\kappa}(\ell)$ from the
non-linear $P_{\delta}(k)$. Therefore, the weak lensing survey can be
used to probe the gravitational growth of the density structure.
While dealing with full-sky measurements, it is useful to note that
the angular power spectrum of weak lensing convergence $\kappa$
and complex shear $\bm{\gamma}$ can be derived
from the spin-$s$ spherical harmonics $_{s}Y_{\ell}^{m}$
\citep{1996astro.ph..9149S}. The relations of power spectra
between the convergence, shear E- and B-mode
\citep{2002A&A...389..729S, 2003NewAR..47..987B, 2003PhRvD..67b3501B,
2010PhRvD..82b3001Z} have been derived by
\citet{2000PhRvD..62d3007H} for an all-sky lensing deformation tensor
field. Here we briefly summarize the spin-$s$ spherical harmonic
decomposition of the full-sky lensing, referring the reader to
\citet{2000PhRvD..62d3007H} for detailed discussions of the power
spectrum in weak lensing.
As reviewed in the appendix of \citet{2013MNRAS.435..115B}, the convergence,
lensing shear and rotation in the distortion matrix (Eq.~\ref{equ:matrixA}) can be
decomposed by the spherical harmonics \citep{2000PhRvD..62d3007H, 2013MNRAS.435..115B},
\begin{equation}
\kappa(\hat{\bm{n}}) = -\frac{1}{2}\sum_{\ell m}\ell(\ell+1)\phi_{\ell m}Y_{\ell}^{m}(\hat{\bm{n}})
\end{equation}
\begin{equation}
\gamma_{1}(\hat{\bm{n}}) \pm i\gamma_{2}(\hat{\bm{n}}) =
\frac{1}{2}\sum_{\ell m}\sqrt{\frac{(\ell+2)!}{(\ell-2)!}} (\phi_{\ell m} \pm i \Omega_{\ell m}) _{\pm2}Y_{\ell}^{m}(\hat{\bm{n}})
\end{equation}
\begin{equation}
\omega(\hat{\bm{n}}) = -\frac{1}{2}\sum_{\ell m}\ell(\ell+1)\Omega_{\ell m}Y_{\ell}^{m}(\hat{\bm{n}}),
\end{equation}
where $\phi$ is the lensing deflection potential, $\Omega$ is
the pseudo-scalar potential \citep[as described by][]{1996astro.ph..9149S},
and $\hat{\bm{n}}$ denotes a given position on the sky.
Consequently, the different power spectra can be related as,
\begin{equation}
C^{\kappa\kappa}(\ell) = \frac{1}{4}\ell^2 (\ell+1)^2 C^{\phi\phi}(\ell)
\end{equation}
\begin{equation}
C^{\omega\omega}(\ell) = \frac{1}{4}\ell^2 (\ell+1)^2 C^{\Omega\Omega}(\ell)
\end{equation}
\begin{equation}
C^{EE}(\ell) = \frac{1}{\ell^2 (\ell+1)^2} \frac{(\ell+2)!}{(\ell-2)!} C^{\kappa\kappa}(\ell)
\end{equation}
\begin{equation}
C^{BB}(\ell) = \frac{1}{\ell^2 (\ell+1)^2} \frac{(\ell+2)!}{(\ell-2)!} C^{\omega\omega}(\ell).
\end{equation}
Thus in the flat-sky limit, one has $\frac{1}{\ell^2 (\ell+1)^2} \frac{(\ell+2)!}{(\ell-2)!} \approx 1$,
showing that $C^{EE}(\ell)$ equals $C^{\kappa\kappa}(\ell)$ at small scales \citep{2016arXiv161104954K, 2017arXiv170205301K}.
\subsection{Cosmic Shear and Intrinsic Alignment}
Weak lensing will induce an additional coherent deformation to the
intrinsic galaxy shape, which means that the measured ellipticity
$\bm{\epsilon}^{({\rm obs})}$
\footnote{We use the complex ellipticity
$\bm{\epsilon}=\epsilon e^{2i \psi}$, where $\epsilon = (1-r)/(1+r)$,
and $r=b/a$ is the ratio between minor and major axes.}
of a galaxy can be expressed as \citep{2001PhR...340..291B, 2006glsw.conf.....M},
\begin{equation}
\bm{\epsilon}^{({\rm obs})} = \bm{g} + \bm{\epsilon}^{({\rm I})} + \bm{\epsilon}^{({\rm rnd})},
\end{equation}
where $\bm{g}$ is the reduced shear, defined as $\bm{g} =
\bm{\gamma}/(1-\kappa)$, and $\bm{\epsilon}^{({\rm rnd})}$
denotes the noise part in galaxy shape measurements, which
is assumed to be uncorrelated with the other components.
In the weak lensing regime, $\kappa$ is
small and the $\bm{\gamma} \simeq \bm{g}$ assumption is often
made. The intrinsic shape of a galaxy is described as
$\bm{\epsilon}^{({\rm I})}$. Ideally, if the intrinsic ellipticities
of galaxies are isotropic, the lensing shear $\bm{g}$ can be derived
by averaging over a population of galaxies. However, it is not the
case for real data, because of the presence of correlated intrinsic
alignment of observed galaxies.
The observed 2-point shear correlation function consists of the
following contributions \citep{2015PhR...558....1T,
2016ApJ...824...77J, 2016MNRAS.456..207K},
\begin{equation} \label{eq:2pcf}
\langle \bm{\epsilon}^{({\rm obs})}_{i} \bm{\epsilon}^{({\rm obs})}_{j} \rangle =
\langle \bm{g}_{i} \bm{g}_{j} \rangle + \langle \bm{\epsilon}^{({\rm I})}_{i} \bm{g}_{j} \rangle + \langle \bm{\epsilon}^{({\rm I})}_{i} \bm{\epsilon}^{({\rm I})}_{j} \rangle,
\end{equation}
where we assume that the two observed galaxies are located
at the redshifts $z_{i}$ and $z_{j}$ (with $z_{i} \leqslant z_{j}$), respectively.
The first term, $\langle \bm{g}_{i} \bm{g}_{j} \rangle$, represents
the shear-shear correlation, GG, which is the weak lensing
signal we want to extract. The correlation $\langle
\bm{\epsilon}^{({\rm I})}_{i} \bm{g}_{j} \rangle$ , often named as
GI, is the cross term between gravitational shear and intrinsic
ellipticity. This correlation comes from the fact that the shape of
a distant galaxy `$j$' is lensed by the foreground gravitational
potential, in which galaxy `$i$' is intrinsically aligned with
the underlying tidal field \citep{2004PhRvD..70f3526H}. Since
nearby galaxies are affected by the same environment,
the intrinsic-intrinsic correlation
$\langle \bm{\epsilon}^{({\rm I})}_{i} \bm{\epsilon}^{({\rm I})}_{j}
\rangle$, often referred as II term, may be non-zero.
Both the II and GI correlations can contaminate our measurements of
cosmic shear, and are important to quantify, particularly in accurate
shear measurements expected from future large lensing surveys
\citep{2016MNRAS.456..207K}.
In order to model the II and GI parts in the measurements,
\citet{2004PhRvD..70f3526H}, \citet{2007NJPh....9..444B} and \citet{2011A&A...527A..26J}
developed a non-linear intrinsic alignment model based on the work of \citet{2001MNRAS.323..713C}.
In this model, the power spectra of the II and GI contributions are related
to the non-linear matter power spectrum as $P_{\rm II}(k, z) = f^{2}(z) P_{\delta}(k, z)$
and $P_{\rm GI}(k, z) = f(z) P_{\delta}(k, z)$, respectively. Here
the modification factor $f(z)$ is defined as
\begin{equation}
f(z) = -A_{\rm IA} C_{1} \rho_{\rm c} \frac{\Omega_{\rm m}}{D(z)}
\left( \frac{1+z}{1+z_{0}}\right)^{\eta}
\left( \frac{L}{L_{0}}\right)^{\beta},
\label{equ:IA}
\end{equation}
where $A_{\rm IA}$ is a free parameter;
$C_{1}=5\times 10^{-14} h^{-2}{\rm M}_{\sun}^{-1}{\rm Mpc}^{3}$;
$\rho_{c}$ is the critical density at the present; and $D(z)$ is the linear
growth factor (normalized to unity at $z=0$).
The free parameters $\eta$ and $\beta$ account for the dependence
on redshift and luminosity around the pivot redshift $z_{0}$ and luminosity $L_{0}$.
Following the discussion of \citet{2017MNRAS.465.2033J} based on the CFHTLenS
data, we fix $\eta=0$ and $\beta=0$ in our model fitting.
These formulas are used in section 5.3.2 to fit the measurements of GI and II terms
from our simulation. There we will see that the sign of $A_{\rm IA}$
actually depends on galaxy type.
\section{Numerical simulations}
In this section, we describe the $N$-body simulations
(\ref{simulation:Nbody}), the spherical ray-tracing technique
(\ref{simulation:rt}), and the comparison between the measured power spectra from our lensing
simulations and those from the non-linear model predictions.
\subsection{$N$-body Simulation} \label{simulation:Nbody}
\begin{table*}
\centering
\caption{The parameters of the two $N$-body simulations. Cosmological parameters are given as $\Omega_{{\rm m}}, \Omega_{\Lambda}$, $h$ and $\sigma_{8}$. $L_{{\rm box}}$ is the box-size, $m_{{\rm p}}$ is the particle mass, $l_{{\rm soft}}$ is the softening length.}
\label{tab:Nbody}
\begin{tabular}{cccccccc}
\hline
Simulation & $\Omega_{{\rm m}}$ & $\Omega_{\Lambda}$ & $h$ &$\sigma_{8}$ & $L_{{\rm box}}/h^{-1}{\rm Mpc}$ & $m_{{\rm p}}/(10^{10}h^{-1}{\rm M}_{\sun})$ & $l_{{\rm soft}}/h^{-1}{\rm kpc}$ \\
\hline
PS-I & 0.260 & 0.740 & 0.710 & 0.80 & 1000 & 0.249 & 7.0 \\
L500 & 0.282 & 0.718 & 0.697 & 0.82 & 500 & 0.034 & 3.5 \\
\hline
\end{tabular}
\end{table*}
We use two sets of different $N$-body cosmological simulations.
The first one is a part of the ELUCID project
\citep{2014ApJ...794...94W, 2016ApJ...831..164W,
2016RAA....16..130L, 2017arXiv170403675T}, which is run with
$3072^{3}$ dark matter particles in a cubic box with $L_{{\rm box}} =
500 h^{-1} {\rm Mpc}$ on each side. This simulation is referred as
L500 in the following. The cosmological parameters of L500 are from the WMAP9
cosmology \citep{2013ApJS..208...19H}. The second simulation is the
Pangu simulation (PS-I), performed by the Computational Cosmology
Consortium of China \citep{2012ApJ...761..151L}, which has the same
number of particles as L500, but with a box size of $1000 h^{-1}{\rm Mpc}$ on each side. The
cosmological parameters of PS-I are from the WMAP7
\citep{2011ApJS..192...18K}. In Table.~\ref{tab:Nbody}, we list the
parameters of the two $N$-body simulations. Both simulations
were run using the GADGET-2 code \citep{2005MNRAS.364.1105S}.
With its higher mass resolution, the L500 simulation is used to
generate galaxies from a semi-analytical model
\citep{2016MNRAS.458..366L}. This model is based on the L-Galaxies
model developed by the Munich group
\citep[e.g.,][]{2013MNRAS.428.1351G} (see Section 4.1 for more details).
We do not produce mock galaxies using the PS-I simulation due to its lower mass resolution, but use it
as a reference to check our calculation of the convergence power spectrum on large scales.
\subsection{Spherical Ray-tracing Simulation} \label{simulation:rt}
To perform a ray-tracing simulation with full-sky coverage,
we follow the multi-plane algorithm
developed by \citet{2008ApJ...682....1D}, \citet{2009A&A...497..335T},
and \citet{2013MNRAS.435..115B}. In order to control the
residual in the solution of the lensing potential, we implement
an iterative spherical harmonic analysis scheme, which is different from the
multi-grid method adopted by \citet{2013MNRAS.435..115B}.
In the following, we briefly summarize the main procedures;
more details can be found in Appendix~\ref{appendix:rt}.
To trace the trajectory of a light beam, we first employ
the $N$-body simulations to build light-cone to redshift
$z_{{\rm max}} \sim 2.0$. In practice the simulation boxes are
divided into sets of small cubic boxes with $\sim 100 h^{-1}{\rm Mpc}$ on
each side. These cell boxes are appropriately piled together so as
to cover the past light-cone from $z=0$ to $z=z_{{\rm max}}$.
For both L500 and PS-I described above, a full-sky
light-cone can be constructed in this manner.
Note that the size of our simulation box is relatively small
compared with the comoving distance to redshift $z_{\rm s} = 2.0$
and periodic effects will show up at several specific directions,
especially along the box axes, but the effects disappear very quickly apart from these directions.
Then each light-cone is divided into a set of spherical shells with a
thickness of $50 h^{-1} \rm Mpc$ centered at the observer, and the dark matter distribution in
the corresponding shells is projected into pixels defined by
the HEALPix \footnote{\url{healpix.jpl.nasa.gov}} tessellation
\citep{2005ApJ...622..759G, 2007MNRAS.381..865C}. The
HEALPix resolution parameter is set to $N_{{\rm side}} = 8192$, which
gives an angular resolution of $\sim 0.43$ arcmin.\footnote{Given the
HEALPix resolution $N_{{\rm side}}$, one can calculate the
pixel scale by ${\rm d}\theta = \sqrt{4 \pi / (12 \times
N^2_{{\rm side}})}$} The projected surface mass densities are
calculated for each shell using the SPH algorithm
\citep{2005ApJ...635..795L, 2010ARA&A..48..391S}. We use the nearest
64 particles to define the kernel size, but keep the smoothing
length larger than two HEALPix cells in high-density regions. The
lensing potential for the $n$-th shell, $\phi^{(n)}_{\ell m}$, is then
obtained, using the Poisson equation, from the mass density shell
after applying an iterative spherical harmonic transformations (refer
to as HEALPix predefined functions),
\begin{equation}
-\ell(\ell+1)\phi^{(n)}_{\ell m} = 2\kappa^{(n)}_{\ell m}.
\end{equation}
To perform multi-sphere ray-tracing simulations, we set
the initial positions of ray-beams at the centers of the HEALPix
cells, and propagate light rays from the observer to a desired
redshift applying deflection angle,
\begin{equation}\label{equ:alpha1}
\alpha^{(n)}_{\ell m} = -\sqrt{\ell(\ell+1)}\phi^{(n)}_{\ell m}.
\end{equation}
\begin{figure}
\centering
\includegraphics[width=0.98\columnwidth]{./fig1.pdf}
\caption{One realization of the convergence map from PS-I light-cone for sources at $z_{\rm s} = 1.0$.}
\label{fig:kmap}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.98\columnwidth]{./fig2_C4.pdf}
\quad
\includegraphics[width=0.98\columnwidth]{./fig2_L500.pdf}
\caption{Comparison between the power spectrum from ray tracing
with the model predictions. The top panels show the angular power
spectrum of convergence (red solid line), shear E-mode (blue) and B-mode (magenta),
and rotation (cyan) for sources at $z_{s} = 1$ from the PS-I (left panel)
and L500 (right panel). The measured convergence power
from the Born approximation (grey dashed line) and revised Halofit model
\citep{2012ApJ...761..152T} predictions (black dashed line) are also shown
for comparison. The relative deviations of the measured convergence
statistics and from predictions are presented in the middle panels.
Lower panels show the ratios of the E- and B-modes with to
the convergence and rotation modes, respectively.}
\label{fig:PS1}
\end{figure*}
From our lensing simulation, we evaluate the
distortion matrix $\mathcal{A}$ on each lensing shell and construct
the full-sky map of the convergence and lensing shear. As an illustration,
Fig.~\ref{fig:kmap} shows one realization of our simulated full-sky
convergence map, $\kappa$, for sources at redshift $z_{s} = 1.0$.
In Fig.~\ref{fig:PS1}, we show the power spectra measured from
PS-I (left panels) and L500 simulation (right panels). The top panels show the angular power
spectra of the convergence (red solid line), the shear E-mode (blue) and B-mode (magenta), and the rotation mode (cyan).
We also show the prediction from the Born approximation
\citep{2002ApJ...574...19C} by stacking density field along the line-of-sight
in our mock light-cone as the grey dashed line
and the theoretical prediction from the revised non-linear Halofit
\citep{2012ApJ...761..152T, 2014ascl.soft02032P} as the black dashed line.
The middle panels of Fig.~\ref{fig:PS1} show the
relative deviations between convergence powers measured from our
ray-tracing simulation and the theoretical predictions, like that
the relative deviation between the ray-tracing simulation and Halofit model is defined by
$\Delta_{\rm Halofit} = \left[ C^{\kappa\kappa}(\ell) - C^{\rm Halofit}(\ell) \right]/ C^{\rm Halofit}(\ell)$.
The measured convergence power from our ray-tracing simulation
agrees well with the theoretical prediction of Halofit, and the
relative error is less than 10 per cent at $\ell \lesssim 4000$.
At large scales, $\ell \lesssim 10$, the
power from the PS-I simulation is in better agreement with theoretical
predictions than the L500 simulation, as is expected from the fact that
PS-I has a larger box to represent the matter power on large scales.
The prediction from Born approximation is closer to
that of Halofit at small scales than our ray-tracing
simulation, because both Born approximation and Halofit
are based on the first-order approximations. This good
agreement, within 10 per cent for $\ell \lesssim 6000$, roughly
on scales larger than the smoothing scale in our simulation, indicates
that the revised Halofit model \citep{2012ApJ...761..152T} provides a
good approximation to the non-linear matter power spectrum.
In addition, since the power spectra from our Born
approximation and full-sky ray-tracing simulation are both
based on the same convergence $\kappa$ maps, the difference between
the two is not caused by the smoothing effect.
The grey lines in the middle panels of Fig.~\ref{fig:PS1} show that
Born approximation can cause a deviation of more than 10 per cent
for $\ell > 6000$.
The bottom panels show the ratio of shear E-mode and B-mode power spectra
relative to the measured convergence and rotation mode spectra, respectively.
As shown in Section~\ref{sec:introPS}, at high-$\ell$ (small
scales), one has $C^{EE}(\ell) = C^{\kappa\kappa}(\ell)$ from
the full-sky weak lensing. Our corresponding measurements on small scales
are indeed in consistent with this expectation.
At large scales (low-$\ell$), we must take account
of the extra factor of $(\ell-1)(\ell+2)/\ell/(\ell+1)$ to explain the
difference between the power spectrum of the shear E-mode
and convergence spectrum.
As discussed in \citet{2013MNRAS.435..115B}, we
also measure the power spectra of the B-mode and the rotation mode
from our lensing simulation. We find that the B-mode power
is effectively suppressed relative to the E-mode by more than four
orders of magnitude. Moreover, the power ratio between the B-mode and the
rotation mode shows that the extra numerical B-mode in our simulation
is negligible, and the accuracy of our shear map is only limited by the smoothing length at small scales.
We refer the reader to Fig.~\ref{fig:dcl2} in Appendix A for more
details. Compared with the \citet{2013MNRAS.435..115B} results, the predictions
of our simulations for the convergence and shear power spectra are more
accurate extending to higher $\ell$.
\section{GALAXY MODELING}\label{sec:GM}
One important merit of our work is to include model galaxies in
the N-body simulations and to predict shear correlation functions
that can be compared directly to observations. In this section, we
describe how we model the physical properties of galaxies and show
the intrinsic alignment of the model galaxies.
\subsection{Semi-analytic Models}
\begin{figure}
\centering
\includegraphics[width=0.98\columnwidth]{./fig3.pdf}
\caption{Cumulative probability distributions of the
bulge-to-total mass ratio of model galaxies at three different redshifts.}
\label{fig:btCDF}
\end{figure}
The model galaxies are produced using the semi-analytical model of
\citet{2016MNRAS.458..366L} which is based on the \citet{2013MNRAS.428.1351G} model,
one version of the Munich Semi-analytical model which is called L-Galaxies.
In a semi-analytical model, galaxy population is assigned to dark matter haloes
on the basis of simple assumptions of many physical processes. As a first step in our implementation
of L-Galaxies, dark matter haloes are identified in the $N$-body simulation using the standard Friends-of-Friends
(FOF) algorithm. Only haloes that contain at least 20 particles are used.
The subhaloes within each FOF halo are identified with the SUBFIND
algorithm \citep{2001MNRAS.328..726S, 2005Natur.435..629S}.
Merger trees of these dark matter (sub)haloes can be constructed by linking
progenitors of a halo in different snapshots.
Galaxies are assumed to form at the centers of the dark matter haloes
according to analytical prescriptions of the relevant physical processes,
such as gas cooling, star formation, supernova and black hole
feedback. For the details of L-Galaxies, we refer the reader to \citet{2013MNRAS.428.1351G}.
\citet{2016MNRAS.458..366L} improved the prescription for low-mass galaxies,
especially satellite galaxies, by including additional physics about cold gas
stripping and an analytical modeling of orphan galaxies. In this model, the stellar mass
function, HI and H$_{2}$ mass functions are tuned to match the observational data
\citep{2003ApJ...582..659K, 2005MNRAS.359L..30Z, 2008MNRAS.388..945B, 2009MNRAS.398.2177L}.
The fraction and spatial distributions of central versus satellite galaxies
are reproduced roughly correctly by the model, as shown in \citet{2016MNRAS.458..366L}
and \citet{2013MNRAS.428.1351G}.
To describe the morphology of a galaxy, we use the ratio between
the bulge and the total mass ($B/T$), which can be predicted from
the semi-analytical model,
as the classification of `early-type' and `late-type' galaxies \citep{2009MNRAS.396.1972P}.
Following \citet{2013MNRAS.431..477J}, we adopt $B/T = 0.6$ to classify
the model galaxies into early or late types. Fig.~\ref{fig:btCDF}
shows the cumulative probability distribution of
$B/T$ of our simulated galaxies at three different redshifts.
More than $80\%$ of all galaxies are late-types, and
the fraction is slightly higher at high redshifts. This
fraction of early/late-types is consistent with that found in other studies
\citep[e.g.,][]{2013MNRAS.428.1351G}. However, compared
to hydro-dynamical simulations, semi-analytical models are less
powerful in predicting the shapes of galaxies. To proceed,
we have to assign shapes to galaxies and their images with
some simplified prescriptions, as described below.
\subsection{Galaxy Shape Measurement}\label{sec:galaxyshape}
A common assumption is that the shape of an elliptical galaxy roughly
follows that of its host dark matter halo, while the rotation axis of a spiral galaxy
is determined by the spin of its halo \citep[e.g.,][]{2007MNRAS.378.1531K, 2009ApJ...694..214O,
2010ApJ...709.1321A}. \citet[][hereafter
'J13']{2013MNRAS.436..819J} used this assumption and studied
the alignment of galaxies from the Millennium Simulation
\citep{2005Natur.435..629S}. Here we follow J13 to assign
shapes to model galaxies.
To assign a shape to a model galaxy using the mass distribution of its dark
matter halo, we need to distinguish between central and satellite
galaxies. A central galaxy is assumed to be located at the center of a
dark matter halo and its shape may be related to that of the host halo.
The shape of a dark matter halo is usually defined using the inertia
tensor $I_{ij}$ \citep{2005ApJ...627..647B},
\begin{equation}
I_{ij}= \sum^{N_{\rm p}}_{n=1}{m_{p} x_{i,n}, x_{j,n}},
\end{equation}
where $N_{\rm p}$ denotes the particle number of the FOF halo, and
${\bf x}_{n}$ is the position of the $n$-th particle with respect to the
center of the halo. By diagonalizing the inertia tensor $\bm{I}$, one
can get the eigenvalues $\lambda_{1} \leqslant \lambda_{2} \leqslant
\lambda_{3}$ and the corresponding eigenvectors that define
a triaxial ellipsoid and its orientation. It has been argued that
ta minimum number of $N_{\rm p}=300$ is needed to ensure an
accurate measurement of the halo shape
\citep{2002MNRAS.335L..89J, 2007MNRAS.376..215B}.
Constrained by the resolution of our simulation, we have to reduce
the number limit to $100$. As shown in J13, a minimum number of $100$
can lead to $\sim10\%$ deviation in the axis ratio and $10^{\circ}$
deviation in the orientation angle. Since the magnitude limitation, the
corresponding haloes always have more than 100 particles in our mock catalogs
for DLS and KiDS-450.
Once the three-dimensional shapes and orientations of galaxies are obtained,
we project them into the sky to obtain the projected ellipses
\citep{1983Ap&SS..92..335G, 1985MNRAS.212..767B}.
Details about how to make the projections can be found in
appendix~\ref{appendix:proj}.
In addition to the above model in which perfect alignments are assumed
between elliptical galaxies and their host halos, J13 also considered a mis-alignment
model in which the major axis of the central elliptical galaxy is mis-aligned
with that of the halo, with the mis-alignment angle obeying a gaussian
distribution with zero mean and a dispersion of $35^{\circ}$. This is motived
by the finding that such a mis-alignment is needed to explain the alignment
between luminous red galaxies on large scales
\citep[e.g.,][]{ 2007MNRAS.378.1531K,
2009RAA.....9...41F, 2009ApJ...694..214O, 2013ApJ...768...20L}.
We will come back in Section~\ref{sec:GIIIterm} to discuss the effect of such mis-alignment
on shear correlation functions.
For a central late-type galaxy, defined by $B/T < 0.6$, J13 assigned its
shape according to the angular momentum vector of the host halo,
\begin{equation}
\bm{L} = \sum^{N_{\rm p}}_{n=1}{m_{p} \bm{x}_{n} \times \bm{v}_{n}},
\end{equation}
where $\bm{v}_{n}$ is the velocity of the $n$-th halo particle relative
to the halo center. The angular momentum $\bm{L}$ defines a circular disc in
the halo, and the complex ellipticity is obtained by
projecting the disc along the line of sight. As described in J13, the apparent
axis ratio of the projected ellipse is,
\begin{equation}
r = \frac{|L_{{\rm los}}|}{|\bm{L}|} + r_{d}\sqrt{1-\frac{L_{{\rm los}}^2}{|\bm{L}|^2}},
\end{equation}
where $r_{d}$ is the ratio between the disc thickness and diameter,
and we set $r_{d}=0.25$ following J13; $L_{{\rm los}}$ is
the component of $\bm{L}$ along the line of sight. The
ellipticity of the mocked galaxy is then given by
\begin{equation}
\epsilon = (1-r)/(1+r).
\end{equation}
For a satellite galaxy, on the other hand, the original dark matter halo
associated with it may have suffered strong mass loss after it is accreted
into a big halo, depending on the infall time and orbit \citep{2000MNRAS.319..168C}.
It is thus unclear how the shapes of satellites are connected to those
of the dark matter subhaloes associated with them. Some earlier
investigations \citep[e.g.,][]{2008ApJ...672..825P} have shown that the tidal torque
of the host halo can induce a correlation between subhalo orientation
and its direction to the host center.
Therefore, J13 assigned the shape of an early-type satellite galaxy by randomly
choosing the three-dimensional axis ratios from a halo sample with more than 300
particles and then made its major axis point to the central galaxy.
For a late-type satellite, its spin is assumed to
be perpendicular to the line connecting the satellite to the central galaxy, and
the ellipticity is obtained by projecting the disk onto the sky, as done for the
central spiral galaxies.
The above assumptions for the shapes and orientations of
satellite galaxies are clearly too idealistic. In fact, while
orbiting in their host halos, satellite galaxies may have
their radial alignments scrambled. Observationally,
measurements of the shapes of faint satellite galaxies are
difficult and sensitive to the methods used to derive galaxy shapes
\citep[e.g.,][]{2011ApJ...740...39H}. Currently, there is no consensus
on the alignments of satellites. Some studies have reported detection
of radial alignment of satellites
\citep{2005ApJ...627L..21P, 2006ApJ...644L..25A, 2007ApJ...662L..71F,
2014MNRAS.445..726C, 2015MNRAS.450.2195S, 2016MNRAS.463..222H},
while others have not found such an alignment
\citep{2009arXiv0903.2264S, 2011ApJ...740...39H, 2013MNRAS.433.2727S,
2015A&A...575A..48S}. Because of this uncertainty, we also consider
a simple case in which the orientations of satellites, regardless of
their types, are random distributed in their host haloes. We will show
in Section 5.3 that satellite alignment has a stronger effect on the
shear correlation on smaller scales.
\subsection{Intrinsic Shape Correlations of Mock Galaxies}
\begin{figure}
\centering
\includegraphics[width=0.98\columnwidth]{./fig4.pdf}
\caption{The intrinsic ellipticity correlation, $\eta(r)$,
for early-type (upper panel) and late-type (lower panel) central
galaxies. Only central galaxies within haloes of mass
$m_{{\rm halo}} \geq 3.4\times10^{10} {\rm M}_{\sun}/h$ (i.e.
with 100 particles or more) are used.
Early-type galaxies show
significant shape alignments and redshift dependence, while for
late-types the correlations of their intrinsic ellipticities
are very weak. Here errors are estimated using the jackknife method.}
\label{fig:eta0}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.98\columnwidth]{./fig5.pdf}
\caption{Correlation function $\eta(r)$ of early-type galaxies at $z = 1.0$.
The top panel shows the halo mass dependence of the correlation for galaxies
in a given luminosity bin.
The bottom panel shows the correlation of galaxies in a given small range of
halo mass but with different luminosities. The figure shows that {the intrinsic ellipticity}
correlation is mainly determined by halo mass.}
\label{fig:eta1}
\end{figure}
The intrinsic shape correlation function of galaxies, $\eta(r)$,
is defined as
\begin{equation}\label{equ:eta}
\eta(r) = \langle \epsilon_{t}(\bm{x}) \epsilon_{t}(\bm{x}+\bm{r}) + \epsilon_{\times}(\bm{x}) \epsilon_{\times}(\bm{x}+\bm{r}) \rangle_{\bm{x}},
\end{equation}
where $r$ is the three-dimensional comoving separation between two galaxies
\citep[e.g.,][]{2006MNRAS.371..750H}. The quantities
$\epsilon_{t}$ and $\epsilon_{\times}$ are the tangential and cross components of
the galaxy ellipticity:
\begin{equation}
\epsilon_{t}+{\rm i}\epsilon_{\times} = -\bm{\epsilon} {\rm e}^{-2{\rm i}\varphi}
\end{equation}
where $\varphi$ is the angle between the separation vector
of a given galaxy pair and the horizontal axis \citep{2001PhR...340..291B}.
Following \citet{2013MNRAS.436..819J}, we first measure the
intrinsic shape correlation of galaxies in our simulation by projecting
semi-analytic galaxies along the line of sight parallel to the edges of
the simulation box. In order to estimate the error bars in our measurements,
we divide our simulation box into eight equal-sized cubic sub-boxes of
$250\,{\rm Mpc}/h$, and use the Jackknife method to estimate the errors.
Fig.~\ref{fig:eta0} shows the redshift dependence of the
correlation function $\eta(r)$ for early-type (upper) and late-type
(lower) central galaxies in our simulation. It is seen that early-type
galaxies have a strong correlation and the correlation is stronger at
higher redshifts. In contrast, late-type galaxies do not show any
significant correlation of their projected ellipticities, although
some weak positive correlation signals can be seen at small scales,
$r \lesssim 8 {\rm Mpc}/h$. The weak/null correlation for spiral galaxies in
our simulation is consistent with the non-detection in both
observations \citep[e.g.,][]{2011MNRAS.410..844M} and in simulation results
\citep[e.g.,][]{2013MNRAS.436..819J}.
In Fig.~\ref{fig:eta1} we further investigate the dependence of the
shear correlation $\eta(r)$ on halo mass and luminosity for
early-type galaxies at $z = 1.0$. To separate the two dependencies,
we select galaxies in a small
ranges of luminosity and halo mass, and divide galaxies into two
subsamples in halo mass (the top panel) and in galaxy luminosity
(the bottom panel ). It can be seen that there is no significant
luminosity dependence for fixed halo mass, but a significant
dependence on halo mass is seen at a fixed luminosity.
This dependence of {the intrinsic ellipticity} correlation on mass and
luminosity in our simulation is similar to the results found in J13.
\section{COSMIC SHEAR AND COMPARISON WITH OBSERVATION}
\subsection{Shear Correlation Function}\label{sec:xipm}
Weak lensing induces small correlated distortions in observed
galaxy shapes. This correlation can be quantified using different
statistics. Observationally, the most direct measurement of the
lensing signal is the two-point shear correlation function.
The shear-shear correlation between galaxies at a given
separation $\vartheta$ is estimated as
\begin{equation}
\xi_{tt}(\vartheta) = \frac{\sum_{i,j} w_{i}w_{j}\epsilon_{t,i}\epsilon_{t,j}}{\sum_{i,j}w_{i}w_{j}}
\end{equation}
and
\begin{equation}
\xi_{\times\times}(\vartheta) = \frac{\sum_{i,j} w_{i}w_{j}\epsilon_{\times,i}\epsilon_{\times,j}}{\sum_{i,j}w_{i}w_{j}},
\end{equation}
where $w_{i}$ is the ellipticity weight of the $i$-th galaxy and $\vartheta$ is
the angular separation between the galaxy pair.
The two linear combinations of $\xi_{tt}$ and $\xi_{\times\times}$
that are also frequently used in the lensing analysis are
\begin{equation}\label{equ:xipm}
\xi_{\pm} = \xi_{tt} \pm \xi_{\times\times}.
\end{equation}
The convergence power spectrum $C^{\kappa\kappa}(\ell)$ can be related to
the estimator $\xi_{\pm}$ as
\begin{equation}
\xi_{\pm}(\vartheta) = \frac{1}{2\pi}\int^{\infty}_{0}{\rm d}\ell \ell J_{0,4}(\ell\vartheta) C^{\kappa\kappa}(\ell),
\end{equation}
where $J_{0,4}(\ell\vartheta)$ denotes the zeroth and fourth Bessel function for $\xi_{+}$ and $\xi_{-}$,
respectively \citep{2002A&A...396....1S}.
\subsection{Tomographic Cosmic Shear}
Tomographic measurement is capable of utilizing the
redshift-dependence of cosmic shear signals to reveal both
the cosmological structure growth and the redshift-dependent geometry in the
universe \citep{2003A&A...398...23K}, and
has been widely used in weak lensing observations, such as
CFHTLens \citep{2013MNRAS.432.2433H}, DLS \citep{2016ApJ...824...77J},
and KiDS-450 \citep{2017MNRAS.465.1454H}. In order to
perform similar tomographic analysis of cosmic shear in our mock
observation and compare the prediction with the KiDS and DLS observations,
we employ the tomographic redshift bins as used by
\citet{2017MNRAS.465.1454H} for KiDS-450 and
\citet{2016ApJ...824...77J} for DLS, to mimic their measurements of cosmic
shears.
Fig.~\ref{fig:KIDSandDLSphotoz} shows the redshift distributions
of the source galaxies in the two surveys,
where $z_{\rm B}$ denotes the Bayesian point estimates of
the photo-$z$ \citep{2016ApJ...824...77J, 2017MNRAS.465.1454H}.
It is seen that the two distributions are quite different.
In KiDS-450 the shape of the distribution is not regular, with more overlaps
between different redshift bins, while the distribution for DLS is more regular and
different bins are more clearly separated.
For a consistent comparison between model predictions and observations,
we adopt their redshift distributions for the source galaxies respectively,
and we truncate source galaxies at $z=2.0$.
To compare with the survey results, we first divide the
full-sky into a set of small patches with sizes $\sim
3.6^{\circ}\times3.6^{\circ}$, and then randomly select $35$ patches
in total to cover a field of $\sim 450$ square degrees to match the sky
coverage of KiDS-450. By setting a limiting magnitude of
$\sim 24.5^{{\rm th}}$ mag in the $r$-band, the effective number density
in our light-cone is $n \sim 8\ {\rm arcmin}^{-2}$, similar to
that in the KiDS-450 observations. The DLS is much deeper, with a
magnitude $\sim 27^{\rm th}$ in $r$-band, producing an effective
number density $\sim 11\ {\rm arcmin}^{-2}$ of the source population
in 5 tomographic bins. Given that our lensing simulation is
performed to redshift $z_{\rm max} \sim 2.0$, we discard the 5th redshift bin,
using only tomographic bins 1-4 of DLS, which gives
$n \sim 8\ {\rm arcmin}^{-2}$. Thus, a set of mock galaxy catalogue
is constructed to mimic the sky coverage and galaxy number
density for each of KiDS-450 and DLS. For each mock we produce 100
realizations by randomly sampling the patches at different positions
to estimate the uncertainties of tomographic shear
correlations due to the cosmic variance and sampling noise.
\begin{figure}
\centering
\includegraphics[width=0.98\columnwidth]{./fig6.pdf}
\caption{Redshift distributions of source galaxies
in the two weak lensing surveys, DLS (left) and KiDS-450 (right).
These redshift distributions are used in our light-cone to select
source galaxies and used for the tomographic analyses.}
\label{fig:KIDSandDLSphotoz}
\end{figure}
\subsection{Results}
We measure the
auto-correlation and cross-correlation functions $\xi^{(ij)}_{\pm}$ using
the public code {\tt
Athena} \footnote{\url{http://www.cosmostat.org/software/athena}},
which estimates the second-order shear correlation functions from
equation~(\ref{equ:xipm}). The superscript $(ij)$ denotes different
redshift bins used for the calculation of the
correlation function. In our case there are four redshift bins labeled
from $1$ to $4$ with increasing redshift.
\subsubsection{Model predictions and comparison with observations}
\label{sec:modelcomparison}
\begin{figure*}[t]
\centering
\includegraphics[width=1.98\columnwidth]{./fig7.pdf}
\caption{Tomographic measurements of $\xi_{+}$ from our mock KiDS-450
catalogue with galaxy number density $n = 8.0\ {\rm arcmin}^{-2}$ and
total sky coverage of $\sim 450$ square degrees. Red circles represent the shear
correlations from the gravitational field (GG), and black circles are the
total shear correlations (GG+II+GI), which can be directly compared with
the observational results (blue circles) from the KiDS-450 \citep{2017MNRAS.465.1454H}.}
\label{fig7}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=1.98\columnwidth]{./fig7_1.pdf}
\caption{Similar to figure~\ref{fig7}, but for the tomographic
measurements of $\xi_{-}$.}
\label{fig8}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=1.98\columnwidth]{./fig8.pdf}
\caption{Similar to Fig.~\ref{fig7}, but here
satellite galaxies are assumed to be radially aligned
with the central galaxy while in Fig.~\ref{fig7} satellites are
assumed to have random orientations. See the text for
details of how we assign galaxy shape and orientation.}
\label{fig9}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1.98\columnwidth]{./fig9.pdf}
\caption{Tomographic shear correlations for the DLS mock catalogue with galaxy
density $n=8.0\ {\rm arcmin}^{-2}$ and total sky coverage $\sim 20$ square degrees.
It is seen that the data points are higher than our model predictions.}
\label{fig10}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1.98\columnwidth]{./fig10.pdf}
\caption{Top panels show the contributions of
different components to the shear correlations in
a few redshift bins, as obtained
from our mocked KiDS-450. The lower
panels show the ratios of II and GI to the real
gravitational shear correlation (GG). In the
top panels, solid lines show the results obtained from
the best fit to the non-linear alignment model (Eq.~\ref{equ:IA}).
Dashed lines show the contributions of GG, II and GI
in the model fitting.}
\label{fig11}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=1.98\columnwidth]{./fig11.pdf}
\caption{The ratios of the II and GI terms with the gravitational
shear correlation (GG) in two ideal cases.
Top panel: all central galaxies are assumed
to be spirals, and their spins are assumed to follow that of the
dark matter haloes. Bottom panel: all central galaxies
are assumed to be ellipticals, and their shapes
are assumed to follow their host haloes but with a
mis-alignment that is given by a
Gaussian distribution with a dispersion of $35^\circ$.
It is clearly seen that the GI contribution is positive in the
first case, but negative in the second.}
\label{fig12}
\end{figure*}
In Figs.~\ref{fig7} and \ref{fig8} we show the tomographic shear
correlations $\xi_{\pm}$ from our model and compare them with the
KiDS-450 results. We note that here the
orientation of the central galaxy is assumed to follow that of the dark matter halo.
Namely for an elliptical central, its major axis follows that of the halo,
while for spiral central its orientation is determined by halo spin. For
satellite galaxies, regardless of ellipticals or spirals, their
orientations are randomly distributed on the sky. The black circles are the
model predictions for the ellipticity correlation, which can be directly compared with
the data (blue circles). As mentioned in Section 2.3, the predicted
ellipticity correlations are combinations of the GG, GI and II
correlations. For simplicity, here we only show the GG terms as the red
circles connected by the red lines. We will show the contributions of
the II and GI terms in Section 5.3.2.
Figs.~\ref{fig7} and \ref{fig8} show that in general
the model predictions (black circles) agree well with the KiDS-450
results. To quantify the difference between the model and the
data, we calculate the reduced $\chi^{2}$ defined by
\begin{equation} \label{eq:eq29}
\chi^{2}_{\pm} = \frac{1}{n}\sum\frac{(\xi^{\rm M}_{\pm}-\xi^{\rm D}_{\pm})^{2}}{(\sigma^{\rm D}_{\pm})^{2}}.
\end{equation}
Here $n$ is the number of data points in the tomographic measurements;
$\xi^{\rm M}_{\pm}$ and $\xi^{\rm D}_{\pm}$ represent the
predicted and observed tomographic correlations, respectively;
$\sigma^{\rm D}_{\pm}$ is the error in the data.
The error bars we predicted only contain intrinsic ellipticity
dispersion of galaxy, cosmic variance and shot noise, and they represent
the dispersion between the results of our 100 realizations, i.e. the
uncertainties of one realization. The uncertainties of $\xi_{\pm}^{\rm M}$
which we predicted (the mean of 100 realization) are very low, so we do
not take them into account in Eq.~\ref{eq:eq29}.
We then find the reduced $\chi^{2}_{+(-)} = 1.70\ (1.82)$ between
our model prediction and the KiDS-450 results.
If the correlation between $\xi_{+}$ and $\xi_{-}$ is taken into account,
we should estimate the reduced $\chi^{2}$ from the correlation matrix,
\begin{equation} \label{eq:eq30}
\chi^{2} = \frac{1}{n}\sum_{i,j}\Delta \xi_{i}C_{ij}^{-1} \Delta \xi_{j},
\end{equation}
where $C_{ij}$ is the covariance matrix of the data \citep{2017MNRAS.465.1454H};
$\Delta \xi_{i}$ is the difference between the model prediction and the data
in the $i$-th separation bin. This gives a reduced $\chi^{2}=1.36$ for the full
data vector of KiDS-450, which is slightly higher than the reduced $\chi^{2}=1.33$
in the fiducial analysis of KiDS-450 \citep{2017MNRAS.465.1454H}.
If we calculate the reduced $\chi^{2}$ for $\xi_{+}$
and $\xi_{-}$ separately, we find $\chi^{2} = 1.23$ and $1.61$, respectively.
The good agreement between our model and the KiDS-450 data is
encouraging, as this is the first time to reproduce the observational
results using lensed images of mock galaxies in N-body simulations
combined with a realistic model of galaxy formation. However,
inspecting Fig.~\ref{fig7} carefully, one can see that for some bins,
such as 12, 24 and 34, the model predictions are slightly higher than the data on large
scales. Note that the error bars are also larger in these bins and
the data points are not well described by the best fitting model of
\citet{2017MNRAS.465.1454H}. To further quantify the systematic deviation between
our model prediction and the data, we define the weighted mean deviation as,
\begin{equation}\label{eq:eq31}
\Delta_{\rm m} = \frac{1}{n}\sum\frac{\Delta \xi_{i}}{\sigma_{i}}.
\end{equation}
Here only errors from the data are used as the weight $\sigma_{i}$,
{and the correlation between the measurements is not accounted for.}
{Note that the correlation between the measurements will sightly reduce
the deviation, similar to previous analyses for the reduced $\chi^{2}$.}
The systematic bias between the simulations and the observed data
can then be estimated in terms of the standard deviation as
\begin{equation}\label{eq:eq32}
S = \frac{\Delta_{\rm m}}{\sigma_{\rm m}},
\end{equation}
where $\sigma_{\rm m} = 1/\sqrt{n}$ is the scatter of $\Delta_{\rm m}$.
In this way, we find that the systematic bias between our model
prediction and KiDS-450 result is $S = 1.80$ and $1.92$ for $\xi_{+}$
and $\xi_{-}$, respectively.
This positive deviation suggests that our predicted correlations
are slightly, but systematically higher than the KiDS-450 data with a significance of $\sim 1.8 \sigma$.
As the error bars are often correlated between different redshift
bins, our analyses of these derivations might be too simplistic.
However, we do not intend to quantify the difference between the model
and the data in detail, but would like to point out that such a difference
could be due to the cosmological parameter,
$\sigma_{8}\sqrt{\Omega_{\rm m}/0.3}=0.79$, adopted in our simulation,
which is slightly larger than that derived one, $0.745 \pm 0.039$, from the KiDS-450 data
\citep{2017MNRAS.465.1454H}.
The results shown in Figs.~\ref{fig7} and \ref{fig8} assume
that satellites have random orientations. In
\citet{2013MNRAS.436..819J} the orientations of
satellite galaxies are assumed to be radially aligned with
central galaxies. Thus, for an early-type satellite, its major axis is
assumed to point towards to the central galaxy, while for a late-type satellite,
its spin is assumed to be perpendicular to the line connecting the satellite to
the central galaxy. In Fig.~\ref{fig9} we compare the model results obtained
by assuming such radial alignments to the KiDS-450 results.
We can see that the shear correlations in diagonal panels at lower redshift
are much higher than the data on small scales, except for the highest redshift,
4-4 bin, where the model prediction is close to the data. This indicates
that the radial alignment model of satellites produces too strong
correlation on small scales. In addition, this model also leads to a
strong negative GI term, suppressing the measured total signal on
small scales in the cross-correlation of shears, as seen in the
off-diagonal bins (13, 14, 24).
This strong positive correlation at small scales in the
auto-correlation bins, and the strong negative correlation in
the cross bins can be explained by the balance between the contributions
of the II term and the GI term. With the J13 assumption for satellites,
the II term is positive and GI term is negative. In the auto-correlation bins,
the II term is stronger than the GI term, so as to make a strong positive correlation
of total signal at small scales, as shown in Fig.~\ref{fig9}. On the other hand,
in the cross-correlation bins, the contribution of the II term is reduced
relative to the GI term, so that the correlation at the small scales shows
a strong negative correlation in the cross-correlation bins.
The reduced $\chi^{2}$ can be calculated from the covariance matrix of the
data, and we find $\chi^{2} = 3.73$ for the full data vector of KiDS.
Calculating the $\chi^{2}$ separately for $\xi_{+}$ and $\xi_{-}$,
we get $\chi^{2} = 4.63$ and $2.76$, respectively.
The results in Fig.~\ref{fig9}, therefore, suggest that the radial alignment
model for satellite galaxies can be rejected. In what follows, we will only show
model predictions in which satellites are assumed to have random orientations.
DLS is another weak lensing survey completed recently
\citep{2013ApJ...765...74J, 2016ApJ...824...77J,
2016AAS...22730707J}. Compared to KiDS-450, DLS has a smaller sky
coverage of 20 square degrees. We produce mock DLS catalogues following
its sky coverage, galaxy number density and redshift distribution
of source galaxies (see Fig.~\ref{fig:KIDSandDLSphotoz}).
Our model predictions and comparisons with DLS are shown in
Fig.~\ref{fig10}.
Considering that the data of $\xi_{-}$ is not available for DLS \citep{2016ApJ...824...77J},
here we only present the results of tomographic
correlations $\xi_{+}$. It is seen that the model results (black
circles) are lower than the DLS data (blue circles), especially in the
lower redshift bins. Note that the error bars in the model
is slightly larger. For a DLS-like survey, the reduced
$\chi_{+}^{2} = 1.56$, and the systematic bias between
the simulation and the data is $S = -2.57$ for $\xi_{+}$.
The lower reduced $\chi_{+}^{2}$ seems to indicate that the
agreement between our model and DLS is slightly better than the
agreement with KiDS-450. However, it is clear
that the lower reduced $\chi_{+}^{2}$ is also related
to the fact the error-bars in DLS data are much larger
than those of the KiDS. The large error bars in the DLS data
are partly due to its small sky coverage and the smaller sample
of galaxies. In the tomographic analysis the KiDS-450 sample
is more than 10 times as large as DLS in terms of the total number
of galaxies. The strong negative systematic bias $S=-2.57$ indicate that
the DLS data are systematically higher than our model predictions.
Since the cosmic variance has been taken into account in the error bar of DLS,
such a large systematic deviation can hardly be explained by the cosmic variance only.
It is unclear what causes the discrepancy ($\sim 2\sigma$ in $S_{8}$) between the
observational results of DLS and KiDS-450. One potential cause
might be from the estimations of photo-$z$. For KiDS
\citep{2017MNRAS.465.1454H}, the DIR method is used to estimate the
redshift distribution of galaxies, while for DLS \citet{2016ApJ...824...77J}
the BPZ method is adopted. As briefly discussed in \citet{2017MNRAS.465.1454H},
the $\chi^{2}$ can increase by $\sim 10$ when switching from the DIR
redshift distribution to the BPZ
distribution. They also argued that the deeper DLS data is harder to be calibrated.
It is beyond our scope to discuss the discrepancy between the KiDS-450 and DLS results in detail.
We refer the reader to the paper cited above for more discussions.
\subsubsection{The contributions of II and GI terms}
\label{sec:GIIIterm}
Our previous model results in Fig.~\ref{fig7} show that
there is a difference around 10 per cent between the GG term and the total shear correlation.
The difference is due to a combination of II and GI terms.
In Fig.~\ref{fig11} we show the contributions from the two components
separately. As some data points are negative, we plot the $\theta\xi_{+}$ in
linear scales, and for clarity, we do not show the observational
data. The top panel shows our fiducial results for the KiDS-450 mock in
some tomographic redshift bins (the same as in
Fig.~\ref{fig7}), and the lower panel shows the ratios of GI and II
with the GG term. As one can see from the plot, the II term is very weak,
consistent with zero. This is expected because most galaxies in our model are late-types
and their intrinsic alignment is very weak
(Fig.~\ref{fig:eta0}). Moreover, the GI term is basically positive and
its contribution could be as large as 15\% on large scales. Following
the procedures usually adopted in observational work to determine the free
parameter $A_{\rm IA}$ \citep[e.g.,][]{2013MNRAS.432.2433H}, we fit
the total signals (black circles) from simulation using the non-linear intrinsic
alignment model (Eq.~\ref{equ:IA}).
Note that in our calculation, the GG signal is given using the non-linear
theoretical power spectrum with given cosmological parameters, and the red dashed
line in Fig.~\ref{fig11} shows that the theoretical prediction agrees well with
the measured GG term from our simulation. The best fit to total signal (GG+II+GI)
gives $A_{\rm IA} = -0.972 \pm 0.217$.
The fit to each component is also shown as the dashed line in the Fig.~\ref{fig11}.
{Compared with the result by \citet{2017MNRAS.465.1454H}, they constrained the
amplitude of the intrinsic alignment to a positive $A_{\rm IA}=1.10 \pm 0.64$ in their
fiducial analysis of KiDS-450, which gives a negative GI term in the measurements.}
While a negative $A_{\rm IA}$ here indicates the contribution from GI term is
positive and acts to increase the overall correlation signals.
In fact, a positive GI signal is not surprising and has been reported in
major weak lensing surveys. For example, \citet{2008A&A...479....9F} found that
$A_{\rm IA}=-2.2^{+4.6}_{-3.8}$ from the third-year CFHTLenS data, and
\citet{2013MNRAS.432.2433H} reported that $A_{\rm IA}=-1.18^{+0.96}_{-1.17}$
from the final CFHTLenS data. \citet{2017MNRAS.465.2033J} found that
$A_{\rm IA} = -3.6 \pm 1.6$ from the re-analysis of CFHTLenS data.
\citet{2017MNRAS.465.1454H} found a positive GI term with $A_{\rm IA}=-1.10_{-0.70}^{+0.96}$
for the KiDS-450 data if they used the BPZ method to estimate galaxy photometric redshift.
Recently, \citet{2017arXiv170801538T} found, from the DES data, that for spiral
galaxies, the GI term is also positive with $A_{\rm IA}=-0.8$ at a
84 per cent confidence level.
Since most galaxies in our model are late-types, we conclude that the
positive GI signal is contributed by spirals. As a further
test on the contributions from spiral and elliptical galaxies,
we show, in Fig.~\ref{fig12}, the ratios of the
II and GI terms with the GG term for two ideal cases. In the top panel,
we assume that all central galaxies in our model are spirals and
their spins follow the spins of their host dark matter halos.
In the bottom panel, we assume all central galaxies are
ellipticals and their shapes follow the shapes of the host halos
defined with the inertia tensor, but with a misalignment given
by a gaussian distribution with a dispersion of $35^{\circ}$. In the two cases all
satellite galaxies are assumed to have random orientations.
Fig.~\ref{fig12} shows that the GI term is indeed positive on all scales
for spiral galaxies, although their {II contribution} is close to zero.
For elliptical galaxies, their II term is positive and the GI term
is negative.
{Note here the error bars are different for the two different ratios.
This can be simply explained from Eq.~\ref{eq:2pcf} by considering the noise in the correlation of GI and II terms.
As show by the GG correlation, the gravitational shear can be accurately measured in the mock.
So if we consider the noise (${\bm N}$) in the measurement of intrinsic shapes of galaxies,
GI and II terms can be expressed as
$\left< {\bm \epsilon}_{i}^{\rm (I)} {\bm g}_{j}\right> + \left< {\bm N}_{i} {\bm g}_{j}\right>$
and
$\left< {\bm \epsilon}_{i}^{\rm (I)} {\bm \epsilon}_{j}^{\rm (I)}\right> + 2 \left< {\bm \epsilon}_{i}^{\rm (I)} {\bm N}_{j}\right> + \left< {\bm N}_{i} {\bm N}_{j}\right>$, respectively.
The effect of shape noise can contribute the additional correlations as
$\left< {\bm N}_{i} {\bm g}_{j}\right>$ for GI term
and $2 \left< {\bm \epsilon}_{i}^{\rm (I)} {\bm N}_{j}\right> + \left< {\bm N}_{i} {\bm N}_{j}\right>$ for II term.
Combining with the definition of correlation $\xi_+$, this difference
can be used to explain the different error bars for the two different ratios.}
A positive GI term from spiral galaxies is not expected from the tidal
field model. From linear theory
\citep[e.g.,][]{2004PhRvD..70f3526H}, the GI term is found to be
negative, which has been used as a fiducial model in weak lensing data analyses
\citep[e.g.,][]{2011A&A...527A..26J, 2013MNRAS.432.2433H}. One
important assumption in the linear model
is that the shape of foreground galaxies are radially aligned with the
nearby over-dense region. This is on average true for elliptical galaxies.
But for spiral galaxies, where alignments are mainly determined by
angular momenta of the dark matter halos through
large-scale tidal field \citep[see][and references therein]{2009IJMPD..18..173S},
this may not be true. Observationally, it is
found the spins of spiral galaxies tend to align with the nearby
filaments, but the short axes of ellipticals are perpendicular to
filaments \citep{2010MNRAS.408..897J, 2013ApJ...775L..42T}. Both
hydro-dynamical simulations \citep[e.g.,][]{2015MNRAS.448.3391C} and N-body
simulations \citep[e.g.,][]{2015ApJ...813....6K, 2017MNRAS.468L.123W} also confirmed
such a dependence on galaxy types. In particular,
\citet{2015MNRAS.454.2736C} have found from hydro-dynamical
simulations that spiral galaxies have a significant tendency to be
tangentially aligned with over-density regions. Their figure 10
demonstrates clearly the alignment of spiral galaxies around
over-density regions and the origin of a positive GI term.
Finally we note that the II and GI terms are close related to how we
model galaxy shapes and orientations. In this paper we simply assume that
spiral galaxies follow the spins of dark matter halos. However, as
shown in, e.g., \citet[][]{2010MNRAS.404.1137B}, galaxy
spins have a broad distribution of mis-alignment with dark matter
haloes. This mis-alignment will reduce the positive GI
terms. Furthermore, the total GI and II terms in real data
depend also on the fraction of spiral and elliptical galaxies,
as well as on galaxy luminosities and redshift.
More comprehensive analyses on these factors are needed
to quantify their impacts on GI and II terms.
This paper, which makes use of both $N$-body simulation
and galaxies from a semi-analytical model, is a step towards this goal.
But here we only focus on a first comparison of the predicted
shear correlations with the data.
We will present a more comprehensive investigation on the contribution of GI and
II terms in a future paper.
\section{CONCLUSIONS}
It is well known that the intrinsic alignment of galaxy and its
associated correlation with the gravitational shear is one of the
dominant contaminations to the weak lensing survey. Numerous efforts have
been devoted to modeling galaxy-galaxy intrinsic alignment,
gravitational-galaxy intrinsic alignment and their impacts on the measure
cosmic shear correlation \citep[for a review, see][]{2015PhR...558....1T,
2015SSRv..193....1J, 2015SSRv..193..139K, 2015SSRv..193...67K}. One
useful and direct way to judge these alignment models and their impacts on the
measured galaxy shear correlations is to produce mocked galaxy images
using ray-tracing simulations which can be directly compared with the observational data.
In this work we make a first attempt to use a large cosmological
$N$-body simulation, ELUCID, and a semi-analytical model for galaxy
formation, to perform a full-sky ray tracing, so as to produce mocked
galaxies images and associated gravitational shear field. We compare
our results on the tomographic shear correlation with data from two
recent weak-lensing surveys, KiDS-450 and DLS. The main results are
summarized in the following.
To produce galaxies images on a curved sky, which is needed for a survey
with a large sky coverage, we follow the methods of \citet{2013MNRAS.435..115B},
and perform a high-spatial and mass resolution ray tracing
with an iterative scheme of spherical harmonic analysis. We compare
the measured power spectrum of convergence and shear with the
analytical Halofit model and the Born approximation. It is found that
the measured power spectrums of convergence and shear E-/B-mode have
good agreement with the revised non-linear Halofit prediction
\citep{2012ApJ...761..152T}. The prediction from Born approximation gives
higher power at small scales than ray-tracing simulation with $> 10\%$ for $\ell \ge 6000$.
We follow \citet{2013MNRAS.431..477J, 2013MNRAS.436..819J} to assign
shapes to model galaxies. For an early-type central galaxy, its major
axis is assumed to align with that of the host dark matter halo, and for
late-type central galaxy its spin follows that of the halo with major
axis determined by projecting the circular disk on the sky. For
early-type satellite galaxies, they are radially aligned with the
central galaxy, and for late-type satellites their spin lies in the
plane perpendicular to the radial direction to the central galaxy. We
also consider an additional model in which satellite galaxies have
random orientations. Using this modes for galaxy shapes, we find that
early-type central galaxies have strong {intrinsic ellipticity correlation} but
late-type galaxies have very weak alignment, in broad agreement with
observations.
To compare with the observational data of KiDS-450 and DLS, we produce
mock surveys by mimicking their sky coverage, galaxy number density
and redshift distribution of source galaxies. It is found that our
model with random orientation of satellites agrees well with
KiDS-450. Using the covariance matrix of the data \citep{2017MNRAS.465.1454H},
we can follow the Eq.~\ref{eq:eq30} to give the reduced $\chi^{2} = 1.36$
for the full data vector of KiD-450. This reduced $\chi^{2}$ is slightly
higher than that in the fiducial analysis of KiDS-450,
where they obtain a reduced $\chi^{2} = 1.33$ \citep{2017MNRAS.465.1454H}.
In addition, we also calculate the reduced $\chi^{2}$ for $\xi_{+}$ and $\xi_{-}$
separately, and we find the reduced $\chi^{2}$ is 1.23 and 1.61 respectively.
To further quantity the difference between our model prediction and the data,
we estimate this systematic bias $S$ by Eq.~\ref{eq:eq32}. The result shows
that the systematic bias between our model prediction and KiDS-450 result is
$S=1.80$ and $1.92$ for $\xi_{+}$ and $\xi_{-}$, respectively. In other
words, the $\xi_{\pm}$ we predicted are systematically higher than what KiDS measured
with a significance of $\sim 1.8 \sigma$.
On other hand, considering that the data of $\xi_{-}$ is not available for the DLS data,
we only compare our model prediction with the data in $\xi_{+}$. Following Eq.~\ref{eq:eq29},
we calculate the reduced $\chi_{+}^{2}$ between the simulation and the data is $1.56$ for
the DLS data, which seems to be acceptable in our work. While we also find a strong
negative systematic bias $S=-2.57$ between the model prediction and the DLS result,
which indicates that the DLS data are systematically higher than our model predictions.
Since the cosmic variance has been taken into account in the error bar of DLS, such
a large systematic deviation can hardly be explained by the cosmic variance only.
Moreover, assuming that there is no scatter in the alignment angle,
we rule out the model in which satellite galaxies are radially aligned
with central galaxies, as it produces too strong power on small scales.
We also study the contributions of the II and GI terms on the total shear correlations.
It is found that the II term is consistent with zero,
as in our model most galaxies are spirals and they have very weak
intrinsic alignment. Most importantly, we detect a positive GI term
which is mainly contributed by spiral galaxies. The GI term can be
up to 15 per cent on large scales, and so its effect on the total shear
correlation cannot be neglected. A positive GI term is a result of
the correlation between the spins of spirals and the large-scale
structure, where it is found that spiral galaxies are significantly
tangential aligned with the nearby over-dense regions.
This alignment is different from that of elliptical galaxies, which are radially
aligned with the over-dense regions, and produces a negative GI
term.
Finally, we note that in our simulation the shape orientation of model
galaxies is determined by the host dark matter halo, which is
probably too simplistic. In fact, there should be mis-alignments for both elliptical
and spiral galaxies. Quantifying these mis-alignments and their dependence on
galaxy properties with observations or hydro-dynamical simulations is crucial.
Our results suggest that an accurate model of GI term is very important for weak-lensing survey,
and it must include the dependence on galaxy type.
\acknowledgments
The authors would like to thank M. R. Becker for making \texttt{CALCLENS} available
and H. Hildebrandt for providing the redshift distribution of KiDS-450. Also thanks
are given to Catherine Heymans, John Peacock, Yanchuan Cai , Jun Zhang, Zuhui Fan,
Liping Fu and Xiangkun Liu for useful comments and suggestions.
We also thank the anonymous referee for constructive reports
which significantly improve the quality of the paper.
The work is supported by the NSFC (No.11333008), the 973 program (No. 2015CB857003,
2015CB857002, 2013CB834900), the NSFC (11273179,
11673065, 11273061, 11233005, 11621303, 11522324, 11421303, and 11473053),
and the NSF of Jiangsu province (No. BK20140050).
|
\section{Introduction}
\label{sec:introduction}
Probabilistic numerics \citep{Hennig2015} proposes approaching problems of numerical analysis from the point of view of statistics. In particular, Bayesian probabilistic numerical methods approach this problem from a Bayesian point of view, and can provide posterior distributions on the solutions of numerical problems (e.g. in the case of this paper, the solution of some integral). These posterior distributions represent our epistemic uncertainty about these quantities of interest. In the case of quadrature rules, the uncertainty is due to the fact that we only have a finite number of function evaluations and therefore are uncertaint about the value of the integral. The notion of \emph{Bayesian probabilistic numerical method} was independently introduced by several authors \citep{Larkin1972,Kadane1985,Diaconis1988,OHagan1992}, but only recently formalised by \cite{Cockayne2017}.
Apart from the uncertainty quantification property described above, these methods have several other advantages over ``classical" (i.e. non-Bayesian) numerical methods (although some of the classical and Bayesian methods coincide \citep{Diaconis1988}). First of all, they allow the user to formulate all of its prior knowledge in the form of a prior, making all of the assumptions of the numerical scheme explicit. Second of all, they can allow for coherent propagation of numerical uncertainties through chains of computation; see \citep{Cockayne2017,Oates2017hydrocyclones}.
However, one property which has not been studied so far is the possibility of jointly inferring several quantities of interest. In this paper, we study the problem of numerically integrating a sequence of functions $f_1,\ldots,f_D$ (which are correlated to one another) with respect to some probability measure $\Pi$, and hence propose to build a model for joint inference of $\int f_1 \mathrm{d}\Pi,\ldots,\int f_D \mathrm{d}\Pi$. Such a joint model allows for better finite-sample performance, and can also lead to more refined posterior distributions on each of the individual integrals.
To tackle this problem, we extend the well-known Bayesian quadrature \citep{OHagan1991} algorithm and study the performance of the proposed methodology from a theoretical and experimental point of view. In particular, we provide asymptotic convergence results for the marginal posterior variance on each of the integrals, both in the case of a well specified and misspecified prior. We also demonstrate the performance of our algorithm on some toy problems from the engineering literature on multi-fidelity models, and on a challenging problem from the field of computer graphics.
\section{Methodology}
\paragraph{Bayesian Quadrature}
Let $(\mathcal{X},\mathcal{B},\Pi)$ be a probability space and consider some function $f:\mathcal{X}\rightarrow \mathbb{R}$ where $\mathcal{X}\subseteq \mathbb{R}^p, p \in \mathbb{N}_+$. The classical problem of numerical integration is concerned with approximating the integral:
\begin{eqnarray*}
\Pi[f] & := & \int_{\mathcal{X}} f(\bm{x}) \Pi(\mathrm{d}\bm{x}),
\end{eqnarray*}
where we assume $\int_{\mathcal{X}} f^2(\bm{x}) \Pi(\mathrm{d}\bm{x}) < \infty$. Under fairly general conditions on $f$, one can show that an optimal algorithm (in terms of worst-case integration error in some function space) takes the form of a quadrature (or cubature) rule $\hat{\Pi}[f] = \sum_{i=1}^N w_i f(\bm{x}_i)$ for some weights $\{w_i\}_{i=1}^N \in \mathbb{R}$ and samples $\{\bm{x}_i\}_{i=1}^N \in \mathcal{X}$ (see \cite{Bakhvalov1971}). These are also sometimes denoted in vectorised form as $\Pi[f]=\bm{w}^\top f(\bm{X})$ where $\bm{w}=(w_1,\ldots,w_N)^\top$, $\bm{X}=(\bm{x}_1,\ldots,\bm{x}_N)^\top$ and $f(\bm{X})=(f(\bm{x}_1),\ldots,f(\bm{x}_N))^\top$. The notation $\hat{\Pi}[f]$ is motivated by the fact that we can see this object as an exact integral with respect to a discrete measure $\hat{\Pi}=\sum_{i=1}^N w_i \delta_{\bm{x}_i}$, where $\delta_{\bm{x}_i}$ denotes the Dirac delta measure taking value $1$ at $\bm{x}_i$ and $0$ otherwise. Many popular numerical integration methods take this form, including Newton--Cotes rules, Gaussian quadrature, Monte Carlo methods and sparse grids.
Let $(\Omega,\mathcal{F},\mathbb{P})$ be another probability space. \textit{Bayesian quadrature} (BQ), introduced by \cite{OHagan1991}, proposes to approach the problem of numerical integration by first formulating a prior stochastic model $g:\mathcal{X}\times \Omega \rightarrow \mathbb{R}$ for the integrand $f$ (where $\forall \omega \in \Omega$, $g(\cdot,\omega)$ represents a realisation of $g$). This prior model is then conditioned on the vector of observations $f(\bm{X})$ to obtain a posterior model for $f$, which is then pushed forward through the integral operator to give a posterior on $\Pi[f]$.
A popular choice of prior is a Gaussian Process (GP) $\mathcal{GP}(m,k)$ with $m:\mathcal{X}\rightarrow \mathbb{R}$ denoting the mean function (i.e. $m(\bm{x}) = \mathbb{E}_{\omega}[g(\bm{x},\omega)]$), and $c:\mathcal{X}\times \mathcal{X}\rightarrow \mathbb{R}$ denoting the covariance function/kernel (i.e. $c(\bm{x},\bm{x}') = \mathbb{E}_{\omega}[(g(\bm{x},\omega)-m(\bm{x}))(g(\bm{x}',\omega)-m(\bm{x}'))]$). Let us assume that $m=0$ (this can be done without loss of generality since the domain can be re-parametrized to be centred at $0$). After conditioning on $X$, we have a new Gaussian process $g_N$ with mean and covariance:
\begin{eqnarray*}
m_N(\bm{x}) & = & c(\bm{x},\bm{X})c(\bm{X},\bm{X})^{-1}f(\bm{X}), \\
c_N(\bm{x},\bm{x}') & = & c(\bm{x},\bm{x}') - c(\bm{x},\bm{X})c(\bm{X},\bm{X})^{-1}c(\bm{X},\bm{x}'),
\end{eqnarray*}
for all $\bm{x},\bm{x}' \in \mathcal{X}$. Here, $c(\bm{X},\bm{X})$ is the Gram matrix with entries $(c(\bm{X},\bm{X}))_{ij} = c(\bm{x}_i,\bm{x}_j)$ and $c(\bm{x},\bm{X})=(c(\bm{x},\bm{x}_1),\ldots,c(\bm{x},\bm{x}_N))$ whilst $c(\bm{X},\bm{x}) = c(\bm{x},\bm{X})^\top$ . The push-forward of this posterior through the integral operator is a Gaussian distribution with mean and variance:
\begin{eqnarray*}
\mathbb{E}\left[\Pi[g_N]\right] & = & \Pi[c(\cdot,\bm{X})]c(\bm{X},\bm{X})^{-1} f(\bm{X}),\\
\mathbb{V}\left[\Pi[g_N]\right] & = & \Pi\bar{\Pi} \left[c\right] - \Pi[c(\cdot,\bm{X})]c(\bm{X},\bm{X})^{-1}\bar{\Pi}[c(\bm{X},\cdot)],
\end{eqnarray*}
where $\Pi[c(\cdot,\bm{X})] = (\Pi[c(\cdot,\bm{x}_1)],\ldots,\Pi[c(\cdot,\bm{x}_N)])$. These expression can be obtained in closed-form if the \textit{kernel mean} $\Pi[c(\cdot,\bm{x})]=\int_{\mathcal{X}} c(\bm{x}',\bm{x}) \Pi(\mathrm{d}\bm{x}')$ (also called the representer of integration) and \textit{initial error} $\Pi\bar{\Pi}[c] = \int_{\mathcal{X}\times \mathcal{X}} c(\bm{x},\bm{x}') \Pi(\mathrm{d}\bm{x}) \Pi(\mathrm{d}\bm{x}')$ can be obtained in closed form (here $\bar{\Pi}$ indicates that the integral is taken with respect to the second argument).
The choice of covariance function $c$ can be used to encode prior beliefs about the function $f$, such as smoothness or periodicity, and is very important to obtain good performance in practice. A popular example is the family of Mat\'ern kernels
\begin{equation}\label{eq:matern_kernels}
\begin{split}
c_{\alpha}(\bm{x},\bm{x}') \; & = \; \lambda^2 \frac{2^{1-\alpha}}{\Gamma(\alpha)} \left(\sqrt{2 \alpha} \frac{\|\bm{x}-\bm{x}'\|_{2}^{2}}{\sigma^2}\right)^{\alpha} \\
& \qquad \qquad \qquad \times J_{\alpha} \left(\sqrt{2 \alpha} \frac{\|\bm{x}-\bm{x}'\|_{2}^{2}}{\sigma^2}\right),
\end{split}
\end{equation}
for $\sigma,\lambda>0$ where $J_{\alpha}$ is the Bessel function of the second kind and $\alpha>0$ gives the smoothness of the kernel. On $\mathcal{X}=\mathbb{R}^p$, this will give an RKHS norm-equivalent to the Sobolev space $W_2^{\alpha}(\mathbb{R}^d)$\footnote{We say that two norms $\|\cdot\|_{1}$ and $\|\cdot\|_{2}$ on a vector space are \emph{norm-equivalent} if and only if $\exists C_1,C_2>0$ such that $C_1 \|\cdot\|_{2} \leq \|\cdot\|_{1} \leq C_2 \|\cdot\|_{2}$.}. Examples of infinitely smooth kernels include the squared-exponential kernel $c(\bm{x},\bm{x}') = \exp(-\|\bm{x}-\bm{x}'\|_2^2/\sigma^2)$ where $\sigma>0$, the multi-quadric kernel $c(\bm{x},\bm{x}') = (-1)^{\ceil{\beta}}(\sigma^2+\|\bm{x}-\bm{x}'\|_2^2)^{\beta}$ for $\beta,\sigma>0, \beta \not\in \mathbb{N}$ and the inverse multi-quadric kernel $c(\bm{x},\bm{x}') = (\sigma^2+ \|\bm{x}-\bm{x}'\|_2^2)^{-\beta}$ for $\beta,\sigma>0$.
In practice, numerical inversion can be challenging since the Gram matrix tends to be nearly singular, and so one may wish to regularise the matrix using a Tikhonov penalty. The inverses above can also potentially render the computation of the BQ estimator computationally expensive (up to $\mathcal{O}(N^3)$ cost in the most general settings), although this can be alleviated in specific cases \cite{Karvonen2017symm}. Even if this is not the case, the additional cost can be worthwhile regardless since the method has been shown to attain fast convergence rates \cite{Briol2015,Briol2015a,Kanagawa2016,Kanagawa2017,Bach2015} when the target integrand and the kernel used are smooth.
Recent research directions in BQ include efficient sampling algorithms (for the point set $\bm{X}$) to improve the performance of the method \cite{Rasmussen2003,Huszar2012,Gunter2014,Briol2015,Karvonen2017,Briol2017}, asymptotic convergence results \cite{Briol2015,Briol2015a,Kanagawa2016,Bach2015} and equivalence of BQ with known quadrature rules for certain choices of point sets and kernels \cite{Sarkka2015,Karvonen2017}. Furthermore, there has also been a wide range of new applications, including to other numerical methods in optimization, linear algebra and functional approximation \cite{Kersting2016,Fitzsimons2017}, inference in complex computer models \cite{Oates2016}, and problems in econometrics \cite{Oettershagen2017} and computer graphics \cite{Brouillat2009,Marques2013,Briol2015a}.
Although other stochastic processes could of course be used \citep{Cockayne2017}, GPs are popular due to their conjugacy properties, and the terminology Bayesian quadrature usually refers to this case. Note that other names for BQ with GP priors include Gaussian-process quadrature \citep{Sarkka2015} or kernel quadrature \citep{Bach2015,Briol2017,Kanagawa2017}. In fact, a well-known alternative view of the posterior mean provided by BQ is that of an optimally-weighted quadrature rule in a reproducing kernel Hilbert spaces (RKHS) in the classical worst-case setting \cite{Ritter2000}. Let $\mathcal{H}_k$ be an RKHS with inner product and norm denoted $\langle\cdot,\cdot\rangle_k$ and $\|\cdot\|_k$ respectively; i.e. a Hilbert space with an associated symmetric and positive definite reproducing kernel $k:\mathcal{X}\times \mathcal{X} \rightarrow \mathbb{R}$ such that $f(\bm{x}) = \langle f, k(\cdot,\bm{x})\rangle_k$ (see \cite{Berlinet2004} for a detailed study). Suppose that our integrand $f\in \mathcal{H}_k$ and that $\int_{\mathcal{X}} k(\bm{x},\bm{x}) \Pi(\mathrm{d}\bm{x}) < \infty$. In that case, using the Cauchy--Schwarz inequality, the integration error can be decomposed as:
\begin{eqnarray*}
\left|\Pi[f] - \hat{\Pi}[f]\right| & \leq & \left\|f\right\|_{k} \left\| \Pi\left[k(\cdot,\bm{x})\right] - \hat{\Pi}\left[k(\cdot,\bm{x})\right] \right\|_{k}.
\end{eqnarray*}
The corresponding worst-case error over the unit ball of the space $\mathcal{H}_k$ is given by:
\begin{multline*}
e \left(\mathcal{H}_k,\hat{\Pi},\bm{X}\right)
\; = \; \sup_{\|f\|_k \leq 1} \left|\Pi[f] - \hat{\Pi}[f]\right| \\
= \; \left\| \Pi\left[k(\cdot,\bm{x})\right] - \hat{\Pi}\left[k(\cdot,\bm{x})\right] \right\|_{k} \hspace{23mm} \\
= \; \Big(\bm{w}^\top k(\bm{X},\bm{X}) \bm{w} - 2\Pi[k(\cdot,\bm{X})]^\top \bm{w} + \Pi \bar{\Pi}[k]\Big)^{\frac{1}{2}}.
\end{multline*}
This final expression can be minimised in closed form over $\bm{w} \in \mathbb{R}^{N}$ to show that the optimal quadrature rule has weights $\bm{w} = \Pi[k(\cdot,\bm{X})] k(\bm{X},\bm{X})^{-1}$. This corresponds exactly to the weights for the BQ posterior mean if we take our prior on $f$ to be a $\mathcal{GP}(0,k)$, whilst the worst-case error can be shown to correspond to the posterior variance squared. The BQ estimator with prior $\mathcal{GP}(0,c)$ is therefore optimal in the classical worst-case sense for the RKHS $\mathcal{H}_c$.
\paragraph{Multi-output Bayesian Quadrature}\label{sec:multioutput_BQ}
We now extend the set-up of our problem. Suppose we have a sequence of probability spaces $(X_d,\mathcal{B}_d,\Pi_d)$ and functions $f_d:\mathcal{X}_d \rightarrow \mathbb{R}$ for which we are interested in numerically computing integrals of the form $\Pi_d[f_d]$ for $d=1,\ldots,D$. In many applications where we are faced with this type of problem, we also have prior knowledge about correlations between the individual $f_d$. However, this information is often ignored and the integrals are approximated individually. This is not a principled approach from a Bayesian point of view since it means we are not conditioning on all available information. In this section, we extend the BQ algorithm to solve this problem by building a joint model of $f_1,\ldots,f_D$ in order to obtain a joint posterior on the integrals $\Pi_1[f_1],\ldots,\Pi_D[f_D]$.
For notational convenience, we will restrict ourselves to the case where all of the input domains are identical and denoted $\mathcal{X}$, all of the probability measures are identical and denoted $\Pi$, and the input sets $\bm{X} = \{\bm{X}_d\}_{d=1}^D$ consist of $N$ points $\bm{X}_d = (\bm{x}_{d,1},\ldots,\bm{x}_{d,N})$ per output function $f_d$ (note the setup can be made more general if necessary). We re-frame the integration problem as that of integrating some vector-valued function $\bm{f}:\mathcal{X} \rightarrow \mathbb{R}^D$ such that $\bm{f}(\bm{x}) = (f_1(\bm{x}),\ldots,f_D(\bm{x}))^\top$; i.e. we want to estimate $\Pi[\bm{f}] = (\Pi[f_1],\ldots,\Pi[f_D])^\top$. In this multiple-integral setting, we can have generalised quadrature rules of the form:
\begin{eqnarray*}
\hat{\Pi}[f_d] & = & \sum_{d'=1}^D \sum_{i=1}^N (\bm{W}_{i})_{dd'} f_{d'}(\bm{x}_{d',i})
\end{eqnarray*}
where $\bm{W}_{i} \in \mathbb{R}^{D \times D}$ are weight matrices and $(\bm{W}_{i})_{dd'}$ gives the influence of the value of $f_{d'}$ at $\bm{x}_{d',i}$ on the estimate of $\Pi[f_d]$. The quadrature rule for $\bm{f}$ can be re-written in compact form as $\hat{\Pi}[\bm{f}] = \bm{W}^\top \bm{f}(\bm{X})$ for some weight matrix $\bm{W} \in \mathbb{R}^{ND \times D}$ (a concatenation of $\{\bm{W}_i\}_{i=1}^N$) and function-evaluations vector $\bm{f}(\bm{X}) = (f_1(\bm{x}_{1,1}),\ldots,f_1(\bm{x}_{1,N}),\ldots,f_D(\bm{x}_{D,1}),\ldots,f_D(\bm{x}_{D,N}))^\top$.
These generalised quadrature rules encompass popular Monte Carlo methods such as control variates or functionals \cite{Glasserman2004,Oates2017}, multilevel Monte Carlo \cite{Giles2015} and multi-fidelity Monte Carlo \cite{Peherstorfer2016MC}. However, it is important to point out that these methods can only deal with very specific relations between integrands, usually requiring $(\int_{\mathcal{X}} (f_d(\bm{x})-f_{d'}(\bm{x}))^2 \Pi(\mathrm{d}\bm{x}))^{\frac{1}{2}}$ to be small for all pairs of integrands $f_d,f_{d'}$. Our method will be able to make use of much more complex relations.
We propose to approach this problem using an extended version of BQ, where we impose a prior $\bm{g}:\mathcal{X}\times \Omega \rightarrow \mathbb{R}^D$ which is a $\mathcal{GP}(\bm{0},\bm{C})$ on the extended space (this is often called a multi-output GP or co-kriging model \cite{Alvarez2011}) where now $\bm{C}$ is matrix-valued and $(C(\bm{x},\bm{x}'))_{dd'} = \mathbb{E}_{\omega \sim \mathbb{P}}[g_d(\bm{x},\omega)g_{d'}(\bm{x}',\omega)]$. In this case, after conditioning on $\bm{X}$, we have a GP $\bm{g}_N$ with vector-valued mean $\bm{m}_N:\mathcal{X}\rightarrow \mathbb{R}^D$ and matrix-valued covariance $\bm{C}_N: \mathcal{X} \times \mathcal{X}\rightarrow \mathbb{R}^{D \times D}$:
\begin{align*}
\bm{m}_N(\bm{x}) & = \bm{C}(\bm{x},\bm{X})\bm{C}(\bm{X},\bm{X})^{-1} \bm{f}(\bm{X}), \\
\bm{C}_N(\bm{x},\bm{x}') & = \bm{C}(\bm{x},\bm{x}') - \bm{C}(\bm{x},\bm{X})\bm{C}(\bm{X},\bm{X})^{-1}\bm{C}(\bm{X},\bm{x}'),
\end{align*}
for $\bm{C}(\bm{x},\bm{X}) = (C(\bm{x},\bm{x}_1),\ldots,C(\bm{x},\bm{x}_N)) \in \mathbb{R}^{D \times ND}$ and Gram matrix $\bm{C}(\bm{X},\bm{X}) \in \mathbb{R}^{ND \times ND}$ is:
\begin{equation*}
\bm{C}(\bm{X},\bm{X}) =
\begin{bmatrix}
\left(\bm{C}(\bm{X}_1,\bm{X}_1)\right)_{1,1} & \dots & \left(\bm{C}(\bm{X}_1,\bm{X}_D)\right)_{1,D} \\
\left(\bm{C}(\bm{X}_2,\bm{X}_1)\right)_{2,1} & \vdots & \left(\bm{C}(\bm{X}_2,\bm{X}_D)\right)_{2,D} \\
\vdots & \vdots & \vdots \\
\left(\bm{C}(\bm{X}_D,\bm{X}_1)\right)_{D,1} & \dots & \left(\bm{C}(\bm{X}_D,\bm{X}_D)\right)_{D,D}
\end{bmatrix},
\end{equation*}
where $C(\bm{X}_d,\bm{X}_{d'})_{d,d'}$ is an $N\times N$ matrix. The posterior on the value of the integral vector $\Pi[\bm{f}]$ can also be obtained whenever the kernel mean $\Pi[\bm{C}(\cdot,\bm{x})]$ and initial error $\Pi\bar{\Pi} \left[\bm{C}\right]$ are available in closed form, which is potentially a restrictive condition. The authors of \cite{Briol2015a} give a table of closed-form expressions of these quantities for popular kernels in the uni-output case, and we envision the same type of table being necessary for future extensions of multi-output BQ. Alternatively, \cite{Oates2017,Oates2016CF2} proposed a kernel which is tailored to the target probability measure $\Pi$ and which could also be extended to the multi-output case.
\begin{proposition}\label{prop:multioutput_BQ}
Consider multi-output Bayesian Quadrature with a $\mathcal{GP}(\bm{0},\bm{C})$ prior on $\bm{f}=(f_1,\ldots,f_D)^\top$. The posterior distribution on $\Pi[\bm{f}]$ is a $D$-dimensional Gaussian distribution with mean and covariance matrix:
\begin{eqnarray*}
\mathbb{E} \left[\Pi[\bm{g}_N]\right] & = & \Pi[\bm{C}(\cdot,\bm{X})]\bm{C}(\bm{X},\bm{X})^{-1} \bm{f}(\bm{X}),\\
\mathbb{V} \left[\Pi[\bm{g}_N]\right] & = & \Pi\bar{\Pi} \left[\bm{C}\right] - \Pi[\bm{C}(\cdot,\bm{X})]\bm{C}(\bm{X},\bm{X})^{-1}\bar{\Pi}[\bm{C}(\bm{X},\cdot)].
\end{eqnarray*}
\end{proposition}
All proofs can be found in Appendix \ref{sec:appendix_proofs}. In this case, we clearly end up with a generalised quadrature rule with weight matrix:
$\bm{W}^{\text{BQ}} = (\Pi \left[\bm{C}(\cdot,\bm{X})\right]\bm{C}(\bm{X},\bm{X})^{-1})^{\top} \in \mathbb{R}^{ND \times D}$.
In general, the computational cost for computing the posterior mean and variance is now of order $\mathcal{O}(N^3D^3)$. However, several choices of kernels can reduce this cost significantly, and it is also possible to obtain sparse GP approximations; see e.g. \cite{Alvarez2011convolved}.
The choice of kernel $\bm{C}$ is of course once again of great importance since it encodes prior knowledge about each of the integrand and their correlation structure and should be made based on the application considered. We also remark that matrix valued kernels $\bm{C}$ can be described in term of some scalar-valued kernel $r$ on the extended space $\mathcal{X}\times\{1,\ldots,D\}$ as $(\bm{C}(\bm{x},\bm{x}'))_{dd'} = r((\bm{x},d),(\bm{x}',d'))$. We now present two choices of covariance functions which are popular in the literature and will be used in this paper:
\begin{itemize}
\item The \textit{separable} kernel is of the form
\begin{eqnarray*}
\bm{C}(\bm{x},\bm{x}') & = & \bm{B} c(\bm{x},\bm{x}'),
\end{eqnarray*}
where $\bm{B} \in \mathbb{R}^{D \times D}$ is symmetric and positive definite, and $\bm{c}:\mathcal{X} \times \mathcal{X} \rightarrow \mathbb{R}$ is a scalar-valued reproducing kernel. This treats the kernel as the product of two scalar-valued reproducing kernels, one defined on $\mathcal{X}$ and the other on $\{1,\ldots,D\}$. A particular case of interest is the \textit{linear model of coregionalization} (LMC) where the matrix is of the form $(\bm{B})_{dd'} = \sum_{i=1}^{R} a^i_{d} a^i_{d'}$ for some $a^i_{d} \in \mathbb{R}$. This type of kernel can lead to a lower computational cost of order $\mathcal{O}(N^3 + D^3)$ when evaluating all $f_d$ on the same input set and using tensor product formulations (see Appendix \ref{sec:appendix_implementation}).
\item The \textit{process convolution} kernel \citep{VerHoef1998,Higdon2002,Alvarez2011} models the individual functions $f_1,\ldots,f_D$ as blurred transformations of $R \in \mathbb{N}_+$ underlying functions. It is given by:
\begin{eqnarray*}
(\bm{C}(\bm{x},\bm{x}'))_{d,d'} & = & c_{d,d'}(\bm{x},\bm{x}') + c_{w_d}(\bm{x},\bm{x}') \delta_{d,d'},
\end{eqnarray*}
where $\delta_{dd'} = 1$ if $d=d'$ and $0$ else. Here there are two parts of the kernel, first $c_{d,d'}:\mathcal{X}\times \mathcal{X}\rightarrow \mathbb{R}$ defined as:
\begin{eqnarray*}
c_{d,d'}(\bm{x},\bm{x}') & = & \sum_{i=1}^{R} \int_{\mathcal{X}} G^i_{d}(\bm{x}-\bm{z}) \times \\ & & \; \int_{\mathcal{X}} G^i_{d'}(\bm{x}'-\bm{z}') c_i(\bm{z},\bm{z}') \mathrm{d}\bm{z}' \mathrm{d}\bm{z},
\end{eqnarray*}
and $c_{w_d}:\mathcal{X}\times \mathcal{X}\rightarrow \mathbb{R}$ representing covariance inherent to the $d^\text{th}$ function and $G_{d}^i:\mathcal{X}\rightarrow \mathbb{R}$ is a blurring kernel\footnote{Note that the term ``blurring kernel" does not mean the function is a reproducing kernel.} which is a continuous function either having compact support or being square integrable. Notice that taking $G^i_{d}(\bm{x}-\bm{z}) = a^i_{d} \delta(\bm{x}-\bm{z})$ (where $\delta(\cdot)$ represents a Dirac function) gives back the LMC case.
\end{itemize}
Note that it is also common to combine kernels, by summing them (i.e. $\bm{C}(\bm{x},\bm{x}') = \sum_{q=1}^Q \bm{C}_q(\bm{x},\bm{x}')$) in order to obtain more flexible models. The kernel means and initial error, as well as other details for implementation are provided in Appendix \ref{sec:appendix_implementation}.
\section{Theoretical results}
\label{sec:MOBQ_theory}
In this section, we begin by exploring properties of multi-output BQ with $\mathcal{GP}(\bm{0},\bm{C})$ prior as an optimally-weighted quadrature algorithm in vector-valued RKHS $\mathcal{H}_{\bm{C}}$.
Let $\mathcal{H}_{\bm{K}}$ be a vector-valued RKHS with norm and inner product denoted $\|\cdot\|_{\bm{K}}$ and $\langle \cdot, \cdot \rangle_{\bm{K}}$ respectively. These spaces were extensively studied in \cite{Pedrick1957,Micchelli2008,Carmeli2006,Carmeli2008,DeVito2013}, and generalise the notion of RKHS to vector-valued functions. In the multi-output case, there is also a one-to-one correspondance between the RKHS $\mathcal{H}_{\bm{K}}$ and the kernel $\bm{K}$. Theorem 3.1 in \cite{Micchelli2008} shows that the minimizer of the variational problem:
\begin{eqnarray*}
\min_{\bm{h} \in \mathcal{H}_{\bm{K}}} \left\{ \|\bm{h}\|_{\bm{K}}^2 \; : \; \bm{h}:\mathcal{X}\rightarrow \mathbb{R}^D, \bm{h}(\bm{x}_i) = \bm{f}(\bm{x}_i) \; \forall \bm{x}_i \in \bm{X} \right\}
\end{eqnarray*}
takes the form of the multi-output posterior GP mean $\bm{m}_N$ obtained after conditioning a $\mathcal{GP}(\bm{0},\bm{K})$ on some data set $\bm{X}$. We can therefore extend a well-known result from the uni-output case to show that $\hat{\Pi}_{\text{BQ}}[f_d]$ is an optimally weighted quadrature rule for all $f_d$ in terms of their worst-case integration error, denoted:
\begin{eqnarray}\label{eq:MOBQ_WCE}
e(\mathcal{H}_{\bm{C}},\hat{\Pi},\bm{X},d)
& := &
\sup_{\|\bm{f}\|_{\bm{C}}\leq 1} \left| \Pi[f_d] - \hat{\Pi}[f_d]\right|.
\end{eqnarray}
\begin{proposition}[Optimally weighted quadrature rule in $\mathcal{H}_{\bm{C}}$]\label{prop:optimal_weights}
For a fixed point set $\bm{X}$, denote by $\hat{\Pi}[\bm{f}] = \bm{W}^\top \bm{f}(\bm{X})$ any quadrature rule for the vector-valued function $\bm{f}=(f_1,\ldots,f_D)$ and by $\hat{\Pi}_{\text{BQ}}[\bm{f}]=\bm{W}^\top_{\text{BQ}}\bm{f}(\bm{X})$ the BQ rule with $\mathcal{GP}(\bm{0},\bm{C})$ prior. Then, $\forall d=1,\ldots,D$:
\begin{eqnarray*}
\bm{W}_{\text{BQ}} & = & \argmin_{\bm{W} \in \mathbb{R}^{ND \times D}} e(\mathcal{H}_{\bm{C}},\hat{\Pi},\bm{X},d).
\end{eqnarray*}
\end{proposition}
In specific cases, it is also possible to characterise the rate of convergence of the worst-case error for each element $f_d$. This is for example the case with the separable kernel introduced in Sec. \ref{sec:multioutput_BQ}, as will be demonstrated in the Theorem \ref{theorem:convergence_separable} below. First, we introduce some technical definitions which will be required for the statement of the theorem.
We say that a domain $\mathcal{X} \subset \mathbb{R}^p$ satisfies an \textit{interior cone condition} if there exists an angle $\theta \in (0,\frac{\pi}{2})$ and a radius $r>0$ such that $\forall \bm{x} \in \mathcal{X}$, a unit vector $\bm{\xi}(\bm{x})$ exists such that the cone $\{\bm{x}+\lambda \bm{y}: \bm{y} \in \mathbb{R}^p, \|\bm{y}\|_2 =1, \bm{y}^\top \bm{\xi}(\bm{x}) \geq \cos \theta, \lambda \in [0,r]\}$ is a subset of $\mathcal{X}$.
For a point set $\bm{X}$, we call $h_{\bm{X},\mathcal{X}} := \sup_{\bm{x} \in \mathcal{X}} \inf_{\bm{x}_j \in \bm{X}} \|\bm{x} - \bm{x}_j\|_2$ the \emph{fill distance}, $q_{\bm{X}} := \frac{1}{2} \min_{j \neq k} \|\bm{x}_j -\bm{x}_k\|_2$ the \emph{separation radius} and $\rho_{\bm{X},\mathcal{X}} := h_{\bm{X},\mathcal{X}}/q_{\bm{X}}$ the \emph{mesh ratio}. We will assume we evaluate all integrands on the same point set $\bm{X}$ which satisfies either of these assumptions:
\begin{enumerate}[label=(\subscript{A}{{\arabic*}})]
\item $\bm{X}$ consists of independently and identically distributed (IID) samples from some probability measure $\Pi'$ which admits a density $\pi'>0$ on $\mathcal{X}$.
\item $\bm{X}$ is a \emph{quasi-uniform} grid on $\mathcal{X} \subset \mathbb{R}^p$ (i.e. satisfies $h_{\bm{X},\mathcal{X}} \leq C_1 N^{-\frac{1}{p}}$ for some $C_1>0$) and satisfies $h_{\bm{X},\mathcal{X}} \leq C_2 q_{\bm{X},\mathcal{X}}$ for some $C_2>0$.
\end{enumerate}
Examples of point sets satisfying $(A_2)$ include uniform grid points in some hypercube.
\begin{theorem}[Convergence rate for BQ with separable kernel]\label{theorem:convergence_separable}
Suppose we want to approximate $\Pi[\bm{f}]$ for some $\bm{f}:\mathcal{X}\rightarrow \mathbb{R}^D$ and $\hat{\Pi}_{\text{BQ}}[\bm{f}]$ is the multi-output BQ rule with the kernel $\bm{C}(\bm{x},\bm{x}') = \bm{B} c(\bm{x},\bm{x}')$ for some positive definite $\bm{B} \in \mathbb{R}^{D \times D}$ and scalar-valued kernel $c:\mathcal{X}\times \mathcal{X}\rightarrow \mathbb{R}$. Then, $\forall d=1,\ldots,D$, we have:
\begin{eqnarray*}
e(\mathcal{H}_{\bm{C}},\hat{\Pi}_{\text{BQ}},\bm{X},d)
& = &
\mathcal{O}\left( e(\mathcal{H}_c,\hat{\Pi}_{\text{BQ}},\bm{X} )\right).
\end{eqnarray*}
In particular, assume that $\mathcal{X} \subset \mathbb{R}^p$ satisfies an interior cone condition with Lipschitz boundary\footnote{Formally defined in Appendix \ref{appendix:additional_background_MOBQ} for completeness.} and $\bm{X}$ satisfies assumption $(A_1)$ or $(A_2)$. Then, the following rates hold:
\begin{itemize}
\item If $\mathcal{H}_c$ is norm-equivalent to an RKHS with Mat\'ern kernel of smoothness $\alpha>\frac{p}{2}$, we have $\forall d=1,\ldots,D$:
\begin{eqnarray*}
e(\mathcal{H}_{\bm{C}},\hat{\Pi}_{\text{BQ}},\bm{X},d)
& = &
\mathcal{O}\left(N^{-\frac{\alpha}{p}+\epsilon}\right),
\end{eqnarray*}
for $\epsilon>0$ arbitrarily small.
\item If $\mathcal{H}_c$ is norm-equivalent to the RKHS with squared-exponential, multiquadric or inverse multiquadric kernel, we have $\forall d=1,\ldots,D$:
\begin{eqnarray*}
e(\mathcal{H}_{\bm{C}},\hat{\Pi}_{\text{BQ}},\bm{X},d)
& = &
\mathcal{O}\left(\exp\left(-C_1 N^{\frac{1}{p}-\epsilon}\right)\right),
\end{eqnarray*}
for some $C_1>0$ and for some $\epsilon>0$ arbitrarily small.
\end{itemize}
\end{theorem}
\begin{proposition}[Convergence rate for sum of kernels]\label{cor:sum_separable_kernels}
Suppose that $\bm{C}(\bm{x},\bm{x}') = \sum_{q=1}^Q \bm{C}_q(\bm{x},\bm{x}')$. Then:
\begin{eqnarray*}
e(\mathcal{H}_{\bm{C}},\hat{\Pi}_{\text{BQ}},\bm{X},d)
& = & \argmax_{q \in \{1,\ldots,Q\}} \mathcal{O}\left( e(\mathcal{H}_{\bm{C}_q},\hat{\Pi}_{\text{BQ}},\bm{X},d)
\right).
\end{eqnarray*}
\end{proposition}
We clarify that the notation with $\epsilon$ is common in the numerical integration literature, and is used to hide powers of $\log n$ terms since these do not have a significant influence on the asymptotic convergence rate.
It is interesting to note that the rate of convergence for multi-output BQ is the same as that of uni-output BQ \citep{Briol2015a}. This can be explained intuitively by the fact that, when adding a new integrand, we can only gain by a constant factor since we always evaluate the functions at the same input points. In fact the proof of Thm. \ref{theorem:convergence_separable} provides an expression for this improvement factor (in terms of WCE) for any integrand $f_d$, and this depends explicitly on its correlation with the other functions: $|\sum_{i,j=1}^D (\bm{B}^{-1})_{ij} \bm{B}_{id} \bm{B}_{jd}|$. From a practitioner's viewpoint, this can clearly be used to balance the value of using several integrands with the additional computational cost incurred by using multi-output BQ.
We now give a result in the misspecified setting when the function $f$ is assumed to be smoother than it is. In this case, it is still possible to recover the optimal convergence rate:
\begin{theorem}[Misspecified Convergence Result for Separable Kernel]\label{thm:misspecified_multioutputBQ} Let $c_{\alpha}$ be a kernel norm-equivalent to a Mat\'{e}rn kernel of smoothness $\alpha$ on some domain $\mathcal{X}$ with Lipschitz boundary and satisfying an interior cone condition. Consider the BQ rule $\hat{\Pi}_{\text{BQ}}[\bm{f}]$ corresponding to a separable kernel $\bm{C}_{\alpha}(x,x') = \bm{B} c_{\alpha}(x,x')$ with $\bm{X}$ satisfying $(A_2)$, and suppose that $\bm{f} \in \mathcal{H}_{\bm{C}_{\beta}}$ where $\frac{p}{2} \leq \beta \leq \alpha$. Then, $\forall d = 1,\ldots,D$:
\begin{eqnarray*}
\left| \Pi[f_d] - \hat{\Pi}_{\text{BQ}}[f_d]\right| & = & \mathcal{O}\left(N^{-\frac{\beta}{p}+\epsilon}\right),
\end{eqnarray*}
for some $\epsilon > 0$.
\end{theorem}
This last theorem demonstrate that the method is rate adaptive as long as we choose a kernel which is too smooth. However, it also demonstrates a drawback of the separable kernels: if one of the integrands is rough but all other are smooth, then the worst-case error could potentially converge slowly for all of them.
Finally, we note that studying the method in other information complexity settings than the worst-case would also be interesting. For example, it is trivial to show that the method above satisfies the definition of Bayesian probabilistic numerical method of \cite{Cockayne2017} (Def. 2.5). Furthermore, optimality conditions for this method could also be obtained in a game-theoretic setting (in terms of a two-player mixed strategies game) by extending the theory on gamblets by \cite{Owhadi2017}.
\section{Applications}\label{sec:application}
\paragraph{Multi-fidelity modelling}
\begin{figure}[t!]
\begin{center}
\includegraphics[trim={5cm 0 0 0},clip,width=0.55\textwidth]{Figures/Figure1_1}
\includegraphics[trim={5cm 0 0 0},clip,width=0.55\textwidth]{Figures/Figure1_2}
\caption{Multi-fidelity modelling: Plot of the Step function (top), Forrester function (bottom) for the low fidelity (left) and high fidelity (right). Each plot gives the true function (blue) and their unit-output (dashed, red), LMC-based multi-output (dashed, yellow) and PC-based multi-output (dotted purple) approximations.
\vspace{-3mm}}
\label{plot:multifidelity_functions}
\end{center}
\end{figure}
Consider some function $f^{\text{high}}:\mathcal{X}\rightarrow \mathbb{R}$ representing some complex engineering model of interest, which we would like to use for some task such as statistical inference or optimization. These models usually require the simulation of underlying physical systems, which can make each evaluation prohibitively expensive and will therefore limit $N$ to the order of tens or hundreds. To tackle this issue, multi-fidelity modelling proposes to build cheap, but less accurate, alternatives $f^{\text{low}}_1,\ldots,f^{\text{low}}_{D-1}:\mathcal{X}\rightarrow \mathbb{R}$ to $f^{\text{high}}$, and use the cheaper models in order to accelerate computation for the task of interest. This can be done using surrogate models (e.g. support vector machines, GPs or neural networks), projection-based models (Krylov subspace or reduced basis methods) or models where the underlying physics is simplified; see \cite{Peherstorfer2016} for an overview.
In this section, we consider the problem of numerical integration in such a multi-fidelity setup. Note that two related methods for Monte Carlo estimation are the multi-fidelity Monte Carlo estimator \citep{Peherstorfer2016} and the multilevel Monte Carlo of \citep{Giles2015}, both of which are based on control variate identities.
\begin{figure}[t!]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
Model & BQ & LMC-BQ & PC-BQ \\ [0.5ex]
\hline\hline
Step (l) & 0.02 (0.22) & 0.02 (0.21) & 0.02 (0.52) \\
Step (h) & 0.41 (0.03) & 0.09 (0.09) & 0.04 (0.15)\\
\hline
For. (l) & 0.08 (4.91) & 0.08 (4.95) & 0.07 (33.95)\\
For. (h) & 3.96 (3.98) & 2.86 (27.01) & 1.06 (63.80)\\
\hline
\end{tabular}
\end{center}
\caption{Multi-fidelity modelling: Performance of uni-output BQ and multi-output BQ (with LMC and PC kernels) on the step function (Step) and the Forrester function with jump (For.) in the low fidelity (l) and high fidelity (h) cases. The values given are absolute errors with the posterior variance in brackets.}
\label{table:multifidelity}
\end{figure}
We approach this problem with multi-output BQ on the vector-valued function $\bm{f} = (f^{\text{high}},f^{\text{low}}_{1},\ldots,f^{\text{low}}_{D-1})^\top$. Note that multi-output Gaussian processes were already proposed for multi-fidelity modelling in
\cite{Perdikaris2016,Parussini2017}, and we extend their methodologies to the task of numerical integration. We consider two toy problems from this literature \cite{Raissi2016} to highlight some of the advantages and disadvantages of our methodology
\begin{enumerate}
\item A \textit{step function} on $\mathcal{X} = [0,2]$:
\begin{equation*}
f_1^{\text{low}}(x) =
\begin{cases}
0, \; x \leq 1\\
1, \; x > 1
\end{cases}
f^{\text{high}}(x) =
\begin{cases}
-1, \; x \leq 1\\
2, \; x > 1
\end{cases}
\end{equation*}
\item The \textit{Forrester function with Jump} on $\mathcal{X} = [0,1]$:
\begin{equation*}
f_1^{\text{low}}(x) =
\begin{cases}
\frac{(3x-1)^2 \sin(12x-4)}{4} +10(x-1), \; x \leq \frac{1}{2}\\
3+ \frac{(3x-1)^2\sin(12x-4)}{4} +10(x-1), \; x > \frac{1}{2}
\end{cases}
\end{equation*}
\begin{equation*}
f^{\text{high}}(x) =
\begin{cases}
\quad 2 f^{\text{low}}(x)-20(x-1), \; x \leq \frac{1}{2}\\
4+2 f^{\text{low}}(x)-20(x-1), \; x > \frac{1}{2}
\end{cases}
\end{equation*}
\end{enumerate}
The functions and conditioned GPs are given in Fig. \ref{plot:multifidelity_functions}, whilst the uni-output and multi-output BQ estimates for integration of these functions against a uniform measure $\Pi$ are given in the table in Fig. \ref{table:multifidelity}. In both cases, $20$ equidistant points are used, with point number $4,10,11,14$ and $17$ used to evaluate the high fidelity model and the others used for the low fidelity model. The choice of kernel hyperparameters is made by maximising the marginal likelihood (often called empirical Bayes). Further details, and an additional test function can be found in Appendix \ref{sec:appendix_multifidelity}.
Note that both of these problems are challenging for several reasons. Firstly, due to their discontinuity, the integrands are not in the RKHS $\mathcal{H}_{\bm{C}}$ corresponding to the kernel $\bm{C}$ used in multi-output BQ. In particular, the problems are misspecified in the sense that the true function is not in the support of the prior. It is therefore difficult to interpret the posterior distribution on $\Pi[\bm{f}]$, and we end up with credible intervals which are too wide. This is for example illustrated in the values of the posterior variance for the high-fidelity Forrester function. Secondly, in each case, the high and low-fidelity models are defined on different scales and so require tuning of several kernel hyper-parameters. This can of course make it challenging for multi-output BQ since the number of function evaluations $N$ is small and empirical Bayes will tend to be inefficient in those cases.
However, despite these two issues, it is interesting to note that both of the multi-output BQ methods manage to significantly outperform uni-output BQ in terms of point estimate, as the sharing of information allows the multi-output models to better represent the main trends in the functions. Furthermore, the multi-output BQ does not suffer from the issues of overconfident posterior credible intervals present in uni-output BQ; contrast for example the posterior variances for the high-fidelity step function.
\paragraph{Global illumination}
In this section, we apply multi-output BQ to a challenging numerical integration problem from the field of computer graphics, known as global illumination. BQ was previously applied to this problem in several papers \cite{Brouillat2009,Marques2013,Briol2015a}, but we propose to extend these results using multi-output BQ.
Global illumination is a problem which occurs when trying to obtain realistic representation of light interactions for the design of virtual environments (e.g. a video game). One model of the amount of light coming from an object towards the camera (representing the current viewpoint on this environment) is given by the following equation:
\begin{eqnarray*}
L_0(\omega_0) & = & L_e(\omega_0) + \int_{\mathbb{S}^2} L_i(\omega_i) \rho(\omega_i,\omega_0) [\omega_i \cdot n]_{+} \mathrm{d}\Pi(\omega_i).
\end{eqnarray*}
where $[x]_+ = \text{max}(0,x)$. The function $L_0:\mathbb{S}^2 \rightarrow \mathbb{R}$ evaluated at $\omega_0$ is called the outgoing radiance in direction $\omega_0$ (the angle of the outgoing light from the object normal $n$), $L_e(\omega_0):\mathbb{S}^2 \rightarrow \mathbb{R}$ is the amount of light emitted by the object, and $L_i:\mathbb{S}^2 \rightarrow \mathbb{R}$ evaluated at $\omega_i$ is the amount of light reflected by the object (which originated from an angle $\omega_i$ from the object's normal $n$). Here, $\mathbb{S}^2 = \{\bm{x} = (x_1,x_2,x_3) \in \mathbb{R}^3 : \|\bm{x}\|_{2} = 1\}$ and $\rho(\omega_i,\omega_0):\mathbb{S}^2 \times \mathbb{S}^2 \rightarrow \mathbb{R}$ is called the bidirectional reflectance distribution and represents the proportion of light being reflected.
We follow \cite{Briol2015a} and consider the problem as $\Pi[h^{\omega_0}] = \int_{\mathbb{S}^2} h^{\omega_0}(\omega_i) \Pi(\mathrm{d}\omega_i)$ where $\Pi$ is the uniform measure on $\mathbb{S}^2$, and $h^{\omega_0}(\omega_i) = L_i(\omega_i)\rho(\omega_i,\omega_0)[\omega_i\cdot \omega_0]_{+}$ is a function which can be evaluated by making a call to an environment map (which we consider to be a black box). One scenario which is common in these type of problems is to look at an object from different angles $\omega_0$, with the camera moving. In this case, it is reasonable to assume that the different integrands $h^{\omega_0}$ will be very similar when the difference in the angle $\omega_0$ is small, and it is therefore natural to consider jointly estimating their integrals. In the experiments we consider five integrands $f_i = h^{\omega_0^i}$ for $i=1,\ldots,5$ where $\omega_0^1,\ldots,\omega_0^5$ are on a great circle of the sphere at intervals determined by an angle of $0.005 \pi$.
We therefore consider two-output and five-output BQ with independent and identically distributed (Monte Carlo) samples $\bm{X}$ from the uniform measure $\Pi$. We propose to use a separable kernel with scalar-valued RKHS $\mathcal{H}_c$ being a Sobolev space of smoothness $\frac{3}{2}$ over $\mathbb{S}^2$ and has kernel $c(\bm{x},\bm{x}') = \frac{8}{3}-\|\bm{x}-\bm{x}'\|_2^2$. For the matrix $B$ representing the covariance between outputs, we propose to make this covariance proportional to the difference in angle at which the camera looks at the object. In particular we choose $(B)_{ij} = \exp(\omega_0^i \cdot \omega_0^j -1)$ for simplicity, but this could be generalised to include a lengthscale and amplitude hyperparameter to be learnt together with the hyperparameters of the scalar-valued kernel $c$.
\begin{figure}[t!]
\begin{center}
\includegraphics[trim={2cm 0 2cm 0},clip,width=0.23\textwidth]{Figures/f1.png}
\includegraphics[trim={2cm 0 2cm 0},clip,width=0.23\textwidth]{Figures/f2.png} \\
\includegraphics[trim={2cm 0 2cm 0},clip,width=0.23\textwidth]{Figures/f1_uni.png}
\includegraphics[trim={2cm 0 2cm 0},clip,width=0.23\textwidth]{Figures/f2_uni.png}\\
\includegraphics[trim={2cm 0 2cm 0},clip,width=0.23\textwidth]{Figures/f1_icm.png}
\includegraphics[trim={2cm 0 2cm 0},clip,width=0.23\textwidth]{Figures/f2_icm.png}\\
\caption{Global illumination: Plot of $f_1$ (left) and $f_2$ (right) in the case of the red channel. The plots give the true functions being approximated (top), the uni-output approximations (middle) and the five-output approximations with LMC kernel (bottom).
\vspace{-3mm}}
\label{fig:sphere_plots}
\end{center}
\end{figure}
The GP means for the one-output and five-output cases are given in Fig. \ref{fig:sphere_plots}, and we can clearly notice a significant improvement in approximation accuracy with the larger number of outputs.
Results for integration error are given in Fig. \ref{plot:global_illum_error}. As noticed, the integration error (for a fixed number of evaluations $N$ of each integrand) is significantly reduced by increasing the number of outputs $D$. The individual posterior variances for this problem (see Appendix \ref{sec:appendix_global_illum} Fig. \ref{plot:global_illum_WCE}) are also smaller, reflecting the fact that our uncertainty is reduced due to use of observations from other integrands.
In fact, a small extension of Thm. \ref{theorem:convergence_separable} (combined with the rate for the scalar-valued kernel in \cite{Briol2015a}) allows us to obtain an asymptotic convergence rate for the posterior variance on each integral $\Pi[f_d]$:
\begin{corollary}
Let $\mathcal{X}$ be the sphere $\mathbb{S}^2$ and $\bm{X}$ be IID uniform points on $\mathcal{X}$. Assume $\bm{C}$ is a separable kernel with $c$ defined above. Then $e(\mathcal{H}_{\bm{C}},\hat{\Pi}_{\text{BQ}},\bm{X},d) = \mathcal{O}_P \left(N^{-\frac{3}{4}}\right)$.
\end{corollary}
The same rate with improved rate constant was observed in \citep{Briol2015a} when using QMC point sets, and similar gains could be obtained in this multi-output case.
\begin{figure}[t!]
\begin{center}
\includegraphics[trim={2cm 0 1cm 0},clip,width=0.495\textwidth]{Figures/Figure4.png}
\caption{Global illumination: Plot of error estimates for $f_1$ (top) and $f_2$ (bottom), in the case of the red, green and blue channels. The log-error is plotted for uni-output BQ (red), two-output BQ based on LMC (blue), five-output BQ based on LMC (magenta) and standard Monte Carlo (dotted black).
\vspace{-3mm}}
\label{plot:global_illum_error}
\end{center}
\end{figure}
We note that there a significant potential further gains for the use of multi-output BQ in this setting. Similar integration problems need to be computed for three colors in every pixel of an image, and for every image in a video. This is challenging computationally and limits the use of Monte Carlo methods to a few dozen points. Designing specific matrix-valued kernels could provide enormous gains since we end up with thousands of correlated integrands. Furthermore, the weights only depend on the choice of kernel and not on function values, so that all of the weights could be pre-computed off-line to be later used in real-time.
\section{Conclusion}
\label{sec:conclusion}
We have proposed an extension of Bayesian Quadrature to the case where we are interested in numerically computing the integral of several functions which are related. In particular we have proposed a new algorithm based on jointly modelling the integrands with a Gaussian prior. Then, we provided a theoretical study of the rate of convergence for the case where the kernel is separable and illustrated the potential of our methodology on applications in multi-fidelity modelling and computer graphics. Our main contribution however, has been to highlight the natural extension of Bayesian probabilistic numerical methods to the joint estimation of the solution of several numerical problems (in this case, numerical integration problems).
There are several possible extensions of multi-output BQ which we reserve for future work. One important question remaining is that of the choice of sampling distribution. In the uni-output case, it is well known that obtaining an optimal sampling distribution with respect to the $\mathbb{V}_n[\Pi[f]]$ is intractable in most cases. \cite{Briol2017} proposed an algorithm to approach such a distribution, and \cite{Kanagawa2017} provided conditions on the point sets to guarantee fast convergence. In the multi-output case, the problem is even more complex due to the interaction between the different integration problems. However, the literature on the design of experiments for co-kriging/multi-output GPs may be of interest, and the use of more advanced sampling distributions will certainly provide significant gains.
\subsection*{Acknowledgements}
The authors are grateful to Alessandro Barp, Aretha Teckentrup, Chris Oates and Motonobu Kanagawa for helpful discussions. FXB was supported by the EPSRC grants [EP/L016710/1, EP/R018413/1, EP/N510129/1]. MG was supported by the EPSRC grants [EP/J016934/3, EP/K034154/1, EP/P020720/1, EP/R018413/1, EP/N510129/1], an EPSRC Established Career Fellowship, the EU grant [EU/259348] and the Lloyds Register Foundation Programme on Data-Centric Engineering. The authors would like to thank the Isaac Newton Institute for Mathematical Sciences for support and hospitality during the programme on ``Uncertainty Quantification for Complex Systems: Theory and Methodologies". This work was supported by EPSRC grant no [EP/K032208/1]. Finally, this material was also based upon work partially supported by the National Science Foundation under Grant DMS-1127914 to the Statistical and Applied Mathematical Sciences Institute.
|
\section{INTRODUCTION}
There is a fast-growing interest in using first-principles computational methods to interpret x-ray spectroscopies for characterizations of materials and thereby enhance our basic understanding of electronic structure \cite{haerle2001sp, nilsson2004chemical, wernet2004structure, prendergast2006x, debeer2010calibration, woicik2007ferroelectric, rehr2009ab, zhao2011visualizing, tan2012unraveling, liu2012phase, drisdell2013probing, eelbo2013adatoms, t2013compensation, velasco2014structure, pascal2014x, mcdonald2015cooperative, wernet2015orbital, lu2016quantitative, drisdell2017determining, liu2017highly}.
Fulfilling this task requires a reliable prediction of possible atomic structures that could lead to the observed spectra, and more challengingly, a generic theory that can predict accurate x-ray spectral fingerprints for given systems.
Central to a first-principles spectroscopic theory is solving the dynamics of a many-electron Hamiltonian upon excitation of a core electron by an x-ray photon, for realistic systems ranging from molecules to solids, in an efficacious manner.
From a fundamental viewpoint, the approaches to tackle a many-body problem fall into two major categories.
Quantum-field-theoretical methods \cite{wen2004quantum, bruus2004many, tsvelik2007quantum, mattuck2012guide} focus on describing the trajectories of a many-body system.
Through computing the path integrals of all trajectories from one many-body state to another,
one obtains the transition probability between the two.
The field-theoretical approach has given rise to a set of powerful first-principles tools such as the $GW$ and Bethe-Salpeter-Equation (BSE) method \cite{hybertsen1986electron, rohlfing2000electron, onida2002electronic, vinson2011bethe}.
In current implementations of these methods, only a finite set of diagrams are incorporated, due to the daunting complexity of evaluating all them.
The other category of approaches focuses on the description of many-body wave functions based on Slater determinants \cite{knowles1984new, szabo2012modern, shao2006advances}.
This leads to methods that are used prevalently in quantum chemistry such as the full configuration interaction (FCI) approach and the coupled-cluster technique \cite{pople1987quadratic, bartlett2007coupled, booth2013towards},
or exact diagonalization for solving strongly-correlated systems \cite{caffarel1994exact, dagotto1994correlated}.
Currently these methods are mostly applied to systems with $10-20$ electrons limited by the exponential growth of the configuration space.
For x-ray excitations and associated spectra, we have witnessed the success of the constrained-occupancy density functional theory ($\Delta$SCF)
\cite{taillefumier2002x, prendergast2006x, drisdell2013probing, pascal2014x, velasco2014structure, ostrom2015probing, drisdell2017determining},
which approximates an x-ray excited state with one empty Kohn-Sham (KS) orbital in the final state.
Recently, we highlighted the shortcomings in this single-particle (1p) approach for a class of $3d$ transition metal oxides (TMOs)
and indicated an lack of generation to higher-order excitations involving multiple electron-hole ($\emph{e}$-$\emph{h}${}) pairs \cite{liang2017accurate}.
Driven by these deficiencies, we proposed a better many-body wavefunction ansatz
that approximates the initial and final state with a \emph{single} Slater determinant.
The initial-state Slater determinant is constructed from the KS orbitals of the ground-state system,
while the Slater determinant for a specific final-state is derived from the KS orbitals of the core-excited system.
Within this approximation,
the transition amplitude can also be expressed as a determinant \cite{anderson1967infrared, dow1980solution, stern1983many, ohtaka1990theory, liang2017accurate} comprising transformation coefficients between the two KS basis sets.
We find this determinant approach can rectify the deficiency of the 1p $\Delta$SCF approach for a few TMOs \cite{liang2017accurate}.
It is natural to ask:
(a) does this formalism provide a good approximation for x-ray near-edge structures in general?
(b) is it practicable for calculations of extended systems, given the huge configuration space?
(c) can this approach permit access to higher-order excitations and describe various many-body x-ray spectral features beyond the BSE?
In this work, we answer these questions by demonstrating an efficient yet simple approach to explore the large configuration space in the determinant formalism.
A crucial first step is to relate similar determinants to one another via exterior algebra \cite{yokonuma1992tensor, winitzki2010linear}
and then evaluate them via updates, rather than from scratch.
Even so there are still $10^6$ to $10^9$ many-body states to consider for configurations with double $\emph{e}$-$\emph{h}${} pairs .
However, only a small portion of these determinants have significant transition amplitudes, due to the spatially localized nature of core-level excitations,
as can be tested by brute-force calculations.
Motivated by this observation, we adopt a breadth-first search (BFS) algorithm \cite{knuth1998art, cormen2009introduction} to look for nontrivial configurations rather than exhausting the entire configuration space.
The BFS algorithm is a basic algorithm for traversing a tree structure, finding the shortest path \cite{skiena1990dijkstra, zeng2009finding}, solving a maze \cite{moore1959shortest}, and other combinatorial search problems.
Although the BFS algorithm cannot guarantee answers within a polynomial time, substantial speed-up can often be achieved via heuristically pruning the search tree \cite{pearl1984heuristics, zhou2006breadth}.
For the many-body configuration problem, we design the BFS to search for active \lq\lq pathways\rq\rq{} from the initial state to many excited-state configurations.
Instead of directly accessing a large number of high-order configurations,
the search algorithm first visits its ascendant configurations with fewer $\emph{e}$-$\emph{h}${} pairs.
If multiple pathways to an ascendant configuration interfere destructively and result in a small transition amplitude,
the search algorithm will discard the configuration before more high-order configurations are generated.
We will show that this tree-pruning technique can typically lead to at least 100-fold speed-up in the calculation of x-ray spectra.
This search algorithm is generic and can be generalized to any kind of sudden perturbation.
The determinant formalism is an exact solution to the Mahan-Nozi\'eres-De Dominicis (MND) model \cite{mahan1967excitons, nozieres1969singularities} in which multiple electrons interact with a core hole.
Hence, this approach can naturally incorporate all many-electron processes in the MND theory,
which includes
the direct and exchange diagrams as in the BSE \cite{rohlfing2000electron, onida2002electronic, vinson2011bethe},
the zig-zag diagrams,
and the diagrams with a core hole dressed by many $\emph{e}$-$\emph{h}${} bubbles.
While the BSE diagrams mainly describe $\emph{e}$-$\emph{h}${} attraction, or excitonic effects,
the zigzag or bubble diagrams describe higher order $\emph{e}$-$\emph{h}${} excitations
that lead to shakeup features \cite{brisk1975shake, mcintyre1975x, ohtaka1990theory, enkvist1993c, calandra2012k, mahan2013many, lemell2015real}
or many-body effects due to reduced wave-function overlap.
A reduction in many-body wave function overlap is the origin of the Anderson orthogonality catastrophe \cite{anderson1967infrared, mahan2013many}.
If one were to include all of these effects using the diagrammatic approach, a comprehensive set of techniques, such as solving BSE-like equations and using a cumulant expansion \cite{kas2015real, kas2016particle}, would be required.
Here, the determinant formalism, in conjunction with the first-principles KS orbitals, provides a efficient means to investigate all many-electron effects within the MND model rigorously, for a wide energy range, within a simple unified framework.
This new determinant formalism has already shown great practicality to address realistic problems in materials characterization.
We systematically study the \ce{O} $K$-edge ($1s\rightarrow np$) x-ray absorption spectra (XAS) of various TMOs
and find this approach can faithfully reproduce the experimental x-ray line shapes for most of the investigated systems.
This can be immediately applied to study various energy conversion and storage systems involving oxides \cite{yabuuchi2011detailed, hu2013origin, suntivich2011design, lin2016metal, luo2016charge, strasser2010lattice, matsukawa2014enhancing, lebens2016direct, de2016mapping},
where the interpretation of x-ray spectra can be challenging, and the conclusions often depend sensitively on intricate near-edge line shapes.
This work is organized as follows.
Sec. \ref{sec:model} revisits the many-body effects captured by the MND theory in terms of Feynman diagrams.
Sec. \ref{sec:det} and \ref{sec:comparison} provides a solution to the MND model from the perspective of many-electron wave functions and introduces the determinant formalism.
Sec. \ref{sec:sol} introduces exterior algebra to elucidate the linear dependence of the determinants that is encoded in the so-called $\zeta$-matrix,
followed by a BFS algorithm for an efficient evaluation in Sec. \ref{sec:bfs}.
Sec. \ref{sec:dscf} discusses how to combine this algorithm with DFT simulations and its validity in the presence of $\emph{e}$-$\emph{e}${} interactions.
The simulated XAS of a variety of oxides are shown in Sec. \ref{SEC:results},
with analysis of spectra obtained from different level of approximations.
The many-body aspects beyond the BSE as captured by this method will be discussed in Sec. \ref{sec:more_eh} and \ref{sec:spin_conv}, using the half-metal \ce{CrO2} as an example.
Finally, the numerical details and efficiency of the newly introduced algorithm are analyzed in Sec. \ref{sec:comp}.
\section{THEORETICAL MODELS AND METHODS}
\subsection{Independent-electron model and diagrammatic approaches}
\label{sec:model}
We first revisit the conceptually simple MND model from the perspective of Feynman diagrams.
The incorporation of first-principles calculations will be deferred to Sec. \ref{sec:dscf}.
In the MND model \cite{nozieres1969singularities, ohtaka1990theory, mahan2013many}, the electrons only interact with the core hole and electron-electron ($\emph{e}$-$\emph{e}${}) interactions are neglected.
Consider a supercell with one of the atoms replaced by its core-excited version.
This is typically a good approximation to a core-excited system at low photon flux.
Assume there are $N$ valence electrons in its ground state and there is only one core level.
The MND Hamiltonian without $\emph{e}$-$\emph{e}${} interactions reads
\begin{align}
\begin{split}
\mathcal{H}&=\mathcal{H}_0+\mathcal{H}_I\\
\mathcal{H}_0&=
\sum_{c}\varepsilon_{c} a^\dagger_{c} a_{c}
- \varepsilon_h h^\dagger h\\
\mathcal{H}_I&=
\sum_{cc'} V_{cc'} a^\dagger_{c} a_{c'}
h^\dagger h\\
\end{split}
\label{eq:ham_mnd}
\end{align}
where the diagonal part $\mathcal{H}_0$ is composed of the valence orbitals ($c$ iterates over both occupied and empty valence orbitals) and the core level ($h$).
$a^\dagger_c$ and $h^\dagger$ are electron and hole creation operators respectively.
The only two-body term in $\mathcal{H}$ is the Coulomb interaction between the valence orbitals and the core level, as described by $\mathcal{H}_I$,
in which the core-hole potential $V_{\alpha\beta}$ is defined by
\begin{align}
V_{\alpha\beta}&=\int d^3 r d^3 r' \psi^*_\alpha(\textbf{r})\psi_\beta(\textbf{r})V(\textbf{r}, \textbf{r}')
\psi^*_h(\textbf{r}')\psi_h(\textbf{r}')
\label{eq:v}
\end{align}
where $\psi_i$'s are the 1p wave functions and $V(\textbf{r}, \textbf{r}')$ is the (effective) Coulomb potential.
The two-body interaction $V_{\alpha\beta}$ accounts for the electron scattering from orbital $\beta$ to $\alpha$ due to the core-hole potential.
The x-ray photon field can be described by a current operator \cite{mahan2013many} that promotes one core electron to a valence orbtial
\begin{align}
\begin{split}
\hat{J}=\sum_{c} a^\dagger_c h^\dagger\langle\psi_c|\hat{j}|\psi_h\rangle + h.c.
\label{eq:J}
\end{split}
\end{align}
The transition operator is the electric field polarization-projected position operator that couples the core level to valence orbitals: $\hat{j} = \bm{\epsilon}\cdot\bm{r}$, in the limit of zero-momentum transfer and within the dipole approximation \cite{prendergast2006x, de2008core}.
In principle, the transition operator $\hat{j}$ can be any other local sudden perturbation, not necessarily limited to a core hole.
The independent-electron model was originally considered by the MND theory \cite{mahan1967excitons, nozieres1969singularities, mahan2013many} using diagrammatic techniques.
The time-evolution of the many-electron system after photon absorption is described by the Kubo current-current correlation function
\begin{align}
\begin{split}
\Pi(t) &= -\frac{i}{\mathcal{V}} \langle \Psi_i| \mathcal{T} \hat{J}(t) \hat{J}(0)] \Psi_i\rangle\\
&=\frac{1}{\mathcal{V}}\sum_{cc'} \langle\psi_c|\hat{j}|\psi_h\rangle \langle\psi_h|\hat{j}|\psi_{c'} \rangle L_{cc'}(t)\\
&=\frac{1}{\mathcal{V}} \sum_{cc'} w_c w^*_{c'} L_{cc'}(t)
\end{split}
\label{eq:pi_t}
\end{align}
where $w_c = \langle\psi_c|\hat{j}|\psi_h\rangle$ is the vertex that represents the absorption of a photon to create an $\emph{e}$-$\emph{h}${} pair ($w^*_c$ represents the opposite process).
The x-ray absorption spectrum (XAS) $A(\omega)$ is the spectral function of the photon self-energy in the frequency domain
\begin{align}
\begin{split}
\Pi(\omega) &= \int^\infty_{-\infty} dt e^{i\omega t} \Pi(t)\\
A(\omega) & = -\frac{1}{\pi}\text{Im} \Pi(\omega)
\end{split}
\end{align}
In the following discussion, we focus on the $\emph{e}$-$\emph{h}${} correlation function as defined in Eq. (\ref{eq:pi_t})
\begin{align}
\begin{split}
L_{cc'}(t) = -i \langle \Psi_i|\mathcal{T} h(t) a_c(t) a^\dagger_c(0) h^\dagger(0) |\Psi_i\rangle
\end{split}
\end{align}
which includes all the many-electron processes in x-ray absorption.
\begin{figure}[t]
\centering
\includegraphics[angle=0, width=0.90\linewidth]{./Lcc_2nd.png}
\caption{
Four distinct types of $\emph{e}$-$\emph{h}${} processes in the second-order Feynman diagrams in the MND theory.
There are exactly two Coulomb lines (at $t_1$ and $t_2$) in each diagram, as marked by vertical dashed lines.
}
\label{fig:lcc_2nd}
\end{figure}
We exemplify these many-electron processes by four types of second-order Feynman diagram of $L_{cc'}(t)$, as shown in Fig. \ref{fig:lcc_2nd}.
The time axis runs from left to right and the Coulomb lines are vertical due to the neglect of dynamical effects in the Coulomb interaction $\mathcal{H}_I$.
The BSE captures two kinds of processes: direct $\emph{e}$-$\emph{h}${} attraction as described by the ladder diagram Fig. \ref{fig:lcc_2nd} (a), and $\emph{e}$-$\emph{h}${} exchange as described by the diagram in Fig. \ref{fig:lcc_2nd} (b).
In these diagrams, there is only one $\emph{e}$-$\emph{h}${} pair present at any time of the propagation.
However, there are other diagrams with more $\emph{e}$-$\emph{h}${} pairs present at a time, e.g., the zigzag diagram in Fig. \ref{fig:lcc_2nd} (c).
The corresponding process involves a core hole causing the ground state to decay into a valence $\emph{e}$-$\emph{h}${} pair ($c'$ and $v$) at $t_1$.
At a later time $t_2$, the core hole assists the newly generated valence hole ($v$) to recombine with incoming electron ($c$), leaving an outgoing electron ($c'$) and the core hole.
Lastly, it is also possible that the valence $\emph{e}$-$\emph{h}${} pair ($c'$ and $v$) generated earlier does not correlate with the incoming electron at all and simply annihilates at a later time $t_2$.
This leads to a bubble diagram with a freely propagating electron and a core hole dressed by $\emph{e}$-$\emph{h}${} bubbles as shown in Fig. \ref{fig:lcc_2nd} (d).
These $\emph{e}$-$\emph{h}${} bubbles tend to reduce the many-body wave function overlap and are the causes for the Anderson orthogonality catastrophe \cite{anderson1967infrared}.
The MND theory \cite{mahan1967excitons, nozieres1969singularities, mahan2013many} systematically studies and estimates the impact of these diagrams on the near-edge structure of x-ray spectra.
In essence, it is found that denominators in the BSE diagrams involve $\varepsilon_{c} - \varepsilon_{h}$, which is roughly the energy required to create an electron-core-hole excitation, while the denominators in the zigzag or bubble diagrams involve an offset of $\varepsilon_c - \varepsilon_v$, the energy required to create an additional (valence) $\emph{e}$-$\emph{h}${} pair.
This means the zigzag processes or the bubble diagrams can become significant in a metallic system where $\varepsilon_c - \varepsilon_v$ can be vanishingly small, or if the photon energy is sufficiently high to be in resonance with double $\emph{e}$-$\emph{h}${} excitations.
In practical first-principles BSE calculations however, $\emph{e}$-$\emph{e}${} interactions are taken into account
and the bare Coulomb interactions in the direct (Fig. \ref{fig:lcc_2nd} (a) )
diagram are replaced by screened Coulomb interactions.
The screened Coulomb interactions are typically modeled with the empty-bubble diagrams within the random-phase approximation \cite{rohlfing2000electron, onida2002electronic, vinson2011bethe},
which, to some extent, describe the many-electron screening effects in x-ray excitations.
\subsection{An alternative MND solution based on many-body wave functions}
\label{sec:det}
In the last section we have discussed the diagrammatic approach, or many-body perturbation theory (MBPT), for solving the MND model [Eq. (\ref{eq:ham_mnd})].
However, this Hamiltonian is essentially quadratic and exactly solvable.
For the initial state, no core hole is excited and $\langle h^\dagger h \rangle = 0$ and hence the initial-state Hamiltonian $\mathcal{H}_i$ is simply $\mathcal{H}_0$
For the final state, there is exactly one core hole, i.e., $\langle h^\dagger h \rangle = 1$, and the final-state Hamiltonian $\mathcal{H}_f$ also becomes quadratic
\begin{align}
\begin{split}
\mathcal{H}_f& = \mathcal{H}_i + \sum_{cc'}V_{cc'} a^\dagger_{c} a_{c'}
\end{split}
\label{eq:hi_hf}
\end{align}
Within the quadratic forms, it is straightforward to construct the many-body wave functions of the initial- and final-state.
The initial state is simply a Slater determinant that consists of $N$ valence electrons occupying the $N$ lowest-lying orbitals:
\begin{align}
\begin{split}
|\Psi_i\rangle = \big( \prod^N_{\mu=1} a^\dagger_\mu \big) h |0\rangle
\end{split}
\label{eq:psi_i}
\end{align}
where $\mu$ goes over all the occupied valence orbitals, $h$ annihilates the core hole (fills the core level with one electron), and $|0\rangle$ is the null state with no electrons.
The final-state XAS wave functions can be expressed in a similar manner, but using
the eigenvectors of $\mathcal{H}_f$
\begin{align}
\begin{split}
|\Psi_f\rangle = \prod^{N+1}_{\mu=1} \tilde{a}^\dagger_{f_{\mu}} |0\rangle
\end{split}
\label{eq:psi_f0}
\end{align}
where the index $f$ is a tuple: $f=(f_1, f_2, \dots, f_{N+1})$, which denotes the \emph{valence} $N+1$ orbitals that the $N+1$ electrons will occupy in the final state.
$\tilde{a}_i$ (with tilde) correspond to the eigenvectors of $\mathcal{H}_f $ so that $\mathcal{H}_f = \sum_i \tilde{\varepsilon}_i \tilde{a}^\dagger_i \tilde{a}_i$.
To apply the Fermi's Golden rule, one needs to work within the same basis set.
We express the final-state basis set in terms of the initial-state one:
\begin{align}
\begin{split}
|\tilde{\psi}_i \rangle &= \sum_j \xi_{ij} | \psi_j \rangle\\
\tilde{a}^\dagger_i &= \sum_j \xi_{ij} a^\dagger_j
\end{split}
\label{eq:xi}
\end{align}
where $\xi_{ij}$'s are the transformation coefficients: $\xi_{ij} = \langle \psi_j | \tilde{\psi}_i \rangle$.
With these expressions for $|\Psi_i\rangle$ and $|\Psi_f\rangle$, the many-body transition matrix element for any one-body operator $\mathcal{O}$ has been calculated in previous work \cite{stern1983many, ohtaka1990theory, liang2017accurate}
\begin{align}
\begin{split}
\langle \Psi_f | \mathcal{O} | \Psi_i \rangle = \sum_{c} (A^{f}_c)^* \langle \psi_c | o | \psi_h \rangle
\end{split}
\label{eq:mat_elem}
\end{align}
in which the transition amplitude also takes a determinantal form
\begin{align}
\begin{split}
A^{f}_c=
\det
\begin{bmatrix}
\xi_{f_1, 1} & \xi_{f_1, 2} & \cdots & \xi_{f_1, N} & \xi_{f_1, c} \\
\xi_{f_2, 1} & \xi_{f_2, 2} & \cdots & \xi_{f_2, N} & \xi_{f_2, c} \\
\vdots & & \ddots & & \vdots\\
\xi_{f_{N+1}, 1} & \xi_{f_{N+1}, 2} & \cdots & \xi_{f_{N+1}, N} & \xi_{f_{N+1}, c} \\
\end{bmatrix}
\end{split}
\label{eq:afc}
\end{align}
The row index goes over $N+1$ occupied final-state orbitals $f_i$,
the column index over the lowest-lying $N$ initial-state orbitals plus one \emph{empty} orbital labeled by $c$
(This empty orbital is coupled to the core level with the one-body operator $\mathcal{O}$).
This determinantal form reflects how these $N+1$ electrons transit from the initial to final state in the x-ray excitation process.
All the possible electronic pathways are taken into account by the transformation matrix in $A^f_c$.
The transition amplitude of an individual electron is quantified by the matrix elements, i.e., the initial-final orbital overlap $\xi_{ij} = \langle \psi_j | \tilde{\psi}_i \rangle$.
The interference of these pathways is lumped into a determinant due to the fermionic nature of electrons.
For the quadratic $\mathcal{H}_f$, the energy of a final-state $|\Psi_f\rangle$ can be obtained
by direct summation of 1p-orbital energies
\begin{align}
\begin{split}
E_f=\sum^{N+1}_{j=1}\tilde{\varepsilon}_{f_j}
\label{eq:Ef}
\end{split}
\end{align}
where $\tilde{\varepsilon}_{f_j}$ are taken from the diagonalized $\mathcal{H}_f$.
A relative energy $\Omega_f = E_f - E_\text{th}$ may also be defined for later discussion,
where $E_\text{th}$ is the energy of the lowest-lying $|\Psi_f\rangle$:
$E_\text{th}=\sum^{N+1}_{j=1}\tilde{\varepsilon}_{j} $.
For ease of calculation, previously we have also regrouped the final-state multi-electron configurations according to the convention in quantum chemistry \cite{dow1980solution, szabo2012modern, bartlett2007coupled}.
The configuration $f = (1, 2, \cdots, N, c_0)$ with $c_0 > N$ is dubbed as a \emph{single} or a $f^{(1)}$ configuration
because it has one electron-(core-)hole pair.
A shorthand notation for an $f^{(1)}$ configuration can be employed, using $(c)$ to denote the the orbital of the excited valence electron.
$f=(1,2,\cdots, v_1 - 1, v_1 + 1, \cdots, N, c_0, c_1)$ with $v_1 \leq N$ and $c_1 > c$ is dubbed as a \emph{double} or $f^{(2)}$ configuration because it has one extra (valence) $\emph{e}$-$\emph{h}${} pair
as defined by the electron (hole) index $c_1$ ($v_1$).
The shorthand notation for $f^{(2)}$ is $(c_0, v_1, c_1)$.
This definition can be extended to higher orders such as triples and so forth.
For unique indexing, we require $c_0 < c_1 < c_2 < \cdots < c_{n - 1}$ and $v_1 > v_2 > \cdots > v_{n - 1}$ in a $f^{(n)}$ index.
Examples of final-state $f^{(n)}$ are shown in Fig. \ref{fig:fn} (schematics on the second row).
\begin{figure}[t]
\centering
\includegraphics[angle=0, width=0.90\linewidth]{./fn.png}
\caption{
Definitions of the multi-electron configurations used in the initial(i)- and final(f)-state picture according to the convention in quantum chemistry.
A final-state configuration (a single Slater determinant) at the order of $f^{(n)}$ can be hybridized from a number of initial-state configurations at multiple orders,
as shown by the thick opaque downarrows, which illustrates the spirit of Eq. (\ref{eq:psi_f0}).
The solid uparrows in a configuration indicate one possible multi-electronic pathway to access that configuration from the ground-state.
The dashed uparrows show the other possible pathway to access the $f^{(2)}$ configuration.
}
\label{fig:fn}
\end{figure}
\subsection{Interpretation of the final-state many-body approach from an initial-state perspective}
\label{sec:comparison}
In this section, we provide a comparison between the outlined determinant formalism and MBPT using Feynman diagrams.
While the determinant formalism constructs many-electron states using both initial- and final-state orbitals, MBPT, such as BSE, relies on initial-state quantities only.
To relate the two theories, we can express the MND many-electron final states
$| \Psi_f \rangle$ in Eq. (\ref{eq:psi_f0}) using only the initial-state orbitals.
Rewriting final-state operators $\tilde{a}_i$ according to a linear combination of the initial-state operators $a_i$ [Eq. (\ref{eq:xi})] and expressing the wave function $|\Psi_f\rangle$ in terms of $|\Psi_i\rangle$:
\begin{align}
\begin{split}
|\Psi_f\rangle &= \prod^{N+1}_{\mu=1} \sum_{j_\mu} \xi_{f_\mu, j_\mu} a^\dagger_{j_\mu} |0\rangle \\
&= \prod^{N+1}_{\mu=1} ( \sum_{j_\mu} \xi_{f_\mu, j_\mu} a^\dagger_{j_\mu} ) \big( \prod^N_{\nu=1} a_\nu \big) h^\dagger |\Psi_i\rangle
\end{split}
\label{eq:psi_f_all}
\end{align}
Expanding the product of the operators and regrouping like terms,
\begin{align}
\begin{split}
|\Psi_f\rangle
&= \sum_{c\in\text{unocc}} A^f_c a^\dagger_c h^\dagger |\Psi_i\rangle \\
&+ \sum_{\substack{c, c'\in\text{unocc} \\ v \in \text{occ}}} B^f_{cc',v} (a^\dagger_c h^\dagger) (a^\dagger_{c'} a_v) |\Psi_i\rangle \\
&+ \cdots
\end{split}
\label{eq:psi_f}
\end{align}
The leading-order term comprises linear combinations of single electron-(core-)hole pairs, because there are $N+1$ creation operators $a_i^\dagger$ and $N$ destruction operators $a_i$ in Eq. (\ref{eq:psi_f_all}),
leaving at least one creation operator $a^\dagger_c$ for an unoccupied state.
For this term, $N$ out of $N+1$ indices $j_\mu$ are chosen from $1,2,\cdots, N$ so that $N$ $a^\dagger_i$'s can cancel with $N$ $a_i$'s.
There are $(N+1)!$ such permutations, and reordering the fermionic operators gives rise to the determinantal form of the coefficients, as previously stated in Eq. (\ref{eq:afc}).
The next term in Eq. (\ref{eq:psi_f}) is a double term $(a^\dagger_c h^\dagger) (a^\dagger_{c'} a_v)$,
which has one additional valence \emph{e-h} pair $(a^\dagger_{c'} a_v)$ generated on the top of the electron-core-hole pair.
This term takes into account the second-order many-electron processes: the valence \emph{e-h} excitations induced by the core-hole potential, which are also known as the \emph{shakeup} excitations \cite{brisk1975shake, mcintyre1975x, ohtaka1990theory, enkvist1993c, calandra2012k, mahan2013many, lemell2015real},
because an additional amount of energy is required to create these valence excitations.
As the series expansion proceeds, each term will have one more valence $\emph{e-h}$-pair than the last,
and more complicated shakeup processes with multiple \emph{e-h}-pairs are included.
A full schematic for the relation of one single final-state configuration $|\Psi_f\rangle$ (written as one Slater determinant using final-state orbitals) in terms of initial-state configurations is shown in Fig. \ref{fig:fn}.
Within MBPT, the configuration series in Eq. (\ref{eq:psi_f}) is typically truncated,
and the coefficients are solved by expanding the Hamiltonian over the restricted configuration space and solving the eigenvalue problem.
In the BSE, for instance, the final-state Hamiltonian is expanded over the single-\emph{e-h}-pair space $a^\dagger_c h^\dagger |\Psi_i\rangle$
and the eigenvector coefficients (analogous to $A^f_c$) refer to this single-$\emph{e}$-$\emph{h}${} basis.
In some sense, this approximation corresponds to the ladder and exchange diagrams: at any point in time of the propagation, there is only one \emph{e-h} pair involved.
By contrast, the determinant formalism does not restrict the number of \emph{e-h}-pairs in the final-state configuration space.
When $|\Psi_f\rangle$ is projected onto $|\Psi_i\rangle$ as in Eq. (\ref{eq:psi_f}), a superposition of single, double, and high-order terms naturally arises,
although only the leading-order coefficients $A^f_c$ are relevant for calculating matrix elements of one-body operator.
In this way, the zig-zag and bubble diagrams, present within MND theory, which involve multiple \emph{e-h}-pair generation, are automatically incorporated.
\subsection{Efficient evaluation of determinantal transition amplitudes}
\label{sec:sol}
The above determinantal formalism provides an alternative solution to the MND model in Eq. (\ref{eq:ham_mnd})
without using diagrammatic approaches.
If a sufficient number of final states are included, one may expect the determinantal method to give the spectrum as solved from the MND model.
However, an brute-force calculation is rarely used because the many-electron configuration space grows factorially with the number of electrons.
It does not seem to be practical to compute the large number of determinants that would represent all configurations.
For a half-filled system with $M$ orbitals and $N$ ($N \approx M / 2$) electrons, even the $f^{(2)}$ group has $(M - N) (M - N - 1) N \approx M^3 / 8$ configurations.
Iterating the index $c$ of $A^f_c$ [Eq. (\ref{eq:afc})] over all empty initial-state orbitals multiplies the time complexity by a factor of $M / 2$.
Calculating the determinant for each configuration requires a computational cost of $\mathcal{O}(M^3)$.
With all the three factors combined, obtaining the determinants for all of the $f^{(2)}$ configurations gives rise to a time complexity of $\mathcal{O}(M^7)$.
For metallic systems where the fermi surfaces are susceptible to the core-hole potential, higher-order terms such as $f^{(3)}$ are typically needed for testing convergence, which leads to a higher time complexity of $\mathcal{O}(M^9)$.
Such a brute-force calculation that scales up quickly with number of states is not very practical for realistic core-hole calculations in which there could easily be $10^2$ to $10^3$ orbitals.
In this section, we introduce an efficient algorithm at much lower computational cost to access the determinants that are important for determining the x-ray spectrum.
The $\mathcal{O}(M^3)$ determinant calculation needs to be performed only once for a given configuration, and subsequently the determinants for other configurations can be derived from it.
More importantly, a BFS algorithm is employed to identify the important determinants above a specified threshold, largely reducing the number of configurations to be visited.
An apparent first step is to move the summation over $c$ in Eq. (\ref{eq:mat_elem}) into the definition of the transition amplitude coefficient, so that for each final-state configuration $f$, obtaining $A^f = \langle \Psi_f |\mathcal{O}| \Psi_i \rangle$ requires calculating only one determinant.
More specifically, we rewrite $A^f$ as
\begin{align}
\begin{split}
A^f & = \det \bm{A}^f\\
\bm{A}^f & =
\begin{bmatrix}
\xi_{f_1, 1} & \xi_{f_1, 2} & \cdots & \xi_{f_1, N} & \sum_c \xi_{f_1, c} w^*_c\\
\xi_{f_2, 1} & \xi_{f_2, 2} & \cdots & \xi_{f_2, N} & \sum_c \xi_{f_2, c} w^*_c \\
\vdots & & \ddots & & \vdots\\
\xi_{f_{N+1}, 1} & \xi_{f_{N+1}, 2} & \cdots & \xi_{f_{N+1}, N} & \sum_c \xi_{f_{N+1}, c} w^*_c \\
\end{bmatrix}
\end{split}
\label{eq:Af}
\end{align}
where $w_c = \langle \psi_c |o| \psi_h \rangle$.
The summation in the $(N+1)^{th}$ column of $\bm{A}^f$ can be calculated first before obtaining the determinant.
This reduces the overall time complexity by a factor of $M$.
Secondly, when considering transitions to various final-state configurations, the determinants of interest in fact have many common rows
so one can make use of the multilinearity of determinants to speed up the calculations significantly.
For example, the tuple for a double configuration $(1,2,\cdots, v_1 - 1, v_1 + 1, \cdots, N, c, c_1)$ only differs from the ground-state one $(1,2,\cdots, N, N + 1)$ by 3 indices,
meaning their corresponding determinants $A^f$ only differ by 3 rows.
This observation motivates us to choose the determinant for the ground state as a reference, and evaluate other determinants for excited states via
a low-rank updating technique.
To demonstrate this technique, it is most transparent to express the determinant in terms of the wedge (exterior) product \cite{yokonuma1992tensor, winitzki2010linear} of its row/column vectors.
The wedge product is anticommutative and has similar algebra to the Fermionic operators.
Suppose an arbitrary matrix $\bm{A}$ has $n$ row/column vectors $a_1, a_2, \cdots, a_n$, its determinant can be expressed as
\begin{align}
\begin{split}
\det \bm{A} = a_1 \wedge a_2 \wedge \cdots \wedge a_n
\end{split}
\end{align}
Assume $\det \bm{A}$ has been calculated from scratch and is nonzero (assume $\bm{A}$ is full-rank).
If $a_n$ is replaced by a new vector $a_{n+1}$,
which can be considered as a rank-1 update,
the updated determinant can be obtained by expanding $a_{n+1}$ in terms of $a_1, a_2, \cdots, a_n$
\begin{align}
\begin{split}
\det \bm{A}'
&\equiv a_1 \wedge a_2 \wedge \cdots \wedge a_{n+1} \\
& = a_1 \wedge a_2 \wedge \cdots \wedge \sum_{i=1}^{n} \zeta_{n+1, i} a_i \\
& = \zeta_{n + 1, n} a_1 \wedge a_2 \wedge \cdots \wedge a_n \\
& = \zeta_{n + 1, n} \det \bm{A}
\end{split}
\label{eq:Ap}
\end{align}
where $\zeta_{ij}$ is the expansion coefficient defined as
\begin{align}
\begin{split}
a_{n+1} = \sum_{i=1}^{n} \zeta_{n+1, i} a_i
\end{split}
\label{eq:zeta_intro}
\end{align}
$\zeta_{n + 1, i}$ can be obtained via the matrix inversion of $\bm{A}$: $\zeta_{n + 1, i} = \sum_{j} a_{n + 1, j}(\bm{A}^{-1})_{ji}$.
When multiplied by $a_1 \wedge a_2 \cdots \wedge a_{n-1}$, only $a_n$ survives in the summation because $a_i \wedge a_i = 0$.
Then the new determinant $\det \bm{A}'$ is simply the product of an expansion coefficient $\zeta_{n+1, n}$ and the already-known $\det \bm{A}$.
Now if the last two lines of $\bm{A}$ are replaced by two new row vectors $a_{n+1}$ and $a_{n+2}$,
the rank-2 updated determinant is
\begin{align}
\begin{split}
&\det \bm{A}'' \\
\equiv & a_1\wedge \cdots \wedge a_{n - 2} \wedge a_{n+1} \wedge a_{n+2} \\
= & a_1 \wedge \cdots \wedge a_{n - 2} \wedge \sum_{i=1}^{n} \zeta_{n+1, i} a_i \wedge \sum_{j=1}^{n} \zeta_{n+2, j} a_j \\
= & a_1 \wedge \cdots \wedge a_{n - 2} \wedge ( \zeta_{n+1, n-1} \zeta_{n+2, n} a_{n-1} \wedge a_n \\
+ & \zeta_{n+1, n} \zeta_{n+2, n-1} a_{n} \wedge a_{n-1} ) \\
= & a_1 \wedge \cdots \wedge a_{n - 2} \wedge ( \zeta_{n+1, n-1} \zeta_{n+2, n} a_{n-1} \wedge a_n \\
- & \zeta_{n+1, n} \zeta_{n+2, n-1} a_{n-1} \wedge a_n ) \\
= &\det
\begin{bmatrix}
\zeta_{n+1, n-1} & \zeta_{n+1, n} \\
\zeta_{n+2, n-1} & \zeta_{n+2, n} \\
\end{bmatrix}
\det \bm{A}
\end{split}
\label{eq:App}
\end{align}
The minus sign arises from the anticommutative property of the wedge product: $a_i \wedge a_j = - a_j \wedge a_i$.
Thus the new determinant is the product of a $2\times2$ determinant composed of the expansion coefficients and $\det A$.
The above procedure can be carried out to more general situations where more row/column vectors are replaced.
This remove the need to calculate the new determinant from scratch using the $\mathcal{O}(n^3)$ algorithm.
For a rank-$r$ update, one only needs to compute the product of the reference determinant $A^\text{ref}\equiv\det\bm{A}$ and a small $r\times r$ determinant containing $\zeta_{ij}$,
at the cost of $\mathcal{O}(1)$.
In the context of the determinantal formalism as in Eq. (\ref{eq:Af}),
we define the row vector corresponding to the $i^{th}$ final-state orbital as:
\begin{align}
\begin{split}
a_i =
\begin{bmatrix}
\xi_{i, 1} & \cdots & \xi_{i, N} & \sum_c \xi_{i, c} w^*_c
\end{bmatrix}
\end{split}
\end{align}
Then the ground-state reference determinant can be expressed as $A^\text{ref} = a_1 \wedge a_2 \wedge \cdots \wedge a_{N} \wedge a_{N + 1}$.
To access the determinants for excited states via this updating method,
we formally introduce the auxiliary $\zeta$-matrix ($\bm{\zeta}$) for a system with M orbitals and N valence electrons ($M > N$),
which is the transformation matrix from $a_1, a_2, \cdots, a_{N + 1}$ to $a_{N+1}, a_{N+2}, \cdots, a_M$:
\begin{align}
\begin{split}
\begin{bmatrix}
a_{N+1} \\ a_{N+2} \\ \vdots \\ a_M
\end{bmatrix}
=
\begin{bmatrix}
0 & 0 & \cdots & 1 \\
\zeta_{N+2, 1} & \zeta_{N+2, 2}& \cdots & \zeta_{N+2, N+1} \\
\vdots & \vdots & & \vdots \\
\zeta_{M, 1} & \zeta_{M, 2}& \cdots & \zeta_{M, N+1}
\end{bmatrix}
\begin{bmatrix}
a_{1} \\ a_{2} \\ \vdots \\ a_{N+1}
\end{bmatrix}
\end{split}
\label{eq:zeta_def}
\end{align}
Rewrite the above matrix multiplication in a compact form, we have $\bm{A}^\text{new} = \bm{\zeta} \bm{A}^\text{ref}$,
where $(\bm{\zeta})_{ij} = \zeta_{N + i, j}$.
Then $\bm{\zeta}$ can be obtained easily via matrix inversion and multiplication
\begin{align}
\begin{split}
\bm{\zeta} = \bm{A}^\text{new} (\bm{A}^\text{ref})^{-1}
\end{split}
\label{eq:zeta_cal}
\end{align}
Note that $\bm{A}^\text{ref}$ is a $(N+1)\times(N+1)$ matrix and we find it is typically invertible in practical calculations.
$\bm{\zeta}$ is of size $(M - N)\times (N + 1)$.
Its column indices map onto the lowest-lying $(N+1)$ orbitals
while the row indices map onto the $(N+1)^{th}$ to $M^{th}$ orbitals.
A $f^{(n)}$ determinant can be obtained from the product of $A^\text{ref}$,
a $n\times n$ minor of $\bm{\zeta}$,
and an overall $\pm$ sign due to permutation of rows (trivial to consider in the single-determinant case).
The last column of this $n\times n$ minor must be taken from the \emph{last} (the $(N+1)^{th}$) column of $\bm{\zeta}$,
because there are $n$ electron orbitals and $(n - 1)$ hole orbitals in a $f^{(n)}$ configuration,
and the extra one electron can be viewed as removed from the hole on the $(N+1)^{th}$ orbital.
The $n \times n$ minor reflects the interference effect of $n!$ pathways to access the $f^{(n)}$ configuration via permuting $n$ empty orbitals.
The rows (columns) of the $n\times n$ minor indicate the electrons (holes) that are excited in the given $f^{(n)}$ configuration:
the minor formed by rows $i_1, i_2, \cdots, i_n$ ($i_1 < i_2 < \cdots < i_n$) and columns $j_1, j_2, \cdots, j_{n-1}, N+1$ ($j_1 < j_2 < \cdots < j_{n-1} < N+1$)
corresponds to the configuration $(c_0, v_1, c_1, \cdots, v_{n - 1}, c_{n - 1}) = (i_1 + N, j_{n-1}, i_2 + N, \cdots, j_1, i_n + N)$.
\subsection{Pruning the configuration space using the breadth-first search algorithm}
\label{sec:bfs}
With this updating technique, we can access the determinants of many configurations without repeatedly carrying out the full determinant calculation for each.
However, the number of configurations still grows exponentially with the order $n$.
Even the $f^{(3)}$ group grows rapidly as $M^5$, and a system with $M=10^3$ orbitals may have $10^{15}$ $f^{(3)}$ configurations.
The problem now becomes how to efficiently find all the significant minors of $\bm{\zeta}$ at all orders.
Enumerating all of these minors will definitely be a hard problem that can not be solved within a polynomial time,
and the question is whether it is necessary to visit all of them.
In fact, we find that for the systems studied in this work (introduced in Sec. \ref{sec:results}) $\bm{\zeta}$ is sparse, with its non-vanishing elements concentrated in some regions, as will be shown in Sec. \ref{sec:comp}.
A more efficient algorithm should be possible given the sparsity of $\bm{\zeta}$.
To make the best use of the sparsity of $\bm{\zeta}$, we investigate its minor determinants in a bottom-up and recursive manner.
According to the Laplace (cofactor) expansion, an $n \times n $ determinant can be expanded into a weighted sum of $n$ minors of size $(n-1)\times(n-1)$.
The $n \times n $ determinant is non-vanishing only when at least one of these $(n-1)\times(n-1)$ minors is non-vanishing.
Physically, this means that a transition to an $f^{(n)}$ configuration is only probable when at least one of its parent $f^{(n-1)}$ configurations is probable,
otherwise the transition to that $f^{(n)}$ configuration is forbidden.
Assume that $\bm{\zeta}$ is sparse, and one can keep a short list of non-vanishing $(n-1)\times(n-1)$ minors.
When proceeding to $n^{th}$ order, one can construct the $n \times n$ determinant from the short list of non-vanishing $(n-1)\times(n-1)$ minors instead of exhaustively listing all of them.
This recursive construction of $n$-minors from the $(n-1)$-minors leads us to an ultimate improvement to the efficiency of the determinantal approach.
We employ the breadth-first search (BFS) algorithm to enumerate all important minors of $\bm{\zeta}$.
A $f^{(n)}$ configuration can be considered as a descendent of $f^{(n-1)}$ via creating one more $\emph{e}$-$\emph{h}${} pair with the $f^{(n-1)}$ configuration.
Through arranging $f^{(n)}$ according to this inheritance relation, a tree-like structure of the many-body expansion is formed, as illustrated in Fig. \ref{fig:tree_bfs}.
The BFS algorithm visits this tree-like structure in ascending order of $f^{(n)}$.
Note that a $f^{(n)}$ configuration can be accessed from its multiple $f^{(n-1)}$ parents via different pathways.
If these pathways to the $f^{(n)}$ configuration interfere destructively such that the transition amplitude is vanishingly small,
the BFS algorithm will discard this $f^{(n)}$ configuration, hence reducing the search space for the next order.
Here is the detailed algorithm
\begin{algorithm}[H]
\caption{Breadth-First Search for Pathways}\label{algo:bfs}
\begin{algorithmic}[1]
\State initialize $f^{(1)}$ * \label{algo:bfs:f1}
\State $n \gets 2$
\Repeat
\For{$f \in f^{(n - 1)}$}
\State extract the indices of $f$: $(c_0, v_1, c_1, \cdots, v_{n - 2}, c_{n - 2})$
\ForAll{$\zeta_{cv}$ satisfying $|\zeta_{cv}| > \zeta_\text{th}$} *\label{algo:bfs:zth}
\If{$c\notin \{c_0, c_1, \cdots, c_{n - 2}\} $ and $v < v_{n - 2}$} *\label{algo:bfs:v}
\State Obtain a composite index at $f^{(n)}$ order:
\Statex \hspace{\algorithmicindent} \hspace{\algorithmicindent} \hspace{\algorithmicindent} $\: f' \gets (c'_0, v'_1, c'_1, \cdots, v'_{n - 1}, c'_{n - 1})$ *\label{algo:bfs:index}
\If{$f' \notin f^{(n)}$}
\State Add $f'$ to $f^{(n)}$
\State $A^{f'} \gets 0$
\State $E_{f'}\gets E_f + (\tilde{\varepsilon_c} - \tilde{\varepsilon_v})$
\EndIf
\State $A^{f'} \gets A^{f'} + (-1)^p \zeta_{cv} A^f $ *\label{algo:bfs:af}
\EndIf
\EndFor
\EndFor
\For{$f \in f^{(n)}$}
\If{$|A^f|^2 < I_\text{th}$}
Delete $f$ *\label{algo:bfs:ith}
\EndIf
\EndFor
\State Calculate the spectral contribution from $f^{(n)}$
\State $n \gets n + 1$
\Until{the spectrum converges}
\end{algorithmic}
\end{algorithm}
Below are further instructions on the lines marked by asterisks.
\begin{enumerate}
\item[L\ref{algo:bfs:f1}:]
$A^f$ of $f^{(1)}$ can be simply taken from the nonzero matrix elements on the last column of $\bm{\zeta}$.
\item[L\ref{algo:bfs:zth}:]
$\zeta_\text{th}$ is a threshold for small matrix elements.
One can set $\zeta_\text{th} = r_\text{th} \zeta_m$,
where $\zeta_m\equiv \max{|\zeta_{ij}|} $ and $r_\text{th}$ is a user-defined relative threshold.
\item[L\ref{algo:bfs:v}:]
The $n\times n$ determinant of $f^{(n)}$ is constructed via a Laplace expansion along its \emph{first column}.
$v < v_{n - 2}$ ensures the chosen matrix element $\zeta_{cv}$ is always on the first column of the $n\times n$ determinant.
\item[L\ref{algo:bfs:index}:]
Compare to $f$, $f'$ contains one more $\emph{e}$-$\emph{h}${} pair labeled by $c$ and $v$.
Because we require the ordering of $c_0 < c_1 < \cdots $ and $v_1 > v_2 > \cdots$ for unique indexing,
the new index $(c'_0, v'_1, c'_1, \cdots, v'_{n - 1}, c'_{n - 1})$ must obey the same order.
The new sequence $(c'_i)$ can be obtained by this procedure:
first place the new $c$ in front of the $(c_i)$ sequence of $f^{(n-1)}$ that is already increasingly sorted,
and then shift $c$ to the right by swapping indices till the whole sequence is also sorted.
Define $p$ to be the number of swaps performed for deciding signs.
$(v'_i)$ can be obtained simply by placing $v$ at the end of $(v_i)$.
\item[L\ref{algo:bfs:af}:]
$(-1)^p \zeta_{cv}$ is the cofactor of the Laplace expansion of a $f^{(n)}$ determinant.
where $p$ is the proper position for inserting $c$ into $(c_i)$, as defined above.
At the end of the $\zeta_{cv}$ loop, there are at most $n$ contributions to the total amplitude $A^{f'}$ of a specific $f^{(n)}$ configuration,
corresponding to the transition amplitudes of $n$ different pathways from its parent $f^{(n-1)}$ configuration.
\item[L\ref{algo:bfs:ith}:]
$I_\text{th}$ is a threshold for removing state with small oscillator strengths.
Similar to $\zeta_\text{th}$, $I_\text{th}$ can be set to $I_\text{th} = R_\text{th} I_m$,
where $I_m$ is the maximal oscillator strength and $R_\text{th}$ is a user-defined relative threshold.
$I_m$ can be chosen to be the maximal intensity within the $f^{(1)}$ group which typically have the strongest oscillator strengths among all $f^{(n)}$ groups.
$R_\text{th}$ can be related to the previously defined relative matrix-element threshold $r_\text{th}$.
If the contribution from a small $a_{cv}$ were not added to $A^{f'}$,
its intensity would be $|A^{f'} - (-1)^p \zeta_{cv} A^f |^2 = | A^{f'} |^2 - 2 (-1)^p \text{Re}[\zeta_{cv} A^{f'} (A^f)^*] + \mathcal{O}(|\zeta_{cv}|^2)$.
Replacing $\zeta_{cv}$ with $0$ will lead to an error of $\sim |\zeta_{cv} | |A^{f'}| |A^f | \leq r_\text{th} |\zeta_m | I_m$.
Therefore, choosing a $r_\text{th}$ such that $R_\text{th} \sim \zeta_m r_\text{th}$ can guarantee error in intensities smaller than $I_\text{th} = R_\text{th} I_m$.
In practice, one can lower $R_\text{th}$ till convergence is achieved.
\end{enumerate}
The detailed implementation of this search algorithm can be found at Ref. \cite{mbxaspy} within an open-source PYTHON simulation package.
\begin{figure*}
\centering
\includegraphics[angle=0, width=0.90\linewidth]{./tree_conf.png}
\caption{The search tree in the BFS algorithm for finding all nontrivial minors of $\bm{\zeta}$.
The digits in the bracket denote the configuration, e.g., $(53629)$ means $(c=5, v_1 = 3, c_1 = 6, v_2 = 2, c_2 = 9)$.
The semi-transparent configurations are discarded in the search process so that they don't spawn any child configuration.
}
\label{fig:tree_bfs}
\end{figure*}
Here we demonstrate the BFS algorithm with a toy model with $M = 9$ orbitals and $N = 4$ valence electrons.
Suppose $\bm{\zeta}$ of the system is
\begin{align}
\begin{split}
\begin{bmatrix}
& 1 & 2 & 3 & 4 & 5 \\
5 & 0 & 0 & 0 & 0 & 1 \\
6 & 0 & \zeta_{62} & \zeta_{63} & 0 & \zeta_{65} \\
7 & \zeta_{71} & 0 & 0 & 0 & 0 \\
8 & 0 & 0 & \zeta_{83} & 0 & \zeta_{85} \\
9 & 0 & \zeta_{92} & \zeta_{93} & 0 & 0 \\
\end{bmatrix}
\end{split}
\label{eq:zeta_demo}
\end{align}
The BFS algorithm for this example of $\bm{\zeta}$ is carried out as follows:
First, the non-zero $f^{(1)}$ configurations are initialized, (5), (6), and (8),
whose determinants are simply the matrix elements: $1$, $\zeta_{65}$, and $\zeta_{85}$, respectively.
These configurations are considered as the roots of the BFS trees,
as is shown in Fig. \ref{fig:tree_bfs}.
Next, the $f^{(2)}$ configurations are constructed based on the obtained $f^{(1)}$ configurations.
Take the configuration $(8)$ in $f^{(1)}$ for example.
There are 5 non-zero matrix elements that are to the left of $(8)$ and are not on the same row as $(8)$,
which are $\zeta_{63}$, $\zeta_{93}$, $\zeta_{62}$, $\zeta_{92}$, and $\zeta_{71}$.
Paired up with these matrix elements, the $(8)$ configuration spawns 5 $f^{(2)}$ configurations:
$(638)$, $(839)$, $(628)$ , $(829)$, and $(718)$ (comma omitted due to the single-digit indices).
Likewise, $(5)$ and $(6)$ spawn 6 and 4 $f^{(2)}$ configurations respectively.
Both $(6)$ and $(8)$ give rise to $(638)$ and the contributions from the $f^{(1)}$ configurations are merged:
A(638) = $\zeta_{63}\zeta_{85} - \zeta_{83} \zeta_{65}$.
The two possible pathways are:
(1) the core electron is first promoted to orbital $6$ and then coupled with the $\emph{e}$-$\emph{h}${} pair formed by orbital $3$ and $8$;
(2) the core electron is first promoted to orbital $8$ and then coupled with the $\emph{e}$-$\emph{h}${} pair formed by orbital $3$ and $6$.
If $A(638)$ is vanishingly small ($\zeta_{63}\zeta_{85}$ happens to be close with $\zeta_{83} \zeta_{65}$) due to the destructive interference of the two pathways,
Then $(638)$ will be removed from the $f^{(2)}$ list because it cannot contribute to the transition amplitude of any higher-order configuration.
When the search process for $f^{(2)}$ is completed, $13$ nontrivial configurations are found.
Proceeding to the third order, the 13 $f^{(2)}$ configurations spawn $14$ $f^{(3)}$ configurations.
Paring $(536)$ with $\zeta_{92}$ and $(539)$ with $\zeta_{62}$ both lead to $(53629)$, whose determinant is $\zeta_{62}\zeta_{93} - \zeta_{63}\zeta_{92}$.
Paring $(538)$($=\zeta_{83}$) with $\zeta_{62}$ leads to $(53628)$ (the other two pathways are forbidden because $\zeta_{52} = \zeta_{82} = 0$).
If $\zeta_{83}$ and $\zeta_{62}$ are small numbers such that their product is smaller than the specified threshold $I_\text{th}$,
then $(53628)$ will be removed from $f^{(3)}$.
The above process can be repeated until all new determinants are small enough or no new determinants can be found.
If one brute-forcely enumerates all possible determinants,
there are $\binom{5}{2}\binom{5}{1} = 50$ $f^{(2)}$ and $\binom{5}{3}\binom{5}{2} = 100$ $f^{(3)}$ determinants to examine for the above $\zeta$-matrix.
By contrast, the BFS algorithm only visits the nontrivial determinants and only $14$ $f^{(2)}$ and $14$ $f^{(3)}$ determinants are computed.
\subsection{Incorporating first-principles calculations into the determinant formalism}
\label{sec:dscf}
In the above sections, we have demonstrated an efficient solution to the MND model using many-body wave functions
for simulating x-ray transition amplitudes.
However, in order to simulate reliable x-ray spectra without fitting parameters from experimetns,
we still need accurate approximations to the initial and final states and their energies.
To this end, we rely on DFT calculations to obtain the KS eigenstate energies (for $\tilde{\varepsilon}_f$) and wavefunctions (for both $|\tilde{\psi}_i\rangle$ and $|\psi_j\rangle$)
as input for constructing the transformation matrix $A^f_c$ (Eq. (\ref{eq:afc}))
and computing the energies of many-electron excited states (Eq. (\ref{eq:Ef})).
For the final state, we employ the standard $\Delta$SCF core-hole approach to obtain the KS orbitals and eigenenergies.
The core-excited atom is treated as an isolated impurity embedded in the pristine system,
and typical supercell settings for finite \cite{england2011hydration} and extended \cite{drisdell2013probing, velasco2014structure, pascal2014x, liang2017accurate} systems can be employed.
To simulate a electron-core-hole pair,
the core-excited atom is modeled by a modified pseudopotential with a core hole,
and an electron is added to the supercell system and constrained to one specific empty orbital.
In principle, a $\Delta$SCF iteration needs to be performed for each case of constraint occupancy (for all $f=(1,2,\dots,N,c)$)
which may lead to an expensive computational cost.
As a trade-off, the electron is only placed onto the lowest unoccupied orbital ($f=(1,2,\dots,N,N+1)$),
which we have dubbed the excited-state core-hole (XCH) method.
After the $\Delta$SCF calculation is done, the KS equation with a converged charge density is used for $\mathcal{H}_f$.
Another important variation of the XCH method is the full core-hole (FCH) approach,
in which the ground-state occupation $f=(1,2,\dots,N)$ without the additional electron is used.
The advantage of FCH is that it does not bias towards the lowest excited state and treat all of them on equal footing.
For the initial state, the same supercell as in the final state is used except that the core-excited atom is replaced by a ground-state atom,
using the occupation $f=(1,2,\dots,N)$.
A standard DFT calculation can be done to obtain the KS orbitals $|\psi_i\rangle$.
With the KS orbitals $|\tilde{\psi}_i\rangle$ and $|\psi_j\rangle$ obtained from the $\Delta$SCF core-hole calculation,
we can compute the orbital overlap integral $\xi_{ij} = \langle \psi_j |\tilde{\psi}_i \rangle$ for computing the determinantal amplitudes.
To reduce computational cost,
we employ projector-augmented-wave (PAW) form of ultrasoft pseudopotentials \cite{kresse1999ultrasoft, blochl1994projector, taillefumier2002x} to model electron-ion interactions.
The excited atom potential has deeper energy levels and more contracted orbitals so its PAW construction differ from the ground-state atom.
In Appendix \ref{sec:paw}, we derive a formalism to calculate expectation values between
$|\psi_j\rangle$ of the ground-state system and
$|\tilde{\psi}_i\rangle$ of the core-excited system.
Initial dipole matrix elements $\langle \psi_c | o | \psi_h \rangle$ are also evaluated within this PAW formalism \cite{blochl1994projector, taillefumier2002x}.
As in typical DFT impurity calculations, some low-lying excited states in the core-hole approach could be bound to the core-excited atom, resembling mid-gap localized electronic states near an impurity.
In this situation, the electronic structure is well described by using a single k-point (the $\Gamma$ point) to sample the Brillouin zone (BZ).
However, for the purpose of spectral simulations, which include delocalized scattering states well above the band edges,
we find that employing $k$-point sampling is necessary to improve the accuracy of the calculated line shape.
Therefore, we perform the determinantal calculation individually for each k-point and
take the k-point-weighted average spectrum as the final spectrum.
The band structure and orbitals are interpolated accurately and efficiently using an optimal basis set proposed by Shirley \cite{shirley1996optimal, prendergast2009bloch},
whose size is much smaller than a plane-wave basis.
In the Shirley construction, the periodic parts of Bloch wave functions across the first BZ are represented using a common basis which spans the entire band structure.
Because there is only one optimal basis to represent the Bloch states for all k-points,
the overlap $\xi$-matrix for every k-point can be computed quickly as in Appendix \ref{sec:obf}.
After the XAS is calculated by the first-principles determinantal approach,
the established formation-energy calculation can be adopted to align spectra for core-excited atoms in different chemical contexts, using the XCH method to determine the excitation energy of the first transition \cite{england2011hydration, jiang2013experimental}.
Although there is no valence $\emph{e}$-$\emph{e}${} interaction terms in the MND theory,
which results in a single-determinant solution to the many-body wave functions,
we argue that this first-principles determinantal approach \emph{does not} entirely neglect valence $\emph{e}$-$\emph{e}${} interactions.
The self-consistent-field (SCF) procedure in the DFT updates the total charge density and KS orbitals simultaneously,
and hence takes into account some degree of valence electronic screening.
That said, the $\Delta$SCF approach should lead to a more realistic equilibrium total charge density
for both the ground state and x-ray excited states,
whereas the charge density in MBPT is only treated perturbatively.
Moreover, a more accurate charge density may lead to a better approximation to quasiparticle (QP) wave functions.
In fact, KS orbitals based on a converged SCF are often employed in MBPT to construct the Green's functions and compute optical oscillator strengths \cite{hybertsen1986electron, rohlfing2000electron, onida2002electronic},
which is typically a good approximation within the Fermi-liquid picture.
Finally, corrective DFT (DFT + $U$ or DFT with exact-exchange functionals) or the self-consistent $GW$ approximation \cite{bruneval2006effect, van2006quasiparticle, kang2010enhanced, sun2017x}
can also improve QP energies and wave functions to be used in the determinantal approach.
We imagine the many-body effects captured in this framework can be described by the bolded version of the MND diagrams in Fig. \ref{fig:lcc_2nd}
in which all the Green's functions become dressed,
and the bare Coulomb lines are replaced by the screened core-hole potential described with the chosen exchange-correlation functional.
\subsection{Comparison with the one-body $\Delta$SCF core-hole approach}
\label{sec:onebody}
Before the determinantal formalism, the many-body transition amplitudes in the $\Delta$SCF core-hole approach are often approximated with 1p matrix elements
\begin{equation}
\langle\Psi_f| \bm{\epsilon}{\cdot}\bm{R} |\Psi_i\rangle \approx S\langle\tilde{\psi_f}| \bm{\epsilon}{\cdot}\bm{r} |\psi_h\rangle
\label{eq:dscf_mat_elem}
\end{equation}
where the core orbital $|\psi_h\rangle$ is in the initial state while the electron orbital $|\tilde{\psi}_f\rangle$ is in the final states,
both of which can be taken from DFT calculations.
$S$ represents the response of rest of the many-electron system (excluding the electron-core-hole pair) due to the core hole,
and it is normally assumed to be a constant for ease of calculations.
This 1p form of the matrix element implies that:
(a) the transition from the initial core level $|\psi_h\rangle$ to the final electron orbital $|\tilde{\psi}_f\rangle$ occurs instantaneously
with the response of many other electrons in the system, with no particular time ordering;
(b) the core-level transition and the many-electron response are not entangled.
This is also the so called sudden or frozen approximation.
We know that from the diagrammatic interpretation of the x-ray many-body processes in Fig. \ref{fig:lcc_2nd},
the photon first decays into an initial-state $\emph{e}$-$\emph{h}${} pair instantaneously,
and then the other electrons see the core-hole potential and begin to relax over a finite period of time.
This physical reality can also be seen in the determinant formalism, in which
the core hole is only coupled to an initial-state orbital, and the subsequent many-electron response is described by the determinantal amplitude.
So the question is why the simpler 1p matrix element in the frozen approximation still works for a good number of systems in the past.
In this section, we approach this question theoretically by relating the determinantal amplitude to the 1p matrix element.
To do this, we first express the $(N+1)\times(N+1)$ determinantal amplitude $A^{f}_c$ in terms of is $N\times N$ minors
(wavefunction overlaps of $N$-electron systems, such as $S$) by Laplace expansion along its last column
\begin{align}
\begin{split}
A^{f}_c = \sum^{N+1}_{i=1} M^f_{i} \xi_{f_i, c}
\end{split}
\label{eq:afc_minor}
\end{align}
where $\xi_{f_i, c}$ are the matrix elements on the last column of $A^f_c$ as in Eq. (\ref{eq:afc}) and $M^f_{i}$ is the minor complementary to $\xi_{f_i, c}$.
Since in the one-body core-hole approach only the $f^{(1)}$ terms are summed, we limit our analysis here to the many-body $f^{(1)}$ terms and condense the configuration tuple into a single index:
$(1, 2, \cdots, N, f) \mapsto f$.
Then the matrix elements $\langle \Psi_f | O | \Psi_i \rangle$ can be written as
\begin{align}
\begin{split}
&\sum_{c \in \text{empty}} (A^{f}_c)^* \langle \psi_c | o | \psi_h \rangle\\
=& (M^f_{N+1})^* \sum_{c \in \text{empty}} \langle \tilde{\psi}_f | \psi_c \rangle \langle \psi_c | o | \psi_h \rangle\\
+ & \sum^{N}_{i=1}(M^f_{i})^* \sum_{c \in \text{empty}} \langle \tilde{\psi}_i | \psi_c \rangle \langle \psi_c | o | \psi_h \rangle
\end{split}
\label{eq:mat_decomposed}
\end{align}
First, for systems with significant band gaps (insulators and semiconductors), we could expect that the overlap of the occupied final state orbitals with the unoccupied initial state orbitals could be quite small. For many orbitals unaffected by the localized core-hole perturbation, for example, we might expect the final state occupied orbitals to closely resemble their initial state counterparts, which would render $\langle \tilde{\psi}_v | \psi_c \rangle$ identically zero by orthogonality. Therefore, the sum over $v$ in Eq. (\ref{eq:mat_decomposed}) may only be significant in cases where the transformation matrix $\mathbf{\xi}$ indicates mixing of unoccupied initial state character into the occupied final state orbitals, which might easily be the case for orbitals close to the Fermi level in a metal or otherwise open-shell system.
The first term in Eq. (\ref{eq:mat_decomposed}) is more directly relevant to our previous one-body approximation. Here, $M^f_{N+1}$ is the minor of $(\xi_{ij})_{N\times N}$, the transformation matrix without its $(N+1)^\textrm{th}$ column and row. It reflects the $N$-electron many-body overlap between the initial and final state occupied orbitals and should reflect the extent to which the electron density is modified by the core-hole perturbation.
Since $M^f_{N+1}$ does not depend on $f$,
we can relate it to the many-body prefactor that appears in the final-state rule of Eq. (\ref{eq:dscf_mat_elem}): $S=(M^f_{N+1})^*$.
Using the completeness relation: $\sum_{c \in \text{empty}} | \psi_c \rangle \langle \psi_c | = \mathbb{1} - \sum_{v \in \text{occ}} | \psi_v \rangle \langle \psi_v |$, the first term in the expansion of Eq. (\ref{eq:mat_decomposed}) can be expressed as
\begin{align}
\begin{split}
S \big[ \langle \tilde{\psi}_f | o | \psi_h \rangle - \sum_{v \in \text{occ}} \langle \tilde{\psi}_f | \psi_v \rangle \langle \psi_v | o | \psi_h \rangle \big]
\end{split}
\label{eq:compare_to_f}
\end{align}
If it happened that $\langle \psi_v | \tilde{\psi}_f \rangle = 0$, then this expression would amount to the final state matrix element as defined in the one-body final-state rule (Eq. (\ref{eq:dscf_mat_elem})). By the same arguments made above, for systems with limited mixing of orbital character across a significant band gap, then we might easily expect orthogonality (zero overlap) between occupied initial state and unoccupied final state orbitals. By the same token, we should be wary of limitations in the one-body approach when this is not the case.
It appears useful to focus on $\langle \psi_v | \tilde{\psi}_f \rangle$ to reveal the role of hybridization in modulating near-edge spectral intensity.
To quantify the contribution of the second term in Eq. (\ref{eq:compare_to_f}), we introduce the projection spectrum
\begin{align}
\begin{split}
\sigma_{fi} (E)
&= \sum_{f} | \langle \tilde{\psi}_f | P_c o|\psi_h\rangle|^2 \delta(E - \tilde{\varepsilon}_f) \\
\end{split}
\label{eq:sigma_fi}
\end{align}
in which the single index $f$ sums over all empty final-state orbitals, and $P_c \equiv \sum_{c \in \text{empty}} | \psi_c \rangle \langle \psi_c |$.
The matrix element is nothing but Eq. (\ref{eq:compare_to_f}) or the first term in Eq. (\ref{eq:mat_decomposed}) with $S=1$.
However, it is easier to calculate Eq. (\ref{eq:compare_to_f}) because summation over all empty orbitals is avoided.
\section{RESULTS AND DISCUSSION}
\label{SEC:results}
\subsection{Applications to transition-metal oxides}
\label{sec:results}
In this section, we discuss an important application of the determinantal approach to computing core-excited state transition amplitudes, that is, to predict the x-ray absorption spectra (XAS) for transition metal oxides (TMOs).
This is also our original motivation for proposing the determinantal approach \cite{liang2017accurate}, which can be used to overcome the deficiency of the one-body core-hole approach.
It has been found for a number of TMOs, that the one-body approach systematically underestimates the intensity of near-edge features at the O $K$ edge that correspond to orbitals with hybridization between oxygen $p$-character and TM $3d$-character.
This underestimation can prevent reliable interpretations of the X-ray absorption spectra for this important class of materials.
We use the newly developed determinantal approach to predict the XAS for eight TMOs:
the rutile phase of
\ce{TiO2},
\ce{VO2} ($>340$ K),
and \ce{CrO2},
the corundum \ce{Fe2O3},
the perovskite \ce{SrTiO3},
\ce{NiO},
and \ce{CuO}.
\ce{SiO2} is also chosen for a comparative study.
Their experimental XAS are extracted from Refs. \cite{yan2009oxygen, koethe2006transfer, stagarescu2000orbital, shen2014surface, zhu2012bonding, lin2014hierarchically, jiang2013experimental, dudarev1998electron, ma1992soft}.
The chosen TMOs cover a wide range of electronic and magnetic properties and therefore they are used as benchmark materials for the determinantal approach.
The \ce{O} $K$ edges are investigated here, i.e., the transitions from the \ce{O} $1s$ level to $np$ shells.
For TMOs, the \ce{O} $2p$ orbitals are covalently hybridized with the transition metal $3d$ orbitals,
and hence the \ce{O} $K$-edge spectra can serve as an informative and sensitive probe for the $d$-electron physics \cite{de1989oxygen, yabuuchi2011detailed, hu2013origin, suntivich2011design, lin2016metal, luo2016charge, strasser2010lattice, matsukawa2014enhancing, lebens2016direct, de2016mapping}.
Moreover, unlike transition metal $L_{2,3}$ edges ($2p$-to-$3d$ transitions), in which atomic multiplet effects split spectral features into many closely space lines\cite{de2008core}, the \ce{O} $K$ edges can provide a picture of the electronic density-of-states related to the $d$ shell more easily interpretable in terms of band theory or effective 1p states.
The angularly-averaged (except in \ce{CrO2}, where the polarization is perpendicular to the hard axis) \ce{O} $K$-edge spectra for the chosen TMOs are shown in Fig. \ref{fig:results} (a).
The very near-edge part of the spectra, i.e., the spectral features below $535$ eV contain the most useful information for $3d$ material characterization.
For these TMOs, the near-edge spectral fine structure exhibits two main peaks corresponding to the splitting of the $d$-orbitals into a $t_{2g}$ and an $e_g$ manifold in the (quasi-)octahedral crystal field.
Our goal is to produce reliably all the spectral features, especially the very near-edge part, so that one can interpret the spectra on a first-principles basis.
More specifically, we use the ratio of the intensity of the first (lowest-lying) peak to that of the second (unless otherwise specified) as a metric for the accuracy of different levels of approximation.
\floatsetup[figure]{subcapbesideposition=top}
\begin{figure*}
\centering
\sidesubfloat[(a)]{
\includegraphics[angle=0, width=0.9\linewidth]{./all_pub.eps}
}\\
\sidesubfloat[(b)]{
\includegraphics[angle=0, width=0.3\linewidth]{./tmos_info.eps}
}
\sidesubfloat[(c)]{
\includegraphics[angle=0, width=0.5\linewidth]{./if_schematics.png}
}
\caption{ (Color online)
\textbf{(a)} XAS for the selected crystal structures obtained from experiments (black), one-body FCH approach (blue), and the many-electron determinant approach (red) introduced in this work.
The XAS calculated with the $f^{(1)}$ configurations are shown by dashed orange curves.
The energy axes for \ce{NiO} and \ce{SiO_2} are relative.
\textbf{(b)} Comparison of experimental peak intensity ratios compared with the ones predicted by the one-body (circles) and the many-electron (triangles) formalism.
Each color represents the result for one system.
The peak intensity ratio refers to the ratio of the lowest-energy maxima to the second of the spectrum, unless otherwise specified by the numbers in (a).
The spectra are broadened to the best as compared with experimental broadenings.
\textbf{(c)} Schematics showing how the one-body and the many-electron formalism treats x-ray excitations, using a metal-$3d$-\ce{O}-$2p$ molecular model in both
the initial (i) and final (f) state.
The one-body approach mainly relies on the single-particle (1p) matrix element and has skipped (red arrow) the dynamics of the many-electron charge relaxation,
while the many-electron formalism considers the actual multiple-step (blue arrows) excitation process that involves all the electrons in the system.
}
\label{fig:results}
\end{figure*}
We first calculate the XAS for the chosen compounds using the conventional 1p FCH approach \cite{taillefumier2002x, prendergast2006x, liang2017accurate} described above.
A modified pseudopotential generated with the configuration $1s^1 2s^2 2p^4$ is used for the $1s$-core-excited \ce{O}.
We choose supercell dimensions of approximately $10$\AA{} that is sufficient to separate the effect of the core-hole impurity from its neighboring periodic images.
The FCH calculations are performed using the DFT+$U$ theory \cite{dudarev1998electron} with the $U$ value adopted from Ref. \cite{wang2006oxidation}.
A uniform $5\times 5 \times 5$ $\bm{k}$-point grid of the supercell BZ is employed to sample a continuous density-of-states at higher energies.
As we have demonstrated by calculations before \cite{liang2017accurate}, the 1p FCH approach universally underestimates the peak intensity ratio for all selected TMOs (blue curves in Fig. \ref{fig:results} (a)).
This includes the newly added cases: \ce{MnO2}, \ce{NiO}, and \ce{CuO}, where the peak intensity ratios are just $50\%$ of the experimental ones.
The failure of the 1p FCH approach motivated us to use the determinant formalism in Eq. (\ref{eq:afc}) as a better approximation to the dipole matrix elements \cite{liang2017accurate}.
In this work, we implement the determinant approach with the efficient procedures discussed in Sec. \ref{sec:dscf} and the BFS algorithm.
We use exactly the same final-state SCF as in the 1p FCH approach and an initial-state supercell of the same dimensions.
Besides employing the BFS algorithm to reduce the computational cost, we separate the two spin channels to speed up the calculations.
In the absence of spin-orbital coupling, the $\xi$-matrix is block-diagonalized and each transition either occurs within the spin-up or spin-down manifold,
and the BFS algorithm can be performed over each spin manifold with a reduced $\xi$-matrix.
The total absorption spectra can be obtained from combining individual spectra from the two spin channels (for the collinear case) using the spectral convolution theorem in Appendix \ref{sec:conv}
\begin{align}
\begin{split}
\sigma_{\text{XAS}}(E) &=\int d E' \sigma_{\text{XAS}, \uparrow}(E - E') \sigma_{\text{XPS}, \downarrow} (E') + \{\uparrow \rightleftharpoons\downarrow\}
\end{split}
\label{eq:xas_conv}
\end{align}
where $\sigma_{\text{XAS}, \mu}$ is the XAS of an individual spin channel $\mu$.
$\sigma_{\text{XPS}, \mu}$ is the core-hole spectral function of spin $\mu$
\begin{align}
\begin{split}
\sigma_{\text{XPS}, \mu} (E)= \sum_{f}\langle \Psi^{N_\mu}_{\mu, f} | \Psi^{N_\mu}_{\mu, i} \rangle \delta(E-(E_{\mu, f} - \min E_{\mu, f}))
\end{split}
\label{eq:ch_A}
\end{align}
Here $|\Psi^{N_\mu}_{\mu, f}\rangle$ ($|\Psi^{N_\mu}_{\mu, i}\rangle$) is the $N_\mu$-electrons many-body wave function
of final state $f$ (initial state $i$) within the spin manifold $\mu$,
and $N_{\mu}$ is the number of electrons in its initial state ($N_{\uparrow}\neq N_{\downarrow}$ for a ferromagnetic system).
Because the core-hole spectral function
is analogous to the corresponding x-ray photoemission spectrum (XPS) \cite{kas2015real}, we dub the former as $\sigma_{\text{XPS}}$ hereafter.
The calculation of $\sigma_{\text{XPS}, \mu}$ entirely resembles that of $\sigma_{\text{XAS}, \mu}$ and the prominent matrix elements $\langle \Psi^{N_\mu}_{\mu, f} | \Psi^{N_\mu}_{\mu, i} \rangle$ are also found by the BFS algorithm as in Sec. \ref{sec:sol}.
The spectra calculated with the determinantal approach up to $f^{(2)}$ order are shown in Fig. \ref{fig:results} (a).
There is substantial improvement in the peak intensity ratios and the overall line shapes for the TMOs being investigated.
In particular, the peak intensity ratios of \ce{TiO2}, \ce{CrO2}, \ce{Fe2O3}, \ce{CuO}, \ce{NiO}, and \ce{SrTiO3} are in excellent agreement with experiments [Figs. \ref{fig:results} (a) and (b) ].
The peak intensity ratio of \ce{VO2} is still underestimated, however, this may be related to missing contributions to the leading edge from the nearby \ce{V} $L$ edge, which is not included in our simulation.\cite{koethe2006transfer}.
The prediction of the peak intensity ratio of \ce{MnO2} is less satisfactory partly because we simulate its spectrum using a rutile unit cell with colinear antiferromagnetic order, whereas its actual magnetic order is found to be helical and has a larger periodicity \cite{tompsett2012importance, lim2016improved}.
The lack of anisotropy in the Hubbard $U$ interactions in our current calculation may also explain why the simulated spectrum deviates from experiments.
More advanced treatment of strongly correlated materials, using hybrid functionals, for example,\cite{heyd2003hybrid, chai2008long, paier2009cu, paier2008dielectric}, could be coupled with the determinantal formalism to produce more accurate results.
In principle, any effective 1p orbital basis can be used in this formalism.
\subsection{Origins of XAS intensity underestimation using one-body approaches}
\label{sec:intunder}
In a nutshell, the underestimation of the peak intensity ratios by the one-body approach can be understood from a three energy-level model.
Consider a molecule with one single metal level (M) hybridized with an \ce{O} $2p$ level, plus one \ce{O} $1s$ core level, as is shown in the schematics in Fig. \ref{fig:results} (c). Hybridization within the empty $(c)$ and filled $(v)$ states can be expressed using a unitary transformation of the corresponding atomic orbitals:
$(|\tilde{\psi}_c\rangle, |\tilde{\psi}_v\rangle)^T = R(\theta_i) (|\text{M}_{3d}\rangle, |\text{O}_{2p}\rangle)^T $, where
$R(\theta_i)$ is a 2D rotation matrix
\begin{align}
\begin{split}
R(\theta)=
\begin{bmatrix}
\cos\theta & -\sin\theta \\
\sin\theta & \cos\theta
\end{bmatrix}
\end{split}
\label{eq:model_1}
\end{align}
Initially the system is half filled and its hybridization represented by an angle $\theta_i\in [0,\pi/2]$.
The final state can be expressed likewise using its own angle $\theta_f$:
$(|\tilde{\psi}_c\rangle, |\tilde{\psi}_v\rangle)^T = R(\theta_f) (|\text{M}_{3d}\rangle, |\text{O}_{2p}\rangle)^T $.
Phenomenologically, we expect the initial and final states to differ in their degree of hybridization of these two atomic levels.
The core-hole potential lowers the energy of the oxygen $2p$ orbital in the final state,
enhancing the $ |\text{O}_{2p}\rangle$ component of the occupied final-state orbital $v$ and reducing the same for the unoccupied final-state orbital $c$.
Hence, $0 < \theta_f < \theta_i$.
Within this minimal model of just two electrons, there is only one available core-excited transition,
i.e., the excitation from $i=(h,v)$ to the final state $f=(\tilde{v},\tilde{c})$.
The exact spectral intensity calculated by the many-electron formalism as in Eq. (\ref{eq:afc}) is
\begin{align}
\begin{split}
|\langle\Psi_f| \bm{\epsilon}{\cdot}\bm{R} |\Psi_i\rangle |^2
&= |\det [ R(\theta_i-\theta_f) ] \langle \psi_c|\bm{\epsilon}\cdot\bm{r}|\text{O}_{1s}\rangle|^2\\
&= |1\times\langle \psi_c|\bm{\epsilon}\cdot\bm{r}|\text{O}_{1s}\rangle|^2\\
&= \sin^2 \theta_i |\langle \text{O}_{2p}|\bm{\epsilon}\cdot\bm{r}|\text{O}_{1s}\rangle|^2
\end{split}
\label{eq:model_2}
\end{align}
However, using the one-body core-hole approximation, working with final-state orbitals only, we find
\begin{align}
\begin{split}
|\langle\Psi_f| \bm{\epsilon}{\cdot}\bm{R} |\Psi_i\rangle |^2
&\approx |\langle \tilde{\psi}_c|\bm{\epsilon}\cdot\bm{r}|\text{O}_{1s}\rangle|^2 \\
&= \sin^2 \theta_f |\langle \text{O}_{2p}|\bm{\epsilon}\cdot\bm{r}|\text{O}_{1s}\rangle|^2
\end{split}
\label{eq:model_3}
\end{align}
Therefore, based on the smaller value of $\theta_f$, the one-body final-state intensity is necessarily weaker than the many-electron intensity.
The origin of this underestimation lies in erroneously formulating the excitation as a single-step transition from the core level to the final-state empty orbital,
which contains a reduced \ce{O} $2p$ component due to core-hole attraction [as illustrated in Fig. \ref{fig:results} (c)].
On the other hand, the many-electron formalism takes the correct time-ordering into account, describing a multi-step transition:
the electron is promoted to the unperturbed initial-state empty orbital followed by a many-electron charge transfer.
By this argument, the absorption intensity is the same as in the initial-state picture, $\sin^2 \theta_i |\langle \text{O}_{2p}|\bm{\epsilon}\cdot\bm{r}|\text{O}_{1s}\rangle|^2$.
Note, however, that the energy of the \emph{final-state} configuration should be used in the Fermi's golden rule.
For the two-peak near-edge fine structure in TMOs, we can also make use of the above two-electron model.
Let us define an energy dependent hybridization within the unoccupied orbitals between metal $3d$ and \ce{O} $2p$ character according to $\sin^2\theta = \frac{t^2}{t^2 + \Delta ^ 2}$,
where $t$ is the intrinsic hybridization strength, $\Delta(\varepsilon) = (\varepsilon + \sqrt{\varepsilon ^ 2 + t ^ 2})$, and
$\varepsilon=\varepsilon_{3d} - \varepsilon_{2p} > 0$. Within quasi-octahedral symmetry, we would expect lower intrinsic hybridization values for the $t_{2g}$ orbitals vs. the $e_g$, but the $e_g$ orbital energies should lie above those of the $t_{2g}$. For a two-peak near-edge, we can define the peak intensities using: $t_1$ and $\varepsilon_1 = \varepsilon_{t_{2g}} - \varepsilon_{2p}$ for the lower energy $t_{2g}$ peak and $t_2$ and $\varepsilon_2 = \varepsilon_{e_g} - \varepsilon_{2p}$ for the higher energy $e_g$ peak, assuming $0 < t_1 < t_2$ and $0 < \varepsilon_1 < \varepsilon_2$.
Assume, without loss of generality, that within the initial state picture the $t_{2g}$ and $e_{g}$ peaks have the same intensity:
$\sin^2\theta_{i_1} = \frac{t_1^2}{t_1^2 + \Delta(\varepsilon_1) ^ 2} = \frac{t_2^2}{t_1^2 + \Delta(\varepsilon_2) ^ 2}=\sin^2\theta_{i_2}$.
For the purposes of illustration, we can use the following numerical values: $\varepsilon_{t_{2g}} = 1.0$, $\varepsilon_{e_g} = 4.0$, $\varepsilon_{2p} = -4.0$, and $\sin^2\theta_{i_1} = \sin^2\theta_{i_2} = 0.2$
such that $t_1 = 2.5$ and $t_2 = 4.0$ (a comparable energy unit could be eV), with the expected ordering.
If the core hole deepens the \ce{O} $2p$ orbital energy, $\varepsilon_{2p}$, to $\tilde{\varepsilon}_{2p}$,
then the one-body final-state intensities will change and the intensity ratio decreases, as shown numerically in Table \ref{tab:sin2}.
It can be seen from this example that a one-body final-state estimate of the $3d$ peak-intensity ratio ($\sin^2\theta_{f_1} / \sin^2\theta_{f_2}$) always decreases with increasing core-hole binding.
\renewcommand{\arraystretch}{1.5}
\begin{table}
\setlength{\tabcolsep}{12pt}
\begin{tabular}{ c | c c c c }
\hline
$\tilde{\varepsilon}_{2p}$ & -4.0 ($\varepsilon_{2p}$) & -6.0 & -8.0 & -10.0 \\
\hline
$\sin^2\theta_{f_1}$ & 0.2 & 0.113 & 0.072 & 0.049 \\
$\sin^2\theta_{f_2}$ & 0.2 & 0.138 & 0.100 & 0.075 \\
ratio & 1.0 & 0.82 & 0.72 & 0.65 \\
\hline
\end{tabular}
\caption{The relative near-edge peak intensities in a simple two-electron system with two available empty orbitals having \ce{O} $2p$ hybridization and energies consistent with $t_{2g}$ and $e_g$ orbitals and their dependence on the final state orbital energy $\tilde{\varepsilon}_{2p}$.}
\label{tab:sin2}
\end{table}
\subsection{Charge-transfer effects and impact on simulated spectra}
\label{sec:afi}
\begin{figure}
\centering
\includegraphics[width=0.90\linewidth]{./afi.eps}
\caption{comparison of the final-state spectra ($\sigma_i$) and the projection spectra ($\sigma_{fi}$) for \ce{TiO2}, \ce{SiO2}, and \ce{CrO2}.
The final-state spectra and the spectra from the determinantal approach (up to $f^{(2)}$) are taken from Fig. \ref{fig:results}.
}
\label{fig:afi}
\end{figure}
While the one-body approach fails systematically in predicting the XAS for the chosen TMOs,
it produces a satisfactory lineshape for \ce{SiO2}.
This is consistent with the previous success with using the one-body approach for a wide variety of systems \cite{taillefumier2002x, prendergast2006x, drisdell2013probing, pascal2014x, velasco2014structure, ostrom2015probing, drisdell2017determining} that are not TMOs.
We make use of the connections between the one-body and many-body approaches outlined in Sec. (\ref{sec:onebody}) to understand why this is the case here.
A comparison of spectra obtained in different ways is shown in Fig. \ref{fig:afi}.
The projection spectrum is more intense than the final-state spectrum in all cases,
indicating the hybridization term $\langle \psi_v | \tilde{\psi}_f \rangle$ is not neligible.
However, the spectra of the chosen systems are affected in different manners by this term.
For \ce{SiO2}, the projection spectrum $\sigma_{fi}$ is in proportion to the final-state spectrum $\sigma_f$ (multiplication by the many-body overlap, $S$, correctly renormalizes the spectrum).
On the other hand, the near-edge spectral profiles in \ce{TiO2} and \ce{CrO2} are substantially modified from the one-body approximation by the projection onto empty orbitals,
in particular for \ce{CrO2} where the first peak is partly retrieved in terms of its relative intensity with respect to the second peak (around 532.5 eV).
This indicates that the projection defined in Eq. (\ref{eq:sigma_fi}) plays an important role in retrieving some key absorption features,
which makes this definition an efficient means to determine whether the final-state rule is sufficient for obtaining a satisfactory XAS.
Although the projection spectrum can rectify the deficiency of the final-state rule to some extent,
it is still necessary to employ the determinant formalism for a correct and physical spectrum.
For \ce{CrO2}, the projection spectrum still deviates significantly from experiments, even after it is rescaled by $S$.
This suggests the many-electron effects described by the second terms in Eq. (\ref{eq:mat_decomposed}) are not trivial and should be included.
\begin{figure*}
\centering
\includegraphics[angle=0, width=0.95\linewidth]{./cro2_details.png}
\caption{
\textbf{(a)} Decomposed contributions from the single $f^{(1)}$ and double $f^{(2)}$ configurations to the \ce{O} $K$ edge XAS of \ce{CrO2}.
For each case, the spectrum is decomposed into an spin-up ($\uparrow$) and an spin-down ($\downarrow$) channel.
All the spectra are plotted with the same intensity scale, with sticks, i.e., oscillator strengths of the final states, in the background.
Only $10\%$ states with the strongest oscillator strengths are shown.
The major sticks are highlighted with black bars.
\textbf{(b)} \ce{O} $1s$ XPS of \ce{CrO2}.
The energy of the final \emph{ground-state} (with the least binding energy) is aligned with zero.
\textbf{(c)} Comparison of the peak-intensity ratios of the initial-state spectrum ($\sigma_{A\uparrow} *\sigma_{P\downarrow} + \sigma_{A\downarrow} *\sigma_{P\uparrow}$, black), the final spectrum convoluted from the two spin-channels (red), and the fictitious spectrum without convolution ($\sigma_{A\uparrow} + \sigma_{A\downarrow}$, red).
The first peaks are rescaled to the same height.
\textbf{(d)} Charge difference $\rho_f - \rho_i$ of the $N$-electron charge-transfer (CT) state and relevant 1p orbitals.
$e_i$ and $h_i$ denotes empty and occupied orbitals respectively.
Final-state orbitals are annotated with \emph{tilde}.
$\bm{a}$ and $\bm{b}$ are two hard axes of \ce{CrO2} and $\bm{c}$ is the easy axis.
The photon polarization is in the hard-plane.
Note that the CT state is shown from a perspective different from the 1p orbitals.
For the CT plot, the charge gain (loss) is shown in orange (green).
For the orbital plots, yellow and cyan indicate the phases of the spatial wave functions.
}
\label{fig:cro2_details}
\end{figure*}
We consider the XAS of \ce{CrO2} in more detail.
Fig. \ref{fig:cro2_details} (a) shows the spin-dependent $f^{(1)}$ and $f^{(2)}$ contributions to the spectrum separately,
together with the oscillator strengths of some main transitions ($>10\%$ of the strongest transitions) presented as ``sticks''.
We begin with an analysis of the $f^{(1)}$ terms that consist of only a single electron-core-hole pair.
Because the core hole is fixed, an $f^{(1)}$ term can be mapped to a single empty \emph{final-state} orbital
\begin{align}
\begin{split}
A \mapsto \tilde{e}_1 \uparrow, B \mapsto \tilde{e}_3 \uparrow, C\mapsto \tilde{e}_3\downarrow, D\mapsto \tilde{e}_4 \uparrow
\end{split}
\end{align}
where $\tilde{e}_3 \uparrow$ and $\tilde{e}_3 \downarrow$ closely resemble one another, only one of which is shown in \ref{fig:cro2_details} (d).
The orbitals defining A, B, C, and D correspond to a $t_{2g}$ $d_{xy}$, an $e_{g}$ $d_{z^2} \uparrow $, an $e_{g}$ $d_{z^2} \downarrow$, and an unbound itinerant ($p$-like) orbital respectively.
Hereafter, $\uparrow$ is omitted unless for spin-down orbitals.
What do these transitions have in common?
They all reflect projections of the initial (ground) state, mediated by the photon electric field, onto final states that share a common \ce{O} $1s$ core-hole excitation and its associated perturbing potential.
The core hole attracts electron density towards the excited \ce{O} site, as can be seen from the plotted isosurface of the charge-density difference $\rho_f - \rho_i$ in Fig. \ref{fig:cro2_details} (d) (top left).
This charge transfer results from the response of the $N$-electron system to the core-hole potential.
It is computed as the deviation of final-state DFT charge density $\rho_f$ (without the excited electron as in the FCH approximation) from the one of the initial state $\rho_i$.
According to the first term in Eq. (\ref{eq:mat_decomposed}), there is a single prefactor common to all final states for this component of the $f^{(1)}$ transitions, also denoted $S$ in Eq. (\ref{eq:compare_to_f}).
This $N$-electron determinant is yet another way of representing the CT state.
Generally speaking, all $(N+1)$-electron final states, within this MND single-determinant picture, only differ by a few composite single-particle orbitals that slightly modulate this CT density.
The $f^{(1)}$ states differ by the addition of just one final-state unoccupied orbital.
Close examination of the final state orbitals in \ref{fig:cro2_details} (d) reveals, surprisingly, that the brightest transition of the entire spectrum originates from state $A$,
even though its excited electron orbital, $\tilde{e}_1$ , does \emph{not overlap} with the excited \ce{O} atom [marked by \lq\lq X\rq\rq in Fig. \ref{fig:cro2_details} (d)].
As a result, the one-body final-state rule gives a transition amplitude of only
\begin{align}
\begin{split}
|\langle \tilde{e}_1 |x| \psi_h\rangle| = 9.16 \times10^{-6} (\text{a.u.})
\end{split}
\end{align}
which explains the lack of any significant first peak in the simulated one-body XAS of \ce{CrO2} in Figs.~\ref{fig:results} and~\ref{fig:afi}.
This small amplitude is due to Pauli-blocking resulting from the charge transfer -- in other words, the core-hole potential has lowered some initially unoccupied \ce{O} $2p$ orbital character below the Fermi level of this half-metal, rendering it inaccessible within this 1p picture.
By contrast, the many-body determinantal amplitude of $A$ is a few orders of magnitude larger:
\begin{align}
\begin{split}
|\sum_{c\in\text{empty}} (A^f_c)^* \langle\psi_c |x| \psi_h\rangle| = 1.08\times10^{-2}
\end{split}
\end{align}
To understand why the many-body state $A$ still has a strong oscillator strength,
an analysis can be provided based on the Laplace expansion of the determinantal amplitude in Eq. (\ref{eq:mat_decomposed}).
By inspection, we find that the most important contributions to the amplitude of $A$ are from $\tilde{\psi}_i=\tilde{e}_1$, $\tilde{h}_4$, and $\tilde{h}_3$.
They have substantial overlap (integrals tabulated in Tab. \ref{tab:A}) with a number of \emph{initial-state} empty orbitals that exhibit $p$ character at the excited \ce{O} atom,
such as $e_1$, $e_2$, and $e_3$, as shown in Fig. \ref{fig:cro2_details} (d) (top row).
Consequently, the projection amplitudes $| \langle \tilde{\psi}_i | P_c x | \psi_h \rangle|$ of $\tilde{e}_1$, $\tilde{h}_4$, and $\tilde{h}_3$ are still significant (i.e., similar in magnitude to the amplitude of $A$),
although these final-state orbitals may have small overlap with the core hole.
Furthermore, the corresponding many-electron overlaps, $M^f_i$, are not small (Tab. \ref{tab:A}).
Therefore, the combined contribution $\sum_i (M^f_i)^* \langle \tilde{\psi}_i | P_c x | \psi_h \rangle$ for $\tilde{\psi_i}$ in $\{ \tilde{e}_1$, $\tilde{h}_4$, $\tilde{h}_3 \}$ is significant: $6.91\times 10^{-3}$, comprising $64\%$ of the total amplitude of $A$.
From this example, it can be seen that empty initial-state orbitals and a multi-orbital picture are crucial for understanding the brightness of near-edge transitions in metallic systems.
\renewcommand{\arraystretch}{1.5}
\begin{table}
\setlength{\tabcolsep}{3.5pt}
\begin{tabular}{ c | c c c c c c}
\hline
$\tilde{\psi}_i$
& $| \langle \tilde{\psi}_i | P_c x | \psi_h \rangle|$
& $ |\langle \tilde{\psi}_i | e_1 \rangle |$
& $ |\langle \tilde{\psi}_i | e_2 \rangle |$
& $ |\langle \tilde{\psi}_i | e_3 \rangle |$
& $ |M^f_{i}| $ & \\
\hline
$\tilde{e}_1$ & $3.79\times 10^{-3}$ & $0.30$ & $0.51$ & $0.05$ & $0.34$\\
$\tilde{h}_4$ & $9.93\times 10^{-3}$ & $0.28$ & $0.08$ & $0.24$ & $0.24$\\
$\tilde{h}_3$ & $1.23\times 10^{-2}$ & $0.27$ & $0.07$ & $0.28$ & $0.26$\\
\hline
\end{tabular}
\caption{Quantities relevant for analyzing the expansion in Eq. (\ref{eq:mat_decomposed} for state A.)}
\label{tab:A}
\end{table}
\subsection{Shake-up effects in half-metallic \ce{CrO2}}
\label{sec:more_eh}
The determinantal approach introduced in this work does not set any constraint on the number of $\emph{e}$-$\emph{h}${} pairs to be included
and is capable of considering more complex excitations than in the BSE.
Higher-order $\emph{e}$-$\emph{h}${}-pair production (so-called shake-up effects due to the core-hole perturbation) should be less costly from an energy perspective in systems with smaller band gaps, and therefore more evident in the near-edge fine structure.
This section discusses these effects for the half-metallic \ce{CrO2}, whose majority-spin channel is metallic, while the minority-spin channel is insulating.
The interplay of the two spin channels in x-ray excitations gives rise to intriguing physics that cannot be simply explained by excitonic effects.
We will discuss how the measured XAS takes shape to illustrate additional many-body effects that are captured within the determinantal approach, beyond those already highlighted above for the $f^{(1)}$ transitions.
For \ce{CrO2}, the $f^{(2)}$ XAS contribution becomes comparable to that of $f^{(1)}$ at $\sim4.0$ eV above the absorption onset (Fig. \ref{fig:results} (a)).
The $f^{(2)}$ configurations can be considered as \emph{shake-up} excitations derived from $f^{(1)}$.
Below is the composition of some major $f^{(2)}$ configurations outlined in Fig. \ref{fig:cro2_details} (a)
\begin{align}
\begin{split}
E\mapsto(\tilde{e}_1, \tilde{h}_4, \tilde{e}_2)&, \ F\mapsto(\tilde{e}_1, \tilde{h}_3, \tilde{e}_3),\\
G\mapsto(\tilde{e}_1, \tilde{h}_3, \tilde{e}_3\downarrow)&, \ H\mapsto(\tilde{e}_1, \tilde{h}_3, \tilde{e}_4)
\end{split}
\end{align}
They can be derived from the $f^{(1)}$ states by adding one more $\emph{e}$-$\emph{h}${} pair
\begin{align}
\begin{split}
E\mapsto A + (\tilde{e}_2, \tilde{h}_4)&, \ F\mapsto B + (\tilde{e}_1, \tilde{h}_3), \\
G\mapsto C + (\tilde{e}_1,\tilde{h}_3)&, \ H\mapsto D + (\tilde{e}_1, \tilde{h}_3)
\end{split}
\end{align}
where $\tilde{h}_3$, $\tilde{h}_4$, $\tilde{e}_1$, and $\tilde{e}_2$ are $t_{2g}$ orbitals close to the Fermi level.
As is shown in Fig. \ref{fig:cro2_details} (d),
orbital $\tilde{e}_1$ has significant spatial overlap with $\tilde{h}_3$ (sharing the $d_{xz}$ character at the \ce{Cr} atom next to the excited \ce{O}),
and so does orbital $\tilde{e}_2$ with $\tilde{h}_4$ (near the oxygens at the corners of the plot), albeit weaker.
This overlap makes $E$, $F$, $G$, and $H$ also bright transitions.
There are alternative pathways to access these states with two $\emph{e}$-$\emph{h}${} pairs.
For instant, $F$ can also be mapped to $A + (\tilde{e}_3, \tilde{h}_3)$, i.e., $A$ coupled with an $\emph{e}$-$\emph{h}${} pair $(\tilde{e}_3, \tilde{h}_3)$ (a shake-up $d-d$ transition).
The shake-up excitations can also be found in the satellite features of XPS, as shown in Fig. \ref{fig:cro2_details} (b).
Recently these excitations were investigated with a cumulant expansion technique \cite{kas2015real, kas2016particle}.
Here, we show that these satellite features can also be included naturally within the determinant formalism of the non-interacting MND theory (albeit poorly approximating their energies due to missing additional interactions between these extra $\emph{e}$-$\emph{h}${} pairs).
The strongest transition (labelled as state $I$) originates from the overlap of the $N$-electron states, describing the initial ground state valence system and the final core-excited valence system (assuming the excited electron has escaped, approximated using the full-core-hole approach): $\langle \Psi^N_{f, \text{FCH}} | \Psi^N_{i, \text{GS}} \rangle$.
This corresponds to the charge-transfer state in Fig. \ref{fig:cro2_details} (d).
We may define $I$ as the only zero-order configuration ($f^{(0)}$) of XPS.
$f^{(1)}$ configurations emerge at larger binding energies and appear as satellite features in the XPS profile.
Two representative states are $J$ and $K$
\begin{align}
\begin{split}
J\mapsto(\tilde{e}_1, \tilde{h}_2), \ K\mapsto(\tilde{e}_1, \tilde{h}_1)
\end{split}
\end{align}
which are shake-up excitations from $\tilde{h}_2$ (a \ce{Cr} $3d$ - \ce{O} $2p$ hybrid with mixed bonding and anti-bonding character) and $\tilde{h}_1$ (a deep \ce{O} $2p$ orbital) to the $\tilde{e}_1$ orbital, respectively.
The charge transfer associated with $K$ is particularly strong.
\subsection{Many-body wavefunction overlap effects in \ce{CrO2}}
\label{sec:spin_conv}
As shown in Sec. \ref{sec:afi}, the projection onto empty initial-state orbitals alone cannot account for the XAS lineshape for \ce{CrO2},
and one must employ the determinant formalism.
This suggests that there are important many-electron effects in the determinantal amplitude that lead to the ultimate peak-intensity ratio of $\sim1.7$ between the first and second absorption features.
To explain this, we rewrite the spectrum as the convolution defined in Eq. (\ref{eq:xas_conv})
\begin{align}
\begin{split}
\sigma_{A}&=\sigma_{A\uparrow} *\sigma_{P\downarrow} + \sigma_{A\downarrow} *\sigma_{P\uparrow}
\end{split}
\end{align}
where
$\sigma_{A}\equiv\sigma_\text{XAS}, \sigma_{P}\equiv\sigma_\text{XPS}$,
$*$ represents the convolution integral in Eq. (\ref{eq:xas_conv}),
$\sigma_{A\mu}$ and $\sigma_{P\mu}$ are spectra of one-spin channel before convolution.
Then the spectral functions $\sigma_{P\mu}$ can be considered as weighting factors of the two absorption channels $\sigma_{A\mu}$.
If the weighting factors are not considered, the hypothetical spectrum
\begin{align}
\begin{split}
\sigma'_{A}&=\sigma_{A\uparrow} + \sigma_{A\downarrow}
\end{split}
\end{align}
has a peak-intensity ratio of $\sim 1.3$ that still deviates significantly from experiment (Fig. \ref{fig:cro2_details} (c)).
This implies that the modulation effects of $\sigma_{P\uparrow}$ and $\sigma_{P\downarrow}$ on their counter-spin channel are quite different.
\begin{figure}
\centering
\includegraphics[angle=0, width=0.95\linewidth]{./xps_spin.eps}
\caption{Spin-wise spectral function $\sigma_{P\uparrow}$ and $\sigma_{P\downarrow}$.
The x-axis is the binding energy ($E_B$).
$E_B = 0$ is aligned with the threshold.
}
\label{fig:cro2_xps}
\end{figure}
The spectral functions, $\sigma_{P\uparrow}$ and $\sigma_{P\downarrow}$, are shown in Fig. \ref{fig:cro2_xps}.
In both cases, most spectral weight is concentrated at zero binding energy, $E_B = 0$.
But for the metallic $\uparrow$ channel, more spectral weight is transferred to shake-up satellites at higher energies
because its lack of a band gap makes $\emph{e}$-$\emph{h}${} pair production easier.
As a result, $\sigma_{P\downarrow}$ is more intense than $\sigma_{P\uparrow}$ near $E_B = 0$.
The integrated intensity of $\sigma_{P\uparrow}$ is $\sim 2/3$ of $\sigma_{P\downarrow}$ for $E_B < 1.7$ eV (shaded areas).
The more intense $\sigma_{P\downarrow}$ enhances the contribution of $\sigma_{A\uparrow}$,
especially the lowest-energy peak defined by $t_{2g}\uparrow$ orbitals,
leading to a peak-intensity ratio of $\sim 1.7$ as measured.
To conclude, the three contributing factors leading to the near-edge lineshape of \ce{CrO2} are:
(a) the core-level excitonic effect in the metallic screening environment lead to a mild increase in the edge intensity
(the initial-state spectrum is also shown in Fig. \ref{fig:cro2_details});
(b) shakeup excitations in the spin-up channel reduces the many-body wave function overlap $\sigma_{P\uparrow}$ at $E_B=0$;
(c) the smaller wave function overlap (orthogonality effects) reduces the intensity of the spin-down channel that mainly contributes to the second absorption feature,
leading to a even stronger first peak versus the second.
\section{Numerical considerations and computational efficiency}
\label{sec:comp}
\begin{figure*}
\centering
\includegraphics[angle=0, width=0.98\linewidth]{./numerics.png}
\caption{
\textbf{(a)}
Schematic showing the relation of the subset of initial(final)-state orbitals chosen in practical calculations to the full Hilbert space.
\textbf{(b)}
Histograms for the distribution of the eigenvalues of the square $\bm{\xi}'s$.
Counts of eigenvalues are in logarithm scale.
The bar widths (above $0.9$) are $\frac{2}{3}\times10^{-3}$.
\textbf{(c)}
$\xi$-matrices for \ce{SiO2}, \ce{TiO2}, and \ce{CrO2} $\uparrow$.
The dashed lines mark the Fermi level of the initial (vertical) and final (horizontal) state.
The right panels are the regions enclosed by the bolded squares on the left ones near the crossings of the two Fermi levels.
Within these regions, the \ce{CrO2} has large matrix elements in all four quadrants while the large matrix elements are mainly located within the $vv-$ or $cc-$block for \ce{SiO2} and \ce{TiO2}.
\textbf{(d)}
$\zeta$-matrices that correspond to the $\xi$-matrices in (c).
Rows iterates over empty-orbital indices with $1$ being the lowest empty one.
Columns iterates over occupied-orbital indices with $1$ being the lowest occupied one.
Right panels are enlarged views of the square regions in the left ones.
Both (c) and (d) display the absolute values of the complex matrix elements in logarithm scale.
}
\label{fig:numerics}
\end{figure*}
\subsection{Properties of the $\xi$-matrix}
\label{sec:xi}
One primary concern of the determinantal approach is the numerical accuracy of the $\xi$-matrix ($\bm{\xi}$).
In practice, one can only choose a finite number of orbitals (bands) in first-principles calculations and
this set of orbitals can not span the full 1p Hilbert space, as illustrated in Fig. \ref{fig:numerics} (a).
Therefore, the initial-state orbital set may not overlap with the final-state one,
resulting in a $\bm{\xi}$ that is projective rather than unitary.
Furthermore, it may be worrisome if the numerical error in the matrix elements of $\bm{\xi}$ is accumulative, leading to determinant values that are either vanishingly small or unrealistically large.
Here, we demonstrate that using the optimal basis set for expanding 1p wave functions can produce a $\xi_{ij}$ matrix close to unitary,
such that the spectral weight of the determinantal spectrum is on the same order of magnitude as the 1p final-state spectrum as compared in Sec. \ref{sec:afi}.
When constructing the Shirley optimal basis sets, we include a sufficient number of bands (Tab. \ref{tab:comp}) so that
the optimal basis functions can cover a range of 1p wave functions, from localized $3d$-orbitals to delocalized states.
We measure the quality of a transformation matrix by its eigenvalues.
A close-to-unitary transformation matrix should have eigenvalues that are close to $1$ predominantly.
Through examining $\bm{\xi}$ of the studied systems,
we find more than $90\%$ of the eigenvalues are larger than $0.995$, with a maximum below $1.0001$,
which suggests these $\bm{\xi}$'s are close to unitary.
A typical statistics of the eigenvalues of $\bm{\xi}$ using \ce{Fe2O3} and \ce{CrO2} $\uparrow$ as examples is provided in Fig. \ref{fig:numerics} (b).
The second concern regarding the practicality of the determinantal approach is how many configurations are relevant for a converged lineshape.
From the analysis of the BFS algorithm, we know that this depends on the sparsity of $\bm{\zeta}$ and how many non-vanishing minors one can extract from $\bm{\zeta}$.
We first analyze the properties of $\bm{\xi}$.
Fig. \ref{fig:numerics} (c) displays the $\bm{\xi}$ for the three representative cases,
the large-band-gap \ce{SiO2} ($690\times690$),
the semiconducting \ce{TiO2} ($800\times800$),
and the metallic spin channel ($\uparrow$) of \ce{CrO2} ($1200\times1200$).
All $\bm{\xi}$'s are \emph{quasi-block-diagonal}, which indicates the core-hole-induced hybridization mainly occurs within orbitals of similar energies.
Overall, the $\bm{\xi}$ of \ce{SiO2} and \ce{TiO2} has more off-diagonal matrix elements compared to \ce{CrO2}
because electronic screening of the core hole is weaker in an insulator/semiconductor than in a metal.
In the region near the Fermi levels, however, the $\bm{\xi}$ of \ce{SiO2} and \ce{TiO2} has less off-diagonal matrix elements than \ce{CrO2}:
for \ce{SiO2} and \ce{TiO2}, the significant matrix elements are mainly concentrated at the vv-(occupied-to-occupied) and cc-(empty-to-empty) blocks;
but for \ce{CrO2}, there are more non-vanishing matrix elements in the vc- or cv-block, especially in the vicinity of the Fermi-level crossing.
This is because the Fermi surface of a metallic system is susceptible to the core-hole potential, which strongly rehybridizes the orbitals near the Fermi surface.
$\bm{\xi}$ of \ce{SiO2} is also significantly different from those of \ce{TiO2} and \ce{CrO2}.
The distribution of nontrivial matrix elements is more homogeneous within the $vv$ and $cc$ block for \ce{SiO2} compared to \ce{TiO2} or \ce{CrO2}.
This is also consistent with the analysis with projection spectra in Sec. \ref{sec:afi}:
the conduction bands of \ce{SiO2} hybridize uniformly with the valence bands due to the core hole,
leading to very similar lineshapes in the 1p, projection, and determinantal spectrum,
whereas the $cv$-hybridization in \ce{TiO2} or \ce{CrO2} is less uniform and orbital-dependent,
leading to a few-body molecular description of x-ray excitations as in Sec. \ref{sec:afi} and \ref{sec:more_eh}.
\renewcommand{\arraystretch}{1.5}
\begin{table*}
\begin{tabular}{ c | c c | c c | c c c c}
\hline
\hline
System & Calc. $E_g$ (eV) & $|\det|$ & \#Orb. & \#Elec. & \# $f^{(2)}$ (M) & \# Significant $f^{(2)}$ (M) & Ratio ($\%$) \\
\hline
\ce{TiO2} & 1.79 & 0.8077 & 800 & 288 & 37.7 & 0.0567 & 0.15\\
\ce{SrTiO3} & 2.26 & 0.8125 & 1200 & 540 & 117 & 0.891 & 0.76\\
\ce{Fe2O3} & 1.10 & 0.7922 & 1000 & 400 & 71.9 & 0.197 & 0.27\\
\ce{VO2} & 0.00 & 0.7594 & 800 & 300 & 37.4 & 0.0345 & 0.09\\
\ce{CrO2$\uparrow$} & 0.00 & 0.3397 & 1200 & 336 & 125 & 0.242 & 0.19\\
\ce{CrO2$\downarrow$} & 3.68 & 0.8474 & 1200 & 288 & 119 & 0.282 & 0.24\\
\ce{MnO2} & 0.09 & 0.8047 & 800 & 324 & 36.6 & 0.708 & 1.9\\
\ce{NiO} & 3.33 & 0.8122 & 500 & 256 & 7.59 & 0.0812 & 1.1\\
\ce{CuO} & 0.12 & 0.4871 & 1024 & 544 & 62.5 & 0.604 & 0.97\\
\ce{SiO2} & 6.19 & 0.8370 & 690 & 192 & 23.7 & 0.179 & 0.75\\
\hline
\hline
\end{tabular}
\caption{
(Initial-state) band gaps $E_g$ obtained on the DFT ($+U$) level;
the absolute values of the determinants for the transformation matrix from the initial to final state for the $N$-electron systems, i.e., $|\langle\Psi^N_i | \Psi^N_f \rangle|(|\det|)$, of individual spin channels without the photoelectron;
numbers of all orbitals and those of the occupied ones;
numbers of all $f^{(2)}$ configurations and the prominent ones that contribute to converged lineshapes and their proportions among the whole;
for the systems being studied;
}
\label{tab:comp}
\end{table*}
\subsection{Properties of the $\zeta$-matrix}
\label{sec:zeta}
Consider an ideal situation where there is no hybridization induced between the occupied and empty orbitals as the core-hole potential is introduced.
The $\xi$-matrix is exactly block diagonal and $\bm{\zeta}$ only has non-zero matrix elements in its last column.
The actual $\zeta$-matrix can be considered as a deviation from this ideal situation.
How much it deviates depends on the hybridization of the occupied and empty orbitals.
Fig. \ref{fig:numerics} (d) displays the $\zeta$-matrices for \ce{SiO2}, \ce{TiO2}, and \ce{CrO2} $\uparrow$,
for the region that spans the lowest 170 unoccupied orbitals (rows) and the topmost 16 occupied orbitals plus the lowest unoccupied orbital (columns).
Near the Fermi levels, the $\xi$-matrices of \ce{SiO2} and \ce{TiO2} are quasi-block-diagonal, which leads to a $\zeta$-matrix with significant matrix elements mainly located on its last column.
There are relatively a small number of non-vanishing $2\times2$ or high-order minors, and
therefore the XAS converges mostly at the $f^{(1)}$ order. We can also see that the more uniform, reduced coupling between occupied and unoccupied orbitals in \ce{SiO2} leads to a $\zeta$-matrix with a more dominant final column. By contrast, the hybridization across the band gap in \ce{TIO2} exhibits less uniformity, reflecting the existence of more localized orbitals subspaces affected by the core-hole potential, and the corresponding $\zeta$-matrix exhibits more significant terms outside the final column, indicating that the many-body approach may be more accurate for \ce{TiO2}.
For \ce{CrO2} with strong hybridization, $\bm{\zeta}$ has more significant matrix elements beyond the last column.
These matrix elements form several strips with widths of a few columns, leading to more nontrivial high-order minors.
\subsection{Computational overhead}
\label{sec:overhead}
The computational complexity of the BFS depends on how many nontrivial minors can be found from $\bm{\zeta}$.
A statistics of the computational effort required to converge XAS is shown in Table. \ref{tab:comp}.
The XAS is simulated with a supercell with dimensions around 10 \AA{} and several hundred ($N_v$) electrons.
To cover an energy window up to $20$ eV above onset, another few hundred ($N_c$) empty orbitals are also included.
Since the investigated XAS converges at the $f^{(2)}$ order, we use the number of nontrivial $f^{(2)}$ configurations as a measure of the computational costs.
There are $N_c (N_c - 1) N_v / 2$ $f^{(2)}$ configurations in total,
whose numbers are from tens to hundreds of millions for the investigated systems.
The number of the nontrivial $f^{(2)}$ configurations as found by the BFS algorithm is typically around $1\%$ of the total.
In all of the investigated systems, this translates to at least a 100-fold speed-up of calculations, thanks to the BFS algorithm that screens out configurations of weak transition amplitudes.
For insulators such as diamond or \ce{TiO2}, even fewer configurations are needed to achieve convergence.
The overall trend for the computational cost is: the smaller the band gap ($E_g$), the more the valence orbitals tend to hybridize with the empty orbitals (due to the core-hole potential), the smaller the determinant of the overlap matrix between the initial- and final-state ($\langle\Psi^N_i | \Psi^N_f \rangle$), and more configurations and computational efforts are required.
\section{Conclusions and Outlooks}
In conclusion, we have implemented an efficient algorithm for simulating x-ray absorption spectra (XAS) employing transition amplitudes computed within a many-body determinantal ansatz.
The core of the algorithm exploits the linear dependence of the determinants representing various electronic configurations for a fixed number of electrons and
a breadth-first search (BFS) graph algorithm that efficiently and controllably neglects configurations whose contributions are insignificant to computed XAS, as defined by some numerical tolerance.
The new methodology has been applied to study a series of transition metal oxides (TMOs), and this simulation technique can be readily used for interpreting XAS of these technologically important materials. In the majority of cases, this approach provides an accuracy comparable to or exceeding Bethe-Salpeter equation (BSE) solutions and naturally includes electronic confgurations representing higher-order excitations beyond the subset of Feynman diagrams accessible within the BSE.
The determinantal approach can be extended to other types of x-ray spectra besides XAS, such as X-ray photoemission spectroscopy (XPS) and resonant inelastic x-ray scattering (RIXS),
using a similar linear algebra technique and search algorithm.
It will be worthwhile to compare this new method with recent studies that apply a cumulant expansion to capture the charge-transfer satellites in XPS \cite{kas2015real, kas2016particle, zhou2017cumulant}.
And it will be interesting to test the efficiency of the current approach to produce 2D RIXS spectra that provide rich information for materials characterization.
The main drawback of the current approach relates to its approximation of the various final state configurations, which are currently derived from a single (core-orbital excited-state) self-consistent field and its associated valence Kohn-Sham orbitals. The spectrum of excitation energies within this orbital space neglects additional valence-orbital excited-state electron interactions. Therefore it cannot describe further excitonic final-state effects resulting from the shake-up of additional valence electron-hole pairs nor coupling with many-body collective modes, such as plasmon excitations. These effects can be captured within the cumulant expansion through accurate determination of the valence dielectric response function beyond the random-phase approximation.
However, this is an excellent approximation for higher-order contributions to the spectra of metallic or semi-metallic systems, as demonstrated here for CrO$_2$, and future work will explore solutions for an interacting picture to refine our description of higher-order excited states of semiconductors and insulators and their associated spectral features.
\label{sec:outlook}
\begin{appendix}
\section{PAW formalism for obtaining the overlap matrix elements}
\label{sec:paw}
To obtain the transition amplitude $A^f_c$, a prerequisite is to find the overlap integral between the initial- and final-state Kohn-Sham orbitals, i.e., the matrix elements $\xi_{ij} = \langle\psi_j |\tilde{\psi}_i \rangle$.
In our implementation of the $\Delta$SCF calculations, we employ a plane-wave basis and the electron-ion interaction is modeled using Vanderbilt's ultrasoft pseudopotentials.
The computational efficiency gain through the use of a smaller plane-wave energy cutoff compared to what might be required when using norm-conserving pseudopotentials is offset by some additional steps in the formalism which account for using non-orthogonal projections in the pseudopotential.
In the above calculations with the many-electron method, we have used the PAW formalism to find the overlap matrix elements $\xi_{ij}$ and here we provide the details for finding these quantities.
In the PAW formalism, the real (all-electron, AE) wave function is reconstructed from the pseudo (PS) wave function via a linear transformation $\mathcal{T}$
\begin{align}
\begin{split}
|\psi^{\text{AE}} \rangle= \mathcal{T}|\psi^\text{PS}\rangle
\end{split}
\end{align}
In practice, there is one such $\mathcal{T}$ for each pseudized atom. To simplify notation, we will omit the sum over atomic indices, $I$, for most of what follows, until it is necessary to the discussion.
$\mathcal{T}$ is responsible for correcting the wave function within the augmented spherical region $\Omega$ centered at the atom of interest.
First, $\mathcal{T}$ projects the pseudo wave function onto the preselected projectors $|p_l\rangle$ of a particular angular momentum $l$;
then $\mathcal{T}$ corrects the wave function in the augmented region using the difference of the real and pseudo atomic wave functions of the corresponding $l$, i.e., $|\phi_l^\text{AE}\rangle - |\phi_l^\text{PS}\rangle$, and scales the wave function difference with the projection amplitude.
$|\phi_l^\text{AE}\rangle$ and $|\phi_l^\text{PS}\rangle$ and the associated projectors are all determined when generating the pseudopotential.
Put together, the linear transformation reads
\begin{align}
\begin{split}
\mathcal{T}=1 + \sum_{l}(|\phi_l\rangle-|\phi^\text{PS}_l\rangle)\langle p_l |
\end{split}
\end{align}
For simplicity, \quotes{AE} is dropped and only \quotes{PS} is kept to indicate a wave function is pseudo.
A PAW construction satisfies the following conditions:
(i) the projector functions $\langle \bm{r} | p_l \rangle$ are zero outside the augmented region $\Omega$;
(ii) the difference of the atomic wave functions of the same $l$, i.e., $\langle \bm{r} | \phi_l \rangle - \langle \bm{r} | \phi^\text{PS}_l \rangle$, is also zero outside $\Omega$;
(iii) and we have an orthogonality and completeness relation: $\langle \phi^\text{PS}_i | p_j \rangle = \delta_{ij}$, for all $i$ and $j$, and $ P=\sum_l | \phi^\text{PS}_l\rangle \langle p_l| $ is the identity operator over $\Omega$.
It should be noted that in the PAW formalism each angular momentum may have more than one channel so $\langle \phi_l | \tilde{\phi}_{l'} \rangle = \delta_{ll'}$ may not hold in general.
As stated, there is one such linear transformation $\mathcal{T}$ for each type of atom (i.e., for each element) and the projections should include a structure (phase) factor to account for different atomic positions within an extended, periodic context.
In the x-ray core-hole approach, however, we introduce a new type of atom. We have, as before, the initial-state (ground-state) atoms and one new type to describe the final-state atom with an excited core hole.
In practice, this means that there are two sets of projectors and atomic wave functions involved for this particular atom.
If one wants to obtain the overlap matrix elements $\xi_{ij}$, it is necessary to obtain overlap integrals of two wave functions that are reconstructed from two different PAW constructions.
Here, we focus on the single-atom case and find the expression for the overlap.
Consistent with the notation in the rest of the manuscript, we use a tilde to denote quantities related to the final (excited) state.
Omitting the irrelevant indices, the overlap between an initial-state and a final-state orbital is
\begin{align}
\begin{split}
\langle \psi | \tilde{\psi} \rangle = \langle \psi^\text{PS} | \mathcal{T}^\dagger \tilde{\mathcal{T}} | \tilde{\psi}^\text{PS} \rangle
\end{split}
\label{eq:psipsi}
\end{align}
where $\psi$ and $\tilde{\psi}$ are reconstructed from two different linear transformations, $\mathcal{T}$ and $\tilde{\mathcal{T}}$.
Expanding the operator product, we find
\begin{align}
\begin{split}
&\mathcal{T}^\dagger \tilde{\mathcal{T}} \\
&= 1
+ \sum_{l} | p_l \rangle ( \langle \phi_l | - \langle \phi^\text{PS}_l | )
+ \sum_{l'} (| \tilde{\phi}_{l'} \rangle - | \tilde{\phi} ^ \text{PS}_{l'} \rangle)\langle \tilde{p}_{l'} | \\
&+ \sum_{ll'} | p_l \rangle ( \langle \phi_l | - \langle \phi^\text{PS}_l | ) (| \tilde{\phi}_{l'} \rangle - | \tilde{\phi} ^ \text{PS}_{l'} \rangle)\langle \tilde{p}_{l'} |
\end{split}
\label{eq:tt}
\end{align}
This expansion can be regrouped and simplified by making use of the properties of the projectors and PAW atomic wave functions in conjunction with the completeness relation.
First, the last summation in Eq. (\ref{eq:tt}) can be further expanded so as to make use of the projection operators $ P=\sum_l | \phi^\text{PS}_l\rangle \langle p_l| $ and $\tilde{P}$, as follows:
\begin{align}
\begin{split}
&\sum_{ll'} | p_l \rangle ( \langle \phi_l | - \langle \phi^\text{PS}_l | ) (| \tilde{\phi}_{l'} \rangle - | \tilde{\phi} ^ \text{PS}_{l'} \rangle) \langle \tilde{p}_{l'} | \\
=&\sum_{ll'} | p_l \rangle ( \langle \phi_l | \tilde{\phi}_{l'} \rangle - \langle \phi^\text{PS}_l | \tilde{\phi} ^ \text{PS}_{l'} \rangle ) \langle \tilde{p}_{l'} | \\
- &\sum_{l} | p_l \rangle ( \langle \phi_l | - \langle \phi^\text{PS}_l | ) \big( \sum_{l'} | \tilde{\phi} ^ \text{PS}_{l'} \rangle \langle \tilde{p}_{l'} | \big) \\
- &\big( \sum_{l} | p_l \rangle \langle \phi^\text{PS}_l | \big) \sum_{l'} (| \tilde{\phi}_{l'} \rangle - | \tilde{\phi} ^ \text{PS}_{l'} \rangle)\langle \tilde{p}_{l'} |
\end{split}
\label{eq:last_sum}
\end{align}
The last two terms can be regrouped with the two single-summations over $l$ and $l'$ in Eq. (\ref{eq:tt}).
For instance, the second summation in Eq. (\ref{eq:last_sum}) can be combined with the second term in Eq. (\ref{eq:tt}) as
\begin{align}
\begin{split}
\sum_{l} | p_l \rangle ( \langle \phi_l | - \langle \phi^\text{PS}_l | ) (1 - \tilde{P})
\end{split}
\label{eq:proj}
\end{align}
While the operator $| p_l \rangle ( \langle \phi_l | - \langle \phi^\text{PS}_l | ) $ is only non-zero within $\Omega$, $(1 - \tilde{P})$ projects the wave function onto the region complementary to $\tilde{\Omega}$.
Therefore, we can consider the union of the augmented regions for the ground-state and the core-excited atom, $\Omega \cup \tilde{\Omega}$ as a volume within which the product of these operators will zero out any wave function, and the operator in Eq. (\ref{eq:proj}) is a zero operator.
In practice, we can set the radial limit for atomic integrals, like $\langle \phi | \tilde{\phi} \rangle$ to the maximum of the cutoff radii used when generating the pseudopotentials for the ground-state and the core-excited atom. More often than not, these cut-off radii are identical and $\Omega = \tilde{\Omega}$.
And so, the second and third summations in Eq. (\ref{eq:last_sum}) cancel exactly with the second and third terms in Eq. (\ref{eq:tt}).
With all terms combined, the final expression for the operator product is simplified as
\begin{align}
\begin{split}
\mathcal{T}^\dagger \tilde{\mathcal{T}} = 1 + \sum_{ll'} | p_l \rangle ( \langle \phi_l | \tilde{\phi}_{l'} \rangle - \langle \phi^\text{PS}_l | \tilde{\phi} ^ \text{PS}_{l'} \rangle ) \langle \tilde{p}_{l'} |
\end{split}
\label{eq:tt_final}
\end{align}
In a multi-atomic system, the overlap matrix elements in Eq. (\ref{eq:psipsi}) can be written as
\begin{align}
\begin{split}
\langle \psi | \tilde{\psi} \rangle
&= \langle \psi^\text{PS} | \tilde{\psi}^\text{PS} \rangle \\
&+ \sum_{I, ll'} \langle \psi^\text{PS} | p^I_l \rangle ( \langle \phi^I_l | \tilde{\phi}^I_{l'} \rangle - \langle \phi^{I,\text{PS}}_l | \tilde{\phi} ^ {I,\text{PS}}_{l'} \rangle ) \langle \tilde{p}^I_{l'} | \tilde{\psi}^\text{PS} \rangle
\end{split}
\label{eq:psipsi_final}
\end{align}
in which the index $I$ goes over all the PAW atoms in the system.
Here, we only consider one core-excited atom within a given supercell, and so, for all but one of the atoms, the initial and final state PAW projections are identical (i.e., we can drop the tildes).
The first term in Eq. (\ref{eq:psipsi_final}) can be obtained efficiently using the pseudo wave functions in their native plane-wave basis.
The routines to evaluate the projection amplitude $\langle \psi^\text{PS} | p^I_l \rangle$ are already required to obtain the core-level position matrix operator at the core-excited atom (as we have done in the past for the one-body final state approach).
The same procedure can be trivially extended to estimate projection amplitudes for all atoms and for both the initial and final state, using outputs from the pseudopotential generation.
An additional routine is needed for the atomic overlap term $S^I_{ll'} \equiv \langle \phi^I_l | \tilde{\phi}^I_{l'} \rangle - \langle \phi^{I,\text{PS}}_l | \tilde{\phi} ^ {I,\text{PS}}_{l'} \rangle$,
which can be obtained beforehand using the atomic wave functions from two given PAW constructions. All of these quantities can be computed and stored in advance for an established set of pseudopotentials and then used for any number of further periodic calculations.
\section{Optimal Basis Set for Obtaining Electronic Structure over Dense $\bm{k}$-Grid}
\label{sec:obf}
Generating electronic states over a dense enough $\bm{k}$-grid within the first Brillouin zone (BZ) is an essential step for producing continuous spectral functions that respect the continuity in the electronic density of states.
This is particularly important for simulating X-ray absorption spectra, especially when excited states extend into the continuum, either beyond the ionization potential in a non-infinite system or into the Bloch-periodic states of extended periodic systems.
Although setting up a supercell for simulating XAS is equivalent to using some $\bm{k}$-point sampling over the BZ of the primitive unit cell, generating a $\bm{k}$-grid on the top of the supercell setups in some occasions does further improve the quality of simulated spectra, particularly at higher energies.
Similarly, a metallic system may have a large number of extended states near the Fermi level, which need to be included to accurately reproduce the near-edge fine structure.
The number of extended states is proportional to and limited to the size of the supercell that can be realistically simulated in the $\Delta$SCF core-hole calculation.
In this circumstance, using $\bm{k}$-point sampling over the supercell BZ may partially compensate for the disadvantage of using a supercell that is not quite large enough. However, $\bm{k}$-point sampling will not correct for a model of the final state within which the charge-density response to the core-excited state has not sufficiently converged within the supercell.
Previously, we have studied, implemented, and tested an efficient calculation scheme for obtaining band structure on a dense $\bm{k}$-grid.
By employing so-called optimal basis sets \cite{shirley1996optimal, prendergast2009bloch}, one can first generate the band structure on a coarse $\bm{k}$-grid and then reproduce band energies and wave functions at any $\bm{k}$-point with much less computational effort.
The optimal basis set is the minimal basis for representing the periodic components of Bloch-waves across the BZ, constructed by removing linear dependence between these vectors (with the assumption that these functions vary smoothly throughout the BZ).
Similar to plane-wave basis sets or maximally-localized Wannier functions \cite{marzari1997maximally}, the optimal basis functions, denoted as $\{B_i\}$, can be used to expand a Bloch-periodic wave function $|\psi_k\rangle=e^{ik \cdot r}|u_k\rangle$, in terms of its periodic component: $\langle B_i | u_k \rangle$, but are not limited to extended or localized states, no more than the actual Kohn-Sham orbitals themselves.
Moreover, the number of optimal basis functions required is much smaller than the number of plane-waves for expanding these orbitals.
The size of a good optimal basis set ranges from $10^3$ to $10^4$, which can be at easily $1000$ times smaller than the plane-wave basis set of equivalent accuracy. The energies and eigenstates at a given k-point are obtained from diagonalization of a representation of the original (k-dependent) Kohn-Sham Hamiltonian in this much smaller basis.
Now we revisit the quantities needed for computing the overlap matrix elements in Eq. (\ref{eq:psipsi_final}),
$ \langle \psi^\text{PS} | \tilde{\psi}^\text{PS} \rangle $ and $ \langle p^I_{l} | \psi^\text{PS} \rangle (\langle \tilde{p}^I_{l} | \tilde{\psi}^\text{PS} \rangle) $,
which will benefit greatly from using optimal basis sets.
First, the pseudo overlap matrix element (carried out at every k-point independently) can be computed as
\begin{align}
\begin{split}
\langle \psi^\text{PS}_{nk} | \tilde{\psi}^\text{PS}_{mk} \rangle = \sum_{ij} \langle u^\text{PS}_{nk} | B_i \rangle \langle B_i | \tilde{B}_j \rangle \langle \tilde{B}_j | \tilde{u}^\text{PS}_{mk} \rangle
\end{split}
\end{align}
where $\langle u^\text{PS}_{nk} | B_i \rangle$ ($\langle \tilde{B}_j | \tilde{u}^\text{PS}_{mk} \rangle$) are the eigensolutions (Hermitian conjugates) of the k-dependent Hamiltonian in their corresponding optimal bases. Although each optimal basis set is constructed to be orthonormal, $\langle B_i | B_j \rangle = \delta_{ij}$ and $\langle \tilde{B}_i | \tilde{B}_j \rangle = \delta_{ij}$, note that the k-independent overlap matrix is not, in general: $\langle B_i | \tilde{B}_j \rangle \neq \delta_{ij}$, because we employ different optimal basis sets to represent initial- and final-state systems. We could in principle employ a sub-optimal basis to describe both systems, but it has not been attempted here.
Although optimal basis functions themselves are represented in a plane-wave basis, $\{G_i\}$, the relatively expensive calculation,
$\langle B_i | \tilde{B}_j \rangle = \sum_{i'} \langle B_i | G_{i'} \rangle\langle G_{i'} | \tilde{B}_j \rangle $,
only needs to be computed once, and the matrix is universally applicable to any $\bm{k}$-point.
Similarly, $ \langle p^I_{l} | \psi^\text{PS} \rangle $ can be obtained by inserting the optimal basis set, in the same manner used to construct the same projectors in the non-local pseudopotential within the Hamiltonian.
This procedure has been implemented in the one-body core-hole approach and it simply needs to be extended to all atoms in the system.
\section{Spectral Convolution Theorem}
\label{sec:conv}
In practice, we may encounter a situation where a many-electron system can be factorizable into subsystems that are not entangled with each other, and inter-system transitions are forbidden.
For example, in a system where electron spins are collinear, and each electron can be associated with either a spin-up or -down state, then a many-body transition operator which can be similarly partitioned cannot induce transitions from the spin-up subsystem to the spin-down subsystem. The many-body dipole operator, which is the sum of one-body dipole operators, behaves in this way, and so, light-induced transitions of spin-collinear systems (within the dipole approximation) cannot effect spin cross-over.
In general, if a spectrum reflects a multidimensional integral over a function factorizable for each independent variable (or, equivalently, over some partitioning of the same space), then we can take advantage of the spectral convolution theorem. For two independent variables, $x$ sampling subsystem $A$ and $y$ sampling subsystem $B$, suppose $f(x,y)=f_A(x) f_B(y)$, and we define a spectral function
\begin{align}
\begin{split}
\sigma(E) & = \int \int f(x,y) \delta(E-(x+y)) dx dy \\
& = \int \left( \int f_A(x) \delta((E-y)-x) dx \right) f_B(y) dy \\
& = \int \sigma_A(E-y) f_B(y) dy \\
& = \int \sigma_A(E-E') \sigma_B(E') dE'
\label{eq:specconv}
\end{split}
\end{align}
where we have just changed variable ($E'=y$) in the last line and defined the following subsystem spectral functions for each subset $I$ in the partition $\{A, B\}$:
\begin{align}
\begin{split}
\sigma_I(E)=\int_I f_I(E') \delta(E-E')
\end{split}
\end{align}
The general case, for many subsystems $I$ in $\{I_n\}$ can be written as a set of nested integrals over each subsystem,
\begin{align}
\begin{split}
\sigma(E)= & \int dE_1 \sigma_{I_1}(E-E_1) \\
& \times \int dE_2 \sigma_{I_2}(E_1-E_2) \\
& \dots \times \int dE_n \sigma_{I_{n-1}}(E_{n-1}-E_n) \sigma_{I_n} (E_n)
\label{eq:nestedconv}
\end{split}
\end{align}
Let us assume that the many-body wave functions are factorizable and limit our discussion to two subsystems, $A$ and $B$, so that $|\Psi\rangle = |\Psi^A\rangle \otimes |\Psi^B\rangle$.
Then the transition amplitude can be factorized by considering final states where the transition probes each subsystem at a time, assuming that the transition operator can also be partitioned, for example, $\mathcal{O} = \sum_{i \in A} \mathcal{O}_i + \sum_{j \in B} \mathcal{O}_j$. Note that, in practice, if the symmetry of the system causes final states in different subsystems to be distinct, we should do a separate $\Delta$SCF calculation to define each final state orbital subspace. Here, let us focus on the components of the transition operator which act directly on subsystem $A$, inducing a many-body response in subsystem $B$, and index each final state by similarly partitioning the orbital configuration vector: $f = (f^A, f^B)$ (same for $i$) as follows:
\begin{align}
\begin{split}
\langle \Psi_f |\mathcal{O}|\Psi_i\rangle = \langle \Psi_{f^A} |\mathcal{O}|\Psi_{i^A}\rangle
\langle \Psi_{f^B} | \Psi_{i^B}\rangle
\end{split}
\end{align}
Then the total spectrum can be written using Eq.~\ref{eq:specconv}, but recognizing a subtle difference between the subset spectral functions:
\begin{align}
\begin{split}
\sigma^A(E) &= \sum_{f_A} |\langle \Psi^A_{f_A} |\mathcal{O}|\Psi^A_i\rangle|^2 \delta(E - \Delta E_{f_A})
\end{split}
\end{align}
this includes the transition operator, while
\begin{align}
\begin{split}
\sigma^B(E) &= \sum_{f_B} |\langle \Psi^B_{f_B} | \Psi^B_i\rangle|^2 \delta(E - \Delta E_{f_B})
\end{split}
\end{align}
reflects the response of subsystem $B$ to the excitation in $A$.
$\Delta E_f = \Delta E_{f_A} + \Delta E_{f_B}$ is the energy required to make the transition.
This theorem is particularly useful for combining spectra from opposite spin orientations and different k-points by performing each calculation separately.
In the many-electron formalism, the size of determinants for each subsystem is much smaller than the determinants for the entire system with spins taken into account, and hence one can compute a spectrum for each subsystem at much lower memory cost and time complexity and then obtain the resulting total spectrum via the nested spectral convolution outlined in Eq.~\ref{eq:nestedconv}.
\end{appendix}
|
\section{Introduction}
An interesting and fundamental question in plasma physics is how
thermal equilibrium is determined in an essentially collisionless plasma. In such a system,
there is no reason \textit{a priori} to assume that the relative temperatures of the
component species will be equal: On the contrary, there are numerous collisionless
heating mechanisms that have been identified - shocks, magnetic reconnection,
cyclotron resonance~\citep{cranmerApJ99}, and various forms of turbulent heating~\citep{quataertApJ99a,howesJGR08,chandranApJ10a,barnesPRL12}, to name a few - and
each of these mechanisms may drive temperature separation instead of equilibration.
Indeed, evidence from observations of collisionless space and astrophysical plasmas, e.g.,
the solar wind and accretion flows onto compact objects,
suggests that more massive charged particles may be preferentially heated; cf.
~\citep{schmidtGRL80,collierGRL96,kohlSP97,quataertApJ99a}.
A thorough treatment of this problem would require exploration of a parameter space
spanning a wide range of plasma $\beta$, electron-ion temperature ratio, energy injection
mechanisms, etc.,
while including all potentially relevant physics present from the macroscopic to microscopic
space-time scales. As such a study is infeasible, we choose to focus on a single heating
mechanism - turbulent Joule heating - in a simplified
system: a two-component plasma immersed in a straight, homogenous magnetic field
with a cross-field density gradient. This system supports electron drift waves and is
known to be susceptible to the so-called `universal' instability~\citep{galeevJETP63,krallPoF65,landremanPRL15,helanderPoP15}.
The universal instability serves as an energy injection mechanism for turbulence at the ion
Larmor scale that gives rise to plasma fluctuations that drive cross-field
transport and Joule heating (or cooling)
of the component plasma species. For sufficiently infrequent collisions, this turbulent
heating and transport determines the thermal equilibrium of the plasma.
To guide our study, we note that the universal instability is derived analytically in the limit of disparate ion and electron thermal speeds.
As noted in Ref.~\citep{helanderPRL14}, the universal instability is eliminated for an equal
temperature pair plasma, or, more generally, when the ion and electron thermal speeds are equal.
In the following sections we analytically and numerically calculate the dependence of the
universal instability and resultant turbulent heating on electron-ion temperature ratio
for two cases: pair plasmas and conventional mass ratio plasmas. The former case is relevant
for proposed experiments in which
a small population of positrons is to be added to a pure electron plasma of a different temperature~\citep{pedersenPRL02};
the latter may provide insight into how astrophysical plasmas can achieve disparate electron and ion
temperatures.
We introduce the gyrokinetic model used for the analysis
in Sec. 2, and we calculate linear growth rates and quasilinear heating and cross-field flux
estimates in Sec. 3 before concluding.
\section{Model system}
We consider a collisionless plasma immersed in a straight, homogeneous magnetic field
$\mathbf{B}=B\hat{\mathbf{z}}$
and with a fixed density gradient perpendicular to the field in the $x$-direction. The fixed density
gradient could be, e.g., the result of gravitational equilibrium in an astrophysical plasma or of
an external particle source in a laboratory plasma.
We restrict our attention to electrostatic fluctuations whose frequency is small compared
with the Larmor frequency and adopt the gyrokinetic ordering:
\begin{equation}
\epsilon \doteq \frac{\omega}{\Omega_s} \sim \frac{k_{\parallel}}{k_{\perp}} \sim \frac{\rho_s}{L_n} \sim \frac{\delta f_s}{f_s} \sim \frac{q_s\varphi}{T_s}\ll 1,
\end{equation}
where $\varphi$ is the electrostatic potential fluctuation, $f_s$ is the distribution function for species $s$, $\delta f$ is the fluctuating
component of $f$, $\omega$ is the characteristic frequency of the fluctuations,
$k_{\parallel}$ and $k_{\perp}$ are the associated wavenumbers along and across the mean field,
$\Omega_s$ is the Larmor frequency for species $s$, $\rho_s$ its thermal Larmor radius,
$T_s$ its temperature, $q_s$ its charge, and $L_n$ is the mean density gradient scale
length.
Applying these orderings to the Fokker-Planck equation and averaging over the rapid gyration
of particles about the mean magnetic field results in the gyrokinetic equation,
\begin{equation}
\pd{h_s}{t} + v_{\parallel}\pd{h_s}{z} + \frac{c}{B}\{\gyroR{\varphi},h_s\} = -q_s \pd{\gyroR{\varphi}}{t}\pd{F_{0,s}}{E_s} + \frac{c}{B}\pd{\gyroR{\varphi}}{y}\pd{F_{0,s}}{x} + C[h_s],
\label{eqn:gke}
\end{equation}
where $h_s = \delta f_s - q_s\varphi (\partial F_{0,s}/\partial E_s)$ is the non-Boltzmann piece of $\delta f_s$, $E_s = m_s v^2 / 2$ is the kinetic energy of species $s$, $m_s$ is species mass,
$t$ is time, $v_{\parallel}$ is the parallel component of the particle velocity, $c$ is the speed of light,
$F_{0,s}$ is the mean component of $f_s$,
$\{\}$ indicates a Poisson bracket, $\gyroR{\varphi}$ is the
average of $\varphi$ over Larmor angle at fixed guiding center position $\mbf{R}_s$,
and $C[h_s]$ represents the effect of Coulomb collisions on species $s$.
The gyrokinetic system is closed by coupling to Poisson's equation:
\begin{equation}
4\pi\sum_s q_s \int d^3 v \delta f_s = -\nabla_{\perp}^2 \varphi,
\label{eqn:poisson}
\end{equation}
with $\nabla_{\perp}^2 = \partial_x^2 + \partial_y^2$. If the
Debye length $\lambda_s \doteq (T_s/4\pi n_s q_s^2)^{1/2}$ is much smaller than
the electron Larmor radius, the righthand side of Eq.~(\ref{eqn:poisson}) can be neglected.
In this limit Poisson's equation reduces to the quasineutrality constraint that the total
charge density of the plasma is zero. We note that a non-relativistic treatment of finite Debye
length effects requires a small plasma beta, $\beta_s=8\upi p_s/B^2 \lesssim (\vths{s}/c)^2$, consistent
with the electrostatic approximation we employ.
\subsection{Turbulent heating and heat transport}
By taking fluid moments of the Fokker-Planck equation and closing the system with the gyrokinetic
ordering, one obtains an equation for the slow
evolution of the mean temperature $T_{s}$:
\begin{equation}
\frac{3}{2}n_s\frac{dT_s}{dt} + \pd{Q_s}{x} = H_s,
\label{eqn:pressure}
\end{equation}
where the cross-field turbulent heat flux $Q_s$ and turbulent heating $H_s$ are given by
\begin{equation}
Q_s = \frac{1}{V}\int d^3 R\int_{\mbf{R}} d^3 v \left(\frac{m_s v^2}{2}-\frac{3}{2}T_s\right)\overline{\left(h_s v_E\right)}
\label{eqn:qflx}
\end{equation}
and
\begin{equation}
H_s = \frac{1}{V}\int d^3 R\int_{\mbf{R}} d^3 v q_s \overline{\left(h_s \pd{\gyroR{\varphi}}{t}\right)}.
\label{eqn:xchange0}
\end{equation}
Here $V$ is the volume of the region over which the spatial integration is performed,
the subscript $\mbf{R}$ on the velocity integration indicates that it is carried out at
fixed guiding center, $v_E=-(c/B)\partial\varphi/\partial y$ is the x-component of the $E\times B$ drift
velocity, and the overline indicates an average over time scales long compared
to the fluctuation time $1/\omega$ but short compared to the equilibrium time scale.
Note that we have neglected collisional temperature equilibration
as we are considering systems with collisional mean free path much longer than
any other scales of interest.
An alternative expression for $H_s$ is obtained by integrating Eq.~(\ref{eqn:xchange0}) by parts in
time~\citep{candyPoP13}:
\begin{equation}
H_s = \frac{1}{2V}\int d^3 R\int_{\mbf{R}} d^3 v q_s \overline{\left(h_s \pd{\gyroR{\varphi}}{t}-\pd{h_s}{t} \gyroR{\varphi}\right)}.
\label{eqn:xchange}
\end{equation}
Substitution of Poisson's equation~(\ref{eqn:poisson}) in Eq.~(\ref{eqn:xchange})
immediately indicates that the net (species-summed) turbulent heating is zero in the absence of an external energy injection mechanism; i.e., for
a two-component plasma, $\overline{H_i} = -\overline{H_e}$. If trace minority ions are present, one can obtain mass- and charge-dependent scalings for their turbulent heating rates relative to the main ions~\citep{barnesPRL12}. Expressing Eqs.~(\ref{eqn:qflx}) and~(\ref{eqn:xchange}) in terms of Fourier modes with $h_s = \sum_{\mbf{k}} \tilde{h}_{\mbf{k},s}(v_{\parallel},v_{\perp},t) \exp(\textnormal{i}\mbf{k}\cdot\mbf{R}_s)$ gives
\begin{equation}
Q_s = \textnormal{i}\frac{c}{B}\sum_{\mbf{k}} k_y\int d^3 v J_0(\alpha_{k_{\perp},s}) \frac{m_s v^2}{2}\overline{\left(\tilde{h}_{\mbf{k},s} \tilde{\varphi}_{\mbf{k}}^*\right)} \doteq \sum_{\mbf{k}}\tilde{Q}_{\mbf{k},s}
\label{eqn:qflxk}
\end{equation}
and
\begin{equation}
H_s = \frac{q_s}{2}\sum_{\mbf{k}} \int d^3 v J_0(\alpha_{k_{\perp},s}) \overline{\left(\pd{\tilde{\varphi}_{\mbf{k}}^*}{t} \tilde{h}_{\mbf{k},s} - \tilde{\varphi}_{\mbf{k}}^* \pd{\tilde{h}_{\mbf{k},s}}{t}\right)} \doteq \sum_{\mbf{k}} \tilde{H}_{\mbf{k},s},
\label{eqn:xchangek}
\end{equation}
where $\alpha_{k_{\perp},s} \doteq k_{\perp} v_{\perp}/\Omega_s$ and $J_0$ is a Bessel function of the first kind.
\subsection{Connection to particle transport}
The effect of cross-field particle transport is encapsulated in the continuity equation:
\begin{equation}
\pd{n_s}{t} + \pd{\Gamma_s}{x} = 0,
\label{eqn:continuity}
\end{equation}
where $n_s$ is particle density and
\begin{equation}
\Gamma_s = \frac{1}{V}\int d^3 R\int_{\mbf{R}} d^3 v \overline{\left(h_s v_E\right)}
\label{eqn:pflx}
\end{equation}
is the cross-field particle flux.
Expanding $\varphi$ and $h$ in terms of Fourier modes gives
\begin{equation}
\Gamma_s = \textnormal{i}\frac{c}{B}\sum_{\mbf{k}} k_y\int d^3 v J_0(\alpha_{k_{\perp},s}) \overline{\left(\tilde{h}_{\mbf{k},s} \tilde{\varphi}_{\mbf{k}}^*\right)} \doteq \sum_{\mbf{k}}\tilde{\Gamma}_{\mbf{k},s}.
\label{eqn:pflxk}
\end{equation}
The particle flux $\Gamma$ can be related to the turbulent heating of Eq.~(\ref{eqn:xchange0}) by
multiplying the gyrokinetic equation~(\ref{eqn:gke}) by $\varphi$ and averaging over the phase space:
\begin{equation}
H_s = -\frac{1}{V} \int d^3 R\int_{\mbf{R}} d^3 v \left(\frac{T_s h_s}{F_{0,s}}C[h_s]\right) + T_{s} \pd{\ln n_{s}}{x} \Gamma_s
\label{eqn:xchange2}
\end{equation}
Note that the effect of collisions is retained in Eq.~(\ref{eqn:xchange2}), but collisional temperature
equilibration is neglected in the energy equation~(\ref{eqn:pressure}). This is because the turbulent
fluctuations undergo filamentation in phase space and thus tiny collisional deflections in velocity
have immediate
impact on the turbulent fluctuations -- even when the amount of energy that is exchanged
in the course of the deflection is negligible.
At the small phase space scales where collisional dissipation occurs,
the dissipation is effectively diffusive. This gives rise to
heating that is positive definite for each species. Furthermore,
Poisson's equation~(\ref{eqn:poisson}) applied to our two-component plasma dictates that $\Gamma_i=\Gamma_e \doteq \Gamma$ and $n_{i}=n_{e}\doteq n$. Substituting these expressions in Eq.~(\ref{eqn:xchange2}) and enforcing $H_i = -H_e$
results in the constraint that $\Gamma (\partial \ln n/\partial x) \leq 0$ and thus
$|H_s| \leq T_s |\Gamma (\partial \ln n/\partial x)|$. Comparing this
inequality with Eqs.~(\ref{eqn:pressure}) and~(\ref{eqn:continuity}), one finds that the
timescale associated with turbulent particle transport is always at least as fast as that associated with
turbulent heating. One would therefore expect that when turbulent heating drives temperatures apart (as we find below), the effect would be bounded by the rate of transport down the density gradient.
Depending on the system under consideration, this may be set by the rate at which plasma is introduced.
\section{Linear analysis}
If we consider small amplitude perturbations, we may neglect the quadratic nonlinearity and
carry out a linear analysis of the gyrokinetic equation. Upon assuming solutions of the form
$h_s = \sum_{\mbf{k}}\hat{h}_{\mathbf{k},s}(v_{\parallel}, v_{\perp})\exp(i \mathbf{k}\cdot\mathbf{R}_s-i\omega t)$, one obtains
\begin{equation}
\hat{h}_{\mathbf{k},s} = \frac{q_s \hat{\varphi}_{\mbf{k}}}{T_s}J_0(\alpha_{k_{\perp},s})\left( \frac{\omega + \omega_{*,s}}{\omega-k_{\parallel} v_{\parallel}}\right)F_{M,s},
\label{eqn:hks}
\end{equation}
where
\begin{equation}
\omega_{*,s} = \frac{k_y \rho_s \vths{s}} {2L_{n,s}}\frac{q_s}{|q_s|},
\end{equation}
$\vths{s}=2T_s/m_s$, $m_s$ is species mass, $\rho_s = \vths{s}/|\Omega_s|$, and $1/L_{n,s}=-\partial \ln n_s/\partial x$.
From Poisson's Equation~(\ref{eqn:poisson}),
\begin{equation}
\int d^3 v \left(h_i - h_e\right) = \frac{e n}{T_e}\left(1 + \tau - \lambda_e^2\nabla_{\perp}^2\right) \varphi,
\label{eqn:QN}
\end{equation}
where $\tau \doteq T_e/T_i$, and we have restricted our attention to a single ion species
with proton charge $e$. Substitution of Eq.~(\ref{eqn:hks}) into Eq.~(\ref{eqn:QN}) results
in the dispersion relation
\begin{equation}
\epsilon(\omega,\mbf{k})=1+\tau + k_{\perp}^2\lambda_e^2 + \left(\zeta_e - \frac{k_y\rho_e}{2k_{\parallel} L_n }\right)
\Upsilon(k_y\rho_e) Z\left(\zeta_e\right)
+ \tau \left(\zeta_i + \frac{k_y{\rho_i}}{2k_{\parallel} L_n}\right)\Upsilon\left(k_y\rho_i\right)Z\left(\zeta_i\right) = 0,
\label{eqn:disp}
\end{equation}
where $\zeta_s \doteq \omega/k_{\parallel}\vths{s}$,
\begin{equation}
Z(x) \doteq \textnormal{i} \sqrt{\upi} \textnormal{e}^{-x^2} \ \erfc{-\textnormal{i} x}
\end{equation}
is the plasma dispersion function, $\textnormal{erfc}$ is the complementary error function, and
\begin{equation}
\Upsilon(x) \doteq \exp\left(-\frac{x^2}{2}\right) I_0\left(\frac{x^2}{2}\right),
\end{equation}
with $I_0$ a modified Bessel of the first kind. Note that we have used quasineutrality to
set $L_{n,i}=L_{n,e}\doteq L_n$.
\subsection{Quasilinear energy exchange}
Using Eq.~(\ref{eqn:hks}) for $\hat{h}_{\mathbf{k},s}$ in Eq.~(\ref{eqn:xchangek})
we get a quasilinear approximation for the energy exchange:
\begin{equation}
\begin{split}
\hat{H}_{\mbf{k},s} &\doteq \frac{\tilde{H}_{\mbf{k},s}}{n T_i} \frac{|L_n|}{\vths{i}}\frac{T_i^2}{e^2 |\tilde{\varphi}_{\mbf{k}}|^2} \\
&=\Upsilon(k_{\perp}\rho_s)k_{\parallel} |L_n|\frac{T_i}{T_s} \real{\zeta_i}
\imag{\left(\zeta_s + \frac{k_y\rho_s}{2k_{\parallel} L_n}\frac{q_s}{|q_s|}\right) Z(\zeta_s)},
\label{eqn:HQL}
\end{split}
\end{equation}
It is straightforward to verify that summing this expression over species and using the
dispersion relation of Eq.~(\ref{eqn:disp}) leads to zero net heating.
This quasilinear estimate provides information about linear phase relationships that indicate the sign
of the heating contribution as a function of wavelength. It predicts neither the spectrum nor the saturated heating amplitude in steady-state. As we are primarily interested in the sign of the heating, it is
enough to make some assumption about the turbulent heating spectrum; in particular, we assume that the steady-state heating has the same
sign as the quasilinear estimate for the mode(s) with the largest linear growth rate.
In the following subsections we obtain numerical and approximate analytical solutions
for the mode frequencies and associated quasilinear heating in various limits.
\subsection{Comparable thermal speeds}
We first show that there is no instability, and thus no turbulent heating, if $\vths{e}=\vths{i}$. To find the condition for marginal stability, we seek solutions for which $\gamma\doteq\imag{\omega}=0$.
In this case, the plasma dispersion function simplifies to $Z(x)=\sqrt{\upi} \exp(-x^2) (\textnormal{i} - \erfi{x})$, with $\textnormal{erfi}$ the imaginary error function and $x=\real{\zeta}$. The constraint $\imag{\epsilon}=0$ then gives
\begin{equation}
\left(x_e - \frac{k_y\rho_e}{2k_{\parallel} L_n }\right)
\Upsilon(k_y\rho_e) \exp(-x_e^2)=
-\tau \left(x_i + \frac{k_y{\rho_i}}{2k_{\parallel} L_n}\right)\Upsilon\left(k_y\rho_i\right)\exp(-x_i^2).
\end{equation}
Substituting this expression into the constraint $\real{\epsilon}=0$ gives
\begin{equation}
0 = 1+\tau +k_{\perp}^2\lambda_e^2+\tau \left(x_i + \frac{k_y{\rho_i}}{2k_{\parallel} L_n}\right)\Upsilon\left(k_y\rho_i\right)\upi^{1/2}\exp(-x_i^2) \left(\erfi{x_e}-\erfi{x_i}\right).
\label{eqn:ReEps}
\end{equation}
When $\vths{e} = \vths{i}$, then $x_e = x_i$, and Eq.~(\ref{eqn:ReEps}) has no solution. Such a plasma is therefore either always stable or always unstable, independent of wavenumber and density gradient.
As there is no instability for zero density gradient, the plasma must therefore be always stable for
$\vths{e}=\vths{i}$.
Next, we consider how marginal stability is modified when $\vths{e} = \vths{i}\left(1+\delta\right)$, with $|\delta| \ll 1$. In this limit, the constraint Eq.~(\ref{eqn:ReEps}) becomes
\begin{equation}
0 \approx 1 +\frac{k_{\perp}^2\lambda_e^2}{2}- x_i \Upsilon(k_y\rho_i) \left(x_i + \frac{k_y\rho_i}{2k_{\parallel} L_n}\right)\delta,
\end{equation}
with solutions given by
\begin{equation}
x_i = -\frac{k_y\rho_i}{4k_{\parallel} L_n}\pm \sqrt{\left(\frac{k_y\rho_i}{4k_{\parallel} L_n}\right)^2
+ \frac{1+k_{\perp}^2\lambda_e^2/2}{\Upsilon(k_y\rho_i)\delta}}.
\label{eqn:xi}
\end{equation}
In order for the solutions to be real (as we assumed when we considered marginal stability),
the term inside the square root must be positive definite. This constraint is satisfied when $\delta>0$ or when $(k_y\rho_i)^2 \Upsilon(k_y\rho_i) |\delta| > 16(k_{\parallel} L_n)^2(1+k_{\perp}^2\lambda_e^2/2)$. The latter constraint can always be satisfied for sufficiently long parallel wavelengths.
Consequently, an unbounded system can be unstable for all finite values of $\delta$; i.e., for all plasmas
with $\vths{i}\neq \vths{e}$.
We now proceed to obtain the sign of the turbulent heating driven by instabilities with
$\delta <0$ and $\delta >0$, respectively. Since $|\delta| \ll 1$, we
can can approximate $\zeta_s \approx x_s$ in the quasilinear heating expression~(\ref{eqn:HQL}) to obtain:
\begin{equation}
\hat{H}_{\mbf{k},i} = k_{\parallel} |L_n| x_i
\Upsilon(k_{\perp}\rho_i) \left(x_i + \frac{k_y\rho_i}{2k_{\parallel} L_n}\right)\frac{\sqrt{\upi}}{2}\exp(-x_i^2).
\label{eqn:HQLapp}
\end{equation}
When $\delta < 0$, $x_i$ given by Eq.~(\ref{eqn:xi}) satisfies $-k_y \rho_i/2k_{\parallel} L_n < x_i < 0$.
Substituting this range of $x_i$ values in Eq.~(\ref{eqn:HQLapp}) results in the constraint
$\hat{H}_{\mbf{k},i} < 0$. When $\delta > 0$, $x_i$ given by Eq.~(\ref{eqn:HQL})
satisfies $x_i > 0$ or $x_i < -k_y\rho_i / 2k_{\parallel} L_n$. Substituting this range of $x_i$ values
in Eq.~(\ref{eqn:HQLapp}) results in the constraint $\hat{H}_{\mbf{k,}i} > 0$.
Combining these two constraints gives the general expression $\textnormal{sgn}(\hat{H}_{\mbf{k},i})=\textnormal{sgn}(\delta)$. Thus we see that for small deviations from stability,
the quasilinear turbulent heating acts to equalize the ion and electron thermal speeds
and thus stabilize the mode.
For the case of pair plasmas ($m_i = m_e$) our quasilinear analysis indicates that turbulent heating
driven by the electron drift wave acts to equalize the ion and electron temperatures.
We can also use a symmetry of the gyrokinetic-Poisson system of equations
to see how pair plasma heating depends on temperature ratio.
In particular, we consider how the equations
are modified under an interchange of the electron and ion temperatures.
First we note that the average over Larmor angle is unaffected by the interchange, as the Larmor radius
of each particle is independent of temperature. Denoting the solutions when $\tau=\tau_0$ as
($h_s$, $\varphi$) and the solutions when $\tau = 1/\tau_0$ as ($\hswap{s}$, $\dphi^{\leftrightarrow}$), we have
\begin{equation}
\pd{\hswap{s}}{t} + v_{\parallel}\pd{\hswap{s}}{z} + \frac{c}{B}\{\left<\varphi^{\leftrightarrow}\right>_{\mathbf{R}},h_s^{\leftrightarrow}\} = \frac{q_s}{T_{s'}} \pd{\left<\dphi^{\leftrightarrow}\right>_{\mathbf{R}}}{t} F_{0,s'}+ \frac{c}{B}\pd{\left<\dphi^{\leftrightarrow}\right>_{\mathbf{R}}}{y}\pd{F_{0,s'}}{x} + C[\hswap{s}]
\end{equation}
and
\begin{equation}
\int d^3v \left(\hswap{i} - \hswap{e}\right) = \left(\sqrt{\tau_0} + \sqrt{\frac{1}{\tau_0}} - \frac{T_0}{4\pi e^2}\nabla^2\right) \frac{e\dphi^{\leftrightarrow}}{T_0},
\end{equation}
with $T_0 \doteq (T_e T_i)^{1/2}$, $s'=i$ when $s=e$, and $s'=e$ when $s=i$. This admits
solutions ($\hswap{s}(x,y,z,v_{\parallel},v_{\perp},t) = -h_{s'}(-x,-y,z,v_{\parallel},v_{\perp},t)$, $\dphi^{\leftrightarrow}(x,y,z,t)=\varphi(-x,-y,z,t)$). The linear growth rates for pair plasmas are thus symmetric under interchange of $T_i$ and
$T_e$, as seen in Fig.~(\ref{fig:gamma_pair}).
Applying these symmetries to the heating expression~(\ref{eqn:xchange0}) gives
\begin{equation}
\begin{split}
H_s^{\leftrightarrow} &= \frac{1}{V}\int d^3 R\int_{\mbf{R}} d^3 v q_s \overline{\left(\hswap{s}\pd{\gyroR{\dphi^{\leftrightarrow}}}{t}\right)} \\
&=\frac{1}{V}\int d^3 R\int_{\mbf{R}} d^3 v q_{s'} \overline{\left(h_{s'}(-x,-y,z,v_{\parallel},v_{\perp},t)\pd{\left<\varphi(-x,-y,z,t)\right>_{\mathbf{R}_{s'}}}{t}\right)} \\
&= H_{s'} = -H_s,
\end{split}
\end{equation}
where the last equality follows from the fact that the species-summed heating is zero. Thus
the heating of each species of a pair plasma is anti-symmetric under interchange of $T_i$ and $T_e$,
and there can be no turbulent heating when $T_i = T_e$.
\subsection{Disparate thermal speeds}
For plasmas with $m_i \gg m_e$, the ions and electrons will have
disparate thermal speeds $\vths{i} \ll \vths{e}$
for temperature ratios satisfying $m_e/m_i \ll \tau \ll m_i/m_e$. In this case,
we can look for solutions to the dispersion relation that satisfy
$|\zeta_e| \ll 1 \ll |\zeta_i|$. With this restriction, the plasma dispersion functions
appearing in Eq.~(\ref{eqn:disp}) can be greatly simplified. In particular, we use
\begin{equation}
Z\left(\zeta_e\right) \approx \textnormal{i}\sqrt{\upi}
\end{equation}
and
\begin{equation}
Z(\zeta_i) \approx -1/\zeta_i,
\end{equation}
giving the following approximate dispersion relation:
\begin{equation}
\zeta_i\left(1+\tau + k_{\perp}^2\lambda_e^2\right) + \zeta_i\left(\zeta_e - \frac{k_y\rho_e}{2k_{\parallel} L_n }\right)
\Upsilon(k_y\rho_e) \textnormal{i}\sqrt{\upi}
- \tau \left(\zeta_i + \frac{k_y{\rho_i}}{2k_{\parallel} L_n}\right)\Upsilon\left(k_y\rho_i\right)= 0.
\label{eqn:disp_vte_gg_vti}
\end{equation}
We next consider wavelengths much shorter than the ion Larmor radius but much longer than
the electron Larmor radius; i.e., $k_{\perp}\rho_i \gg 1 \gg k_{\perp}\rho_e$. In this limit
we can approximate $\Gamma(k_y\rho_e) \approx 1$ and $\Gamma(k_y\rho_i) \approx 1/(\sqrt{\upi}k_{\perp}\rho_i)$ in
Eq~(\ref{eqn:disp_vte_gg_vti}) to obtain
\begin{equation}
\zeta_i\left(1+\tau + k_{\perp}^2 \lambda_e^2\right) + \zeta_i\left(\zeta_e - \frac{k_y\rho_e}{2k_{\parallel} L_n }\right) \textnormal{i}\sqrt{\upi}
- \frac{\tau}{\sqrt{\upi}} \left(\frac{\zeta_i}{k_{\perp}\rho_i} + \frac{k_y}{k_{\perp}}\frac{1}{2k_{\parallel} L_n}\right)= 0.
\end{equation}
Seeking solutions for which $\zeta_i \sim k_{\perp}\rho_e / k_{\parallel} L_n$ allows us
to neglect the $\zeta_e$ and $\zeta_i/k_{\perp}\rho_i$ terms. The resultant solution
for $\omega$ is
\begin{equation}
\omega = \frac{\tau k_{\parallel}\vths{i}}{\sqrt{\upi}}\frac{2k_{\parallel} L_n\left(1+\tau+k_{\perp}^2\lambda_e^2\right)\textnormal{sgn}(k_y) + \textnormal{i} \sqrt{\upi} k_{\perp}\rho_e}{\left(1+\tau+k_{\perp}^2\lambda_e^2\right)^2\left(2k_{\parallel} L_n\right)^2 + \upi k_{\perp}^2\rho_e^2}.
\label{eqn:omconventional}
\end{equation}
In Fig.~(\ref{fig:tungsten}) we show an example comparison of the exact and analytical expressions for $\omega$ for a plasma with $k_{\perp}\lambda_e=0$ and $m_i/m_e = 337824$ (corresponding to the
mass of tungsten).
\begin{figure}
\centerline{\includegraphics[height=1.9in]{figures/freqtungsten.pdf}
\caption{Comparison of exact (solid lines) and approximate analytical (dashed lines) growth rates for the case $\tau=1$ and $m_i/m_e = 337824$ (tungsten ions).}
\label{fig:tungsten}
\end{figure}
Eq.~(\ref{eqn:omconventional}) indicates that for $k_{\parallel} > 0$ there is an instability with peak growth rate $\gamma$ at wavelengths $\mbf{k}_m$ satisfying $(\partial \gamma/\partial k_{\parallel})|_{\mbf{k}_{m}} = (\partial \gamma/\partial k_{\perp})|_{\mbf{k}_m} = 0$. These constraints give
\begin{equation}
\frac{2}{\sqrt{\upi}}k_{\parallel,m} |L_n| \left(1+\tau + k_{\perp,m}^2\lambda_e^2\right) = k_{\perp,m}\rho_e,
\end{equation}
with the largest growth rate for $k_{\perp,m}\lambda_e = 0$. Using these results
in Eq.~(\ref{eqn:omconventional}) gives
\begin{equation}
\omega_m = \frac{\tau}{4\sqrt{\upi}\left(1+\tau\right)}\frac{\vths{i}}{|L_n|}
\left(\textnormal{i} + \textnormal{sgn}(k_y L_n)\right),
\label{eqn:ommax}
\end{equation}
with $\omega_m\doteq \omega(\mbf{k}_m)$ the complex frequency evaluated at the wavevector $\mbf{k}_m$ that
maximizes the linear growth rate.
Plugging Eq.~(\ref{eqn:ommax}) for $\omega_m$ into Eq.~(\ref{eqn:HQL}) for turbulent heating
and using the appropriate approximations for $\Upsilon(k_y\rho_s)$ and $Z(\zeta_s)$ gives
\begin{equation}
\begin{split}
\hat{H}_{\mbf{k},i} = -\hat{H}_{\mbf{k},e} &\approx \frac{k_y\rho_e}{k_{\parallel} \vths{i}}\real{\omega}\frac{|L_n|}{L_n} \frac{\sqrt{\upi}}{4} \\
& = \frac{\tau}{8\sqrt{\upi}} > 0.
\end{split}
\end{equation}
So ions are heated and electrons are cooled; i.e., the instability acts to equalize
$\vths{e}$ and $\vths{i}$ and thus stabilize the mode.
\section{Simulation results}
We now provide numerical data to verify our analytical predictions. All simulations were conducted
using the local, Eulerian gyrokinetic code $\texttt{GS2}$~\citep{kotschCPC95,dorlandPRL00}
with kinetic electrons and a single ion species
immersed in a straight, uniform magnetic field and with a cross-field density gradient. The fluctuations
are constrained to be purely electrostatic, and we take $k_{\perp}\lambda_e = 0$. Each simulation used 32 points along the magnetic field direction ($z$), and the velocity space is sampled on a polar grid~\citep{barnesPoP10a}, with 12 points in speed $v$ and 16 points in pitch angle $v_{\parallel}/v$.
First we consider the case of an electron-positron plasma. The linear growth rates, maximized
over $k_{\parallel}$ and $k_{\perp}$, are plotted against $T_e/T_i$ in Fig.~(\ref{fig:gamma_pair}). They
are normalized by $|L_n|/v_0$, with $v_0 \doteq \sqrt{\vths{i}\vths{e}}$, in order to make manifest the expected symmetry of the growth rates with respect to the interchange of $T_i$ and $T_e$. As predicted,
the plasma is stable only when $T_e = T_i$, and the growth rates are symmetric about $T_e/T_i=1$.
Also shown in Fig.~(\ref{fig:gamma_pair}) are the quasilinear heating estimates defined in Eq.~(\ref{eqn:HQL}) and the heating determined from nonlinear simulations as a function of $T_e/T_i$.
The nonlinear simulations employed 47 modes (after de-aliasing) in both directions perpendicular to the magnetic field ($x$ and $y$). As we showed analytically, the heating (both quasilinear and nonlinear)
is antisymmetric about $T_e/T_i=1$, with the heating acting to equilibrate the electron and positron
temperatures and thus shut off the linear instability.
\begin{figure}
\includegraphics[height=1.9in]{figures/gamma_positron.pdf}
\includegraphics[height=1.9in]{figures/heating_positron.pdf}
\caption{
Normalized linear growth rates maximized over $k_{\parallel}$ and $k_y$ (left) and corresponding turbulent heating (right)
from \texttt{GS2} simulations as a function of electron-ion temperature ratio for an electron-positron plasma. The quasilinear turbulent ion heating $\hat{H}_i$ (black circles and line) given by Eq.~(\ref{eqn:HQL}) is weighted by the linear growth rate $\gamma |L_n|/v_0$ to qualitatively estimate saturated fluctuation amplitudes. The turbulent ion heating $(H_i |L_n| / n T_0 v_0) (|L_n|/\rho_0)^2$
obtained from nonlinear simulations is also given (red squares). Here $\rho_0 = v_0 / \Omega_i$, with
$v_0$ the geometric mean of the ion and electron thermal speeds.
}
\label{fig:gamma_pair}
\end{figure}
We next show the growth rates and turbulent ion heating as a function of electron-ion temperature
for an electron-proton plasma in
Fig.~(\ref{fig:gamma_proton}). Again we see that the plasma is unstable for $\vths{e}\neq \vths{i}$.
Furthermore, the ion turbulent heating is positive definite for $\vths{e} > \vths{i}$, in agreement
with our approximate analytic result.
\begin{figure}
\includegraphics[height=1.9in]{figures/gamma_proton.pdf}
\includegraphics[height=1.9in]{figures/heating_proton.pdf}
\caption{Normalized linear growth rates maximized over $k_{\parallel}$ and $k_y$ (left) and corresponding turbulent heating (right)
from \texttt{GS2} simulations as a function of electron-ion temperature ratio for an electron-proton plasma. The quasilinear turbulent ion heating $\hat{H}_i$ (black circles and line) given by Eq.~(\ref{eqn:HQL}) is weighted by the linear growth rate $\gamma |L_n|/\vths{i}$ to qualitatively estimate saturated fluctuation amplitudes. The turbulent ion heating $(H_i |L_n| / n T_i \vths{i}) (|L_n|/\rho_i)^2$
obtained from nonlinear simulations is also given (red squares)}
\label{fig:gamma_proton}
\end{figure}
\section{Conclusions}
The analytical and numerical results shown in this paper indicate that turbulent heating
driven by the electron drift wave instability present in an inhomogeneous,
magnetized plasma acts to stabilize the mode. Stabilization occurs when the ion and
electron thermal speeds are equal. For a conventional mass ratio
plasma with $T_i \sim T_e$, this leads to the ions being heated and the electrons cooled;
for a pair plasma, the turbulent heating acts to equalize the ion and electron temperatures.
While turbulent heating acts to stabilize the mode, it is not the only stabilization mechanism
in the system. We showed that the instability drives particle transport that flattens the driving
density gradient on a time scale at least as fast as the heating influences the electron-ion temperature
ratio. Consequently, for the turbulent heating to fully set the thermal equilibrium there must be additional
physics that fixes the density gradient; e.g., an external density source or a
lowest order equilibrium set by gravitational forces.
In this case, thermal equilibrium would correspond to $T_i /T_e = \sqrt{m_i/m_e}$.
The authors would like to thank F. I. Parra, A. A. Schekochihin and A. Zocco for useful discussions. M. Barnes was supported in part by STFC grant ST/N000919/1. The authors also acknowledge the use of ARCHER through the Plasma HEC Consortium EPRSC grant number EP/L000237/1 under project e281-gs2 and the use of
the EUROfusion High Performance Computer (Marconi-Fusion) under project MULTEI.
\bibliographystyle{jpp}
|
\section{Introduction}
Missing data is common in many fields of applications. One way to deal with the missing data problem is to delete observations
containing missing data. In doing so we may produce biased estimates and erroneous conclusions, depending on the data missing mechanism. If data are missing completely at random, standard estimation and inference procedures are still consistent when the missing data observations are ignored, see \cite{heitjan1996distinguishing}, \cite{little1988test} among others. If data are missing at random (MAR) in the sense that the propensity of missingness depends only on the observed covariates, consistent estimation can still be obtained through covariate balancing, see \cite{rubin1976comparing, rubin1976inference}, \cite{little1989analysis}, \cite{robins1995semiparametric}, \cite{robins1995analysis}, \cite{bang2005doubly}, \cite{qin2007empirical}, \cite{chen2008semiparametric}, \cite%
{tan2010bounded}, \cite{rotnitzky2012improved}, \cite{little2014statistical}
among others. In many applications, data are missing not at random (MNAR). For example, the income question in sample surveys is often not answered by people at the top end of the distribution, that is, their response frequency depends on an outcome variable that is often the key focus. An investigator is examining the effect of sleep on pain by calling subjects daily to ask them about last night's sleep and their pain today. Patients who are experiencing severe pain are more likely to not come to the phone leaving the data missing for that particular day; again this would violate the MAR assumption. From political science, roll-call votes, which measure legislatures ideological positions, are subject to non-ignorable nonresponse because, unsurprisingly, politicians behave strategically.
In the MNAR case, the parameter of interest may not even be identified (e.g., \cite{robins1997toward}), let
alone be consistently estimated. To be more specific, let $T\in \{0,1\}$ denote the binary random variable indicating the missing status of the outcome variable $Y$: $Y$ is observed if $T$ takes the value one and $Y$ is not observed if $T$ takes the value zero. Let $\boldsymbol{X}$ denote a vector of explanatory variables, let $\pi (\boldsymbol{x},y)=\mathbb{P}(T=1|%
\boldsymbol{X}=\boldsymbol{x},Y=y)$ denote the propensity score function and let $f_{Y|\boldsymbol{X}}(y|\boldsymbol{x})$ denote the conditional density function of $Y$ given $\boldsymbol{X}$. \cite{robins1997toward} shows that
if both the propensity score function and the conditional density function are completely unknown, the joint \ distribution of $(T,Y)$ given $\boldsymbol{X}$ is
not point identifiable. In this case, a necessary identification condition is the parameterization of either the propensity score function or the conditional density function. Molenberghs and Kenward (2007) proposes the parameterization of both the propensity score function and the conditional density function as an identification strategy, while \cite{sverchkov2008new} and \cite%
{riddles2016propensity} parameterize the propensity score function and only a component of the conditional density function: $f_{Y|\boldsymbol{X},T}(y|\boldsymbol{x},T=1)$. \\
If the joint distribution is not the parameter of interest, the identification strategy above can be modified. For example, if the parameter of interest is the conditional density of $Y$ given $\boldsymbol{X}$ (i.e., $f_{Y|\boldsymbol{X}}(y|\boldsymbol{x})$),
parameterization of the propensity score function is not needed. However, parameterization of $f_{Y|\boldsymbol{X}}(y|\boldsymbol{x})$ in this case is not sufficient for identification due to missing data. \cite{tang2003analysis}
suggests parameterization of the marginal density $f_{\boldsymbol{X}}(\boldsymbol{x})$ as well, while \cite{zhao2015semiparametric} imposes an exclusion
retriction. In both studies, $f_{Y|\boldsymbol{X}}(y|\boldsymbol{x})$ is identified and consistently estimated. \\
We consider estimation of the parameter $\theta _{0}=\mathbb{E}[U(\boldsymbol{X},Y)]$,
where $U(\cdot )$ is any known function. We suppose that the propensity score $\pi$ is parameterized but do not restrict the conditional density function of the outcome variable.
In earlier work in this framework, either the coefficients in the propensity score
function are known or consistently estimated from an external sample (\cite{kim2011semiparametric}) or an exclusion restriction is imposed (\cite{wang2014instrumental} and \cite{shao2016semiparametric}). \cite{morikawa2016semiparametric} study the
efficient estimation of $\theta _{0}$. They derive the efficient score
function (and hence the semiparametric efficiency bound) for $\theta _{0}$ in this model. They propose to estimate the
efficient score function by estimating $f_{Y|\boldsymbol{X},T}(y|,%
\boldsymbol{x},1)$ by a working parametric model (MK1) or by kernel nonparametric estimation (MK2). Their approach MK1 is not efficient unless the working parametric model is correct, although it is consistent.
Their method MK2 suffers from the curse of dimensionality (their smoothness conditions depend on the dimensionality of the covariates through their conditions C14) and the bandwidth selection problem (about which they give no guidance). \\
We study the same estimation problem as in \cite{morikawa2016semiparametric} but propose a simpler yet equally efficient estimation procedure. Our proposed method does not require explicit nonparametric estimation and hence does not suffer
from the curse of dimensionality. The proposed
estimator is motivated by the key insight that the model parameter
satisfies a parametric conditional moment restriction, of which the
semiparametric efficiency bound is identical to the bound derived in \cite%
{morikawa2016semiparametric}. The conditional moment restriction is then turned
into an expanding set of unconditional moment restrictions and the parameter
of interest is estimated by applying the widely available and easy to
compute GMM estimation (see Hansen (1982)). Under some sufficient conditions, we establish that the proposed
estimator is consistent and asymptotically normally distributed even if the
set of unconditional moment restrictions does not expand, thereby freeing us
from the curse of dimensionality and the bandwidth selection problem; when the set does
expand, the proposed estimator attains the semiparametric efficiency bound.
This is in contrast with the MK2 method of \cite{morikawa2016semiparametric}%
, which is inconsistent if the bandwidth does not go to zero at a certain rate.\\
The paper is organized as follows. Section 2 describes the estimation,
Section 3 derives the large sample properties of the estimator, Section 4 provides a consistent asymptotic variance estimator, Section 5 suggests two data driven approaches to determine the number of unconditional moment restrictions, Section 6 reports on a small scale simulation study, followed by some concluding remarks in Section 7. All technical proofs are relegated to the Appendix.
\section{Basic Framework and Estimation}
We begin by setting up the basic framework. Denote $\boldsymbol{Z}=(%
\boldsymbol{X}^{\top },Y)^{\top }$. The following assumption shall be maintained throughout the paper:
\ \\
\textbf{Assumption 2.1}. (i) Parameterization of data missing mechanism: $\mathbb{P}(T=1|Y,\boldsymbol{X})=\pi (Y,\boldsymbol{X};\gamma _{0})=\pi (%
\boldsymbol{Z};\gamma _{0})$ holds for some\ known function $\pi (.;.)$, where $\gamma _{0}\in \mathbb{R}^{p}$ for some known $p\in \mathbb{N}$ is the true (unknown) value; (ii) exclusion restriction: there exists some
nonresponse instrument variables $\boldsymbol{X}_{1}$ in $\boldsymbol{X}=(%
\boldsymbol{X}_{1}^{\top },\boldsymbol{X}_{2}^{\top })^{\top }$ so that $%
\boldsymbol{X}_{2}$ is independent of $T$ given both $\boldsymbol{X}_{1}$
and $Y$; and (iii) the parameter of interest is $\theta _{0}=\mathbb{E}[U(%
\boldsymbol{Z})]$ for some known function $U(\cdot )$.\\
Under Assumption 2.1 and by applying the law of iterated expectations, we obtain the following conditional moment restrictions:
\begin{align}
& \mathbb{E}\left[ 1-\frac{T}{\pi (\boldsymbol{Z};\gamma _{0})}\bigg|\boldsymbol{X%
}\right] =0, \label{sequential1} \\
& \mathbb{E}\left[ \theta _{0}-\frac{T}{\pi (\boldsymbol{Z};\gamma _{0})}U(%
\boldsymbol{Z})\right] =0, \label{sequential2}
\end{align}%
which will form the basis for the proposed estimation. We notice that the parameters of interest in (\ref{sequential1})-(\ref{sequential2}) are finite dimensional (and there is no explicit infinite dimensional nuisance parameter) and can be easily
estimated with GMM estimation. We also notice that it is a special case of
the model studied in \cite{ai2012semiparametric}. By applying their result (Remark
2.1, pp. 446), we obtain the semiparametric efficiency bound for model (\ref{sequential1})-(%
\ref{sequential2}), which is identical to the bound derived in \cite%
{morikawa2016semiparametric}, thereby suggesting a simple and efficient estimation. \\
The (nuisance) parameter $\gamma _{0}$ is identified by (\ref{sequential1}) and the parameter of interest $\theta _{0}$ is identified by (\ref{sequential2}). The following condition shall also be maintained throughout the paper:
\ \\
\textbf{Assumption 2.2}. The parameter space $\Gamma $ is a compact subset
of $\mathbb{R}^{p}$. The true value $\gamma _{0}$ lies in the interior of $%
\Gamma $ and is the only solution to (\ref{sequential1}). {\color{black}The parameter space $\Theta$ is a compact subset of $\mathbb{R}$ and the true value $\theta_0$ lies in the interior of $\Theta$.} \\
To estimate model (\ref{sequential1})-(\ref{sequential2}), we first turn it
into a set of unconditional moment restrictions. We work with a set of known basis functions: for each integer $\ K\in
\mathbb{N}\ \text{with}\ K\geq p\ $, let
\begin{equation*}
u_{K}(\boldsymbol{X})=(u_{1K}(\boldsymbol{X}),\ldots ,u_{KK}(\boldsymbol{X}%
))^{\top }.\
\end{equation*}
{\color{black}Discussion on the choice of $u_K(\boldsymbol{X})$ and its properties can be found in Section \ref{sec:uK} in Appendix}. Model (\ref{sequential1})-(\ref%
{sequential2}) {\color{black}implies the unconditional moment restrictions}:
\begin{align}
& \mathbb{E}\left[ \left( 1-\frac{T}{\pi (\boldsymbol{Z};\gamma _{0})}%
\right) u_{K}(\boldsymbol{X})\right] =0, \label{uncond1} \\
& \mathbb{E}\left[ \theta _{0}-\frac{T}{\pi (\boldsymbol{Z};\gamma _{0})}U(%
\boldsymbol{Z})\right] =0. \label{uncond2}
\end{align}%
To avoid redundant moment restrictions, we require $\mathbb{E}\left[ u_{K}(%
\boldsymbol{X})u_{K}(\boldsymbol{X})^{\top }\right] $ to be nonsingular for
every $K$. The following somewhat stronger identification condition shall be maintained throughout the paper:
\bigskip
\textbf{Assumption 2.2'}. The parameter space $\Gamma $ is a compact subset
of $\mathbb{R}^{p}$. The true value $\gamma _{0}$ lies in the interior of $%
\Gamma $ and is the only solution to (\ref{uncond1}). {\color{black}The parameter space $\Theta$ is a compact subset of $\mathbb{R}$ and the true value $\theta_0$ lies in the interior of $\Theta$.}\\
We can estimate the parameter of interest by the GMM
method. Let $\{T_{i},\boldsymbol{Z}_{i}\}_{i=1}^{N}$ denote an $i.i.d%
\text{.}$ sample drawn from the joint distribution of $(T,\boldsymbol{Z})$.
Denote
\begin{eqnarray*}
\boldsymbol{G}_{K}(\gamma ,\theta ) :&=&\left( \sum_{i=1}^{N}\left[ 1-\frac{%
T_{i}}{\pi (\boldsymbol{Z}_{i};\gamma )}\right] u_{K}(\boldsymbol{X}%
_{i})^{\top },\sum_{i=1}^{N}\left[ \theta -\frac{T_{i}}{\pi (\boldsymbol{Z}%
_{i};\gamma )}U(\boldsymbol{Z}_{i})\right] \right) ^{\top } \\
&=&\sum_{i=1}^{N}g_{K}(T_{i},\boldsymbol{Z}_{i};\gamma ,\theta )\text{,}
\end{eqnarray*}
{\color{black} where $g_{K}(T,\boldsymbol{Z};\gamma ,\theta ):=\left( \left[ 1-\frac{%
T}{\pi (\boldsymbol{Z};\gamma )}\right] u_{K}(\boldsymbol{X})^{\top }, \theta -\frac{T}{\pi (\boldsymbol{Z};\gamma )}U(\boldsymbol{Z})\right) ^{\top }$.} The GMM estimator of $\gamma _{0}$ and $\theta _{0}$ is defined as
\begin{equation*}
(\check{\gamma},\check{\theta})=\arg \underset{\gamma \in \Gamma ,\theta\in\Theta }{%
\min }\boldsymbol{G}_{K}(\gamma ,\theta )^{T}\cdot \mathbf{W}\cdot
\boldsymbol{G}_{K}(\gamma ,\theta )
\end{equation*}%
where $\mathbf{W}$ is a $(K+1)\times (K+1)$ symmetric weighting matrix. For every fixed $K\geq p$, \cite{hansen1982large} shows that, under some regularity conditions, the estimator
\begin{equation}
{(\check{\gamma}-\gamma _{0},\check{\theta}-\theta _{0})=\color{black}O_{p}({N}^{-1/2})}
\label{Hansen1}
\end{equation}%
{\ is asymptotically normally distributed, but generally not the best unless
the best weighting matrix is used. The best weighting matrix is the inverse
of }%
\begin{equation*}
\boldsymbol{D}_{(K+1)\times (K+1)}:=\mathbb{E}\left[ g_{K}(T,\boldsymbol{Z};\gamma
_{0},\theta _{0})g_{K}(T,\boldsymbol{Z};\gamma _{0},\theta _{0})^{\top }%
\right] .
\end{equation*}%
The best estimator (within the class defined by the specific unconditional moments) is defined as
\begin{equation*}
(\overline{\gamma },\overline{\theta })=\arg \underset{\gamma \in \Gamma
,\theta\in\Theta }{\min }\boldsymbol{G}_{K}(\gamma ,\theta )^{T}\cdot \boldsymbol{D}%
_{(K+1)\times (K+1)}^{-1}\cdot \boldsymbol{G}_{K}(\gamma ,\theta ).
\end{equation*}%
Suppose that the propensity score function is differentiable with respect to
$\gamma $. Denote
\begin{align*}
\boldsymbol{B}_{(K+1)\times (p+1)} =\nabla _{\gamma ,\theta }\mathbb{E}%
\left[ \frac{1}{N}\boldsymbol{G}_{K}(\gamma _{0},\theta _{0})\right] =\left(
\begin{array}{cc}
\mathbb{E}\left[ u_{K}(\boldsymbol{X})\frac{\nabla _{\gamma }\pi (%
\boldsymbol{Z};\gamma _{0})^{\top}}{\pi (\boldsymbol{Z};\gamma _{0})}\right]
\text{,}\vspace{4mm} & \mathbf{0}_{K\times 1} \\
\mathbb{E}\left[ U(\boldsymbol{Z})\frac{\nabla _{\gamma }\pi (\boldsymbol{Z}%
;\gamma _{0})^{\top}}{\pi (\boldsymbol{Z};\gamma _{0})}\right] \text{,} & 1
\end{array}
\right)
\end{align*}
and
\begin{align*} \boldsymbol{V}_{K} =\left\{ \left( \boldsymbol{B}_{(K+1)\times
(p+1)}\right) ^{\top }\boldsymbol{D}_{(K+1)\times (K+1)}^{-1}\left(
\boldsymbol{B}_{(K+1)\times (p+1)}\right) \right\} ^{-1}.
\end{align*}
Hansen (1982) shows that, for every fixed $K\geq p$,
\begin{equation}
\boldsymbol{V}_{K}^{-1/2}\binom{\sqrt{N}(\overline{\gamma }-\gamma _{0})}{%
\sqrt{N}(\overline{\theta }-\theta _{0})}\rightarrow N\left(
0,I_{(p+1)\times (p+1)}\right) \text{ in distribution.} \label{Hansen2}
\end{equation}
Since the best weighting matrix depends on the unknown parameter value, the
best estimator $(\overline{\gamma },\overline{\theta })$ is infeasible.
Hansen (1982) suggests the following two-step procedure:\\
\qquad Step I. Compute the initial $\sqrt{N}$-consistent estimator
{\color{black}\begin{align*}
&\widehat{\boldsymbol{W}}_0:=\begin{pmatrix}
\frac{1}{N}\sum_{i=1}^Nu_K(\bold{X}_i)u_K(\bold{X}_i)^{\top} & \boldsymbol{0}_{K\times 1} \\[2mm]
\boldsymbol{0}_{K\times 1}^{\top} & 1
\end{pmatrix}\ ,\\
&(\check{\gamma},\check{\theta})=\arg \underset{(\gamma,\theta) \in \Gamma \times \Theta }{%
\min }\boldsymbol{G}_{K}(\gamma ,\theta )^{T}\cdot \widehat{\boldsymbol{W}}^{-1}_0\cdot \boldsymbol{G}_{K}(\gamma
,\theta ).
\end{align*}}
\qquad \qquad\ \qquad
\qquad Step II. Compute the best weighting matrix and the best estimator {\
\begin{equation*}
\hat{\boldsymbol{D}}_{(K+1)\times (K+1)}:=\frac{1}{N}\sum_{i=1}^{N}g_{K}(T_{i},%
\boldsymbol{Z}_{i};\check{\gamma},\check{\theta})g_{K}(T_{i},\boldsymbol{Z}%
_{i};\check{\gamma},\check{\theta})^{\top }\;\text{,}
\end{equation*}%
}
\begin{equation*}
(\widehat{\gamma },\widehat{\theta })=\arg \underset{\gamma \in \Gamma
,\theta \in \Theta }{\min }\boldsymbol{G}_{K}(\gamma ,\theta )^{T}\cdot
\widehat{\boldsymbol{D}}_{(K+1)\times (K+1)}^{-1}\cdot \boldsymbol{G}%
_{K}(\gamma ,\theta ).
\end{equation*}
Hansen (1982) establishes that, for {every fixed $K\geq p$},
\begin{equation}
\boldsymbol{V}_{K}^{-1/2}\binom{\sqrt{N}(\widehat{\gamma }-\gamma _{0})}{%
\sqrt{N}(\widehat{\theta }-\theta _{0})}\rightarrow N\left( 0,I_{(p+1)\times
(p+1)}\right) \text{ in distribution.} \label{Hansen3}
\end{equation}%
Moreover, denote
\begin{align*}
\widehat{\boldsymbol{B}}_{(K+1)\times (p+1)} :=\left(
\begin{array}{cc}
N^{-1}\sum_{i=1}^{N}u_{K}(\boldsymbol{X}_{i})\frac{\nabla _{\gamma }\pi (%
\boldsymbol{Z}_{i};\widehat{\gamma })^{\top}}{\pi (\boldsymbol{Z}_{i};\widehat{%
\gamma })}\text{,}\vspace{4mm} & \mathbf{0}_{K\times 1} \\
N^{-1}\sum_{i=1}^{N}U(\boldsymbol{Z}_{i})\frac{\nabla _{\gamma }\pi (%
\boldsymbol{Z}_{i};\widehat{\gamma })^{\top}}{\pi (\boldsymbol{Z}_{i};\widehat{%
\gamma })}\text{,} & 1%
\end{array}%
\right)
\end{align*}
and
\begin{align*}
\widehat{\boldsymbol{V}}_{K} :=\left\{ \left( \widehat{%
\boldsymbol{B}}_{(K+1)\times (p+1)}\right) ^{\top }\widehat{\boldsymbol{D}}%
_{(K+1)\times (K+1)}^{-1}\left( \widehat{\boldsymbol{B}}_{(K+1)\times
(p+1)}\right) \right\} ^{-1}.
\end{align*}%
Hansen (1982) proves that, for {every fixed $K\geq p$},
\begin{equation}
\widehat{\boldsymbol{V}}_{K}\rightarrow \boldsymbol{V}_{K}\text{ in
probability.} \label{Hansen4}
\end{equation}
The best estimator (within the class defined by the specific unconditional moments) is generally not semiparametrically efficient. To obtain the efficient
estimator, we shall allow $K$ to increase with the sample size at the rate {\color{black}$
o(N^{1/3})$} so that $\{u_{K}(\boldsymbol{X})\}$ span the space of measureable
functions (see also \cite{geman1982nonparametric} and \cite%
{newey1997convergence}). In the next two sections, we shall establish that results in (\ref{Hansen1})-(\ref{Hansen4}) still hold with expanding {\color{black}$K=o(N^{1/3})$}. \\
The advantage of our proposed estimator over the existing estimators is evident. Our
estimation problem is a parametric one, requiring no modeling of or
nonparametric estimation of $f_{Y|\boldsymbol{X},T}(y|x,1)$. In contrast,
the estimators proposed in \cite{riddles2016propensity} and \cite%
{morikawa2016semiparametric} could be inconsistent if $f_{Y|\boldsymbol{X}%
,T}(y|x,1)$ is incorrectly specified or suffers from the curve of
dimensionality and bandwith selection problem of the nonparametric
estimation of $f_{Y|\boldsymbol{X},T}(y|x,1)$.
\section{Asymptotic Theory}
In this section, we show that results in {\ (\ref{Hansen1})- (%
\ref{Hansen3}) still hold with expanding }$K$, all technical proof can be found in the supplemental material \cite{alz2018}. First, we establish the convergence rate of the first step estimator $(\check{\gamma}%
,\check{\theta})$.
\begin{theorem}
Under Assumptions 2.1-2.2 and Assumptions \ref{as:id}, \ref{as:suppX}, \ref{as:eigen}, \ref{as:iid}, \ref{as:pi}, and \ref{as:K&N} listed in Appendix,
with $K=o(N^{1/3})$, the first step estimator satisfies
\begin{equation*}
(\check{\gamma}-\gamma _{0},\check{\theta}-\theta _{0})=O_{p}\left(N^{-1/2}\right).
\end{equation*}
\end{theorem}
Next, we establish the large sample properties of the infeasible
best estimator $(\overline{\gamma },\overline{\theta })$ without imposing
the smoothness Assumptions \ref{as:proj_smooth} and \ref{as:boundS} listed in Appendix.
\begin{theorem}
Under Assumptions 2.1-2.2 and Assumptions \ref{as:id}, \ref{as:suppX}, \ref{as:eigen}, \ref{as:iid}, \ref{as:pi}, and \ref{as:K&N} listed in
Appendix, with $K=o(N^{1/3})$, the infeasible best estimator satisfies
\begin{equation*}
\boldsymbol{V}_{K}^{-1/2}\binom{\sqrt{N}(\overline{\gamma }-\gamma _{0})}{%
\sqrt{N}(\overline{\theta }-\theta _{0})}\rightarrow N\left(
0,I_{(p+1)\times (p+1)}\right) \text{ in distribution.}
\end{equation*}
\end{theorem}
{If in addition the smoothness Assumptions \ref{as:proj_smooth} and \ref%
{as:boundS} are
satisfied, the next result shows that $\boldsymbol{V}_{K}\rightarrow
\boldsymbol{V}_{eff}$} in probability, where $\boldsymbol{V}_{eff}$ is the
semiparametric efficiency bound of $(\gamma _{0},\theta _{0})$ {\color{black}derived in \cite{morikawa2016semiparametric}, see Lemma 1 in Section 8.3 of Appendix.}
\begin{theorem}
Under Assumption 2.1-2.2 and Assumption 1-\ref{as:K&N} listed in Appendix, with $%
K=o(N^{1/3})$, we obtain
\begin{equation*}
\boldsymbol{V}_{K}\rightarrow \boldsymbol{V}_{eff}\text{ in probability.}
\end{equation*}
\end{theorem}
By Theorem 1-3, the infeasible best estimator attains the semiparametric efficiency bound. The next result establishes
the equivalence between the best estimator $(\widehat{\gamma },\widehat{%
\theta })$ and the infeasible best estimator $(\overline{\gamma },\overline{%
\theta })$, implying that the best estimator also attains the semiparametric
efficiency bound.
\begin{theorem}
Under Assumption 2.1-2.2 and Assumption 1-\ref{as:K&N} listed in Appendix, with $%
K=o(N^{1/3})$, we obtain
\begin{equation*}
\binom{\sqrt{N}(\overline{\gamma }-\widehat{\gamma })}{\sqrt{N}(\overline{%
\theta }-\widehat{\theta })}=o_{p}(1).
\end{equation*}
\end{theorem}
\section{Variance Estimation}
In order to conduct statistical inference, we need a consistent covariance
estimator. Notice that {\ (\ref{Hansen1}) } implies that $\widehat{\boldsymbol{V}}_{K}
$ is a consistent estimator of $\boldsymbol{V}_{K}$ for every fixed $K\geq p$%
. We now show that this result still holds true with expanding $K$, thereby
providing a consistent covariance estimator.
\begin{theorem}
Under Assumption 2.1-2.2 and Assumption 1-\ref{as:K&N} listed in Appendix, with $%
K=o(N^{1/3})$, we obtain%
\begin{equation*}
\widehat{\boldsymbol{V}}_{K}\rightarrow \boldsymbol{V}_{K}\text{ in
probability.}
\end{equation*}
\end{theorem}
We notice that our covariance estimator is much simpler and more natural
than the one suggested in \cite{morikawa2016semiparametric}, which requires
nonprametric estimation of $f_{Y|\boldsymbol{X},T}(y|x,1)$ and tends to have
poor performance in finite samples. Our covariance estimator is the GMM
covariance estimator and is easily computed by existing statistical
packages.
\section{Selection of $K$}
The large sample properties of the proposed estimator established in the
previous sections allow for a wide range of values for $K$, and theoretically the sensitivity of the estimator
to the choice of $K$ is not so pronounced, it affects higher order terms in a way
that does not affect consistency and asymptotic normality. Nevertheless, there may be some
higher order effect of the choice of $K$ on perfomance. In this section, we present two data-driven approaches to
select $K$. Both approaches strike a balance between bias and variance. \\
\textbf{Covariate balancing approach}. The first approach attempts to
balance the distribution of the covariates between the whole population and
the non-missing population through weighting. Notice that
\begin{equation*}
\mathbb{E}\left[ \frac{T}{\pi (\boldsymbol{Z};\gamma _{0})}I(X_{j}\leq x_{j})%
\right] =\mathbb{E}[I(X_{j}\leq x_{j})]\ ,\ j\in \{1,...,r\}\ ,
\end{equation*}%
where $X_{j}$ is the $j^{th}$ component of $\boldsymbol{X}$ and $I(X_{j}\leq
x_{j})$ is the indicator function. Obviously the propensity score function $\pi (%
\boldsymbol{Z};\gamma _{0})$ plays the role of balancing. Notice that the
estimator $\hat{\gamma}$ depends on $K.$ For a given $K$, we compute
\begin{equation*}
\hat{F}_{N,K}^{j}(x_{j}):=\frac{1}{N}\sum_{i=1}^{N}\frac{T_{i}}{\pi (%
\boldsymbol{X}_{i};\hat{\gamma})}I(X_{ij}\leq x_{j}),\;j\in \{1,\ldots ,r\}%
\text{.}
\end{equation*}%
We compute the empirical distributions of the covariates
\begin{equation*}
\tilde{F}_{N}^{j}(x_{j}):=\frac{1}{N}\sum_{i=1}^{N}I(X_{ij}\leq
x_{j}),\;j\in \{1,\ldots ,r\}\text{.}
\end{equation*}%
We choose the lowest $K$ so that the difference between $\{\hat{F}%
_{N,K}^{j}\}_{j=1}^{r}$ and $\{\hat{F}_{N}^{j}\}_{j=1}^{r}$ is small. Denote
the upper bound of $K$ by $\bar{K}$ (e.g. $\bar{K}=7$ in our simulation
studies). We choose $K\in \{1,...,\bar{K}\}$ to minimize the aggregate
Kolmogorov-Smirnov distance between $\{\hat{F}_{N,K}^{j}\}_{j=1}^{r}$ and $\{%
\hat{F}_{N}^{j}\}_{j=1}^{r}$:
\begin{equation*}
\hat{K}=\arg \min_{K\in \{1,...,\bar{K}\}}{D}_{N}(K)=\sum_{j=1}^{r}%
\sup_{x_{j}\in \mathbb{R}}\left\vert \tilde{F}_{N}^{j}(x_{j})-\hat{F}%
_{N,K}^{j}(x_{j})\right\vert .
\end{equation*}
{\color{black}
\textbf{Higher order MSE approach}. The second approach chooses $K$ to minimize the mean-squared error of the estimator. \cite{donald2009choosing} derives the
higher-order asymptotic mean-square error (MSE) of a linear combination $\mathbf{t}^{\top } \hat{\gamma }$ for some fixed $\mathbf{t} \in \mathbb{R}^{p}$. \\
Let $\check{\gamma }$ be some preliminary estimator. Define:
\begin{align*} & \widehat{\Pi } (K ;\boldsymbol{t}) =\sum _{i =1}^{N}\hat{\xi }_{i i} \rho (T_{i} ,\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma }) \cdot (\mathbf{t}^{\top } \hat{\Omega }_{p \times p}^{ -1} \tilde{\mathbf{\eta }}_{i})\text{,} \\
& \hat{\Phi } (K ;\boldsymbol{t}) =\sum _{i =1}^{N}\hat{\xi }_{i i} \left \{\mathbf{t}^{\top } \hat{\Omega }_{p \times p}^{ -1} \left [\widehat{\mathbf{D}}_{i}^{ \ast } \rho (T_{i} ,\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma })^{2} - \nabla _{\gamma }\rho (T_{i} ,\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma })\right ]\right \}^{2} \\
& \text{\quad \quad \quad \quad \quad \quad } -\mathbf{t}^{\top } \hat{\Omega }_{p \times p}^{ -1} (\hat{\Gamma }_{K \times p})^{\top } \hat{\Upsilon }_{K \times K}^{ -1} \hat{\Gamma }_{K \times p} \hat{\Omega }_{p \times p}^{ -1} \boldsymbol{t}\;\text{.}\end{align*}
where $\rho(T_i,\boldsymbol{X}_i,Y_i;\check{\gamma})$, $\hat{\Omega}_{p\times p}$, $\tilde{\eta}_i$, $\hat{\xi}_{ii}$, $\hat{\boldsymbol{D}}_i^*$, $\hat{\Gamma}_{K\times p}$, and $\hat{\Upsilon}_{K\times K}$ are defined in Section 8.2 of Appendix. Notice that $\widehat{\Pi}(K;\boldsymbol{t})^2/N$ is an estimate of the squared bias term derived in \cite{newey2004higher} and $\hat{\Phi } (K;\boldsymbol{t})$ is the asymptotic variance. \\
The second approach chooses $K$ to minimize the following higher-order MSEs of $\hat{\gamma }_{j} ,j =1 ,\ldots ,p$:
\begin{align}S_{G M M} (K) =\sum _{j =1}^{p}\left \{\frac{1}{N} \widehat{\Pi } (K ;e_{j})^{2} +\hat{\Phi } (K ;e_{j})\right \}\;\text{,} \label{eq:criteria_K}\end{align}where $e_{j}$ is the $j^{t h}$ column of the $p$-dimensional identity matrix. In practice, we set the upper bound $\bar{K}$ and then choose $K \in \{1 ,2 ,\ldots ,\bar{K}\}$ to minimize the criteria \eqref{eq:criteria_K}}
.
\begin{table}
\centering
\caption{Simulation results under Scenorio I}
\begin{threeparttable}
{\fontsize{12pt}{15pt} \selectfont
\begin{tabular}{c|ccccccccccc}
\hline \hline
\multicolumn{10}{c}{$n = 200$} \\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & 0.028 & -0.125 & 0.039 & 0.055 & 0.120 & 0.106 & -0.997 & 0.167 & 0.301 \\
Stdev & 0.254 & 0.413 & 0.129 & 0.229 & 0.272 & 0.118 & 0.197 &0.266 & 0.101 \\
MSE & 0.065 & 0.186 & 0.018 & 0.055 & 0.088 & 0.025 & 1.033 & 0.099 & 0.101 \\
CP & --- & --- & 0.908 & --- & --- & 0.908 &--- &--- & 0.22 \\
\hline \hline
\multicolumn{10}{c}{$n = 500$} \\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & 0.011 & -0.067 & 0.016 & 0.048 & 0.058 & 0.063 & -0.966 & 0.220 & 0.299 \\
Stdev & 0.161 & 0.282 & 0.090 & 0.151 & 0.193 & 0.077 & 0.126 & 0.160 & 0.063 \\
MSE & 0.026 & 0.084 & 0.008 & 0.025 & 0.040 & 0.010 & 0.949 & 0.074 &0.093 \\
CP & --- & --- & 0.928 & --- & --- & 0.892 &--- &--- & 0.034 \\
\hline \hline \multicolumn{10}{c}{}
\\[1mm]
\hline \hline
\multicolumn{10}{c}{$n = 1000$} \\ \hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$\\[1mm]
\hline\\
Bias & 0.005 & -0.040 & 0.008 & 0.034 & 0.023 & 0.040 & -0.962 & 0.235 & 0.298\\
Stdev & 0.103 & 0.187 & 0.065 & 0.102 & 0.132 & 0.055 & 0.078 & 0.099 & 0.045 \\
MSE & 0.010 & 0.036 & 0.004 & 0.011 & 0.018 & 0.004 &0.932 &0.065 &0.091\\
CP & --- & --- & 0.934 & --- & --- & 0.906 & --- & --- &0.012 \\
\hline \hline
\end{tabular}}
{\fontsize{9.5pt}{12pt} \selectfont Stdev: standard deviation ; MSE: mean squared error;
CP: coverage probability. The bandwith used in computing the nonparametric kernel estimators $(\hat{\alpha}_{MK},\hat{\beta}_{MK},\hat{\theta}_{MK})$ is $h=0.15$. }
\end{threeparttable}
\end{table}
\begin{table}
\centering
\caption{Simulation results under Scenorio II}
\begin{threeparttable}
{\fontsize{12pt}{15pt} \selectfont
\begin{tabular}{c|ccccccccccc}
\hline \hline
\multicolumn{10}{c}{$n = 200$} \\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & -0.208 & 0.096 & 0.084 & -0.552 & 0.588 & 0.173 & -2.053 & 1.215 & 0.530 \\
Stdev & 0.646 & 0.555 & 0.201 & 0.372 & 0.245 & 0.125 & 0.809 & 0.148 & 0.205 \\
MSE & 0.462 & 0.318 & 0.047 & 0.443 & 0.406 & 0.045 & 4.873 & 1.498 &0.323 \\
CP & --- & --- & 0.95 & --- & --- & 0.784 &--- &--- & 0.138 \\
\hline \hline
\multicolumn{10}{c}{$n = 500$} \\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & -0.081 & 0.040 & 0.044 & -0.313 & 0.392 & 0.122 & -1.924& 1.203 & 0.583 \\
Stdev & 0.406 & 0.363 & 0.131 & 0.261 & 0.186 & 0.085 & 0.175 & 0.064 & 0.132 \\
MSE & 0.171 & 0.134 & 0.019 & 0.166 & 0.188 & 0.022 &3.732 & 1.451& 0.357 \\
CP & --- & --- & 0.932 & --- & --- & 0.764 &--- &--- & 0.06 \\
\hline \hline \multicolumn{10}{c}{}
\\[1mm]
\hline \hline
\multicolumn{10}{c}{$n = 1000$} \\ \hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$\\[1mm]
\hline\\
Bias & -0.036 & 0.019 & 0.019 & -0.198 & 0.268 & 0.085 &-1.900 & 1.201 & 0.590\\
Stdev & 0.260 & 0.225 & 0.086 & 0.203 & 0.164 & 0.061 & 0.086 & 0.044 & 0.078 \\
MSE & 0.069 & 0.051 & 0.007 & 0.080 & 0.098 & 0.011 & 3.618 & 1.445 & 0.354 \\
CP & --- & --- & 0.932 & --- & --- & 0.768 & --- & --- & 0.018 \\
\hline \hline
\end{tabular}}
{\fontsize{9.5pt}{14pt} \selectfont Stdev: standard deviation ; MSE: mean squared error;
CP: coverage probability. The bandwith used in computing the nonparametric kernel estimators $(\hat{\alpha}_{MK},\hat{\beta}_{MK},\hat{\theta}_{MK})$ is $h=0.05$.
}
\end{threeparttable}
\end{table}
\begin{table}
\caption{Simulation results under Scenorio III}
\centering
\begin{threeparttable}
{\fontsize{12pt}{15pt} \selectfont
\begin{tabular}{c|ccccccccccc}
\hline \hline
\multicolumn{10}{c}{$n = 200$} \\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & 0.155 & -0.171 & 0.003 & 0.047 & 0.015 & 0.071 & -2.794 & 0.954 & -1.146 \\
Stdev & 0.584 & 0.585 & 0.155 & 0.376 & 0.190 & 0.131 & 1.395 & 0.396 & 0.263 \\
MSE & 0.365 & 0.372 & 0.024 & 0.144 & 0.036 & 0.022 & 9.758 & 1.069 & 1.384 \\
CP & --- & --- & 0.934 & --- & --- & 0.884 &--- &--- & 0.032 \\
\hline \hline
\multicolumn{10}{c}{$n = 500$} \\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & 0.034 & -0.036 & 0.000 & 0.012 & 0.012 & 0.034 & 0.782 & 0.355 & 0.123 \\
Stdev & 0.305 & 0.224 & 0.103 & 0.250 & 0.128 & 0.085 & 0.433 & 0.113 & 0.101 \\
MSE & 0.094 & 0.051 & 0.010 & 0.062 & 0.016 & 0.008 & 0.799 & 0.139 & 0.025 \\
CP & --- & --- & 0.902 & --- & --- & 0.894 & ---& --- & 0.698 \\
\hline \hline \multicolumn{10}{c}{}
\\[1mm]
\hline \hline
\multicolumn{10}{c}{$n = 1000$} \\ \hline\\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & 0.009 & -0.010 & 0.002 & 0.002 & 0.009 & 0.017 & 0.728 & 0.372 & 0.126 \\
Stdev & 0.215 & 0.157 & 0.069 & 0.167 & 0.083 & 0.056 & 0.302 &0.078 &0.067 \\
MSE & 0.046 & 0.024 & 0.004 & 0.028 & 0.007 & 0.003 & 0.621 & 0.144 & 0.020 \\
CP & --- & --- & 0.932 & --- & --- & 0.934 & --- &--- & 0.454 \\
\hline \hline
\end{tabular}}
{\fontsize{9.5pt}{14pt} \selectfont Stdev: standard deviation ; MSE: mean squared error;
CP: coverage probability. The bandwith used in computing the nonparametric kernel estimators $(\hat{\alpha}_{MK},\hat{\beta}_{MK},\hat{\theta}_{MK})$ is $h=0.1$.
}
\end{threeparttable}
\end{table}
\begin{table}
\caption{Simulation results under Scenorio IV}
\centering
\begin{threeparttable}
{\fontsize{12pt}{15pt} \selectfont
\begin{tabular}{c|ccccccccccc}
\hline \hline
\multicolumn{10}{c}{$n = 200$} \\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & 0.097 & -0.114 & 0.005 & -0.018 & 0.027 & 0.043 & -1.002 & 1.003 & 0.136 \\
Stdev & 1.140 & 0.721 & 0.118 & 0.308 & 0.185 & 0.103 & 0.081 & 0.139 & 0.348 \\
MSE & 1.310 & 0.533 & 0.014 & 0.095 & 0.035 & 0.013 & 1.011 & 1.026 & 0.139 \\
CP & --- & --- & 0.914 & --- & --- & 0.92 &--- &--- & 0.998 \\
\hline \hline
\multicolumn{10}{c}{$n = 500$}\\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & -0.001 & -0.026 &0.003 & -0.042 & 0.041& 0.022 & -1.003 & 1.000 &0.146 \\
Stdev & 0.203 & 0.139 & 0.071 & 0.172 & 0.100 & 0.067 & 0.048 & 0.088 & 0.199 \\
MSE & 0.041 & 0.020 & 0.005 & 0.031 & 0.011 & 0.005 & 1.010 & 1.009 &0.061 \\
CP & --- & --- & 0.944 & --- & --- & 0.946 & ---&--- & 1.000 \\
\hline \hline \multicolumn{10}{c}{}
\\[1mm]
\hline \hline
\multicolumn{10}{c}{$n = 1000$}\\
\hline \\
& $\hat{\alpha}$& $\hat{\beta}$ & $\hat{\theta}$ & $\hat{\alpha}_{MK}$ & $\hat{\beta}_{MK}$ & $\hat{\theta}_{MK}$ & $\tilde{\alpha}_{MAR}$ & $\tilde{\beta}_{MAR}$ & $\tilde{\theta}_{MAR}$ \\[1mm]
\hline\\
Bias & 0.010 & -0.034 & -0.001 & -0.027 & 0.024 & 0.011 & -1.000 & 0.997 & 0.134 \\
Stdev & 0.262 & 0.264 & 0.052 &0.122 &0.070 & 0.048 & 0.035 & 0.065 & 0.148 \\
MSE & 0.068 & 0.070 & 0.002 & 0.015 & 0.005 &0.002 & 1.003& 1.000 & 0.039\\
CP & --- & --- & 0.936 & --- & --- & 0.932 & ---& ---& 1.000\\
\hline\hline
\end{tabular}}
{\fontsize{9.5pt}{14pt} \selectfont Stdev: standard deviation ; MSE: mean squared error;
CP: coverage probability. The bandwith used in computing the nonparametric kernel estimators $(\hat{\alpha}_{MK},\hat{\beta}_{MK},\hat{\theta}_{MK})$ is $h=0.2$.
}
\end{threeparttable}
\end{table}
\section{Simulations}
After establishing the large sample properties of the proposed estimator, we now evaluate its finite sample performance through a small scale simulation
study. We consider four scenarios. In all scenarios, the parameter of
interest is $\theta _{0}=\mathbb{E}[Y]$ and the sample size is set
respectively at $N=200,500$ and $1000$.
\begin{itemize}
\item {\color{black} \textbf{Scenario I}: $X$ is generated from the normal distribution $N(0,1)$, and the outcome $Y$ is generated from the normal distribution with mean $X+1$ and unit variance, i.e. $Y\sim N(X+1,1)$. The relationship between the outcome variable and the covariate is linear, and the distribution of outcome is normal. The missing mechanism is modeled by
$$\mathbb{P}(T=1|Y,X)=[1+\exp(\alpha_0 +\beta_0 Y)]^{-1}\ ,$$
with the true value $(\alpha_0,\beta_0)=(0,-1.2)$. The true value of the parameter of interest is $\theta_0=\mathbb{E}[Y]=1$}.\\
\item \textbf{Scenario II}: $X$ is generated from the normal distribution $%
N(0,1)$, and the outcome $Y$ is generated from the normal distribution with
mean $X^{2}+1$ and unit variance, i.e. $Y\sim N(X^{2}+1,1)$. Thus the
relationship between the outcome variable and the covariate is nonlinear,
and the distribution of outcome is non-normal. The missing mechanism is
modeled as
\begin{equation*}
\mathbb{P}(T=1|Y,X)=[1+\exp (\alpha _{0}+\beta _{0}Y)]^{-1}\
\end{equation*}%
with the true value $(\alpha _{0},\beta _{0})=(1.25,-1.2)$. The
true value of the parameter of interest is $\theta _{0}=\mathbb{E}[Y]=2$. \\
\item \textbf{Scenario III}. The design follows \cite{qin2002estimation}. We
generate the outcome from
\begin{equation*}
Y=0.1X^{2}+ZX^{1/2}/5\ ,
\end{equation*}%
where $Z$ and $X$ are independent, $Z$ is standard normal random variable,
and $X$ follows the $\chi _{(6)}^{2}/2$ distribution. The missing mechanism
is modeled as
\begin{equation*}
\mathbb{P}(T=1|Y,X)=\left[ 1+\exp (\alpha _{0}+\beta _{0}Y)\right] ^{-1}\
\end{equation*}%
with the true value $(\alpha _{0},\beta _{0})=(3,-1)$. The true value of the
target parameter is $\theta _{0}=\mathbb{E}[Y]=1.2$.\\
\item \textbf{Scenario IV}. The design is similar to that in \cite{kang2007demystifying}. $\boldsymbol{Z}=(Z_{1},Z_{2})$ is generated from the standard
bivariate normal distribution, and $Y$ is generated from the normal
distribution with mean $2+Z_{1}$ and unit variance. The missing mechanism is
modeled as
\begin{equation*}
\mathbb{P}(T=1|Y,X_{1},X_{2})=[1+\exp (\alpha _{0}Z_{1}+\beta _{0}Y)]^{-1}\
\end{equation*}%
with $(\alpha _{0},\beta _{0})=(1,-1)$. The true value of the parameter of
interest is $\theta _{0}=\mathbb{E}[Y]=2$. Instead of directly observing
covariates $\boldsymbol{Z}$, we observe a non-linear transformation of $%
\boldsymbol{Z}$: $X_{1}=\exp (Z_{1}/2)$ and $X_{2}=Z_{2}/(1+\exp (Z_{1}))$.
\end{itemize} \
In all scenarios, we generate $J=500$ random samples, and for each sample, we compute the following three estimators:
\begin{enumerate}
\item Naive estimator. We compute the missing at random estimator $(%
\tilde{\alpha}_{MAR},\tilde{\beta}_{MAR},\tilde{\theta}_{MAR})$ as
\begin{equation*}
\tilde{\theta}_{MAR}=\frac{1}{N}\sum_{i=1}^{N}\frac{T_{i}}{\pi (\boldsymbol{X}%
_{i};\tilde{\alpha}_{MAR},\tilde{\beta}_{MAR})}Y_{i}\ ,
\end{equation*}%
where $\pi (\boldsymbol{X}_{i};\tilde{\alpha}_{MAR},\tilde{\beta}_{MAR})$ is
an estimated response model. In Scenarios I, II \& III, $%
\pi (\boldsymbol{X}_{i};\tilde{\alpha}_{MAR},\tilde{\beta}_{MAR})=\left[
1+\exp (\tilde{\alpha}_{MAR}+\tilde{\beta}_{MAR}X_{i})\right] ^{-1}$ and
in Scenario IV $\pi (\boldsymbol{X}_{i};\tilde{\alpha}_{MAR},\tilde{\beta}%
_{MAR})=\left[ 1+\exp (\tilde{\alpha}_{MAR}Z_{1i}+\tilde{\beta}_{MAR}X_{2i})%
\right] ^{-1}$, where $(\tilde{\alpha}_{MAR},\tilde{\beta}_{MAR})$ are
estimated by GMM.
\item MK2 estimator. We compute $(\hat{\alpha}_{MK},\hat{\beta}_{MK},\hat{%
\theta}_{MK})$ using the approach of \cite{morikawa2016semiparametric}, i.e. $(\hat{\alpha}_{MK},\hat{\beta}_{MK},\hat{%
\theta}_{MK})$ is the solution of $$\sum_{i=1}^N\left(\hat{\boldsymbol{S}}_{1}(T_i,\boldsymbol{Z}_i;\alpha,\beta)^{\top}, \hat{S}_2(T_i,\boldsymbol{Z}_i;\alpha,\beta,\theta)\right)^{\top}=0\ ,$$ where
\begin{align*}
&\hat{\boldsymbol{S}}_{1}(T,\boldsymbol{Z};\alpha,\beta)=-\left(1-\frac{T}{\pi(\boldsymbol{Z};\alpha,\beta)}\right)\mathbb{E}^\star\left[\frac{\nabla_{\gamma}\pi(\boldsymbol{Z};\alpha,\beta)}{1-\pi(\boldsymbol{Z};\alpha,\beta)}\bigg|\boldsymbol{X}\right]\ , \\
&\hat{S}_2(T,\boldsymbol{Z};\alpha,\beta,\theta)=-\frac{T}{\pi(\boldsymbol{Z};\alpha,\beta)}U(\boldsymbol{Z})+\theta-\left(1-\frac{T}{\pi(\boldsymbol{Z};\alpha,\beta)}\right) \mathbb{E}^\star\left[U(\boldsymbol{Z})|\boldsymbol{X}\right]\ ,
\end{align*}
and for any function $g(\boldsymbol{Z})$ the quantity $\mathbb{E}^\star[g(\boldsymbol{Z})|\boldsymbol{X}]$ is defined by
\begin{align*}
&\mathbb{E}^\star[g(\boldsymbol{Z})|\boldsymbol{X}=\boldsymbol{x}]:=\frac{\sum_{j=1}^NT_jK_h(\boldsymbol{x}-\boldsymbol{X}_j)T_j\pi(\boldsymbol{Z}_j;\alpha,\beta)^{-1}O(\boldsymbol{x},Y_j;\alpha,\beta)g(\boldsymbol{x},Y_j)}{\sum_{j=1}^NK_h(\boldsymbol{x}-\boldsymbol{X}_j)T_j\pi(\boldsymbol{Z}_j;\alpha,\beta)^{-1}O(\boldsymbol{x},Y_j;\alpha,\beta)}\ ; \\
&O(\boldsymbol{z};\alpha,\beta)=\frac{1-\pi(\boldsymbol{z};\alpha,\beta)}{\pi(\boldsymbol{z};\alpha,\beta)} \ ,
\end{align*}
$K_h(\boldsymbol{x}-\boldsymbol{w})=K\left((\boldsymbol{x}-\boldsymbol{w}/h)\right)$, $K(\cdot)$ is Gaussian kernel function and $h$ is the bandwidth.
\item Our GMM estimator. We compute $(\hat{\alpha},\hat{\beta},\hat{\theta})$
using the proposed approach and the covariate balancing approach to select $K
$, {\color{black}with $\bar{K}=7$ in Scenarios I, II, III, and with $\bar{K}=10$ in Scenario IV. Here $\bar{K}$ is the maximal number of candidate moments to be considered.}
\end{enumerate}
The simulation results (the bias, the standard deviation (Stdev), the mean squared error (MSE), and the coverage probability (CP) (for significance level $\alpha =0.05$) of the point estimates) for all scenarios are reported in Tables 1, 2, 3, and 4 respectively. {\color{black} The histogram of selected $K'$s (based on $500$ Monte Carlo samples) in all scenarios is reported in Figure 1.} Glancing at these tables, we find:
\begin{enumerate}
\item In all scenarios, the naive estimator using the missing at random assumption has a large bias, because this assumption does not hold.
\item In all scenarios, our proposed estimator of $\mathbb{E}[Y]$
out-performs the MK estimator.
\item In all scenarios, our proposed variance estimator has coverage probability close to $95\%$, even the sample size is small. The MK's variance estimator performs well in Scenario IV, but badly in other scenarios: in Scenarios I, the coverage probability using MK's approach converges to $90\%$ rather than $95\%$; in Scenarios II, the CP values are far from $95\%$ in Scenario 2 no matter the sample size is small or large; in Scenarios III, the MK's variance estimaotr is consistent only when the sample size is large.
\item {\color{black}When the sample size is small the optimal $K$ tends to be $2$. When the sample size is large, the optimal $K$ tends to be $3$. The growing rate of $K$ is extremely slow comparing to that of the sample size $n$, which is consistent with our theoretical Assumption \ref{as:K&N}.}
\end{enumerate}
These results clearly show that the proposed approach has better finite sample performance.
\begin{figure}[!htp]
\caption{Histogram of $K$}
\begin{center}
\includegraphics[height=21cm,width=15cm, scale=1.0]{hist.png}
\footnotesize \\[4mm] The Monte Carlo sample size used to plot the histogram of $K$ is $J=500$.
\end{center}
\end{figure}
\newpage
\section{Discussion}
The data missing not at random problem is common in applications. \cite
{morikawa2016semiparametric} studies the efficient estimation of a class of missing not at random problems. But their approach requires nonparametric estimation of the conditional density function and thus suffers from the curse of dimensionality and smoothing parameter selection problem. In this
paper, we study the same class of missing not at random problems but present a much simpler and more natural efficient estimator. Our approach is based
on a parametric moment restriction model that does not require nonparametric estimation and hence does not suffer from the curse of dimensionality problem nor the bandwidth selection problem. Indeed the simulation results confirm that the proposed approach out-performs their approach in finite samples. The GMM approach is also easy to adapt to stratified sampling and other sampling schemes common in survey data. \\
Both approaches require correct parameterization of the propensity score function. If the propensity score function is misspecified, then both approaches yield inconsistent estimates. There is some attempt in the literature to mitigate this problem. For instance, Zhao and Shao (2015) introduce a partial linear index to model missing mechanism. The proposed approach can be extended in this direction. Such extension shall be pursued in a future study.
\section{Appendix}
\subsection{Assumptions}
We first introduce the smoothness classes of functions used in the nonparametric estimation; see e.g. \cite{stone1982optimal, stone1994use}, \cite{robinson1988root}, \cite{newey1997convergence}, \cite{horowitz2012semiparametric} and \cite{chen2007large}. Suppose that $\mathcal{X}$ is the Cartesian product of $r$-compact intervals. Let $0<\delta \leq 1$. A fucntion $f$ on $\mathcal{X}$ is said to satisfy a H$\ddot{\text{o}}$lder condition with exponent $\delta$ if there is a positive constant $L$ usch that $\|f(\boldsymbol{x}_1)-f(\boldsymbol{x}_2)\|\leq L \|\boldsymbol{x}_1-\boldsymbol{x}_2\|^{\delta}$ for all $\boldsymbol{x}_1,\boldsymbol{x}_2\in\mathcal{X}$. Given a $r$-tuple $\boldsymbol{\alpha}=(\alpha_1,...,\alpha_r)$ of nonnegative integer, denote $[\boldsymbol{\alpha}]=\alpha_1+\cdots+\alpha_r$ and let $D^{\boldsymbol{\alpha}}$ denote the differential operator defined by $D^{\boldsymbol{\alpha}}=\frac{\partial^{[\boldsymbol{\alpha}]}}{\partial x_1^{\alpha_1}\cdots \partial x_r^{\alpha_r}}$, where $\boldsymbol{x}=( x_1,..., x_r)$.
\begin{definition}
\emph{Let $s$ be a nonnegative integer and $s:=s_0+\delta$. The function $f$ on $\mathcal{X}$ is said to be} $s$-smooth \emph{if it is $s$ times continuously differentiable on $\mathcal{X}$ and $D^{\boldsymbol{\alpha}}f$ satisfies a H$\ddot{\text{o}}$lder condition with exponent $\delta$ for all $\boldsymbol{\alpha}$ with $[\boldsymbol{\alpha}]=s_0$}.
\end{definition}
The following notations are needed for presenting the efficiency bounds:
\begin{align} & O (\boldsymbol{Z}):=\frac{1 -\pi (\boldsymbol{Z} ;\gamma _{0})}{\pi (\boldsymbol{Z} ;\gamma _{0})} ,\text{\quad }\boldsymbol{S}_{0} (\boldsymbol{Z}):= -\frac{ \nabla _{\gamma }\pi (\boldsymbol{Z} ;\gamma _{0})}{1 -\pi (\boldsymbol{Z} ;\gamma _{0})} \label{def:os}\ , \\
&m(\boldsymbol{X}):=\frac{\mathbb{E}[O(\boldsymbol{Z})\boldsymbol{S}_{0} (\boldsymbol{Z})\vert \boldsymbol{X}]}{\mathbb{E}[O(\boldsymbol{Z})\vert \boldsymbol{X}]} ,\text{\quad }R (\boldsymbol{X}):=\frac{\mathbb{E}[O(\boldsymbol{Z})U (\boldsymbol{Z})\vert \boldsymbol{X}]}{\mathbb{E}[O(\boldsymbol{Z})\vert \boldsymbol{X}]}\; \label{def:osmr}\ , \\
& \boldsymbol{S}_{1} (T ,\boldsymbol{Z} ;\gamma _{0}) :=\left (1 -\frac{T}{\pi (\boldsymbol{Z} ;\gamma _{0})}\right ) m (\boldsymbol{X})\;\text{,} \label{def:S1}\\
& S_{2} (T ,\boldsymbol{Z} ;\gamma _{0} ,\theta _{0}) := -\frac{T}{\pi (\boldsymbol{Z} ;\gamma _{0})} U (\boldsymbol{Z}) +\theta _{0} -\left (1 -\frac{T}{\pi (\boldsymbol{Z} ;\gamma _{0})}\right ) R (\boldsymbol{X})\;\text{.} \label{def:S2}
\end{align}
The following assumptions are required in this paper:
\begin{assumption}
\label{as:id} There exists a nonresponse instrumental variable $\mathbf{X}_{\mathbf{2}}$, i.e., $\mathbf{X} =(\mathbf{X}_{1}^{^{ \intercal }} ,\mathbf{X}_{2}^{^{ \intercal }})^{^{ \intercal }}\text{,}$ such that $\boldsymbol{X}_{2}$ is independent of $T$ given $\boldsymbol{X}_{1}$ and $Y$; furthermore, $\boldsymbol{X}_{2}$ is correlated with $Y$.
\end{assumption}
\begin{assumption}
\label{as:suppX} The support of $\mathbf{X}$, which is denoted by $\mathcal{X}$, is a Cartesian product of $r$-compact intervals, and we denote $\boldsymbol{X}=(X_1,...,X_r)^{\top}$.
\end{assumption}
\begin{assumption}
\label{as:proj_smooth} The functions $\mathbb{E}[O(\boldsymbol{Z})S_{0} (\boldsymbol{Z})\vert
\bold {X}
=\boldsymbol{x}]$, $\mathbb{E}[O(\boldsymbol{Z})U (\boldsymbol{Z})\vert
\bold {X}
=\boldsymbol{x}]$ and $\mathbb{E}[O(\boldsymbol{Z})\vert \boldsymbol{X} =\boldsymbol{x}]$ are $s$-smooth in $\boldsymbol{x}$, where $s > 0$.
\end{assumption}
\begin{assumption}\label{as:eigen}
There exists two finite positive constants $\underline{a}$ and $\overline{a}$ such that the smallest (resp. largest) eigenvalue of $\mathbb{E}[u_K(\boldsymbol{X})u_K^{\top}(\boldsymbol{X})]$ is bounded away from $\underline{a}$ (resp. $\overline{a}$) uniformly in $K$, i.e., $$0<\underline{a}\leq \lambda_{\min}(\mathbb{E}[u_{K}(\boldsymbol{X})u_{K} (\boldsymbol{X})^{\top }])\leq \lambda_{\max}(\mathbb{E}[u_{K}(\boldsymbol{X})u_{K} (\boldsymbol{X})^{\top }])\leq \overline{a}<\infty\ .$$
\end{assumption}
\textbf{Remark}: Asssumption \ref{as:eigen} implies that following results:
\begin{enumerate}
\item \begin{align}\label{bound:Eu_K^2}
\mathbb{E}[\|u_K(\boldsymbol{X})\|^2]=\tr\left(\mathbb{E}\left[u_K(\boldsymbol{X})u_K(\boldsymbol{X})^{\top}\right]\right)=O(K)\ ;
\end{align}
\item the matrices
$\bar{a}\cdot I_{K\times K}-\mathbb{E}[u_{K}(\boldsymbol{X})u_{K} (\boldsymbol{X})^{\top }]$ and $\mathbb{E}[u_{K}(\boldsymbol{X})u_{K} (\boldsymbol{X})^{\top }]-\underline{a}\cdot I_{K\times K}$ are positive definite, and
\begin{align}\label{bound_u_kK}
\underline{a}\leq \inf_{k\in\{1,...,K\}}\mathbb{E}[u_{kK}(\boldsymbol{X})^2]\leq \sup_{k\in\{1,...,K\}}\mathbb{E}[u_{kK}(\boldsymbol{X})^2] \leq \overline{a}\ .
\end{align}
\end{enumerate}
\begin{assumption}
\label{as:iid}The full data $\{(T_{i} ,\boldsymbol{X}_{i} ,Y_{i})\}_{i =1}^{N}$ are independently and identically distributed.
\end{assumption}
\begin{assumption}
\label{as:boundS}$ \boldsymbol{S}_{eff}(T,\boldsymbol{Z};\gamma,\theta%
):=(\boldsymbol{S}_{1}^{^{\intercal}}(T,\boldsymbol{Z};\gamma
),S_{2}(T,\boldsymbol{Z};\gamma,\theta))^{^{\intercal}}$
is continuously differentiable
at each $(\gamma ,\theta ) \in \Gamma \times \Theta $ with probability one, and $\mathbb{E} \left [ \partial \boldsymbol{S}_{e f f} (\gamma ,\theta )/ \partial (\gamma ^{\top } ,\theta )\right ]$ is nonsingular at $(\gamma _{0} ,\theta _{0})$.
\end{assumption}
\begin{assumption}
\label{as:pi} The response probability $\pi $ satisfies the following conditions:
\begin{enumerate}
\item there exists two positive constants $\bar{c}$ and $\underline{c}$ such that $0 <\underline{c} \leq \pi (\boldsymbol{x} ,y ;\gamma ) \leq \bar{c} <1$ for all $\gamma \in \Gamma $ and $(\boldsymbol{x} ,y) \in \mathcal{X} \times \mathbb{R}$;
\item the propensity score $\pi (\boldsymbol{x},y;\gamma )$ is twice continuously differentiable in $\gamma \in \Gamma $, and the derivatives are uniformly bounded.
\item for any $\gamma\in \Gamma$, the conditional functions $\mathbb{E}\left[1-\frac{T}{\pi(\boldsymbol{Z};\gamma)}|\boldsymbol{X}=\boldsymbol{x}\right]$ and $\mathbb{E}\left[\frac{\nabla_{\gamma}\pi(\boldsymbol{Z};\gamma)}{\pi(\boldsymbol{Z};\gamma)}\bigg|\boldsymbol{X}=\boldsymbol{x}\right]$ are $s$-smooth in $\boldsymbol{x}$, where $s> 0$.
\end{enumerate}
\end{assumption}
\begin{assumption}
\label{as:K&N} Suppose $K \rightarrow \infty $ and {\color{black}$K^3/N \rightarrow 0$}.
\end{assumption}
The Assumption \ref{as:id} is used for the identification
of the model, which was discussed in \cite{wang2014instrumental}. Assumptions
\ref{as:suppX} and \ref{as:proj_smooth} are required for uniform boundedness
of approximations. Assumption \ref{as:eigen} is a standard assumption used in nonparametric sieve approximation, see also \cite{newey1997convergence}. Assumption \ref{as:iid} is a standard condition for statistical sampling. Assumptions \ref{as:boundS}-\ref{as:pi}
are required for the convergence of our estimator as well as the boundness of the asymptotic variance. Assumption \ref{as:K&N}
is the same as Assumption 2 in \cite{newey2003instrumental}, it is required
for controlling the stochastic order of the residual terms, which is desirable in practice because $K$ grows very slowly with $N$ so a relatively small number of moment conditions is sufficient for the method proposed to perform well.
{\color{black}\subsection{Discussion on $u_K$}\label{sec:uK}
To construct the GMM estimator, we need to specify the matching function $u_K(\boldsymbol{X})$. Although the approximation theory is derived for general sequences of approximating functions, the most common class of functions
are power series. Suppose the dimension of covariate $\boldsymbol{X}$ is $r\in\mathbb{N}$, namely $\boldsymbol{X}=(X_1,...,X_r)^{\top}$. Let $\lambda =(\lambda _{1} ,\ldots \lambda _{r})^{\top }$ be an $r$-dimensional vector of nonnegative integers (multi-indices), with norm $\vert \lambda \vert =\sum _{j =1}^{r}\lambda _{j}$. Let $(\lambda (k))_{k =1}^{\infty }$ be a sequence that includes all distinct multi-indicesand satisfies $\vert \lambda (k)\vert \leq \vert \lambda (k +1)\vert $, and let $\boldsymbol{X}^{\lambda } =\prod _{j =1}^{r}X_{j}^{\lambda _{j}}$. For a sequence $\lambda (k)$ we consider the series $u_{k K} (\boldsymbol{X}) =\boldsymbol{X}^{\lambda (k)}$, $k\in\{1,...,K\}$. \cite{newey1997convergence}
showed the following property for the power series: there exists a universal constant $C >0$ such that
\begin{align}\zeta (K):=\underset{\boldsymbol{x} \in \mathcal{X}}{\sup }\Vert u_{K} (\boldsymbol{x})\Vert \leq C K \label{eq:zeta}\ , \end{align}where $\Vert \cdot \Vert $ denotes the usual matrix norm $\Vert A\Vert =\sqrt{\tr (A^{\top } A)}$. \\
Another important issue is choosing the number of the matching function $K$ in finite sample experiment. \cite{donald2009choosing} proposed a strategy for an appropriate choice of $K$ by minimizing the higher order MSE defined in \eqref{eq:criteria_K}, and the following notations are needed to describe this criteria:
\begin{align*} & \rho (T_{i} ,\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma }) =1 -\frac{T_{i}}{\pi (\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma })} ,\text{\quad }\hat{\Upsilon }_{K \times K} =\frac{1}{N} \sum _{i =1}^{N}\rho (T_{i} ,\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma })^{2} u_{K} (\boldsymbol{X}_{i})^{ \otimes 2}\text{,} \\
& \widehat{\Gamma }_{K \times p} =\frac{1}{N} \sum _{i =1}^{N}u_{K} (\boldsymbol{X}_{i}) \nabla _{\gamma }\rho (T_{i} ,\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma })^{\top } ,\text{\quad }\hat{\Omega }_{p \times p} =(\hat{\Gamma }_{K \times p})^{\top } \hat{\Upsilon }_{K \times K}^{ -1} \hat{\Gamma }_{K \times p}\text{,} \\
& \tilde{\mathbf{d}}_{i} =(\hat{\Gamma }_{K \times p})^{\top } \left (\frac{1}{N} \sum _{j =1}^{N}u_{K} (\boldsymbol{X}_{j})^{ \otimes 2}\right )^{ -1} u_{K} (\boldsymbol{X}_{i}) ,\text{\quad }\tilde{\mathbf{\eta }}_{i} = \nabla _{\gamma }\rho (T_{i} ,\boldsymbol{X}_{i} ,Y_{i} ;\check{\gamma }) -\tilde{\mathbf{d}}_{i}\text{,}
\end{align*}
\begin{align*}
& \hat{\xi }_{i j} =\frac{1}{N} u_{K} (\boldsymbol{X}_{i})^{\top } \hat{\Upsilon }_{K \times K}^{-1} u_{K} (\boldsymbol{X}_{j}) ,\text{\quad }\widehat{
\bold {D}
}_{i}^{ \ast } =(\hat{\Gamma }_{K \times p})^{\top } \hat{\Upsilon }_{K \times K}^{ -1} u_{K} (\boldsymbol{X}_{i})\;\text{.}\end{align*}
\subsection{Semiparametric Efficiency Bounds}
The following lemma is Theorem 1 in \cite{morikawa2016semiparametric}.
\begin{lemma}[\cite{morikawa2016semiparametric}]
\label{lemma:MK} The efficient variance bounds of $(\gamma _{0} ,\theta _{0})$ is $\boldsymbol{V}_{eff}:=\mathbb{E}\left[\boldsymbol{S}_{eff}(T,\bold{Z};\gamma_0,\theta_0)^{\otimes 2}\right]^{-1}$, where $\boldsymbol{S}_{eff}=(\boldsymbol{S}_1^{\top},S_2)^{\top}$ and $\boldsymbol{S}_1$, $S_2$ are defined in \eqref{def:S1} and \eqref{def:S2} respectively.
\end{lemma}
Let $\boldsymbol{V}_{\gamma _{0}}$ (resp. $V_{\theta _{0}}$) be the efficient variance bound of $\gamma_0$ (resp. $\theta_0$). After some simple computation, we can find out
\begin{align}\boldsymbol{V}_{\gamma _{0}} = & \mathbb{E} \left [\frac{1 -\pi (\boldsymbol{Z} ;\gamma _{0})}{\pi (\boldsymbol{Z} ;\gamma _{0})} m (\boldsymbol{X})^{ \otimes 2}\right ]^{ -1} \label{effbound_gamma}\end{align}
and
\begin{align}V_{\theta _{0}} =V a r \left (S_{2} (T ,\boldsymbol{Z} ;\gamma _{0} ,\theta _{0}) -\mathbf{\kappa }^{^{ \intercal }} \boldsymbol{S}_{1} (T ,\boldsymbol{Z} ;\gamma _{0})\right )\;\text{.} \label{effbound_theta}\end{align}
where
\begin{align}\mathbf{\kappa }^{^{ \intercal }}:=\mathbb{E} \left [\frac{ \nabla _{\gamma }\pi (\boldsymbol{Z} ;\gamma _{0})^{^{ \intercal }}}{\pi (\boldsymbol{Z} ;\gamma _{0})} \{R (\boldsymbol{Z}) -U (\boldsymbol{X})\}\right ] \cdot \mathbb{E} \left [\frac{m (\boldsymbol{X})}{\pi (\boldsymbol{Z} ;\gamma _{0})} \nabla _{\gamma }\pi (\boldsymbol{Z} ;\gamma _{0})^{^{ \intercal }}\right ]^{ -1}\;\text{.} \label{def:kappa}\end{align}
\bibliographystyle{dcu}
|
\section{Introduction}\label{sec_introduction}
Compact smooth toric varieties $X$, and their topological
counterparts, toric manifolds\footnote{Though Davis and
Januszkiewicz \cite{DJ} introduced the name toric manifolds,
in recent literature they are sometimes called {\em quasitoric
manifolds\/}.}, have their integral cohomology torsion free
and concentrated in even degrees. Consequently, the action
of the compact $n$-dimensional torus $T = (S^1)^n$ on $X$,
allows for a satisfying description of the integral cohomology
ring arising from the collapsing Serre spectral sequence of
the canonical fibration
\begin{equation}
X \overset{\iota}{\hookrightarrow} ET \times_{T} X \xrightarrow{\pi} BT\cong (\mathbb{C}P^\infty)^{n}.
\end{equation}
Franz and Puppe note in \cite[Theorem 1.1]{FP}, that for
compact smooth toric varieties, or more generally, a finite
$T$-CW complexes, this is equivalent to the map $\iota^{\ast}$
inducing an isomorphism
\begin{equation}\label{btfree}
H^{\ast}\big(ET \times_{T} X\big)\otimes_{H^{\ast}(BT)}\mathbb{Z} \longrightarrow H^{\ast}(X).
\end{equation}
\noindent The term {\em integrally equivariantly formal\/} suggests
itself for $X$, by analogy with the case of rational coefficients,
see for instance \cite{GKM}.
In this smooth setting, \eqref{btfree} leads to a description of
the ring $H^{\ast}(X)$ as a quotient of a Stanley--Reisner ring
by a linear ideal.
The situation for {\em singular\/} toric spaces is not so nice;
examples with non-vanishing cohomology in odd degrees
abound, \cite{Fis}, \cite[Chapter 12]{CLS}. In the case of
orbifolds however, the results of \cite{BNSS} establish
tractable sufficient conditions ensuring that the integral
cohomology is torsion free and concentrated in even degree.
So, for such spaces an integral equivariant formality holds
and the ring structure of the cohomology can be identified
explicitly, in a manner entirely analogous to the smooth case
described above; the Stanley--Reisner ring is replaced by a ring
of certain {\em piecewise polynomials} or \emph{weighted
Stanley--Reisner rings}, see \cite[Proposition 2.2]{BFR} and
\cite[Theorem 5.3]{BSS}.
In this sequel to \cite{BNSS}, we extend the program to a
class of infinite families of toric orbifolds, derived from a given
one by a combinatorial construction, known variously as:
{\em the simplicial wedge construction\/} \cite{PB}, \cite{Ew},
{\em the doubling construction\/} \cite{LDM} and
{\em the J-construction\/} \cite{BBCG18, BBCG15}.
The construction associates to a sequence of positive
integers $J = (j_1,j_2,\ldots,j_m)$ and an $(n-1)$-dimensional
simplicial complex $K$ on $m$ vertices, a new simplicial complex
$K(J)$ on $d(J) = j_1 + j_2 + \cdots + j_m$ vertices, of dimension
$d(J)-m+n-1$. Equally well, it associates to a simple polytope
$Q$ of dimension $n$ with $m$ facets, a new simple polytope
$Q(J)$ of dimension $d(J)-m+n$ having $d(J)$ facets.
(Notice that $m-n$ = $d(J) - (d(J)-m+n)$.)
As outlined in Section \ref{sec_toric_orb_orb_lens} below,
a $2n$-dimensional toric orbifold $X(Q,\lambda)$ is specified by an
$\mathcal{R}$-characteristic pair $(Q,\lambda)$ where $Q$ is an
$n$-dimensional simple polytope and $\lambda\colon \mathcal{F}(Q)
\to \mathbb{Z}^{n}$ is a function from the set of facets
of $Q$ which satisfies certain conditions.
Section \ref{sec_J-const} describes the way in which each sequence
$J$, determines from $X(Q,\lambda)$ a new
$2\big(d(J)-m+n\big)$-dimensional toric orbifold
$X_{(J)} \mathrel{\mathop:}= X(Q_{(J)},\lambda_{(J)})$.
Our goal here is two-fold:\\[-6mm]
\begin{enumerate}
\item To confirm that if $X(Q,\lambda)$ satisfies the sufficiency
conditions of \cite[Theorem 1.1]{BNSS}, which ensure that the
integral cohomology is torsion free and concentrated in even
degree, then as $J$ varies, the spaces $X(Q_{(J)},\lambda_{(J)})$
yield an infinite family of similarly integrally equivariantly formal
orbifolds.
\item To relate the integral cohomology ring of $X_{(J)}$ to that of $X$.
\end{enumerate}
The elementary theory of toric orbifolds is reviewed in Section
\ref{sec_toric_orb_orb_lens}, drawing from the expositions to be
found in \cite{DJ}, \cite{PS} and \cite{BSS}. The emphasis is on
tracking the finite isotropy groups and the singularities resulting
from the failure of the $\mathcal{R}$--characteristic function
$\lambda$ to satisfy the regularity condition which ensures the
smoothness of a toric space.
The orbifold analogue of a CW-complex, called a {\bf q}-CW
complex, developed rationally in the toric space setting by
Poddar and Sarkar \cite{PS} and integrally in \cite{BNSS},
is reviewed in Section \ref{sec_bdseq_and_q-cell}.
The {\bf q}-CW complex structure is connected to the underlying
combinatorics by the notion of a {\em retraction sequence\/}
for a simple polytope, introduced in \cite{BSS}. The outcome
of these observations allows for an iterated construction of
the toric orbifold via a sequence of cofibrations which keep
track of the isotropy, and hence singularities, as they arise.
A condition involving the divisibility of all the orders of the
isotropy groups, which emerge from {\em particular\/} retraction
sequences, proves sufficient to establish the integral
equivariant formality of the toric orbifold. The cornerstone
of several of the results presented here, is the next theorem.
\skp{0.05}
\noindent {\bf Theorem \ref{thm_no-torsion}} \cite[Theorem 4.6]{BNSS}
{\em Let\/} $X := X(Q,\lambda)$ {\em be a toric orbifold.\/}
{\em If for each prime number $p$ there is a retraction sequence\/}
$\{(B_j, E_j, b_j)\}_{j=1}^{\ell}$ {\em such that\/}
gcd$\{p,|G_{E_j}(b_j)|\} = 1$ {\em for all\/} $j$, {\em then\/}
$H_{\ast}(X; \mathbb{Z})$ {\em has no torsion and
$H_{\text{odd}}(X; \mathbb{Z})$ is trivial\/}.
\skp{0.05}
In Section \ref{sec_J-const}, various formulations of the simplicial
wedge construction are introduced on both simplicial complexes
and simple polytopes. This is followed by a description of the
transition of $\mathcal{R}$-characteristic pairs
$$X(Q,\lambda) \rightsquigarrow X(Q_{(J)},\lambda_{(J)})$$
\noindent for each sequence $J = (j_1,j_2,\ldots,j_m) \in \mathbb{N}^m$.
Included here is the verification, in Lemma
\ref{lem_lambda(J)_satisfies_the_condition}, that if the
$\mathcal{R}$-characteristic map $\lambda$ satisfies the
orbifold condition \eqref{def_R-char_fun}, then $\lambda_{(J)}$
does too. This is a modification of the argument for the smooth
case, \cite[Theorem 3.2]{BBCG15}. The construction of
$Q_{(J)}$ from $Q$ is an iterative process involving a sequence
of ``doubling'' operations. Though it is well known that the result
is independent of the particular sequence chosen, we include here
Proposition \ref{prop_(2,1,...1)(2,1,...,1)=(3,1,...,1)} and the details
of a proof for completeness.
Section \ref{sec_homology_of_X_{(J)}} is devoted to the homology
groups of the toric orbifold $X_{(J)}:=X(Q_{(J)}, \lambda_{(J)})$;
the main theorem is the following.
\skp{0.05}
\noindent {\bf Theorem \ref{thm_X_satisfies_assump_then_X(J)_eq_formal}}
{\it Let $X := X(Q,\lambda)$ be a toric orbifold satisfying the
assumption of Theorem $3.6$. Then, $H_{\ast}(X_{(J)}; \mathbb{Z})$
is torsion free and $H_{\text{odd}}(X_{(J)}; \mathbb{Z})$ vanishes
for arbitrary $J = (j_1,j_2,\ldots,j_m) \in \mathbb{N}^m$}.
\skp{0.05}
The paper concludes with a discussion of the integral cohomology
ring of the toric orbifolds $X_{(J)}$. When the toric orbifold
$X(Q,\lambda)$ satisfies the hypothesis of Theorem
\ref{thm_no-torsion}, then Theorem \ref{thm_cohom_of_X(J)}
characterizes the cohomology of $X_{(J)}$ as
the quotient of a certain {\em weighted\/} Stanley--Reisner ring
$w\mathcal{SR}[Q_{(J)},\lambda_{(J)}]$ by a linear ideal which
depends on the $\mathcal{R}$--characteristic map $\lambda$.
The final result relates the ring
$w\mathcal{SR}[Q_{(J)},\lambda_{(J)}]$ to the ring $w\mathcal{SR}[Q,\lambda]$.
\subsection*{Acknowledgments}
This work was supported in part by grants 210386 and 426160 from Simons Foundation.
The third author has been supported by
the POSCO Science Fellowship of POSCO TJ Park Foundation.
\section{Toric orbifolds}
\label{sec_toric_orb_orb_lens}
In this section, we review the basic theory of toric orbifolds \cite{DJ, PS} constructed from
a combinatorial information called an $\mathcal{R}$-characteristic pair \cite{BSS}.
Given an $n$-dimensional simple convex polytope $Q$, let
$V(Q)=\{v_1, \dots, v_\ell\}$ be the set of vertices and
$\mathcal{F}(Q)=\{F_1, \dots, F_m\}$ the set consisting of codimension
1 faces called \emph{facets} of $Q$.
\begin{definition}
A pair $(Q, \lambda)$ consisting of an $n$-dimensional simple polytope $Q$ and
a function $\lambda \colon \mathcal{F}(Q)\to \mathbb{Z}^n$,
is called an $\mathcal{R}$-\emph{characteristic pair} if the following condition is satisfied:
\begin{align}\label{def_R-char_fun}
\{\lambda(F_{i_1}), \dots, \lambda(F_{i_k})\}\text{ is a linearly independent set, whenever }
\bigcap_{j=1}^k F_{i_j}\neq \emptyset.
\end{align}
In this case, we call $\lambda$ an $\mathcal{R}$-\emph{characteristic function} on $Q$.
\end{definition}
An $\mathcal{R}$-characteristic function is often represented by an
$n\times m$ matrix whose $i$-th column vector is the transpose of $\lambda(F_i)$
for $i=1, \dots, m$. We call
this matrix the \emph{characteristic matrix} associated to $\lambda$.
Given an $(n-k)$-dimensional face $E= F_{i_1}\cap \dots \cap F_{i_k}$ of $Q$
for $k\geq 1$,
let $M(E)$ be the $\mathbb{Z}$-submodule of $\mathbb{Z}^n$
generated by the set $\{ \lambda(F_{i_1}), \dots, \lambda(F_{i_k})\}$.
Then, $M(E)$ induces a free $\mathbb{Z}$-submodule
$\left(M(E) \otimes_\mathbb{Z} \mathbb{R} \right) \cap \mathbb{Z}^n$ of rank $k$ in $\mathbb{Z}^n$.
Hence, we can define a natural projection
$$\rho_E \colon \mathbb{Z}^n \to \mathbb{Z}^{n-k} \cong \mathbb{Z}^n/
( \left(M(E) \otimes_\mathbb{Z} \mathbb{R} \right) \cap \mathbb{Z}^n).$$
Next, set
$$ \mathcal{F}(E):=
\{E \cap F_j \mid F_j\notin \{F_{i_1}, \dots, F_{i_k}\} \text{ and }
F_j\cap E\neq \emptyset\}.$$
Now, we define a function
\begin{equation}\label{eq_lambda_E}
\lambda_E \colon \mathcal{F}(E) \to \mathbb{Z}^{n-k}
\quad \text{by}\quad \lambda_E(E\cap F_j) = { prim}((\rho_E\circ \lambda)(F_j)),
\end{equation}
where $ {prim}((\rho_E\circ \lambda)(F_j))$ denotes the primitive vector of
$(\rho_E\circ \lambda)(F_j).$
One can check that the pair $(E, \lambda_E)$ is an $\mathcal{R}$-characteristic
pair. We also associate $M(E)$ with a $k$-dimensional subtorus $T_E$ of
standard $n$-dimensional real torus $T^n$
generated by the images of $\lambda(F_{i_1}), \dots, \lambda(F_{i_k})$
under the map $M(E) \hookrightarrow \mathbb{Z}^n \xrightarrow{\exp} T^n$.
An $\mathcal{R}$-characteristic pair $(Q, \lambda)$
determines a $2n$-dimensional orbifold
with an action of $n$-dimensional torus $T^n$.
To be more precise,
for each point $x\in E$, let $E(x)$ be the face of $Q$ containing $x$
in its interior. Now we consider the following quotient space
$$X(Q, \lambda):=(T^n \times Q)/\sim,$$
where
\begin{equation}\label{eq_equivalence_rel}
(t,x)\sim (s,y) ~~ \mbox{if and only if} ~~ x=y ~~\mbox{and} ~~ t^{-1}s\in T_{E(x)}.
\end{equation}
The space $X(Q, \lambda)$ is equipped with an action of $T^n$
given by the multiplication on the first factor. The orbit map
is induced by the projection onto the second factor,
\begin{equation}\label{eq_orbit_map}
\pi\colon X(Q, \lambda) \to Q
\end{equation}
defined by $[t, x]_{\sim} \mapsto x.$
A detailed verification that $X(Q,\lambda)$ has an orbifold structure,
including an explicit description of the orbifold charts, is
contained in \cite[Section 2]{PS}. The authors also give an
axiomatic description and show that it agrees with the
construction above up to equivariant homeomorphism.
The space $X(Q, \lambda)$ is known as a
\emph{toric orbifold}\footnote{Davis
and Januszkiewicz \cite{DJ} first called them toric orbifolds.
In recent literature, they are sometimes called \emph{quasitoric orbifolds}.}.
\begin{remark}
If the collection $\{\lambda(F_{i_1}), \dots, \lambda(F_{i_k})\}$ in
\eqref{def_R-char_fun} is unimodular whenever $F_{i_1}\cap \dots\cap F_{i_k}\neq \emptyset$,
then $\lambda$ is called a \emph{characteristic function} and the
resulting space $X(Q, \lambda)$ is a smooth \emph{toric manifold},
see \cite[Section 1]{DJ}.
\end{remark}
\begin{remark}\label{rmk_preimage_of_face}
An $\mathcal{R}$-characteristic pair $(Q, \lambda)$
induces another $\mathcal{R}$-characteristic pair $(E, \lambda_E)$ for each
face $E$ of $Q$, which defines another toric orbifold
$X(E, \lambda_E)$. On the other hand, $\pi^{-1}(E)$ is an invariant
suborbifold of $X(Q, \lambda)$ with respect to the action of $T^n$
on $X(Q, \lambda)$. Indeed, $\pi^{-1}(E)$ is the suborbifold fixed by the subtorus $T_E$ of
$T^n$. The residual torus $T^n/T_E \cong T^{\dim E}$ acts on the
suborbifold $X(E, \lambda_E)$ and $\pi^{-1}(E)$.
It is shown in \cite[Section 2.3]{PS} that
$X(E, \lambda_E)$ is equivariantly homeomorphic to $\pi^{-1}(E)$.
In particular, when $E$ is a $1$-dimensional face of $Q$, then
$\pi^{-1}(E)$ is homeomorphic to $S^2$.
\end{remark}
For simplicity, we summarize notation that we introduced above as follows.
\begin{enumerate}
\item $(Q, \lambda)$ is an $\mathcal{R}$-characteristic pair
with $\dim Q=n$ and
$X(Q, \lambda)$ is the associated toric orbifold of dimension $2n$.
\item $E=F_{i_1} \cap \dots \cap F_{i_k}$ is an $(n-k)$-dimensional
face of $Q$, for $k \geq 1$.
\item $\lambda_E \colon \mathcal{F}(E) \to \mathbb{Z}^{n-k}$ is an
$\mathcal{R}$-characteristic function induced from $(Q, \lambda)$.
\item $X(E, \lambda_E)$ is the toric orbifold associated to the $\mathcal{R}$-characteristic pair
$(E, \lambda_E)$.
\item\label{item_(5)} For a vertex $v$ in $E$, we denote by $\Lambda_{E, v} \colon \mathbb{Z}^{n-k} \to \mathbb{Z}^{n-k}$
the linear map given by the square matrix
$$\Lambda_{E,v}=\left[ \begin{array}{c|c|c}
\lambda_E(E\cap F_{s_1})^t& \dots& \lambda_E(E\cap F_{s_{n-k}})^t
\end{array} \right],$$
where $v=\bigcap_{a=1}^{n-k} (E\cap F_{s_a})$.
In particular, when $E=Q$ and $v=\bigcap_{a=1}^n F_{s_a}$,
$$\Lambda_{v}:=\Lambda_{Q, v}=\left[ \begin{array}{c|c|c} \lambda(F_{s_1})^t& \dots& \lambda( F_{s_{n}})^t \end{array} \right].$$
\item $G_E(v) := \ker (\exp \Lambda_{E,v} \colon T^{n-k} \twoheadrightarrow T^{n-k}).$
\label{page_local_group}
\end{enumerate}
Notice that the notation $G_E(v)$ is used for $\operatorname{coker} \Lambda_{E,v}$ in
\cite[Section 4]{BSS}. However, the following commutative diagram and
snake lemma, see for instance \cite{Ati}, shows that those two finite groups are
isomorphic.
\begin{equation}\label{eq_kerexp=coker}
\begin{tikzcd}[row sep=small]
&&0\dar &\ker (\exp \Lambda_{E, v}) \dar & \\
0 \rar &\mathbb{Z}^{n-k} \arrow{dd}{\Lambda_{E, v}} \rar&
\mathbb{Z}^{n-k} \otimes_\mathbb{Z} \mathbb{R} \rar \arrow{dd}{\cong} &
T^{n-k}\rar \arrow{dd}{\exp \Lambda_{E, v}} &0 \\
&&&&\\
0 \rar &\mathbb{Z}^{n-k} \rar\dar & \mathbb{Z}^{n-k} \otimes_\mathbb{Z} \mathbb{R} \rar \dar & T^{n-k}\rar &0 \\
& \operatorname{coker} \Lambda_{E, v} &0 & &
\end{tikzcd}
\end{equation}
\begin{remark}\label{rmk_order_of_local_group}
One can see from \eqref{eq_kerexp=coker} that the order $|G_E(v)|$
of the finite group $G_E(v)$ is exactly $|\det \Lambda_{E, v}|$.
\end{remark}
\begin{example}\label{ex_induced_char_ftn}
Let $Q$ be a 3-dimensional prism and $F_1, \dots, F_5$ its facets illustrated in
Figure \ref{fig_prism_and_char_ftn}.
\begin{figure}
\begin{tikzpicture}[scale=0.7]
\draw[dashed] (0,0)--(1,-1)--(3,1)--cycle;
\draw (0,2)--(1,1)--(3,3)--cycle;
\draw (0,0)--(1,-1)--(1,1)--(0,2)--cycle;
\draw (1,1)--(1,-1)--(3,1)--(3,3)--cycle;
\draw[dashed] (0,0)--(3,1)--(3,3)--(0,2)--cycle;
\node at (1/2, 1/2) {$F_1$};
\node at (2.3, 1.2) {$F_2$};
\node at (4/3, 2) {$F_4$};
\node[right] at (4,2) {$F_3$};
\node[right] at (4,0) {$F_5$};
\draw[dotted, thick, ->] (4,2.3) to [out=120, in=90] (2.5,2);
\draw[dotted, thick, ->] (4,0) to [out=240, in=330] (4/3,0);
\draw plot [mark=*, mark size=2] coordinates{(1,1)};
\draw plot [mark=*, mark size=2] coordinates{(0,2)};
\node[left] at (0,2) {$v_1$};
\node[right] at (1,1) {$v_2$};
\end{tikzpicture}
\caption{}
\label{fig_prism_and_char_ftn}
\end{figure}
We assign $\mathcal{R}$-characteristic vectors by
$\lambda(F_1)=(1,0,0),~\lambda(F_2)=(0,1,0),~\lambda(F_3)=(0,0,1),~
\lambda(F_4)=(1,2,4)$ and $\lambda(F_5)=(-1,-1,-1)$.
Observe that
\begin{align*}
G_Q(v_1) &=\{(t_1, t_2, t_3) \in T^3 \mid t_1t_3=t_3^2=t_2t_3^4=1\}\\
&= \langle (1,1,1), (-1, 1, -1) \rangle \cong \mathbb{Z}/2\mathbb{Z},
\end{align*}
Similarly we have
\begin{align*}
G_Q(v_2) &=\{ (t_1, t_2, t_3) \in T^3 \mid t_1t_3=t_2t_3^2=t_3^4=1\} \\
&= \langle (1,1,1), (-i, -1, i), (-1, 1, -1), (i, -1, -i)\rangle \cong \mathbb{Z}/4\mathbb{Z},\\
G_Q(v) & = \langle (1,1,1) \rangle \cong 1,~ \text{ for } v\in V(Q) \setminus \{v_1, v_2\}.
\end{align*}
Choosing the face $F_4$, we consider $\lambda(F_4)=(1,2,4),~(0,1,0)$ and $(0,0,1)$ as a basis
of $\mathbb{Z}^3$. Then, the induced $\mathcal{R}$-characteristic function
$$\lambda_{F_4} \colon \{F_4 \cap F_1, F_4\cap F_2, F_4\cap F_3\} \to \mathbb{Z}^2$$
on $F_4$ is given by
\begin{align*}
\lambda_{F_4} (F_4 \cap F_1) = (-1,-2), ~~
\lambda_{F_4} (F_4 \cap F_2) = (1,0) ~~ \mbox{and} ~~
\lambda_{F_4} (F_4 \cap F_3) = (0,1).
\end{align*}
To be more precise, since $\lambda(F_1)=(1,0,0)=(1,2,4)-2(0,1,0)-4(0,0,1),$
we have $(\rho_{F_4}\circ \lambda)(F_1)=(-2,-4)$ whose primitive vector gives
$\lambda_{F_4}(F_4 \cap F_1)$. A similar computation gives other two induced
$\mathcal{R}$-characteristic vectors. Moreover, we have
\begin{align*}
G_{F_4}(v_1)&=\{ (t_1, t_2)\in T^2 \mid t_1^{-1}=t_1^{-1}t_2=1\}
=\left< (1,1)\right>\cong 1,\\
G_{F_4}(v_2)&=\{ (t_1, t_2)\in T^2 \mid t_1^{-1}t_2=t_2^{-2}=1\}
=\left< (1,1), (-1, -1)\right>\cong \mathbb{Z}/2\mathbb{Z}.
\end{align*}
Finally, $G_{F_4}(v)$=1, where $v=F_2\cap F_3\cap F_4$ is considered as a vertex of $F_4$.
\end{example}
We finish this section by the following proposition which we shall use in
Section \ref{sec_homology_of_X_{(J)}}. One can see this property in
Example \ref{ex_induced_char_ftn}.
\begin{proposition}\cite[Proposition 4.3]{BSS}\label{prop_sub_local_group}
Let $E$ and $E'$ be two faces of $Q$ containing a vertex $v$ such that
$E$ is a face of $E'$. Then, $|G_E(v)|$ divides $|G_{E'}(v)|$.
\end{proposition}
\section{Building sequences and homology of toric orbifolds}
\label{sec_bdseq_and_q-cell}
The goal of this section is to discuss integral homology of toric orbifolds.
After Poddar--Sarkar \cite{PS} developed the notion of $\mathbf{q}$-CW complex,
the authors of \cite{BNSS} introduced a \emph{building sequence}
which enables us to detect $p$-torsion freeness of an orbifold
having $\mathbf{q}$-CW complex structure for each prime number $p$.
In case of toric orbifolds, one can derive a building sequence via a \emph{retraction
sequence} of a simple polytope introduced in \cite{BSS}. In this subsection, we
review the definition of a \emph{building sequence} and a
\emph{retraction sequence} and study their relation.
\subsection{Building sequence}
Let $\bar{D}^n$ be a closed $n$-dimensional disc and $G$ a finite group
acting linearly on $\bar{D}^n$. We call the quotient $\bar{D}^n/G$ a
\emph{$\mathbf{q}$-disc}. Now, a $\mathbf{q}$-CW complex is defined inductively
in the similar manner as a usual CW complex, but we use $\mathbf{q}$-discs instead
of $\bar{D}^n$.
To be more precise, we start with a discrete set $X_0$ of
$0$-dimensional $\mathbf{q}$-cells. Next, assuming $X_{i-1}$ is defined,
we define
$$X_i := X_{i-1} \cup_{\{f_\alpha\}} \{\bar{D}^{i}/G_\alpha\},$$
where $f_\alpha \colon \partial (\bar{D}^{i}/G_\alpha) \to X_{i-1}$ is the
\emph{attaching map} for a $\mathbf{q}$-cell $\bar{D}^{i}/G_\alpha$ for
finitely many $\alpha$.
A \emph{building sequence} for a $\mathbf{q}$-CW complex $X$ is a sequence
$\{Y_j\}_{j=1}^\ell$ of $\mathbf{q}$-CW subcomplexes of $X$ such that
\begin{equation}\label{eq_attaching_cell_in_building_seq}
Y_j\setminus Y_{j-1} \cong D^{k_j} / G_j
\end{equation}
for some $k_j$-dimensional open $\mathbf{q}$-disc $D^{k_j}/G_j$. We denote by
$0_j$ the image of
origin of \eqref{eq_attaching_cell_in_building_seq} and call it the \emph{free special
point} in $Y_j$.
When we need to emphasize the free special point, we denote the
building sequence by $\{(Y_j, 0_j)\}_{j=1}^\ell$, see \cite[Section 2]{BNSS} for
more details.
\begin{example}
Let $Y_1$ be a point and $Y_2= Y_1 \cup _{f_1} \bar{D}^1$ a circle obtained
by the canonical attaching map $f_1 \colon S^0 \to \{pt\}$. Consider two
$\mathbf{q}$-cells $\bar{D}^2/\mathbb{Z}_p$ and $\bar{D}^2/\mathbb{Z}_q$, where we regard
$\mathbb{Z}_p$ as a finite group generated by $p$-th root of unity and it acts on
$\bar{D}^2\subset \mathbb{C}$ by the multiplication. Similarly we consider
$\mathbb{Z}_q$-action on $\bar{D}^2$. Now we define
$$Y_3: =Y_2 \cup_{f_2} \bar{D}^2/\mathbb{Z}_p \text{ and }
Y_4: =Y_3 \cup_{f_3} \bar{D}^2/\mathbb{Z}_p,$$
where
$f_2 \colon S^1/\mathbb{Z}_p \to Y_1\cong S^1$, respectively
$f_3 \colon S^1/\mathbb{Z}_q \to Y_1 \subset Y_2 $,
defined by $f_2([e^{2\pi i x}]) = e^{2\pi i px}$, respectively $f_3([e^{2\pi i y}])=e^{2\pi i qy}$, for
$0 \leq x,y \leq 1$. Then, the resulting space $Y_4$ is a $2$-dimensional orbifold sphere
of football type $\mathbb{C}P^1_{p,q}$, see Figure \ref{fig_football}.
\begin{figure}[h!]
\begin{tikzpicture}[scale=0.8]
\draw[fill] (-1.5,0) circle (0.05);
\draw[fill] (0,0) circle (0.05);
\node at (-1.5,-1.5) {\footnotesize$Y_1$};
\node at (1,-1.5) {\footnotesize$Y_2=Y_1 \cup_{f_1} \bar{D}^1$};
\begin{scope}[yscale=0.7]
\draw (0,0) [out=90, in=90] to (2,0) [out=270, in=270] to (0,0);
\end{scope}
\begin{scope}[xshift=110, yscale=0.7]
\draw[fill] (0,0) circle (0.05);
\draw[dashed] (0,0) [out=90, in=90] to (2,0) [out=270, in=270] to (0,0);
\draw[fill=yellow, opacity=0.5]
(1,2) [out=315, in=90] to (2,0) [out=270, in=270] to (0,0) [out=90, in=225] to (1,2);
\draw(1,2) [out=315, in=90] to (2,0) [out=270, in=270] to (0,0) [out=90, in=225] to (1,2);
\node at (1,-2.5) {\footnotesize$Y_3=Y_2 \cup_{f_2} \bar{D}^2/\mathbb{Z}_p$};
\end{scope}
\begin{scope}[xshift=220, yscale=0.7]
\draw[fill] (0,0) circle (0.05);
\draw[dashed] (0,0) [out=90, in=90] to (2,0) [out=270, in=270] to (0,0);
\draw[fill=blue, opacity=0.4, dashed]
(1,-2) [out=60, in=270] to (2,0) [out=90, in=90] to (0,0) [out=270, in=120] to (1,-2);
\draw
(0,0) [out=270, in=120] to (1,-2) [out=60, in=270] to (2,0) ;
\draw[dashed] (0,0) [out=90, in=90] to (2,0);
\draw[fill=yellow, opacity=0.5]
(1,2) [out=315, in=90] to (2,0) [out=270, in=270] to (0,0) [out=90, in=225] to (1,2);
\draw (1,2) [out=315, in=90] to (2,0) [out=270, in=270] to (0,0) [out=90, in=225] to (1,2);
\node at (1,-2.5) {\footnotesize$Y_4=Y_3 \cup_{f_3} \bar{D}^2/\mathbb{Z}_q$};
\end{scope}
\end{tikzpicture}
\caption{A building sequence of $\mathbb{C}P^1_{p,q}$.}
\label{fig_football}
\end{figure}
\end{example}
\begin{remark}\label{rmk_trivial_q-cell}
A $\mathbf{q}$-disc $\bar{D}^k/G$ is homeomorphic to
the usual disc $\bar{D}^k$ if $k=0,1$ or $2$. Indeed,
the football-type orbifold sphere $\mathbb{C}P^1_{p,q}$ is
homeomorphic to $S^2$ for arbitrary positive integers $p$ and $q$.
\end{remark}
\subsection{Retraction sequence}
Given an $n$-dimensional simple polytope $Q$, as a polytopal
complex (see \cite[Definition 5.1]{Zie}), we define a sequence of
polytopal subcomplexes of $Q$ which will determine a building
sequence of a toric orbifold whose orbit space is $Q$. We
refer to \cite[Section 2]{BSS} for more details
about retraction sequences of a simple polytope.
Let $B$
be a connected polytopal subcomplex of $Q$. A vertex $v$ of $B$
is called a \emph{free vertex} if $v$ has a neighborhood homeomorphic to
$\mathbb{R}^N_{\geq 0}$ for some $N\in \{1, \dots, \dim B\}$ as manifolds with corners. In this case, there
exists a unique maximal face $E$ of $B$ containing $v$ as a vertex. We write $FV(B)$
as the set of all free vertices of $B$.
For instance, every vertex $v$ of $Q$ is a free vertex and $Q$ itself is the
unique (non-proper) face of $Q$ containing $v$. In particular, $FV(Q)=V(Q)$.
However, for a polytopal subcomplex $B$ of $Q$, $FV(B)$ is a subset of $V(B)$
in general.
A retraction sequence $\{(B_j, E_j, b_j)\}_{j=1}^\ell$ for $Q$ is a sequence of
triples consists of a polytopal subcomplex $B_j$ of $Q$, a free vertex $b_j$
of $B_j$ and the unique face $E_j$ of $B_j$ containing a free vertex $b_j$,
which is defined inductively such that
$$B_{j+1}=\bigcup \{E\mid E \text{ is a face of }B_j \text{ with } b_j\notin E\}.$$
The next term $(B_{j+1}, E_{j+1}, b_{j+1})$ is given by the choice of a free
vertex $b_{j+1}$ of $B_{j+1}$ and the unique face $E_{j+1}$
determined by the edges of $B_{j+1}$ containing $b_{j+1}$.
Finally, the sequence ends up with $(B_\ell, E_\ell, b_\ell)=(b_\ell, b_\ell, b_\ell)$
for some vertex $b_\ell$ of $Q$.
We note that every simple polytope has at least one retraction sequence.
It can be shown by realizing a simple polytope $Q$ as a convex subset in
$\mathbb{R}^n$ and taking a linear function $f\colon \mathbb{R}^n \to \mathbb{R}$
such that $f(v_i)\neq f(v_j)$ whenever two vertices $v_i$ and $v_j$ of $Q$
are distinct. We refer to \cite[Proposition 2.3]{BSS} for the completeness of arguments.
\begin{example}\label{ex_n-gon}
Let $Q$ be an $\ell$-gon and $v_1, \dots, v_\ell$ its vertices with counterclockwise
order. Let $F_i$ be the $1$-dimensional face connecting $v_i$ and $v_{i+1}$ for
$i=1, \dots, \ell-1$ and $F_{\ell}$ the $1$-dimensional face with vertices $v_{\ell}$ and
$v_1$. Then, one canonical choice of a retraction sequence
$\{(B_j, E_j, b_j)\}_{j=1}^{\ell}$
is given by
$$(B_j, E_j, b_j)=\begin{cases} (Q, Q, v_1) & \text{if } j=1;\\
( F_j\cup \dots \cup F_{\ell-1}, F_j, v_j)& \text{if } j=2, \dots, \ell-1;\\
(v_\ell, v_\ell, v_\ell) & \text{if } j=\ell,
\end{cases}$$
as in Figure \ref{fig_ret_of_polygon} for $5$-gon.
\end{example}
\begin{figure}
\begin{tikzpicture}[scale=0.8]
\draw[fill=blue!30] (18:1cm) \foreach \x in {90, 90+72, 90+144, 90+216, 18}
{ -- (\x: 1cm)} ;
\foreach \x/\l/\p in
{18/{\blue{\small$v_1$}}/right,
90/{\small$v_2$}/above,
90+72/{\small$v_3$}/left,
90+144/{\small$v_4$}/left,
90+216/{\small$v_5$}/right}
\node (\x:1cm) [circle, draw, fill, inner sep = 1pt, label = {\p:\l}] at (\x:1cm) {};
\begin{scope}[xshift=100]
\draw[very thick, blue] (90:1cm)--(90+72:1cm);
\draw (90:1cm) \foreach \x in {90+72, 90+144, 90+216} { -- (\x: 1cm)} ;
\foreach \x/\l/\p in
{90/{\blue{\small$v_2$}}/above,
90+72/{\small$v_3$}/left,
90+144/{\small$v_4$}/left,
90+216/{\small$v_5$}/right}
\node (\x:1cm) [circle, draw, fill, inner sep = 1pt, label = {\p:\l}] at (\x:1cm) {};
\end{scope}
\begin{scope}[xshift=200]
\draw[very thick, blue] (90+144:1cm)--(90+72:1cm);
\draw (90+72:1cm) \foreach \x in {90+144, 90+216} { -- (\x: 1cm)} ;
\foreach \x/\l/\p in
{
90+72/{\blue{\small$v_3$}}/left,
90+144/{\small$v_4$}/left,
90+216/{\small$v_5$}/right}
\node (\x:1cm) [circle, draw, fill, inner sep = 1pt, label = {\p:\l}] at (\x:1cm) {};
\end{scope}
\begin{scope}[xshift=300]
\draw[very thick, blue] (90+144:1cm)--(90+216:1cm);
\draw (90+144:1cm) \foreach \x in {90+144, 90+216} { -- (\x: 1cm)} ;
\foreach \x/\l/\p in
{
90+144/{\blue {\small$v_4$}}/left,
90+216/{\small$v_5$}/right}
\node (\x:1cm) [circle, draw, fill, inner sep = 1pt, label = {\p:\l}] at (\x:1cm) {};
\end{scope}
\begin{scope}[xshift=350]
\foreach \x/\l/\p in
{
90+216/{\small$v_5$}/right}
\node (\x:1cm) [circle, draw, fill, inner sep = 1pt, label = {\p:\l}] at (\x:1cm) {};
\end{scope}
\end{tikzpicture}
\caption{A retraction sequence for $5$-gon.}
\label{fig_ret_of_polygon}
\end{figure}
The following proposition shows a relation between a retraction sequence of a
simple polytope $Q$ and a building sequence of the associated toric orbifold
$X(Q, \lambda)$.
\begin{proposition}\cite[Proposition 4.4]{BNSS}\label{prop_ret_build}
Let $(Q, \lambda)$ be an $\mathcal{R}$-characteristic pair and $\pi \colon
X(Q, \lambda) \to Q$ be the orbit map.
Then, a retraction sequence $\{(B_j, E_j, b_j)\}_{j=1}^\ell$ for $Q$ induces a building
sequence $\{(Y_{j}, 0_{j})\}_{j=1}^\ell$ as follows:
\begin{itemize}
\item $Y_1=0_i=\pi^{-1}(b_\ell)$,
\item $Y_j=\bigcup_{s=\ell-j+1}^\ell \pi^{-1}(U_s)$, where $U_s$ is the open subset of
$E_s$ obtained by deleting faces of $E_s$ which does not contain $b_s$, for $j=2, \dots, \ell$.
\end{itemize}
In particular, $Y_\ell= X(Q, \lambda)$ and $0_j=\pi^{-1}(b_j)$ for all $j=1, \dots, \ell$.
\end{proposition}
\begin{remark}
Though a building sequence induced from a retraction
sequence may require the image of an attaching
map for a q-cell $\overline{D}^k/G$
to be in a higher dimensional skeleton, a $G$-equivariant
triangulation \cite[Theorem 3.6]{Il} of $\overline{D}^k/G$ together with
an application of the cellular approximation theorem
allows for a deformation into the appropriate dimension.
\end{remark}
Next two theorems can be obtained by combining
\cite[Theorem 1.1, 1.2]{BNSS} and Proposition \ref{prop_ret_build},
which describe a sufficient condition for ($p$-)torsion freeness and vanishing odd degree
(co)homology of a toric orbifold.
\begin{theorem}\label{thm_no_p-torsion}
Let $X:=X(Q, \lambda)$ be a toric orbifold and $p$ a prime number.
If there is a retraction $\{(B_j, E_j, b_j)\}_{j=1}^{\ell}$ such that
$\gcd\{p, |G_{E_j}(b_j)|\}=1$ for all $j$, then $H_{\ast}(X; \mathbb{Z})$
has no $p$-torsion and $H_{odd}(X; \mathbb{Z}_p)$ is trivial.
\end{theorem}
\begin{theorem}\label{thm_no-torsion}
Let $X(Q, \lambda)$ be a toric orbifold. If for each prime $p$,
there is a retraction $\{(B_j, E_j, b_j)\}_{j=1}^{\ell}$ such that
$\gcd\{p, |G_{E_j}(b_j)|\}=1$ for all $j$, then $H_{\ast}(X; \mathbb{Z})$
has no torsion and $H_{odd}(X; \mathbb{Z})$ is trivial.
\end{theorem}
Next two examples are applications of Theorem \ref{thm_no-torsion}
to toric orbifolds over a polygon and a simplex, respectively. Though
these results are well known from the literature, for example
\cite{Fis, Jor, KMZ} and \cite{Ka},
the same conclusions follow from the results above.
\begin{example}
Let $Q$ be an $\ell$-gon and $F_1, \dots, F_\ell$ facets of $Q$ as in Example
\ref{ex_n-gon}. Let $\lambda(F_i):=(a_i, b_i)\in \mathbb{Z}^2$ be characteristic
vectors for $i=1, \dots, \ell$. Hence, we have
$$|G_Q(v_1)|=\left| \det \begin{bmatrix} a_\ell & a_{1} \\ b_\ell & b_{1} \end{bmatrix}\right|
\quad \text{and} \quad
|G_Q(v_i)|=\left| \det \begin{bmatrix} a_{i-1} & a_{i} \\ b_{i-1} & b_{i} \end{bmatrix}\right|
\text{ for } i=2, \dots, \ell.$$
We refer to Remark \ref{rmk_order_of_local_group}.
Moreover, since the facets are $1$-dimensional, $G_{F_i}(v_i)$ is trivial for each
facet $F_i$. Indeed, one can easily check that the induced $\mathcal{R}$-characteristic
function
$$\lambda_{F_i} \colon \{v_i, v_{i+1}\} \to \mathbb{Z}$$
can be always defined by
$\lambda_{F_i}(v_i)=\lambda_{F_i}(v_{i+1})=\pm 1$, see
Remark \ref{rmk_preimage_of_face} and Remark \ref{rmk_trivial_q-cell}.
Now, we assume that
$$\gcd\{|G_Q(v_1)|, \dots, |G_Q(v_{\ell})| \} =q.$$
Then, one can always choose a retraction sequence $\{(B_j, E_j, b_j)\}_{j=1}^{\ell}$
of $Q$ such that $\gcd\{p, |G_{E_j}(b_j)|\}=1$ unless $p$ is a
factor of $q$. Hence, we conclude
that if $H_{\ast}(X(Q, \lambda); \mathbb{Z})$ has a non-trivial torsion part, then it must have
a $p$-torsion for some $p$ dividing $q$.
In particular, if $q=1$, then $H_{\ast}(X(Q, \lambda); \mathbb{Z})$ is torsion free.
\end{example}
\begin{example}\label{ex_wCP}
Consider an $n$-simplex $\Delta^n$ and let $\chi:=(\chi_1, \dots, \chi_{n+1})\in \mathbb{N}^{n+1}$.
An $\mathcal{R}$-characteristic function
$$\lambda \colon \mathcal{F}(\Delta^n):=\{F_1, \dots, F_{n+1}\} \to \mathbb{Z}^n$$
satisfying
$\sum_{i=1}^{n+1} \chi_i \lambda(F_i)=\mathbf{0}$ and
${\rm span}_\mathbb{Z} \{\lambda(F_1), \dots, \lambda(F_{n+1})\}=\mathbb{Z}^n.$
It is well-known that this $\mathcal{R}$-characteristic pair $(\Delta^n, \lambda)$
defines a weighted projective space $\mathbb{C}P^n_{\chi}$,
see \cite[Example 3.1.17]{CLS} or \cite[Section 2.2]{Ful}.
Without loss of generality, we may assume that
$\gcd\{\chi_1, \dots, \chi_{n+1}\}=1$. Indeed, for some
$k\in \mathbb{N}$, two vectors
$(\chi_1, \dots, \chi_{n+1})$ and $(k\chi_1, \dots, k\chi_{n+1})$
yield homeomorphic weighted projective spaces, see for example
\cite[Theorem 1.1]{BFNR}.
Note that $\chi_i$ is the order of singularity at
$[0, \dots, 0,\underset{i\text{-th}}{1},0,\dots 0]\in \mathbb{C}P^n_{\chi}$, which is same as
$|G_{\Delta^n}(v_i)|$ defined from the $\mathcal{R}$-characteristic pair
$(\Delta^n, \lambda)$, where
$v_i=F_1\cap \cdots \cap F_{i_1} \cap F_{i+1} \cap \cdots F_{n+1}$.
Now, following the proof of \cite[Proposition 4.5]{BNSS},
one can always
find retraction sequences $\{(B_j, E_j, b_j)\}_{j=1}^{n+1}$ which
satisfies the assumption of Theorem \ref{thm_no-torsion}. Hence,
we conclude that the (co)homology of any weighted projective space is
torsion free and concentrated in even degrees.
\end{example}
\section{$J$-construction of toric orbifolds}\label{sec_J-const}
For \emph{toric manifolds},
the authors of \cite{BBCG15} use a construction introduced in \cite{PB}
to construct new toric manifolds from a given one.
Indeed, this was done
by producing a new \emph{characteristic pair} from the original
one in a canonical way. The process for making a new polytope from the
given one is called \emph{simplicial wedge construction}.
Moreover, by a successive procedure,
a countably infinite family of new toric manifolds can
arise from the original manifold. In \cite{BBCG15}, where
the construction was analyzed in the
context of polyhedral products, the process is described efficiently
by using a vector
$J=(j_1, \dots, j_m) \in \mathbb{N}^m$, where
$m$ is the number of facets in the original polytope.
To be more precise, given a positive integral vector
$J=(j_1, \dots, j_m) \in \mathbb{N}^m$, one can obtain a new
toric manifold $M(J)$ from the original toric manifold
$M$. We refer this procedure as the \textit{$J$-construction}
and apply it to toric orbifolds.
As an example, we shall investigate the class of spaces
that can be produced from a weighted projective spaces.
\subsection{The simplicial wedge construction}
Let $K$ be a simplicial complex with vertex set $V(K)=\{w_1, \dots, w_m\}$.
We call a subset $\sigma\subset V(K)$ \emph{a non-face of} $K$
if $\sigma$ is not a simplex in $K$.
A non-face $\sigma$ is called \emph{minimal} if every proper subset
of $\sigma$ is a simplex in $K$.
Then, the combinatorial type of $K$ is determined by the set
of minimal non-faces of $K$.
For arbitrary positive integral vector $J=(j_1, \dots, j_m)\in \mathbb{N}^m$,
a new simplicial complex $K_{(J)}$ is defined on
$V(K_{(J)}) = \{w_{11}, \ldots, w_{1j_1}, \ldots, w_{1m}, \ldots, w_{mj_m} \}$
with minimal non-faces of the form $V_{i_1} \cup \cdots \cup V_{i_k}$,
where $V_i = \{w_{i1}, \ldots, w_{ij_i}\}$, whenever $\{w_{i_1}, \ldots, w_{i_k}\}$
is a minimal non-face of $K$.
In the special case $J=(1,\dots, 1,\underset{\substack{\uparrow \\ i-\text{th}}}{2},1,\dots, 1)$,
we denote $K_{(w_i)} := K_{(J)}$, and refer to it as \emph{the simplicial wedge construction} of
$K$ on $w_i \in V$, see \cite{PB}. The following representation is a useful
combinatorial description of the simplicial wedge construction.
\begin{equation}\label{eq_wedge_of_K}
K_{(w_i)}=
\big[\{w_{i1}, w_{i2}\} \ast {\rm link}_{K}\{w_i\}\big] \cup
\big[\{\{w_{i1}\}, \{w_{i2}\}\}\ast (K \setminus \{w_i\})\big],
\end{equation}
where $\ast$ denotes the join of two simplicial complexes
and we identify $w_{r1}$ with $w_{r}$ for $r\neq i$.
In this paper, we focus on the case when $K$ is dual to the boundary $\partial Q$ of
a simple polytope $Q$, which we denote by $K_Q$ and refer to as the
\emph{nerve complex} of $Q$, see for instance \cite[Section 2.2]{BP-book}.
Notice that the vertex set
$V(K_Q)=\{w_1, \dots, w_m\}$ bijectively corresponds to the set of facets
$\mathcal{F}(Q)=\{F_1, \dots, F_m\}$.
\begin{example} \label{ex_simplicial_wedge}
\begin{enumerate}
\item Let $K=K_{\Delta^n}$ be the nerve complex of an $n$-simplex $\Delta^n$
and $V(K)=\{w_1,\dots,w_{n+1}\}$ its vertex set. Then, there exists
only one minimal non-face $\sigma=\{w_1,\dots ,w_{n+1}\}$.
For an arbitrary $J=(j_1, \dots, j_{n+1}) \in \mathbb{N}^{n+1}$, $K_{(J)}$
is the simplicial complex on the vertex set
$$\{w_{11}, \dots, w_{1j_1}, \dots, w_{n+1,1}, \dots , w_{n+1, j_{n+1}}\}$$
with the unique minimal non-face
$\{w_{11}, \dots, w_{1,j_1}, \dots, w_{n+1,1}, \dots , w_{n+1,j_{n+1}}\}$.
Hence, we get $K_{(J)}=K_{\Delta^{d(J)}}$, where
$d(J):= {\sum_{i=1}^{n+1} j_i}$.
Figure \ref{Fig_simp_wedge_of_simplex} describes the case
when $n=2$, $J=(1,1,2)$ and the decomposition by \eqref{eq_wedge_of_K}.
\begin{figure}[h]
\begin{tikzpicture}[scale=0.3]
\draw (0,9.5)--(4,9.5)--(2,13)--cycle;
\node[left] at (0,9.5) {\small{$w_1$}};
\node[right] at (4,9.5) {\small{$w_2$}};
\node[above] at (2,13) {\small{$w_3$}};
\draw[line join=round, decorate, decoration=
{zigzag, segment length=4,
amplitude=.9,post=lineto, post length=2pt}, ->] (7,11)--(9,11);
\draw (12,9.5)--(16,9.5)--(13.5,13.5)--cycle;
\draw [dotted] (12,9.5)--(16,9.5)--(16.5,13)--cycle;
\draw (13.5,13.5)--(16.5,13)--(16,9.5);
\node[below] at (12,9.5) {\small$w_{11}$};
\node[below] at (16,9.5) {\small$w_{21}$};
\node[above] at (13.5,13.5) {\small$w_{31}$};
\node[above] at (16.5, 13) {\small$w_{32}$};
\begin{scope}[xshift=250]
\node at (10, 11) {$=$};
\draw (13.5,13.5)--(16,9.5)--(16.5, 13)--cycle;
\draw [dotted] (288/107+12, 224/107+9.5)--(16.5, 13);
\draw (288/107+12, 224/107+9.5)--(12,9.5)--(13.5,13.5);
\node[below] at (12,9.5) {\small$w_{11}$};
\node[below] at (16,9.5) {\small$w_{21}$};
\node[above] at (13.5,13.5) {\small$w_{31}$};
\node[above] at (16.5, 13) {\small$w_{32}$};
\end{scope}
\begin{scope}[xshift=500]
\node at (10, 11) {$\cup$};
\draw (12,9.5)--(16,9.5)--(13.5,13.5)--cycle;
\draw [dotted] (12,9.5)--(288/107+12, 224/107+9.5);
\draw (288/107+12, 224/107+9.5)--(16.5,13)--(16, 9.5);
\node[below] at (12,9.5) {\small$w_{11}$};
\node[below] at (16,9.5) {\small$w_{21}$};
\node[above] at (13.5,13.5) {\small$w_{31}$};
\node[above] at (16.5, 13) {\small$w_{32}$};
\end{scope}
\end{tikzpicture}
\caption{$(K_{\Delta^2})_{(w_3)}=K_{\Delta^3}$.}
\label{Fig_simp_wedge_of_simplex}
\end{figure}
\item Consider next the nerve complex
$$K:=K_{\Delta^{n_1} \times \Delta^{n_2}}=
K_{\Delta^{n_1}} \ast K_{\Delta^{n_2}}.$$
of product of two simplices $\Delta^{n_1}$ and $\Delta^{n_2}$.
Suppose that $\{v_1, \dots, v_{n_1+1}\}$ and
$\{w_1, \dots, w_{n_2+1}\}$ are vertex sets of $K_{\Delta^m}$
and $K_{\Delta^n}$, respectively.
Then, $K$ is a simplicial complex on the vertex set
$\{v_1, \dots, v_{n_1+1},w_1, \dots, w_{n_2+1}\}$
with two minimal non-faces
$\{v_1, \dots, v_{n_1+1}\}$ and $\{w_1, \dots, w_{n_2+1}\}$.
Taking $J=(2,1, \dots, 1)\in \mathbb{N}^{n_1+n_2+2}$, we obtain
$$K_{(J)}=K_{(v_1)}=K_{\Delta^{n_1+1}}\ast K_{\Delta^{n_2}}.$$
See Figure \ref{fig_J-fication_of_simp.cpx} for the case when $n_1=n_2=1$.
\begin{figure}[h]
\begin{tikzpicture}[scale=0.3]
\draw (0,2)--(2,4)--(4,2)--(2,0)--cycle;
\node[left] at (0,2) {\small{$v_1$}};
\node[above] at (2,4) {\small{$w_1$}};
\node[right] at (4,2) {\small{$v_2$}};
\node[below] at (2,0) {\small{$w_2$}};
\draw[line join=round, decorate, decoration={zigzag, segment length=4,
amplitude=.9,post=lineto, post length=2pt}, ->] (6.5,2)--(8.5,2);
\begin{scope}[xshift=-20]
\draw (14.5,5)--(17,2)--(14.5,0)--(12.5,1.5)--(12,3)--cycle;
\draw (17,2)--(12.5,1.5)--(14.5,5);
\draw[dotted] (17,2)--(12,3)--(14.5,0);
\node[above] at (14.5,5) {\small{$w_1$}};
\node[above] at (17.3,2) {\small{$v_2$}};
\node[below] at (14.5,0) {\small{$w_2$}};
\node[left] at (12.5,1.5) {\small{$v_{11}$}};
\node[left] at (12,3) {\small{$v_{12}$}};
\end{scope}
\begin{scope}[xshift=270]
\node at (9, 2) {$=$};
\draw (12.5,1.5)--(14.5,5)--(12,3)--(12.5,1.5)--(14.5, 0);
\draw[dotted] (12,3)--(18/59+12.5, 63/118+3/2);
\draw (18/59+12.5, 63/118+3/2)--(14.5,0);
\node[above] at (14.5,5) {\small{$w_1$}};
\node[below] at (14.5,0) {\small{$w_2$}};
\node[left] at (12.5,1.5) {\small{$v_{11}$}};
\node[left] at (12,3) {\small{$v_{12}$}};
\end{scope}
\begin{scope}[xshift=500]
\node at (8.5, 2) {$\cup$};
\draw[dotted] (18/59+12.5, 63/118+3/2)--(14.5,0);
\draw[dotted] (95/78+12, -19/78+3)--(17,2);
\draw (12.5, 1.5)--(14.5,5)--(17,2)--cycle;
\draw (12.5,1.5)--(14.5, 0)--(17,2);
\draw (14.5, 5)--(12,3)--(18/59+12.5, 63/118+3/2);
\draw (12,3)--(95/78+12, -19/78+3);
\node[above] at (14.5,5) {\small{$w_1$}};
\node[below] at (14.5,0) {\small{$w_2$}};
\node[left] at (12.5,1.5) {\small{$v_{11}$}};
\node[left] at (12,3) {\small{$v_{12}$}};
\node[above] at (17.3,2) {\small{$v_2$}};
\end{scope}
\end{tikzpicture}
\caption{$(K_{\Delta^1} \ast K_{\Delta^1})_{(w_1)}$.}
\label{fig_J-fication_of_simp.cpx}
\end{figure}
\item In general, if $K_1$ and $K_2$ are simplicial complexes on
$\{v_1, \dots, v_{m_1}\}$ and $\{ w_1, \dots, w_{m_2}\}$, respectively, then
\begin{equation}\label{eq_wedge_of_join}
(K_1 \ast K_2)_{(v_i)}={K_1}_{(v_i)}\ast K_2 \text{ and }
(K_1 \ast K_2)_{(w_{j})}=K_1\ast {K_2}_{(w_j)}
\end{equation}
for $i\in \{1, \dots, m_1\}$ and $j\in \{1, \dots, m_2\}$. Indeed, one can see these
relations by comparing minimal non-faces.
\end{enumerate}
\end{example}
\subsection{The polytopal wedge construction}\label{subsec_polytopal_wedge}
According to \cite[page 582]{PB}, if $K$ is a dual to the boundary of a simple polytope $Q$,
then $K_{(w_i)}$ is again a simplicial
complex which is dual to the boundary of a simple polytope.
Notice that, for an arbitrary $J\in \mathbb{N}^m$, $K_{(J)}$ can be
obtained by the iterated procedure of simplicial wedge
construction. Hence, we can see that $K_{(J)}$ is also dual to
the boundary of a simple polytope which we denote by $Q_{(J)}$.
In particular, when $J=J':=(1,\dots,1,2,1,\dots,1)\in \mathbb{N}^m$, $Q_{(J')}$
is homeomorphic to
\begin{equation}\label{eq_polytopal_wedge}
Q_{(J')}:=(Q\times I)/_{\sim_{F_i}}, \text{ where } (x, t)\sim_{F_i} (y, 0) \text{ if } x=y \in F_i
\end{equation}
as manifolds with corners. Indeed, $(Q\times I)/_{\sim_{F_i}}$ has the following
facets
$$\big\{Q^+, Q^-\big \} \cup
\big\{(F\times I)/_{\sim_{F_i}} ~\mid~ F\in \mathcal{F}(Q)\setminus\{F_i\}\big\}$$
where
$Q^+:=Q\times \{1\}$ and $Q^-:=Q\times \{0\}$
intersect at $(F_i \times I)/_{\sim_{F_i}} \cong F_i$ and each of
$Q^+$ and $Q^-$ intersects
all other facets $\big\{(F\times I)/_{\sim_{F_i}} \mid F\in \mathcal{F}(Q)\setminus\{F_i\}\big\}$.
Notice that this observation is exactly the dual
representation of \eqref{eq_wedge_of_K}
given by associating $Q^-$, $Q^+$, $F\times I /_{F_i}$ for
$F\in \mathcal{F}(Q)\setminus \{ F_i\}$
with $w_{i1}$, $w_{i2}$ and $w\in V(K)\setminus \{w_i\}$, respectively,
where $w$ is the dual of $F$.
We call $Q_{(J')}$ the \emph{polytopal wedge construction} of $Q$
with respect to $F_i\in \mathcal{F}(Q)$. We may also denote
$Q_{(J')}$ by $Q_{(F_i)}$ to emphasize the chosen facet $F_i$.
See Figure \ref{fig_polytopal_wedge}
for the example of the polytopal wedge construction of $5$-gon.
\begin{figure}[h]
\begin{tikzpicture}[scale=0.4]
\begin{scope}
\draw (0,2)--(3,4)--(6,2)--(4.5,1)--(1.5,1)--cycle;
\node at (3,-1) {$Q$};
\draw[very thick] (4.5,1)--(1.5,1);
\node[above] at (3,1) {\footnotesize$F_i$};
\end{scope}
\begin{scope}[xshift=280]
\draw[fill=yellow, opacity=0.5] (4.5,1)--(1.5,1)--(1.5,3.5)--(4.5, 3.5)--cycle;
\draw (6,2)--(4.5,1)--(1.5,1)--(0,2);
\draw[dotted] (0,2)--(3,4)--(6,2);
\draw (0,4.5)--(3,6.5)--(6,4.5)--(4.5,3.5)--(1.5,3.5)--cycle;
\draw[dotted] (0,2)--(3,4)--(6,2);
\draw[dotted] (3,6.5)--(3,4);
\draw (0,4.5)--(0,2);
\draw (1.5,3.5)--(1.5,1);
\draw (4.5,1)--(4.5,3.5);
\draw (6,2)--(6,4.5);
\begin{scope}[>=latex]
\draw[thick,->] (3.8,3.3)--(3.8,1);
\draw[thick, ->] (2.2,3.3)--(2.2,1);
\end{scope}
\draw[very thick] (4.5,1)--(1.5,1);
\draw[very thick] (4.5,3.5)--(1.5,3.5);
\node[above] at (3, 3.5) {\footnotesize$F_i\times \{1\}$};
\node[below] at (3, 1) {\footnotesize$F_i\times \{0\}$};
\node at (3,-1) {$Q\times I$};
\end{scope}
\begin{scope}[xshift=560]
\draw (6,2)--(4.5,1)--(1.5,1)--(0,2);
\draw[dashed] (0,2)--(3,4)--(6,2);
\draw (0,4.5)--(3,6.5)--(6,4.5)--(4.5,1)--(1.5,1)--cycle;
\draw[dashed] (0,2)--(3,4)--(6,2);
\draw[dashed] (3,6.5)--(3,4);
\draw (0,4.5)--(0,2);
\draw (6,2)--(6,4.5);
\node at (3,-1) {$(Q\times I)/_{\sim_{F_i}}$};
\end{scope}
\end{tikzpicture}
\caption{A polytopal wedge construction of $5$-gon.}
\label{fig_polytopal_wedge}
\end{figure}
The following example is dual to Example \ref{ex_simplicial_wedge}.
\begin{example}\label{ex_polytopal_wedge}
\begin{enumerate}
\item $\Delta^n_{(F)}=\Delta^{n+1}$, where $F$ is any facet of $\Delta^n$.
\item $(\Delta^{n_1} \times \Delta^{n_2})_{(E\times \Delta^{n_1})}=
\Delta^{n_1+1} \times \Delta^{n_2}$,
where $E$ is a facet of $\Delta^{n_1}$, see Figure \ref{fig_J-fication_of_simp.cpx} for the case
when $n_1=n_2=1$.
\item In general, given two simple polytopes $P$ and $Q$, any facet of $P\times Q$
is of the form $E\times Q$ for some facet $E$ of $P$, or $P\times F$ for some
facet $F$ of $Q$. Then, the relation
$K_{P\times Q} = K_P \ast K_Q$
together with \eqref{eq_wedge_of_join}
leads us to the following:
\begin{equation*}\label{eq_wedge_of_product_of_polytopes}
(P\times Q)_{(E\times Q)}= P_{(E)}\times Q \quad \text{and} \quad
(P\times Q)_{(P\times F)}= P\times Q_{(F)}.
\end{equation*}
\end{enumerate}
\end{example}
\subsection{A new $\mathcal{R}$-characteristic function}
Let $(Q, \lambda)$ be an $\mathcal{R}$-characteristic pair and $K$
the simplicial complex dual to $\partial Q$ as above.
As in previous sections, $m$ and $n$ denote
the number of facets of $Q$ and the dimension of $Q$, respectively.
Given a vector $J=(j_1, \dots, j_m)\in \mathbb{N}^m$, we define a matrix
$\Lambda_{(J)}$ of size
$(d(J)-m+n)) \times d(J) $ as follows, where
$d(J):= \sum_{i=1}^m j_i$ ;
\begin{align}\label{eq_lambda(J)_matrix}
\Lambda_{(J)}= \tiny{\left[ \begin{array}{ccc|ccc|ccc|ccc|ccccc}
&&&&&&&&&&&&-1&& \\
& I_{j_1-1}& & & & & & & & & & & \vdots && & \\
&&&&&&&&&&&&-1&&& \\ \hline
&&&&&&&&&&&&&-1&&\\
&&&&I_{j_2-1}&&&&&&&&&\vdots& &\\
&&&&&&&&&&&&&-1&&\\ \hline
&&&&&&&&&&&&&&&\\
&&&&&&&\ddots&&&&&&&\vdots& \\
&&&&&&&&&&&&&&&\\ \hline
&&&&&&&&&&&&&&&-1\\
&&&&&&&&&&I_{j_m-1}&&&&&\vdots\\
&&&&&&&&&&&&&&&-1\\ \hline
&&&&&&&&&&&&&&&\\
&&&&&&&&&&&&&\Lambda& & \\
&&&&&&&&&&&&&&&
\end{array}\right]},
\end{align}
where all the columns of the matrix are indexed respectively by
\begin{equation}\label{eq_column_index_of_lambda(J)}
\{ w_{12}, \ldots , w_{1{j_1}}, w_{22}, \ldots , w_{2{j_2}}, \ldots , w_{m2},
\ldots , w_{m{j_m}}, w_{11}, \ldots , w_{m1} \},
\end{equation}
all of the entries in the empty spaces are zero and $\Lambda$ is the original
characteristic matrix associated to $\lambda$. Notice that
the indexing in \eqref{eq_column_index_of_lambda(J)} bijectively corresponds to
the vertex set $V(K_{(J)})$.
Let $F_{ik}$ be the facet of $Q_{(J)}$ dual to the vertex $w_{ik}$,
where $i\in\{1, \dots, m\}$ and $k\in \{1, \dots, j_i\}$.
Then, the matrix $\Lambda_{(J)}$ defines a function
\begin{equation}\label{eq_lambda(J)}
\lambda_{(J)} \colon \mathcal{F}(Q_{(J)}) \to \mathbb{Z}^{d(J)-m+n},
\end{equation}
by assigning to the facet $F_{ik}$ the transpose of the column
vector of $\Lambda_{(J)}$ indexed by $w_{ik}$.
If $\lambda$ satisfies Davis and Januszkiewicz's regularity condition
$(\ast)$, \cite[p.423]{DJ},
then so does $\lambda_{(J)}$ for all $J\in \mathbb{N}^m$, see \cite[Theorem 3.2]{BBCG15}.
The same proof goes through by replacing condition $(\ast)$
$$\det \left[\begin{array}{c|c|c}
\lambda(F_{i_1})^t & \cdots &\lambda(F_{i_{n}})^t \end{array}\right]=\pm 1$$
with the orbifold condition \eqref{def_R-char_fun}
$$\det \left[\begin{array}{c|c|c}
\lambda(F_{i_1})^t & \cdots &\lambda(F_{i_{n}})^t \end{array}\right]\neq 0,$$
to give the next lemma.
\begin{lemma}\label{lem_lambda(J)_satisfies_the_condition}
Let $(Q, \lambda)$ be an $\mathcal{R}$-characteristic pair. Then,
for arbitrary $J\in \mathbb{N}^m$, the function \eqref{eq_lambda(J)} satisfies the
orbifold condition \eqref{def_R-char_fun}.
\end{lemma}
\noindent Hence from $(Q, \lambda)$, one can obtain an infinite family of toric orbifolds
\begin{equation*}
X_{(J)}:=X(Q_{(J)}, \lambda_{(J)})
\end{equation*}
for arbitrary $J\in \mathbb{N}^m$.
\begin{example}\label{ex_J-cons_of_w_proj_sp}
Let $Q$ be an $n$-simplex and consider the $\mathcal{R}$-characteristic
function from Example \ref{ex_wCP}.
As in Example \ref{ex_simplicial_wedge}-(1), for an arbitrary vector
$J=(j_1, \dots, j_{n+1})\in \mathbb{N}^{n+1}$,
$$\Delta^n_{(J)}=\Delta^{d(J)-1}.$$
Then the function $\lambda_{(J)}$, defined by \eqref{eq_lambda(J)}, satisfies
the equation
$$\sum_{k=1}^{j_1} \chi_1 \lambda_{(J)}(F_{1k}) + \dots +
\sum_{k=1}^{j_{n+1}}\chi_n\lambda_{(J)}(F_{n+1,k}) =\mathbf{0},$$
and one can show that the new characteristic vectors span the whole
lattice $\mathbb{Z}^{d(J)-1}$, since the original $\mathcal{R}$-characteristic
vectors in Example \ref{ex_wCP} span $\mathbb{Z}^n$.
Hence, we conclude that $(\mathbb{C}P^n_{\chi})_{(J)}$ is the weighted
projective space $\mathbb{C}P^{d(J)-1}_{\chi_{(J)}}$,
where
$$\chi_{(J)}=
( \underbrace{\chi_1,\ldots,\chi_1}_{j_1}, \ldots ,
\underbrace{\chi_k,\ldots,\chi_k}_{j_k}, \ldots,
\underbrace{\chi_{n+1},\ldots,\chi_{n+1}}_{j_{n+1}})\in \mathbb{N}^{d(J)}.$$
\end{example}
In the special case $J' := (1,\dots,1,2,1,\dots,1)$, the matrix of \eqref{eq_lambda(J)_matrix}
takes a particularly simple form,
\begin{equation}\label{eq_char_matrix_J'}
\Lambda_{(J')}=\begin{bmatrix}
1 & 0 & \cdots & -1 & \cdots & 0\\
0 & & & & & \\
\vdots & \lambda(F_1) & \cdots& \lambda(F_i) &
\cdots &\lambda(F_m)\\
0 & & & & &
\end{bmatrix}.
\end{equation}
Hence, the characteristic function $\lambda_{(J')} \colon \mathcal{F}(Q_{(J')}) \to \mathbb{Z}^{n+1}$
is defined by
\begin{align*}
\lambda_{(J')}(Q^+)&=(1, 0, \dots, 0)^t,\\
\lambda_{(J')}(Q^-)&=(-1, \lambda(F_i))^t \text{ and }\\
\lambda_{(J')}(F_{s1})&=(-1, \lambda(F_s))^t,
\text{ for } s\in \{1, \dots, i-1, i+1, \dots, m\}.
\end{align*}
\begin{remark}
The two induced $\mathcal{R}$-characteristic functions ${\lambda_{(J')}}_{Q^+}$ and
${\lambda_{(J')}}_{Q^-}$ coincide with $\lambda$.
This implies that $X_{(2, 1,\dots, 1)}$ has two copies of original orbifold
$X$ as suborbifolds defined in Section \ref{sec_toric_orb_orb_lens}.
Ewald \cite{Ew} called $X_{(2,1,\dots,1)}$ the \emph{canonical extension} of $X$,
when $X$ is a toric variety. We refer to \cite{BBCG18,BBCG15}, \cite{CP, CP2nd}
and \cite{Ew} for more topological and geometrical observations about the
wedge operation on toric manifolds.
\end{remark}
The following proposition confirms that $X_{(J)}$ for arbitrary
$J=(j_1, \dots, j_m) \in \mathbb{N}^m$ can be constructed from
iterated wedge operations as mentioned in \cite[Remark 3.1]{BBCG15}.
Here, we give an explicit proof.
\begin{proposition}\label{prop_(2,1,...1)(2,1,...,1)=(3,1,...,1)}
Let $X$ be the toric orbifold associated to an $\mathcal{R}$-characteristic pair $(Q, \lambda)$.
Then, two toric orbifolds $X_{(3,1,\dots,1)}$ and $Y_{(2,1,\dots,1)}$ where
$Y=X_{(2,1,\dots, 1)}$ are homeomorphic.
\end{proposition}
\begin{proof}
Let $\{w_1, \dots, w_m\}$ be the vertices of $K:=K_Q$.
Then, one can see from the definition of $K_{(J)}$ or \eqref{eq_wedge_of_K} that
both $K_{(3,1,\dots,1)}$ and $(K_{(2,1,\dots,1)})_{(2,1,\dots,1)}$ have the same number of
vertices and have the same minimal non faces.
Therefore, the simple polytopes
$Q_{(3,1,\dots,1)}$ and $(Q_{(2,1\dots,1)})_{(2,1,\dots,1)}$
determined by
$K_{(3,1,\dots,1)}$ and $(K_{(2,1,\dots,1)})_{(2,1,\dots,1)}$, respectively,
are homeomorphic as manifolds with corners.
The two characteristic matrices $\Lambda_{(3,1,\dots, 1)}$ and
$(\Lambda_{(2,1,\dots,1)})_{(2,1,\dots,1)}$ differ by an element
$$\begin{pmatrix}
1& -1 & 0 & \cdots & 0\\
0& 1 & 0 & \cdots & 0\\
\vdots & & \ddots & &\vdots \\
\vdots & & & 1& 0 \\
0 &\cdots&\cdots &0&1
\end{pmatrix} \in SL_{n+2}(\mathbb{Z}),$$
which induces an automorphism
$$\phi \colon T^{n+2} \to T^{n+2}$$
given by $\phi(t_1, \dots, t_{n+2})= (t_1t_2^{-1}, t_2, \dots, t_{n+2})$.
Finally, the map
$$\phi \times id \colon T^{n+2} \times Q_{(3,1,\dots, 1)} \to
T^{n+2} \times (Q_{(2,1\dots,1)})_{(2,1,\dots,1)}$$
induces a homeomorphism from $X_{(3,1,\dots, 1)}$ to
$(X_{(2,1,\dots,1)})_{(2,1,\dots,1)}$.
\end{proof}
\section{Homology of $X_{(J)}$}\label{sec_homology_of_X_{(J)}}
In this section, we shall see that the homology of $X_{(J)}$ depends on
$J$ and the homology of $X$.
First we compare retraction sequences for the two polytopes $Q$ and $Q_{(J)}$.
It suffices to consider the case when $J'=(1, \ldots, 1, 2, 1, \ldots, 1)$,
because $Q_{(J)}$ can be constructed by the iterations of
the polytopal wedge construction, as in Subsection \ref{subsec_polytopal_wedge}.
We assume that the entry $2$ appears in $i$-th coordinate of $J'$, hence
it corresponds to the $i$-th facet $F_i$ of $Q$.
Let $V(Q)=\{v_1, \dots, v_\ell\}$ be the vertices of $Q$ and
$V(F_i):=\{v_{i_1}, \ldots, v_{i_k}\}$ the vertices of $F_i$.
Now, the vertices of $Q_{(J')}$ are identified as
\begin{equation}\label{eq_vertices_of_Q_tilde}
V(Q_{(J')})=\left\{ v^+, v^- \mid v\in V(Q)\setminus V(F_i) \right\} \cup
\big\{v_{i_1}^-, \ldots, v_{i_k}^-\big\},
\end{equation}
where we write $v^+:=v\times \{1\}$ and $v^-:=v\times \{0\}$ for notational convenience.
Now, we introduce the following two lemmas about the finite group
$G_{Q_{(J')}}(u)$, as defined in Section \ref{sec_toric_orb_orb_lens},
associated to each vertex $u$ of $Q_{(J')}$.
\begin{lemma}\label{lem_loc_goup_isom_for_v_0}
For each vertex $v_{i_r} \in V(F_i)$, $r\in \{1, \dots,k\}$,
the finite group $G_{Q_{(J')}}(v_{i_r}^-)$ is
isomorphic to $G_Q(v_{i_r}).$
\end{lemma}
\begin{proof}
Given a vertex $v_{i_r} \in V(F_i)$,
assume that $v_{i_r}=F_i \cap F_{s_1} \cap \cdots \cap F_{s_{n-1}}$
for some $\{s_1, \ldots, s_{n-1}\} \subset \{1, \ldots, m\}$.
Then, we have
\begin{equation}\label{eq_v_in_F_i_intersection_of_facets}
v_{i_r}^- = Q^+ \cap Q^- \cap \bigcap_{a=1}^{n-1} {F_{s_a}}_{(F_i\cap F_{s_a})},
\end{equation}
where ${F_{s_a}}_{(F_i\cap F_{s_a})}$ is the polytopal wedge construction, as in
Subsection \ref{subsec_polytopal_wedge}, by regarding the facet $F_{s_a}$
as a simple polytope, and $F_i\cap F_{s_a}$ as a facet of $F_{s_a}$ for $a=1, \dots, n-1$.
Restricting to the facets which meet at $v_{i_r}^-$, we get from \eqref{eq_char_matrix_J'}
\begin{align*}
{\Lambda_{(J')}}_{v_{i_r}^-}
&=\left[\begin{array}{c|c|c|c|c}
\lambda_{(J')}(Q^{+})^t & \lambda_{(J')}(Q^{-})^t &
\lambda_{(J')}({F_{s_1}}_{(F_i\cap F_{s_1})})^t&
\cdots &
\lambda_{(J')}({F_{s_{n-1}}}_{(F_i\cap F_{s_{n-1}})})^t
\end{array}\right]\\
&=\left[ \begin{array}{ccccc}
1 & -1 & 0 & \cdots &0 \\
0&&&&\\
\vdots &\lambda(F_{i})^t&\lambda(F_{s_1})^t&\cdots&\lambda(F_{s_{n-1}})^t\\
0 &&&& \end{array}\right],
\end{align*}
which induces an endomorphism
$\exp ({\Lambda_{(J')}}_{v_{i_r}^-}) \colon T^{n+1} \to T^{n+1}$.
According to the definition in Page \pageref{page_local_group} of
Section \ref{sec_toric_orb_orb_lens}, we have
\begin{align*}
G_{Q_{(J')}}(v_{i_r}^-) &=
\ker \left( \exp ({\Lambda_{(J')}}_{v_{i_r}^-}) \colon T^{n+1}
\twoheadrightarrow T^{n+1}\right) \\
&=\{ (t_1, \dots, t_{n+1}) \in T^{n+1} \mid t_1=t_2,~ (t_2, \dots, t_{n+1}) \in G_Q(v_{i_r})\}\\
&\cong G_Q(v_{i_r}).
\end{align*}
\end{proof}
Next, we consider vertices away from $F_i$.
\begin{lemma}\label{lem_loc_group_isom_for_v_pm}
Let $v$ be a vertex in $V(Q) \setminus V(F_i)$. Then,
the finite groups $G_{Q_{(J')}}(v^+)$
and $G_{Q_{(J')}}(v^-)$ are isomorphic to $G_Q(v).$
\end{lemma}
\begin{proof}
Suppose $v=F_{s_1}\cap \dots \cap F_{s_n}$ with $i\notin \{s_1, \dots, s_n\}$.
Then, we have
\begin{equation}\label{eq_vertex_in_Q(J)_not_in_F_i}
v^-=Q^-\cap (F_{s_1}\times I)\cap \cdots \cap (F_{s_n}\times I).
\end{equation}
The $\mathcal{R}$-characteristic function $\lambda_{(J')}$ yields the
square matrix
\begin{align*}
{\Lambda_{(J')}}_{v^-} &=
\left[\begin{array}{c|c|c|c}
{\lambda_{(J')}}(Q^-)^t &
{\lambda_{(J')}}(F_{s_1}\times I)^t &
\cdots &
{\lambda_{(J')}}(F_{s_n}\times I)^t
\end{array}\right]\\
&= \left[\begin{array}{c|c|c|c}
-1& 0 & \cdots & 0 \\
\lambda(F_i)^t & \lambda(F_{s_1})^t & \cdots &\lambda(F_{s_n})^t
\end{array}\right]\\
&= \left[ \begin{array}{c|ccc}
-1 & 0 & \cdots & 0\\ \hline
&&& \\
\lambda(F_i)^t & & \Lambda_v &\\
&&&
\end{array}\right].
\end{align*}
Hence the kernel of the endomorphism
$\exp({\Lambda_{(J')}}_{v^-})\colon T^{n+1} \to T^{n+1}$
of tori induced by
${\Lambda_{(J')}}_{v^-}$ is
$$\{ (t_1, \dots, t_{n+1}) \in T^{n+1} \mid t_1=1,~ (t_2, \dots, t_{n+1}) \in G_Q(v)\}$$
which is isomorphic to $G_Q(v)$.
Similarly one can show that $G_{Q_{(J')}}(v^+)$ is also isomorphic to
$G_Q(v)$.
\end{proof}
The observations above allow us now to adapt the
hypothesis of Theorem \ref{thm_no_p-torsion} to the
$\mathcal{R}$-characteristic pair $(Q_{(J)}, \lambda_{(J)})$.
\begin{lemma}\label{lem_ret_simj}
Given an $\mathcal{R}$-characteristic pair $(Q, \lambda)$ and a prime number $p$,
suppose that there exist a retraction sequence $\{(B_r, E_r, b_r)\}_{r=1}^{\ell}$
of $Q$ satisfying the condition $\gcd\{p, |G_{E_r}(b_r)|\}=1$ for $r=1, \ldots, \ell$. Then
for an arbitrary $J=(j_1, \dots, j_m)\in \mathbb{N}^m$, there exists a retraction sequence
$\{(B^{\prime}_s, E^{\prime}_s, b_s^{\prime})\}_{s=1}^{\ell'}$ for $Q_{(J)}$
which satisfies $\gcd\{p, |G_{E^{\prime}_s}(b_s^{\prime})|\}=1$ for $s=1, \ldots, \ell'$,
where $\ell':=|V(Q_{(J)})|$.
\end{lemma}
\begin{proof}
It is enough to consider the case when $J=J'=(1,\dots, 1,2,1,\dots, 1)$
from the discussion in Section \ref{sec_J-const} and the opening
remark of Section \ref{sec_homology_of_X_{(J)}}.
We assume the entry $2$ appears in $i$-th coordinate of $J$ and let
$b_{\beta_1} \dots, b_{\beta_k}$ be the vertices of the facet $F_i$ of $Q$,
where $\beta_1 < \cdots < \beta_k$.
Given a prime number $p$ and a retraction sequence
$\{(B_r, E_r, b_r)\}_{r=1}^{\ell}$ for $Q$ such that
$\gcd\{p, |G_{E_r}(b_r)|\}=1$ for $r=1, \ldots, \ell$,
we now construct a retraction sequence
$\{(B^{\prime}_s, E^{\prime}_s, b_s^{\prime})\}_{s=1}^{2\ell-k}$
for $Q_{(J')}$ satisfying the hypothesis.
To accomplish this, consider the following sequence of vertices of $Q_{(J')}$:
\begin{align}\label{eq_Q(J)_seq_of_free_vertices}
\begin{split}
b_1^+ \to b_1^-\to \cdots \to b_{\beta_1-1}^+\to b_{\beta_1-1}^-\to b_{\beta_1}^-\to
b_{\beta_1+1}^+\to b_{\beta_1+1}^-\to\cdots\\
\cdots \to b_{\beta_k-1}^+\to b_{\beta_k-1}^-\to
b_{\beta_k}^{-}\to b_{\beta_k+1}^+\to b_{\beta_k+1}^-\to \cdots\to b_\ell^+\to b_\ell^-.
\end{split}
\end{align}
The sequence \eqref{eq_Q(J)_seq_of_free_vertices}
begins with $b_1^-$ if $b_1$ is a vertex of $F_i$, i.e., $\beta_1=1$.
Now, we construct a retraction sequence satisfying the hypothesis
using the sequence \eqref{eq_Q(J)_seq_of_free_vertices} above.
\begin{enumerate}
\item[(Case 1)]
We first assume that $b_1$ is a vertex of $F_i$. Then, we take
$b_1'=b_1^-$ as in \eqref{eq_Q(J)_seq_of_free_vertices} and set
$(B_1', E_1', b_1')= (Q_{(J')}, Q_{(J')}, b_1^-)$. Next,
the choice of $b_{1}^{-}$ as a free vertex of $B_1'$ gives
\begin{align*}
B_2'=\bigg(\bigcup_{E\cap F_i =\emptyset} E\times I\bigg) \cup
\bigg( \bigcup_{\substack{E\cap F_i \neq \emptyset \\ b_1\notin E}} E_{(E\cap F_i)} \bigg),
\end{align*}
where $E$ is a face of $Q$, and $E_{(E\cap F_i)}$ is the polytopal wedge of $E$ by
considering $E$ as a simple polytope and $E\cap F_i$
as a facet of $E$. In general, if a face $E$ of a simple
polytope $Q$ intersects a facet $F$ of $Q$, then $E$ is a face
of $F_i$ or $E\cap F_i$ is a facet of $E$.
Observe that the face structure of $B_2'$
is naturally inherited from the face structure of $Q_{(J')}$.
In particular, neither $Q^+$ nor $Q^-$ is a face of $B_2'$.
Next, we consider the following two possibilities:
(i) $b_2\in V(F_i)\setminus \{b_1\}$, i.e., $\beta_2=2$, and (ii) $b_2\notin V(F_i)$.
If $b_2\in V(F_i)\setminus \{b_1\}$, we set $b_2'=b_2^-$ and
$E_2'={E_2}_{(E_2\cap F_i)}$.
If $b_2\notin V(F_i)$, we set $b_2' =b_2^+$ and $E_2'=E_2\times I$.
Then, $b_2'$ has a neighborhood
homeomorphic to $\mathbb{R}^{\dim E_2+1}_{\geq 0}$ in $E_2'$,
because $b_2$ has the appropriate neighborhood in $E_2$.
Hence, we can define the second term $(B_2', E_2', b_2')$.
The first two retraction sequences in Figure \ref{fig_ret_wedge_5-gon}
illustrate this case when $Q$ is a pentagon.
\begin{figure}
\begin{tikzpicture}[scale=0.25]
\begin{scope}[yshift=-460]
\node at (-4, 3) {\small{Case 2:}};
\draw[fill] (6,4) circle (0.2);
\node[above] at (7,4) {\scriptsize$b_1'=b_1^+$};
\draw (1.5,1)--(0,4)--(3,6)--(6,4)--(4.5,1)--cycle;
\draw (1.5,1)--(0,2)--(0,4);
\draw (4.5,1)--(6,2)--(6,4);
\draw[dashed] (0,2)--(3,4)--(6,2);
\draw[dashed] (3,6)--(3,4);
\draw[->] (7.3,3)--(9.3,3);
\begin{scope}[xshift=300]
\draw[fill] (6,2) circle (0.2);
\node[below] at (7,2) {\scriptsize$b_2'=b_1^-$};
\draw (1.5,1)--(0,2)--(3,4)--(6,2)--(4.5,1)--cycle;
\draw (3,4)--(3,6)--(0,4);
\draw (0,2)--(0,4)--(1.5,1);
\draw[->] (7.3,3)--(9.3,3);
\end{scope}
\begin{scope}[xshift=600]
\draw[fill] (3,6) circle (0.2);
\node[right] at (3,6) {\scriptsize$b_3'=b_2^+$};
\draw (4.5,1)--(1.5,1)--(0,2)--(3,4);
\draw (3,4)--(3,6)--(0,4);
\draw (0,2)--(0,4)--(1.5,1);
\draw[->] (7.3,3)--(9.3,3);
\end{scope}
\begin{scope}[xshift=900]
\draw[fill] (3,4) circle (0.2);
\node[above] at (3,4) {\scriptsize$b_4'=b_2^-$};
\draw (4.5,1)--(1.5,1)--(0,2)--(3,4);
\draw (0,2)--(0,4)--(1.5,1);
\draw[->] (7.3,3)--(9.3,3);
\node at (11,3) {$\cdots$};
\end{scope}
\end{scope}
\begin{scope}[yshift=-230]
\node at (-4, 3) {\small{Case 1-(ii):}};
\draw[fill] (4.5,1) circle (0.2);
\node[right] at (4.5,0.5) {\scriptsize$b_1'=b_1^-$};
\draw (1.5,1)--(0,4)--(3,6)--(6,4)--(4.5,1)--cycle;
\draw (1.5,1)--(0,2)--(0,4);
\draw (4.5,1)--(6,2)--(6,4);
\draw[dashed] (0,2)--(3,4)--(6,2);
\draw[dashed] (3,6)--(3,4);
\draw[->] (7.3,3)--(9.3,3);
\begin{scope}[xshift=300]
\draw[fill] (6,4) circle (0.2);
\node[above] at (7,4) {\scriptsize$b_2'=b_2^+$};
\draw (1.5,1)--(0,2)--(3,4)--(6,2);
\draw (3,4)--(3,6)--(0,4);
\draw (0,2)--(0,4)--(1.5,1);
\draw (6,2)--(6,4)--(3,6);
\draw[->] (7.3,3)--(9.3,3);
\end{scope}
\begin{scope}[xshift=600]
\draw[fill] (6,2) circle (0.2);
\node[below] at (6,2) {\scriptsize$b_3'=b_2^-$};
\draw (1.5,1)--(0,2)--(3,4)--(6,2);
\draw (3,4)--(3,6)--(0,4);
\draw (0,2)--(0,4)--(1.5,1);
\draw[->] (7.3,3)--(9.3,3);
\end{scope}
\begin{scope}[xshift=900]
\draw[fill] (3,6) circle (0.2);
\node[right] at (3,6) {\scriptsize$b_4'=b_3^+$};
\draw (1.5,1)--(0,2)--(3,4);
\draw (3,4)--(3,6)--(0,4);
\draw (0,2)--(0,4)--(1.5,1);
\draw[->] (7.3,3)--(9.3,3);
\node at (11,3) {$\cdots$};
\end{scope}
\end{scope}
\draw[fill] (4.5,1) circle (0.2);
\node at (-4, 3) {\small{Case 1-(i):}};
\node[right] at (4.5,0.5) {\scriptsize$b_1'=b_1^-$};
\draw (1.5,1)--(0,4)--(3,6)--(6,4)--(4.5,1)--cycle;
\draw (1.5,1)--(0,2)--(0,4);
\draw (4.5,1)--(6,2)--(6,4);
\draw[dashed] (0,2)--(3,4)--(6,2);
\draw[dashed] (3,6)--(3,4);
\draw[->] (7.3,3)--(9.3,3);
\begin{scope}[xshift=300]
\draw[fill] (1.5,1) circle (0.2);
\node[right] at (1.5,1) {\scriptsize$b_2'=b_2^-$};
\draw (1.5,1)--(0,2)--(3,4)--(6,2);
\draw (3,4)--(3,6)--(0,4);
\draw (0,2)--(0,4)--(1.5,1);
\draw (6,2)--(6,4)--(3,6);
\draw[->] (7.3,3)--(9.3,3);
\end{scope}
\begin{scope}[xshift=600]
\draw[fill] (0,4) circle (0.2);
\node[above] at (-.8,4.2) {\scriptsize$b_3'=b_3^+$};
\draw (0,2)--(3,4)--(6,2);
\draw (3,4)--(3,6)--(0,4);
\draw (0,2)--(0,4);
\draw (6,2)--(6,4)--(3,6);
\draw[->] (7.3,3)--(9.3,3);
\end{scope}
\begin{scope}[xshift=900]
\draw[fill] (0,2) circle (0.2);
\node[below] at (0,2) {\scriptsize$b_4'=b_3^-$};
\draw (0,2)--(3,4)--(6,2);
\draw (6,2)--(6,4)--(3,6)--(3,4);
\draw[->] (7.3,3)--(9.3,3);
\node at (11,3) {$\cdots$};
\end{scope}
\end{tikzpicture}
\caption{Three retraction sequences of the wedge of $5$-gon.}
\label{fig_ret_wedge_5-gon}
\end{figure}
\item[(Case 2)] Here we assume that $b_1\notin V(F_i)$.
Then, we take $b_1'=b_1^+$ by \eqref{eq_Q(J)_seq_of_free_vertices},
which gives
\begin{equation*
B_2'=Q^- \cup \bigg(\bigcup_{E\cap F_i =\emptyset} E\times I\bigg) \cup
\bigg(\bigcup_{\substack{E\cap F_i \neq \emptyset \\ b_1\notin E}} E_{(E\cap F_i)}\bigg),
\end{equation*}
where $E$ is a face of $Q$.
Next, we take $b_2'=b_1^-$ and $E_2'=Q^-$ which is the unique maximal
face of $Q_{(J')}$ containing $b_2'$.
Now, $B_3'$ is naturally defined by deleting faces of $Q^-$ which contains
$b_2'$ from $B_2'$. Observe that neither $Q^+$ nor $Q^-$ is a face
of $B_3'$, because a vertex in each of $Q^+$ and $Q^-$ has been removed.
See the third retraction sequence in Figure \ref{fig_ret_wedge_5-gon} for an
example.
\end{enumerate}
In both (Case1) and (Case2), one can see that
$$\gcd\{p, |G_{E_1'}(b_1')|\}=\gcd\{p, |G_{E_2'}(b_2')|\}=1$$
by Proposition \ref{prop_sub_local_group},
Lemma \ref{lem_loc_goup_isom_for_v_0} and
Lemma \ref{lem_loc_group_isom_for_v_pm}.
Finally, the remaining terms of the desired retraction sequence
$\{(B^{\prime}_s, E^{\prime}_s, b_s^{\prime})\}_{s=1}^{2\ell-k}$ can be obtained
by setting the vertices in the sequence \eqref{eq_Q(J)_seq_of_free_vertices}
as the desired sequence of free vertices $\{b_s'\}_{s=1}^{2\ell-k}$.
Indeed, it is enough to verify the following claims:
\begin{enumerate}
\item[(a)] Assume that $b_r\notin V(F_i)$, $b_s'=b_r^+$ and $b_{s+1}'=b_r^-$.
Then, $b_s'$ and $b_{s+1}'$ are free vertices in $B_s'$ and $B_{s+1}'$,
respectively. In particular, $E_s'$ and $E_{s+1}'$ are determined by
$E_r\times I$ and $E_r\times \{0\}$, respectively.
\item[(b)] Assume that $b_r\in V(F_i)$ and $b_s'=b_r^-$.
Then, $b_s'$ is a free vertex in $B_s'$ with a unique maximal face
$E_s'={E_r}_{(E_r\cap F_i)}$.
\item[(c)] $\gcd\{p, |G_{E_s'}(b_s')|\}=1$ for each $s=1, \dots, \ell'$.
\end{enumerate}
The claims follow by repeating the arguments in (Case 1) and (Case 2) above.
\end{proof}
Now, Theorem \ref{thm_no_p-torsion}, \ref{thm_no-torsion} and Lemma \ref{lem_ret_simj}
concludes the following two theorems.
\begin{theorem}
Let $X:=X(Q, \lambda)$ be a toric orbifold satisfying the assumption of
Theorem \ref{thm_no_p-torsion}. Then
the cohomology $H^\ast(X_{(J)};\mathbb{Z})$ has no $p$-torsion and
$H_{odd}(X_{(J)}; \mathbb{Z}_p)$ is trivial for arbitrary $J=(j_1, \dots, j_m)\in \mathbb{N}^m$.
\end{theorem}
\begin{theorem}\label{thm_X_satisfies_assump_then_X(J)_eq_formal}
Let $X:=X(Q, \lambda)$ be a toric orbifold satisfying the assumption of
Theorem \ref{thm_no-torsion}.
Then, $H^\ast(X_{(J)};\mathbb{Z})$ is torsion free and $H_{odd}(X_{(J)}; \mathbb{Z})$ vanishes
for arbitrary $J=(j_1, \dots, j_m)\in \mathbb{N}^m$.
\end{theorem}
\section{Application to the cohomology ring}\label{sec_cohom_ring}
Motivated by the results of \cite{BFR}, a notion of \emph{weighted Stanley--Reisner ring}
was introduced and used in \cite{BSS} to explicitly
compute the singular cohomology ring with integer coefficients of
spaces identified as integrally equivariantly formal projective toric orbifolds.
In this section, we briefly introduce their results and
study the integral cohomology ring of $X_{(J)}$.
\subsection{Cohomology ring of $X$}
In this subsection, we briefly summarize the theory of weighted Stanley--Reisner
ring of a
\emph{simple lattice polytope}\footnote{A simple polytope in $\mathbb{R}^n$ whose vertices
belong to the lattice $\mathbb{Z}^n \subset \mathbb{R}^n$. } $Q$
which defines a projective normal toric variety, see \cite[Section 5]{BSS}.
The original definition of weighted
Stanley--Reisner ring is based on a polytopal fan, but we translate the notation
of \cite{BSS} to a simple lattice polytope which defines a polytopal fan.
Let $\mathcal{F}(Q)=\{F_1, \dots, F_m\}$ be the set of facets of $Q$.
Since $Q$ is a lattice polytope, we may choose a primitive inward normal vector
$\lambda_i$ of each facet $F_i$. Moreover, if $F_{i_1}\cap \dots \cap F_{i_k}=\emptyset$,
then $\lambda_{i_1}, \dots, \lambda_{i_k}$ is linearly independent, because $Q$
is simple. Hence, the set of primitive inward normal vectors forms
an $\mathcal{R}$-characteristic pair, say $(Q, \lambda)$.
Next, for each vertex $v=F_{s_1} \cap\dots \cap F_{s_n} \in V(Q)$,
we recall from Section \ref{sec_toric_orb_orb_lens}, (item (5) in page \pageref{item_(5)}),
the $(n\times n)$-matrix associated to $v$,
that is
$$\Lambda_{v} :=
\left[ \begin{array}{c|c|c} \lambda(F_{s_1})^t & \cdots & \lambda(F_{s_n})^t\end{array} \right].$$
For each vertex $v=F_{s_1}\cap \dots \cap F_{s_n}$, we define a vector
$$z^v:=(z^v_1, \dots, z^v_m) \in \bigoplus_m \mathbb{Q}[u_1, \dots, u_n],$$
by the following rule:
\begin{enumerate}\label{int_vectors}
\item[(i)] $z^v_j=0$ if $j\notin \{s_1, \dots, s_n\}$,
\item[(ii)] $\begin{bmatrix}
z^v_{s_1}\\ \vdots\\ z^v_{s_n}
\end{bmatrix} = \Lambda_{v}^{-1} \cdot \begin{bmatrix}u_1 \\ \vdots \\ u_n\end{bmatrix}$
\end{enumerate}
where the operation on the right hand side is the usual matrix multiplication.
Next, we define a subset of $\mathbb{Z}[x_1, \dots, x_m]$ as follows:
$${\rm Int}[Q, \lambda]:=\{ f(x_1, \dots, x_m) \mid f(z^v)\in \mathbb{Z}[u_1, \dots, u_n],
~\text{ for all } v\in V(Q)\}.$$
\begin{remark}\label{rmk_u_i-variables}
The variables $u_1, \dots, u_n$ stand for the basis of $H^2(BT^n;\mathbb{Z})$,
where $T^n$ is the $n$-dimensional torus acting on $X(Q, \lambda)$.
Indeed, one may regard polynomial rings $\mathbb{Q}[u_1, \dots, u_n]$ and $\mathbb{Z}[u_1, \dots, u_n]$
as $H^\ast(BT^n;\mathbb{Q})$ and $H^\ast(BT^n;\mathbb{Z})$, respectively.
We refer to \cite[Section 5.2]{BSS}.
\end{remark}
The next proposition highlights critical properties of ${\rm Int}[Q,\lambda]$.
\begin{proposition}
\begin{enumerate}
\item\label{prop_subring} The subset ${\rm Int}[Q, \lambda]$ is a subring of
$\mathbb{Z}[x_1, \dots, x_m]$.
\item\label{prop_SR-ideal} The \emph{Stanley--Reisner ideal}
$\mathcal{I}_Q :=\big\langle \prod_{j=1}^k x_{i_j} \mid F_{i_1} \cap \dots \cap F_{i_k}
= \emptyset \big\rangle$
of the ring $\mathbb{Z}[x_1, \dots, x_m]$ for $Q$ is again an ideal of ${\rm Int}[Q, \lambda]$.
\end{enumerate}
\end{proposition}
\begin{proof}
The first statement is almost obvious.
Next, if a monomial $f(x_1, \dots x_m)=\prod_{j=1}^k x_{i_j}$
is an element of $\mathcal{I}_Q$, it follows from item (i) above
that $f(z^v)=0$ for all $v\in V(Q)$.
Hence, $\mathcal{I}_Q$ is not only a subset,
but also an ideal of ${\rm Int}[Q, \lambda]$.
\end{proof}
\begin{definition}\cite[Section 5]{BSS}
The {\it weighted Stanley--Reisner ring} $w\mathcal{SR}[Q, \lambda]$ of an
$\mathcal{R}$-characteristic pair $(Q, \lambda)$, associated to a simple lattice polytope,
is the subring of the Stanley--Reisner ring $\mathcal{SR}[Q]$ of $Q$ defined by
the quotient:
$$w\mathcal{SR}[Q, \lambda]:= {\rm Int}[Q, \lambda] / \mathcal{I}_Q.$$
\end{definition}
We remark that if a simple lattice polytope is Delzant, i.e., normal vectors associated to
facets intersecting a vertex form a $\mathbb{Z}$-basis, then $\Lambda_{v}^{-1}$ has integer entries.
Hence, $f(z^v)$ is polynomial with integer coefficients for all $v\in V(Q)$, which says
that $w\mathcal{SR}[Q, \lambda]$ is the usual Stanley--Reisner ring of a simple lattice polytope. Hence,
$w\mathcal{SR}[Q, \lambda]$ contains geometric data about $X(Q,\lambda)$,
including singularities, in addition to the combinatorial information about $Q$.
\begin{theorem}\cite[Theorem 5.3]{BSS}\label{thm_cohom_of_X}
Let $X(Q, \lambda)$ be a projective toric variety over a simple lattice polytope $Q$
with $H^{odd}(X(Q, \lambda);\mathbb{Z})=0$. Then, there is an isomorphism
$$H^\ast(X(Q, \lambda);\mathbb{Z}) \cong w\mathcal{SR}[Q, \lambda]/\mathcal{J},$$
where $\mathcal{J}$ is the ideal generated by the linear elements
$ \sum_{i=1}^m \left< \lambda_i, \mathbf{e}_j \right>x_i$ for $j=1, \dots, n,$
where $\mathbf{e}_j$ denotes the $j$-th standard unit vector in $\mathbb{Z}^n$.
\end{theorem}
\subsection{The cohomology ring of $X_{(J)}$}
In this final section, we study the relationship between the cohomology ring of
$X$ and that of $X_{(J)}:=X(Q_{(J)}, \lambda_{(J)})$.
The next theorem follows directly
from Theorem \ref{thm_no-torsion} and Theorem \ref{thm_cohom_of_X}.
\begin{theorem}\label{thm_cohom_of_X(J)}
Let $(Q, \lambda)$ be an $\mathcal{R}$-characteristic pair satisfying
the hypothesis of Theorem \ref{thm_no-torsion}. Then,
$$H^\ast(X_{(J)};\mathbb{Z}) \cong w\mathcal{SR}[Q_{(J)}, \lambda_{(J)}] / \mathcal{J}_{(J)}, $$
where $\mathcal{J}_{(J)}$ is the ideal generated by
\begin{equation}\label{eq_ideal_I(J)}
\left\{\sum_{i=1}^m \langle \lambda_i, e_j\rangle x_{i1}~\Big|~ j=1, \dots, n\right\}\cup
\left\{ x_{it}=x_{i1}\mid t=2, \dots, j_i \right\}.
\end{equation}
\end{theorem}
If $X$ is a smooth toric manifold, then the cohomology ring $H^\ast(X_{(J)};\mathbb{Z})$
can be dramatically simplified, because
$w\mathcal{SR}[Q_{(J)}, \lambda_{(J)}]=\mathcal{SR}[Q_{(J)}, \lambda_{(J)}]$, which
tells us that the ring is generated by degree $2$ elements
and the second part of \eqref{eq_ideal_I(J)} reduces the degree $2$ elements
to the same generators as $w\mathcal{SR}[Q, \lambda]$. See \cite[Section 4]{BBCG15} for details.
However, in general, the cohomology ring $H^\ast(X;\mathbb{Z})$ of a toric orbifold $X$ is not
generated by degree $2$ elements, hence the multiplication structure has plenty of
divisibility because of singularities.
Recall that $X_{(J)}$ can be obtained from
$X$ by a sequence of simplicial wedge constructions.
Consider now the polytopal wedge
$Q_{(J')}:=Q_{(F_i)}$ of the original simple lattice polytope $Q$ for
some facet $F_i$ of $Q$.
We finish this paper by studying the vectors $\{z^{v^\epsilon} \mid v^\epsilon \in V(Q_{(J')}) \}$,
with respect to the $\mathcal{R}$-characteristic pair $(Q_{(J')}, {\lambda_{(J')}})$.
Recall $V(Q_{(J')})$ from \eqref{eq_vertices_of_Q_tilde}.
Theorem \ref{thm_J-fied_int_cond} below tells us how to get
$$z^{v^{\epsilon}}=(z^{v^\epsilon}_0, z^{v^\epsilon}_1, \dots, z^{v^\epsilon}_m) \in
\bigoplus_{m+1} \mathbb{Z}[u_0, u_1, \dots, u_n]$$
from $\{z^v \mid v\in V(Q)\}$, for each $v\in V(Q)$ and $\epsilon=+$ or $-$.
\begin{remark}
The polynomial ring $\mathbb{Z}[u_0, u_1, \dots, u_n]$ stands for $H^\ast(BT^{n+1};\mathbb{Z})$,
where $T^{n+1}$ is the $(n+1)$-dimensional torus acting on $X_{(J')}$.
The canonical embedding of $T^n$, the acting torus on $X(Q, \lambda)$, into
the last $n$-coordinates of $T^{n+1}$ yields a canonical surjection
$\mathbb{Z}[u_0, u_1, \dots, u_n] \twoheadrightarrow \mathbb{Z}[u_1, \dots, u_n]$. See
Remark \ref{rmk_u_i-variables}.
\end{remark}
According to \eqref{eq_wedge_of_K} and \eqref{eq_vertices_of_Q_tilde},
we have the following 3 types of vertices in $Q_{(J')}$;
\begin{align}
\label{eq_type1}&v^-_{i_r}=Q^+ \cap Q^- \cap \bigcap_{a=1}^{n-1}{F_{s_a}}_{(F_i \cap F_{s_a})}
\text{ for some } v=F_i \cap \bigcap_{a=1}^{n-1} F_{s_a} \in V(Q), \\
\label{eq_type2}& v^+= Q^+ \cap \bigcap_{a=1}^n (F_{s_a} \times I)
\text{ for some } v=\bigcap_{a=1}^n F_{s_a} \text{ and } v\notin V(F_i), \\
\label{eq_type3}&v^-= Q^- \cap \bigcap_{a=1}^n (F_{s_a} \times I)
\text{ for some } v=\bigcap_{a=1}^n F_{s_a} \text{ and } v\notin V(F_i).
\end{align}
We refer also to \eqref{eq_v_in_F_i_intersection_of_facets} and
\eqref{eq_vertex_in_Q(J)_not_in_F_i} for the indexing of facets in the above three cases.
\begin{theorem}\label{thm_J-fied_int_cond}
The weighted Stanley--Reisner ring of $(Q_{(J')}, \lambda_{(J')})$ is related to
that of $(Q, \lambda)$ as follows.
For each vertex of one of the three types \eqref{eq_type1},
\eqref{eq_type2} and \eqref{eq_type3} above, we have:
\begin{enumerate}
\item\label{item_(1)} $z^{v_{i_r}^-}= (u_0+z^{v_{i_r}}_i, z^{v_{i_r}}_1, \dots, z^{v_{i_r}}_m)$ for $r=1, \dots, k$;
\item\label{item_(2)} $z^{v^+}= (u_0, z^{v}_1, \dots, z^{v}_m)$;
\item\label{item_(3)} $z^{v^-}=(0, z^{v^-}_1, \dots, z^{v^-}_m)$, where
$$z^{v^-}_\ell=\begin{cases} 0 & \text{ if } \ell \notin \{s_1, \dots, s_n\} \\
\gamma_\ell u_0 + z^{v}_\ell & \text{ if } \ell \in \{s_1, \dots, s_n\} \end{cases} ~ \text{ and }~
\begin{bmatrix} \gamma_{1} \\ \vdots \\ \gamma_{n} \end{bmatrix}
=\Lambda_{v}^{-1}\cdot \lambda(F_i)^t.$$
\end{enumerate}
\end{theorem}
\begin{proof}
The proof follows from the direct computation of the inverse of
${\Lambda_{(J')}}_{v^\epsilon}$
for each $\epsilon=+$ and $-$. If $v$ is a vertex in $F_i$, i.e., $v=v_{i_r}$
for some $i_r\in \{i_1, \dots, i_k\}$, and
$v_{i_r}=F_{i} \cap F_{s_1} \cap \dots \cap F_{s_{n-1}}$,
then $\Lambda_{v}=
\left[ \begin{array}{c|c|c|c} \lambda(F_i)^t & \lambda(F_{s_1})^t &
\cdots & \lambda(F_{s_{n-1}})^t \end{array}\right]$
and
$$
{\Lambda_{(J')}}_{v^-_{i_r}}=
\left[\begin{array}{ccccc}1&-1& 0&\cdots& 0 \\
0& & & & \\
\vdots &\lambda(F_i)^t & \lambda(F_{s_1})^t
&\cdots & \lambda(F_{s_{n-1}})^t \\
0& & & & \end{array} \right].$$
Its inverse is
$${\Lambda_{(J')}}_{v^-_{i_r}}^{-1}
=\left[\begin{array}{c|ccc}
1 & d_{1}&\cdots & d_{n}\\ \hline
0 &&&\\
\vdots&& \Lambda_{v}^{-1}& \\
0 && & \end{array}\right],$$
where $(d_1, \dots, d_n)$ is the first row of $\Lambda_{v}^{-1}$.
Hence, by (i) and (ii) in page \pageref{int_vectors}, we conclude \eqref{item_(1)}.
Next, we assume that $v=F_{s_1} \cap \dots \cap F_{s_n}$ with $i\notin \{s_1, \dots, s_n\}$.
The square matrices corresponding to
$v^+=Q^+ \cap \bigcap_{a=1}^n (F_{s_a} \times I)$ and
$v^-=Q^- \cap \bigcap_{a=1}^n (F_{s_a} \times I)$ are
\begin{align*}
{\Lambda_{(J')}}_{v^+}&=\left[\begin{array}{c|ccc}1 & 0 & \cdots & 0 \\\hline
0& & & \\
\vdots & & \Lambda_{v} & \\
0& & & \end{array}\right]=
\left[\begin{array}{cccc}
1& 0& \cdots &0 \\
0&&&\\
\vdots& \lambda(F_{s_1})^t &\cdots& \lambda(F_{s_n})^t \\
0&&&
\end{array}\right], \\
{\Lambda_{(J')}}_{v^-} &=
\left[ \begin{array}{c|ccc}
1 & 0 & \cdots & 0\\ \hline
&&&\\
\lambda(F_i)^t & & \Lambda_v & \\
&&&
\end{array}
\right]=
\left[\begin{array}{cccc}
-1& 0& \cdots &0 \\
&&&\\
\lambda(F_i)^t& \lambda(F_{s_1})^t &\cdots& \lambda(F_{s_n})^t \\
&&&
\end{array}\right].
\end{align*}
respectively. Their inverses are
\begin{equation}\label{eq_type_2_inverse}
{\Lambda_{(J')}}_{v^+}^{-1}=
\left[\begin{array}{c|ccc}1 & 0 & \cdots & 0 \\ \hline
0& & & \\
\vdots & & \Lambda_{v}^{-1} & \\
0& & & \end{array}\right]
\end{equation}
and
\begin{equation}\label{eq_type_3_inverse}
{\Lambda_{(J')}}_{v^-}^{-1} = \left[ \begin{array}{c|ccc}
-1 & 0 & \cdots & 0\\ \hline
\gamma_1 & & &\\
\vdots & &\Lambda_v^{-1} &\\
\gamma_n & & &
\end{array}\right], \text{ where }
\begin{bmatrix} \gamma_1 \\ \vdots \\ \gamma_n \end{bmatrix}
=\Lambda_{v}^{-1}\cdot \lambda(F_i)^t.
\end{equation}
Now, the results for \eqref{item_(2)} and \eqref{item_(3)} are straightforward from
\eqref{eq_type_2_inverse} and \eqref{eq_type_3_inverse}, respectively.
\end{proof}
|
\section{Introduction}
Amidst rising costs of operating rooms (ORs), hospitals strive to satisfy high surgery demand with increasingly limited budgets. Costs in ORs largely depend on the allocation and scheduling decisions \citep{Denton2007}. Also, in literature relevant to surgery allocation and scheduling problems (\textit{e.g.}, \cite{denton2003sequential} and \cite{Ho-YinMak2014appointment}), surgery duration is considered as an important input. Hence, estimation of surgery duration is critical to hospital management. Evidence suggests that surgery duration depends on clinical factors, \textit{e.g.,} the surgery type (details of the procedure), the patient's anamnesis \citep{Lowndes2016}, as well as the surgeon's experience \citep{Zheng2008}. However, non-clinical impacts have not been investigated to the same extent. In this paper, we would like to investigate the impacts on surgery duration from a broad perspective, \textit{i.e.}, the clinical and non-clinical factors as well as the interactions between them.
The term ``surgery duration'' in this paper is somewhat different from that in literature. We define it explicitly as follows. The whole surgery process can be divided into four parts. The first part is the interval from the time a patient enters the OR to the start of anaesthesia, which is labelled preparation period. The second part is the interval from start to the end of anaesthesia, which we label anaesthesia period. The third part is the interval from the start (\textit{i.e.}, the end of anaesthesia) to the end of the surgical procedure, which we label surgery period. The last part is the interval from the end of the surgical procedure to the time the patient leaves the OR, which we call wake-up period. The term ``surgery duration'' throughout the paper means the length of the third period, \textit{i.e.}, that of the surgical procedure itself, since this time period is dominated by the surgeon and critical for surgical outcome quality.
In this paper, non-clinical factors refer to day-of-the-week, surgeon workload, the workload in an OR, and the surgery position in a sequence of surgeries. The workload of a surgeon means the number of surgeries a surgeon performs in a day. We find that surgery duration will increase with around 10 minutes if a surgeon performs one more surgery. The workload in an OR refers to the number of surgeries scheduled in an OR in a day. Interestingly, we find a non-linear relationship between workload in ORs and surgery duration. When there are no more than 4 surgeries in an OR, surgery duration decreases with around 10 minutes if one more surgery is allocated to the OR; and when there are more than 4 surgeries in an OR, surgery duration increases with around 5 minutes for each additional surgery. Surgery position is the position of a surgery in a sequence of surgeries a surgeon performs in a day. We find evidence of a relationship between its position and its duration. In other words, the duration of a surgery varies with its position in a sequence. Furthermore, we investigate the effects of the interactions among surgery type, position, and surgeon, by which we aim to uncover whether different surgeons exhibit different patterns on the effects of surgery type and position. Indeed, we find evidence suggesting that for certain surgery types, its position is linked to surgeons working faster (or slower). Note that we do not claim the generalizability of the findings in this paper, since they are obtained based on a particular data set. Instead, our paper provides insights into the surgery duration analysis in ORs. In particular, similar analysis can be repeated for any OR to find the clinic and non-clinical imparts on surgery duration. Note that the results in the paper have been shown in \cite{wang2017predicting}. In this paper, we present details of our data and the methodologies used to obtain the results.
Our work makes the following contributions to literature on the impacts on surgery duration:
\begin{itemize}
\item[(1)]
Usually, researchers use surgery duration as an input to make surgery allocation and scheduling decisions. However, we find that allocation and scheduling decisions in turn influence surgery duration.
\item[(2)] We investigate the effects of two kinds of workloads on surgery duration, \textit{i.e.}, surgeon workload and workload in ORs.
\item[(3)] We propose a new method to find the interactions between predictors. That is, the regression tree is used to indicate the possible interactions.
\item[(4)] We find that the surgery position in a sequence of surgeries a surgeon performs in a day impacts the surgery duration.
\item[(5)] We find that surgeon performance is related not only to surgery types but also to the surgery position in a day.
\end{itemize}
The remainder of this paper is organized as follows. Section 2 gives a review of literature. Section 3 develops hypotheses. Section 4 describes the data structure. In Section 5, we propose statistical models and illustrate the results. Section 6 presents concluding remarks along with some possible future work.
\section{Literature Review}
We review the literature relevant to the factors that impacts surgery duration, \textit{e.g.}, the day of the week, the surgeon performing the procedure, and the surgeon workload.
The weekday or weekend effects in healthcare have been investigated. \cite{cram2004effects} find that the mortality rate of patients admitted to hospitals on weekends is slightly higher than that of patients admitted on weekdays. \cite{kostis2007} conclude that patients with myocardial infarction admitted on weekends experience higher mortality and lower use of invasive cardiac procedures.
The impacts of the workload have been studied for decades. In industry, the empirical study of \cite{Schultz1999} illustrates that employees tend to work faster when they face rising inventory and the work is standardized. \cite{Oliva2001} find that employees can ``cut corners'' by omitting tasks to reduce waiting time for incoming customers. In healthcare, plenty of researches have investigated the effects of occupancy in medical ward on the length of stay (LOS), readmission rate, and mortality rate in hospitals. Some papers conclude that high occupancy levels lead to decreased LOS, \textit{e.g.}, \cite{Anderson2011} and \cite{Kc2012msom}. This kind of speed-up may have consequences such as high readmission rates\citep{Kc2012msom} and high mortality rates \citep{Kc2009}. It is worth noting that the workload in ORs discussed in this paper is different from that in medical wards. First, the workload or occupancy of medical wards in the above papers is represented by the number of patients. The workload is work needed simultaneously by the patients in the ward, while the workload of a surgeon in a day in this paper is the number of patients who need surgery sequentially. Hence the workload of a surgeon should be time-dependent. Specifically, a surgeon may experience workload pressure during surgeries earlier in the day, but not during the latter surgeries. Or it may be the other way around. Second, different from patients in medical wards, a surgery is often performed by one surgeon or a surgeon team. Hence, the surgery duration is often related to the surgeon performing the procedure.
In the literature, surgery duration has often been linked to the surgeon performing the procedure. In particular, it is influenced by the performing surgeon's level of experience \citep{Zheng2008}, or to that of the assistant surgeon \citep{MolinaPariente2015}, or to that of the surgical team as a whole \citep{Zheng2008, Eijkemans2010}. Team size has also been investigated in relation to surgery duration \citep{Cassera2009}.
Some scholars found evidence that surgery duration is related to procedure start time. \cite{Peskun2012} found that surgery duration increases with later surgery start time, but the small differences found (7.1 minutes) were likely not clinically significant. For procedures that started after 5 PM, \cite{Cassera2009} found no significant increase in surgery duration. Different from the papers that focus on start time, in this paper, we study the relevant effects on the surgery duration from another perspective, i.e., the position of a surgery in a sequence of surgeries performed by a surgeon. It is also an aspect to illustrate whether surgery duration depends on the surgeon performing the procedure.
\section{Hypothesis Development}\label{sec:hypo}
Generally, it is believed that surgery duration depends on surgery type, and on the details of the surgical procedure, \textit{e.g.,} how many tumours need to be removed, how deep are they, etc. However, we suspect that some non-clinical factors influence surgery duration. In this section, we develop hypotheses which will be tested in section \ref{sec:ana}.
\subsection{Hypothesis 1: the Effects of Weekdays }
Literature has investigated the effects of weekdays on mortality, \textit{e.g.}, \cite{cram2004effects} and \cite{crowley2009influence}. It is concluded that the effects of weekends are often significant, \textit{e.g.,} the mortality rate of patients admitted during the weekend is higher. In this paper, we investigate whether effects of the weekdays influence surgery duration significantly. We formally present the hypothesis as follows.
\begin{itemize}
\item [\textbf{Hypothesis 1:}] The effects of weekdays influence the mean of surgery duration.
\end{itemize}
The corresponding \textit{null hypothesis} is that the effects of weekdays do not affect the mean of surgery duration.
\subsection{Hypothesis 2: the Effects of Workload} \label{sec:hypo-2}
We focus on two kinds of workload: surgeon workload and workload in ORs. A surgeon might perform a different number of surgeries in one day. We conjecture that surgery duration decreases with the number of surgeries performed by the surgeon in a day. More specifically, surgeons under high workload will accelerate their surgeries. On the other hand, the number of surgeries that scheduled in an OR varies day by day. Hence, we also conjecture that if many surgeries are scheduled in an OR, the workload pressure in the OR will also cause surgeons to accelerate their surgeries. We present these conjectures formally in the following hypothesis.
\begin{itemize}
\item [\textbf{Hypothesis 2:}] the mean of surgery duration decreases with (1) the number of surgeries a surgeon performs in a day, and (2) the number of surgeries scheduled in the OR.
\end{itemize}
The corresponding \textit{null hypothesis} is that the mean of surgery duration does not vary with the number of surgeries a surgeon preforms in a day, or with the number of surgeries scheduling in the OR.
\subsection{Hypothesis 3: the Effects of Orderings}
We also suggest that surgery duration is affected by the position of the surgery in the sequence of surgeries that a surgeon performs in a day. Specifically, the duration of prior surgeries will be shorter, while that of the later surgeries will increase. Our reasoning is twofold. Firstly, if a surgeon performs multiple surgeries in a day, he/she may experience higher workload pressure when performing the earlier surgeries, knowing that there are still a number of surgeries to come. Hence, the surgeon will accelerate his/her work and the surgery duration will decrease. For latter surgeries we suppose the opposite. Secondly, if a surgery is one of the first procedures for a surgeon that day, it is usually one of the first surgeries in the OR in which multiple surgeries are scheduled. Hence, the earlier surgeries suffer more OR workload pressure than the later ones. We formally present our conjectures about the effects of orderings in Hypothesis 3.
\begin{itemize}
\item[\textbf{Hypothesis 3:}] The mean of surgery duration decreases (increases), if a surgery takes place earlier (later) in the sequence of surgeries performed by a surgeon.
\end{itemize}
The corresponding \textit{null hypothesis} is that the mean of surgery duration does not decrease or increase with its position in the sequence of surgeries performed by a surgeon.
\subsection{Hypothesis 4: the Effects of Surgeons}
Though the surgery duration mainly depends on surgery type, we suspect that different surgeons exert different performance. For example, some surgeons are senior specialists who may need less time than junior surgeons for the same procedures. Surgery duration may also be linked to a surgeon's personal style, or preference for the use of certain surgical tools. We present the hypothesis as follows.
\begin{itemize}
\item [\textbf{Hypothesis 4:}] The mean of surgery duration is related to the surgeon performing the procedure.
\end{itemize}
The corresponding \textit{null hypothesis} is that the mean of surgery duration has nothing to do with the surgeon performing the procedure.
\subsection{Hypothesis 5: the Effects of Interactions}
We conjecture that surgeons are not homogeneous. Firstly, a surgeon may be good at certain types of surgeries, but less so at others, which indicates a potential interaction between surgeon and surgery type. Secondly, different surgeons may have different effects of orderings, \textit{i.e.}, an interaction between surgeons and orderings. Additionally, other interactions may also be argued, \textit{e.g.}, the interaction between orderings and surgery types, hence we will investigate the effects of interactions. We present the hypothesis formally as follows.
\begin{itemize}
\item [\textbf{Hypothesis 5:}] The interactions between some factors, \textit{e.g.}, surgeons, orderings and surgery types, influence surgery duration.
\end{itemize}
The corresponding \textit{null hypothesis} is that there are no interactions of surgeons, orderings and surgery types that affect mean of surgery duration.
\section{Data Collection and Description} \label{sec:dataDescr}
Our data come from a large hospital in China from January 2014 through October 2016. There are 46 ORs in the hospital, which together host more than 20,000 surgeries a year. We describe the data structure from the following perspectives.
\begin{table}[h]
\centering
\caption{Surgery statistics I: Frequency distribution of the number of surgeries performed by each surgeon in a day}
\label{tab:sta1}
\begin{tabular}{crrrrrrrrr}
\toprule[1.5pt]
\multirow{2}{*}{Surgeon} & \multicolumn{1}{c}{Number} & \multicolumn{1}{c}{\multirow{2}{*}{Persent}} & \multicolumn{1}{c}{} & \multicolumn{6}{c}{Number of surgeries in a day} \\ \cmidrule[0.7pt]{5-10}
& \multicolumn{1}{c}{of surgeries} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{1} & \multicolumn{1}{c}{2} & \multicolumn{1}{c}{3} & \multicolumn{1}{c}{4} & \multicolumn{1}{c}{5} & \multicolumn{1}{c}{6} \\
\midrule[1pt]
A & 919 & 37.49\% & & 128 & 308 & 332 & 143 & 8 & 0 \\
& & & & 13.93\% & 33.51\% & 36.13\% & 15.56\% & 0.87\% & 0.00\% \\
B & 507 & 20.69\% & & 150 & 200 & 89 & 52 & 10 & 6 \\
& & & & 29.59\% & 39.45\% & 17.55\% & 10.26\% & 1.97\% & 1.18\% \\
C & 484 & 19.75\% & & 98 & 265 & 103 & 8 & 10 & 0 \\
& & & & 20.25\% & 54.75\% & 21.28\% & 1.65\% & 2.07\% & 0.00\% \\
D & 296 & 12.08\% & & 104 & 119 & 56 & 15 & 1 & 1 \\
& & & & 35.14\% & 40.20\% & 18.92\% & 5.07\% & 0.34\% & 0.34\% \\
E & 170 & 6.94\% & & 103 & 56 & 11 & 0 & 0 & 0 \\
\multicolumn{1}{l}{} & & & & 60.59\% & 32.94\% & 6.47\% & 0.00\% & 0.00\% & 0.00\% \\
F & 75 & 3.06\% & & 21 & 29 & 21 & 4 & 0 & 0 \\
\multicolumn{1}{l}{} & & & & 28.00\% & 38.67\% & 28.00\% & 5.33\% & 0.00\% & 0.00\% \\
\midrule[1pt]
\multicolumn{1}{l}{} & 2451 & & & 604 & 977 & 612 & 222 & 29 & 7 \\
\multicolumn{1}{l}{} & & & & 24.64\% & 39.86\% & 24.97\% & 9.06\% & 1.18\% & 0.29\% \\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
\subsection{Surgeons}
We limit our data to one department: Thoracic Surgery department which performs the largest number of surgeries during our data collection period. We focus on 6 surgeons who performed the most surgeries. The number of surgeries performed by each surgeon is listed in the second column of Table \ref{tab:sta1}, in which the anonymous surgeons are denoted by letters, ``A", ``B", ``C", ``D", ``E'' and ``F". Surgeon A, B and C perform the most number of surgeries, which take up more than $75\%$ of the total number.
\begin{figure}[h]
\centering
\begin{tabular}{cc}
\includegraphics[width = 8cm]{durNumInDay2} & \includegraphics[width=8cm]{durNumInOR2} \\
\text{\footnotesize (a)} & \text{\footnotesize (b)} \\%[4pt]
\includegraphics[width=8cm]{durSurgeryType1} & \includegraphics[width=8cm]{durWeekday1} \\
\text{\footnotesize (c) } & \text{\footnotesize (d)}
\end{tabular}
\caption{\label{fig:surgeyDur} Relationship between surgery duration and other surgery factors such as (a) Surgery type, (b) Surgeon workload (c) Workload in OR (d) Weekday/weekend}
\end{figure}
\subsection{Number of Surgeries a Surgeon Performs in a Day}
A surgeon may perform multiple surgeries in a day. For the surgeons in consideration, the number a surgeon performs in a day varies from 1 to 6. The columns labelled 1-6 in Table \ref{tab:sta1} illustrates these statistics. The labels in the title row ($1, 2, \cdots, 6$) represent the number of surgeries that are performed by the surgeon in a day. Take the number in the row labelled ``Surgeon A'' and the column labelled ``2'', 308, for an example. It means that there are 308 surgeries each of which is performed by Surgeon A in the day with 2 surgeries finished by the surgeon. The last row lists the sum of the above numbers in each column. Note that the number in the last row might not be divisible by the number in the title row. This is because some instances are not included due to erroneous or incomplete records. In most cases surgeons perform fewer than four surgeries in a day. Also, different surgeons show different patterns. Specifically, Surgeon A performs two or three surgeries a day, while in most of the cases Surgeon B, C, D, and F tend to perform two surgeries in a day, especially for Surgeon B. Surgeon E is the only surgeon who usually performs only one surgery in a day. Figure \ref{fig:surgeyDur}(a) demonstrates 6 box-plots of surgery durations, corresponding to the cases in which 1-6 surgeries are performed in one day, respectively. There are only a few instances in which six or more surgeries are performed in one day, hence six or more is combined as such in the sixth box-plot of Figure \ref{fig:surgeyDur}(a). It is shown that surgery duration decreases with the number of surgeries performed by a surgeon in a day. This is reasonably expected, since surgeons may accelerate if they face higher work pressure, as was also shown in \cite{Kc2009}. Also, there are many outliers if a surgeon performs two or three surgeries in a day, because there are many cases for the two scenarios. In our model, we exclude the cases with six or more surgeries a day, since there are only seven such cases, representing only 0.29\% of the data.
\subsection{Number of Surgeries Scheduled in An OR}\label{sec:numInOR}
\begin{table}[h]
\centering
\caption{Surgery statistics II: Frequency distribution of the number of surgeries scheduled in ORs performed by each surgeon }
\label{tab:numInOR}
\begin{tabular}{crrrrrrrrrr}
\toprule[1.5pt]
\multicolumn{1}{l}{\multirow{2}{*}{Surgeon}} & \multicolumn{1}{c}{Number} & \multicolumn{1}{c}{\multirow{2}{*}{Percent}} & \multicolumn{1}{c}{} & \multicolumn{7}{c}{Number of surgeries in an OR in a day} \\ \cmidrule[0.7pt]{5-11}
\multicolumn{1}{l}{} & \multicolumn{1}{c}{of surgeries} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{1} & \multicolumn{1}{c}{2} & \multicolumn{1}{c}{3} & \multicolumn{1}{c}{4} & \multicolumn{1}{c}{5} & \multicolumn{1}{c}{6} & \multicolumn{1}{c}{7} \\
\midrule[1pt]
A & 919 & 37.49\% & & 76 & 473 & 284 & 55 & 23 & 8 & 0 \\
& & & & 8.27\% & 51.47\% & 30.90\% & 5.98\% & 2.50\% & 0.87\% & 0.00\% \\
B & 507 & 20.69\% & & 36 & 236 & 140 & 73 & 14 & 5 & 3 \\
& & & & 7.10\% & 46.55\% & 27.61\% & 14.40\% & 2.76\% & 0.99\% & 0.59\% \\
C & 484 & 19.75\% & & 43 & 290 & 103 & 31 & 10 & 5 & 2 \\
& & & & 8.88\% & 59.92\% & 21.28\% & 6.40\% & 2.07\% & 1.03\% & 0.41\% \\
D & 296 & 12.08\% & & 59 & 174 & 47 & 11 & 1 & 4 & 0 \\
& & & & 19.93\% & 58.78\% & 15.88\% & 3.72\% & 0.34\% & 1.35\% & 0.00\% \\
E & 170 & 6.94\% & & 31 & 86 & 31 & 17 & 3 & 1 & 1 \\
& & & & 18.24\% & 50.59\% & 18.24\% & 10.00\% & 1.76\% & 0.59\% & 0.59\% \\
F & 75 & 3.06\% & & 11 & 37 & 27 & 0 & 0 & 0 & 0 \\
& & & & 14.67\% & 49.33\% & 36.00\% & 0.00\% & 0.00\% & 0.00\% & 0.00\% \\
\midrule[1pt]
& 2451 & & & 256 & 1296 & 632 & 187 & 51 & 23 & 6 \\
& & & & 10.44\% & 52.88\% & 25.79\% & 7.63\% & 2.08\% & 0.94\% & 0.24\% \\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
Similar to the illustration in Table \ref{tab:sta1}, Table \ref{tab:numInOR} illustrates the number of surgeries scheduled in the OR in a day. It is shown that in most cases there are no more than three surgeries in an OR, and surgeons have similar patterns that most of the surgeries are done in an OR with two surgeries in a day. Figure \ref{fig:surgeyDur}(b) is the box-plot of the surgery duration under a different number of surgeries in an OR. As in Figure \ref{fig:surgeyDur}(a), we combine the surgeries performed in an OR of seven or more in a day. There is a tipping point. Specifically, in the cases that there are four or fewer surgeries in an OR, the surgery duration decreases with the number of surgeries, while the duration increases if the number of surgeries in an OR is five or more. We will test the existence of a tipping point later.
\subsection{Surgery Types}
\begin{table}[h]
\centering \small
\caption{Summary statistics of surgery duration for surgery types}
\label{tab:DurSta}
\begin{tabular}{clrrrrrr}
\toprule[1.5pt]
Index & Surgery Name & Frequency & Percent & Median & Min & Mean & Max \\\midrule[1pt]
1 & Surgery for lung cancer & 99 & 4.04\% & 148.85 & 47.17 & 154.05 & 440.05 \\
2 & Thoracoscopic pulmonary bullous resection & 61 & 2.49\% & 80.00 & 30.33 & 92.05 & 299.83 \\
3 & Thoracoscopic partial pulmonary lobectomy & 443 & 18.07\% & 128.62 & 37.00 & 133.14 & 308.70 \\
4 & Thoracoscopic interior pulmonary lobectomy & 930 & 37.94\% & 121.60 & 33.67 & 129.12 & 410.00 \\
5 & Total pneumonectomy & 27 & 1.10\% & 185.00 & 88.67 & 194.05 & 362.03 \\
6 & Partial pulmonary lobectomy & 280 & 11.42\% & 136.93 & 52.00 & 140.54 & 332.33 \\
7 & Thoracoscopic exploration & 32 & 1.31\% & 114.75 & 57.00 & 145.58 & 499.52 \\
8 & Pulmonary wedge resection & 58 & 2.37\% & 111.30 & 48.50 & 122.31 & 241.37 \\
9 & Oesophageal cancer & 50 & 2.04\% & 257.46 & 156.52 & 278.82 & 522.00 \\
10 & Mediastinal tumour resection & 142 & 5.79\% & 97.09 & 36.17 & 111.02 & 335.03 \\
11 & Pulmonary tumour resection & 329 & 13.42\% & 117.72 & 29.33 & 125.54 & 304.17 \\\midrule[1pt]
& & 2451 & 100.00\% & 125.00 & 29.33 & 133.53 & 522.00\\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
The surgeons in the department are specialized in thoracic surgeries, which can be categorized into 11 types. Table \ref{tab:DurSta} presents surgery types and the corresponding statistics. The fourth type of surgery has the highest occurrence and is the main surgical procedure that the department performs. Figure \ref{fig:surgeyDur}(c) shows the box plot of surgery duration for each surgery type. Surgery 9 (oesophageal cancer) has the longest duration. The duration of another cancer, lung cancer (Surgery 1), is also longer than most surgeries. For the main surgery in the department (pulmonary lobectomy), total pneumonectomy (Surgery 5) is reasonably more time-consuming than partial/interior pulmonary lobectomy (Surgery 3 and 4). The mean of Surgery 2 duration is the smallest. The duration of Surgery 3 and 4 have similar means and medians. They also have the highest number of outliers, which is may be due to their high occurrence.
\subsection{Number of Surgeries in the day of the week}\label{sec:weekday}
\begin{table}[h]
\centering
\caption{Number surgeries performed in the day of the week}
\label{tab:weekday}
\begin{tabular}{crr}
\toprule[1.5pt]
Weekday & Number & Percent \\
\midrule[1pt]
Sun & 30 & 1.22\% \\
Mon & 44 & 1.80\% \\
Tue & 509 & 20.77\% \\
Wen & 491 & 20.03\% \\
Thu & 435 & 17.75\% \\
Fri & 563 & 22.97\% \\
Sat & 379 & 15.46\% \\
\midrule[1pt]
& 2451 & 100.00\%\\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
We illustrate the number of surgeries in each day of the week in Table \ref{tab:weekday}, which shows that there are only a small portion of surgeries scheduled on Sunday and Monday. Hence, to some extent, Sunday and Monday represent the ``weekends'' for the surgeons. Figure \ref{fig:surgeyDur}(d) shows the box-plot of surgery duration for each day of the week. The mean of surgery duration does not seem significantly different for different days of the week.
\section{Statistical Modelling and Analysis}\label{sec:ana}
In this section, we propose statistical models to test the hypotheses developed in Section \ref{sec:hypo}.
\subsection{Variables}
The variables in consideration are listed in Table \ref{tab:variables}. The terms, ``variable'', ``predictor'', and ``factor'', are used interchangeably throughout the paper.
We divide the variables into two categories: clinical variables and non-clinical variables. Clinical variable is $SURGRYTYPE$, while the non-clinical variables include $SURGEON$, $ANESTHETIST$, $DAY$, $W\_SURGEON$ (the number of surgeries the surgeon performs in a day), $ORDER$ (the position of a surgery in the sequence of the surgeries performed by the surgeon in a day) and $W\_OR$ (the number of surgeries scheduled in the OR). $ORDER$ is a categorical variable, taking the value of $ONLYONE$ if a surgery is the only surgery a surgeon performed in a day, ``$2\sim1$'' if the surgery is the first surgery that is done by a surgeon who performs two surgeries on that day, ``$2\sim 2$'', ``$3\sim 1$'', and so forth.
\begin{table}[h]
\centering
\caption{\label{tab:variables} Description of variables used in the models}
\label{tab:varIntro}
\begin{tabular}{r|p{12cm}}
\toprule[1.5pt]
\textbf{Variable} & \textbf{Description}\\
\midrule[1pt]
$SURGEON$ & the surgeon performing the procedure\\
$ANESTHETIST$ & the surgery anesthetist \\
$SURGRYTYPE$ & the surgery type, one of 11 types shown in Table \ref{tab:DurSta} \\
$DAY$ & the day of the week when the surgery is performed\\
$W\_SURGEON$ & the number of surgeries the surgeon performs in a day\\
$ORDER$ & the position of a surgery in the sequence of surgeries performed by the surgeon in a day \\
$W\_OR$ & the number of surgeries scheduled in the OR where the surgery is performed\\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
\subsection{Simple Models}
We first use simple models (the models without interactions) to test Hypothesis 1-4. In this paper, we regard the effects of the predictors listed in Table \ref{tab:variables} as the main effects. In Figure 1 we observe that the the response variable is very skewed. This is against the assumptions that we make when we fit a linear model and has negative impact in the estimation of the model. To avoid this problem, we take the log of surgery duration and use it as the response. Since the variables $ORDER$ and $W\_SURGEON$ are collinear, the following models (Model I and II) with the two variables are developed respectively.
\begin{align*}
& \begin{array}{rl}
\textbf{Model I: }log(DURATION_i) =& \beta_0 + \beta_1 DAY_i+ \beta_2 W\_SURGEON_i + \beta_{4} W\_OR_i + \beta_5 SURGEON_i\\
&+ \beta_6 SURGERYTYPE_i + \beta_7 ANESTHETIST_i
\end{array}\\
& \begin{array}{rl}
\textbf{Model II: }log(DURATION_i) =& \beta_0 + \beta_1 DAY_i+ \beta_3 ORDER_i + \beta_{4} W\_OR_i + \beta_5 SURGEON_i\\
&+ \beta_6 SURGERYTYPE_i + \beta_7 ANESTHETIST_i
\end{array}
\end{align*}
Inspired by Figure \ref{fig:surgeyDur}(c), we would like to test different effects of $W\_OR$ below and above the tipping point (denoted as $TP$). Hence, we use the following piecewise linear function to substitute $\beta_{4} W\_OR_i$ in Model I and II.
\begin{equation*}
\beta_{41} W\_OR1_i + \beta_{42} W\_OR2_i = \beta_{41} W\_OR_i + \beta_{42} \max\{W\_OR_i - TP,0\}
\end{equation*}
which indicates that $\beta_{41}$ estimates the slope below the tipping point, and $\beta_{41}+\beta_{42}$ estimates the slope above the tipping point. Thereby we have the following two models.
\begin{align*}
& \begin{array}{rl}
\textbf{Model III: }log(DURATION_i) =& \beta_0 + \beta_1 DAY_i+ \beta_2 W\_SURGEON_i+\beta_{41}\min\{ W\_OR_i,TP\}\\
&+ \beta_{42} \max\{W\_OR_i - TP,0\}+ \beta_5 SURGEON_i\\
&+ \beta_6 SURGERYTYPE_i + \beta_7 ANESTHETIST_i
\end{array}\\
& \begin{array}{rl}
\textbf{Model IV: }log(DURATION_i) =& \beta_0 + \beta_1 DAY_i+ \beta_3 ORDER_i+\beta_{41}\min\{ W\_OR_i,TP\}\\
&+ \beta_{42} \max\{W\_OR_i - TP,0\}+ \beta_5 SURGEON_i \\
&+ \beta_6 SURGERYTYPE_i + \beta_7 ANESTHETIST_i
\end{array}
\end{align*}
The regression results are illustrated in Table \ref{tab:simM}. The terms relevant to the factor $ANESTHETIST$ are not included because it has 50 categories and it will add too many lines to the table. The factor $SURGERYTYPE$ is also excluded since by using the simple models we would like to test Hypotheses 1-4, which do not refer to surgery types.
Hypothesis 1 is supported by all models. To show the effect of the day of the week, we set Saturday as the baseline. In all models, the effects of Tuesday and Friday are positively significant, indicating that duration of surgeries on these two days in a week is longer than those on Saturday. If we compare Table \ref{tab:weekday} and Table \ref{tab:simM}, we see that the two busiest days, which are Tuesdays and Fridays, have slightly longer average operation time. They are followed by Wednesday and then the other week days. Besides, as mentioned in Section \ref{sec:weekday}, Sunday and Monday represent the weekends for the surgeons. Hence, longer duration in Tuesday might reveal a weekday effect, \textit{i.e.}, surgeons appear to work at a lower pace when they go back to work after the weekend. Numerically, the duration of surgeries on Tuesday or Friday is around 8 minutes longer than those on Saturday.
\begin{table}[h]
\centering
\caption{Regression results for models using only the main effect and without interaction terms (simple models). Anesthetist and surgery type were excluded because of space.}
\label{tab:simM}
\begin{tabular}{lllll}
\toprule[1.5pt]
& \multicolumn{4}{c}{Coefficient} \\\cmidrule[0.7pt]{2-5}
Variable & Model I & Model II & Model III & Model IV \\ \midrule[1pt]
Sunday & -0.0294 & -0.0313 & -0.0412 & -0.0445\\
Monday & 0.0557 & 0.0661 & 0.0652 & -0.0766\\
Tuesday & 0.0600* & 0.0613* & 0.0561* & 0.0572* \\
Wednesday & 0.0418 & 0.0442 & 0.0328 & 0.0348\\
Thursday & 0.0072 & 0.0084 & 0.0074 & 0.0087\\
Friday & 0.0552* & 0.0544* & 0.0543* & 0.0533*\\
Order 2$\sim$1 & & -0.081*** & & -0.0748** \\
Order 2$\sim$2 & & -0.0756** & & -0.0669** \\
Order 3$\sim$1 & & -0.1835*** & & -0.1693*** \\
Order 3$\sim$2 & & -0.1558*** & & -0.1452*** \\
Order 3$\sim$3 & & -0.1722*** & & -0.1631*** \\
Order 4$\sim$1 & & -0.2343*** & & -0.206*** \\
Order 4$\sim$2 & & -0.1688** & & -0.1413* \\
Order 4$\sim$3 & & -0.1528** & & -0.1289* \\
Order 4$\sim$4 & & -0.1752** & & -0.1483** \\
Order 5$\sim$1 & & -0.3204* & & -0.2874. \\
Order 5$\sim$2 & & -0.08 & & -0.051 \\
Order 5$\sim$3 & & -0.1702 & & -0.1318 \\
Order 5$\sim$4 & & -0.4697** & & -0.4426* \\
Order 5$\sim$5 & & -0.1469 & & -0.1055 \\
W\_SURGEON & -0.0686*** & & -0.0613*** & \\
W\_OR1 & & & -0.0662*** & -0.0667*** \\
W\_OR2 & & & 0.1033** & 0.1093** \\
W\_OR & -0.0377*** & -0.0371*** & & \\
Surgeon B & -0.0978*** & -0.1016*** & -0.0936*** & -0.0977*** \\
Surgeon C & 0.0569* & 0.0577* & 0.0579* & 0.0586* \\
Surgeon D & -0.0037 & -0.0054 & -0.006 & -0.0081 \\
Surgeon E & -0.039 & -0.0433 & -0.0345 & -0.0391 \\
Surgeon F & -0.2761*** & -0.2746*** & -0.2722*** & -0.2704*** \\\midrule[0.5pt]
Log likelihood &\multicolumn{1}{c}{-1079} & \multicolumn{1}{c}{-1073} & \multicolumn{1}{c}{-1070} & \multicolumn{1}{c}{-1064} \\
Adjusted $R^2$ & \multicolumn{1}{c}{0.198} & \multicolumn{1}{c}{0.197} & \multicolumn{1}{c}{0.203} & \multicolumn{1}{c}{0.203}\\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
Hypothesis 2(1) is supported by Model I and III, \textit{i.e.}, surgery duration decreases with the number of surgeries a surgeon performs in a day. On average, surgery duration will decrease by 10 minutes if a surgeon performs one more surgery. It seems that Hypothesis 2(2) is supported by Model I and II, \textit{i.e.}, surgery duration decreases with the number of surgeries scheduled in an OR. However, Model III and IV with tipping point demonstrate a better fit, with larger log likelihood and adjusted $R^2$ values. Hence, we can conclude that a tipping point of the workload in OR exists. That is, surgery duration decreases with the number of surgeries allocated to an OR if there are no more than 4 surgeries in the OR, while the duration increases with the number if there are more than 4 surgeries allocated to the OR. As we mentioned in Section \ref{sec:hypo-2}, high workload in an OR results in decreased surgery duration. However, the regression result demonstrates that surgery duration will increase if the number of surgeries scheduled in an OR in a day is beyond a certain value (say 4 for our data set). One possible reason is that if too many surgeries are allocated in an OR, the OR becomes disordered, which results in longer surgery duration. Numerically, when there are no more than 4 surgeries in an OR, surgery duration decreases around 8 minutes if one more surgery is allocated into the OR; and when there are more than 4 surgeries in an OR, surgery duration increases around 12 minutes if one more surgery is added to the OR.
\begin{figure}[h]
\centering
\begin{centering}
\subfloat[]{\begin{centering}
\includegraphics[width=8cm]{coeffOrder}
\par\end{centering}
\label{fig:coeffOrder}}
\subfloat[]{\begin{centering}
\includegraphics[width=8cm]{coeffSurgeon}
\par\end{centering}
\label{fig:coeffSurgeon}
}
\par\end{centering}
\caption{\label{fig:coeff}Estimation of (a) Surgeon effects and (b) Ordering position effects with error bars representing their standard errors}
\end{figure}
As for Hypothesis 3, it is partially supported by Model II and IV. Though some items are not significant when a surgeon performs five surgeries in a day, they only take up 1.18\% of all cases (see Table \ref{tab:sta1}). Note that the base line is $ONLYONE$. All coefficients of orderings are negative, which also support Hypothesis 2(1). The coefficients of different orderings do not always monotonically increase, but are very close for a certain number of surgeries a surgeon performs in a day (see the coefficients of Order $3\sim \cdot$ and Order $4\sim \cdot$). We can only conclude that the duration of the first surgery decreases more than later surgeries when a surgeon has two, three or four surgeries in a day. Figure \ref{fig:coeffOrder} shows a bar plot of coefficient estimations of surgery positions. It shows that the standard errors increase with the number of surgeries performed by a surgeon in a day, especially when a surgeon performs five surgeries. For the extend to which the duration decreases, duration of surgeries in position (4, 1) suffers the biggest drop around 28 minutes (for the cases where there are no more than 4 surgeries per surgeon in a day).
Hypothesis 4 is supported by all models, \textit{i.e.}, the mean of surgery duration is related to the surgeon performing the procedure. We set Surgeon A as the baseline for the categorical variable $SURGEON$. It indicates that the effects of Surgeon B, C and F are significant. The duration of surgeries performed by Surgeon B and F (especially Surgery F) is less than that performed by Surgeon A, while the duration of Surgeon C's surgeries is longer than that of Surgeon A. Figure \ref{fig:coeffSurgeon} is the bar plot of coefficient estimations of surgeons, indicating the differences among the surgeons. Surgeon C appears to be slowest, with average surgery duration approximately 13 minutes more than that of Surgeon A. Surgeon F appears to be fastest, whose surgery duration is around 34 minutes less than that of Surgeon A, but has the largest standard error.
To summarize, in this subsection we tested Hypothesis 1-4. In other words, we investigated effects of the non-clinical factors on surgery duration, including day of the week, surgeon workload, workload in an OR, the position of the surgery, and surgeon. Specifically, Hypothesis 1, 2(1), 3, and 4 are supported by the regression results. For Hypothesis 2(2), we find that surgery duration does not monotonically decrease with the workload in the OR. Instead, there exists a tipping point. Surgery duration will increase if the workload in the OR is beyond the tipping point.
\subsection{Models with Interactions}
\begin{figure}[h]
\centering
\includegraphics[width = 10cm]{tree1}
\caption{\label{fig:tree}Regression tree of Surgery duration by the model predictors. The objective of this tree is to uncover potential interactions terms for the model}
\end{figure}
In this subsection, we would like to test Hypothesis 5, \textit{i.e.}, the effects of some interactions on surgery duration. The question is: which interactions should be taken into account? We use the \verb|rpart()| function in R to build a regression tree of surgery duration by the model predictors. It is illustrated in Figure \ref{fig:tree}, which provides some evidence that there may be interactions among $SURGERYTYPE$, $SURGEON$ and $ORDER$. Hence, there are three kinds of interactions we need to consider.
We formulate the models with interactions based on Model IV as follows, since it is the best among the simple models.
\begin{align*}
& \begin{array}{rl}
\textbf{Model V: }log(DURATION_i) = \beta_0 + \beta_1 DAY_i+ \beta_3 ORDER_i+\beta_{41}\min\{ W\_OR_i,TP\}&\\
+ \beta_{42} \max\{W\_OR_i - TP,0\}+ \beta_5 SURGEON_i + \beta_6 SURGERYTYPE_i &\\
+ \gamma_1 SURGEON_i: SURGERYTYPE_i + \beta_7 ANESTHETIST_i&
\end{array} \\
& \begin{array}{rl}
\textbf{Model VI: }log(DURATION_i) = \beta_0 + \beta_1 DAY_i+ \beta_3 ORDER_i+\beta_{41}\min\{ W\_OR_i,TP\}&\\
+ \beta_{42} \max\{W\_OR_i - TP,0\}+ \beta_5 SURGEON_i + \beta_6 SURGERYTYPE_i &\\
+ \gamma_2 SURGEON_i:ORDER_i + \beta_7 ANESTHETIST_i&
\end{array} \\
& \begin{array}{rl}
\textbf{Model VII: }log(DURATION_i) = \beta_0 + \beta_1 DAY_i+ \beta_3 ORDER_i+\beta_{41}\min\{ W\_OR_i,TP\}&\\
+ \beta_{42} \max\{W\_OR_i - TP,0\}+ \beta_5 SURGEON_i + \beta_6 SURGERYTYPE_i &\\
+ \gamma_3 SURGERYTYPE_i:ORDER_i + \beta_7 ANESTHETIST_i&
\end{array}
\end{align*}
Next, we need to examine whether the new models are superior to the previous ones, \textit{i.e.}, whether the effects of interactions are significant. Hence, we run \verb|anova()| to perform analysis of variance for the results of Model IV and those of the new models, respectively. The results listed in Table \ref{tab:anova} indicate that the three kinds of interactions are all significant.
\begin{table}[h]
\centering
\caption{Results of analysis of variance for Model IV, V, VI, and VII}
\label{tab:anova}
\begin{tabular}{lcccccc}
\toprule[1.5pt]
& Res.Df & RSS & Df & Sum of Sq & F & Pr($\textgreater$F) \\
\midrule[1pt]
Model IV & 2351 & 341.76 & & & & \\
Model V & 2304 & 329.02 & 47 & 12.736 & 1.8976 & 0.0002* \\
Model VI & 2261 & 325.85 & 90 & 15.908 & 1.2264 & 0.0761. \\
Model VII & 2299 & 330.92 & 52 & 10.841 & 1.4485 & 0.0204*\\
\bottomrule[1pt]
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{The regression results relevant to $SURGEON$, $SURGERYTYPE$ and their interaction obtained by using Lasso for Model V}
\label{tab:lassoModelV}
\begin{tabular}{lrrrrl}
\toprule[1.5pt]
\multicolumn{1}{c}{Variable} & \multicolumn{1}{c}{Coefficient} & \multicolumn{1}{c}{Std. Error} & \multicolumn{1}{c}{t value} & \multicolumn{1}{c}{Pr($>|$t$|$)} & \multicolumn{1}{c}{} \\
\midrule[1pt]
Surgeon B & -0.1153 & 0.0231 & -4.986 & 6.62e-07 & *** \\
Surgeon C & 0.0454 & 0.0227 & 2.003 & 0.0453 & * \\
Surgeon F & -0.2138 & 0.0501 & -4.271 & 2.03e-05 & *** \\
SurgeryType 1 & 0.1749 & 0.0402 & 4.349 & 1.43e-05 & *** \\
SurgeryType 2 & -0.3559 & 0.0786 & -4.526 & 6.31e-06 & *** \\
SurgeryType 5 & 0.4288 & 0.0784 & 5.468 & 5.02e-08 & *** \\
SurgeryType 6 & 0.1027 & 0.0257 & 4.004 & 6.43e-05 & *** \\
SurgeryType 8 & -0.0841 & 0.0508 & -1.655 & 0.0980 & . \\
SurgeryType 9 & 0.8218 & 0.0703 & 11.696 & \textless2.00e-16 & *** \\
SurgeryType 10 & -0.1654 & 0.0337 & -4.914 & 9.51e-07 & *** \\
SurgeryType 11 & -0.1161 & 0.0308 & -3.769 & 0.0002 & *** \\
Surgeon B : SurgeryType 2 & 0.2508 & 0.1221 & 2.054 & 0.0401 & * \\
Surgeon B : SurgeryType 3 & 0.0895 & 0.0567 & 1.577 & 0.1149 & \\
Surgeon B : SurgeryType 7 & -0.4774 & 0.1709 & -2.793 & 0.0053 & ** \\
Surgeon B : SurgeryType 11 & 0.1594 & 0.0685 & 2.329 & 0.0199 & * \\
Surgeon C : SurgeryType 11 & 0.3472 & 0.2290 & 1.517 & 0.1295 & \\
Surgeon C : SurgeryType 9 & 0.1566 & 0.0608 & 2.575 & 0.0101 & * \\
Surgeon D : SurgeryType 5 & -0.4320 & 0.2309 & -1.871 & 0.0615 & . \\
Surgeon D : SurgeryType 9 & -0.2881 & 0.1172 & -2.459 & 0.0140 & * \\
Surgeon D : SurgeryType 11 & 0.1500 & 0.0667 & 2.249 & 0.0246 & * \\
Surgeon E : SurgeryType 1 & -0.4140 & 0.1934 & -2.14 & 0.0325 & * \\
Surgeon E : SurgeryType 2 & -0.3372 & 0.1232 & -2.737 & 0.0062 & ** \\
Surgeon E : SurgeryType 7 & 0.4288 & 0.1265 & 3.389 & 0.0007 & *** \\
Surgeon F : SurgeryType 2 & -0.3618 & 0.2367 & -1.529 & 0.1265 & \\
Surgeon F : SurgeryType 3 & -0.2657 & 0.1424 & -1.866 & 0.0622 & . \\
\midrule[1pt]
log likelihood value & -1040 & & & & \\
Adjusted $R^2$ value & 0.2276 & & & &\\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
Given these results, it is meaningful to investigate the effects of the three interactions. However, there are too many variables in the models with interaction terms. We perform variable selection by applying the LASSO method.
In table \ref{tab:lassoModelV}, we report the coefficients relevant to the two predictor variables in Model V, \textit{i.e.}, $SURGEON$, $SURGERYTYPE$ and their interaction. The coefficients of other predictors (\textit{e.g.}, $WEEKDAY$ and $ORDER$) are not listed in the table, since they are quite similar to those of Model IV. Recall that Surgeon A is the baseline for $SURGEON$. Generally, the mean duration of surgeries performed by Surgeon B is lower than that of Surgeon A. Surgeon B needs more time to perform Surgery 2 and 11, but less time for Surgery 7. The mean of the duration of Surgeon C's surgeries is more than that of Surgeon A. Surgeon C is slower when he/she is performing Surgery 9. The mean durations of Surgeon D and E's surgeries are not significantly higher or lower than that of Surgeon A. However, Surgeon D spends less time on Surgery 5 and 9, but more time on Surgery 11. Surgeon E needs more time to perform Surgery 7, but less time to perform Surgery 1 and 2. Surgeon F is generally much faster than Surgeon A, especially when he/she is performing Surgery 3. Alternatively, from a surgery perspective, the reference is Surgery 4. It is illustrated that the mean of the duration of Surgery 1, 5, 6, and 9 is larger than that of Surgery 4, while that of Surgery 2, 8, 10, and 11 is smaller. In addition, Surgery 2 seems hard for Surgeon B, requiring more time, but easy for Surgeon E and F, and Surgery 11 seems hard for Surgeon B, C, and D.
\begin{table}[h]
\centering
\caption{The regression results relevant to $SURGEON$, $ORDER$ and their interaction obtained by using Lasso for Model VI}
\label{tab:lassoModelVI}
\begin{tabular}{lrrrrl}
\toprule[1.5pt]
\multicolumn{1}{c}{Variable} & \multicolumn{1}{c}{Coefficient} & \multicolumn{1}{c}{Std. Error} & \multicolumn{1}{c}{t value} & \multicolumn{1}{c}{Pr($>|$t$|$)} & \multicolumn{1}{c}{} \\
\midrule[1pt]
Order 3$\sim$1 & -0.1547 & 0.0319 & -4.848 & 1.33e-06 & *** \\
Order 3$\sim$3 & -0.0847 & 0.0311 & -2.722 & 0.0065 & ** \\
Order 4$\sim$1 & -0.1649 & 0.0528 & -3.123 & 0.0018 & ** \\
Order 4$\sim$2 & -0.1018 & 0.0555 & -1.833 & 0.0669 & . \\
Order 4$\sim$4 & -0.1002 & 0.0507 & -1.975 & 0.0484 & * \\
Order 5$\sim$1 & -0.4219 & 0.1696 & -2.487 & 0.0129 & * \\
Surgeon C & 0.0743 & 0.0222 & 3.349 & 0.0008 & *** \\
Surgeon F & -0.2692 & 0.0456 & -5.908 & 3.96e-09 & *** \\
Surgeon B : Order 2$\sim$1 & -0.1406 & 0.0393 & -3.582 & 0.0003 & *** \\
Surgeon B : Order 2$\sim$2 & -0.1410 & 0.0406 & -3.471 & 0.0005 & *** \\
Surgeon B : Order 3$\sim$2 & -0.2284 & 0.0701 & -3.259 & 0.0011 & ** \\
Surgeon B : Order 3$\sim$3 & -0.1713 & 0.0752 & -2.280 & 0.0227 & * \\
Surgeon B : Order 4$\sim$3 & -0.4338 & 0.1063 & -4.082 & 4.62e-05 & *** \\
Surgeon B : Order 5$\sim$5 & -0.3933 & 0.2673 & -1.471 & 0.1414 & \\
Surgeon C : Order 3$\sim$1 & 0.1501 & 0.0713 & 2.104 & 0.0355 & * \\
Surgeon C : Order 3$\sim$2 & -0.2581 & 0.0687 & -3.758 & 0.0002 & *** \\
Surgeon C : Order 5$\sim$4 & -0.7556 & 0.2687 & -2.812 & 0.0050 & ** \\
Surgeon D : Order 5$\sim$1 & 0.9796 & 0.4132 & 2.371 & 0.0178 & * \\
Surgeon E : Order 3$\sim$2 & -0.3612 & 0.2186 & -1.652 & 0.0987 & . \\
Surgeon E : Order 3$\sim$3 & -0.4276 & 0.1726 & -2.478 & 0.0133 & * \\
\midrule[1pt]
log likelihood value & -1055 & & & & \\
Adjusted $R^2$ value & 0.2215 & & & & \\
\bottomrule[1.5pt]
\end{tabular}
\end{table}
In Table \ref{tab:lassoModelVI}, we report the coefficients of $SURGERON$, $ORDER$, and their interaction in Model VI. It is shown that Surgeon B and C are flexible. That is, the duration of these two surgeons' surgeries is significantly influenced by the surgery positions. In particular, Surgeon B (C) work much faster for the surgeries in position $4\sim 3$ ($5\sim 4$).
Surgeon C (D) performs surgeries in position $3\sim 1$ ($5\sim 1$) slowly. Surgeon E needs more time for the second and third surgeries when he/she has three surgeries in a day. In addition, the interaction between Surgeon F and surgery position is not significant.
\begin{table}[h]
\centering
\caption{The regression results relevant to $ORDER$, $SURGERYTYPE$ and their interaction obtained by using Lasso for Model VII}
\label{tab:lassoModelVII}
\begin{tabular}{lrrrrl}
\toprule[1.5pt]
\multicolumn{1}{c}{Variable} & \multicolumn{1}{c}{Coefficient} & \multicolumn{1}{c}{Std. Error} & \multicolumn{1}{c}{t value} & \multicolumn{1}{c}{Pr($>|$t$|$)} & \\
\midrule[1pt]
Order 3$\sim$1 & -0.1245 & 0.0330 & -3.776 & 0.0002 & *** \\
Order 3$\sim$2 & -0.0759 & 0.0281 & -2.700 & 0.0070 & ** \\
Order 4$\sim$1 & -0.0770 & 0.0548 & -1.406 & 0.1599 & \\
Order 4$\sim$4 & -0.0903 & 0.0509 & -1.773 & 0.0763 & . \\
SurgeryType 1 & 0.1651 & 0.0392 & 4.214 & 2.60e-05 & *** \\
SurgeryType 2 & -0.3132 & 0.0521 & -6.008 & 2.17e-09 & *** \\
SurgeryType 5 & 0.4212 & 0.0753 & 5.591 & 2.51e-08 & *** \\
SurgeryType 6 & 0.1262 & 0.0254 & 4.961 & 7.51e-07 & *** \\
SurgeryType 9 & 0.7586 & 0.0558 & 13.594 & \textless2.00e-16 & *** \\
SurgeryType 10 & -0.1614 & 0.0335 & -4.823 & 1.51e-06 & *** \\
SurgeryType 2 : Order 3$\sim$3 & -0.6483 & 0.1960 & -3.308 & 0.0010 & *** \\
SurgeryType 3 : Order 3$\sim$1 & 0.1381 & 0.0646 & 2.137 & 0.0327 & * \\
SurgeryType 3 : Order 5$\sim$4 & -0.6374 & 0.2677 & -2.381 & 0.0174 & * \\
SurgeryType 5 : Order 3$\sim$2 & -0.8241 & 0.3849 & -2.141 & 0.0324 & * \\
SurgeryType 6 : Order 3$\sim$1 & -0.2692 & 0.1319 & -2.041 & 0.0414 & * \\
SurgeryType 6 : Order 4$\sim$1 & -0.4893 & 0.2252 & -2.173 & 0.0299 & * \\
SurgeryType 7 : Order 3$\sim$1 & 0.4146 & 0.2682 & 1.546 & 0.1223 & \\
SurgeryType 7 : Order 3$\sim$3 & -0.5595 & 0.2664 & -2.100 & 0.0358 & * \\
SurgeryType 7 : Order 4$\sim$4 & 0.5464 & 0.3797 & 1.439 & 0.1503 & \\
SurgeryType 8 : Order 3$\sim$3 & -0.4080 & 0.1887 & -2.162 & 0.0307 & * \\
SurgeryType 9 : Order 4$\sim$4 & 0.8681 & 0.3842 & 2.260 & 0.0239 & * \\
SurgeryType 10 : Order 4$\sim$1 & -0.4181 & 0.2741 & -1.526 & 0.1272 & \\
SurgeryType 11 : Order 3$\sim$3 & -0.3565 & 0.0761 & -4.685 & 2.95e-06 & *** \\
SurgeryType 11 : Order 5$\sim$1 & -1.4652 & 0.3764 & -3.892 & 0.0001 & *** \\
\midrule[1pt]
log likelihood value & -1050 & & & & \\
Adjusted $R^2$ value & 0.224 & & & & \\
\midrule[1.5pt]
\end{tabular}
\end{table}
We report the coefficients relevant to $SURGERTYPE$, $ORDER$, and their interaction in Table \ref{tab:lassoModelVII}. As we know, surgery duration greatly depends on surgery type. However, this table shows that the effects of surgery type are different for surgeries with different surgery positions.
For example, the duration of Surgery 2 decreases if it is the last surgery in a day with three surgeries performed by a surgeon that day. A similar interpretation can be done for all the interaction items. It is worth noting that Surgery 3 is flexible, since its duration increases if it is placed in the position $3\sim 1$, while it decreases a lot if it is in the position $5\sim 4$.
In sum, the results of the models with interactions support Hypothesis 5, \textit{i.e.}, the interaction between some factors influence surgery duration.
The log likelihood and adjusted $R^2$ values of the models with interactions are larger than those of simple models, which indicates the necessity to consider these interactions.
We find that surgeons exhibit different performance on different surgery types, as well as on surgeries in different positions. Also, the interaction between surgery type and surgery position impacts the duration.
\section{Conclusions and Future Work}
In this paper, we investigate the impact of clinical and
non-clinical factors on surgery duration. It is found that surgery duration is influenced by surgeons' workload and workload in the OR where the surgery is performed. The duration decreases with surgeons' workload. However, it does not monotonically decrease with the workload in the OR. It decreases with the OR workload if the number of surgeries in the OR in a day is not more than four, while it increases with the OR workload if the number is beyond four. Also, we find the duration of a surgery is slightly impacted by the position of the surgery in a sequence of surgeries a surgeon performs in a day. In addition, different surgeons exhibit different patterns on the impacts of surgery types and surgery positions. More specifically, some surgeons perform certain surgery types faster than others. Surgeon B and C are flexible, whose performance is largely dependent on surgery positions. Furthermore, it is found that the effect of interactions between surgery types and positions is significant. In other words, for some kinds of surgeries, the duration is different if it is placed in different positions. These results all indicate reason to incorporate non-clinical factors in OR planning and scheduling. More accurate schedules may be attained, with fewer delays and less re-scheduling as a result.
Note that the surgeon performance of surgery duration is not equivalent to the quality of surgery from a medical perspective. Hence, an intuitive idea on future work is to investigate the impacts of the factors discussed in this paper on healthcare quality, \textit{e.g.}, readmission and mortality rates. Also, the impart of surgery duration on healthcare quality should be take into consideration. Furthermore, as mentioned before, we only focus on the third part of the whole surgery process, \textit{i.e.}, surgery duration. Hence, another future work could be to investigate the impacts of the factors mentioned in this paper on the other time intervals, as well as the impacts of the length of the intervals on healthcare quality.
\section*{Acknowledgments}
This work was supported in part by Research Grants Council (RGC) Theme-Based Research Scheme under Grant T32-102/14-N, and in part by the National Natural Science Foundation of China (NSFC) under Grant 71701132 and Grant 61603321.
\bibliographystyle{ormsv080}
|
\section{\label{sec:introduction}Introduction}
The Lattice Boltzmann Method (LBM) is a mesoscale discrete velocity model that has become an increasingly popular method for simulating fluid flows, particularly in complex geometries like porous flow (see \cite{Liu2016} for a recent review). It employs a carefully coordinated discretization of physical space, velocity space and time, to track the evolution of a vector valued mass expectation distribution, ${\bf f}$ (whose vector components are sometimes referred to as ``populations.") This evolution is carried out in a cycle of ``streaming'' and ``collision'' steps (see \S \ref{sec:LBM}).
Historically, the LBM evolved \cite{McNamara1988,Higuera1989,Higuera1989a,Chen1991,Benzi1992,Chen1992,Qian1992} from the Lattice Gas Automata (LGA \cite{Rivet2005,Chopard2005}) to address undesirable features of the LGA such as statistical noise. However, a second interpretation of the LBM is that it is a finite difference form of the continuous Boltzmann Equation \cite{He1997,He1997a,He1997c,Banda2006}. The finite difference view of the LBM allowed researchers to explore various aspects of the method by performing the discretization using various quadratures and lattices.
Despite this progress it became clear that the LBM can suffer from numerical instabilities \cite{Sterling1996} and so increasing the stability of the LBM became the focus of significant effort. Indeed, a growing number of researchers felt that the lack of unconditional stability (particularly with thermal Lattice Boltzmann Methods) was due to the lack of a so-called ``H-Theorem" for the LBM. Such an H-Theorem would draw inspiration from Boltzmann's H-Theorem for a classical gas \cite{BOLTZMANN2003}.
The first attempts to equip the LBM with an H-Theorem was to
retain the Single Relaxation Time (SRT-LBM) collision step \Ref{bgkcollision} and replace the Maxwell-Boltzmann \Ref{cnts-MB} based polynomial equilibrium \cite{Qian1992} with an equilibrium that minimizes some entropy function \cite{karlin1998,karlin1998a,renda1998,Frapolli2015,Frapolli2016,Zadehgol2014,Zadehgol2016}. Karlin et al \cite{Karlin1999} proposed that any such entropy function should be convex and, by employing its corresponding minimizer as the equilibrium in the collision step, one should recover the Navier-Stokes equations (up to second order in the macroscopic velocity, ${\bf u}$). Such entropy functions are called ``perfect entropy functions." In addition, Succi \cite{Succi2002} further suggested that the entropy's minimizer should be realizable (bounded between 0 and 1), solvable (expressible as an explicit function of the local macroscopic properties) and lead to Galilean invariant evolution equations.
An alternative path to equipping the LBM with an H-Theorem would be to find a novel collision step entirely. The most popular of these alternative collision steps is the Entropic Lattice Boltzmann Method (ELBM) which was first described by Karlin et al. in 1999 \cite{Karlin1999,Ansumali2000,Ansumali2002a,Karlin1999,Ansumali2003a} and recently extended to thermal, compressible flows \cite{Frapolli2017EntropicFlows}. More recently a modified collision step that incorporates a local entropic stabilizer parameter, $\lambda$, into the higher order moments was proposed by Karlin, B{\"o}sch and Chikatamarla \cite{Karlin2014,Bosch2017EntropicDynamics}. This entropic stabilizer is locally specified to minimize an entropy function. These models are sometimes called ``KBC" models and have since been explored further \cite{Bosch2015,Dorschner2016,Mattila2015,Dorschner2017TransitionalMethod,Dorschner2017Fluid-StructureMethod,Dorschner2017EntropicDimensions,Flint2017AnMHD}.
In addition to attempts to equip the LBM with an H-Theorem, other entropic methods have been explored such as non-equilibrium entropy limiters \cite{Brownlee2008}, artificial dissipation \cite{Brownlee2007} and Ehrenfests' coarse-graining \cite{Gorban2006,Brownlee2006} and viscosity filters \cite{Ricot2009}.
Excellent reviews of the early work on equipping the LBM with an H-Theorem can be found in \cite{Succi2002} as well as \cite{Tosi2008} and of other stabilization methods in \cite{Brownlee2011}.
The assumption that the LBM should (or even {\it could}) incorporate entropic principles is not universally accepted. A wide range of other attempts have been made to improve the stability of the LBM. Chief among them is the Multiple Relaxation Time Lattice Boltzmann Method (MRT-LBM) \cite{DHumieres1992,DHumieres1994,Lallemand2000,DHumieres2002,McCracken2005,Du2006,Chen2010,Du2012}.
We share the view that a notion of entropy plays an important role in the stability of the LBM and that entropy violations are a cause of numerical instabilities. Moreover, we consider the fundamental quantity in LBM simulations to be the {\em continuous} mass expectation density, $f$, and we treat the discrete ${\bf f}$ of the LBM to be a particular sampling of it. This allows us to connect the evolution of ${\bf f}$ to information theory. In turn it allows us to employ entropy in the information theoretic sense rather than the thermodynamic sense, avoiding the need for a well defined temperature.
In this paper we describe a novel, third interpretation of the LBM. This new interpretation is based on the Principle of Minimum Discrimination Information (or Minimum Cross Entropy) MinxEnt. We will call this LBM based on MinxEnt, ``MinxEnt-LBM."
This paper is organized in the following way: the LBM, and some of its current variations are described in Section \ref{sec:LBM}. Section \ref{maxent} outlines the foundation of our method, MinxEnt and describes the MinxEnt-LBM framework. Section \ref{sec:numerical} gives some numerical results and Section \ref{conclusions} offers final thoughts and conclusions.
\section{\label{sec:LBM}The Lattice Boltzmann Method}
In kinetic theory \cite{Liboff2003}, the evolution of macroscopic properties often involves understanding the behaviour of a mass expectation density, $f$, (hereafter called a ``distribution"),
$$f= f({\bf x},{\bf v},t).$$
The LBM aims to understand $f$ by tracking a related sampled version of it. This section aims to explain how $f$ is sampled via discretization and how its evolution is simulated.
\subsection{\label{sec:LBMsimsec}LBM Simulation Procedure}
Consider the discretization of time by finite time step of length $\delta_t$ such that $t_{n+1}=t_{n}+\delta_t$, $n \in \mathbb{N}$. Further consider a discrete set of $D$ dimensional velocities,
${\bf V} = \{{\bf v}_1,{\bf v}_2,\dots,{\bf v}_b\ | {\bf v}_{\alpha} \in \mathbb{R}^D\},$ and a discrete set of positions organized into a regular lattice, $\Lambda$. The positions on the lattice are such that ${\bf x}_j \in \Lambda$ if and only if
${\bf x}_i = {\bf x}_j + {\bf v} \, \delta t$ for some ${\bf v} \in {\bf V}$ and ${\bf x}_j \in \Lambda$.
Having discretized velocity, space and time we define the vector-valued distribution, ${\bf f}({\bf x}_j,t_n)$ (denoted ${\bf f}$ hereafter), by,
\begin{equation} \label{continuous_to_discrete}
f_{\alpha}({\bf x}_j,t_n)= W_{\alpha} \frac{f({\bf x}_j,{\bf v}_{\alpha},t_n) }{\omega({\bf v}_{\alpha})} \qquad i \in \{1,2,\dots,b \}.
\end{equation}
where $W_{\alpha}$ and $\omega({\bf v_{\alpha}})$ are quadrature weights corresponding to the particular choice of ${\bf V}$. A particular example of $W_{\alpha}$ and $\omega({\bf v}_{\alpha})$ is given in \S \ref{KLDiv}.
Given a distribution ${\bf f}$ at time $t_n$ we approximate ${\bf f}$ at time $t_{n+1}$ with a two-step process; an instantaneous local ``collision step" followed by a ``streaming step.'' This process can be summarized by the equation,
\begin{equation}
f_{\alpha}( {\bf x}_j+ {\bf v}_{\alpha} \, \delta t, t_{n+1})= \Delta_{\alpha} [ {\bf f}({\bf x}_j,t_n)]
\qquad \forall i \in \{1, \dots, b\}
\end{equation}
for some choice of local collision rule ${\bf \Delta}$.
With knowledge of ${\bf f}$, macroscopic quantities at ${\bf x}_j$ are calculated via,
\begin{equation}
\int f({\bf x}_j,{\bf v},t_n) \phi({\bf v}) \, d{\bf v} \approx \sum_{i=1}^b f_{\alpha}({\bf x}_j,t_n) \; \phi({\bf v}_{\alpha}) \label{continuous_moment}
\end{equation}
where $\phi({\bf v})$ is some function of ${\bf v}$. For example, to calculate the local density, $\phi({\bf v}) = 1$, resulting in,
$$\rho({\bf x}_j,t_n) = \sum_{i=1}^b f_{\alpha}({\bf x}_j,t_n)$$
and local momentum, $\phi({\bf v}) = {\bf v}$, resulting in,
$$\rho({\bf x}_j,t_n){\bf u}({\bf x}_j,t_n) = \sum_{\alpha=1}^b f_{\alpha}({\bf x}_j,t_n)\; {\bf v}_{\alpha}.$$
Henceforth, to ease notation we will omit the arguments in the local density, $\rho$ and macroscopic velocity ${\bf u}$.
\subsection{Collision Rules}
\subsubsection{Single Relaxation Time, SRT-LBM}
\label{SRT_CR}
The collision step accounts for changes to the components of the
distribution arising from collisions between fluid
particles as specified by a collision rule.
The most popular collision rule is based the linearization of
the kinetic collision term of the LGA \cite{Higuera1989,Higuera1989a} and further approximation by assuming a single relaxation time $\tau$ \cite{Chen1991}:
\begin{align}
{\bf \Delta}= {\bf f}({\bf x}_j,t_n) + \frac{1}{\tau}\left( {\bf f}^{\rm eq}({\bf x}_j,t_n)
- {\bf f}({\bf x}_j,t_n) \right ).
\label{bgkcollision}
\end{align}
where $\tau$ is some predetermined relaxation time that is related to the fluid viscosity and ${\bf f}^{\rm eq}$ is appropriately chosen ``equilibrium'' distribution (see \S \ref{eqmdists}). The relationship between $\tau$ and viscosity is lattice dependent and is shown for a specific lattice in \Ref{visco-tau}. This single relaxation time approach was made more popular in \cite{Qian1992} and assumed the name Lattice BGK (LBGK) owing to its similarity to the Bhatnagar-Gross-Krook kinetic equation \cite{BGK1954}. For this reason it is common in the literature to refer to LBMs using the single relaxation time collision step as ``LBGK."
\subsubsection{Multiple Relaxation Time, MRT-LBM}
\label{MRT_CR}
Because of its single adjustable parameter, $\tau$, different fluid properties (such as viscosity and the Prandtl number) cannot be independently specified in SRT-LBM simulations. In an attempt to rectify this issue and improve numerical stability researchers returned to the more general linearized collision term of \cite{Higuera1989,Higuera1989a} which allowed for multiple relaxation times during the collision step
\cite{DHumieres1992,DHumieres1994,Lallemand2000,DHumieres2002,McCracken2005,Du2006,Chen2010,Du2012}. This is accomplished via the collision rule:
\begin{align}
{\bf \Delta} = {\bf f}({\bf x}_j,t_n)
+{\bf T}^{-1} {\bf B} {\bf T} \left({\bf f}^{\rm eq}({\bf x}_j,t_n)
- {\bf f}({\bf x}_j,t_n) \right )
\label{mrtcollision}
\end{align}
where ${\bf B}$ is a diagonal matrix of relaxation times and ${\bf T}$
is an invertible matrix transforming the vector ${\bf f}$ into a vector of ``moments'' in ``moment space".
Note that the SRT-LBM collision step \Ref{bgkcollision} can be recovered from the MRT-LBM collision by assuming ${\bf B} = \frac{1}{\tau} {\bf I}$. The class of ``Two Relaxation Time'' (TRT-LBM) collision steps was suggested by Ginzburg et al. \cite{Ginzburg2008} and is related to accuracy at boundaries \cite{Ginzbourg1994,Luo2011}. In the TRT-LBM, the diagonal entries of ${\bf B}$ can take only one of two values.
\subsubsection{SRT-LBM with Ehrenfest Steps (EF-LBM)}
\label{EF_CR}
One attempt to stabilize the LBM using entropic ideas is based on Ehrenfest coarse graining \cite{Ehrenfest1990}. The EF-LBM collision rule equips the LBM with an entropy limiter which monitors
the simulation for lattice points at which some type of local entropic criteria is violated \cite{Brownlee2006}. This approach has been shown to be successful in reducing instabilities in 1-D shock tube \cite{Brownlee2006,Brownlee2006a,Brownlee2007,Brownlee2007a,Brownlee2008,Brownlee2011} and 2D lid driven cavity flow \cite{Brownlee2011} simulations.
Given a choice for entropy, $S$, EF-LBM monitors the local nonequilibrium entropy, $\delta S$, defined as,
\begin{align}
\delta S({\bf f}) := S({\bf f}^{\rm eq})-S({\bf f}). \label{EFlimit}
\end{align}
$\delta S$ serves as an indicator of locations where the non-equilibrium entropy may be too large. If $\delta S$ is below some threshold, a regular SRT-LBM \Ref{bgkcollision} collision step is taken. Otherwise, if the threshold is exceeded, the collision step is altered at that location. For example the collision step in a common version of EF-LBM is,
\begin{subequations}
\label{EFrule}
\begin{empheq}[left={{\bf \Delta} =\empheqlbrace\,}]{align}
{\bf f}({\bf x}_j,t_n) +& \frac{1}{\tau}\left( {\bf f}^{\rm eq}({\bf x}_j,t_n)
- {\bf f}({\bf x}_j,t_n) \right ) \nonumber \\
&\mbox{if } \delta S \left ({\bf f} \right) < \mbox{threshold} \\
{\bf f}({\bf x}_j,t_n) +& \frac{1}{2\tau}\left( {\bf f}^{\rm eq}({\bf x}_j,t_n)
- {\bf f}({\bf x}_j,t_n) \right ) \nonumber \\
&\mbox{otherwise.} \label{gentle-ef-rule}
\end{empheq}
\end{subequations}
We can see from \Ref{EFrule} that if $\delta S$ is above the threshold, the approach of EF-LBM is to locally modify the effective relaxation time and reduce the change in ${\bf f}$ that occurs during the collision step. This effectively makes the collision step more ``gentle." However, the effect of modifying the effective relaxation time is to locally modify the viscosity.
\subsection{Equilibrium Distributions}
\label{eqmdists}
${\bf f}^{\rm eq}$ is a sampled version of a continuous equilibrium distribution which is to be chosen by the user. For systems involving classical fluid particles, the most widely used equilibrium distribution \cite{Qian1992} is based on the continuous Maxwell-Boltzmann
Distribution:
\begin{align}
f^{\rm MB}({\bf x},{\bf v},t)=\frac{\rho}{(2 \pi R {T})^{D/2}} e^{-\frac{|{\bf v}-{\bf u}|^2}{2R{T}}}.
\label{cnts-MB}
\end{align}
Alternatively, as discussed in the introduction, one path to equipping an LBM with an H-Theorem is to abandon this approach and instead choose an equilibrium that maximizes an entropy function subject to some physical constraints. Explicit examples of these two cases are given in \S \ref{simsetup}
\section{\label{maxent}The MinxEnt Collision Rule}
\subsection{\label{sec:minxent}Principle of Minimum Cross Entropy (MinxEnt)}
Having given a brief description of the general LBM in \S \ref{sec:LBM}, we now turn to our specific contribution: the MinxEnt collision rule. The MinxEnt collision rule comes from an information theoretic approach which we now discuss.
From kinetic theory, one can show that the mass expectation distribution, $f$, is directly related to the probability, $p$, of finding a particle moving with velocity ${\bf v}$ at location ${\bf x}$ and time $t$. Using the same discretization as in \Ref{continuous_to_discrete}, one finds that,
\begin{align}
{\bf f}({\bf x}_j,t_n)=\rho({\bf x}_j,t_n) \, {\bf p}({\bf x}_j,t_n).
\label{fandpdef}
\end{align}
With this relationship in mind we take the approach that the fundamental quantity of interest should be these {\it probability} distributions. The question then becomes: during an LBM collision step, how does the pre-collision probability distribution ${\bf p}^{\rm pre}$ change to the post-collision probability distribution ${\bf p}^{\rm post}$? We propose appealing to the Principle of Maximum Entropy (MaxEnt) as described by Jaynes in his seminal paper from 1957 \cite{Jaynes1957}. By adopting this approach we will derive a new LBM collision rule.
In his 1957 paper, Jaynes argues that the probability distribution of an event represents our uncertainty of its outcome. As such we should always assign the probability that incorporates all available knowledge of the event and then maximizes our uncertainty. To do otherwise would introduce bias for which we do not have evidence to support. It is this argument that we take to be the foundation of our work on the Lattice Boltzmann Method.
To quantify our uncertainty it is natural to appeal to the Shannon Entropy. Consider an event with $n$ possible outcomes with discrete probabilities $p_1, p_2,...,p_n$. The Shannon Entropy is given by,
\begin{align}
H(p_{1},p_{2},...,p_{n})=-K\sum_{i=1}^{n}p_{i} \ln p_{i} \label{shanent}
\end{align}
where $K$ is a constant. It was proven by Khinchin in 1957 \cite{Khinchin1957} that \Ref{shanent} was the only function that is
\begin{enumerate}
\item non-negative
\item continuous and symmetric in $p_i$
\item additive for independent sources of uncertainty
\item attains its maximum value when all outcomes are equally likely.
\end{enumerate}
These are the properties that are desirable to have for a function that quantifies uncertainty.
However, even though the LBM tracks discrete distributions ${\bf f}$ (equivalently {\bf p} via \Ref{fandpdef}), we must remember that we are treating these as discrete samples of an underlying continuous functions $f$ and $p$. Thus we need a functional that generalizes the Shannon Entropy for situations where the functional acts on continuous probability distributions.
One such functional is the Kullback-Leibler
Divergence (KLD) \cite{Kullback1951,Ihara1993},
$$
H_{\rm KL} [p | q]= \int_{\mathbb{R}^D} p({\bf v}) \ln \left (\frac{p({\bf v})}{q({\bf v})} \right) \, d {\bf v},
$$
where $q$ is some predetermined reference probability distribution. Note that unlike the Shannon Entropy, the Kullback-Leibler Divergence, $H_{\rm KL}$, does not have a negative sign and so the maximization problem for $H$ becomes a minimization problem for $H_{\rm KL}$. This is called the Principle of Minimum Cross Entropy.
To incorporate our available knowledge of the system we utilize a system of constraints. Let $\{C_1,C_2,...,C_k\}$ be a set of $k$ constraint functionals specific to the physical system of interest,
$C_i: \mathcal{P} \to \mathbb{R} \qquad i=1,...,k
$
with $\mathcal{P}$ the space of all probability distributions on $\mathbb{R}^D$. Let $\{c_1,c_2,...c_k\}$ be the set of $k$ constraint values.
Generally speaking the constraint values $c_1,...c_k$ will usually depend on the pre-collision probability distribution $p^{\rm pre}$. For example, if momentum is conserved during a collision step we would have $D$ functionals, $C_i(p) = \int {\bf v}_i \, p({\bf v}) \, d{\bf v}$ and $D$ constraint values $c_i = {\bf u}_i$. That is,
$$C_i(p) = c_i \Leftrightarrow \int {\bf v}_i \, p({\bf v}) \, d{\bf v} = {\bf u}_i \Leftrightarrow \int {\bf v}_i \, f({\bf v}) \, d{\bf v} = \rho{\bf u}_i,$$
where $i \in {1,..,D}$.
Armed with $H_{KL}$ and a set of constraints, the MinxEnt collision rule
is,
\begin{align}\notag
p^{\rm post} =&\arg\!\min_{p\in \gamma} \quad H_{KL}[p|q]\\
\gamma=& \left \{ p \; | \; p \in \mathcal{P}, \;C_1(p) = c_1,...,C_k(p)=c_k \right \} \label{minxentcollision-prob}
\end{align}
\subsection{\label{sec:minxent-collision}MinxEnt-LBM}
The main premise of the MinxEnt-LBM is to apply the Principle of Minimum Cross Entropy as the local collision step within the usual LBM framework (see \S \ref{sec:LBM}). That is, in MinxEnt-LBM the only modification to standard LBM methods will be to choose the post-collision
probability distribution, ${\bf p}^{\rm post}$,
as the probability distribution that minimizes the Kullback-Leibler Divergence subject to physical constraints. The collision rule will then be constructed from ${\bf p}^{\rm post}$ via \Ref{fandpdef},
$${\bf \Delta}= \rho({\bf x}_j,t_n) \, {\bf p}^{\rm post}({\bf x}_j,t_n).$$
We can discretize the entropy using a given velocity scheme,
\begin{align}
H_{KL}[p| q] &=\int_{\mathbb{R}^D} p({\bf v}) \ln \left ( \frac{p({\bf v})}{q({\bf v})} \right ) \, d {\bf v} \nonumber\\
&\approx \sum_{\alpha=1}^{b} p_{\alpha} \ln \left (\frac{p_{\alpha}}{q_{\alpha}} \right ), \nonumber \\
&:= \mathcal{H}_{\bf q}({\bf p}) \label{discreteEntropy}
\end{align}
where $p({\bf v})$ and ${\bf p}$ are related via \Ref{continuous_to_discrete} and \Ref{fandpdef}, as are $q({\bf v})$ and ${\bf q}$.
This leads us to the general MinxEnt-LBM collision rule,
\begin{align}\notag
{\bf \Delta} &= \rho \, {\bf p}^{\rm post}, \qquad
{\bf p}^{\rm post} =\arg\!\min_{{\bf p}\in \tilde{\gamma}} \quad \mathcal{H}_{\bf q}({\bf p})\\
\tilde{\gamma}&= \left \{ {\bf p} \; | \; \widetilde{C_1}({\bf p}) = c_1,...,\widetilde{C_k}({\bf p})=c_k \right \} \label{minxentLBM-collision}
\end{align}
where the $\widetilde{C_k}$ are the discretized versions of the corresponding continuous constraint functionals, $C_k$ and ${\bf p}$ is related to $p \in \mathcal{P}$ via \Ref{continuous_to_discrete}.
\section{\label{sec:numerical}Numerical Applications}
To demonstrate the MinxEnt-LBM method, a number of numerical simulations of an athermal 2-D, isotropic Newtonian fluid were carried out. For comparison, simulations were also carried out with MRT-LBM, SRT-LBM, TRT-LBM and EF-LBM.
\subsection{Discretization Scheme}
\label{KLDiv}
In our work we use the popular D2Q9 (two dimensional, nine velocity) whose velocities and quadrature weights are described in Table \ref{d2q9table1}.
\begin{table}[h!]
\centering
\caption{D2Q9 Velocity Scheme:}
\label{d2q9table1}
\begin{tabular}{c||c|c|c|c|c|c|c|c|c}
$\alpha$ & 1 & 2 & 3 & 4& 5& 6 & 7& 8& 9\\
\hline
${\bf v}_{\alpha}$ & (1,0) & (0,1) & (-1,0) & (0,-1) & (1,1) & (-1,1) & (-1,-1)& (1,-1) & (0,0) \\
$W_{\alpha}$ & $\frac{1}{9}$ & $\frac{1}{9}$ & $\frac{1}{9}$& $\frac{1}{9}$ &$\frac{1}{36}$ & $\frac{1}{36}$ & $\frac{1}{36}$& $\frac{1}{36}$& $\frac{4}{9}$\\
\end{tabular}
\end{table}
\subsection{Choice of $q({\bf v})$}
If our system is that of a fluid composed of classical particles then it is reasonable that our choice for $q({\bf v})$ is related to the Maxwell-Boltzmann distribution, $q({\bf v})= f^{\rm MB}/\rho$.
Sampling this choice of $q({\bf v})$ according to \Ref{continuous_to_discrete} the discritized entropy \Ref{discreteEntropy} becomes,
\begin{align}
\mathcal{H}({\bf p}) := \sum_{\alpha=1}^{9} p_{\alpha} \ln \left ( \frac{p_{\alpha}}{W_{\alpha}} \right )
\label{discrete_KLD}
\end{align}
It should be noted that this form of the discretized entropy is the frequently used form of the discretized entropy used in some LBM simulations (\cite{Ansumali2006,Ansumali2003a,Ansumali2002d,Brownlee2007a,Brownlee2007,Brownlee2008,Brownlee2011} for example).
\subsection{Choice of Constraints}
Because we are considering an athermal, isotropic, Netwonian fluid our system satisfies conservation of mass, conservation of momentum, and the condition of an isotropic/Newtonian fluid. In D2Q9, conservation of mass and momentum yield the numerical constraints,
\begin{align}
\tilde{C_1}({\bf p}) := \sum_{\alpha=1}^9 p_{\alpha} &= 1 :=c_1 \label{discreteprobconst}\\
\tilde{C_4}({\bf p}) :=\sum_{\alpha=1}^9 p_{\alpha} & v_{\alpha,x} = u_x :=c_4,\\
\label{discretexmomconst}
\tilde{C_6}({\bf p}) :=\sum_{\alpha=1}^9 p_{\alpha} & v_{\alpha,y} = u_y :=c_6,
\end{align}
For an isotropic Newtonian fluid we want to ensure the local stress tensor, $\sigma_{jk}$, takes the form,
\begin{align}
\sigma_{jk}({\bf x},t)=\pi({\bf x},t) \delta_{jk} - 2 \mu \varepsilon_{jk} ({\bf x},t),
\label{newtonian-fluid}
\end{align}
where $\pi$ is the hydrostatic pressure, $\mu$ is the shear viscosity and $\varepsilon_{ij}$ is the strain rate tensor,
\begin{align}
\varepsilon_{jk}=\frac{1}{2}\left [ \frac{\partial u_j}{\partial x_k}+\frac{\partial u_k}{\partial x_j}\right].
\label{strain-rate-tensor}
\end{align}
In terms of the mass expectation distribution, this constraint becomes \cite{Liboff2003},
\begin{align}
\int f({\bf x},{\bf v},t) [{\bf v}-{\bf u}]_j & [{\bf v}-{\bf u}]_k \, d{\bf v} \nonumber\\
& = \pi({\bf x},t) \delta_{jk} - 2 \mu \varepsilon_{jk} ({\bf x},t) \label{viscoconstraint}.
\end{align}
Discretized in the D2Q9 scheme these constraints become,
\begin{align}
\tilde{C_8}({\bf p}) := \sum_{\alpha=1}^9 p_{\alpha}& \left (v_{\alpha,x}^2-v_{\alpha,y}^2 \right) \nonumber\\
&=\sum_{\alpha=1}^{9} \left [ p_{\alpha}^{\rm pre} + \frac{1}{\tau}( p_{\alpha}^{\rm eq}-p_{\alpha}^{\rm pre}) \right ] \left (v_{\alpha,x}^2-v_{\alpha,y}^2 \right):=c_8, \label{discreteviscoconst}\\
\tilde{C_9}({\bf p}) := \sum_{\alpha=1}^9 p_{\alpha}& v_{\alpha,x}v_{\alpha,y} \nonumber\\
&=\sum_{\alpha=1}^{9} \left [ p_{\alpha}^{\rm pre} + \frac{1}{\tau}( p_{\alpha}^{\rm eq}-p_{\alpha}^{\rm pre}) \right ] \left (v_{\alpha,x}v_{\alpha,y} \right):=c_9, \label{discreteisoconst}
\end{align}
One can arrive at the numerical constraints \Ref{discreteviscoconst} and \Ref{discreteisoconst} by assuming the equilibrium takes the form \Ref{numerical-plyeqm} and using the Chapman-Enskog expansion \cite{Wilson2016}. The Chapman-Enskog expansion also provides the relationship between $\tau$ and the viscosity $\nu$, as well as pressure $\pi$ and density,
\begin{align}
\nu=\frac{2 \tau -1}{6}\frac{\delta_x^2}{\delta_t}, \qquad \pi= \frac{\rho}{3} \label{visco-tau}
\end{align}
Additional, numerically motivated, constraints can be added to deal with inaccuracies at the boundary \cite{Ginzbourg1994,Luo2011},
\begin{align}
\tilde{C_5}({\bf p}) :=\sum_{\alpha=1}^9 &p_{\alpha} \left[-5+3 \left(v_{\alpha,x}^2+v_{\alpha,y}^2 \right) \right] v_{\alpha,x} \nonumber\\
&= \sum_{\alpha=1}^{9} \left [ p_{\alpha}^{\rm pre} + \frac{1}{\tau_2}( p_{\alpha}^{\rm eq}-p_{\alpha}^{\rm pre}) \right ], \nonumber \\
&\qquad \qquad \left[-5+3 \left(v_{\alpha,x}^2+v_{\alpha,y}^2 \right) \right] v_{\alpha,x}:=c_5 , \label{extraconstraint1}\\
\tilde{C_7}({\bf p}) :=\sum_{\alpha=1}^9 &p_{\alpha} \left[-5+3 \left(v_{\alpha,x}^2+v_{\alpha,y}^2 \right) \right] v_{\alpha,y} \nonumber\\
&= \sum_{\alpha=1}^{9} \left [ p_{\alpha}^{\rm pre} + \frac{1}{\tau_2}( p_{\alpha}^{\rm eq}-p_{\alpha}^{\rm pre} ) \right ] \nonumber\\
&\qquad \qquad \left[-5+3 \left(v_{\alpha,x}^2+v_{\alpha,y}^2 \right) \right] v_{\alpha,y}:=c_7, \label{extraconstraint2}
\end{align}
where,
\begin{align}
\tau_2 &= \frac{8 \tau - 1}{2 \tau -1} \label{tau2}.
\end{align}
This is in the same spirit as TRT-LBM \cite{Ginzburg2005,Ginzburg2008,Ginzburg2005a}. These constraints are purely numerical to increase accuracy of the simulation near the boundaries.
To summarize, in the D2Q9 scheme, the MinxEnt-LBM collision for an athermal 2D isotropic Newtonian fluid becomes the constrained optimization problem:
\begin{align}\notag
{\bf p}^{\rm post} =\arg\!\min_{{\bf p}} \quad \sum_{\alpha=1}^{9} p_{\alpha} \ln \left ( \frac{p_{\alpha}}{W_{\alpha}} \right )
\end{align}
subject to the constraints \Ref{discreteprobconst}, \Ref{discretexmomconst}, \Ref{discreteviscoconst}, \Ref{discreteisoconst}. We call this version of MinxEnt-LBM with 5 constraints ``MinxEnt4,'' owing to the 4 free parameters remaining to minimize over. We call the version of MinxEnt-LBM with the 7 constraints, \Ref{discreteprobconst}, \Ref{discretexmomconst}, \Ref{discreteviscoconst}, \Ref{discreteisoconst}, \Ref{extraconstraint1} and \Ref{extraconstraint2} ``MinxEnt2".
\subsection{Minimization Procedure: MinxEnt-LBM Using Newton-Raphson in Moment Space}
\label{minxent-NR}
In this work we will choose to minimize the entropy using the Newton-Raphson minimization procedure. Before proceeding however we find it convenient to turn the constrained minimization into an unconstrained minimization by moving into moment space using the invertible matrix ${\bf T}$: ${\bf M}:={\bf T}{\bf p}$. Given the form of the constraints and velocities from Table \ref{d2q9table1} it is convenient to define some rows of ${\bf T}$ to correspond to the discrete constraints and thus we choose,
\begin{align}
{\bf T}= \left[ \begin{array}{ccccccccc}
1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 \\
-4 & -1 & -1 & -1 & -1 & 2 & 2 & 2 & 2 \\
4 & -2 & -2 & -2 & -2 & 1 & 1 & 1 & 1 \\
0 & 1 & 0 & -1 & 0 & 1 & -1 & -1 & 1 \\
0 & -2 & 0 & 2 & 0 & 1 & -1 & -1 & 1 \\
0 & 0 & 1 & 0 & -1 & 1 & 1 & -1 & -1 \\
0 & 0 & -2 & 0 & 2 & 1 & 1 & -1 & -1 \\
0 & 1 & -1 & 1 & -1 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 1 & -1 & 1 & -1
\end{array} \right], \label{Tmatrix}
\end{align}
(The ordering of these moments is taken to be consistent with the literature \cite{Lallemand2000}).
Choosing ${\bf T}$ with these properties renders the constraints into a simpler form,
\begin{align}
M_1 &=1 , \label{momentconstraint1}\\
M_4 &=u_x,\label{momentconstraint4}\\
M_5 &= M_5^{\rm pre} + \frac{1}{\tau_2}( M_5^{\rm eq}-M_5^{\rm pre} ),\label{momentconstraint5}\\
M_6 &= u_y,\label{momentconstraint6} \\
M_7 &= M_7^{\rm pre} + \frac{1}{\tau_2}( M_7^{\rm eq}-M_7^{\rm pre} ) \label{momentconstraint7}\\
M_8 &= M_8^{\rm pre} + \frac{1}{\tau}( M_8^{\rm eq}-M_8^{\rm pre} )\label{momentconstraint8}\\
M_9 &= M_9^{\rm pre}+ \frac{1}{\tau}( M_9^{\rm eq}-M_9^{\rm pre}). \label{momentconstraint9}
\end{align}
The second and third rows of ${\bf T}$ remain to be chosen. In principle they are arbitrary, provided that${\bf T}$ is invertible. To be consistent with the literature We will use \Ref{Tmatrix}.
Defining the vector of free, unconstrained moments by ${\bf m}$,
the {\it full} vector of moments (including constraints) as
$${\bf M}=\begin{cases} \langle 1,m_1,m_2, u_x,m_3, u_y,m_4, c_8, c_9 \rangle \qquad &\mbox{MinxEnt4},\\ \langle 1,m_1,m_2, u_x, c_5, u_y, c_7, c_8, c_9 \rangle \qquad &\mbox{MinxEnt2}.\end{cases}$$
To fulfill the MinxEnt-LBM collision step we now seek the vector ${\bf m}$ that minimizes the discretized entropy, \Ref{discrete_KLD}, rewritten in moment space:
\begin{align}
\mathcal{S}({\bf M}):&=\mathcal{H}({\bf T}^{-1}{\bf M})\nonumber \\
&=\sum_{\alpha=1}^{9} ( {\bf T}^{-1} {\bf M} )_{\alpha} \, \ln \left ( \frac{ ( {\bf T}^{-1} {\bf M} )_{\alpha}}{W_{\alpha}} \right ). \label{KL-moment}
\end{align}
Depending on the version of MinxEnt-LBM either five or seven components of ${\bf M}$ are fixed by the constraints, and so the constrained minimization problem involves the gradient of $\mathcal{S}$ with respect to the four or two unconstrained moments respectively,
\begin{align*}
\nabla_k \mathcal{S}({\bf m}) &= \frac{ \partial \mathcal{S}}{\partial m_k}({\bf M})\\
&= \sum_{\alpha=1}^{9} ({\bf T}^{-1})_{\alpha k} \left [ \ln \left ( \frac{ ( {\bf T}^{-1}{\bf M})_{\alpha}}{W_{\alpha}}\right ) +1 \right ]
\\
& \qquad \qquad k \in \{1,...,4\} \, \mbox{ or }\{1,2\}.
\end{align*}
The Hessian is
\begin{align*}
{\bf H}_{jk} ({\bf m})&= \frac{ \partial^2 \mathcal{S}}{\partial m_j \partial m_k}({\bf M})\\
&= \sum_{\alpha=1}^{9} \frac{({\bf T}^{-1})_{\alpha k} ({\bf T}^{-1})_{\alpha j}}{({\bf T}^{-1} {\bf M} )_{\alpha}} \\
& \qquad \qquad j,k \in \{1,...,4\} \, \mbox{ or }\{1,2\}.
\end{align*}
We then perform the Newton-Raphson procedure,
\begin{align*}
{\bf m}^{n+1}={\bf m}^0-{\bf H}^{-1}({\bf m}^n)\, \nabla \mathcal{S}({\bf m}^n)
\end{align*}
For the initial moments in the Newton-Raphson procedure, ${\bf m}^0$ we compute the moments of the discretized equilibrium distribution, ${\bf M}^{0}={\bf T} {\bf p}^{\rm eq}$
and then take,
$${\bf m}^0=\begin{cases} \langle M^0_{2},M^0_{3},M^0_{5},M^0_{7} \rangle \qquad &\mbox{MinxEnt4},\\
\langle M^0_{2},M^0_{3} \rangle \qquad &\mbox{MinxEnt2}.\end{cases}$$
In principle the Newton-Raphson procedure should be continued until some convergence criteria is satisfied, however to reduce computational overhead we terminate after a single step.
This MinxEnt-LBM collision step can be summarized by the following algorithm,
\begin{enumerate}
\item Calculate pre-collision moments ${\bf M}^{0} ={\bf T} {\bf p}^{\rm eq}$ and constraints
\item Calculate ${\bf H}$ and gradient vector $\nabla \mathcal{S}$
\item Perform a single Newton-Raphson step for the unconstrained moments,
$${\bf m}^{1}={\bf m}^0-{\bf H}^{-1}({\bf m}^0)\, \nabla \mathcal{S}({\bf m}^0)$$
\item Construct the full post-collision moment vector, ${\bf M}^{\rm post}$
\item Return to distribution space ${\bf p}^{\rm post} = {\bf T}^{-1} {\bf M}^{\rm post}$
\end{enumerate}
\subsection{General Simulation Setup}
\label{simsetup}
In D2Q9 the Maxwell-Boltzmann based polynomial equilibrium is given by,
\begin{align}
f_{\alpha}^{\rm eq} &=W_{\alpha} \rho \left \{ 1 + 3{\bf v}_{\alpha}\cdot {\bf u}+ \frac92({\bf v}_{\alpha}\cdot {\bf u})^2 - \frac32 \left |{\bf u}\right|^2 \right \}.\label{numerical-plyeqm}
\end{align}
All LBMs simulated here will use this equilibrium with the exception of EF-LBM.
EF-LBM simulations adopt the entropy minimization approach to the choice of equilibrium, taking ${\bf f}^{\rm eq}$ to minimize the discrete entropy function \Ref{discrete_KLD} subject to physical constraints \Ref{discreteprobconst},\Ref{discretexmomconst},
arriving at \cite{Ansumali2003a},
\begin{align}
f_{\alpha}^{\rm eq} &=W_{\alpha} \rho \prod_{j=1}^{2} \left (2 - \sqrt{1 + 3u_j^2} \right ) \left ( \frac{2u_j + \sqrt{1+3u_j^2}}{1-u_j}\right )^{v_{\alpha,j}}. \label{numerical-enteqm}
\end{align}
The MRT-LBM, TRT-LBM, SRT-LBM and EF-LBM collisions are all based on the same rule \Ref{mrtcollision}.
The only difference between the methods is the choice of the matrix ${\bf B}$. The different versions of ${\bf B}$ are shown in Table \ref{table:relaxation-times}. We will utilize the same matrix ${\bf T}$, \Ref{Tmatrix}, for all simulations including MinxEnt-LBM.
\begin{table}[h!]
\begin{center}
\begin{tabular}{|c|c|}
\hline
Collision & ${\bf B}$\\
\hline
MRT-LBM &
$ diag\left (0,1.64,1.54,0,\frac{1}{\tau_2},0,\frac{1}{\tau_2},\frac{1}{\tau},\frac{1}{\tau} \right )$\\
\hline
TRT-LBM & $diag\left (0,\frac{1}{\tau},\frac{1}{\tau},0,\frac{1}{\tau_2},0,\frac{1}{\tau_2},\frac{1}{\tau},\frac{1}{\tau} \right )$ \\
\hline
SRT-LBM &$\frac{1}{\tau} {\bf I}$\\
\hline
EF-LBM & $\begin{cases} \frac{1}{\tau} {\bf I} &\mbox{if } \delta S < \mbox{tolerance} \\
\frac{1}{2 \tau} {\bf I}& \mbox{if otherwise.} \end{cases}$\\
\hline
\end{tabular}
\end{center}
\caption{Relaxation times for the various LBM collisions. The values 1.64 and 1.54 are chosen to agree with \cite{Luo2011,Lallemand2000}. Tolerance values are given in \Ref{tolerances}. $\delta S$ is defined in \Ref{EFlimit}}
\label{table:relaxation-times}
\end{table}
We denote three different versions of EF-LBM by EF1, EF2 and EF3 according to their tolerance values,
\begin{align}
\mbox{tolerance} = \begin{cases}
\infty &\mbox{if EF1}\\ 10^{-3} &\mbox{if EF2}, \\
10^{-5} &\mbox{if EF3}. \end{cases}
\label{tolerances}
\end{align}
Note that although SRT-LBM and EF1 have the same $\frac{1}{\tau}$ timescale, they are different schemes because of they use different equilibriums; \Ref{numerical-plyeqm} and \Ref{numerical-enteqm} respectively.
All no slip boundary, zero velocity conditions are realized by using the full-way bounceback scheme. For distributions on fixed non-zero velocity boundaries, the components of ${\bf f}$ are assigned the equilibrium distribution according to the macroscopic conditions (density and velocity) required at the boundary.
The initial distributions are set to the equilibrium distribution.
\subsection{\label{shockstudies}1D Shock tube (ST)}
A benchmark simulation to test the stability of a simulation is the low viscosity 1D shock tube \cite{Sod1978ALaws}. Simulations were performed with values of $\tau$ close to $0.5$ since the zero viscosity limit occurs when $\tau=0.5$, see \Ref{visco-tau}. The initial condition is chosen to have a density shock located at the centre of the tube.
The simulation was carried out in a 2D geometry because simulations carried out in 1D use the 1DQ3 scheme. The MinxEnt-LBM method would be over constrained and minimization would not be required. To perform MinxEnt-LBM in a 1D geometry with more than three velocities would require a non-uniform spacing of lattice points \cite{Chikatamarla2006a} and destroy its lattice structure. Thus a 2D simulation with periodic boundary conditions was employed with initial data that was taken to be constant in the $y$ direction.
The remainder of the simulation setup is summarized in the first row of Table \ref{table:shocksetup}.
\begin{table}[h!]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
Sim & $(N_x,N_y)$ & IC & SC & $\tau$\\
\hline
ST& (800,4) & $\rho = 1, x \in [0,400]$ & $t_n=400$ & $0.5 + 10^{-9}$\\
\S F& & $ \rho= .5, x \in [401,800]$ & & \\
& & ${\bf u}={\bf 0}$ & & \\
\hline
LDSS& (17,17) & $\rho = 2.7, {\bf u}=0 $ & $t_n=1000$ & varied \\
\S G& &$u_{lid}$=varied & or fail & \\
\hline
LDAS& (257,257) & $\rho = 2.7, {\bf u}=0 $ & $\Delta \psi_{min} < $ & varied \\
\S H&&$u_{lid}$=0.01 or &$10^{-5}$ & \\
& & 0.1 & & \\
\hline
\end{tabular}
\caption{Setup for the simulations. Sim: Simulation Type, $N_x,N_y$: Number of lattice nodes in respective directions, IC: Initial Condition, SC: Stopping Condition}
\label{table:shocksetup}
\end{table}
\begin{figure}[t!]
\centering
\includegraphics[width=.95\columnwidth]{figures/rho-SRT-EF1.eps}
\caption{1D shock tube density profiles after 400 times steps. a) SRT-LBM, b) EF1-LBM, c) EF2-LBM, d) EF3-LBM }\label{fig:STLBGKrho}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.95\columnwidth]{figures/ux-SRT-EF1.eps}
\caption{1D shock tube velocity profiles after 400 times steps. a) SRT-LBM, b) EF1-LBM, c) EF2-LBM, d) EF3-LBM }\label{fig:STLBGKux}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.95\columnwidth]{figures/rho-MRT-Minxent4.eps}
\caption{1D shock tube density profiles after 400 times steps. a) MRT-LBM, b) MinxEnt-LBM, c) TRT-LBM, d) MinxEnt2 }\label{fig:STMRTrho}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=.95\columnwidth]{figures/ux-MRT-Minxent4.eps}
\caption{1D shock tube velocity profiles after 400 times steps. a) MRT-LBM, b) MinxEnt-LBM, c) TRT-LBM, d) MinxEnt2 }\label{fig:STMRTux}
\end{figure}
\subsubsection*{Results}
Results for the 1D shock tube are shown in Figures \ref{fig:STLBGKrho}-\ref{fig:STMRTux}. Of particular interest is the behaviour near the shock front. It is clear from the plots that SRT-LBM, EF1-LBM and TRT-LBM suffer from the worst stability, showing severe oscillation near the shock front in both density and velocity. MRT-LBM has improved stability in both density and velocity. The two MinxEnt-LBM based simulations (Figures 3b,3d,4b and 4d) rival the stability of the density and velocity of the lowest tolerance EF-LBM simulations (Figures \ref{fig:STLBGKrho}d and \ref{fig:STLBGKux}d).
\subsection{\label{stabstudies}Lid-Driven Cavity Flow: Stability Studies (LDSS)}
Another benchmark fluid simulation is 2D lid-driven cavity flow. In lid-driven cavity flow the fluid begins at rest and the lid of the cavity is given a constant velocity in the $x$ direction. The remainder of the simulation setup is summarized in second row of Table \ref{table:shocksetup}. A range of $\tau$ values is considered including values approaching zero viscosity ($\tau \to 0.5$). For each value of $\tau$, simulations were carried out with decreasing lid velocities. Simulations are considered ``stable'' the distribution populations remained finite and non-negative at every lattice node and each of the first 1000 time steps. A distribution population that has negative components frequently precedes instabilities. This is the reason we choose to label simulations with such distribution populations as ``unstable". The maximum lid velocity at which a simulation is stable is noted for each value of $\tau$.
\subsubsection*{Results}
Results of the lid-driven cavity flow stability simulations are shown in Figure \ref{fig:stabilityplot}. The best performing methods are the MRT-LBM and MinxEnt4 with similar stability, with MinxEnt4 consistently faring slightly better than MRT-LBM. The next best performers were TRT-LBM, MinxEnt2; SRT-LBM the worst. MinxEnt2 is consistently somewhat more stable than its counterpart, TRT-LBM. EF3 and EF2 simulations (not shown) were stable at all lid velocities below 1, and for all values of $1/\tau$ between $1.9$ and $2$. Velocities of 1 or larger were not considered because of the form of \Ref{numerical-enteqm}.
\begin{figure}[h!]
\centering
\includegraphics[width=\columnwidth]{figures/paper-stabilityPlot.eps}
\caption{Maximum lid velocity in lid-driven cavity flow allowing simulation to survive 1000 time steps for various viscosity relaxation times. }\label{fig:stabilityplot}
\end{figure}
\subsection{\label{lidstudies}Lid-Driven Cavity Flow: Accuracy Studies (LDAS)}
To assess accuracy, another version of lid-driven flow simulations were carried out and compared to results from commercial CFD software, COMSOL. These simulations were similar to other studies such as SRT-LBM lid-driven flow (examined by Hou et al. in \cite{Hou1995}) and also MRT-LBM lid-driven flow (by Luo et al. \cite{Luo2011}). In addition to these, Brownlee and co-workers \cite{Brownlee2011} studied lid-driven cavity flows at various Reynolds numbers and using various LBM stabilization techniques.
In each simulation, a lattice was constructed with $257^2$ nodes ($65^2$ and $129^2$ simulations were also carried out, see \cite{Wilson2016}). A velocity was imparted on the top of the cavity in the $x$ direction.
To calculate the stream function,
$$\psi = - \int u_y({\bf x},t)\, dx $$
we used Simpson's rule for where the lattice was uniformly spaced and the trapezoidal rule where it was not. We used the same formulas for vorticity and normalization of results as in \cite{Luo2011}.
To standardize the results, the stream function was normalized to the lid velocity, the pressure deviation was normalized to the square of the lid velocity, and the vorticity was normalized to the lid velocity,
$$\hat{\psi}= \frac{\psi}{|{\bf u}_{\rm lid}|}, \qquad \hat{\delta \pi} = \frac{\delta \pi}{{\bf u}_{\rm lid}^2},\qquad \hat{\omega}=\frac{\omega}{|{\bf u}_{\rm lid}|}.$$
\begin{table}[h!]
\small
\centering
\caption{Main Vortex Results, $N_x,N_y=257$. Top Row: Main Vortex, Middle Row: Lower Right Vortex, Bottom Row: Lower Left Vortex.}
\label{table:results}
\begin{tabular}{|c|ccccc|}
\hline
& $\hat{\psi}_{\rm min}$ & x & y & $\hat{\delta \pi}$ & $\hat{\omega}$ \\
\hline
Comsol & -0.11881 & 0.53137 & 0.56445 & -0.074009 & -2.0634 \\
EF3&-0.13746&0.51569&0.55664&-0.10237&-2.3762\\
Minxent4 & -0.11808 & 0.53137 & 0.56445 & -0.073515 & -2.0552 \\
\hline
Comsol & 1.7192 & 0.86471 & 0.11133 & 3.5281 & 1.0996 \\
EF3&1.785&0.86471&0.12695&4.3053&1.2018\\
Minxent4 & 1.7071 & 0.86471 & 0.11133 & 3.4503 & 1.0801 \\
\hline
Comsol & 2.2514 & 0.084314 & 0.076172 & 4.4704 & 3.5089 \\
EF3&2.9686&0.088235&0.076172&5.5588&4.2134\\
Minxent4 & 2.255 & 0.084314 & 0.076172 & 4.382 & 3.4119 \\
\hline
\end{tabular}
\end{table}
\subsubsection*{Results}
The results of the $257 \times 257$ node lid-driven cavity flows for $Re=1000$ were the most visually interesting and are the only results reported here; see Figure \ref{fig:LID1000}. Numerical results are given in Table \ref{table:results}. Presented in the table are the centres (determined by the extrema of the streamfunction) of the main, lower-left, and lower-right vortices. In addition, the pressure deviation and vorticity at these locations is presented. For simulations with $65^2$ and $129^2$ nodes as well as for smaller Reynolds numbers, see \cite{Wilson2016}. Also refer to \cite{Wilson2016} for simulations using other collision rules discussed in this manuscript.
\begin{figure}[h!]
\centering
\includegraphics[width=0.9\columnwidth]{figures/paper_figure.pdf}
\caption{Flow contours of lid-driven cavity flow for Re$=1000$ with $N=257^2$. Left: Pressure deviation, Middle: Stream Function, Right: Vorticity. Top Row: Comsol, Middle Row: SRT-LBM, Bottom Row: EF3-LBM}\label{fig:LID1000}
\end{figure}
\subsection{Discussion}
To explore the the effect on stability of different collision rules, two types of simulations were conducted. In the 1D shock tube simulations, we can see from Figures 1-4 that, although MRT-LBM shows improved stability over SRT-LBM, the only collision rules that maintained a sharp shock front were EF3-LBM, Minxent4 and Minxent2. All other collision rules experienced instability at the shock front. This seems to indicate that EF3-LBM, Minxent4 and Minxent2 are candidates for the most stable collision rule.
Similar conclusions are found in results of the lid driven stability simulations, shown in Figure 5. Though not plotted in Figure 5, simulations employing EF2-LBM and EF3-LBM collision rules remained stable for all lid velocities below 1 and for values of $1/\tau$ between $1.9$ and $2$. Simulations were not carried out for EF3-LBM and EF2-LBM at a lid velocity of 1 owing to the form of the equilibrium \Ref{numerical-enteqm}. It is unsurprising that EF2-LBM and EF3-LBM are able to survive at all tested lid velocities because when $\tau$ is close to $0.5$, the effect of the EF2-LBM and EF3-LBM collision rules \Ref{EFrule} is to return the distribution to near-equilibrium whenever $\delta S$ exceeds tolerance. Effectively, whenever a lattice node was in danger of losing stability (indicated by $\delta S$ above tolerance), the distribution was set to equilibrium. This ensured the simulation always remained stable. Other than EF2-LBM and EF3-LBM, the collision rule that survives at the highest lid velocities is Minxent4 followed by MRT, Minxent2 and TRT. The least stable collision rule was SRT-LBM. From these observations we can further conclude that the collision rules that lead to the most stable simulations are EF3-LBM and Minxent4.
We next tested the accuracy of the these two, most stable, collision rules. To accomplish this we performed lid driven cavity flow simulations for various Reynolds numbers and compared the results to simulations using commercially available software (Comsol) under the same flow conditions. From Figure \Ref{fig:LID1000} we see that EF3-LBM deviates considerably from the Comsol results. As shown, the Minxent4 simulations qualitatively reproduced the Comsol results. These conclusions are also quantitatively supported in Table \ref{table:results} where we see that results using the Minxent4 collision rule are much more similar to Comsol than the results using the EF3-LBM collision rule.
From these stability and accuracy tests we find that Minxent4 offers the best mix of stability and accuracy in the simulations discussed here.
It is worth mentioning that improved accuracy for EF2-LBM and EF3-LBM collision rules have been reported in the literature \cite{Brownlee2006,Brownlee2007}. However, in order to maintain accuracy, these studies limited the number of lattice nodes where the ``more gentle" collision rule \Ref{gentle-ef-rule} was used. That is, using the EF-LBM collision rules, there is a trade-off between accuracy and stability. This trade-off is mediated by the maximum number of lattice nodes permitted to use \Ref{gentle-ef-rule}. The more lattice nodes that are allowed to use \Ref{gentle-ef-rule}, the more stable the simulation is, but the less accurate the simulation. The fewer lattice nodes that use \Ref{gentle-ef-rule}, the more accurate the simulation is but the less stable it is. This means, for EF-LBM, the tolerance and maximum number of lattice nodes permitted to use \Ref{gentle-ef-rule}, are parameters that need to be tuned.
Likewise, MRT and TRT both show improved stability over SRT, but they are both dependent on the choice of relaxation times that are not related to the fluid viscosity. Thus, similar to EF-LBM, MRT and TRT simulations can exhibit improved stability, but require parameters to be tuned. In MRT-LBM, relaxation times associated with the second and third moments could be adjusted to increase stability. In TRT-LBM $\tau_2$ could be adjusted rather than use its prescribed value given in \Ref{tau2}.
Tuning these parameters would be need to be performed on a simulation by simulation basis. Unlike the other collision rules simulated here, This is contrasted with Minxent-LBM methods which able to improve stability, without losing accuracy and without needing to tune any parameters.
\section{\label{conclusions}Conclusions}
In this paper we have derived a novel collision step for the Lattice Boltzmann Method based on the Principle of Minimum Cross Entropy, MinxEnt-LBM.
MinxEnt-LBM was used in numerical simulations and compared to existing LBMs. The only scheme that showed comparable stability to MinxEnt-LBM was the entropy limiting scheme of the LBM based on Ehrenfest Steps (EF-LBM). However, lid driven cavity flow simulations showed that without tuning the parameters involved in EF-LBM schemes they suffered from a degradation of accuracy. We can conclude that, of the LBM schemes tested, MinxEnt-LBM had the best combination of stability and accuracy.
An important practical consideration is that EF-LBM and MRT-LBM require specification of parameters which need to be tuned and optimal values are not known {\it a priori}. This is not the case for the MinxEnt-LBM.
|
\section{Introduction}
\label{intro}
Solar corona is the outer layer of the solar atmosphere. It consists of highly ionized plasma that is structured by the magnetic field. In
the solar corona a number of important but still ill-understood phenomena take place, such as the initiation and acceleration of coronal mass
ejections (CMEs) and the acceleration of the solar wind.
Nowadays the corona is observed by space-borne telescopes and coronagraphs in various spectral ranges: from X-rays and extreme ultraviolet
(EUV) to white light. The EUV spectral range is more suitable for observations of the low corona due to extremely small brightness of plasma
in EUV at high altitudes that decreases with height proportionally to the square of the electron density $n_e^2$, contrary to white-light
brightness proportional to $n_e$. There are occasional observations up to $2 R_\sun$ with regular EUV telescopes that operated in
dedicated long-exposure regimes, e.g. TESIS/CORONAS-PHOTON \citep{Kuzin2011,Reva2014} or SWAP/PROBA-2
\citep{2006AdSpR..38.1807B,Seaton2013,2017JSWSC...7A...7D}, or up to $2.5 R_\sun$ with specialized EUV coronagraphs, e.g.
SPIRIT/CORONAS-F \citep{2003AdSpR..32..473Z,2008AnGeo..26.3007S}.
White-light coronagraphs can provide reliable observations over a significantly larger field of view due to a different mechanism of
emission, primarily Thomson scattering. For example, the field of view of the internally occulted coronagraph COR1/STEREO ranges from
$1.5$ to $4.0R_\sun$ \citep{Howard2008}. The field of view of externally occulted coronagraphs usually starts slightly higher, from $\sim
2.2R_\sun$ in LASCO C2/SOHO \citep{Brueckner1995}, and from $2.5R_\sun$ in COR2/STEREO-A \citep{Howard2008} \citep[see also comparison of
various coronagraphs in][]{Frazin2012}. The very important range of heights from $1.1R_\sun$ to $2.5R_\sun$
\citep{2008ApJ...680.1532Z} is almost not covered in white-light observations from space, except for LASCO C1/SOHO $\sim
1.1-3.0R_\sun$ \citep{Brueckner1995}, which suffered from significant stray light.
Classical internally occulted coronagraphs can not provide such observations as the theoretical limit of diffracted light around $10^{-4} B_\sun$ at
$1.1R_\sun$ exceeds the brightness of the solar corona $10^{-5}B_\sun$ \citep[see review of various coronagraph systems in][hereafter
RR17]{2017A&A...599A...2R}. Scattering of the bright light from the solar disk in the primary objective further worsens the problem.
The intensity of diffracted light in the inner corona region can be decreased by the use of apodized entrance aperture
\citep{2007A&A...467..317A}, but this solution was not yet implemented in a solar coronagraph.
For externally occulted coronagraphs, there are two fundamental factors that complicate observations at low heights: these are drastic
increase of diffracted light and vignetting of the inner field of view \citep[see e.g.][]{Koutchmy1988,JGRA:JGRA53203}. Furthermore,
straightforward ways to compensate both effects are contradictory: diffracted light decreases \citep{Lenskii1988,Fort1978}, whereas
vignetting increases with increase of the external occulter size (or putting it closer to the primary objective). In all the previous
space-borne coronagraphs, the distance from the external occulter to the primary objective was limited by the length of the instrument,
i.e. $\sim 1$~m. This fact does not allow using sufficiently large primary objectives and simultaneously results in almost 50\%
vignetting of the objective throughout the field of view.
ASPIICS (Association of Spacecraft for Polarimetric and Imaging Investigation of the Corona of the Sun) is a novel white-light coronagraph
\citep{2010SPIE.7731E..18L, 2015SPIE.9604E..0AR,Renotte2016}, that will perform regular observations of the corona over the field of view
from $\sim 1.08R_\sun$ up to $3.0 R_\sun$. This will be possible thanks to European Space Agency PROBA-3 mission that will use formation
flying (FF). Such a technology allows virtual enlarging of instrumentation to an unprecedented size: the external occulter will be placed
on one satellite and the optical instrument -- on the second. Both satellites will synchronously fly on a highly elliptical 19.38-hour
orbit and form a precise formation with inter-satellite distance $\sim 150$~m during 6 hours. The external occulter with $\diameter
1.42$~m diameter will produce $\diameter77$~mm shadow, and the telescope with the entrance aperture $\diameter 50$~mm will be placed
in the center with the maximal possible precision. The giant inter-satellite distance allows unvignetted observations of the solar corona
starting already from $\approx 1.1R_\sun$.
In order to provide highest possible accuracy of FF, the satellites are equipped with fine metrology subsystems that include laser
beam/reflector and shadow position sensor (SPS) \citep{Renotte2016,2015SPIE.9604E..0CB}. SPS consists of 8 photodiodes, placed around the entrance aperture
of the telescope in the penumbra region. SPS will measure relative position of the satellites with the accuracy better than $\sim 1$~mm in
longitudinal and transversal directions. The expected FF accuracies are: $\pm 15$~mm in longitudinal and $\pm 5$~mm in transversal $x$-,
and $y$-directions, $30$~arcmin for the attitude of the occulter satellite, and $15$~arcsec for the attitude of the satellite with the
telescope ($3\sigma$ values).
Various analyses \citep{Landini2010,2013A&A...558A.138A} show that the amount of diffracted light reaching the ASPIICS primary objective
is high enough. In order to reduce its level, the external occulter of ASPIICS will have a toroidal shape; technological limitations do not
allow using a superior external occulter like a multi-disk system or a threaded-cone \citep[see][for review of various systems]{Bout2000}.
RR17 used analytical/numerical approach and calculated not only the diffraction level just in front of the primary
objective, but, more importantly, the spatial distribution of diffracted light at the detector. Authors found that the intensity drastically
depends on the sizes of the internal occulter and the Lyot stop, and provided a ``recipe'' for choosing their proper sizes. The analysis,
however, was limited only to the symmetrical case, when the Sun, the external occulter and the coronagraph are on the same axis and are
perfectly co-aligned.
Beside the misalignments introduced by FF, there are additional sources of misalignments expected in ASPIICS. These can be long- or
short-term thermal expansions of mechanical structure, errors of initial co-alignment of the telescope on the satellite etc. In total these
misalignments can be very small, of the order of 10--30~arcsec. Nevertheless, as \citet{Venet2010} concluded, even such low values can be
critical for coronagraphs with extremely small overoccultation.
The aim of the present paper is to investigate how different misalignments influence the intensity of the diffracted light in the final image,
to compare it with the intensity of the corona and to choose the proper size of the internal occulter to ensure reliable rejection for the cases of
possible misalignments. The analysis considers only the effect of diffraction, i.e. the effect that is caused by masking of individual
parts of the wavefront of propagating light. Other effects that will unavoidably be present and degrade the performance of the coronagraph --
scattered light and ghosts, non-ideal lenses, etc., will provide an additional contribution but are not considered here. Whereas the
computations are performed for a particular geometry and optical layout (representative for the ASPIICS coronagraph), we note that the
obtained properties of the diffracted light and its behaviour with various misalignments remain valid for any externally occulted coronagraph.
The paper is structured as follows: in Sect.~\ref{sec2} we describe the optical layout and the basics of the algorithm, in Sect.~\ref{properties} we
discuss some properties of the diffracted light, in Sect.~\ref{types-sec} we consider possible types of misalignments. In
Sect.~\ref{results} we present results of computations, and consider different sizes of IO and its apodization. In Sect.~\ref{discussion-sec} we
discuss the obtained results, and give conclusions in Sect.~\ref{conclusion-sec}. Appendix~\ref{method} contains mathematical details of
the calculations, Appendix~\ref{numerical} details the numerical computations, and Appendix~\ref{sampling-sec} analyzes influence of
samplings.
\section{Optical layout, model, algorithm}
\label{sec2}
\subsection{Optical layout}
\label{layout-sec}
A detailed description of the optical layout of ASPIICS is given by \citet{2015SPIE.9604E..0BG}. In the present paper we use a simplified model
and conventions to denote optical planes described in RR17. Table~\ref{planes-table} summarizes the names and the description of the optical
planes.
\begin{table}
\caption{Key planes of the ASPIICS coronagraph}
\label{planes-table}
\begin{tabular}{l l}
\hline \hline
Plane & Description \\
\hline
$O$ & External occulter plane \\
$A$ & Entrance aperture of the telescope \\
$B$ & Focal plane of the primary objective \\
$O'$ & Conjugate image of plane $O$ (plane of the IO) \\
$C$ & Conjugate image of plane $A$ (plane of the Lyot stop) \\
$D$ & Final focal plane and the detector \\
\hline
\end{tabular}
\end{table}
The optical layout of the ASPIICS coronagraph is given in Fig.~\ref{layout}. The external occulter --
EO (plane $O$) is situated on the occulter satellite placed at the distance $z_0=144\,348$~mm ahead of the coronagraph satellite with the
optical instrument. The telescope consists of the entrance aperture (plane $A$) and the primary lens L1 with the focal length $f$, that makes
an image of the corona in the primary focal plane $B$. The light diffracted at the EO is focused in the $O'$ plane, situated at the distance
$z_1$. This is the plane of the conjugate image of the EO produced by L1, thus it is $\Delta =z_1-f \approx 0.76$~mm further than $B$. The major
part of the diffracted light is cut by the internal occulter, the size of which should be carefully chosen to obtain a good rejection of
diffraction and not to vignette too much. The IO is deposited directly on the surface of the field lens L2. The lens L2 along with L1 make
an image of the entrance aperture on the $C$ plane. The Lyot stop is placed in this plane and rejects the light diffracted at the
entrance aperture. Simultaneously, the lens L2 projects the entrance aperture to the relay lens L3, ensuring that the light propagates
further into the coronagraph and justifying its name -- the field lens. Finally, both L2 and L3 project the primary focus $B$ onto the
detector plane $D$.
\begin{figure*}
\centering
\includegraphics[width=18cm]{optical-layout.eps}
\caption{Optical layout of the ASPIICS coronagraph. Dim yellow light represents a regular coronal beam, orange
arches and lines represent the light that diffracts on the EO, the red lines represent the light diffracted at the entrance aperture.}
\label{layout}
\end{figure*}
In Fig.~\ref{layout} the dim yellow light represents a regular coronal beam, orange arches represent light that diffracts on the EO. After
propagating through the entrance aperture and L1, this light is focused (orange lines) in the $O'$ plane. The red lines represent the
light diffracted at the entrance aperture.
Our optical model contains two main simplifications: we place the entrance aperture and the primary objective into the same plane $A$ (in the real
instrument the objective is $\sim 120$~mm behind), and we change the distances $d$, $l$ and the focal distance of L3 to keep the same spatial scale
in $D$ as in $B$. Both simplifications do not influence the diffracted light.
We perform the calculation for the central wavelength 550~nm of the main (wide-band) passband and discuss possible effects in the
Sect.~\ref{wavelength}.
Table~\ref{geometry-table} summarizes main parameters of the ASPIICS optical layout and other parameters used in our computations.
We consider an infinitely thin (razor edge) external occulter. In reality the edge has a toroidal cross-section both for improvement
of the diffracted light rejection \citep{Landini2010}, and for reduction of the effect of tilting. The simplification used here is linked
to the difficulty of a correct representation of 3D occulters in the framework of the Fresnel diffraction \citep{Sirbu:16}. Currently
attempts are ongoing to take the effect of the toroidal shape of the occulter into account both, experimentally and numerically.
\begin{table}
\caption{Parameters used in our numerical study taken from the ASPIICS configuration. See text for details.}
\label{geometry-table}
\begin{tabular}{l l l}
\hline \hline
Parameter & Symb. & Value \\
\hline
Wavelength & $\lambda$ & $550$~nm \\
Angular radius of the Sun & $R_\sun$ & 16~arcmin \\
Radius of the EO & $R_{EO}$ & 710~mm \\
Distance plane $O$ -- plane $A$ & $z_0$ & $144\,348$~mm \\
Angular radius of the EO & $\omega_{EO}$ & 16.909~arcmin \\
Radius of the entrance aperture & $R_{A}$ & 25.0~mm \\
Focal length of L1 & $f$ & 330.348~mm \\
Distance plane $A$ -- plane $O'$ & $z_1$ & 331.143~mm \\
Radius of the internal occulter & $R_{IO}$ & 1.662~mm\tablefootmark{a} \\
Hole in the internal occulter\tablefootmark{b} & $r_{IO}$ & 0.489~mm \\
Lyot stop radius & $R_C$ & $0.97R_A$ (24.25~mm)\tablefootmark{c} \\
\hline
\end{tabular}
\tablefoot{
\tablefoottext{a}{the value that we use as a baseline for the diffracted light;}
\tablefoottext{b}{a hole in the internal occulter is used to produce images of LEDs from the occulter satellite on the detector;}
\tablefoottext{c}{it is the relative size of the entrance aperture and the Lyot stop that plays a role.}
}
\end{table}
\subsection{Projection of the Sun, EO and IO onto different planes}
It is interesting to compare how different objects -- the Sun, the EO and the IO are projected in the framework of the geometrical optics
onto planes $B$ and $O'$. It turns out that the geometrical approach correctly predicts the relative sizes of images, their relative shifts due to
the tilt of the coronagraph or other misalignments. Since $O$ and $O'$ are conjugate planes, we can project the IO onto $O$ and after that onto $B$.
In the figures below we show the EO and IO with a transparent central part for clarity.
We start with the case of the perfect symmetry, when the coronagraph and the EO are co-aligned and co-centered, and the center of the Sun
lies on the same optical axis $z$. In Fig.~\ref{symmetrical} we present sketches with plane $B$ in the left panel and plane $O'$ in the
right panel.
The Sun is sharply focused in the $B$ plane, and slightly defocused in the $O'$ plane (we denote defocusing by a smooth limb).
The EO is represented by the blue color in both planes. In the $O'$ plane the image of EO is perfectly focused, which is denoted by a thin
ring. In the $B$ plane the image of the EO is defocused, thus the EO is marked as a ring with a finite thickness and gradient filling.
Inner and outer edges of the ring correspond to the extremes in optical vignetting produced by the EO, and the gradient itself represents
the vignetting function (full obscuration inside and full transparency outside). Obviously, the angular size of the EO should be selected
in such a way that its inner vignetting zone is larger than the Sun.
The IO is represented by a green ring. Similarly to the EO, it is perfectly focused in the $O'$ plane and defocused in the $B$ plane. The
reasoning about the inner and the outer edges can be applied here as well. The size of the IO should be selected in such a way that it fully covers
both the Sun and the EO.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{symmetrical-planes.eps}}
\caption{Projections of the Sun, the EO and IO into planes $B$ and $O'$ in symmetrical case.}
\label{symmetrical}
\end{figure}
\subsection{Algorithm of the diffraction calculation}
\label{algorithm}
In our analysis we follow the basic idea of \citet{2013A&A...558A.138A} and RR17. Here we briefly summarize the method, and we give
the full mathematical description in Appendix~\ref{method}.
We setup a reference frame with the $xy$ plane coinciding with the $O$ plane and co-centered with the EO, and the $z$-axis pointing to the
entrance aperture (this implies the co-alignment and co-centering of the coronagraph and the EO). The Sun is considered as a spatially
extended light source that produces a set of non-coherent plane-parallel waves with wave-vectors $\vec k$:
\begin{equation}
\Psi_{\vec k} = A \exp { \left(-i (\omega t - \vec k \vec r) \right) }.
\label{plane-parallel}
\end{equation}
By varying the direction of $\vec k$ we sample the solar disk (so far it needs not to be co-aligned).
The algorithm for calculating the diffracted light in various planes consists of the following:
\begin{enumerate}
\item We select a particular point on the Sun and consider its plane-parallel wave $\Psi_{\vec k}$. The direction of the wave-vector
$\vec k$ can be specified in the $B$ plane either by Cartesian $(\alpha,\beta)$, or by polar $(\rho,\varphi)$ coordinates. The
intensity of the wave depends on the solar brightness and the sampling area: $A^2 = B_{\vec k} \mbox{d}S$.
\item After diffraction at the EO, each wave becomes significantly not a plane-parallel one, and it can not be expressed similarly to
(\ref{plane-parallel}). We express its amplitude in the $A$ plane as $\Psi_{A \vec k}$, where $\vec k$ denotes the orientation of
initial wave.
\item Further propagation of the wave is considered in the framework of the Fourier optics formalism. The fields in the $O'$,
$C$, and $D$ planes are calculated with three successive Fourier-Fresnel transforms over the distances $z_1$, $d$, $l$ and taking
into account apertures $A_A$, $A_{O'}$, and $A_C$.
\item We consider all the possible waves, summarizing over directions of $\vec k$ (i.e. over $(\rho,\varphi)$ or
$(\alpha,\beta)$), and calculate the final image in $D$ as:
\begin{equation}
I_D = \sum_{\vec k} B_{\vec k} |\Psi_{D \vec k}|^2 \, \mbox{d}S.
\end{equation}
\end{enumerate}
We stress that the $\vec k$ vector is determined in the reference frame of the EO and the coronagraph, not that of the
Sun. All the characteristics of the Sun -- the solar brightness, angular size, limb darkening -- are implicitly hidden in $B_{\vec k}$
and possible $\vec k$ directions, whereas the obscuration by the EO and IO is implicitly hidden in $\Psi$.
Calculation of $\Psi_{A \vec k}$ is a separate problem as it involves computing of the Fresnel diffraction on the EO. Again, we follow
approach of \citet{2013A&A...558A.138A} and RR17, who considered an infinitely thin (razor edge) external occulter. The details are
given in Appendix~\ref{method}, and here we note that $\Psi_{A \vec k}$ for a particular $\vec k$ can be calculated from the co-axial
wave $\Psi_{A00}$ ($\rho=0, \, \varphi=0$) by shifting and multiplicating by an arbitrary phase function. We give several examples of
$\Psi_{A \vec k}$ in Fig.~\ref{tilted-waves}, where the three panels correspond to different $\vec k$: the co-axial one with $\rho=0$
(this is essentially $\Psi_{A00}$), and two tilted ones with $\rho=30\arcsec$ and $\rho=60\arcsec$ ($\varphi=35^\circ$ in the latter two
cases). The central bright feature of the co-axial wave -- the Arago spot is displaced in tilted waves. Highly oscillating nature of the
$\Psi_{A \vec k}$ function (see Fig.~2 in RR17) is present but is not visible here in the pdf/hardcopy images. Only the part of the wave
that fits the entrance aperture produces further signal in the instrument. The full-Sun umbra/penumbra pattern in the aperture plane can
be obtained by integrating $\Psi_{A\vec k}$ over the Sun and resembles geometrical umbra/penumbra pattern with some signal in the umbra
region (see Fig.~5 in RR17).
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{tilted_psia_ann.eps}}
\caption{Intensities $|\Psi_{A \vec k}|^2$ of waves in the $A$ plane propagating with different directions of $\vec k_{\rho\varphi}$: a)
$\rho=0$, b) $\rho=30\arcsec$, and c) $\rho=60\arcsec$ ($\varphi=35^\circ$ in the latter two cases). Only the part of $\Psi_A$ that enters into
the aperture produces signal in further planes of the telescope.}
\label{tilted-waves}
\end{figure}
\section{Some properties of diffraction}
\label{properties}
Calculation of the diffracted image in the $O'$ plane $I_{O'}=\sum_{\vec k} B_{\vec k} |\Psi_{O' \vec k}|^2 \mbox{d}S$ reveals that it is a
bright ring with radius $r=z_1 \tan ( \omega_{EO})$, i.e. with the angular size \emph{exactly} coinciding with the angular size of the EO -- $\omega_{EO} =
\arctan \left( R_{EO}/z_0 \right)$.
One of the main properties of the diffraction is the following: every single wave $\Psi_{A\vec k}$ after propagation
into $O'$ highlights there either \emph{a sector} or \emph{the whole ring} (see Fig.~\ref{property} with examples of $|\Psi_{O'\vec
k}|^2$). The highlighted sector has the same polar coordinate $\varphi$ as the initial wave, and its angular width depends on
$\rho$ of the initial wave: the higher $\rho$, the smaller the width. For the co-axial wave $\rho=0$, and the whole ring is
bright (panel a). For the waves with $\rho \approx 10 \arcmin$ and $\rho \approx 16\arcmin$ the bright sector becomes very
narrow (panels b and c). Two-dimensional distributions of intensities have very intense peaks and strongly decreasing wings.
This reasoning allows inferring the second property: if the EO and the coronagraph are on the same optical axis, the relative
displacement of the Sun from this axis \emph{does not} change the geometrical symmetry of the diffraction ring in $O'$. In the case of such
a displacement, individual parts of the diffracted ring become brighter or dimmer, but they
do not move in the $O'$ plane. Similarly, seasonal change of the angular size of the Sun \emph{does not} change the geometry of the
diffracted ring in $O'$.
\begin{figure*}
\includegraphics{plane_o.eps}
\caption{Intensity of the diffracted wave $|\Psi_{O'\vec k}|^2$ in the $O'$ plane for different initial waves $\vec k_{\rho\varphi}$: a)
co-axial wave $\rho=0$, b) moderately tilted wave with $\rho=10.24\arcmin$, c) highly tilted wave with $\rho=16\arcmin$, d) integrated along the
solar radius; $\mbox{d}S = \rho \, \mbox{d} \rho \, \mbox{d}\varphi$ was taken into account during integration, thus the color table is
different; $\varphi=35^\circ$ was used in the latter three cases.}
\label{property}
\end{figure*}
The final image in $D$ is formed by diffracted light that ``propagates behind'' the IO. As long as the relative shift of the diffracted ring
and the IO is small (as expected in ASPIICS), the effect of diffraction determines the final image in $D$. Thus, the third property is:
the position of the diffraction ring on the detector, its size and shape are determined by the IO. Misalignments do not shift or distort
the final diffraction ring; they just make individual parts of the ring brighter or dimmer. It is the variations of the IO size, that can change the size of the final
diffraction ring.
\section{Types of misalignments}
\label{types-sec}
There are various types of misalignments of optical elements that influence the overall intensity and spatial distribution of the
diffracted light. Among them are transversal (in the $xy$-plane) shifts of the EO and the telescope, tilts of the EO and the telescope, shifts
and tilts of internal optical elements inside the coronagraph, change of $z_0$, etc. Influence of some of the misalignments, such as
change of $z_0$ or the axial displacement of the IO, can be analysed using the symmetrical model since they do not break the axial symmetry.
Other misalignments, such as the transversal shift of the EO and the coronagraph, or transversal shifts of apertures must be
considered using full-sampling approach, as the optical layout becomes not axi-symmetrical.
Below we will show that any configuration with transversal shifts of ``external'' optical elements can be reduced to superposition of
just two of them -- tilt of the telescope and shift of the Sun.
We are not going to consider influence of the possible tilt of the lenses (the expected tilts of the order of 0.5--5~arcmin will be
negligible) and the tilt of the EO (due to the reasons explained in Sect.~\ref{layout-sec}).
\begin{figure*}
\includegraphics{types-of-shift.eps}
\caption{Comparison of various misalignments: a) symmetrical case; b) shift of the Sun; c) shift of the EO; d) shift of the coronagraph,
and e) tilt of the coronagraph. In panels c) and d) the red axis goes through centers of the EO and the entrance aperture, and dotted
rectangles represent the coronagraph co-aligned with this new axis.}
\label{misalignments}
\end{figure*}
In the symmetrical case (panel (a) in Fig.~\ref{misalignments}) the $z$-axis goes through the center of the Sun, the EO and the coronagraph.
It is perpendicular to the EO plane and co-aligned with the coronagraph optical axis. The entrance aperture is co-centered with
the umbra/penumbra pattern.
In the case the Sun is shifted (panel (b) in Fig.~\ref{misalignments}), the EO and the coronagraph remain on the initial $z$-axis. However, the
umbra/penumbra pattern is not symmetrical anymore with respect to the entrance aperture. The projections of the Sun, EO and IO into
planes $B$ and $O'$ are shown in Fig.~\ref{shift}: the projection of the Sun is shifted, but the EO remains in the same position: the solar
shift changes the relative intensity of EO (i.e. the diffraction bright ring), but does not change its geometry or its position (this is due to the property discussed in
Sect.~\ref{properties}). The coronagraph and its mathematical model remain axi-symmetrical, thus the waves propagating from different
directions $\vec k_{\rho\varphi_1}$ and $\vec k_{\rho\varphi_2}$ produce \emph{similar} responses in $D$, which are just rotated with
respect to each other by the angle $\varphi_2-\varphi_1$. From the computational point of view the misalignment changes coordinates,
e.g. $(\alpha,\beta) \rightarrow (\alpha+\phi,\beta)$, of all the waves $\Psi_{\vec k}$ that represent the Sun.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{shifted-planes.eps}}
\caption{Projections of the Sun, EO and IO into planes $B$ and $O'$ in case of a shifted Sun (cf. Fig.~\ref{symmetrical}).}
\label{shift}
\end{figure}
The cases where the EO is shifted (panel c) or the coronagraph is shifted (panel d) are similar to each other. The umbra/penumbra pattern in
the aperture plane becomes not symmetrical. We define a new axis (red dash-dotted line in Fig.~\ref{misalignments}) through the centers
of the EO and the coronagraph. The new axis is tilted by the angle $\phi = \sqrt{\Delta x^2+\Delta y^2}/z_0 \le 10$~arcsec with respect to
the original axis (black dash-dotted line) as the maximal expected shift is $\sqrt{\Delta x^2+\Delta y^2} \sim 7.0$~mm. In the new reference
frame we can neglect the tilt of the EO and take into account the tilt of the coronagraph and shift of the Sun (the co-aligned coronagraph is
denoted by the red dashed line). Thus, these two cases can be considered as a simultaneous shift of the Sun and tilt of the coronagraph,
with the two effects enhancing each other.
In the case of the tilt of the coronagraph (panel e), the main effect is that in the $O'$ plane the projections of the Sun and the EO are displaced
with respect to the IO (see Fig.~\ref{tilt}). Then more diffracted light propagates behind the IO on one side, and less -- on the other side.
Another effect is the inclination of the EO image (and thus its defocussing), but is less important due to small expected values of the tilt $\sim
25$~arcsec. Obviously, the overoccultion by the IO should be large enough to cover possible shifts of the EO image. The coronagraph and
its mathematical model become not axi-symmetrical: the two waves propagating from different directions $\vec k_{\rho\varphi_1}$ and
$\vec k_{\rho\varphi_2}$ produce \emph{different} responses in $D$. From the computational point of view, the tilt of the coronagraph
around the $Oy$-axis is equivalent to multiplication of every $\Psi_{A \vec k}$ by the complex factor $\exp (-2 \pi i \frac{\Theta x}{\lambda} )$.
\begin{figure*}
\centering
\includegraphics[width=18cm]{tilt-sketch.eps}
\caption{Tilt of the coronagraph. Left panel: relative displacement of the IO and the axis of the instruments. Middle and right panels:
projections of the Sun, EO and IO onto planes $B$ and $O'$ (cf. Fig.~\ref{symmetrical} and Fig.~\ref{shift}).}
\label{tilt}
\end{figure*}
Transversal displacement of the IO is very similar to the case of the tilt of the coronagraph, so we do not present additional
computations for it. In the case of transversal displacement of the Lyot stop, the entrance aperture remains in the center of the umbra/penumbra pattern (it
receives minimal possible level of the diffracted light) and the bright diffraction ring is co-centered with the IO (i.e. the IO acts with maximal
efficiency). Thus we conclude that the influence of this misalignment will be smaller than that produced by the shift of the Sun and do not consider it
separately.
The case with the change of $z_0$ results in two effects: the change of level of the diffracted light on the entrance aperture,
and change of the size of the diffraction ring in the $O'$ plane. The combined impact result of the two effects in the detector plane is not
obvious and will be analyzed below. From the computational point of view, the misalignment requires recalculation of the
axi-symmetrical function $\Psi_{A00}$ mentioned in Sect.~\ref{algorithm} (see also Appendix~\ref{numerical}).
The effect of the longitudinal displacement of the IO from the $O'$ plane results in more diffracted light potentially propagate
beyond the IO. From the computational point of view, it changes the $\Psi_{O'}$ function (see Appendix~\ref{method}).
\section{Results of computations}
\label{results}
We consider $3\sigma$-values of the misalignments, resulting from the FF accuracy (see Sect.~\ref{intro}): relative shift of the Sun by
$\phi=10$~arcsec (due to the transversal displacement of the satellites), tilt of the coronagraph by $\Theta \sim 10$ (due to the
transversal displacement of the satellites), and 25~arcsec (due to the tilt of the satellite plus the misalignment of the telescope), change of
$z_0$ by $+15$~mm and $-100$~mm (we exaggerate the expected $\Delta z_0 = -15$~mm to the extreme), longitudinal displacement of the IO by
$\Delta z_{IO}=\pm 60$~$\mu$m.\footnote{Since the ASPIICS project is in the middle of the preparation phase, these values are subject to
changes.}
During the computations we use parameters listed in Table~\ref{geometry-table}. We use $4k \times 4k$ arrays to represent $\Psi$ in
each plane and sample the solar disk by $N=50$ points in radial direction and by $M=400$ points in polar direction~(see
Appendix~\ref{sampling-sec} for the analysis of the sampling).
We use the same solar limb darkening function as that used by RR17, which was initially proposed by \cite{1993AJ....106.2096V}:
\begin{equation}
B(\rho)=1-0.762 \left( 1-\sqrt{1-\rho^2} \right) -0.232(1-\rho^2)\log\left( \sqrt{1-\rho^2} \right),
\end{equation}
where $\rho$ is the radial coordinate expressed as a fraction of the solar radius.
Since we are interested in the possibility of registering the signal by the ASPIICS detector, we convert the obtained brightness into
number of photons pixel$^{-1}$. For the conversion we use the following ASPIICS parameters \citep{Renotte2016}: aperture size $A=\pi
R_A^2=19.6$~cm$^2$, angular size of a pixel of $2.8$~arcsec, the exposure time $t_{exp}=0.1$~s, and take the mean solar brightness as
$\mathrm{MSB}=2.08 \cdot 10^{20}$~photons s$^{-1}$ cm$^{-2}$ sr$^{-1}$ (the solar spectrum convolved with the spectral transmission
of ASPIICS).
To compare the diffracted light with the corona, we take the K-corona brightness observed during solar maximum from
\citet{1977asqu.book.....A}. We take into account the geometrical vignetting, which is determined by the size of the IO: we apply linear
vignetting function that starts (transmission $T=0\%$) from the angle $v_{min}=\arctan \left( \frac{R_{IO}^*-R_A}{z_0} \right)$, and
finishes ($T=100\%$) at the angle $v_{max}=\arctan \left( \frac{R_{IO}^*+R_A}{z_0} \right)$ , where $R_{IO}^*$ stands for the size of the IO
projected into $O$. In reality there will be an additional effect at low angles: high level of vignetting significantly
widens the point-spread function of the coronagraph, which not only reduces the intensity, but also smears the image of the lower corona (see RR17
for the analysis of the effect).
Further on we show profiles of the signal in the $D$ plane taken in the direction where the increase of the diffracted light is
maximal, i.e. in the $Ox$ direction in $D$.
\subsection{Qualitative results}
Our numerical computations fully confirm the qualitative behaviour of images in different planes as described in Sect.~\ref{properties}. In
the $O'$, plane the diffracted ring does not move due to solar shift, but it moves due to the tilt of the coronagraph. In the $D$ plane, the
diffracted ring moves neither in the case of the solar shift, nor in the case of the tilt of the coronagraph.
\subsection{Shift of the Sun and tilt of the coronagraph}
Intensities of the diffracted light for the case of the shift of the Sun are presented in Fig.~\ref{shift-profile}. In the top panel the radial
profiles correspond to different values of shift: the black curve -- symmetrical (unshifted) case, the yellow curve -- $\phi=5$~arcsec,
and the blue curve -- $\phi=10$~arcsec. The dash-dotted line shows the brightness of the K-corona. In the bottom panel, ratios of
shifted and symmetrical profiles are given.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{shifted-profile-ratio2.eps}}
\caption{Intensities of the diffracted light and the corona for the case of the shift of the Sun. Top panel: radial profiles computed
for the symmetrical and shifted cases. The black curve denotes the symmetrical case, the yellow curve is for the 5~arcsec shift, the
light blue curve is for the 10~arcsec shift. The dash-dotted line represents the brightness of the K-corona. Bottom panel: ratios of
shifted and symmetrical profiles of the diffracted light.}
\label{shift-profile}
\end{figure}
Intensities of the diffracted light for the case of the tilt of the coronagraph are presented in Fig.~\ref{tilt-profile}. The black curve
denotes the symmetrical case, the red curves is for the shift $\Theta=10$~arcsec, the green curve is for the $\Theta=25$~arcsec. The dash-dotted
line shows the brightness of the K-corona. As mentioned above, the radial position of the diffracted profile maximum does not shift
with the tilt of the coronagraph.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{tilted-profile2.eps}}
\caption{Intensities of the diffracted light and the corona for the case of the tilt of the coronagraph. The black
curve denotes symmetrical case, the red curve is for the 10~arcsec tilt, the green curve is for the 25~arcsec tilt. The dash-dotted line represents the
brightness of the K-corona.}
\label{tilt-profile}
\end{figure}
In Fig.~\ref{images} we present images in the plane $D$ computed for symmetrical, shifted and tilted cases. All four panels are
displayed with the same quasi-logarithmic color table.
\begin{figure*}
\includegraphics{images3.eps}
\caption{Images in the plane $D$ computed for: a) symmetrical case, b) solar shift by 10~arcsec, c) coronagraph tilt by
10~arcsec, and d) coronagraph tilt by 25~arcsec. Color table is expressed in mean solar brightness.}
\label{images}
\end{figure*}
\subsection{Longitudinal displacement of the IO and change of the inter-satellite distance $z_0$}
The results for the relative displacements of the IO along the $z$-axis by $+60$~$\mu$m and $-60$~$\mu$m, and for the change of the inter-satellite
distance $z_0$ by $+15$~mm and $-100$~mm are presented in Fig.~\ref{z1-z0-profile-ratio}. In the top panel, the radial profiles are given,
in the bottom panel -- ratios of misaligned and symmetrical cases. All the changes are relatively small (up to 30\%), and can barely be seen
in the top panel. Whereas the result of reducing $z_0$ by 100~mm is pronounced in the bottom plot, the changes are invisible in the top
panel. This is probably due to oscillatory behavior of the obtained profile. The highest impact is produced by the longitudinal
displacement of the IO by $+60$~$\mu$m further from L1, and the changes are comparable to those in the case of the shift of the Sun by 5~arcsec.
The change of $z_0$ by $15$~mm and inward displacement of the IO almost do not change the overall intensity of the diffracted light.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{z1-z0-profile-ratio.eps}}
\caption{Intensities of the diffracted light and the corona for the case of the longitudinal displacement of the IO
and change of $z_0$. Top panel: radial profiles of the diffracted light (colored curves) and the K-corona (dash-dotted line). Bottom
panel: ratios of the displaced and symmetrical cases.}
\label{z1-z0-profile-ratio}
\end{figure}
\subsection{Choosing the proper IO size}
It is clear that the major impact on the diffracted light is produced by the tilt of the coronagraph and by the misalignments that can
be reduced to such a tilt. Already the tilt of 25~arcsec is strong enough so that the diffracted light exceeds the level of the coronal
intensity near the minimal heights. This is not unexpected if we consider projections of the EO and IO onto the $O'$ plane: the rightmost
edge of the EO has the coordinates $x_{max}=z_1 \tan (25\arcsec + \omega_{EO})=1.669$~mm, which is larger than $R_{IO}=1.662$~mm. In
other words, the bright diffraction ring is not fully blocked by the IO (this situation is shown in Fig.~\ref{tilt}). Whereas
the difference of $7$~$\mu$m is small, a lot of diffracted light propagates further in the instrument.
The impact of the tilt can be reduced if we increase the size of the IO. In Fig.~\ref{tilted-IOs} we compare the diffracted light for
various sizes: $R_{IO}=1.662$~mm, 1.677~mm, and 1.694~mm for the same tilt of 25~arcsec. Already $R_{IO}=1.677$~mm reduces the diffracted
light by an order of magnitude. However, the diffracted light is still high enough at heights $<1.15 R_\sun$. The IO with $R_{IO}=1.694$~mm
almost completely removes the effect of the tilt, as the intensity of the diffracted light becomes almost equal to the symmetrical case with
$R_{IO}=1.662$~mm. We note, however, that due to the increase of the IO both the peak of the diffracted light intensity and the coronal vignetting
shift towards higher altitudes with respect to the symmetrical configuration. In the bottom part of Fig.~\ref{tilted-IOs} the vignetting
functions corresponding to different IO sizes are given.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{tilted-ios.eps}}
\caption{Comparison of the diffracted light corresponding to different IO sizes for the same tilt of 25~arcsec. The green curve
corresponds to $R_{IO}=1.662$~mm, the yellow curve is for the $R_{IO}=1.677$~mm, the blue curve is for the $R_{IO}=1.694$~mm, the
black solid curve denotes symmetrical case with $R_{IO}=1.662$~mm. Thin black and blue lines denote the coronal signal with
vignetting taken into account or the vignetting functions (in the bottom part of the plot.}
\label{tilted-IOs}
\end{figure}
In order to choose the proper size of the IO, one has to take into account several misalignments simultaneously. An example of the
combined effect of various misalignments is presented in Fig.~\ref{simultaneous-misalign}. The internal occulter with
$R_{IO}=1.694$~mm leads to the diffracted light below the coronal intensity in the unvignetted zone, even with all the misalignments
combined. It is clear that the major impact is due to the tilt of the telescope.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{compare-various-misalignm.eps}}
\caption{Simultaneous effect of various misalignments for the IO with $R_{IO}=1.694$~mm. The black curve corresponds to the
symmetrical case, the colored curves correspond to various combinations of misalignments: 25~arcsec tilt alone (blue); 25~arcsec
tilt and 10~arcsec shift of the Sun (green); 25~arcsec tilt, 10~arcsec shift of the Sun and 60~$\mu$m displacement of the internal
occulter (orange); 25~arcsec tilt, 10~arcsec shift of the Sun, 60~$\mu$m displacement of the internal occulter and $\Delta
z_0=-100$~mm change of the inter-satellite distance (red). The thin curve corresponds to the intensity of the K-corona taking
into account the vignetting function shown in the bottom part of the plot.}
\label{simultaneous-misalign}
\end{figure}
\subsection{IO with apodized profiles}
\label{apodized}
\cite{2013A&A...558A.138A} demonstrated that the use of apodized external occulters (i.e. those with variable transmission close to the edge) significantly
decreases the level of diffraction light at the entrance aperture. Unfortunately, making an apodized external occulter for ASPIICS is a technical challenge.
However, we demonstrate that apodizing the \emph{internal} occulter may also lead to a significant (although quantitatively different)
decrease of the diffracted light.
In Fig.~\ref{radial-step} we compare the diffracted light for IOs with various apodized profiles -- with the linear gradient of width 0.1 and
0.2~mm, for the same tilt of 25~arcsec.
The thick blue curve corresponds to the sharp-edge IO with $R_{IO}=1.694$~mm. The thick red curve denotes the IO with $R \le 1.662$~mm opaque center and linear gradient
$\Delta=0.1$~mm apodization, the thick green curve denotes a similar IO but with $\Delta=0.2$~mm apodization. Thin colored lines
represent coronal signal with additional vignetting by IOs taken into account (colors correspond to the colors of diffraction curves), or the vignetting
functions (given in the bottom part).
IOs with apodization have superior performance with respect to larger sharp-edge IOs. Reduction of the diffracted light is significant: the
profile of the IO with $\Delta=0.1$~mm apodization and 25~arcsec tilt corresponds to the symmetrical case with $R_{IO}=1.662$~mm
everywhere except for the innermost region of the corona. The IO with $\Delta=0.2$~mm apodization has the intensity of the diffracted
light in the inner region almost equal to the symmetrical case, and significantly lower level (factor of 50) of diffracted light at
heights $>1.1R_\sun$.
Reduction of the coronal signal occurs in a small region within $1.25R_\sun$, and the number of photons reaching the detector remains
rather high (see thin curves in Fig.~\ref{radial-step}). For example, for the case of $\Delta=0.2$~mm apodization and during
$t_{exp}=0.1$~s exposure time, the coronal signal remains at the average level $\sim 3\cdot 10^4$ photons pixel$^{-1}$ for the range of
coronal heights $1.1-1.25R_\sun$. Such a photon flux allows reliable registration of the signal.
\begin{figure}
\resizebox{\hsize}{!}{\includegraphics{radial-step-profiles.eps}}
\caption{Comparison of IOs with various apodization for the same tilt of 25~arcsec. The blue curve corresponds to sharp-edge IO with
$R_{IO}=1.694$~mm, the red curve is for the IO with $R_{IO}=1.662$~mm opaque part and $\Delta=0.1$~mm gradient, the green curve is
for the with $\Delta=0.2$~mm gradient. Thin colored lines denote coronal signal with corresponding vignetting taken into account and
vignetting functions (in the bottom part of the plot).}
\label{radial-step}
\end{figure}
\section{Discussion}
\label{discussion-sec}
\subsection{Comparison of various misalignments}
The performed computations show that the impact of various misalignments is considerably different for each type. For example, shifts
of the EO and the coronagraph may have both severe and negligible impacts on the diffracted light. Non-symmetrical umbra/penumbra pattern on
the entrance aperture not necessarily produces a severe impact (e.g. solar shift). Alternatively, symmetrical umbra/penumbra pattern on
the entrance aperture does not guarantee good diffracted light performance (e.g. tilt of the coronagraph). Placing the EO either closer
to the coronagraph or further away does not result in a significant increase of the diffracted light for the expected displacement.
However, the displacement of the IO from its nominal position along the $z$-axis (which may resemble the displacement of the EO) either worsens or
improves the performance depending on the position in the field of view.
\subsection{Using apodized IOs}
As we show in Sect.~\ref{apodized}, the IO with 0.2~mm apodization significantly reduces the level of diffracted light in tilted cases.
Obviously, the performance of the apodized IO will be even better at smaller tilts or in symmetrical cases.
The reduction of the coronal signal remains reasonable, as the inner corona is bright enough. Furthermore, the reduction of dynamic range
produced by apodized IOs may further improve the performance of the coronagraph by diminishing the level of ghosts and scattered light in the
instrument.
Using a smaller opaque region in apodized IOs potentially provides additional advantages, as it reduces $v_{min}$ -- the height at which
vignetting starts to disappear. In our case apodized IOs have $v_{min}=1.078R_\sun$, whereas a larger IO with $R=1.694$~mm has
$v_{min}=1.099R_\sun$. From this point view, even the IO with apodization $\Delta=0.1$~mm may have an advantage over larger
sharp-edge IOs with the same level of diffraction.
From a technological point of view, manufacturing of apodized IOs may be easier than that of sharp-edge IOs: producing a very sharp edge
may not be possible, and the remaining edge irregularities may additionally increase diffraction. In any case, accurate measurements of
the IO geometry and transmission are of a vital importance for processing of the registered coronal images.
Potential phase shifts of the wavefront due to propagation through the apodized IOs still need to be investigated.
\subsection{Dependence on the wavelength}
\label{wavelength}
ASPIICS is equipped with a filter wheel with 6 positions, which can be switched between individual exposures \citep{Renotte2016}. There are 3 passbands:
``wide-band'' 535--565 nm, narrow-band 530.4 nm for the \ion{Fe}{XIV} line, and narrow-band 587.7~nm for the \ion{He}{I} line. Three
additional positions are equipped with polarizers (rotated by $60^{\circ}$ with respect to each other) together with the wide-band filter.
There are various effects that influence the performance of ASPIICS in various passbands. First, chromatic aberrations of the primary
objective and the rest of the optics very slightly modify the point-spread-function of the telescope. Second, it is essential that in
the narrow-band filters the observed coronal structures will be bright in both spectral lines \emph{and} continuum, thus the relative effect of the diffraction
will be less important.
For the diffraction-related issues, we use the following reasoning: due to the property of diffraction discussed in
Sect.~\ref{properties}, the diffraction image in the $O'$ plane looks like a bright ring with the angular size of the EO. Obviously, this
property will be valid for \emph{any} wavelength, despite the fact that the diffraction pattern from an individual plane-parallel wave
$\Psi_{A\vec k}$ depends on $\lambda$ and the diffraction at the entrance aperture also depends on $\lambda$. As a consequence,
occulting the beam by the IO will occur with almost the same efficiency, regarding of the particular wavelength. Additional effects, like
different radial decrease of the intensity of the bright ring (the radial size of the Airy spot scales as $1.22 \lambda D/f$), may still be
present, but the major effect of blocking the diffracted light by the IO will be the same.
Thus we conclude that for the multi-wavelength observations diffraction effects do not change. A more detailed analysis of the spectral
dependence of the diffraction will be carried out in the future.
\section{Conclusions}
\label{conclusion-sec}
We analyzed different types of misalignments in the externally occulted ASPIICS coronagraph onboard the PROBA-3 mission, which has a very
small overoccultion of the solar disk. We considered impacts of misalignments from the physical point of view and their computational
realizations. We computed the resulting pattern of the diffracted light in the detector plane and compared it with the coronal
intensity.
Our computations of the diffracted light can be applied to any coronagraph with arbitrary geometry. However, for the case of ASPIICS we show
an exceptional importance of precise co-alignment. The most important misalignment is the tilt of the coronagraph. Special care should be taken to co-align
the external and internal occulters. The true optical axis of the instrument (through the center of the IO and the entrance aperture) should
be pointed to the center of EO as close as possible. The margin for the acceptable tilts $\sim 25$~arcsec is determined by oversizing the IO over
the EO; the margin can be increased in case of an apodized IO. Impacts of other misalignments are significantly smaller. Since the tilt
of the coronagraph can be potentially corrected in flight by correcting the attitude of the coronagraph satellite, there is a possibility to
reduce the size of the IO and, as a consequence, to reduce the minimal observed height of the corona.
We found that apodized IOs have very good diffracted light rejection performance, and especially are resistant to tilts.
\begin{acknowledgements}
We acknowledge support from the Belgian Federal Science Policy Office through the ESA -- PRODEX programme (grant No.~4000117262). Authors
are grateful to Dr. Anton Reva for his help during the preparation of the manuscript. We are grateful to anonymous referee for numerous
suggestions, which allowed to improve the paper.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
\label{sec:introduction}
The purpose of this note is to show that linear payoffs arise as
optimal contracts offered by a principal to agents in a setting with
moral hazard, private information and competition for agents. The
design of compensation schemes in a principal-agent relationship
typically involves a bonus component to incentivize the agent (e.g.\
employee, worker) to act in the principal's (employer's) interest. The
complicated, often highly non-linear form of contracts suggested by
economic models is seldom met in reality where bonus schemes
encountered are simple and frequently linear in the output generated
by the agent. One explanation for this gap between theory and practice
lies in the fact that linear compensation schemes are robust in the
richer, more diffuse real environment, whereas the optimal contract in
a highly stylised economic model fails at the slightest change of
model assumptions or parameters.
Against this backdrop, \citet{Holmstrom1987} show that the optimal
contract for an agent controlling the drift of an output process
driven by a Brownian motion, when the principal observes only the
output process, is a linear function of the terminal output. The
principal's inability to directly observe the action of the agent is a
{\em moral hazard problem}, implying that the agent's effort cannot be
contractually agreed, but instead the agent must be incentivized to
exert effort. Restricting the principal's observability to the
terminal output, \citet{Sung2005} extends this to a setting where
agents have different capabilities that the principal cannot observe
or otherwise directly infer ({\em private information}). In this {\em
adverse selection problem\/}, it is optimal to offer a menu of
linear contracts designed for the different agent types, incentivizing
each agent to choose the contract designed for their type ({\em
screening}).
Screening typically provides agent types with different utilities. As
a consequence, if there are several principals competing for agents,
reservation utilities for agents will be type-dependent (as opposed to
the case, for example, where the outside option is not to work at
all). This setting is found in the recent literature on competition,
e.g.\ \citet{Jullien2000,Benabou2016,Bannier2016}. We prove that the
linearity of contracts in the Holmstrom-Milgrom model carries over to
the case of moral hazard, private information and competition. The
proofs use techniques from stochastic control theory, see e.g.\
\citet{Fleming1975}.
\section{The model}
\label{sec:model}
The setting is similar to \citet{Holmstrom1987},
\citet{Schaettler1993} and \citet{Sung2005}. An agent employed by a
principal exerts (costly) effort, which in turn increases her
output. The principal observes only the output, consisting of the
agent's effort level and random noise ({\em moral hazard}). Hence, the
effort level cannot be contracted and must be incentivized by a bonus
scheme. With agents of different capabilities that are unobservable
by the principal ({\em private information}), the principal may find
it optimal to offer different contracts specifically designed for each
type, and must take care that each contract appeals most to the type
that it is designed for. Finally, with several principals competing
for agents, reservation utilities are derived from the outside option
of being employed by another principal. Just as the utilities offered
to agents of different capabilities differ, so do their reservation
utilities. These must be taken into account by a principal wishing to
employ agents of all types. The principal is hence faced with an
optimization problem taking into account random noise,
unobservable agent types and type-dependent reservation
utilities. This type of problem can be solved using martingale theory
to derive a Hamilton-Jacobi-Bellman equation, which in turn leads to
constant controls that do not depend on time.
First, we formalize the setting and then give the principal's
optimization problem in the next section.
A risk-neutral principal
employs a risk-averse agent whose preferences are expressed by CARA
utility with parameter $\rho$, i.e., $U(x)=1-\e^{-\rho x}$. The
agent's capabilities $\theta\in \{\theta_L,\theta_H\}$ are private
information. The probabilities $\alpha$, resp.\ $1-\alpha$, of meeting
an $H$-type, resp.\ $L$-type, agent are known to the principal.
A type-$k$ agent, by exerting effort $\mu=(\mu_t)_{t\geq 0}$, which is
assumed to be bounded, controls the drift of an output process with
dynamics
\begin{equation*}
\dd Z_t = \mu_t\,\theta_k\,\dd t + \sigma\, \dd W_t, \quad t\geq 0,
\end{equation*}
with $\sigma>0$ and where $W=(W_t)_{t\geq 0}$ is a Brownian motion
independent of $\theta$. The agent observes the Brownian motion $W$,
so that $\mu$ is adapted to the filtration generated by the Brownian
motion. The agent's effort is subject to an instantaneous cost
$c(\mu_t)$, $t\geq 0$, where $c(0)=0$, $c$ is strictly increasing,
convex and continuously differentiable. In addition, we require
$c'''\geq 0$ for Proposition \ref{prop:optimalcontract}. We shall
assume that $\theta_H>\theta_L$, expressing that a type-$H$ agent
generates a higher drift at equal effort cost than a type-$L$ agent.
The principal observes the output process $Z=(Z_t)_{t\geq 0}$, but
neither $W$ nor $\mu$, so the compensation, realised at time $1$, can
be contingent on $(Z_t)_{0\leq t\leq 1}$ only. At time $0$, the
principal offers contracts, consisting of sharing rules
$\{S((Z_t)_{t\in[0,1]},\theta_k), k\in \{H,L\}\}$,
where $S(\cdot,\theta_k)$ denotes the sharing rule {\em designed\/} for
a type-$k$ agent. An agent choosing contract $S(\cdot,\theta_k)$
at time $0$, receives at time $1$
\begin{equation*}
S((Z_t)_{0\leq t\leq 1},\theta_k) - \int_0^1 c(\mu_t)\, \dd t,
\end{equation*}
while the principal receives
\begin{equation*}
Z_1 - S((Z_t)_{0\leq t\leq 1},\theta_k).
\end{equation*}
\section{The principal's problem}
\label{sec:principals-problem}
Denote by $S(\cdot,\theta_m)$ the contract designed for an $m$-type
agent. An agent of type $k\in\{H,L\}$, when choosing contract
$S(\cdot, \theta_m)$ exerts effort $\mu^{k,m}=(\mu_t^{k,m})_{t\geq 0}$
and derives certainty equivalent $w_{k,m}$ at time $1$. Whenever
$k=m$, we write $\mu^k$ and $w_k$. The principal's problem is as
follows:
\begin{problem}
\label{problem}
Choose controls $\{\mu^H, \mu^L\}$ and a menu of contracts
$\{S(\cdot,\theta_H)$, $S(\cdot,\theta_L)\}$ maximising
\begin{equation*}
\alpha \, \E\left[Z_1^H - S((Z_t^H)_{0\leq t\leq 1},
\theta_H)\right]
+ (1-\alpha) \,\E\left[ Z_1^L - S((Z_t^L)_{0\leq t\leq 1}, \theta_L)\right],
\end{equation*}
subject to \smallskip
\begin{enumerate}[(1)]
\addtolength{\itemsep}{5pt}
\item $\dd Z_t^k = \mu_t^k\theta_k\, \dd t + \sigma \dd W_t$,
$k\in \{H,L\}$,
\item
$\mu_t^{k,m}\in \argmax_{(\mu_t)_{0\leq t\leq 1}}
\E\left[U\left(S((Z_t)_{0\leq t\leq 1},\theta_m) - \int_0^1
c(\mu_t)\, \dd t\right)\right]$, where
$\dd Z_t=\mu_t\theta_k \, \dd t + \sigma\, \dd W_t$, and
$k,m\in \{H,L\}$,
\item
$\E\left[U\left(S((Z_t^k)_{0\leq t\leq }, \theta_k) - \int_0^1
c(\mu_t^k)\, \dd t\right)\right]\geq
\E\left[U\left(S((Z_t^{k,m})_{0\leq t\leq 1},\theta_m) - \int_0^1
c(\mu_t^{m,k})\, \dd t\right)\right]$, where
$\dd Z_t^{k,m} = \mu_t^{k,m}\theta_k\, \dd t + \sigma\, \dd W_t$
and $m,k\in \{H,L\}$,\hfill {\em (ICC)}
\item
$\E\left[U\left(S((Z_t^k)_{0\leq t\leq }, \theta_k) - \int_0^1
c(\mu_t^k)\, \dd t\right)\right]\geq U(w_k)$, where
$k\in \{H,L\}$, \hfill {\em (PC)}.
\end{enumerate}
\end{problem}
\medskip
The first constraint defines the dynamics of the output processes of
each type when choosing the contract designed for her. The second
constraint expresses that agents maximise their expected utility. The
third constraint, the incentive compatibility constraint (ICC), makes
each agent optimally choose the contract designed from her. Finally,
the fourth constraint, the participation constraint (PC), ensures that
an agent contracts with the principal instead of choosing her outside
option with certainty equivalent $w_k$ (the general results do not
change if it were unprofitable to attract a particular agent type).
\section{The agent's choice of drift}
We consider the agent's problem when faced with a menu of
contracts. The following result is slightly adapted from
\citet{Holmstrom1987}.
\begin{theorem}
\label{theorem:hm}
The adapted stochastic process $(\mu_t)_{0\leq t\leq 1}$ is
implemented with certainty equivalent $w$ by a type-$k$ agent
by a sharing rule $S((Z_t)_{0\leq t\leq 1},\theta_k)$ only if
\begin{multline}
S((Z_t)_{0\leq t\leq 1}, \theta_k) = w + \int_0^1 c(\mu_t)\, \dd t +
\int_0^1 \frac{c'(\mu_t)}{\theta_k}\, \dd Z_t - \int_0^1 c'(\mu_t)
\mu_t\, \dd t + \frac{\rho}{2}\int_0^1
\left(\frac{c'(\mu_t)}{\theta_k}\right)^2\, \sigma\, \dd t.
\label{eq:7}
\end{multline}
\end{theorem}
\begin{proof}
See Theorem 6 of \citet{Holmstrom1987} and Corollary 4.1 of
\citet{Schaettler1993}.
\end{proof}
The first two terms provide a certainty equivalent of $w$ and a direct
compensation of the effort cost should the agent choose to exert
effort $(\mu_t)_{0\leq t\leq 1}$. The third term incentivizes the
agent to choose effort level $\mu$. The last two terms compensate the
agent for the mean and risk of the output process, i.e., they
correspond to the certainty equivalent of the third term.
\begin{proposition}
\label{prop:effort}
A type-$k$ agent, when choosing the contract designed for the
$m$-type agent, derives expected utility
\begin{equation}
\label{eq:9}
\E\left[U\left(w_m + \int_0^1 c'(\mu^{k,m}_t)\mu_t^{k,m}-
c(\mu_t^{k,m}) - \left(c'(\mu_t^m)\mu_t^m -
c(\mu_t^m)\right)\, \dd t\right)\right],
\end{equation}
where $\mu^{k,m}$ denotes the $k$-type agent's optimal control,
which solves
\begin{equation}
\label{eq:4}
c'(\mu_t^{k,m}) = \frac{\theta_k}{\theta_m} c'(\mu_t^m).
\end{equation}
Furthermore, the $H$-type agent exerts greater effort and derives
greater utility from a given contract than the $L$-type agent.
\end{proposition}
\begin{proof}
Let $S((Z_t)_{0\leq t\leq 1},\theta_m)$ be the contract designed for
an $m$-type agent offering $w_m$ and implementing $\mu^m$. A
type-$k$ agent implementing $\mu$, that is,
$\dd Z_t = \mu_t\theta_k\, \dd t + \sigma\, \dd W_t$, $t\geq 0$,
derives expected utility
\begin{multline*}
\E\left[U\left(w_m + \int_0^1 c(\mu_t^m) - c(\mu_t)\, \dd t +
\int_0^1 \frac{c'(\mu_t^m)}{\theta_m}\, \dd Z_t
- \int_0^1 c'(\mu_t^m) \mu_t^m - \frac{\rho}{2}
\left(\frac{c'(\mu_t^m)}{\theta_m}\right)^2\sigma^2\, \dd
t\right)\right],
\end{multline*}
which can be written as $\E\left[U(X_1^{\mu})\right]$ with state
process
\begin{multline*}
X_u^{\mu}:=w_m\, u + \int_0^u c(\mu_t^m)-c(\mu_t)\, \dd t +
\int_0^u \frac{c'(\mu_t^m)}{\theta_m} \sigma\, \dd W_t \\%
+ \int_0^u c'(\mu_t^m)\mu_t \frac{\theta_k}{\theta_m} -
c'(\mu_t^m) \mu_t^m +
\frac{\rho}{2}\left(\frac{c'(\mu_t^m)}{\theta_m}\right)^2\sigma^2\,
\dd t.
\end{multline*}
By the It\^o formula,
\begin{equation*}
\dd X_t^{\mu} = \left[w_m + c(\mu_t^m) - c(\mu_t) +
c'(\mu_t^m)\left(\mu_t\frac{\theta_k}{\theta_m} - \mu_t^m\right) +
\frac{\rho}{2}
\left(\frac{c'(\mu_t^m)}{\theta_m}\right)^2\sigma^2\right]\, \dd
t +
\frac{c'(\mu_t^m)}{\theta_m}\sigma\, \dd W_t.
\end{equation*}
Define $J(t,x;\mu)=\E\left[U(X_1^{\mu})|X_t^{\mu}=x\right]$, which
is once (twice) continuously differentiable in $t$ ($x$) (applying
Dominated Convergence for differentiating inside the expectation
operator), so that $J(0,x;\mu)=\E[U(X_1^{\mu})]$ corresponds to the
objective function. Because $J$ is a martingale, the following PDE
holds:
\begin{equation*}
J_t + \frac{1}{2}\left(\frac{c'(\mu_t^m)}{\theta_m}\right)^2
\sigma^2 J_{xx} + \left[w_m + c(\mu_t^m) - c(\mu_t) +
c'(\mu_t^m)\left(\mu_t\frac{\theta_k}{\theta_m} - \mu_t^m\right) +
\frac{\rho}{2}
\left(\frac{c'(\mu_t^m)}{\theta_m}\right)^2\sigma^2\right]\,
J_x=0,
\end{equation*}
where $J_t, J_x, J_{xx}$ denote the respective partial first- and
second-order derivatives. Setting $V(t,x)=\sup_{\mu} J(t,x;\mu)$,
the Hamilton-Jacobi-Bellman PDE is
\begin{equation*}
V_t + \frac{1}{2}\sigma^2 V_{xx} + \sup_{\mu}
\left[c'(\mu_t^m) \frac{\theta_k}{\theta_m}\mu_t -
c(\mu_t)\right] V_x + \left[c(\mu_t^m) -
c'(\mu_t^m) \mu_t^m + \frac{\rho}{2}
\left(\frac{c'(\mu_t^m)}{\theta_m}\right)^2 \sigma^2\right]
V_x=0,
\end{equation*}
with boundary condition $V(1,x)=U(x)$. The agent's Hamiltonian is
given by
\begin{equation*}
\mathcal H=-c(\mu_t) +
c'(\mu_t^m)\mu_t \frac{\theta_k}{\theta_m},
\end{equation*}
leading to the optimal effort choice $\mu_t^{k,m}$ fulfilling the
FOC \eqref{eq:4}.
If $k=H$ and $m=L$, the FOC expresses that the $H$-type agent exerts
greater marginal effort on the $L$-type's contract than the
$L$-type. Moreover, because of the convexity of $c$, it follows that
$\mu_t^{H,L}>\mu_t^L$. Conversely, if $k=L$ and $m=H$, then
$\mu_t^{L,H}<\mu_t^H$.
The $k$-type agent's expected utility is given by \eqref{eq:9}.
Because $c$ is strictly increasing and convex, the mean value
theorem implies $c'(x)x > c(x)$, and
\begin{equation*}
\frac{\dd}{\dd x} \left[c'(x) x - c(x)\right] =
c''(x) x>0, \quad x>0.
\end{equation*}
Therefore, when $m=H$ and $k=L$, the integral is strictly positive,
and the $H$-type agent derives a greater utility from the $L$-type's
contract than the $L$-type. Conversely, if $m=L$ and $k=H$, the
$L$-type derives a smaller utility from the $H$-type's contract than
the $H$-type.
\end{proof}
The classical result is obtained that if both types' reservation
utilities are equal, then the $H$-type has an incentive to imitate if
second-best contracts were offered (i.e., contracts with moral hazard
only), in which case the contract designed for the $L$-type needs to
be distorted to prevent the $H$-type from imitating (e.g.\
\citet{Salanie2005}). If reservation utilities are type-dependent,
then the situation may be reversed, and the $H$-type's contract may
need to be distorted to prevent the $L$-type from imitating.
\section{Optimal contracts}
Turning to Problem \ref{problem}, we restrict the analysis to the case
where the $H$-type has an imitation incentive; the case when the
difference of the types' reservation utilities is sufficiently large
for the $L$-type to have an imitation incentive is treated in a
similar way. We omit the proof of the following well-known results
when the $H$-type has an imitation incentive
(e.g. \citet{Salanie2005}): The $L$-type's {\em (PC)\/} is binding
(constraint (4) in Problem \ref{problem}; to attract the $L$-type) and
the $H$-type's {\em (ICC)\/} is binding (constraint (3) in Problem
\ref{problem}; to prevent the high type from imitating), while the
$L$-type's {\em (ICC)\/} is non-binding. The $H$-type's contract
features the second-best (constant) drift rate $\mu^{H,\star}$ (there
is no reason to deviate from the optimum), while the effort level
$\mu^{L,\star}$ in the contract for the $L$-type is distorted to
prevent the $H$-type from imitating.\medskip
The principal thus solves
\begin{equation*}
\sup_{\mu^H, \mu^L, S(\cdot, \theta_H), S(\cdot, \theta_L)}\,
\E\left[\alpha (Z_1^H - S((Z_t^H)_{0\leq t\leq 1}, \theta^H)) +
(1-\alpha) (Z_1^L - S((Z_t^L)_{0\leq t\leq 1}, \theta^L))\right],
\end{equation*}
subject to\smallskip
\begin{enumerate}[(1)]
\addtolength{\itemsep}{3pt}
\item
$\E\left[U\left(S((Z_t^L)_{0\leq t\leq 1}, \theta_L) - \int_0^1
c(\mu_t^L)\, \dd t\right)\right]=U(w_L)$;\hfill {\em (PCL)}
\item $\left.\right.$\vspace*{-3\baselineskip}
\begin{multline*}
\E\left[U\left(S((Z_t^{H,L})_{0\leq t\leq 1}, \theta^L) -
\int_0^t c(\mu_t^{H,L})\, \dd t\right)\right]
= \E\left[U\left(S((Z_t^H)_{0\leq t\leq 1}, \theta^H) -
c(\mu^{H,\star})\right)\right] \\%
= \E\Big[1-\exp\Big(-\rho\Big(\underbrace{w_L + \int_0^1
c'(\mu_t^{H,L})\mu_t^{H,L} - c(\mu_t^{H,L}) -
\left[c'(\mu_t^L)\mu_t^L - c(\mu_t^L)\right]\, \dd
t}_{=:u_1(\mu^L)}\Big)\Big)\Big],
\end{multline*}
with $\mu_t^{H,L}$ given by Equation \eqref{eq:4}; \hfill {\em
(ICCH)}
\item
$\E\left[U\left(S((Z_t^H)_{0\leq t\leq 1}, \theta^H) -
c(\mu^{H,\star})\right)\right] \geq \E[U(w_H)]$;
\hfill {\em (PCH)}
\end{enumerate}
\begin{proposition}
\label{prop:optimalcontract}
Under adverse selection, moral hazard and when reservation
utilities are type-dependent, the optimal effort levels
$\mu^{H,\star}$, $\mu^{L,\star}$ in the contracts designed for the
$H$-type, resp.\ $L$-type agent are constant, and optimal
contracts are linear in the terminal outputs $Z_1^H$ and $Z_1^L$.
\end{proposition}
In the proof it is shown that if {\em (PCH)\/} is non-binding, then
$\mu^{L,\star}$ satisfies
\begin{equation}
\label{eq:5}
c'(\mu^{L,\star}) = \left(\theta_L - \frac{\alpha}{1-\alpha} \left\{
c''(\mu^{H,L,\star}) \mu^{H,L,\star}\frac{\partial}{\partial \mu^{L,\star}}
\mu^{H,L,\star} - c''(\mu^{L,\star})\mu^{L,\star}\right\} \right)\,
\left(1+\frac{\rho\, \sigma^2}{\theta_L^2}
c''(\mu^{L,\star})\right)^{-1},
\end{equation}
By the assumption that $c'''\geq 0$ it follows directly that the
optimal control is smaller than the second best optimal control
without adverse selection (which is obtained when $\alpha=0$). If
{\em (PCH)\/} is binding, then $\mu^{L,\star}$ satisfies
\begin{equation}
\label{eq:2}
w_H - w_L = c'(\mu^{H,L,\star}) \mu^{H,L,\star} - c(\mu^{H,L,\star})
- [c'(\mu^{L,\star}) \mu^{L,\star}- c(\mu^{L,\star})].
\end{equation}
\begin{proof}
First, assume that {\em (PCH)\/} is non-binding (this case arises
when the difference between reservation utilities is small or zero).
From {\em (ICCH)\/}, where the right-hand side depends only on the
effort level incentivized by the principal, it follows that the
$H$-type's certainty equivalent is
$\displaystyle -\frac{\ln \E[\e^{-\rho\, u_1(\mu^L)}]}{\rho}$.
Using Theorem \ref{theorem:hm}, the conditions imply
\begin{multline}
S((Z_t^L)_{0\leq t\leq 1}, \theta_L) \\
= w_L + \int_0^1 c(\mu_t^L)\,\dd t + \int_0^1
\frac{c'(\mu_t^L)}{\theta_L}\, \dd Z_t^L - \int_0^1 c'(\mu_t^L)\,
\mu_t^L\, \dd t + \frac{\rho}{2} \int_0^1
\left(\frac{c'(\mu_t^L)}{\theta_L}\right)^2 \sigma^2\, \dd
t\label{eq:8}
\end{multline}
and
\begin{multline}
S((Z_t^H)_{0\leq 1\leq t}, \theta_H) \\
= \frac{-\ln \E[\e^{-\rho\, u_1(\mu^L)}]}{\rho} + c(\mu^{H,\star})
+ \frac{c'(\mu^{H,\star})}{\theta_H} Z_1^H - c'(\mu^{H,\star})
\mu^{H,\star} + \frac{\rho}{2} \left(\frac{c'(\mu^{H,\star})}
{\theta_H}\right)^2 \sigma^2. \label{eq:1}
\end{multline}
Setting
\begin{multline*}
X_t^{\mu^L}:=\alpha \left(Z_t^H -\left\{ -\frac{\ln\E[\e^{-\rho\,
u_t(\mu^L)}]}{\rho} + c(\mu^{H,\star}) t + \frac{\rho}{2}
\left(\frac{c'(\mu^{H,\star})}{\theta_H}\right)^2 \sigma^2\,
t+ \frac{c'(\mu^{H,\star})}{\theta_H}\sigma\,
W_t\right\}\right) \\
+ (1-\alpha) \left(Z_t^L - \left\{w_L\, t + \int_0^t c(\mu_u^L)\,
\dd u + \frac{\rho}{2} \int_0^t
\left(\frac{c'(\mu_u^L)}{\theta_L}\right)^2\sigma^2\, \dd u +
\int_0^t \frac{c'(\mu_t^L)}{\theta_L}\, \sigma\, \dd
W_t\right\}\right),
\end{multline*}
with $\dd Z_t^L = \mu_t^L\, \theta_L\, \dd t + \sigma\, \dd W_t$ and
$\dd Z_t^H = \mu^{H,\star}\theta_H\, \dd t + \sigma\, \dd W_t$, the
principal's problem is
\begin{equation*}
\sup_{\mu_L} \E[X_1^{\mu^L}].
\end{equation*}
The dynamics of $X^{\mu^L}$ are
\begin{align*}
\dd X_t^{\mu^L}
&= \alpha \Bigg( \left\{\mu^{H,\star} \theta_H +
\frac{\partial}{\partial t} \frac{\ln
\E[\e^{-\rho\, u_t(\mu^L)}]}{\rho} - c(\mu^{H,\star}) -\frac{\rho}{2}
\left(\frac{c'(\mu^{H,\star})}{\theta_H}
\right)^2 \sigma^2\right\}\, \dd t\Bigg. \\
&\phantom{=\alpha\,} +
\left(1-\frac{c'(\mu^{H,\star})}{\theta_H}\right)\sigma\, \dd W_t
\Bigg.\Bigg)
\\
&+ (1-\alpha) \left(\left\{\mu_t^L\theta_L -
w_L - c(\mu_t^L) - \frac{\rho}{2}
\left(\frac{c'(\mu_t^L)}{\theta_L}\right)^2\sigma^2\right\}\,
\dd t + \left(1-\frac{c'(\mu_t^L)}{\theta_L}\right) \sigma\, \dd
W_t\right),
\end{align*}
and the principal's Hamiltonian is
\begin{equation*}
\mathcal H = \alpha\, \frac{\partial}{\partial t} \frac{\ln \E[\e^{-\rho\,
u_t(\mu^L)}]}{\rho} + (1-\alpha)\left\{ \mu_t^L \theta_L -
c(\mu_t^L) - \frac{\rho}{2} \left(\frac{c'(\mu_t^L)} {\theta_L}
\right)^2 \sigma^2\right\},
\end{equation*}
with
\begin{equation}
\label{eq:6}
\frac{\partial}{\partial t} \frac{\ln \E[\e^{-\rho\,
u_t(\mu^L)}]}{\rho}
= -\frac{\E\left[\e^{-\rho\, u_t(\mu^L)}\, \left\{w_L +
c'(\mu_t^{H,L})\mu_t^{H,L}-c(\mu_t^{H,L}) - [c'(\mu_t^L)\mu_t^L -
c(\mu_t^L)]\right\}\right]} {\E[\e^{-\rho\, u_t(\mu^L)}]}.
\end{equation}
Equation \eqref{eq:6} describes the change in certainty equivalent
offered to the $H$-type agent, including the information rent to
prevent her from imitating. This is $\mathcal F_0$-measurable, i.e.,
fixed at time $0$. By the principle of optimality, the optimal
change in certainty equivalent does not depend on any particular
time $t$; likewise the optimal control does not depend on the
particular time $t$, so that Equation \eqref{eq:6} is constant. This
is fulfilled for a deterministic and constant control. A constant
optimal control is necessary as well, as the Hamiltonian is
optimised by a deterministic choice of $\mu_t^L$, which is constant
by the principle of optimality. Hence, \eqref{eq:6} becomes
\begin{equation*}
\frac{\partial}{\partial t} \frac{\ln \E[\e^{-\rho\, u_t(\mu^L)}]}
{\rho} = -\left\{w_L +
c'(\mu^{H,L})\mu^{H,L}-c(\mu^{H,L}) - [c'(\mu^L)\mu^L -
c(\mu^L)]\right\},
\end{equation*}
and the Hamiltonian simplifies to
\begin{multline*}
\mathcal H = -\alpha \Bigg\{
\underbrace{c'(\mu^{H,L})}_{=c'(\mu^L) \theta_H/\theta_L}
\mu^{H,L}-c(\mu^{H,L}) - [c'(\mu^L)\mu^L - c(\mu^L)]\Bigg\} \\
+ (1-\alpha) \left\{ \mu^L \theta_L - c(\mu^L) - \frac{\rho}{2}
\left(\frac{c'(\mu^L)} {\theta_L} \right)^2 \sigma^2\right\}.
\end{multline*}
The optimum is determined via
\begin{multline*}
\frac{\partial}{\partial \mu^L} \mathcal H = -\alpha \left\{
\frac{\theta_H}{\theta_L} c''(\mu^L)
\mu^{H,L}\, \frac{\partial}{\partial \mu^L} \mu^{H,L}
- c''(\mu^L)\mu^L\right\} \\
+ (1-\alpha) \left\{\theta_L - c'(\mu^L) - \frac{\rho\,
\sigma^2}{\theta_L^2} c'(\mu^L) c''(\mu^L)\right\},
\end{multline*}
which is zero if $\mu^{L,\star}$ fulfills Equation \eqref{eq:5},
requiring this is nonnegative. It is easily verified that this is a
minimum by $c'''\geq 0$. The sharing rules \eqref{eq:8} and
\eqref{eq:1} depend only on $Z_1$ instead of $(Z_t)_{0\leq t\leq 1}$
and are linear in $Z_1$.
\medskip
If {\em (PCH)\/} is binding, then
\begin{equation*}
\E[\e^{-\rho\, u_1(\mu^L)}] = \e^{-\rho \, w_H},
\end{equation*}
and
\begin{equation*}
\dd \E[\e^{-\rho\, u_t(\mu^L)}] = \dd \e^{-\rho\, w_H\, t} = -\rho\,
w_H\, \e^{-\rho\, w_H\, t}\, \dd t.
\end{equation*}
The left-hand side is therefore an expectation of an exponential
accruing at a constant rate. Since $u_1(\mu^L)$ is comprised of a
constant and an integral with respect to time, by the principle of
optimality, the exponent itself must be constant. Hence
$\mu^{L,\star}$ is constant. Furthermore, $\mu^{L,\star}$ solves the
binding participation constraint, which can be expressed as in
Equation \eqref{eq:2}.
\end{proof}
\bibliographystyle{abbrvnamed}
|
\section{INTRODUCTION}\label{intro}
The inflationary paradigm is the most widely accepted scenario for seeding the structures in the Universe, so far passing observational tests with flying colors. There is, however, both theoretical and observational room for contributions from alternative, well-motivated scenarios. Among these are perturbations sourced by cosmic topological defects formed at cosmological phase transitions.
In particular, cosmic strings (CS) are theoretically expected to be produced in the early Universe \citep{Kibble:1976sj,Zeldovich:1980gh,Vilenkin:1981iu,
Vachaspati:1984dz,Vilenkin:1984ib,Shellard:1987bv,
Hindmarsh:1994re,Vilenkin:2000jqa,Sakellariadou:2006qs,
Bevis:2007gh,Depies:2009im,Bevis:2010gj,
Copeland:1994vg,Sakellariadou:1997zt,Sarangi:2002yt,
Copeland:2003bj,Pogosian:2003mz,Majumdar:2002hy,
Dvali:2003zj,Kibble:2004hq,HenryTye:2006uv}.
The detection of CS would open a unique window to the physics of the early Universe \citep{Kibble:1976sj,Zeldovich:1980gh,Vilenkin:1981iu,Vilenkin:2000jqa,Firouzjahi:2005dh}. Therefore a lot of effort has been put into developing powerful statistical tools for cosmic string network detection and putting tight upper bounds on the CS tension, parametrized by $G\mu$, where $G$ and $\mu$ represent Newton's constant and the string's tension, respectively.
The string tension is intimately related to the energy of the phase transition epoch,
\begin{equation}
\frac{G\mu}{c^2}=\order{\frac{\varpi^2}{M_{\rm Planck}^2}},
\end{equation}
where $\varpi$ is the symmetry breaking energy scale, $c$ is the speed of light and $M_{\rm Planck}\equiv\sqrt{\hbar c/G}$ represents the Planck mass. In this paper we work in natural units with $\hbar=c=1$.
A CS network would leave various imprints on cosmic microwave background (CMB) anisotropies. The Gott-Kaiser-Stebbins (KS) effect \citep{Kaiser:1984iv,Gott:1985, Stebbins:1988, Bouchet:1988hh, Allen:1997ag,Pen:1997ae,Ringeval:2012tk} corresponds to the integrated Sachs-Wolfe effect caused by moving strings. It produces line-like discontinuities on the CMB temperature anisotropies ~\citep{Hindmarsh:1993pu,Stebbins:1995} of the form
\begin{equation}
\frac{\delta T}{T} \sim 8\pi G\mu v_{\mathrm{s}}.
\end{equation}
Here $v_\mathrm{s}$ is the transverse velocity of the string.
The CS network is also expected to produce extra CMB polarization \citep{Benabed:1999wn,Danos:2010gx,Brandenberger:2011eq,Bevis:2007qz} and dipole modulation \citep{Ringeval:2015ywa}.
CMB-based approaches to search for CS are quite diverse. For example, \cite{Ade:2013xla,Ade:2015xua} use the {\it Planck} temperature power spectrum to get an upper bound of $G\mu<3.0\times10^{-7}$ for Abelian-Higgs strings, which improves with the {\it Planck} polarization\footnote{Note that the {\it Planck} 2015 polarization data is preliminary
at large scales due to residual systematics.} to $G\mu<2.0 \times 10^{-7}$ for Abelian-Higgs strings \citep{Lizarraga:2016onn}, to $G \mu < 1.5 \times 10^{-7}$ for Nambu-Goto strings \citep{Lazanu:2014eya} and to $G\mu<1.1 \times 10^{-7}$ for a multi-parameter fit to the unconnected segment model \citep{Charnock:2016nzm}.
In the search for the CS network, one could exploit the non-Gaussianity of CS-induced fluctuations, e.g., through measuring CMB bispectrum, using Wavelet-based methods, or measurements of the CMB Minkowski functionals. These searches lead to
$G\mu<8.8\times10^{-7}$,
$G\mu<7\times10^{-7}$ and $G\mu<7.8\times10^{-7}$, respectively
\citep{Hindmarsh:2009qk, Hindmarsh:2009es, Ade:2013xla,Regan:2015cfa, Ringeval:2010ca, Ducout:2012it}.
\iffalse
CMB-based approaches to search for CSs are quite diverse: using CMB power spectra (giving, e.g., the upper bound of $G\mu<3.0\times10^{-7}$ with {\it Planck} temperature data for Abelian-Higgs strings, \citep{Ade:2013xla}, which improves with {\it Planck} polarization\footnote{Note that {\it Planck}15 polarization data is preliminary
at large scales due to residual systematics} to $G\mu<1.1 \times 10^{-7}$ for Abelian-Higgs strings \citep{Charnock:2016nzm} and to $G \mu < 1.5 \times 10^{-7}$ for Nambu-Goto strings, \citep{Lazanu:2014eya}, exploring the non-Gaussianity of CS-induced fluctuations \citep{Ringeval:2010ca, Ducout:2012it} e.g., through measuring CMB bispectrum, using Wavelet-based methods, or measurements of the CMB Minkowski functionals, leading to
$G\mu<8.8\times10^{-7}$,
$G\mu<7\times10^{-7}$ and $G\mu<7.8\times10^{-7}$, respectively
\citep{Hindmarsh:2009qk, Hindmarsh:2009es, Ade:2013xla,Regan:2015cfa},
\fi
Examples of real-space-based statistical methods are using the crossing statistics of CMB maps which yields the detectability level of $G\mu\gtrsim 4.0\times 10^{-9}$ for noise-free simulations \citep{Movahed:2010zq}, and using the unweighted Two-Point Correlation Function (TPCF) of CMB
peaks which gives $G\mu\gtrsim 1.2\times 10^{-8}$
for noiseless, 1'-resolution maps \citep{Movahed:2012zt}.
Some methods exploit the specific KS pattern, i.e, the line-like discontinuities of CMB fluctuations.
\cite{Stewart:2008zq} applied edge-detection
algorithms to find a minimum detectability of $G\mu \gtrsim
5.5\times 10^{-8}$ for a South Pole Telescope-like scenario and \cite{Hergt:2016xup} used wavelet
and curvelet methods to claim a detection level of
$G\mu\gtrsim 1.4 \times 10^{-7}$ for the third generation SPT.
In a recent paper we introduced a pipeline that applied various image processing and statistical tools to investigate the detectability of the CS network imprint on CMB temperature maps \citep{Sadr:2017hfm}. We claimed CS detectability for strings with $G\mu\gtrsim 4.3\times 10^{-10}$ for noiseless, $0.9'$-resolution, $7.2^\circ \times 7.2^\circ $ patches , and
with $G\mu\gtrsim 1.2 \times 10^{-7}$ for CMB-S4-like (II) experiments.
There are also the quite recent neural network-based approaches, giving $G\mu \gtrsim 2.3 \times 10^{-9}$ for noiseless arcminute-resolution maps \citep{Ciuca:2017wrk}. \cite{Ciuca:2017gca} use a convolutional neural network to locate the position of the CSs and get a limit of $G\mu \gtrsim 5 \times 10^{-9}$.
The tightest bound on the CS tension, $10^{-14} \le G\mu \le1.5 \times 10^{-10}$, comes from the gravitational wave emission of Nambu-Goto CS loops ~\citep{Ringeval:2017eww, Blanco-Pillado:2017oxo,Blanco-Pillado:2017rnf}. It should be noted, however, that these constraints strongly depend on the string microstructure. Abelian-Higgs field-theory simulations indicate that string loops decay mainly by the emission of massive radiation and emit less gravitational waves than estimated from Nambu-Goto simulations \citep{Hindmarsh:2017qff}, thus weaken the bounds.
Constraints from CMB maps are therefore more robust and conservative.
In this work we propose to use machine learning (ML)-based algorithms to search for the KS imprint of the CS network on CMB data. The goal is to develop a detection strategy capable of putting the tightest upper bound on the CS tension through optimally exploiting the available information accessible to the multi-scale pipeline of \cite{Sadr:2017hfm}. For this purpose, we choose to use two tree-based supervised classifiers: random forest (RF) and gradient boosting (GB).
In the following, after introducing our simulations (Section~\ref{simulation}), we explain our proposed strategy for CS detection from CMB maps, through compressing the map information into feature vectors (Section~\ref{Pre-proc}). The vectors are passed to tree-based ML methods to search for CS imprints (Section~\ref{Post-proc}).
We then describe in detail our proposed strategy in reporting the results in cases with possibly biased measurements (Section ~\ref{stdetection}).
Finally we present the results (Section~\ref{results}) and conclude with the discussion of the results (Section~\ref{sum_conc}).
\section{SIMULATIONS}\label{simulation}
Our simulations of the CMB sky closely follow \cite{Sadr:2017hfm} and consist of three components: the Gaussian contribution $G$ (including the primordial inflationary fluctuations, as well as the secondary lensing effect), CS-induced perturbations given by $G\mu \times S$, with $S$ describing the simulated normalized template for the CS signal ~\citep[using the Bennett-Bouchet-Ringeval
code,][]{Bennett:1990, Ringeval:2005kr} and $G\mu$ setting its overall amplitude, and the experimental noise $N$, described by white Gaussian random fields parametrized by the corresponding $SNR$ (signal to noise ratio). Our 2-Dimensional sky map $T(x,y)$ is thus given by:
\begin{equation}\label{fullmap}
T(x,y)=B[G(x,y)+G\mu \times S(x,y)]+N(x,y).
\end{equation}
$B$ denotes the beam function, here taken to be the model used in some ground-based observations \citep{Fraisse:2007nu,White:1997wq}, with an effective FWHM$\approx0.9'$, as well as a Plank-like Gaussian beam with FWHM$\approx5'$.
The simulated maps are square patches with sides $\Theta=7.2^\circ$, pixelized into squares with resolution $R=0.42'$. This yields a total of $1024\times 1024$ pixels. For more details see \cite{Sadr:2017hfm}.
\section{DETECTION STRATEGY I: Pre-processing} \label{Pre-proc}
\label{detect_strategy}
The CS detection algorithm of this work has two main steps. The pre-processing step compresses information from maps into feature vectors (each with $275 $ elements). The feature vectors are then passed to the classifier unit for classification. These two steps are briefly explained in this section and the following.
The feature extraction step employs three layers of image processors and statistical measures to produce a feature vector as the input for the learning unit (Figure~\ref{fig:pre}).
The first two layers aim at producing maps with enhanced CS detectability (Figure~\ref{fig:image_processing}), and the third layer quantifies the deviation of certain statistical measures of the map from those of the baseline model corresponding to null simulations with no CS imprints. These layers can be briefly described as: \\
(i) decomposers to disintegrate maps into scales relevant to the signal of interest.
The output is labeled as either {\it none} (corresponding to the full map), {\it WL} (or wavelet\footnote{The wavelet used here is the Daubechies db12 \citep{daubechies1990wavelet} with the mother function provided by the {\it PyWavelets} package, \url{https://github.com/PyWavelets}, and with the coefficients low-pass filtered with a threshold of $3$.}),
or one of the three curvelet components $C_5$, $C_6$ and $C_7$, corresponding to the three smallest scales\footnote{We used the {\it Pycurvelet} package \citep{Sadr:2017hfm} as our 2D, discrete version of the curvelet transform \citep{candes2006fast}. This package is the python-wrapped version of {\it CurveLab}, \url{http://www.curvelet.org/}. We chose $n_{\rm scales}=7$ and $n_{\rm angles}=10$ as the curvelet transformation parameters.}\citep{Sadr:2017hfm}.\\
(ii) various filters to enhance edges.
The output is labeled as either {\it none} (corresponding to the full map), {\it der} (or derivative), {\it lap} (or Laplacian), {\it sob} (or Sobel) or {\it sch} (or Scharr). \\
(iii) different statistical measures applied on the filtered, scale-decomposed maps. The measures are {\it pdf} (the probability distribution function), $M_2$ to $M_7$ (the second to seventh statistical moments), {\it cor} (the map correlation function), $\Psi_{pp}$ (the autocorrelation of peaks), $\Psi_{cc}$ (the autocorrelation of upcrossings) and $\Psi_{cp}$ (the peak-upcrossing cross-correlation).
For a thorough description see \cite{Sadr:2017hfm}. See also \citealt{rice1944mathematical,Bardeen:1985tr,Bond:1987ub,Ryden:1988rk,ryden1988collapse,Landy:1993yu,Matsubara:1995wj,Matsubara:2003yt,Ducout:2012it,Pogosyan:2008jb,Gay:2011wz,Codis:2013exa}.
For any given map, the final output of the pre-processor is a feature vector with $275$ elements, corresponding to all combinations of processors from each layer (Figure~\ref{fig:pre}).
The feature vector is then passed to the learning unit for classification, i.e. to RF and GB, to learn from simulations and to estimate $G\mu$ for new maps.
\begin{figure}
\begin{center}
\includegraphics[scale=0.32]{feature_vector_making.jpg}
\end{center}
\caption{A schematic view of the feature vector generation. For a CMB map (input on left side) it produces a $275$-dimensional feature vector, here presented as a $25\times11$ array (right side). The vector includes all possible combinations of decomposers, filters and statistical measures used in this work.\label{fig:pre}}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.18]{25.jpg}
\end{center}
\caption{All of the $25$ outputs of the image processing layers of the algorithm applied to a map with $G\mu=1.0\times10^{-7}$. The color scale is logarithmic. These are then passed to the 11 statistical measures, yielding the full set of 275 features.
}
\label{fig:image_processing}
\end{figure}
\section{DETECTION STRATEGY II: Learning Process} \label{Post-proc}
In this section we develop a machine-based algorithm to estimate the $G\mu$'s of given CMB maps using their feature vectors generated by the pre-processors.
We use supervised classifiers to build the data-driven model which maps the feature vector $\bf{x}$ to the predictor $\bf{y}$.
More specifically, we use the two powerful decision-tree-based ensemble methods: random forest or RF \citep{breiman2001random} and gradient boosting or GB \citep{friedman2001greedy} that combine a set of weak learners to improve the prediction performance \citep{quinlan1986induction,kearns1988thoughts,opitz1999popular,polikar2006ensemble,rokach2010ensemble}.
The tree, with its top-down greedy structure, starts from its root corresponding to the full set of observations and splits successfully into branches producing the prediction space. The branching process is based on dividing samples into homogeneous sets considering the most significant differentiator in input variables.
The RF classifier is based on growing many decision trees and its prediction will be the decision with the highest vote from all trees. The GB classifier, on the other hand, is based on the gradual improvement of a sequence of models toward better prediction, usually with decision trees as their base learners.
This is achieved through improving the model in a stage-wise manner by reducing an arbitrary loss function, here taken to be $\mathcal{L}=\sqrt{\operatorname{}\evdel{(\bf{y}-{\bf y}_{\rm fid})^2}}$
where $\bf{y}$ and ${\bf y}_{\rm fid}$ are the model and true (fiducial) values, respectively \citep{friedman2001greedy}.
It is important to note that not all features are expected to be independent or equally significant.
The tree-based learners report the importance of each feature based on the impact of its changes on the classifying parameter, here $G\mu$. This is called feature importance analysis and we will use it to find elements of the feature vectors with the most significant roles in CS detection.
Feature analysis can help to enormously reduce the dimension of the feature space
without a practical impact on the machine's performance \citep{bermingham2015application}.
Extra care needs to be taken in dimensionality reduction since a too small number of features may lead to experiment-dependent models with little generality.
In this work, we investigate the importance of features by averaging their number of occurrences among the top ten features through all machine learning models (MLMs).
Overfitting is a common problem in non-parametric algorithms due to their extreme flexibility.
In overfitting, noise and random correlations of the training set impact the model and result in reduced sensitivity when confronted with new observations that do not have those spurious features.
Cross-validation or CV, also used in this work, is a common powerful technique to avoid this issue \citep{kohavi1995study}.
The training set is partitioned into smaller training sets as well as a validation set. The model is made using the former while the latter plays the role of a new observation to assess how smoothly the method generalizes to new datasets.
Here we use a $K$-fold CV strategy where the original dataset is randomly divided into $K$ equal subsets with $K-1$ subsets forming the training sets and one the validation set. The process is repeated $K$ times to guarantee each subset is validated once.
We divide the $G\mu$ range used in this work ($2.5\times 10^{-11}<G\mu<5\times10^{-7}$ ) into $N_{\rm class}=18$ classes, with equal separation in $\ln G\mu$. A null class with $G\mu=0$ is also considered.
The machine is trained by applying the RF and GB algorithms
as MLMs to the feature vectors of $N_{\rm train}=1900$ CMB maps, corresponding to $N_{\rm sim}=100$ simulations for each class.
Our training unit has $N_{\rm L}=100$ MLMs with different seeds, each with a $K=10$-fold cross validation.
In each folding 90 maps are used for training and the remaining 10 maps of the class are used as the test set.
The results have been tested for robustness against various foldings.
To get a better control of the overfitting problem, we also generate a separate validation set with ten maps for each class.
This MLM can then be applied to any given CMB map to estimate its level of CS contribution.
The trained MLM assigns to any input map a probability vector $\vec{P}$, corresponding to the $G\mu$ of the classes ($\vec{G\mu}$).
We report the following (Bayesian) weighted average of the $G\mu$ as the {\it predicted} $G\mu$:
\begin{equation}\label{gmu_obs}
G\mu_{\rm pre}=\vec{P}.{\vec{G\mu}}.
\end{equation}
Classifiers often suffer from the limitation that the classes do not necessarily include the underlying parameter of an observation.
The above weighted averaging partially alleviates this problem. It should be noted that a relatively flat $\vec{P}$ would reflect the limited power of the trained MLM in discriminating between classes.
In the next section, we clarify in detail how we report the machine's output and translate it to the language of CS detection and measuring its contribution.
\section{Detecting strings or measuring their tension?}\label{stdetection}
Applying the detection strategy of the previous section yields a distribution of $G\mu_{\rm pre}$ for any of the $G\mu_{\rm fid}$ classes.
This distribution is ideally peaked around the $G\mu_{\rm fid}$, and its dispersion is sourced by cosmic variance, as well as contamination from primordial anisotropies and experimental noise. There is also a subdominant contribution to the fluctuations of $G\mu_{\rm pre}$ caused by the random seed of the MLMs which would decrease as the number of MLMs increases.
We define the minimum {\it detectable} $G\mu$, or $G\mu_{\rm det}$, as the minimum $G\mu$ whose distribution can be distinguished, with a maximum two-tail P-value of $0.0054$, from all other $G\mu$ classes, including the null class.
This minimum detection states there is a significant deviation in the map from the null hypothesis (with no string input). Note that this does not necessarily imply an unbiased measurement of the CS tension.
We therefore define the minimum {\it measurable} $G\mu$, or $G\mu_{\rm mes}$, as the minimum $G\mu$ above which the $G\mu_{\rm pre}$'s are unbiased. More precisely, $G\mu_{\rm mes}$ is the minimum $G\mu$ whose bias, defined as $G\mu_{\rm bf}-G\mu_{\rm fid}$ is smaller than one sigma. Here $G\mu_{\rm bf}$ is the best-fit $G\mu$ in the distribution of $G\mu_{\rm pre}$.
The next section presents the results of applying the proposed strategy to simulated CMB maps corresponding to several experimental cases \citep{Sadr:2017hfm}.
\section{RESULTS}\label{results}
In this work, we simulate CMB maps for five experimental setups: an ideal noise-free case, two CMB S4-like experiments, an ACT-like and a Planck-like case. The Planck-like simulations are smoothed with a Gaussian beam with $FWHM=5'$, while for the other four cases the effective beam is $FWHM=0.9'$.
The details of the experimental settings are given in \cite{Sadr:2017hfm}.
The feature vector for each map is generated through its pre-processing, which is then passed to the tree-based learning unit of the algorithm.
Figure~\ref{features} compares the feature importance of the three pre-processing layers for noise-free, ACT-like and Planck-like cases, and for the two tree-based algorithms considered in this work, namely RF and GB.
\begin{figure*}
\begin{center}
\includegraphics[scale=0.55]{star_plot.jpg}
\end{center}
\caption{Feature importance report: the average number of times each feature appeared among the top ten features, for each layer of the pre-processor, for the RF (top) and GB (bottom) learner. }
\label{features}
\end{figure*}
We find that the sixth and seventh curvelet components of the input maps have the dominant role in the first pre-processing layer for the noise-free case. That is expected since these components contain the small-scale information which is important for CS detection.
On the other hand, the instrumental noise contaminates the small scales most, making part of the CS signal in these higher modes inaccessible. That explains the more important roles of $C_5$ and $WL$
for ACT-like and Planck-like setups.
The middle panels of Figure~\ref{features} indicate that the classifiers have no significant preference for the filters. However, the ACT-like scenario should be excepted where Sobel seems to have a major impact on the results if the RF classifier is used.
In the third layer, the second moment of the filtered maps is clearly the main player in both RF and GB algorithms (right panels of Figure~\ref{features}).
The results from feature analysis could enormously decrease the computational cost of future analysis by helping to limit the training process to the feature subspace with most significant impact on the classification.
Table~\ref{tabledet} presents the predictions of our proposed tree-based detection strategy for the minimum detectable $G\mu_{\rm det}$.
\begin{table}
\centering
\caption{The minimum detectable $G\mu$, or $G\mu_{\rm det}$, for the two tree-based algorithms, GB and RF, and for the five experimental setups.}
\label{tabledet}
\begin{tabular}{ccc}
experiment & $ G\mu_{\rm det}({\rm GB}) ~~~$ & $ G\mu_{\rm det} ({\rm RF}) $ \\
\hline
noise-free & $4.3 \times 10^{-10}$ & $2.1 \times 10^{-10}$ \\
CMB-S4-like (II) & $1.2 \times 10^{-7}$ & $3.0 \times 10^{-8}$ \\
CMB-S4-like (I) & $1.2 \times 10^{-7}$ & $1.2 \times 10^{-7}$ \\
ACT-like & $1.2 \times 10^{-7}$ & $1.2 \times 10^{-7}$ \\
Planck-like & $7.0 \times 10^{-7}$ & $5.0 \times 10^{-7}$ \\
\hline
\end{tabular}
\end{table}
Similarly, Table~\ref{tablemes} reports the predicted $G\mu_{\rm mes}$'s for the various experiments considered in this work.
We find that for a Planck-like experiment the detection limit is $G\mu_{\rm det} \approx 5\times 10^{-7}$, while the $G\mu_{\rm mes}$ is above the upper bound of $G\mu$ range considered in the simulations of the training process. This means that our method is capable of detecting traces of CS with high significance for a $G\mu$ as low as $5\times 10^{-7}$. However, this method, with its current $G\mu$ range and fiducial classes, can not make an unbiased measurement of such small string tensions.
For a noise-free observation of the sky, the algorithm can distinguish the traces of CS networks down to $2.1 \times 10^{-10}$, and can correctly estimate the level of CS contribution for $G\mu$ above $G\mu_{\rm mes} \approx 3.6 \times 10^{-9}$.
Note that Table~\ref{tablemes} only reports the minimum measurable $G\mu$'s and not their associated errors. That is because the uncertainties in our measurements are dominated by the bin size of $G\mu$ classes, and not the statistical error. Therefore, for a class with $G\mu_{\rm fid}$, the uncertainty in the measurement is $\sigma_{G\mu}=\Delta \ln G\mu \times G\mu_{\rm fid}$, irrespective of the experiment.
\begin{table}
\centering
\caption{Similar to Table~\ref{tabledet} but for minimum measurable $G\mu$'s, or $G\mu_{\rm mes}$'s.}
\label{tablemes}
\begin{tabular}{ccc}
experiment & $ G\mu_{\rm mes}({\rm GB})~~~ $ & $ G\mu_{\rm mes} ({\rm RF}) $ \\
\hline
noise-free & $3.6 \times 10^{-9}$ & $3.6 \times 10^{-9}$ \\
CMB-S4-like (II) & $1.2 \times 10^{-7}$ & $1.2 \times 10^{-7}$ \\
CMB-S4-like (I) & $1.2 \times 10^{-7}$ & $2.5 \times 10^{-7}$ \\
ACT-like & $2.5 \times 10^{-7}$ & $2.5 \times 10^{-7}$ \\
Planck-like & $1.0 \times 10^{-6}$ & $1.0 \times 10^{-6}$ \\
\hline
\end{tabular}
\end{table}
\section{DISCUSSION}\label{sum_conc}
We proposed a tree-based machine learning algorithm for detecting and measuring the trace of CS-induced signals on CMB maps, simulated for various observational scenarios. Our simulations consisted of $1900$ maps, passed through the pre-processing unit of the algorithm to form the feature vectors, which are the inputs to the classifiers.
The simulations correspond to $18$ classes of $G\mu$ in the range $G\mu = 2.5\times 10^{-11}$ to $5\times 10^{-7}$, with equal spacing in $\ln G\mu$, and one null class.
Out of these maps, $90\%$ were used for training the classifiers (here taken to be random forest and gradient boosting) and the rest as test sets.
We performed feature analysis on the feature vectors to find the significance of the role of each feature for the classification. The results can be a major help in reducing the computational cost of future analysis by decreasing the dimension of the feature space and limiting the analysis to the most significant features.
As general results we can state that the scale of curvelet components should be matched to the effective resolution of experiments in the presence of experimental noise, larger-scale curvelet components are the more important decomposers. For filters it is difficult to make a definite recommendation,
while the second moment is the most important statistical measure in the classification process.
We find that, for each experimental case, three $G\mu$ regimes can be distinguished, whose boundaries marked by the $G\mu_{\rm det}$ and $G\mu_{\rm mes}$. For $G\mu$'s greater than $G\mu_{\rm mes}$, the algorithm is capable of measuring the CS contribution, with no bias and with an error determined by the bin size of that class. For $G\mu$'s smaller than $G\mu_{\rm mes}$ but larger than $G\mu_{\rm det}$, the algorithm can detect the signal, but cannot always make an unbiased measurement of its level.
For $G\mu$'s smaller than $G\mu_{\rm det}$, the CS signals are not reliably detected by the algorithm.
The predicted $G\mu_{\rm det}$ for a noise-free experiment is $2.1 \times 10^{-10}$.
This bound is, to the best of our knowledge, well below
the claimed detectability levels by other methods on noise-less maps. Compare, e.g., to the detectability bound of $G\mu\gtrsim 4.0\times 10^{-9}$ from crossing statistics \citep{Movahed:2010zq}, $G\mu\gtrsim 1.2\times 10^{-8}$ from the unweighted Two-Point Correlation Function of CMB peaks \citep{Movahed:2012zt}, $G\mu \gtrsim 6.3 \times 10^{-10}$ from Wavelet domain Bayesian denoising algorithm \citep{Hammond:2008fg}, $G\mu \gtrsim 2.3 \times 10^{-9}$ from the Neural network-based approaches \citep{Ciuca:2017wrk} and $G\mu\gtrsim 4.3\times 10^{-10}$ from a multi-scale pipeline for CS detection \citep{Sadr:2017hfm}. The minimum detectable tension in this work for a CMB-S4-like (II) experiment, $G\mu\gtrsim 3.0 \times 10^{-8}$, is a major improvement over the claimed detectability level by the above multi-scale pipeline, $G\mu\gtrsim 1.2 \times 10^{-7}$.
%
For a Planck-like case, the minimum detectable $G\mu$ is $5\times 10^{-7}$, comparable to the current upper bounds from Planck data \citep{Ade:2013xla}. Both classification methods seem to perform at a similar level, with RF appearing slightly more powerful based on the numbers in Tables \ref{tabledet} and \ref{tablemes}.
An important and immediate improvement to this work is to devise and apply debiasing techniques to remove the gap between $G\mu_{\rm mes}$ and $G\mu_{\rm det}$.
Given the continuous nature of the problem, one might also expect that using a regressor would improve the results.
That is because classifiers, by construction, are the method of choice in categorization problems while regressors in general are more suited for parameter estimation with continuous parameter ranges. It should be noted that using Bayesian averaging (Eq. \ref{gmu_obs}) in the parameter measurement step partially converted the classifiers of this work into regressors. We leave the full treatment of regressor-based algorithms for CS detection to future work.
\section*{Acknowledgements}
The authors would like to thank C. Ringeval, F. R. Bouchet and the Planck collaboration as they generously provided the simulations of the cosmic string maps used in this work. A. Vafaei acknowledges helpful discussions with Y. Fantaye. The numerical simulations were carried out on Baobab at the computing cluster of University of Geneva.
|
\section{Introduction}
Through the transition from Value at Risk (VaR) to Expected Shortfall (ES) as the primary market risk measure in the Basel Accord \citep{Basel2016}, there is a great demand for reliable methods for estimating, forecasting and backtesting the ES.
Formally, the ES at level $\tau \in (0,1)$ is defined as the mean of the returns smaller than the respective $\tau$-quantile (the VaR), where $\tau$ is usually chosen to be 2.5\% as stipulated by the Basel Accord.
The ES is introduced into the banking regulation because it overcomes several shortcomings of the VaR, such as not being coherent and its inability to capture tail risks beyond the $\tau$-quantile \citep{Artzner1999,Danielsson2001,Basel2013}.
In contrast to estimation and forecasting of ES where most of the existing models for the VaR can easily be adapted and generalized to the ES, such a generalization is unfortunately not as straight-forward for backtesting of ES forecasts \citep{Emmer2015}.
In general, backtesting of a risk measure is the process of testing whether given forecasts for this risk measure are correctly specified, which is carried out by comparing the history of the issued risk forecasts with the corresponding realized returns.
The primary reason for the difficulty to directly backtest ES is its non-elicitability and non-identifiability \citep{Weber2006, Gneiting2011b, Fissler2016, Fissler2016b} as consequently, there is no analog to the hit sequence which is the natural identification function of quantiles and which lies at the heart of most VaR backtests.%
\footnote{See \cite{Yamai2002, Kerkhof2004, Carver2013, Acerbi2014, Emmer2015, Ziegel2016, Fissler2016b, Nolde2017} for the ongoing discussion on backtestability of the ES.}
As a consequence, most of the proposed procedures in the growing literature on backtesting ES use indirect approaches, e.g. based on forecasts for the entire tail distribution or by linear approximations of the ES with VaR forecasts at several probability levels.
However, these approaches are in fact either joint backtests for a vector of risk measures such as the triple containing the VaR, the ES, and the volatility or even for the whole tail distribution \citep{Nolde2017}.
As the proposed backtests require further input parameters such as forecasts for the volatility, the tail distribution beyond some quantile, or even the entire distribution, they are not applicable for the regulatory authorities because this additional information is not reported by the financial institutions \citep{Aramonte2011,Basel2016,Basel2017}.
In this paper, we propose a novel backtest for ES forecasts which is based on a regression framework which models the conditional ES as a linear function, where we use financial returns as the response variable and ES forecasts as the explanatory variable including an intercept term.
For correct ES forecasts, the intercept and slope parameters should be equal to 0 and 1, respectively.
We use a Wald statistic to test for these parameter values, where we apply both, an asymptotic test using the covariance estimator introduced in \citet{DimiBayer2017} and a bootstrap hypothesis test.
We call this novel test the \textit{bivariate ESR backtest}.
This procedure is the first that backtests the risk measure ES stand-alone, i.e. the first that only uses ES forecasts as input parameters.\footnote{The backtests which come closest to our procedure in this regard are the exceedance residual backtests of \cite{McNeil2000} and the conditional coverage backtests of \cite{Nolde2017} which are in fact joint backtests for the VaR and the ES.}
Through this feature, our new test is the first backtest for the ES which is practicably applicable for the regulatory authorities who only have ES forecasts at hand.
Such regression-based forecast evaluation approaches are already used for testing mean forecasts \citep{MincerZarnowitz1969}, quantile forecasts \citep{Gaglianone2011,Guler2017}, and expectile forecasts \citep{Guler2017}.
In contrast to these functionals where regression techniques are easily available (see e.g. \citealp{Koenker1978}, \citealp{Efron1991}), estimating regression parameters for an ES specific regression equation is more difficult as the ES is not elicitable \citep{Gneiting2011b}.
We overcome this difficulty by estimating the parameters of a joint regression procedure for the quantile and the ES, recently proposed by \cite{DimiBayer2017}, \cite{Patton2017} and \cite{Barendse2017}.
We also introduce a second regression-based ES backtest by fixing the slope parameter in the regression to one, and by only estimating and testing the intercept term, where we call this test the \textit{intercept ESR backtest}.
This second backtest allows for both, one-sided and two-sided hypotheses which contrasts with the first backtest that only allows for a two-sided hypothesis as it is generally unclear how underestimated and overestimated ES forecasts respectively influence the slope and intercept parameters.
Because the capital requirements that the financial institutions must keep as a reserve depend on the reported risk forecasts, the market participants have an incentive to overestimate%
\footnote{%
Throughout the paper, we use the sign convention that losses are denoted by negative numbers and \textit{overestimation} of risk measures is meant in the mathematical sense, i.e. as reporting too large real numbers, which implies that the associated market risk is \textit{underestimated}.
}
the risk forecasts to minimize these expensive capital requirements.
In contrast, underestimation of the forecasts results in too conservative risk forecasts and larger capital reserves, which does not have to be punished by the regulatory authorities.
Thus, the regulators only have to prevent and penalize the overestimation of risk forecasts, which demonstrates the necessity of one-sided testing procedures.
For example, the currently applied traffic light system \citep{Basel1996} is in fact a one-sided VaR backtest.
As the bivariate ESR backtest, this intercept ESR test also has the desired characteristic to only require ES forecasts as input parameters and consequently is the first procedure in the literature that solely backtests the ES against a one-sided alternative.
We introduce several simulation setups to evaluate the empirical properties of our novel
ES backtests and compare them to the existing joint VaR and ES backtests of \citet{McNeil2000} and \citet{Nolde2017}.
In the first setup, we implement the classical size and power analysis for backtesting risk measures, where we simulate data stemming from a realistic data generating process and evaluate the empirical rejection frequencies of the backtests for forecasts stemming from the true and from some misspecified forecasting models.
In the second setup, we introduce a novel technique for evaluating the power of backtests for financial risk measures, where we continuously misspecify certain model parameters of the data generating process to obtain a continuum of alternative models with a gradually increasing degree of misspecification.
Misspecifying the different model parameters separately allows us to misspecify certain model characteristics (such as the reaction to shocks) in isolation, which permits a closer examination of the proposed backtesting procedures.
To the best of our knowledge, this evaluation technique is new to the literature.
From these simulations, we find that the bivariate and the intercept ESR backtests that we propose in this paper are reasonable sized, especially when the tests are applied using the bootstrap.
Moreover, they are more powerful than the existing backtests of \citet{McNeil2000} and \citet{Nolde2017} in almost all of the considered simulation designs for both, testing against one-sided and two-sided alternatives.
Notably, throughout all simulation designs, the two ESR backtests are able to detect the various different misspecifications of the forecasts.
In contrast, the existing backtests sometimes completely fail to detect certain misspecifications, for instance when the forecaster reports risk forecasts for a misspecified probability level.
The rest of this paper is organized as follows.
\Cref{sec:theory} introduces the theory of our new backtests, and \Cref{sec:existing_backtests} reviews the existing ES backtesting techniques.
\Cref{sec:monte_carlo} contains several simulation studies, and \Cref{sec:empirical_application} applies the backtests to the risk forecasts of the S\&P500 index.
\Cref{sec:conclusion} concludes.
\section{Theory}
\label{sec:theory}
\subsection{Setup and Notation}
Let us consider a stochastic process
\begin{align}
\label{eqn:DefinitionStochasticProcess}
{Z} = \bigl\{ {Z}_t: \Omega \to \mathbb{R}^{k+1}, k \in \mathbb{N}, t = 1,\dots, T \bigr\},
\end{align}
defined on some complete probability space $\bigl( \Omega,\, \mathcal{F},\, \mathbb{P} \bigr)$, with the filtration $\mathcal{F} = \bigl\{ \mathcal{F}_t,\, t = 1,\, \dots T \bigr\}$ and $\mathcal{F}_t = \sigma \{{Z}_s,\, s\le t \}$ for all $t = 1,\dots,T$.
We partition the stochastic process ${Z}_t = (Y_t, X_t)$, where $Y_t$ is an absolutely continuous random variable of interest and $X_t$ is a $k$-dimensional vector of explanatory variables.
We denote the conditional cumulative distribution function of $Y_t$ given the past information $\mathcal{F}_{t-1}$ by $F_t(y) = \mathbb{P} ( Y_t \le y \mid \mathcal{F}_{t-1})$ and the corresponding probability density function by $f_t$.
Whenever they exist, the mean and the variance of $F_t$ are denoted by $\mathbb{E}_t[\cdot]$ and $\operatorname{Var}_t(\cdot)$.
For financial applications, the variable $Y_t$ denotes the daily log returns of a financial asset (for instance, a stock or a portfolio), i.e. $Y_t = \log P_t - \log P_{t-1}$, where $P_t$ denotes the price of the asset at day $t = 1,\ldots,T$.
This means that throughout this paper, we use the sign convention that positive returns denote profits, and negative returns denote losses.
The vector $X_t$ contains further variables that are used to produce forecasts for certain functionals (usually risk measures) of the random variable $Y_t$.
We are interested in testing whether forecasts for a certain $d$-dimensional, $d \in \mathbb{N}$ functional (risk measure) $\rho = \rho(F_t)$ of the conditional distribution $F_t$ are correctly specified.
For that, we define the most frequently used functionals for financial risk management in the following.
The conditional quantile of $Y_t$ given the information set $\mathcal{F}_{t-1}$ at level $\tau \in (0,1)$ is defined as
\begin{align}
Q_\tau \bigl(Y_t \mid \mathcal{F}_{t-1} \bigr) = F_t^{-1}(\tau) = \inf \bigl\{ y \in \mathbb{R}: F_{t}(y) \ge \tau \bigr\},
\end{align}
which is called the VaR at level $\tau$ in financial applications.
Furthermore, we define the functional ES at level $\tau$ of $Y_t$ given $\mathcal{F}_{t-1}$ as
\begin{align}
\ES_\tau \bigl( Y_t \mid \mathcal{F}_{t-1} \bigr) = \frac{1}{\tau} \int_0^\tau F_t^{-1}(s) \, \mathrm{d}s.
\end{align}
If the distribution function $F_t$ is continuous at its $\tau$-quantile, this definition can be simplified to the truncated tail mean of $Y_t$,
\begin{align}
\ES_\tau \bigl( Y_t \mid \mathcal{F}_{t-1} \bigr) = \mathbb{E}_t \left[ Y_t \mid Y_t \leq Q_\tau \bigl( Y_t \mid \mathcal{F}_{t-1} \bigr) \right].
\end{align}
We denote an $\mathcal{F}_{t-1}$-measurable one-step-ahead forecast for day $t$ for the risk measure $\rho$ of the distribution $F_t$, stemming from some external forecaster or from some given forecasting model%
\footnote{For recent overviews on VaR and ES forecasting approaches, see \citet{Komunjer2013} and \citet{Nadarajah2014}.}
by $\hat \rho_t = \hat \rho_t(\mathcal{F}_{t-1})$.
Following this notation, we denote forecasts for the $\tau$-VaR by $\hat v_t$ and for the $\tau$-ES by $\hat e_t$ for some fixed level $\tau \in (0,1)$.
For simplicity of the notation, we drop the dependence on $\tau$ as it is a fixed quantity.
As both, the incentive of the forecaster and the underlying method used to generate the forecasts are in general unknown, these forecasts are not necessarily correctly specified.
The focus of this paper is to develop statistical tests for correctness of a given series of forecasts $\bigl\{ \hat \rho_t,\, t=1,\dots,T \bigr\}$ for the risk measure $\rho$ relative to the realized return series $\bigl\{ Y_t,\, t=1,\dots,T \bigr\}$.
This is in the literature usually referred to as \textit{backtesting} of the risk measure $\rho$ without strictly defining this terminology.
We provide such a definition in the following.
The core message of this definition is that besides the realized return series, a backtest for some risk measure is only allowed to require forecasts for this risk measure as input parameters.
\begin{definition}
\label{def::ProperBacktest}
A \textit{backtest} for the series of forecasts $\bigl\{ \hat \rho_t,\, t=1,\dots,T \bigr\}$ for the $d$-dimensional risk measure (functional) $\rho$ relative to the realized return series $\bigl\{ Y_t,\, t=1,\dots,T \bigr\}$ is a function
\begin{align}
f: \mathbb{R}^T \times \mathbb{R}^{T \times d} &\to (0,1),
\end{align}
which maps the return and forecast series onto the respective $p$-value of the test.
\end{definition}
This strict differentiation becomes relevant in the context of backtesting ES as, in contrast to the existing VaR backtests, the recently proposed ES backtests require further input parameters such as forecasts for the VaR, the volatility, or the entire tail distribution.
The demand for these further quantities induces the following practical problems.
First, the regulatory authorities who rely on such backtesting methods do not necessarily receive forecasts from the financial institutions for the additional information required by these tests, which makes such backtests inapplicable for the regulatory authorities.
Second, a rejection of the tests does not necessarily imply that the ES is misspecified, but that the forecasts for any of the input components are misspecified.
Consequently, these tests are in fact not backtests for the ES, but rather backtests for vectors of risk measures, or the entire tail distribution.
The two novel regression based ES backtests we propose in the next section are the first backtests for the ES which follow Definition \ref{def::ProperBacktest} as they only require forecasts for the ES.
This makes these tests the first ES backtests in this sense.
\subsection{The bivariate ESR Backtest}
\label{sec::mz_backtest}
We propose a new backtest for the risk measure ES that tests whether a series of ES forecasts $\{\hat e_t, t = 1,\dots T \}$, stemming from some external forecaster or forecasting model is correctly specified relative to a series of in due course realized returns $\{Y_t, t = 1,\dots T\}$.
For that, we regress the returns $Y_t$ on the forecasts $\hat e_{t}$ including an intercept term by using a regression equation designed specifically for the functional ES,
\begin{align}
\label{eqn::BacktestRegressionEquation}
Y_t = \alpha + \beta \hat e_{t} + u^e_t,
\end{align}
where $\ES_{\tau}(u^e_t \mid \mathcal{F}_{t-1}) = 0$.
Given the structure in (\ref{eqn::BacktestRegressionEquation}) and since $\hat e_t$ is generated by using the information set $\mathcal{F}_{t-1}$, this condition on the error term is equivalent to
\begin{align}
\label{eqn::BacktestFunctionalEquation}
\ES_{\tau} \left( Y_t \mid \mathcal{F}_{t-1} \right) = \alpha + \beta \hat e_{t}.
\end{align}
We then test the hypothesis
\begin{align}
\label{eqn::NullHypothethis}
\mathbb{H}_0: ( \alpha, \beta ) = (0,1) \qquad \text{against} \qquad \mathbb{H}_1: (\alpha, \beta ) \neq (0,1).
\end{align}
Under $\mathbb{H}_0$, the ES forecasts are correctly specified as it holds that $\hat e_t = \ES_{\tau} \left( Y_t \mid \mathcal{F}_{t-1} \right)$
\footnote{
Given that the ES forecasts are correctly specified, i.e. $\hat e_t = \ES_{\tau} \left( Y_t \mid \mathcal{F}_{t-1} \right)$, the correct specification condition (\ref{eqn::BacktestFunctionalEquation}) is equivalent to $\alpha = (1-\beta) \hat e_t$.
This results in the remark of \cite{Holden1990}, who claim that the null hypothesis, given in (\ref{eqn::NullHypothethis}) is only a sufficient, but not a necessary condition for correctly specified forecasts as $\alpha = (1-\beta) \hat e_t$ is the required necessary condition.
However, this more general condition implies that the forecasts $\hat e_t$ are constant for all $t = 1,\dots,T$, which is highly unrealistic given the dynamic nature of financial time series.
Consequently, we employ the hypotheses given in (\ref{eqn::NullHypothethis}) for our backtesting procedure.
}
Since this ES backtest is based on a regression procedure and simultaneously tests the parameters $\alpha$ and $\beta$, we call this test the \textit{bivariate ESR backtest}.
As outlined in \citet{DimiBayer2017}, estimating the parameters $(\alpha, \beta)$ in (\ref{eqn::BacktestRegressionEquation}) by M- or Z/GMM-estimation stand-alone using a semiparametric method without specifying the full conditional distribution of the error term $u^e_t$ is not possible since the functional ES is not elicitable \citep{Gneiting2011b}.
However, these parameters can be estimated through a regression technique which jointly models a regression equation for the quantile and the ES proposed by \citet{DimiBayer2017}, \cite{Patton2017} and \cite{Barendse2017}, which we briefly review in Appendix \ref{sec::JointQuantileESReg}.
We use this joint regression framework for the semiparametric estimation of (\ref{eqn::BacktestRegressionEquation}) by estimating the joint system,
\begin{align}
Y_t &= \gamma + \delta \hat e_{t} + u^q_t, \label{eqn::BacktestJointRegressionEquationQuantile} \\
Y_t &= \alpha + \beta \hat e_{t} + u^e_t, \label{eqn::BacktestJointRegressionEquationES}
\end{align}
where $Q_{\tau}(u^q_t \mid \mathcal{F}_{t-1}) = 0$. and $\ES_{\tau}(u^e_t \mid \mathcal{F}_{t-1}) = 0$.
This means we choose $Y_t$ as the response variable and $\big(1,\hat e_t\big)$ as explanatory variables for this regression procedure.
Because our null hypothesis is based on only testing the parameters $(\alpha,\beta)$ in the ES regression equation given in (\ref{eqn::BacktestJointRegressionEquationES}), we use a Wald statistic which only incorporates these parameters,
\begin{align}
\label{eq:wald_statistic}
T_{\text{ESR}} = \left(\big(\hat\alpha, \hat\beta\big)' - (0,1)'\right)' \, \widehat \Sigma_{\ES}^{-1} \, \left( \big(\hat \alpha, \hat\beta \big)' - (0,1)'\right)',
\end{align}
where $\widehat \Sigma_{\ES}$ is an estimator for the (asymptotic) covariance matrix of the M-estimator of the parameters $(\alpha,\beta)$.
\cite{Patton2017} show consistency and asymptotic normality for the M-estimator of the regression parameters for $\alpha$-mixing time series.
Using this, and given that $\hat \Sigma_{\ES} \stackrel{\mathbb{P}}{\to} \Sigma_{\ES}$, the test statistic asymptotically follows a $\chi^2$ distribution with two degrees of freedom,
\begin{align}
T_{\text{ESR}} \stackrel{d}{\to} \chi^2_2.
\end{align}
We implement both, backtests based on estimates for the asymptotic covariance matrix and based on the bootstrap \citep{Efron1979}.
For the asymptotic version, we employ the \textit{scl-sp} covariance estimation method discussed in \cite{DimiBayer2017}.
We further implement the bootstrap hypothesis testing procedure%
\footnote{This approach provides an asymptotic refinement, i.e. the error in the rejection probability decreases faster compared to both, the asymptotic distribution and the bootstrapped covariance matrices for the test, see e.g. \citet{MacKinnon2009}. In the construction of confidence intervals, this is also known as the percentile-$t$ method.}
where in each bootstrap sample, we estimate the model parameters and the asymptotic covariance matrix to compute a total of $B = 1000$ bootstrap Wald statistics as in (\ref{eq:wald_statistic}), where the bootstrap estimates are centered around the estimate for the original sample.
Finally, the bootstrap $p$-value is the share of the $B$ bootstrap test statistics that are larger than or equal to the test statistic for the original sample.
As neither the underlying loss function of the M-estimator, given in (\ref{eqn::RegressionLossFunction}), nor the asymptotic covariance, given in (\ref{eqn::Lambda11}) - (\ref{eqn::AsyCovMatrixC22}), depend on the temporal ordering of the pairs $(Y_t,\hat e_t)$, we apply the iid bootstrap resampling technique of \citet{Efron1979}.
Similar tests are already implemented for backtesting of forecasts for the mean \citep{MincerZarnowitz1969}, for quantiles \citep{Gaglianone2011} and for expectiles \citep{Guler2017}.
As these functionals are elicitable, M-estimation of regression parameters for mean, quantile \citep{Koenker1978} and expectile \citep{Efron1991} regressions is straight-forward.
This section shows that introducing the same concept for backtesting ES forecasts is possible, but technically more demanding as we have to estimate the regression parameters through a joint system as given in (\ref{eqn::BacktestJointRegressionEquationQuantile}) and (\ref{eqn::BacktestJointRegressionEquationES}).
\subsection{The One-sided Intercept ESR Backtest}
\label{sec:OneSidedTest}
The bivariate ESR backtest introduced in the previous section only allows for testing two-sided hypotheses as specified in (\ref{eqn::NullHypothethis}) because it is generally unclear how too small or too large risk forecasts influence the parameters $\alpha$ and $\beta$.
Because the capital requirements the financial institutions have to keep as a reserve depend on the reported risk forecasts, the market participants have an incentive to overestimate\footnote{
Overestimation of a risk measure is meant in the mathematical sense, which means reporting a too large real number. As the ES forecasts are strictly negative, this implies in fact underestimation of the associated market risk.}
the risk forecasts in order to keep as little capital requirements as possible.
In contrast, underestimation of the risk measures results in too conservative risk forecasts and consequently higher capital requirements, which does not have to be punished by the regulatory authorities.\footnote{One could interpret the higher capital requirements as a punishment for too conservative risk forecasts.}
Thus, the regulatory authorities only have to prevent and consequently penalize the overestimation of risk measures, which can be done by using one-sided backtesting procedures.
For example, the traffic light system \citep{Basel1996}, currently implemented in the Basel Accords, is in fact a one-sided backtest for the hit ratios of VaR forecasts.
Consequently, we also introduce a regression-based backtesting procedure for the ES that allows for both, specifying one-sided and two-sided hypotheses.
This backtest is based on regressing the forecast errors, $Y_t - \hat e_t$, on an intercept term only,
\begin{align}
\label{eqn::RegressionEquationInterceptTest}
Y_t - \hat e_t = \alpha + u^e_t,
\end{align}
where $\ES_\tau (u^e_t\mid\mathcal{F}_{t-1}) = 0$ and testing whether the parameter $\alpha$ is zero.
Estimation of the parameter $\alpha$ in (\ref{eqn::RegressionEquationInterceptTest}) is carried out by computing the empirical ES of the forecast errors $Y_t - \hat e_t$.
By using this restricted regression equation, we can define a one-sided and a two-sided hypothesis,
\begin{align}
\begin{split}
\mathbb{H}_0^{2s}: \alpha = 0 \qquad & \text{against} \qquad \mathbb{H}_1^{2s}: \alpha \neq 0, \quad \text{and} \\
\mathbb{H}_0^{1s}: \alpha \geq 0 \qquad & \text{against} \qquad \mathbb{H}_1^{1s}: \alpha < 0,
\end{split}
\end{align}
which we test by using a $t$-test based on the asymptotic covariance and based on the bootstrap procedure described above.
Note that this is equivalent to setting the slope parameter of the bivariate ESR test given in (\ref{eqn::BacktestRegressionEquation}) to one and only estimating and testing the intercept term.
Consequently, we call this backtest the \textit{intercept ESR backtest}.
Both, the bivariate and the intercept ESR backtests proposed in this paper are implemented in our R package \texttt{esback} \citep{Bayer2017a}.
\section{Existing Backtests}
\label{sec:existing_backtests}
Over the past two decades and especially driven by the recent transition from VaR to ES in the Basel regulatory framework \citep{Basel2016, Basel2017}, a large literature on backtesting the ES has emerged.
These backtests are usually introduced with financial regulators in mind who need to verify the risk forecasts they receive from the financial institutions.
To be applicable for the regulatory authorities, a backtest for the risk measure ES thus follows Definition \ref{def::ProperBacktest} and only requires the observed return series and the ES forecasts as input variables.
However, many of the proposed backtests for the ES fail to have this property.
In particular, several tests require the whole return distribution \citep{Berkowitz2001,Kerkhof2004,Wong2008,Acerbi2014,Graham2014}, the cumulative violation process $\int_{0}^{\tau} \mathds{1}_{\{Y_t \leq \hat v_t(p) \}}\, \mathrm{d} p$ \citep{Costanzino2015,Emmer2015,Du2017,Kratz2016}, the volatility \citep{McNeil2000,Nolde2017,Righi2013,Righi2015}, or the VaR \citep{McNeil2000,Nolde2017} in addition to the ES forecasts.
However, this information (except the VaR) is not reported by the financial institutions and therefore, most of these tests can not be used by the regulators \citep{Aramonte2011,Basel2017}.
Furthermore, when more information than solely the ES forecasts is used for backtesting, a rejection of the null hypothesis does not necessarily imply that the ES forecasts are wrong.
More precisely, a rejection of the null implies that \textit{some} component of the input parameters is incorrect \citep[cf.][]{Nolde2017}.
A related concern is raised by \citet{Aramonte2011}, who note that financial institutions could be tempted to submit forecasts of this additional information chosen such that the tests have particularly low power, so that correctness of their internal model (or their issued ES forecasts) is not doubted.
Strictly following Definition \ref{def::ProperBacktest}, we would have to distinguish between backtests for the ES and joint backtests for the pair VaR and ES.
However, as the ES is strongly intertwined with the VaR (through its definition and through the joint elicitability), sensible forecasts for the ES are based on correctly specified VaR forecasts.
Consequently, it is reasonable to backtest both quantities jointly and thus, we compare the performance of our ES backtests to existing joint VaR and ES backtests in the literature.
In the following, we describe the exceedance residual test of \citet{McNeil2000} and the conditional calibration tests of \citet{Nolde2017} in more detail, since both have versions that only require VaR forecasts in addition to the ES.
\subsection{Testing the Exceedance Residuals}
\label{sec:er_test}
One of the first and still most frequently used tests for the ES is the exceedance residual (ER) backtest of \citet{McNeil2000}.
This approach is based on the ES residuals that exceed the VaR, $er_t = \big( Y_t - \hat e_t \big) \mathds{1}_{\{Y_t \le \hat v_t\}}$, which form a martingale difference sequence given that $\hat v_t$ and $\hat e_t$ are the true $\mathcal{F}_{t-1}$-measurable quantile and ES respectively.
\citet{McNeil2000} furthermore consider a second version that uses exceedance residuals standardized by the volatility, i.e. $er_t / \hat \sigma_t$.
This backtest tests whether the expected value of the (raw or standardized) ER, $\mu = \mathbb{E} [ er_t]$, is zero using the estimate $\hat{\mu} = 1 / (\sum_{t=1}^T \mathds{1}_{\{Y_t \le \hat v_t\}}) \sum_{t=1}^T er_t$ in conjunction with a bootstrap hypothesis test \citep[see][p.~224]{Efron1994}.
In the original paper, \citet{McNeil2000} propose to test $\mu$ against the one-sided alternative that $\mu$ is negative, i.e. that the ES is overestimated.
However, in this paper we discuss both, tests based on one-sided and two-sided hypotheses, so that in addition to the original proposal, we also include a two-sided test,
\begin{align}
\begin{split}
\mathbb{H}_0^{2s}: \mu = 0 \qquad & \text{against} \qquad \mathbb{H}_1^{2s}: \mu \neq 0, \quad \text{and} \\
\mathbb{H}_0^{1s}: \mu \geq 0 \qquad & \text{against} \qquad \mathbb{H}_1^{1s}: \mu < 0.
\end{split}
\end{align}
By Definition \ref{def::ProperBacktest}, the test using the standardized ER is in fact a joint backtest for the triple VaR, ES and volatility, whereas the test using the raw ER is a joint backtest for the pair VaR and ES.
In light of the discussion above, the test using the raw ER is therefore preferred.
Nevertheless, in the simulation studies and the empirical application we apply both approaches and find that they perform alike.
Even though the intercept ESR test introduced in \Cref{sec:OneSidedTest} and the ER backtest appear to be similar, there is a subtle difference between the two test statistics.
For the intercept ESR test, we compute the empirical ES of $Y_t - \hat e_t$, i.e. the average of $Y_t - \hat e_t$ given that $Y_t - \hat e_t$ is smaller than its empirical $\tau$-quantile.
In contrast, the ER backtest computes the average of $Y_t - \hat e_t$, given that $Y_t$ is smaller than the respective forecast for its $\tau$-quantile $\hat v_t$.
This difference seems marginal, but it has severe consequences for the theoretical and empirical properties of the tests.
As we can write $\hat{\mu} = 1/\tilde T \sum_{t=1}^T Y_t \mathds{1}_{\{Y_t \le \hat v_t\}} - 1/\tilde T \sum_{t=1}^T \hat e_t \mathds{1}_{\{Y_t \le \hat v_t\}}$, where $\tilde T = \sum_{t=1}^T \mathds{1}_{\{Y_t \le \hat v_t\}}$, the ER backtest in fact compares the empirical average of $Y_t$ truncated at $\hat v_t$ to the average ES forecast $\hat e_t$, whenever there is a VaR violation.
Thus, this backtest rejects whenever the distance/relation between the VaR and ES-forecasts is incorrect.
However, simultaneous misspecifications of both forecasts, such as e.g. generated by misspecification of the volatility process in location scale models cannot be detected.
In the same spirit, the ER backtest cannot distinguish between correct forecasts for the VaR and ES at level $\tau$ and (correct) forecasts for a misspecified probability level $\tilde{\tau} \neq \tau$, as the given level $\tau$ does not influence the ER test statistic at all.
In contrast, by computing the empirical $\tau$-quantile of $Y_t - \hat e_t$ (instead of using the forecast $\hat v_t$), the intercept ESR test does not suffer from these shortcomings as can be observed in the simulation results in Section \ref{sec::ContinuousMisspecification}.
\subsection{Conditional Calibration Backtests}
\citet{Nolde2017} introduce the concept of conditional calibration (CC) based on strict identification functions (also known as moment conditions or estimating equations) of the respective functional and show that many classical backtests for risk measures can be unified using this concept.
For the pair VaR and ES at level $\tau \in (0,1)$, they choose the strict identification function
\begin{align}
V(Y,\, v,\, e) =
\begin{pmatrix}
\tau - \mathds{1}_{\left\{Y \leq v\right\}}\\
e - v + \mathds{1}_{\left\{Y \leq v\right\}}(v - Y) / \tau
\end{pmatrix},
\end{align}
whose expectation is zero if and only if $v$ and $e$ equal the true VaR and ES of $Y$ respectively.
The CC backtest for forecasts for the VaR, $\hat v_t$ and for the ES, $\hat e_t$ is based on the hypotheses
\begin{align}
\begin{split}
\mathbb{H}_0^{2s}: \mathbb{E} \big[ V(Y_t, \, \hat v_t,\, \hat e_t ) \mid \mathcal{F}_{t-1} \big] = 0 \qquad & \text{against} \qquad \mathbb{E} \big[ V(Y_t,\, \hat v_t,\, \hat e_t ) \mid \mathcal{F}_{t-1} \big] \neq 0, \quad \text{and} \\
\mathbb{H}_0^{1s}: \mathbb{E} \big[ V(Y_t, \, \hat v_t, \, \hat e_t ) \mid \mathcal{F}_{t-1} \big] \geq 0 \qquad & \text{against} \qquad \mathbb{E} \big[ V(Y_t, \, \hat v_t, \, \hat e_t ) \mid \mathcal{F}_{t-1} \big] < 0,
\end{split}
\end{align}
component-wise and almost surely for all $t= 1,\ldots,T$.
This is equivalent to testing $\mathbb{E} \big[ h_t' V(Y_t, \hat v_t, \hat e_t ) \big] = 0$ for all $\mathcal{F}_{t-1}$ measurable $\mathbb{R}^2$-valued functions $h_t$.
As this is infeasible, \citet{Nolde2017} propose to use an $\mathcal{F}_{t-1}$-measurable sequence of $q \times 2$-matrices of test functions $\boldsymbol{h}_t$ for some $q \in \mathbb{N}$ and to use the Wald-type test statistic
\begin{align}
T_{\text{CC}} = T \left(\frac{1}{T}\sum_{t=1}^T \boldsymbol{h}_t V \left(Y_t, \hat v_t,\hat e_t \right) \right)' \widehat{\Omega}^{-1} \left(\frac{1}{T}\sum_{t=1}^T \boldsymbol{h}_t V \left(Y_t, \hat v_t,\hat e_t \right) \right),
\end{align}
where $
\widehat{\Omega} = \frac{1}{T} \sum_{t=1}^{T}
\left( \boldsymbol{h}_t V \left(Y_t, \hat v_t,\hat e_t \right) \right)
\left( \boldsymbol{h}_t V \left(Y_t, \hat v_t,\hat e_t \right) \right)'
$
is a consistent estimator of the covariance of the $q$-dimensional vector $\boldsymbol{h}_t V \left(Y_t, \hat v_t,\hat e_t \right)$.
Under $\mathbb{H}_0$, the test statistic asymptotically follows a $\chi^2_q$ distribution with $q$ degrees of freedom.
\citet{Nolde2017} propose two versions of this test, where the first uses no information besides the risk forecasts (termed \textit{simple CC test}), and where the second additionally requires volatility forecasts (termed \textit{general CC} test).
For the simple CC test, the test function is the identity matrix, $\boldsymbol{h}_t = I_2,$ for both, the one- and two-sided hypotheses.
For the general CC test, they propose to choose
\begin{align}
\boldsymbol{h}_t = \hat \sigma_t \big(\left(\hat e_t - \hat v_t \right) / \tau,\, 1 \big) \quad \text{and} \quad \boldsymbol{h}_t = \begin{pmatrix}
1 & |\hat v_t| & 0 & 0 \\
0 & 0 & 1 & \hat \sigma_t^{-1}
\end{pmatrix}',
\end{align}
for the two-sided and for the one-sided test, respectively, where $\hat \sigma_t$ is a forecast for the volatility.
As with the standardized ER test, the general CC test is strictly speaking a backtest for the triple VaR, ES, and volatility, but we nevertheless include both versions in our empirical comparisons.
We provide implementations of the two ESR backtests proposed in this paper, both ER backtests of \citet{McNeil2000} and both CC backtests of \citet{Nolde2017} in our R package \texttt{esback} \citep{Bayer2017a}.
\section{Monte-Carlo Simulations}
\label{sec:monte_carlo}
In this section, we evaluate the empirical performance of our proposed ES backtests and compare them to the tests of \citet{McNeil2000} and \citet{Nolde2017}.
For that, we first assess the empirical size of the tests, defined as the rejection frequency of the test under the null hypothesis and which should equal the nominal significance level.
Then, we analyze the empirical power of the tests which is defined as the rejection frequency of forecasts stemming from some misspecified model and which is optimally as close to one as possible.
This comparison is conducted using two different approaches.
The first, presented in \Cref{sec::TraditionalSizePower}, follows the typical strategy in the related literature of first assessing the size of the backtests with some realistic location-scale data generating process (DGP), followed by an evaluation of the power by backtesting forecasts stemming from overly simplified model, in our case the Historical Simulation model.
In the second setup, presented in \Cref{sec::ContinuousMisspecification}, we continuously misspecify certain parameters of the true model and thereby obtain alternative models with a continuously increasing degree of misspecification.
This approach of evaluating backtests has two main advantages.
First, we obtain power curves which can be used to draw conclusions how an increasing model misspecification influences the test decisions.
Second, misspecifying the different model parameters separately allows us to misspecify certain model characteristics while leaving the remaining model unchanged.
Thus, we can evaluate the capability of the backtests to identify certain model misspecifications, which allows for a closer examination of the backtesting procedures.
\subsection{Traditional Size and Power Comparisons}
\label{sec::TraditionalSizePower}
For the first simulation study, we simulate returns from an EGARCH(1,1) model \citep{Nelson1991} with $t$-distributed innovations, where the parameter values are calibrated using daily returns of the S$\&$P 500 index.
This model is given by
\begin{align}
\begin{aligned}
\label{eq:sim_1_dgp}
Y_t & = \sigma_t z_t, \\
\log(\sigma_t^2) & = -0.160 - 0.125 z_{t-1} + 0.130 \left(|z_{t-1}| - E [|z_{t-1}|] \right) + 0.983\log (\sigma_{t-1}^2 ), \\
z_t & \sim t_{7.24},
\end{aligned}
\end{align}
where $z_t$ are innovations stemming from the standardized Student-$t$ distribution with $7.24$ degrees of freedom.
As the EGARCH model is highly flexible and due to its calibrated parameter values, this DGP accurately replicates the distributional properties of daily financial returns.
Conditional VaR and ES forecasts at level $\tau$ for the DGP in (\ref{eq:sim_1_dgp}) are given by
\begin{align}\label{eq:sim_1_true}
\hat{v}_t = \hat \sigma_t q_z(\tau) \quad \text{and} \quad \hat{e}_t = \hat \sigma_t \xi_z(\tau),
\end{align}
where $\hat \sigma_t$ is a volatility forecast generated through the model given in (\ref{eq:sim_1_dgp}) and $q_z(\tau)$ and $\xi_z(\tau)$ are the $\tau$-quantile, respectively the $\tau$-ES of the innovations $z_t$.
For the following size and power analysis of the backtests, we simulate the process (\ref{eq:sim_1_dgp}) 10,000 times with sample sizes of 250, 500, 1000, 2500, and 5000 observations and 250 additional pre-sample values required for the power analysis.
As stipulated by the Basel Accords, we forecast the two risk measures for the probability level $\tau=2.5\%$.
In this part of the study, we focus on two-sided hypotheses and defer the one-sided case to \Cref{sec:one_sided_alternative}.
\Cref{tab:mc1_size} presents the empirical sizes of the considered backtests for the different sample sizes and for nominal test sizes of 1\%, 5\%, and 10\%.
We find that in large samples, all backtests display rejection rates close to the respective nominal sizes.
However, in small samples all backtests are oversized and they differ with respect to their speed of convergence.
Looking at the individual tests in greater detail, we find that especially the tests relying on asymptotic quantities (i.e. the ESR and CC tests) are substantially oversized in small samples and converge to the nominal sizes comparably slow.
However, by using the bootstrap for the intercept and bivariate ESR tests (indicated by (b) in the table), the empirical sizes are much closer to the nominal sizes in small samples than for the asymptotic versions.
Comparing the intercept and the bivariate ESR test, we find that the former has better size properties in small samples, presumably because less parameters need be estimated and the covariance is simpler.
Furthermore, also the two ER tests (which also rely on bootstrapping) exhibit good empirical sizes and there are hardly any differences between the raw and the standardized version.
\newcommand{The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}{The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\begin{table}[htb!]\footnotesize\centering
\caption{Empirical sizes of the backtests}
\label{tab:mc1_size}
\begin{tabularx}{\linewidth}{XX *{8}{r}}
\toprule
\theadl{Nominal \\ Size} & \theadl{Sample \\ Size} & \thead{bivariate \\ ESR (b)} & \thead{bivariate \\ ESR} & \thead{intercept \\ ESR (b)} & \thead{intercept \\ ESR} & \thead{General \\ CC} & \thead{Simple \\ CC} & \thead{Std. \\ ER} & \thead{ER} \\
\midrule
\input{empirical_sizes_Oracle_3_0.01_2s.txt}
\midrule
\input{empirical_sizes_Oracle_3_0.05_2s.txt}
\midrule
\input{empirical_sizes_Oracle_3_0.1_2s.txt}
\bottomrule\addlinespace
\multicolumn{10}{p{.97\linewidth}}{\textit{Notes:} The table reports the empirical sizes of the backtests for the EGARCH(1,1)-$t$ process given in (\ref{eq:sim_1_dgp}). The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\end{tabularx}
\end{table}
\noindent
For a comparison of the power of the backtests, we evaluate their ability to reject the null hypothesis for risk models producing incorrect ES forecasts.
We utilize the Historical Simulation approach which forecasts the VaR and ES by using their empirical counterparts from previous trading days,
\begin{align}
\hat{v}_t = \widehat{Q}_{\tau} \left(Y_{t-1}, \, Y_{t-2}, \cdots, Y_{t-w}\right) \quad \text{and} \quad
\hat{e}_t = \frac{1}{\sum_{i=1}^{w} \mathds{1}_{\left\{Y_{t-i} \leq \hat{v}_{t-i}\right\}}} \sum_{i=1}^{w} Y_{t-i} \cdot \mathds{1}_{\left\{Y_{t-i} \leq \hat{v}_{t-i} \right\}},
\end{align}
where $\widehat{Q}_{\tau}$ is the empirical $\tau$-quantile and $w$ is the length of a rolling window, that we set to 250, i.e. one year of data.
Since the standardized ER and the general CC backtest both require forecasts of the volatility, we estimate this quantity with the sample standard deviation of the returns over the same rolling window.
For a meaningful and fair comparison of the power of the backtests to reject the null hypothesis, we compare the \textit{size-adjusted power}%
\footnote{%
A comparison of the \textit{raw power}, i.e. the raw rejection rate of the null hypotheses, could be misleading due to the differences in the empirical sizes of the backtests.
In particular, an oversized test would exhibit unrealistically large rejection rates.
For completeness, \Cref{tab:mc1_raw_power_Historical_Simulation} reports the raw power of the tests.
}
of the backtests \citep{Lloyd2005}.
For this, the original critical values of the tests are either increased or decreased such that the rejection frequencies of the true model equal the nominal test sizes.
The size-adjusted power is then given by the rejection frequencies of the alternative models using these modified critical values.
\Cref{fig:roc_plots} contains the size-adjusted power of the backtests for all empirical sizes in the unit interval for a sample size of 1000.%
\footnote{
These plots are known as the receiver operating characteristic (ROC) curves and origin from the psychometrics literature \citep{Lloyd2005}.
They are an effective presentation method for general binary classification tasks such as hypothesis testing as they show the size-adjusted power simultaneously for all significance levels.
}
The black line depicts the case of equal empirical size and power, which can be seen as a lower bound for any reasonable test: whenever the power is below this line, randomly guessing the test decision is more accurate than performing the test.
We observe that both, the bivariate and intercept version of the ESR backtest clearly dominate the others at almost all empirical sizes, including the most relevant region of test sizes between 1\% and 10\%.
Furthermore, the ESR tests using asymptotic quantities are slightly more powerful than their bootstrap versions (indicated by (b)), but the loss in power is negligible compared to the improvements in the sizes we find in \Cref{tab:mc1_size}.
\begin{figure}[htb]
\begin{subfigure}{.5\linewidth}
\includegraphics{roc_design_3_Historical_Simulation_n_1000_2s.pdf}
\caption{Size-adjusted Power}
\label{fig:roc_plots}
\end{subfigure}
\begin{subfigure}{.5\linewidth}
\includegraphics{auc_design_3_Historical_Simulation_2s.pdf}
\caption{Partial Area Under the Curve}
\label{fig:pauc_plots}
\end{subfigure}
\caption{Size-adjusted power and Partial Area Under the Curve plots against Historical Simulation for a sample size of 1000 days. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\label{fig:MC1_power_plots}
\end{figure}
In order to present results for all considered sample sizes in condensed form for the relevant area of empirical sizes between 1\% and 10\%, we summarize the size-adjusted power by the partial area under the curve (PAUC), as proposed by \citet{Lloyd2005}.
For that, we numerically compute the area under each power curve for the empirical sizes between 1\% and 10\% which is thus the power to reject a false model averaged over the considered test sizes.
In \Cref{fig:pauc_plots}, we present the PAUC for all backtests and sample sizes.
As expected, the average power increases with the sample size, so that using more information leads to more reliable decisions about the quality of a forecast.
We find that for all considered sample sizes, the ESR backtests dominate the other testing approaches.
As a robustness check for these findings, we repeat this simulation experiment with the DGP used by \citet{Gaglianone2011} and find that the results, presented in Appendix \ref{sec:robustness_check}, are similar to the findings of this section.
\subsection{Continuous Model Misspecification}
\label{sec::ContinuousMisspecification}
In the second simulation study, we use a GARCH(1,1) model with standardized Student-$t$ distributed innovations,
\begin{align}\label{eq:mc2_model}
\begin{split}
Y_t & =\sigma_t z_t, \\
\sigma_t^2 & = \gamma_0 + \gamma_1 Y_{t-1}^2 + \gamma_2 \sigma_{t-1}^2, \\
z_t & \sim t_{\nu},
\end{split}
\end{align}
with the parameter values $\gamma_0 = 0.01$, $\gamma_1 = 0.1$, $\gamma_2 = 0.85$, and $\nu=5$ for the true model.
For the analysis of the backtests, we simulate 10,000 times from this model with a fixed sample size of 2500 observations and consider the probability level $\tau=2.5\%$ for the VaR and the ES.
\begin{table}[htb!]
\footnotesize\centering
\caption{Empirical sizes for the second simulation study.}
\label{tab:mc2_size}
\begin{tabularx}{\linewidth}{X *{8}{r}}
\toprule
\theadl{Null Hypothesis} & \thead{bivariate \\ ESR (b)} & \thead{bivariate \\ ESR} & \thead{intercept \\ ESR (b)} & \thead{intercept \\ ESR} & \thead{General \\ CC} & \thead{Simple \\ CC} & \thead{Std. \\ ER} & \thead{ER} \\
\midrule
\input{size.txt}
\bottomrule
\addlinespace
\multicolumn{9}{p{.97\linewidth}}{\textit{Notes:} This table shows the empirical sizes of the backtests for the GARCH(1,1)-$t$ model given in (\ref{eq:mc2_model}), for a nominal test size of 5\% and for both, one-sided and two-sided hypotheses. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}. ~Note that the bivariate ESR test does not permit a one-sided hypothesis and therefore, we only present sizes for the two-sided hypothesis.}
\end{tabularx}
\end{table}
\Cref{tab:mc2_size} presents the empirical sizes of the backtests for a nominal size of 5\% for both, the two- and one-sided hypotheses.
As in the first simulation study, we find that most of the backtests are reasonably sized with rejection frequencies close to the nominal value.
However, the two CC tests reject the true model slightly too often in the two-sided, respectively too rarely in the one-sided case.
For a detailed analysis of the power of the backtests, we continuously misspecify the true model according to the following five designs:
\begin{enumerate}[label=(\alph*), labelindent=0em, labelsep=0.1cm, leftmargin=*]
\item
We misspecify how the conditional variance reacts to the squared returns by varying the ARCH parameter $\gamma_1$.
We choose $\tilde{\gamma}_1$ between 0.03 and 0.2 and let $\tilde{\gamma}_2 = 0.95 - \tilde{\gamma}_1$, such that the persistence of the GARCH process remains constant.
When $\tilde{\gamma}_1 < {\gamma}_1$, there is too little variation in the ES forecasts due to the reduced response to shocks and the GARCH process approaches a constant volatility model.
\item
We alter the unconditional variance of the GARCH process $\mathbb{E}[\sigma_t^2] = \gamma_0 / (1 - \gamma_1 - \gamma_2)$ between $0.5$ and $0.01$ by varying the parameter $\gamma_0$ while holding $\gamma_1$ and $\gamma_2$ constant.
Since the conditional variance is a weighted combination of the unconditional variance, the past squared returns and the past conditional variance, this change implies that the ES is always underestimated when the unconditional variance is larger as its true value, and vice versa.
\item
We vary the persistence of shocks between $0.9$ and $0.999$ by setting $\tilde{\gamma}_1 = c\cdot\gamma_1$ and $\tilde{\gamma}_2 = c\cdot\gamma_2$ for a constant $c$ that we vary, and $\tilde{\gamma}_0 = \mathbb{E} [\sigma_t^2] (1-\tilde{\gamma}_1 - \tilde{\gamma}_2)$ to keep the unconditional variance constant.
A higher persistence causes a stronger and longer reaction to shocks.
\item
We vary the degrees of freedom of the underlying Student-$t$ distribution between 3 and $\infty$.
Since the conditional variance is unaffected, this modification implies a relative horizontal shift of the ES forecasts.
\item
We misspecify the probability level $\tilde{\tau}$ of the ES forecasts between $0.5\%$ and $5\%$.
This represents the scenario that a forecaster submits (accidentally or on purpose) predictions for some level $\tilde{\tau} \neq \tau$.
Similar to changing the degrees of freedom, this modification implies a relative horizontal shift of the ES forecasts.
\end{enumerate}
As an illustrative example of these misspecifications, \Cref{fig:mc2_example_series_1,fig:mc2_example_series_2,fig:mc2_example_series_3,fig:mc2_example_series_4,fig:mc2_example_series_5} show 250 realizations of the returns of the true DGP (\ref{eq:mc2_model}), together with the corresponding ES forecasts of the true model (black dashed line) and two models following the parameter misspecifications described in the points (a) to (e) above.
\begin{figure}[htb!]\centering
\begin{subfigure}{.49\linewidth}
\caption{Changing the reaction to the squared returns}
\label{fig:mc2_example_series_1}
\includegraphics{1_e}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the unconditional variance}
\label{fig:mc2_example_series_2}
\includegraphics{2_e}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the persistence}
\label{fig:mc2_example_series_3}
\includegraphics{3_e}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the degrees of freedom}
\label{fig:mc2_example_series_4}
\includegraphics{4_e}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the probability level}
\label{fig:mc2_example_series_5}
\includegraphics{5_e}
\end{subfigure}
\caption{These plots show exemplary simulated return series with 250 observations for the DGP given in (\ref{eq:mc2_model}) and for the five parameter misspecifications illustrated in the points (a) - (e) in Section \ref{sec::ContinuousMisspecification}. In each of the subfigures, the black dashed line corresponds to the true model parameters.}
\label{fig:mc2_example_series}
\end{figure}
\begin{figure}[htb!]\centering
\begin{subfigure}{.49\linewidth}
\caption{Changing the ARCH parameter}
\label{fig:mc2_rejection_rates_1}
\includegraphics{1_2s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the unconditional variance}
\label{fig:mc2_rejection_rates_2}
\includegraphics{2_2s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the persistence}
\label{fig:mc2_rejection_rates_3}
\includegraphics{3_2s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the degrees of freedom}
\label{fig:mc2_rejection_rates_4}
\includegraphics{4_2s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the probability level}
\label{fig:mc2_rejection_rates_5}
\includegraphics{5_2s}
\end{subfigure}
\caption{Size-adjusted rejection rates for various types of misspecification. The gray vertical line depicts the true model. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\label{fig:mc2_rejection_rates}
\end{figure}
We present the size-adjusted rejection rates plotted against the respective misspecified parameters for these five designs in \Cref{fig:mc2_rejection_rates_1,fig:mc2_rejection_rates_2,fig:mc2_rejection_rates_3,fig:mc2_rejection_rates_4,fig:mc2_rejection_rates_5}.
The true model is indicated by the gray vertical line and, induced by the results of \Cref{fig:mc2_example_series}, the x-axis is oriented such that too small (too risky) ES forecasts are on the left side of the true model.\footnote{Notice that this inequality of the forecast magnitude only holds on average in the cases of \Cref{fig:mc2_rejection_rates_1,fig:mc2_rejection_rates_3} whereas it holds strictly for \Cref{fig:mc2_rejection_rates_2,fig:mc2_rejection_rates_4,fig:mc2_rejection_rates_5}.}
Even though there is no backtest that dominates the others throughout all considered designs, several conclusions can be drawn from this figure.
\noindent
(1) Overall, the bivariate and intercept ESR tests perform similar and in four out of the five considered designs, their performance is superior compared to the general CC and both ER backtesting approaches. (\Cref{fig:mc2_rejection_rates_1,fig:mc2_rejection_rates_2,fig:mc2_rejection_rates_3,fig:mc2_rejection_rates_5}).
The ESR backtests outperform the competitors especially when we misspecify the volatility dynamics of the underlying GARCH process (\Cref{fig:mc2_rejection_rates_1,fig:mc2_rejection_rates_2,fig:mc2_rejection_rates_3}).
This shows that, in contrast to the existing approaches, our new ESR backtests can be used to detect misspecifications in the dynamics used to construct the ES forecasts which go beyond level shifts.
\noindent
(2) The application of the bootstrap for our ESR tests mainly affects the empirical sizes whereas the empirical size-adjusted power of the asymptotic and the bootstrap ESR tests is similar throughout all designs.
\noindent
(3) The two ER tests (and the general CC test that is constructed to be similar to the ER backtest) cannot discriminate between forecasts for the VaR and ES issued through misspecified volatility processes (\Cref{fig:mc2_rejection_rates_1,fig:mc2_rejection_rates_2,fig:mc2_rejection_rates_3}) and through misspecified probability levels $\tilde{\tau} \neq \tau$ (\Cref{fig:mc2_rejection_rates_5}).
This confirms the theoretical results discussed in \Cref{sec:er_test} that these backtests only reject misspecifications which affect the relation (distance) between the VaR and ES forecasts.
In contrast, these backtests perform well in the case of misspecified tails of the residual distribution, which affects the relative distance between the VaR and ES forecasts (\Cref{fig:mc2_rejection_rates_4}).
If these backtests would be used by the regulatory authorities, banks could submit joint VaR and ES forecasts for some level $\tilde{\tau} > \tau$ or some (too small) volatility process in order to minimize their capital requirements without facing the risk of being detected by these backtests.
In comparison, our intercept ESR backtest which is similar to the ER backtests by construction is clearly able to identify these misspecified probability levels.
\noindent
(4) Throughout all five misspecifications, the simple CC backtest also exhibits good power properties, similar to our proposed backtests.
However, our two ESR backtests exhibit much better size properties (see \Cref{tab:mc1_size,tab:mc2_size}) and in contrast to the simple CC test, they do not fail to reject the Historical Simulation forecasts in the first simulation study (see \Cref{fig:MC1_power_plots}).
Together with the results from the first simulation study, these findings show that our proposed ESR backtests are a powerful choice for backtesting ES forecasts.
They are reasonably sized and exhibit good power properties against a variety of misspecifications.
Notably, in contrast to the existing backtests, there is no single type of misspecification where our ESR tests are unable to discriminate between forecasts of the true and the misspecified models.
\subsection{Testing one-sided hypotheses}
\label{sec:one_sided_alternative}
For the regulatory authorities, testing against a one-sided alternative might be more meaningful than the two-sided version we consider in the previous section.
Holding more money than stipulated in the Basel accords is no concern for regulators as it is only important that banks keep enough monetary reserves to cover the risk from their market activities.
As all backtests (with exception of the bivariate ESR test) allow for testing against one-sided alternatives, we assess their ability to reject the null hypothesis that the issued ES forecasts are smaller or equal to the true ES, i.e. that the associated market risk is not underestimated.
In \Cref{fig:mc2_rejection_rates_1_1s,fig:mc2_rejection_rates_2_1s,fig:mc2_rejection_rates_3_1s,fig:mc2_rejection_rates_4_1s,fig:mc2_rejection_rates_5_1s}, we present the size-adjusted rejection rates for the one-sided versions of the considered backtests and for the five continuous parameter misspecifications described in the points (a) - (e) from the previous section.
The structure of these figures is analog to the two-sided case where the x-axis is oriented such that too small (too risky) ES forecasts are on the left side of the true model (vertical gray line).
As it can be seen in \Cref{fig:mc2_example_series_1,fig:mc2_example_series_2,fig:mc2_example_series_3,fig:mc2_example_series_4,fig:mc2_example_series_5}, the five modifications of the true model exhibit clear patterns when they are over-, respectively underestimating the true ES, where the overestimation holds strictly for the cases (b), (d) and (e) and on average for the cases (a) and (c).
Thus, the one-sided backtests should only reject the null hypothesis for ES forecasts that overestimate the true ES, i.e. which are on the right side of the true model in \Cref{fig:mc2_rejection_rates_1_1s,fig:mc2_rejection_rates_2_1s,fig:mc2_rejection_rates_3_1s,fig:mc2_rejection_rates_4_1s,fig:mc2_rejection_rates_5_1s}.
We find that our intercept ESR backtest (in the asymptotic and the bootstrap version) is reasonably sized (compare \Cref{tab:mc2_size}) and clearly dominates the ER and the CC tests in terms of their power in four out of five misspecification designs.
Only when changing the degrees of freedom, the ER test is slightly more powerful than the intercept ESR test.
Surprisingly, we see that in four out of the five cases, the one-sided CC tests (both, the simple and the general version) also reject too small (too risky) ES forecasts, even though these should not be rejected by
the specifications of the one-sided tests.%
\footnote{We verified our implementation of the CC tests with the codes provided by \citet{Nolde2017} at \url{https://github.com/nnolde/Elicitability-and-Backtesting/}.}
Furthermore, as for the two-sided tests, both ER backtests fail to detect misspecifications of the underlying volatility process and of the underlying probability level.
Summarizing these results, the proposed intercept ESR backtest is a powerful backtest with good size properties for testing one-sided hypotheses which clearly dominates the existing one-sided backtesting techniques in the literature.
\begin{figure}[htb!]\centering
\begin{subfigure}{.49\linewidth}
\caption{Changing the reaction to the squared returns}
\label{fig:mc2_rejection_rates_1_1s}
\includegraphics{1_1s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the unconditional variance}
\label{fig:mc2_rejection_rates_2_1s}
\includegraphics{2_1s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the persistence}
\label{fig:mc2_rejection_rates_3_1s}
\includegraphics{3_1s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the degrees of freedom}
\label{fig:mc2_rejection_rates_4_1s}
\includegraphics{4_1s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the probability level}
\label{fig:mc2_rejection_rates_5_1s}
\includegraphics{5_1s}
\end{subfigure}
\caption{Size-adjusted rejection rates for various types of misspecification with a one-sided hypothesis. The gray vertical line depicts the true model. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\label{fig:mc2_rejection_rates_1s}
\end{figure}
\section{Empirical application}
\label{sec:empirical_application}
In the empirical application, we predict the market risk for the daily close-to-close log-returns of the S\&P500 index for the time period from January 3, 2000 to October 18, 2017, totaling up to 4478 days.
We predict the ES (and the VaR for application of the existing tests) for this return series using 10 different risk models.
The first two are the Historical Simulation estimated with a rolling window of 250 days and RiskMetrics.
The other 8 models follow the volatility specifications of the GARCH(1,1) model and the asymmetric GJR-GARCH(1,1) model of \citet{Glosten1993} and use four different assumptions on the conditional distribution of the innovations.
These are the standard normal distribution (abbreviated by N), the standardized Student-$t$ ($t$), the standardized skewed Student-$t$ (skew-$t$) and the semi-parametric filtered historical simulation approach (FHS) of \citet{Barone-Adesi1999}, where the quantile, respectively the ES of the innovations is estimated from the standardized returns.
We estimate these 8 models on a rolling window of 1000 days.
\Cref{tab:empirical_application_results} presents the $p$-values of the different ES backtests (for the two-sided hypothesis), the average losses of the strictly consistent 0-homogeneous loss function for the pair VaR and ES%
\footnote{This is in fact the loss function given in (\ref{eqn::RegressionLossFunction}), applied to a scenario of forecast comparison.} \citep{Fissler2016}, and the $p$-value of the Model Confidence Set (MCS) of \citet{Hansen2011} applied to this loss function.
With the MCS $p$-values, we can determine a set of models having equal predictive ability at a certain significance level with respect to the losses.
The models are sorted according to the average loss.
\begin{table}[htb!]\footnotesize
\tabcolsep5pt
\caption{Results of the empirical application.}
\label{tab:empirical_application_results}
\begin{tabularx}{\linewidth}{X *{8}{r} r rr}
\toprule
\theadl{Model} & \thead{biv. \\ ESR (b)} & \thead{biv \\ ESR} & \thead{int. \\ ESR (b)} & \thead{int. \\ ESR} & \thead{General \\ CC} & \thead{Simple \\ CC} & \thead{Std. \\ ER} & \thead{ER} && \thead{Mean \\ Loss} & \thead{MCS \\ $p$-value}\\
\midrule
\input{empirical_application_table_2s.txt}
\bottomrule
\addlinespace
\multicolumn{12}{p{.97\linewidth}}{\textit{Notes:} In this table, $p$-values smaller than 5\% are printed bold-faced and the models are sorted by the average loss. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}. We compute the MCS $p$-values using the $R$-statistic of \citet{Hansen2011}.
}
\end{tabularx}
\end{table}
From this table we can draw several conclusions.
First, the MCS rejects 7 out of 10 models at the 5\% significance level, i.e. only 3 models have equal predictive power with respect to the joint loss function.
These three (GJR-GARCH-skew-t, -FHS, -$t$) share the same assumption on the volatility process and only differ with respect to the assumption on the innovations.
Moreover, for these three models the null hypothesis of correct forecasts is not rejected by almost all backtests at the 5\% significance level.
Thus, the backtests and the MCS agree on which models predict the ES (and the VaR) well.
Second, the CC and ER tests reject less forecasts at the 5\% significance level than the two ESR backtests, which reflects the findings of the simulation studies where these backtests are often less powerful than our ESR tests.
In particular, the null hypothesis is not rejected for the Historical Simulation model, although this approach yields large losses.
Third, incorporating leverage into the volatility dynamics appears to be important, since mainly the models using the GJR-GARCH are not rejected by the backtests.
Additionally, it is crucial to consider models with flexible tails, e.g. by using the skewed Student-$t$ or the FHS approach, since the models based on conditionally normally distributed returns are collectively rejected by the backtests and the MCS.
\section{Conclusion}
\label{sec:conclusion}
In this paper, we introduce two novel backtests for ES forecasts which regress the realized returns on the issued ES forecasts using an appropriate regression method for the ES introduced in \cite{DimiBayer2017, Patton2017, Barendse2017} and test the resulting parameter estimates.
We introduce a bivariate version, denoted \textit{bivariate ESR backtest}, where we test the intercept and the slope parameters for zero and one, and an intercept version, denoted \textit{intercept ESR backtest}, that only incorporates an intercept term being estimated and tested for zero.
The motivation for the latter test is the possibility to specify a one-sided hypothesis that is particularly relevant for the regulatory authorities.
These backtests can be interpreted as ES-specific versions of the classical \citet{MincerZarnowitz1969} test for evaluating mean forecasts.
A unique feature of the backtests proposed in this paper is that they solely require forecasts for the ES and are consequently the first backtests for the ES stand-alone.
In contrast to that, a common drawback of the existing backtests is that they need forecasts of further input parameters, such as the VaR, the volatility, the tail distribution or even the whole return distribution.
Using more information than the ES forecasts is problematic for two reasons.
First, these tests are not applicable for the regulatory authorities, who receive forecasts of the ES, but not of the additional information required by these tests.
Second, rejecting the null hypothesis does not necessarily imply that the ES forecasts are incorrect as the rejection can be a result of a false prediction of any of the input parameters.
In several simulation studies, we assess the empirical size and power properties of our proposed backtests and compare them to the approaches of \citet{McNeil2000} and \citet{Nolde2017}, which jointly backtest the VaR and the ES.
We find that our regression-based tests are reasonably sized, especially when they are applied using the bootstrap.
Moreover, in almost all simulation designs our two proposed backtests are more powerful than the existing tests.
The backtests from the literature are often not able to distinguish between forecasts stemming from the true model and some misspecified model, for instance when we consider a misspecified volatility process or a wrong probability level of the ES.
In contrast to that, our two ESR backtests detect the misspecifications in all considered simulation experiments.
We provide an implementation of our backtests and of several approaches from the literature in the \texttt{esback} package for R \citep{Bayer2017a}.
This paper contributes to the ongoing discussion about which risk measure is the best in practice in the following way.
As the VaR is criticized for not being subadditive and for not capturing tail risks beyond itself, the recent literature proposes both, the ES and expectiles as alternative risk measures.
Expectiles are suggested as they are coherent, elicitable and are able to capture extreme risks beyond the VaR and thus, they simultaneously overcome the drawbacks of the VaR and the ES \citep{Bellini2104, Ziegel2016}. Unfortunately, as opposed to the VaR and ES, they lack a visual and intuitive interpretation \citep{Emmer2015}.
In contrast, the ES is mainly criticized for its theoretical deficiencies of being not elicitable and not (only with difficulties) backtestable.
However, starting with the joint elicitability result of VaR and ES of \cite{Fissler2016}, there is a growing body of literature using this result for a regression procedure \citep{DimiBayer2017,Barendse2017,Patton2017} and for relative forecast comparison \citep{Fissler2016b,Nolde2017}, which is extended by this paper through introducing the ESR backtests, which are the first sensible backtests for the ES stand-alone.
This shows that, even though technically more demanding, the ES can be modeled, evaluated and backtested in the same way as quantiles and expectiles.
Combining this with its ability to capture extreme tail risks and its intuitive visual interpretation, the ES is an appropriate candidate for being the standard risk measure in practice.
\section*{Acknowledgments}
We thank Tobias Fissler, Lyudmila Grigoryeva, Roxana Halbleib, Phillip Heiler, Ekaterina Kazak, Winfried Pohlmeier, James Taylor, and Johanna Ziegel for suggestions which inspired some results of this paper.
Financial support by the Heidelberg Academy of Sciences and Humanities (HAW) within the project ``Analyzing, Measuring and Forecasting Financial Risks by means of High-Frequency Data'' and by the German Research Foundation (DFG) within the research group ``Robust Risk Measures in Real Time Settings''.
The authors acknowledge support by the state of Baden-Württemberg through bwHPC.
\FloatBarrier
\begin{appendices}
\renewcommand\thefigure{\thesection.\arabic{figure}}
\renewcommand\thetable{\thesection.\arabic{table}}
\section{The Joint Quantile and ES Regression Technique}
\label{sec::JointQuantileESReg}
Assume we have a response variable $Y_t$ and a $k$-dimensional vector of covariates $X_t$ following the definition of the stochastic process in (\ref{eqn:DefinitionStochasticProcess}) and assume that we are interested in a linear regression technique for the ES at level $\tau \in (0,1)$,
\begin{align}
\label{eqn::ESRegressionEquationTheory}
Y_t = X_t'\theta^e + u_t^e,
\end{align}
where $\ES_{\tau}(u_t^e \mid \mathcal{F}_{t-1}) = 0$ for all $t = 1,\dots, T$, $T \ge 1$.
This means that we model the conditional ES through a linear function, $\ES_{\tau}(Y_t \mid \mathcal{F}_{t-1}) = X_t'\theta^e$.
As outlined in \citet{DimiBayer2017}, \cite{Patton2017} and \cite{Barendse2017}, it is infeasible to estimate the parameter vector $\theta^e$ by M- or Z-/GMM- estimation using a semiparametric method without specifying the full conditional distribution of the error term $u_t^e$.
The underlying reason for this is that the functional ES is not elicitable \citep{Gneiting2011b}, i.e. there exists no strictly consistent loss function for the ES, which could be minimized in M-estimation of the regression parameters.%
\footnote{In contrast, classical mean regression is built on minimizing the squared loss function, which is a strictly consistent loss function for the mean.
Furthermore, quantile regression \citep{Koenker2005book} is estimated by minimizing the asymmetric linear loss function, a strictly consistent loss function for quantiles and expectile regression \citep{Efron1991} by the asymmetric squared error loss, a strictly consistent loss function for expectiles.}
However, as the ES and the quantile at common probability level $\tau$ are jointly elicitable \citep{Fissler2016}, the parameters $\theta^e$ in (\ref{eqn::ESRegressionEquationTheory}) can be estimated by jointly modeling a regression equation for the quantile and for the ES,
\begin{align}
\label{eqn::JointRegressionEquationTheory}
Y_t & = X_t'\theta^q + u_t^q \qquad \text{and} \\
Y_t & = X_t'\theta^e + u_t^e,
\end{align}
where $Q_{\tau}(u_t^q \mid \mathcal{F}_{t-1}) = 0$ and $\ES_{\tau}(u_t^e \mid \mathcal{F}_{t-1}) = 0$ for all $t = 1,\dots, T$, $T \ge 1$.
Here, $\theta = \big(\theta^q, \theta^e\big)$ denotes the $2k$-dimensional vector of regression parameters of the joint model and the quantile and ES equations are modelled through the separate $k-$dimensional parameter vectors $\theta^q$ and $\theta^e$.
The M-estimator of the regression parameters $\theta$ is obtained by
\begin{align}
\widehat{\theta}_T = \argmin_{\theta} \frac{1}{T} \sum_{t=1}^{T} \rho(Y_t,X_t,\theta),
\end{align}
where the loss function\footnote{As shown by \citet{DimiBayer2017}, consistent and asymptotically normal M-estimation of these regression parameters can be obtained by employing loss functions from a whole class of functions, originally introduced by \citet{Fissler2016} in the context of forecast evaluation.
However, consensus seems to emerge on the 0-homogeneous loss function presented in (\ref{eqn::RegressionLossFunction}), see e.g. \citet{DimiBayer2017, Taylor2017, Patton2017, Barendse2017} and \citet{Nolde2017}.} is given by
\begin{align}
\label{eqn::RegressionLossFunction}
\rho(Y_t,X_t,\theta) = \frac{1}{-X_t' \theta^e} \left( X_t' \theta^e - X_t' \theta^q + \frac{(X_t' \theta^q - Y_t) \mathds{1}_{\{Y_t \le X_t' \theta^q\}}}{\tau} \right) + \log(-X_t' \theta^e).
\end{align}
Consistency and the asymptotic normality of the M-estimator of $\theta$ is shown by \cite{Patton2017} for an $\alpha$-mixing stochastic process $Z_t = (Y_t, X_t)$.
Under the further technical conditions in Assumption 1 and 2 in \cite{Patton2017}, it holds that
\begin{align}
\label{eqn::AsymptoticNormality}
\sqrt{T} C_T^{-1/2} \Lambda_T \big( \hat \theta_T - \theta_0 \big) \stackrel{d}{\longrightarrow} \mathcal{N} \left(0, I \right),
\end{align}
where $\theta_0$ denotes the unknown true parameter value and where
\begin{align}
\Lambda_T = \begin{pmatrix} \Lambda_{11,T} & 0 \\ 0 & \Lambda_{22,T} \end{pmatrix} \qquad \text{and} \qquad
C_T = \begin{pmatrix} C_{11,T} & C_{12,T} \\ C_{21,T} & C_{22,T} \end{pmatrix},
\end{align}
with
\begin{align}
\Lambda_{11,T} &= -\frac{1}{T} \sum_{t=1}^T \frac{1}{\alpha} \mathbb{E} \left[ (X_t X_t') f_t(X_t' \theta^q_0) /(X_t' \theta^e_0) \right], \label{eqn::Lambda11} \\
\Lambda_{22,T} &= \frac{1}{T} \sum_{t=1}^T \mathbb{E} \left[ (X_t X_t') / (X_t' \theta^e_0)^2 \right], \label{eqn::Lambda22} \\
C_{11,T} &= \frac{1}{T} \sum_{t=1}^T \frac{1-\alpha}{\alpha} \mathbb{E} \left[ (X_t X_t') / (X_t' \theta^e_0)^2 \right], \label{eqn::AsyCovMatrixC11} \\
C_{12,T} &= C_{21,T} = - \frac{1}{T} \sum_{t=1}^T \frac{1-\alpha}{\alpha} \mathbb{E} \left[ (X_t X_t') \big( X_t' \theta^q_0 - X_t' \theta^e_0 \big) / (X_t' \theta^e_0)^3 \right], \label{eqn::AsyCovMatrixC12} \\
C_{22,T} &= \frac{1}{T} \sum_{t=1}^T \mathbb{E} \left[ (X_t X_t') / (X_t' \theta^e_0)^4 \left(\frac{1}{\alpha} \operatorname{Var}_t \big(Y_t - X_t' \theta^q_0 \mid Y_t \le X_t' \theta^q_0 \big) + \frac{1-\alpha}{\alpha} \big(X_t' \theta^q_0 - X_t' \theta^e_0 \big)^2 \right) \right]. \label{eqn::AsyCovMatrixC22}
\end{align}
\section{Robustness Check}
\label{sec:robustness_check}
The DGP used by \citet{Gaglianone2011} is a GARCH(1,1) model with standard normally distributed innovations,
\begin{align}
\begin{split}
Y_t & = \sigma_t z_t, \\
\sigma_t^2 & = 0.05 + 0.05 Y_{t-1}^2 + 0.90\sigma_{t-1}^2, \\
z_t & \sim \mathcal{N}\left(0, 1\right).
\end{split}
\end{align}
For this DGP, \Cref{tab:mc1_size_robustness} and \Cref{fig:pauc_robustness} present the empirical sizes and the PAUC analog to the results provided in \Cref{sec::TraditionalSizePower}.
\begin{table}[htb!]\footnotesize\centering
\caption{Empirical sizes of the backtests}
\label{tab:mc1_size_robustness}
\begin{tabularx}{\linewidth}{XX *{8}{r}}
\toprule
\theadl{Nominal \\ Size} & \theadl{Sample \\ Size} & \thead{bivariate \\ ESR (b)} & \thead{bivariate \\ ESR} & \thead{intercept \\ ESR (b)} & \thead{intercept \\ ESR} & \thead{General \\ CC} & \thead{Simple \\ CC} & \thead{Std. \\ ER} & \thead{ER} \\
\midrule
\input{empirical_sizes_Oracle_2_0.01_2s.txt}
\midrule
\input{empirical_sizes_Oracle_2_0.05_2s.txt}
\midrule
\input{empirical_sizes_Oracle_2_0.1_2s.txt}
\bottomrule\addlinespace
\multicolumn{10}{p{.97\linewidth}}{\textit{Notes:} The table reports the empirical sizes of the backtests for a GARCH(1,1)-N process. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\end{tabularx}
\end{table}
\begin{figure}[htb!]
\begin{subfigure}{.5\linewidth}
\includegraphics{roc_design_2_Historical_Simulation_n_1000_2s.pdf}
\caption{Size-adjusted Power}
\end{subfigure}
\begin{subfigure}{.5\linewidth}
\includegraphics{auc_design_2_Historical_Simulation_2s.pdf}
\caption{Partial Area Under the Curve}
\end{subfigure}
\caption{Size-adjusted power and Partial Area Under the Curve plots against Historical Simulation for a sample size of 1000 days. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\label{fig:pauc_robustness}
\end{figure}
\newpage
\section{Raw Power}
\begin{table}[H]\footnotesize\centering
\caption{Raw empirical power of the backtests against Historical Simulation}
\label{tab:mc1_raw_power_Historical_Simulation}
\begin{tabularx}{\linewidth}{XX *{8}{r}}
\toprule
\theadl{Nominal \\ Size} & \theadl{Sample \\ Size} & \thead{bivariate \\ ESR (b)} & \thead{bivariate \\ ESR} & \thead{intercept \\ ESR (b)} & \thead{intercept \\ ESR} & \thead{General \\ CC} & \thead{Simple \\ CC} & \thead{Std. \\ ER} & \thead{ER} \\
\midrule
\input{empirical_power_Historical_Simulation_3_0.01_2s.txt}
\midrule
\input{empirical_power_Historical_Simulation_3_0.05_2s.txt}
\midrule
\input{empirical_power_Historical_Simulation_3_0.1_2s.txt}
\bottomrule\addlinespace
\multicolumn{10}{p{.97\linewidth}}{\textit{Notes:} The table reports the raw empirical power of the backtests against the Historical Simulation for the EGARCH(1,1) process given in (\ref{eq:sim_1_dgp}). The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\end{tabularx}
\end{table}
\begin{figure}[b!]\centering
\begin{subfigure}{.49\linewidth}
\caption{Changing the ARCH parameter}
\label{fig:mc2_rejection_rates_raw_1}
\includegraphics{raw_1_2s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the unconditional variance}
\label{fig:mc2_rejection_rates_raw_2}
\includegraphics{raw_2_2s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the persistence}
\label{fig:mc2_rejection_rates_raw_3}
\includegraphics{raw_3_2s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the degrees of freedom}
\label{fig:mc2_rejection_rates_raw_4}
\includegraphics{raw_4_2s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the probability level}
\label{fig:mc2_rejection_rates_raw_5}
\includegraphics{raw_5_2s}
\end{subfigure}
\caption{Raw rejection rates for various types of misspecification. The gray vertical line depicts the true model. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\label{fig:mc2_rejection_rates_raw}
\end{figure}
\begin{figure}[hb!]\centering
\begin{subfigure}{.49\linewidth}
\caption{Changing the reaction to the squared returns}
\label{fig:mc2_rejection_rates_raw_1_1s}
\includegraphics{raw_1_1s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the unconditional variance}
\label{fig:mc2_rejection_rates_raw_2_1s}
\includegraphics{raw_2_1s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the persistence}
\label{fig:mc2_rejection_rates_raw_3_1s}
\includegraphics{raw_3_1s}
\end{subfigure}
\\[\baselineskip]
\begin{subfigure}{.49\linewidth}
\caption{Changing the degrees of freedom}
\label{fig:mc2_rejection_rates_raw_4_1s}
\includegraphics{raw_4_1s}
\end{subfigure}
\hfill
\begin{subfigure}{.49\linewidth}
\caption{Changing the probability level}
\label{fig:mc2_rejection_rates_raw_5_1s}
\includegraphics{raw_5_1s}
\end{subfigure}
\caption{Raw rejection rates for various types of misspecification with a one-sided hypothesis. The gray vertical line depicts the true model. The number of Monte-Carlo repetitions is 10,000 and the probability level for the risk measures is $\tau=2.5\%$. ESR refers to the backtests introduced in this paper with (b) indicating the bootstrap version, CC to the conditional calibration tests of \citet{Nolde2017}, and ER to the exceedance residuals tests of \citet{McNeil2000}.}
\label{fig:mc2_rejection_rate_raw_1s}
\end{figure}
\end{appendices}
\FloatBarrier
\addcontentsline{toc}{section}{References}
\setstretch{0.8}
\setlength{\bibsep}{5pt}
\bibliographystyle{apalike}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.